Anda di halaman 1dari 496

Advances in Land Remote Sensing

Advances in Land Remote


Sensing

System, Modeling, Inversion and Application

Shunlin Liang
Editor
Department of Geography, University of Maryland,
College Park, MD, USA

123
Shunlin Liang
University of Maryland
College Park, MD
USA

ISBN 978-1-4020-6449-4 e-ISBN 978-1-4020-6450-0

Library of Congress Control Number: 2007940919

°c 2008 Springer Science+Business Media B.V.


No part of this work may be reproduced, stored in a retrieval system, or transmitted in any form or by
any means, electronic, mechanical, photocopying, microfilming, recording or otherwise, without written
permission from the Publisher, with the exception of any material supplied specifically for the purpose of
being entered and executed on a computer system, for exclusive use by the purchaser of the work.

Cover illustration: From sensors and platforms, to information extraction and applications (compilation
by S. Liang)

Printed on acid-free paper

9 8 7 6 5 4 3 2 1

springer.com
Preface

This book is primarily based on presentations in the three reviewing panels of the
9th International Symposium on Physical Measurements and Signatures in Remote
Sensing held at the Institute for Geographical Sciences and Natural Resource Re-
search, Chinese Academy of Sciences, China, in October 2005. It presents a collec-
tion of review papers on remote sensing sensor systems, radiation modeling and
inversion of land surface variables, and remote sensing applications. Each chapter
summarizes the progress in the past few years and also identifies the research issues
for the near future.

v
Acknowledgements

This symposium series is affiliated with the International Society for Photogramme-
try and Remote Sensing (ISPRS) Commission VII/I Working Group on Fundamen-
tal Physics and Modeling led by Professor Michael Schaepman (Chair, Wageningen
University, the Netherlands), Professor Shunlin Liang (co-Chair, University of
Maryland, USA), and Dr. Mathias Kneubuehler (Secretary, University of Zurich,
CH, Switzerland) (2004–2012). It was sponsored and/or financially supported by
the Institute of Geographical Sciences and Natural Resources Research (IGSNRR)
of Chinese Academy of Sciences (CAS), Institute of Remote Sensing Applications
(IRSA) of CAS, Chinese 973 Project “Quantitative Remote Sensing of Major Fac-
tors for Spatio-temporal Heterogeneity on the Land Surface” undertaken by Beijing
Normal University, US National Aeronautics and Space Administration (NASA),
International Society for Photogrammetry and Remote Sensing, IEEE Geoscience
and Remote Sensing Society, and Scientific Data Center for Resources and Environ-
ment, CAS.
This symposium would not be successful without scientific leadership by the In-
ternational Scientific Committee and effective organization by the local Organizing
Committee.

vii
International Scientific Committee

Professor Guanhua Xu, Minister of the Chinese Ministry of Science and Technol-
ogy, China (honorary Chair)
Professor Shunlin Liang, University of Maryland, USA (Chair)
Dr. Frédéric Baret, INRA, Avignon, France
Professor Mike Barnsley, University of Wales Swansea, UK
Professor Marvin Bauer, University of Minnesota, USA
Professor Jon Benediktsson, University of Iceland, Iceland
Professor Jing Chen, University of Toronto, Canada
Professor Peng Gong, University of California at Berkeley, USA
Dr. David Goodenough, Pacific Forestry Centre, Natural Resources Canada
Dr. Tom Jackson, USDA /ARS at Beltsville, Maryland, USA
Dr. David Jupp, CSIRO Earth Observation Centre, Australia
Dr. Yann Kerr, CNES/CESBIO, France
Dr. Marc Leroy, MEDIAS, France
Dr. Philip Lewis, University College London, UK
Professor Deren Li, Wuhan University, China
Professor Xiaowen Li, Beijing Normal University, China
Professor Jiyuan Liu, IGSNRR, CAS, China
Dr. John V. Martonchik, Jet Propulsion Laboratory, USA
Professor Ranga Myneni, Boston University, USA
Professor Ziyuan Ouyang, Institute of Geochemistry, CAS, China
Dr. Jeff Privette, NASA /GSFC, USA
Dr. Jon Ranson, NASA /GSFC, USA
Professor Michael Schaepman, Wageningen University, The Netherlands
Professor Jose Sobrino, University of Valencia, Spain
Dr. Karl Staenz, Canadian Centre for Remote Sensing, Canada
Professor Alan Strahler, Boston University, USA
Professor Qingxi Tong, Institute of Remote Sensing Applications, CAS, China
Dr. Michel Verstraete, JRC, Ispra, Italy
Dr. Charlie Walthall, USDA /ARS at Beltsville, Maryland, USA
Dr. Zhengming Wan, University of California at Santa Barbara, USA

ix
x

Organizing Committee

Professor Jiyuan Liu, Director of IGSNRR, CAS (co-Chair)


Professor Xiaowen Li, Director of Center for Remote Sensing and GIS, Beijing
Normal University and Director of IRSA, CAS (co-Chair)
Professor Dafang Zhuang, Execute Vice-director of Scientific Data Center for
Resources and Environment, CAS (Vice Chair)
Professor Mingkui Cao, IGSNRR, CAS
Professor Changqing Song, Chinese National Foundation of Sciences
Professor Renhua, Zhang, IGSNRR, CAS
Professor Lixin Zhang, Beijing Normal University
Dr. Keping Du, Beijing Normal University
Professor Mengxue Li, National Remote Sensing Center of China
Professor Boqin Zhu, Institute of Remote Sensing Applications, CAS
Dr. Ronggao Liu, IGSNRR, CAS (General Secretary)

Shunlin Liang

University of Maryland, College Park, MD, USA


Reviewers

Each chapter is anonymously reviewed by at least one reviewer. Their valuable com-
ments and suggestions have greatly helped to improve the quality of the volume.
Jing M. Chen
University of Toronto, Canada
Jan Clevers
Wageningen University, The Netherlands
Ruth DeFries
University of Maryland, USA
Alan R. Gillespie
University of Washington, USA
Hongliang Fang
University of Maryland, USA
Xiuping Jia
The University of New South Wales, Australia
David L.B. Jupp
CSIRO Marine and Atmospheric Research, Australia
Yann H. Kerr
CNES/CESBIO, France
Yuri Knyazikhin
Boston University, USA
Randy Koster
NASA, USA
Eric F. Lambin
University of Louvain, Belgium
Tiit Nilson
Tartu Observatory, Estonia

xi
xii Reviewers

Paolo Pampaloni
IFAC-CNR, Italy
Bernard Pinty
EC Joint Research Centre, Italy
Jeff Privette
NOAA, USA
Contents

1 Recent Advances in Land Remote Sensing: An Overview . . . . . . . . . . 1


Shunlin Liang

Part I Remote Sensing Systems

2 Passive Microwave Remote Sensing for Land Applications . . . . . . . . . 9


Thomas J. Jackson

3 Active Microwave Remote Sensing Systems and Applications


to Snow Monitoring . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
Jiancheng Shi

4 Multi-angular Thermal Infrared Observations


of Terrestrial Vegetation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
Massimo Menenti, Li Jia, and Zhao-Liang Li

5 Terrestrial Applications of Multiangle Remote Sensing . . . . . . . . . . . . 95


Mark J. Chopping

Part II Physical Modeling and Inversion Algorithms

6 Modeling the Spectral Signature of Forests: Application of Remote


Sensing Models to Coniferous Canopies . . . . . . . . . . . . . . . . . . . . . . . . . 147
Pauline Stenberg, Matti Mõttus, and Miina Rautiainen

7 Estimating Canopy Characteristics from Remote Sensing


Observations: Review of Methods and Associated Problems . . . . . . . . 173
Frédéric Baret and Samuel Buis

8 Knowledge Database and Inversion . . . . . . . . . . . . . . . . . . . . . . . . . . . . 203


Jindi Wang and Xiaowen Li

xiii
xiv Contents

9 Retrieval of Surface Albedo from Satellite Sensors . . . . . . . . . . . . . . . . 219


Crystal Schaaf, John Martonchik, Bernard Pinty, Yves Govaerts,
Feng Gao, Alessio Lattanzio, Jicheng Liu, Alan Strahler, and Malcolm
Taberner
10 Modeling and Inversion in Thermal Infrared Remote Sensing
over Vegetated Land Surfaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 245
Frédéric Jacob, Thomas Schmugge, Albert Olioso, Andrew French,
Dominique Courault, Kenta Ogawa, Francois Petitcolin,
Ghani Chehbouni, Ana Pinheiro, and Jeffrey Privette
11 Spectrally Consistent Pansharpening . . . . . . . . . . . . . . . . . . . . . . . . . . . 293
Ari Vesteinsson, Henrik Aanaes, Johannes R. Sveinsson,
and Jon Atli Benediktsson
12 Data Assimilation Methods for Land Surface Variable Estimation . . 313
Shunlin Liang and Jun Qin
13 Methodologies for Mapping Land Cover/Land Use and its Change . . 341
Nina Siu-Ngan Lam
14 Methodologies for Mapping Plant Functional Types . . . . . . . . . . . . . . 369
Wanxiao Sun and Shunlin Liang

Part III Remote Sensing Applications

15 Monitoring and Management of Agriculture with Remote Sensing . . 397


Zhongxin Chen, Sen Li, Jianqiang Ren, Pan Gong, Mingwei Zhang,
Limin Wang, Shenliang Xiao, and Daohui Jiang
16 Remote Sensing of Terrestrial Primary Production
and Carbon Cycle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 423
Maosheng Zhao and Steven W. Running
17 Applications of Terrestrial Remote Sensing to Climate Modeling . . . . 445
Robert E. Dickinson
18 Improving the Utilization of Remotely Sensed Data . . . . . . . . . . . . . . . 465
John R. Townshend, Stephen Briggs, Roy Gibson, Michael Hales,
Paul Menzel, Brent Smith, Yukio Haruyama, Chu Ishida, John Latham,
Jeff Tschirley, Deren Li, Mengxue Li, Liangming Liu,
and Gilles Sommeria
19 Emerging Issues in Land Remote Sensing . . . . . . . . . . . . . . . . . . . . . . . 485
Shunlin Liang, Michael Schaepman, Thomas J. Jackson, David Jupp,
Xiaowen Li, Jiyuan Liu, Ronggao Liu, Alan Strahler,
John R. Townshend, and Diane Wickland

Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 495

CD-ROM included
Contributors

Henrik Aanaes
Informatics and Mathematical Modelling, Technical University of Denmark,
Denmark
haa@imm.dtu.dk
Frederic Baret
UMR1114, INRA-CSE, 84914 Avignon, France
baret@avignon.inra.fr
Jon Atli Benediktsson
Department of Electrical and Computer Engineering, University of Iceland,
Hjardarhaga 2-6, 107 Reykjavik, Iceland
benedikt@hi.is
Stephen Briggs
European Space Agency, Via Galileo Galilei, 00644 Frascati, Rome, Italy
stephen.briggs@esa.int
Sanuel Buis
UMR1114, INRA-CSE, 84914 Avignon, France
Ghani Chehbouni
Institute of Research for the Development, Center for Spatial Studies of the
Biosphere, UMR CESBio, Toulouse, France
ghani.chehbouni@cesbio.cnes.fr
Zhongxin Chen
Key Laboratory of Resource Remote Sensing & Digital Agriculture, Ministry
of Agriculture, Beijing 100081, China
chzhx@yahoo.com
Mark J. Chopping
Department of Earth and Environmental Studies, Montclair State University,
1 Normal Ave, Montclair, NJ 07043, USA
chopping@pegasus.montclair.edu

xv
xvi Contributors

Dominique Courault
National Institute for Agronomical Research, Climate – Soil – Environment Unit,
UMR CSE INRA / UAPV, 84914 Avignon, France
courault@avignon.inra.fr
Robert E. Dickinson
School of Earth and Atmospheric Sciences, Georgia Institute of Technology,
311 Ferst Drive, Atlanta, GA 30332-0340, USA
robted@eas.gatech.edu
Andrew French
United States Department of Agriculture/Agricultural Research Service, US Arid
Land Agricultural Research Center, 21881 North Cardon Lane, Maricopa,
AZ 85238, USA
andrew.french@ars.usda.gov
Feng Gao
Earth Resources Technology, Inc., 8106 Stayton Dr., Jessup, MD 20794, USA
Roy Gibson
EUMETSAT
roy.gibson@wanadoo.fr
Pan Gong
Institute of Agricultural Resources & Regional Planning, Chinese Academy
of Agricultural Sciences, Beijing 100081, China
Yves Govaerts
EUMETSAT, Am Kavalleriesand 31, D-64295 Darmstadt, Germany
yves.govaerts@eumetsat.int
Michael Hales
NOAA, Silver Spring, MD 20910, USA
michael.hales@noaa.gov
Yukio Haruyama
JAXA, 1-8-10 Harumi, Chuo-ku, Tokyo 104-6023, Japan
haruyama.yukio@jaxa.jp
Chu Ishida
JAXA, 1-8-10 Harumi, Chuo-ku, Tokyo 104-6023, Japan
ishida.chu@jaxa.jp
Thomas J. Jackson
USDA ARS Hydrology and Remote Sensing Lab, 104 Bldg. 007 BARC-West,
Beltsville, MD 20705,
tjackson@hydrolab.arsusda.gov
Frédéric Jacob
Institute of Research for the Development, Laboratory for studies on Interactions
between Soils – Agrosystems – Hydrosystems, UMR LISAH SupAgro/INRA/IRD,
Montpellier, France
frederic.jacob@supagro.inra.fr
Contributors xvii

Li Jia
Alterra Green World Research, Wageningen University and Research Centre,
The Netherlands
li.jia@wur.nl
Daohui Jiang
Institute of Agricultural Resources & Regional Planning, Chinese Academy
of Agricultural Sciences, Beijing 100081, China
David Jupp
CSIRO Marine and Atmospheric Research, Canberra ACT 2601 Australia
David.Jupp@csiro.au
Nina Siu-Ngan Lam
Department of Environmental Studies, Louisiana State University, Baton Rouge,
LA 70808, USA
nlam@lsu.edu
John Latham
Food and Agriculture Organization, Rome, Italy
john.latham@fao.org
Alessio Lattanzio
Makalumedia gmbh, Robert-Bosch Strasse 7, 64296 Darmstadt, Germany
Deren Li
Wuhan University, 39 Loyu Road, Wuhan, 430070, China
dli@wtusm.edu.cn
Mengxue Li
NRSCC, 15B, Fuxing Road, Beijing, 100862, China
mengxueli@hotmail.com
Sen Li
Institute of Agricultural Resources & Regional Planning, Chinese Academy
of Agricultural Sciences, Beijing 100081, China
Xiaowen Li
Research Center for Remote Sensing and GIS, Beijing Normal University, No.19
XieJieKouWaiDaJie Street, Beijing 100875, China
lix@bnu.edu.cn
Zhao-Liang Li
Institute of Geographic Sciences and Natural Resources Research, Beijing, China
Shunlin Liang
Department of Geography, University of Maryland, College Park, USA
sliang@umd.edu
Jicheng Liu
Department of Geography and Environment, Boston University,
675 Commonwealth Ave., Boston, MA 02215, USA
xviii Contributors

Jiyuan Liu
Institute for Geographical Sciences and Natural Resource Research, Chinese
Academy of Sciences, Beijing, China
liujy@igsnrr.ac.cn
Liangming Liu
National Remote Sensing Center of China
Ronggao Liu
Institute for Geographical Sciences and Natural Resource Research, Chinese
Academy of Sciences, Beijing, China
liurg@lreis.ac.cn
John Martonchik
Jet Propulsion Laboratory, Mail Stop 169-237, 4800 Oak Grove Dr., Pasadena,
CA 91109, USA
john.v.martonchik@jpl.nasa.gov
Massimo Menenti
TRIO/LSIIT, University Louis Pasteur (ULP), Strasbourg, France and Istituto per i
Sistemi Agricoli e Forestali del Mediterraneo (ISAFOM), Naples, Italy
m.menenti@isafom.cnr.it
Paul Menzel
University of Wisconsin, Space Science and Engineering Center, Madison,
WI 53706, USA
paulm@ssec.wisc.edu
Matti Mõttus
Department of Forest Ecology, FI-00014 University of Helsinki, Finland; Tartu
Observatory, 61602 Tõravere, Tartumaa, Estonia
mottus@ut.ee
Kenta Ogawa
Department of Geo-system Engineering, University of Tokyo and Hitachi Ltd,
Tokyo, Japan
Albert Olioso
National Institute for Agronomical Research, Climate – Soil – Environment Unit,
UMR CSE INRA/UAPV, Avignon, France
olioso@avignon.inra.fr
Francois Petitcolin
ACRI-ST, Sophia Antipolis, France
ptc@acri-st.fr
Ana Pinheiro
Biospheric Sciences Branch, NASA’s GSFC, Greenbelt, MD 20771, USA
ana.pinheiro@gsfc.nasa.gov
Contributors xix

Bernard Pinty
Global Environment Monitoring Unit, IES, EC Joint Research Centre, TP 440,
via E. Fermi, I-21020 Ispra (VA), Italy
bernard.pinty@jrc.it
Jeffrey Privette
NOAA’s National Climatic Data Center, Asheville, NC 28801-5001, USA
jeffrey.privette@noaa.gov
Jun Qin
Institute for Geographical Science and Natural Resource Research, Beijing, China
shuairenqin@gmail.com
Miina Rautiainen
Department of Forest Resource Management, FI-00014 University of Helsinki,
Finland
miina.rautiainen@helsinki.fi
Jianqiang Ren
Key Laboratory of Resource Remote Sensing & Digital Agriculture, Ministry of
Agriculture, Beijing 100081, China and Institute of Agricultural Resources &
Regional Planning, Chinese Academy of Agricultural Sciences, Beijing 100081,
China
Steven W. Running
Numerical Terradynamic Simulation Group, Department of Ecosystem and
Conservation Science, University of Montana, Missoula, MT 59812, USA
swr@ntsg.umt.edu
Crystal Schaaf
Department of Geography and Environment, Boston University,
675 Commonwealth Ave., Boston, MA 02215, USA
schaaf@bu.edu
Michael Schaepman
Centre for Geo-Information, Wageningen University, Wageningen, The Netherlands
Michael.Schaepman@wur.nl
Thomas Schmugge
Gerald Thomas Professor of Water Resources, College of Agriculture New Mexico
State University, Las Cruces, NM, USA
schmugge@nmsu.edu
Jiancheng Shi
Institute for Computational Earth System Science, University of California,
Santa Barbara, CA 93106-3060, USA
shi@icess.ucsb.edu
Brent Smith
NOAA, Silver Spring, MD 20910, USA
brent.smith@noaa.gov
xx Contributors

Gilles Sommeria
World Climate Research Programme, WMO, CH-1211 Geneva, Switzerland
gsommeria@wmo.int
Pauline Stenberg
Department of Forest Resource Management, FI-00014 University of Helsinki,
Finland
Pauline.Stenberg@helsinki.fi
Alan Strahler
Department of Geography and Environment, Boston University,
675 Commonwealth Ave., Boston, MA 02215, USA
alan@bu.edu
Wanxiao Sun
Department of Geography and Planning, Grand Valley State University, USA
sunwa@gvsu.edu
Johannes R. Sveinsson
Department of Electrical and Computer Engineering, University of Iceland,
Hjardarhaga 2-6, 107 Reykjavik, Iceland
sveinsso@hi.is
Malcolm Taberner
Global Environment Monitoring Unit, IES, EC Joint Research Centre, TP 440,
via E. Fermi, I-21020 Ispra (VA), Italy
malcolm.taberner@jrc.it
John R. Townshend
Department of Geography, University of Maryland, College Park, MD 20742, USA
jtownshe@umd.edu
Jeff Tschirley
Food and Agriculture Organization, 00153 Rome, Italy
jeff.tschirley@fao.org
Ari Vesteinsson
Department of Electrical and Computer Engineering, University of Iceland,
Hjardarhaga 2-6, 107 Reykjavik, Iceland
Jindi Wang
Research Center for Remote Sensing and GIS, Beijing Normal University, No.19
XieJieKouWaiDaJie Street, Beijing 100875, China
wangjd@bnu.edu.cn
Limin Wang
Key Laboratory of Resource Remote Sensing & Digital Agriculture, Ministry
of Agriculture, Beijing 100081, China
Diane Wickland
NASA Headquarters, Washington, DC, USA
diane.e.wickland@nasa.gov
Contributors xxi

Shenliang Xiao
Institute of Agricultural Resources & Regional Planning, Chinese Academy
of Agricultural Sciences, Beijing 100081, China
Mingwei Zhang
Institute of Agricultural Resources & Regional Planning, Chinese Academy
of Agricultural Sciences, Beijing 100081, China
Maosheng Zhao
Numerical Terradynamic Simulation Group, Department of Ecosystem
and Conservation Science, University of Montana, Missoula, MT 59812, USA
zhao@ntsg.umt.edu
Chapter 1
Recent Advances in Land Remote Sensing:
An Overview

Shunlin Liang

Earth’s surface is undergoing rapid changes due to urbanization, industrialization


and globalization. Environmental problems such as water shortages, desertification,
soil depletion, greenhouse gas emissions warming the atmosphere, deforestation,
elevated coastal waterway sediment and nutrient fluxes, among other environmental
problems, are increasingly common and troubling consequences of human activ-
ities. Policy decisions about the environment rely on accurate and reliable infor-
mation, especially data and understanding leading to better predictions of natural
hazards, epidemics, impacts of energy choices, and climate variations. Comprehen-
sive, systematic Earth observations are key to forecasting Earth system dynamics.
Predicting future scenarios of our planet’s habitability requires analysis of what has
transpired in the past along with observations of present conditions and processes.
Timely, quality long-term global data acquired through remote sensing is essential
for the ongoing viability and enhancement of human society on Earth.
The field of remote sensing (Earth observation) has developed rapidly. Many
publications have documented its progress (e.g., the special issue of Remote Sens-
ing Reviews with a set of papers reviewing the modeling and inversion of surface
bidirectional reflectance distribution function (BRDF) (Liang and Strahler, 2000),
and the edited or authored books on similar subjects (Liang, 2004; Myneni and Ross,
1991). To systematically summarize the achievements of terrestrial remote sensing
in recent years and to set the research agenda for the near future, the 9th Inter-
national Symposium on Physical Measurements and Signatures in Remote Sensing
(ISPMSRS), held in October 2005 in Beijing, organized three review panels. The
papers compiled in this book are largely from these panels and are organized into
three parts, respectively.
Part I of the book (corresponding to the first panel, chaired by Drs. David Jupp
and Tom Jackson, with Drs. Ralph Dubayah, Michael Schaepman, Jianchen Shi,
Stephen Ungar, and David Le Vine) focuses on remote sensing systems and sen-
sors. As there are many different remote sensing systems (the result of various

Shunlin Liang
University of Maryland, USA

S. Liang (ed.), Advances in Land Remote Sensing, 1–6. 1


c Springer Science + Business Media B.V., 2008
2 S. Liang

permutations of passive vs. active, specific domain of the electromagnetic spec-


trum sampled, and spectral channel bandwidth) currently in operation or planned to
be operational in the future, this panel evaluated the capabilities of these systems
for estimating key land surface variables and how they can best be improved and
combined effectively. The panel presentations covered microwave, thermal-infrared
(IR), hyperspectral optical, and Lidar remote sensing. Based on these discussions,
four chapters are compiled on this subject, including passive microwave, active
microwave, multiangle thermal IR, and multiangle optical remote sensing.
In Chapter 2 on passive microwave remote sensing, Jackson identifies six fac-
tors in passive microwave sensor design that affect the retrieval of land surface
properties: frequency, polarization, view geometry, spatial resolution, temporal cov-
erage, and signal-to-noise ratio. While summarizing the features of three current
relevant satellite instruments and two other satellite sensors, he points out that the
low frequency observations (<6 GHz) and coarse spatial resolution are obstacles-
to-land applications. He further discusses three approaches, currently under investi-
gation, for solving these problems. The first is synthetic aperture radiometry, such
as the planned Soil Moisture Ocean Salinity mission. The second is use of large,
lightweight antennas and the Hydros satellite mission (recently cancelled) which
could have demonstrated this approach. The third is disaggregation by integrating
with higher resolution visible to thermal remote sensing data through data fusion.
The next decade will see several exploratory missions using new technologies,
and innovative approaches to integrating passive microwave with active microwave
measurements.
Shi summarizes, in Chapter 3 on active microwave remote sensing, the significant
developments in mapping snow, inferring snow wetness and snow water equivalence
using active microwave sensors (mainly synthetic aperture radar – SAR). Following
a brief introduction to different SAR systems, including current and future systems,
he discusses three types of SAR measurements: backscattering measurements at a
given frequency and polarization, polarization properties obtained by a fully polari-
metric instrument, and interferometry from repeat-pass sampling. Shi devotes more
space to reviewing various techniques that use these three measurements to map
snow properties, followed by a discussion of the need for future SAR systems to
map snow properties.
Multiangle thermal-IR remote sensing is covered in Chapter 4. Menenti et al.
first present experimental evidence of angular signatures of thermal-IR observa-
tions, then demonstrate that two soil and foliage temperatures normally can account
for most thermal IR observations at the top of the canopy. (further division into
sunlit and shadowed components are needed only under extreme conditions). They
review various geometric-optics and radiative transfer modeling techniques, and
finally describe inversion techniques to estimate component temperatures from mul-
tiangle thermal-IR observations. The related subject on estimating land surface skin
temperature is given in Chapter 10.
In Chapter 5 on multiangle optical remote sensing, Chopping first outlines dif-
ferent sensor systems for acquiring multiangle data, and then reviews a variety of
land applications using multiangle observations, such as mapping canopy openness,
1 Recent Advances in Land Remote Sensing: An Overview 3

clumping index, structural scattering index, land cover and community type, snow
and ice, and dust emissions. He also outlines different modeling techniques for
angular signatures in the optical domain and discusses other factors (e.g., near-
simultaneous and accumulated angular sampling, scale) related to multiangle data
acquisition.
Although these four chapters cover a variety of sensors and applications, two
other remote sensing techniques deserve mention. Lidar has a unique ability to mea-
sure vertical and spatial heterogeneity across multiple scales. Many studies have
demonstrated that lidar can be used for estimating forest structure variables such as
canopy height and profile, canopy cover, growth/successional state, foliar profile,
stem density, basal area, biomass, and canopy base height. Except for IceSat, cur-
rently there are no spaceborne lidar remote sensing systems available. Hyperspectral
remote sensing is another approach that has great potential.
A more challenging issue faced by the remote sensing community is how to
integrate data from different sensor systems in such a way that derivation of land
surface variables is maximized.
Part II of the book presents surface radiation modeling and inversion methods
corresponding to the second panel chaired by Drs. Alan Strahler and Xiaowen Li,
and contributed by Drs. Frédéric Baret, Frédéric Jacob, Philip Lewis, Massimo
Menenti, Pauline Stenberg, and Jindi Wang. There have been considerable invest-
ments on developing physical models to understand the surface radiation regimes.
Some of these models have been incorporated into very useful algorithms for
estimating land surface variables from satellite observations. This panel assessed
the progress of surface radiation modeling and satellite inversion algorithm devel-
opment in many different aspects. The papers in this part are composed of panel
presentations and two additional contributions. Note that Chapters 4 and 5 also con-
tain modeling and inversion mainly for multiangle observations.
Stenberg et al. in Chapter 6 describe different geometric-optical models and ra-
diative transfer models for simulating the reflectances of heteorogenerous canopies,
particularly forests. In radiative transfer modeling, they review mainly the three-
dimensional models. They also discuss the structural and spectral characteristics of
coniferous forests, the clumping effects and reflectance scaling using the p-theory,
and the inputs for simulating the forest reflectance.
Baret and Buis in Chapter 7 review various methods for estimating canopy prop-
erties from remotely sensed data. They first describe the empirical regression-type
algorithms, using either experimental data or radiative transfer simulated data, and
different model inversion algorithms, including optimization algorithms, look-up
table algorithms, and two stochastic inversion algorithms. They then discuss the
nature of ill-posed problem in the inversion of canopy properties, and finally sum-
marize two methods to address the ill-posed issue by using the a priori information
or posing additional spatial, temporal and model coupling constraints.
Wand and Li in Chapter 8 present a knowledge database and discuss its applica-
tions to improving the inversion of land surface information from remotely sensed
data. They first discuss the need for a prior knowledge due to the ill-posed problem
in remote sensing inversion that is also discussed by Baret and Buis in Chapter 7,
4 S. Liang

then describe the database they have developed and finally demonstrate how it has
been used in different inversion studies. They also address the parameter scaling and
validation issues.
Schaaf et al. in Chapter 9 summarize the current algorithms for estimating
land surface broadband albedo and anisotropy from satellite observations. After
providing some general background, they present the specific algorithms for
three representative sensors: MODIS (multispectral), MISR (multiangle) and Me-
teosat (geostationary) and also discuss how to integrate historic and current albedo
products.
In Chapter 10, Jacob et al. comprehensively review various modeling methods
of thermal signatures and inversion techniques for estimating land surface skin tem-
perature from thermal-IR remotely sensed data. They start with the definitions of
temperatures and emissivities, and analysis of spatial, spectral, angular and temporal
signatures in thermal-IR remote sensing. Various modeling techniques and inversion
methods are then evaluated. They further examine current research issues, including
land surface brightness temperature from atmospheric correction or model simula-
tion, ensemble emissivity and radiometric temperature of mixed surface types from
different observation capabilities, aerodynamic temperature, and directional effects
of temperature and emissivity. The potential applications of thermal-IR remote sens-
ing are also discussed.
Vesteinsson et al. in Chapter 11 present a new method to integrate remote sensing
data of different spatial resolutions. After reviewing various image fusion methods
in four major types (Frequency, color transformation, statistical, and hierarchical),
they present a new data fusion framework. They model the fusion issue as a mini-
mization problem where the objective function is the energy of a Gibbs distribution
consisting of three energy terms: spectral consistency constraint, smoothness con-
straint, and imaging physics constraint. Minimization of the objective function is
sought using stochastic optimization. This framework is finally implemented using
two IKONOS satellite images by integrating the panchromatic band with multispec-
tral bands.
Data assimilation methods for estimating land surface variables are presented
by Liang and Qin in Chapter 12. After introducing the basic principles of data
assimilation, they identify a series of critical issues, such as data and products to
be assimilated, parameters to be estimated, assimilation algorithms to be used, error
matrices to be determined, and imperfect numerical models. The latest applications
of data assimilation methods to various fields (e.g., hydrology, carbon cycle, agri-
cultural productivity) are also reviewed.
Lam in Chapter 13 introduces the use of textural/spatial measures to automated
land cover classification and change detection. She first identifies the major cri-
teria for evaluating textural measures and then shows different examples in using
them to improve classification accuracy. She also summarizes the existing meth-
ods into a framework, and then argues that the textural approach has potential for
rapid change detection. She finally demonstrates the use of textural measures solely
(not including spectral information) for change detection in New Orleans before and
after Hurricane Katrina.
1 Recent Advances in Land Remote Sensing: An Overview 5

In Chapter 14, Sun and Liang review the existing methods for mapping plant
functional types (PFTs) at regional to global scale from remote sensing data. Tra-
ditionally, land surface models represent vegetation as discrete biomes that are not
natural vegetation units but are products of classification. Since most land models
are expanding beyond their traditional biogeophysical roots to include biogeochem-
istry, especially photosynthesis and the carbon cycle, the land modeling community
has started using PFTs to represent land surface. There is an urgent need for map-
ping PFT using remote sensing. The authors finally present a multisource evidential
reasoning data fusion framework using high-level land products for improved map-
ping of PFTs from satellite observations.
These chapters in Part II of the book cover a variety of topics related to land
surface radiation modeling and inversion. I must emphasize the importance of vali-
dation here. The new modeling techniques, new inversion methods, and newly gen-
erated products require thorough validation. Though a specific chapter on validation
is not presented here, a recently published special issue on this topic (Morisette
et al., 2006) provides many relevant details.
Part III of the book discusses remote sensing applications corresponding to the
last panel chaired by Drs. John Townshend and Jiyuan Liu, and with contributions
by Drs. Zhongxin Chen, David Goodenough, and Diane Wickland. Remote sensing
science driven by applications will have more widespread use and benefit larger user
communities. Significant disconnections between remote sensing development and
applications persist. Some products developed by remote sensing scientists have
not been widely utilized. Many variables required by surface process models and
decision support systems have not been generated. Product accuracy and application
requirements may not always be consistent. Even the same variables may be defined
differently in remote sensing algorithms and application models. This panel pro-
vided a useful dialogue between remote sensing scientists and application experts.
This part of the book includes some of the discussions and also presents two more
chapters on this topic.
Chen et al. in Chapter 15 review the remote sensing applications pertinent to agri-
cultural monitoring and management. They discuss progress in four subjects: crop
identification and mapping, crop yield estimation and prediction, crop phenology
monitoring, and soil moisture estimation from optical to microwave remote sensing.
They demonstrate the need for multisensor data fusion, use of multiple signatures,
and integration of data and numerical models.
Zhao and Running in Chapter 16 review the historical development and the recent
advances in the application of satellite remote sensing data for estimating terrestrial
productivity and monitoring carbon cycle-related ecosystem dynamics and changes.
They first outline the development of using vegetation index for estimating land sur-
face biophysical variables, then describe the MODIS GPP and NPP products and the
findings from the vegetation-related products. They further discuss the application
of long-term satellite data to the study of terrestrial ecosystems, including phenology
monitoring, changes in regional carbon storage resulting from land use change, car-
bon flux changes induced by climate change, disturbance detection, and validation
of ecosystem models. They also propose a scheme for an integrated study of carbon
dynamics.
6 S. Liang

In Chapter 17, Dickinson discusses applications of terrestrial remote sensing to


climate modeling. Beginning with an introduction to the formulation of the climate
models, he then describes land surface models, their components, and energy and
water balance requirements. He reviews the contributions from many land remote
sensing products (e.g., LAI, albedo and temperature) in terms of roles of solar
radiation in climate model terrestrial system, and summarizes various assessments
of the climate models when using these products. He further considers how terres-
trial remote sensing can better support climate models and eventually be a compo-
nent of climate prediction through data assimilation.
Townshend et al. in Chapter 18 identify and analyze factors either facilitating or
hampering the increased use of satellite data, products and services by compiling
25 case studies. These factors include technical, education and capacity building,
financial and policy. They discover that successful adoption of remote sensing prod-
ucts always involves balanced cooperation between space agencies and users. They
derive a set of principles that will lead to enhanced usage of space-based data and
products and summarize the recommendations to the space agencies.
Liang et al. in Chapter 19 summarize the key questions and issues discussed
by three review panels and identify some emerging issues in land remote sensing,
including sensor networks, modeling complex landscapes, machine learning tech-
niques for inversion, and spatial scaling.
Finally, note that two special issues in Photogrammetric Engineering and Remote
Sensing (October 2007), and Journal of Remote Sensing (September 2007) from the
9th ISPMSRS add to and complement the review papers in this book. Readers are
referred to these special issues for details.

References

Liang S (2004) Quantitative Remote Sensing of Land Surfaces. Wiley, New York, 534pp
Liang S, Strahler A (guest eds) (2000) Land surface bi-directional reflectance distribution function
(BRDF): recent advances and future prospects. Special issue of Remote Sens. Rev. 18:83–511
Morisette J, Baret F, Liang S (guest eds) (July 2006) Global land product validation. Special issue
of IEEE Trans. Geosci. Remote Sens. 44(7):1695–1937
Myneni RB, Ross J (eds) (1991) Photon-vegetation Interactions: Applications in Optical Remote
Sensing and Plant Physiology. Springer, New York
Chapter 2
Passive Microwave Remote Sensing for Land
Applications

Thomas J. Jackson

Abstract Land applications, in particular soil moisture retrieval, have been ham-
pered by the lack of low frequency passive microwave observations and the coarse
spatial resolution of existing sensors. The next decade could see several improved
operational and exploratory missions using new technologies as well as innovative
disaggregation and data fusion approaches that could lead the way to an order of
magnitude improvement in spatial resolution.

Keywords: Passive microwave · soil moisture · land surface

2.1 Introduction

Passive microwave remote sensing has made major contributions in atmospheric and
oceanic sciences. These applications have exploited higher frequencies and used low
frequencies to establish background conditions. Land applications have been ham-
pered by the availability of low frequency observations (<6 GHz) and coarse spatial
resolution. Conventional technologies and approaches to retrievals have limited spa-
tial resolution to the 50 + km range, which has in turn limited the potential usage
to only very large-scale studies. Three factors will affect land applications of pas-
sive microwave remote sensing over the next decade: the operational low frequency
instruments available, exploratory missions using new technologies, and innovative
approaches to disaggregating the coarser resolution passive microwave observations
with active microwave measurements. The current satellite sensors, criteria for mis-
sion design, and future missions are presented in this chapter. Because soil moisture
is one of the most important and challenging problems the discussion will focus on
this topic.

Thomas J. Jackson
USDA ARS Hydrology and Remote Sensing Lab, Beltsville, MD, USA
tjackson@hydrolab.arsusda.gov

S. Liang (ed.), Advances in Land Remote Sensing, 9–18. 9


c Springer Science + Business Media B.V., 2008
10 T.J. Jackson

2.2 Satellite Passive Microwave Sensor Systems

Low frequency passive microwave remote sensing provides information on the


dielectric and temperature properties of the Earth’s atmosphere and surface as well
as some characterization of its geometric features. The dielectric properties are
dependent upon the water content of the target. As a result, passive microwave re-
mote sensing has been useful in studies of the atmosphere, oceans, snow, ice, and
land. The breadth of parameters and variables that can be retrieved is illustrated in
Table 2.1 for the Advanced Microwave Scanning Radiometer (AMSR-E) instrument
on the National Space and Aeronautics Administration (NASA) Aqua satellite.
Land applications of passive microwave remote sensing have included classifica-
tion, temperature, vegetation characteristics, and soil moisture. Current operational
sensors such as the Special Sensor Microwave/Imager (SSM/I) have been used in
land cover type classification (Neale et al., 1990), land surface temperatures estima-
tion and in deriving a general surface wetness parameter (Basist et al., 1998).
System design for land variable retrieval includes at least six considerations:
• Frequency
• Polarization
• Viewing geometry
• Spatial resolution
• Temporal coverage
• Signal to Noise

2.2.1 Frequency Selection

Lower frequencies provide greater sensitivity to changes in soil moisture through


a greater range of vegetation cover conditions. For soil moisture retrieval these are
very important considerations (Jackson and Schmugge, 1991). Figure 2.1 illustrates
this point using a sensitivity parameter (degrees K/% soil moisture) and identifies
several key satellite instruments. For bare soil there is little change in sensitivity as a

Table 2.1 Earth variables derived from AMSR-E


Category Variable

Atmosphere Total integrated water vapor over the ocean


Rainfall over ocean
Oceans Sea surface temperature
Ocean surface wind speed
Sea ice Sea ice concentration
Snow depth over sea ice
Sea ice temperature
Snow Snow-cover water equivalent over land
Land Surface soil wetness
2 Passive Microwave Remote Sensing for Land Applications 11

Fig. 2.1 Sensitivity of High


brightness temperature to SMOS AMSR-E Meteorological
Satellites
soil moisture as a function of
microwave frequency
Bare

Sensitivity
Vegetated

Low
1 2 3 5 10 20 30 50
Frequency (GHz)

function of frequency (although) there is a decrease in sensing depth with increasing


frequency. In the presence of vegetation the impact of frequency is much more sig-
nificant. Surface roughness can also decrease sensitivity. Another aspect of Fig. 2.1
to note is that research and applications were limited to the lowest frequencies of the
meteorological satellites for many years. When AMSR-E and other recent satellite
instruments were launched a significant improvement in soil moisture retrieval was
anticipated.
Another key issue in frequency selection is radio frequency interference. This
has become a significant problem with AMSR-E, primarily at 6.9 GHz in the USA.
Contamination of the signal has rendered the data unusable and as a result applica-
tions have had to adapt to using a less desirable higher frequency (Li et al., 2004;
Njoku et al., 2005). As shown in Fig. 2.1, the shift in frequency from 6.9 GHz to
10.7 GHz for soil moisture retrieval results in a reduction in sensitivity and loss of
information. It should be noted that the 6.9 GHz band is not a protected radio fre-
quency, however, 10.7 GHz is protected and still is contaminated in some regions of
the world.

2.2.2 Polarization

It is well known that horizontal polarization is much more sensitive to changes in


the soil dielectric constant than vertical polarization measurements. Simple com-
putations using radiative transfer equations can readily verify this fact (Ulaby et al.,
1982). However, vertical polarization can be very useful in normalizing the horizon-
tal measurements. Information on the polarization difference has been used to char-
acterize the vegetation in several studies and algorithms (Becker and Choudhury,
1988; Paloscia et al., 2001; Owe et al., 2001; Njoku et al., 2003). With the recent
launch of the WindSat instrument (Gaiser et al., 2004) with a fully polarimetric mi-
crowave radiometer, investigators are beginning to examine what new information
about the land may be contained in the additional channels the sensor provides
(Narvekar et al., 2007).
12 T.J. Jackson

2.2.3 Viewing Geometry

Nearly all satellite microwave radiometers used for land studies have employed
conical scanning. Conical scanning provides constant incidence angle observa-
tions, which simplifies the retrieval process. Higher incidence angles result in wider
swaths and increased temporal coverage. So, there are good reasons for choosing
this design.
However, the highest sensitivity to soil moisture will be at low incidence angles.
Radiative transfer relationships include the angle in computation. In addition, the
path length through the vegetation canopy will increase with angle. At low incidence
angles the path attenuation will be smaller. Some algorithm designs actually exploit
these changes with incidence angle in their retrievals (Wigneron et al., 2000). In
this approach, a correction for vegetation is made in performing the soil moisture
retrieval by using several incidence angle observations.

2.2.4 Spatial Resolution

Spatial resolution has been the challenge for passive microwave remote sensing
of land. Unlike atmospheres and oceans, the land exhibits heterogeneity at several
scales that may be significant. With low frequencies it is extremely difficult to de-
sign an antenna system necessary for a high resolution observing system. This topic
will be covered in depth in a later section.

2.2.5 Temporal Coverage

With conical scanning low-resolution sensors, the temporal coverage has not been a
major issue for applications. Wide swaths of current and future sensors result in 1–3
day global coverage.

2.2.6 Signal to Noise

To date, using conventional antenna technology, signal to noise has not been an issue
for land applications (soil moisture) because the range of response and sensitivity to
the geophysical parameters are quite large.
2 Passive Microwave Remote Sensing for Land Applications 13

2.3 Selected Conical Scanning Instruments

Three current satellite instruments are of interest in passive microwave sensing of


land; SSM/I, AMSR-E, and WindSat. Features of each mission are summarized in
Table 2.2. There have been several other important satellites for land remote sensing
that include the Scanning Multifrequency Microwave Radiometer (SMMR) (Owe
et al., 1992) and the Tropical Rainfall Measurement Mission Microwave Imager
(TMI) (Bindlish et al., 2003).
Worth noting in Table 2.2 is the fact that the SSM/I instrument now has an over
20-year period of record. Interpreting data from the SSM/I to extract surface infor-
mation requires accounting for atmospheric effects on the measurement. When one
considers the atmospheric correction, the significance of vegetation attenuation, and
the shallow contributing depth of soil for these high frequencies, it becomes appar-
ent that the data are of limited value for estimating soil water As a result, SSM/I
data has been applied to only a limited set of land applications including a soil wet-
ness product (Basist et al., 1998). Some successful attempts have been made at soil
moisture retrieval under selected conditions (Jackson, 1997).
Aqua was launched in May 2002 and ADEOS-II was also launched in the same
year. Each included an AMSR instrument. ADEOS-II was lost after a few months
of operation but Aqua is still providing data. As shown in Table 2.1 AMSR-E was to
provide significantly lower frequency measurements than the SSM/I with the same
level of spatial resolution. AMSR-E holds great promise for estimating soil water
content in sparsely vegetated regions and is the best possibility in the near term for
mapping soil water.
As opposed to previous passive microwave satellite missions, both NASA (Njoku
et al., 2003) and the Japanese Aerospace Exploration Agency (JAXA) (Njoku et al.,
2000) have included soil moisture as a mission product from AMSR-E. As noted
previously, it is not expected that AMSR-E can provide a globally reliable soil mois-
ture product. However, it will work in some regions and the efforts at operational
implementation and validation of products are providing valuable lessons for future
missions.

Table 2.2 Selected satellite conical scanning microwave radiometers


Frequency (GHz) Satellite Instrument
SSM/I AMSR-E WindSat

85/91 V, H V, H –
37 V, H V, H V, H, U, F
22/23 V V, H V, H
19 V, H V, H V, H, U, F
10.8 – V, H V, H, U, F
6.9/6.8 – V, H V, H
1.4 – – –
37 IFOV (km) 27 × 38 8 × 14 8 × 13
Observing time Various 1:30 p.m. 6:00 a.m.
Period of record 1985– 2002– 2003–
14 T.J. Jackson

The third satellite instrument is WindSat that launched in 2003 and includes a
multifrequency passive microwave radiometer system with a C band channel. This
is a risk reduction mission for one component of the next generation of operational
polar orbiting satellites that the U.S. will be implementing.
Regarding operational sensors, the Advanced Microwave Scanning Radiometer
(AMSR-E) on Aqua will transition to a component of the long term Japanese Global
Climate Observing Mission (GCOM) and WindSat may evolve to the Conical Mi-
crowave Imaging System (CMIS) on National Polar Orbiting Earth Satellite System
(NPOESS). The lower frequencies and commitments to both data and algorithms
should result in better products for land studies despite the rather coarse spatial res-
olution.
Japan provided the AMSR-E instrument to the Aqua platform and has plans to
follow this up with an improved version on the Global Climate Observing Missions-
Water (GCOM-W) satellites possibly in 2010. It is uncertain what the USA will
be doing as a follow on to WindSat and SSM/I. Until recently plans called for a
Conical Microwave Imaging System (CMIS) as part of the National Polar Orbit-
ing Earth Satellite System (NPOESS). CMIS would be similar to WindSat, offering
multifrequency (including C and X band) as well as fully polarimetric observations.
However, issues with instruments and costs are expected to impact the final design
and delay the current launch date. Another source of passive microwave measure-
ments will be satellites involved in the Global Precipitation Mission that follows the
successful TMI. This mission is scheduled for 2012.

2.4 New Directions and Missions

As noted previously, achieving a spatial resolution that can satisfy land application
requirements (or even demonstrate an approach that could open up future options)
has been the challenge to passive microwave remote sensing. The problem has been
achieving high spatial resolution at lower frequencies requires a large antenna. Get-
ting large and heavy antennas into space is obviously difficult and costly. There are
at least three approaches to solving this problem that are currently under investiga-
tion:
• Synthetic aperture radiometry
• Large lightweight antennas
• Data fusion-disaggregation using higher resolution remote sensing techniques
Synthetic aperture radiometry attempts to solve the spatial resolution dilemma by
replacing the traditional large filled array with a sparse array. It can achieve what
a very large antenna would, but with a much smaller mass and size. Earlier studies
established one dimensional synthetic aperture methods (Le Vine et al., 1994) and
demonstrated that this approach could be used for soil moisture retrieval (Jackson
et al., 1995). One-dimensional methods use a real aperture along track and synthetic
aperture methods across track, essentially a set of long sticks.
2 Passive Microwave Remote Sensing for Land Applications 15

The European Space Agency (ESA) will take synthetic aperture radiometry a
step further in the Soil Moisture Ocean Salinity (SMOS) mission that uses a two di-
mensional design (Kerr et al., 2001). SMOS includes an interferometric radiometer
operating at L band (1.4 GHz) and uses a technique based on the cross-correlation
of observations from all possible combinations of receiver pairs. SMOS is currently
scheduled for launch in late 2008. At an altitude of 763 km, the antenna will view
a swath of almost 3,000 km providing a 40 km global soil moisture product every
2–3 days.
Another key element of SMOS that is tied to its design is the soil moisture re-
trieval algorithm. Each SMOS 40 km footprint is the average of 1.2 s. As the satellite
moves along its orbital path each pixel is observed at many different viewing angles,
which permits the observation of brightness temperature as a function of viewing an-
gle. These multiple angles are used with radiative transfer equations to retrieve soil
moisture as well as vegetation information (Wigneron et al., 2000).
An alternative solution to the size-mass issues of real aperture antennas is the use
of lightweight mesh. This approach is compatible with designs that can be packaged
into smaller volumes for launch and also have much lower mass. A large mesh
antenna was one of the key features of the Hydros satellite mission that was studied
by NASA under its Earth System Science Pathfinder program (Entekhabi et al.,
2004) but was recently cancelled by NASA. Using a 6 m mesh antenna operating
at 1.4 GHz, Hydros would be able to provide a 40 km brightness temperature/soil
moisture product. It would employ conical scanning over a wide swath that results
in 2–3 day global coverage. Conical scanning facilitates some aspects of retrieval.
However, the spinning of the large antenna would require careful engineering. This
concept is currently being re-considered for priority implantation.
The passive microwave instruments of SMOS and Hydros would have demon-
strated two potential paths to achieving higher spatial resolution using only passive
sensors. The 40 km products would satisfy the demands of climatology studies.
There are many additional applications that could utilize soil moisture if it was avail-
able at higher spatial resolution. If successful, each of these technologies could lead
to future missions with resolutions of better than 10 km.
Another approach to better spatial resolution is disaggregation. The concept of
disaggregation to achieve a higher resolution product is based on the assumption that
the passive microwave instrument provides a reliable soil moisture product and that
alternative remote sensing measurements while at higher spatial resolutions have
lower accuracies. This has been explored in the past using visible-near infrared re-
mote sensing. In regions with uniform soil properties and small amounts of vege-
tation this might work. However at these wavelengths the high-resolution data only
represents an extremely thin surface soil layer that easily disconnects from the lower
profile. If there is vegetation this approach will not work unless the vegetation ex-
hibits a response related to the soil moisture state. There is some justification for
this; however, it is also compounded by lags in time between changes in soil mois-
ture and changes in canopy visible-near infrared characteristics.
Using thermal infrared for disaggregation has also been proposed. It too has lim-
itations. However, it is expected that the lag time between soil moisture changes and
16 T.J. Jackson

temperature changes will be shorter than with physiographic features observed with
visible-near infrared.
Disaggregation was explored by Chauhan et al. (2003) as a solution for NPOESS.
It was proposed that a combination of microwave, visible-near infrared, and ther-
mal infrared could be used. It was demonstrated using 25 km SSM/I products and
1 km Advanced Very High Resolution Radiometer (AVHRR) data. This might work
sometimes, however, cloud cover would certainly limit it.
Another approach to disaggregation was central to the Hydros mission design. In
addition to the passive microwave instrument, Hydros would have included a radar
that shared the same antenna. This instrument would provide a 3 km soil moisture
product with possibly limited accuracy. However, the real role of this sensor was
to serve as a disaggregation and data fusion tool that would be combined with the
coarser resolution passive product. It is fairly well established that a radar responds
to soil moisture in much the same way as a passive microwave sensor for smooth
bare soils. However, roughness and vegetation effects can be more difficult to ac-
count for using radar, which leads to a higher uncertainty in radar only soil mois-
ture retrieval algorithms. A potential disaggregation scheme is described in Narayan
et al. (2006).
Not only would the radar serve in disaggregation it would provide the option
of using data fusion of passive and active microwave in soil moisture retrieval.
The concept of data fusion of passive and active could be explored using another
L band mission being implemented by NASA called Aquarius (Koblinsky et al.,
2003). Aquarius would provide (∼2010) both active and passive coarse spatial res-
olution measurements using a push broom approach. It would provide some useful
information for algorithm science but the technology does not provide a pathway to
high spatial resolution.
It was anticipated that with Hydros on its successor a highly reliable 10 km soil
moisture product could result from the integration of passive and active microwave
remote sensing. A 10 km soil moisture product would open the doors to a much
wider range of applications and the demonstration of the approach could lead to
even higher resolutions in the future. There are likely to be limits on the differences
in scale that can be used when disaggregating. There has to be some overlap in the
dominant processes that control variability at the two resolutions. If a passive tech-
nology is demonstrated that can lead to future missions with higher spatial resolu-
tions, then radars with higher resolutions can also be consider that could eventually
result in products <1 km. Schemes for disaggregation that could be compatible with
SMOS, as well as other missions, are described in Merlin et al. (2005) and Pellenq
et al. (2003).

2.5 Summary

Land applications, in particular soil moisture retrieval, have been hampered by the
lack of low frequency passive microwave observations and the coarse spatial resolu-
tion of existing sensors. The next decade will see several exploratory missions using
2 Passive Microwave Remote Sensing for Land Applications 17

new technologies, and innovative approaches to integrating passive microwave ob-


servations with active microwave measurements that could lead the way to an order
of magnitude improvement in spatial resolution.

References

Basist A, Grody NC, Peterson TC, Williams CN (1998) Using the Special Sensor Microwave/
Imager to monitor surface temperatures, wetness, and snow cover. J. Appl. Met. 37:888–911
Becker F, Choudhury, BJ (1988) Relative sensitivity of NDVI and microwave polarization dif-
ference index (MPDI) for vegetation and desertification monitoring. Remote Sens. Environ.
24:297–311
Bindlish R, Jackson TJ, Wood E, Gao H, Starks P, Bosch D, Lakshmi V (2003) Soil moisture esti-
mates from TRMM Microwave Imager observations over the Southern United States. Remote
Sens. Environ. 85:507–515
Chauhan NS, Miller S, Ardanuy P (2003) Spaceborne soil moisture estimation at high resolution:
a microwave-optical/IR synergistic approach. Int. J. Remote Sens. 24:4599–4622
Entekhabi D, Njoku E, Houser P, Spencer M, Doiron T, Belair S, Crow W, Jackson T, Kerr Y,
Kimball J, Koster R, McDonald K, O’Neill P, Pultz T, Running S, Shi JC, Wood E, van
Zyl J (2004) The hydrosphere state (Hydros) mission concept: an Earth System Pathfinder
for global mapping of soil moisture and land freeze/thaw. IEEE Trans. Geosci. Remote Sens.
42:2184–2195
Gaiser PW, St. Germain KM, Twarog EM, Poe GA, Purdy W, Richardson D, Grossman W,
Jones WL, Spencer D, Golba G, Cleveland J, Choy L, Bevilacqua RM, Chang PS (2004) The
WindSat spaceborne polarimetric microwave radiometer: sensor description and early orbit
performance. IEEE Trans. Geosci. Remote Sens. 42:2347–2361
Jackson TJ (1997) Soil moisture estimation using special Satellite Microwave/Imager satellite data
over a grassland region. Water Resour. Res. 33:1475–1484
Jackson TJ, Schmugge TJ (1991) Vegetation effects on the microwave emission from soils. Remote
Sens. Environ. 36:203–212
Jackson TJ, Levine DM, Swift CT, Schmugge TJ, Schiebe FR (1995) Large-area mMapping of
soil-moisture using the ESTAR passive microwave radiometer in Washita92. Remote Sens.
Environ. 54:27–37
Kerr YH, Waldteufel P, Wigneron JP, Font J, Berger M (2001) Soil moisture retrieval from
space: the Soil Moisture Ocean Salinity (SMOS) mission. IEEE Trans. Geosci. Remote Sens.
39:1729–1735
Koblinsky CJ, Hildebrand P, LeVine D, Pellerano F, Chao Y, Wilson W, Yueh S, Lagerloef G
(2003) Sea surface salinity from space: science goals and measurement approach. Radio Sci.
38(4):8064, doi:10.1029/2001RS002584
Le Vine DM, Griffis A, Swift CT, Jackson TJ (1994) ESTAR: a synthetic microwave radiometer
for remote sensing applications. Proc. IEEE. 82:1787–1801
Li L, Njoku EG, Im E, Chang P, St. Germain K (2004) A preliminary survey of radio-
frequency interference over the U.S. in Aqua AMSR-E data. IEEE Trans. Geosci. Remote Sens.
42:380–390
Merlin O, Chehbouni AG, Kerr YH, Njoku EG, Entekhabi D (2005) A combined model-
ing and multi-spectral/multi-resolution remote sensing approach for disaggregation of sur-
face soil moisture: application to SMOS configuration. IEEE Trans. Geosci. Remote Sens.
43:2036–2050
Narayan U, Lakshmi V, Jackson TJ (2006) High-resolution change estimation of soil moisture
using L-band radiometer and radar observations made during the SMEX02 experiments. IEEE
Trans. Geosci. Remote Sens. 44:1545–1554
18 T.J. Jackson

Narvekar PS, Jackson TJ, Bindlish R, Li L, Heygster G, Gaiser P (2007) Observations of land
surface passive polarimetry with the WindSat Instrument. IEEE Trans. Geosci. Remote Sens.
45:2019–2028
Neale CMU, McFarland MJ, Chang K (1990) Land-surface type classification using microwave
brightness temperatures from the Special Sensor Microwave/Imager. IEEE Trans. Geosci. Re-
mote Sens. 28:829–838
Njoku E, Koike T, Jackson T, Paloscia S (2000) Retrieval of soil moisture from AMSR data. In:
P. Pampaloni, S. Paloscia (eds), Microwave Radiometry and Remote Sensing of the Earth’s
Surface and Atmosphere. VSP, Zeist, The Netherlands, pp 525–533
Njoku EG, Jackson TJ, Lakshmi V, Chan TK, Ngheim SV (2003) Soil moisture retrieval from
AMSR-E. IEEE Trans. Geosci. Remote Sens. 41:215–229
Njoku EG, Ashcroft P, Chan T, Li L (2005) Global survey and statistics of radio frequency inter-
ference in AMSR-E land observations. IEEE Trans. Geosci. Remote Sens. 43:938–947
Owe M, van de Griend AA, Chang AT (1992) Surface moisture and satellite microwave observa-
tions in semiarid Southern Africa. Water Resour Res. 28:829–839
Owe M, de Jeu R, Walker J (2001) A methodology for surface soil moisture and vegetation optical
depth retrieval using the microwave polarization difference index. IEEE Trans. Geosci. Remote
Sens. 39:1643–1654
Paloscia S, Macelloni G, Santi E, Koike T (2001) A multifrequency algorithm for the retrieval
of soil moisture on a large scale using microwave data from SMMR and SSM/I. IEEE Trans.
Geosci. Remote Sens. 39:1655–1661
Pellenq J, Kalma JD, Boulet G, Saulnier G, Wooldridge S, Kerr Y, Chehbouni A (2003) A scheme
for disaggregation of soil moisture along topography and soil depth. J. Hydrology. 276:112–127
Ulaby F, Moore R, Fung A (1982) Microwave Remote Sensing: Active and Passive, vol. II,
Addison-Wesley, Reading, MA
Wigneron JP, Waldteufel P, Chanzy A, Calvet JC, Kerr Y (2000) Two-dimensional microwave
interferometer retrieval capabilities over land surfaces (SMOS mission). Remote Sens. Environ.
73:270–282
Chapter 3
Active Microwave Remote Sensing Systems
and Applications to Snow Monitoring

Jiancheng Shi

Abstract This chapter describes research activities made over terrestrial snow
covered areas with active microwave instruments, especially with Synthetic
Aperture Radar (SAR). Even though an exhaustive review of all studies could
not be considered in the framework of the book, the chapter consists of many signif-
icant developments which have been made recently in the space-borne and air-borne
based active microwave remote sensing of snow properties including techniques of
mapping snow, inferring snow wetness and snow water equivalence (SWE). The
abilities and existing problems with the current sensors will be discussed.

3.1 Introduction

Seasonal snow cover is one of the most important components in predicting global
water- and energy-cycle consequences of Earth-system variability and change. Sea-
sonal snow cover and its subsequent melt can dominate local-to-regional climate and
hydrology. It is the major source of fresh water over wide areas of the mid-latitudes.
Understanding, characterizing, and predicting snow-related processes across spatial
scales in coupled atmospheric and hydrologic models requires improved capabil-
ity for accurately monitoring spatial and temporal distributions of seasonal snow
properties on land, especially snow water equivalence (SWE) and snow wetness.
In situ measurements provide direct characterization, but at limited spatial and tem-
poral extent and resolution, and frequently must be acquired under challenging or
dangerous field conditions. Furthermore, because of the high spatial and temporal
variability of snow cover, snow properties derived from in situ measurements often
do not provide the reliable spatial and temporal characterization of distributed snow
properties across model domains at the required range of resolutions.

Jiancheng Shi
Institute for Computational Earth System Science, University of California, Santa Barbara, USA
shi@icess.ucsb.edu

S. Liang (ed.), Advances in Land Remote Sensing, 19–49. 19


c Springer Science + Business Media B.V., 2008
20 J. Shi

Because visible and near-infrared radiation does not penetrate snow, and be-
cause the optical properties of ice and water are similar, snow reflectance in this
part of the electromagnetic spectrum is not sensitive to snow depth (except for very
shallow snow) or free liquid water in the snowpack. For large, flat regions, passive
microwave data at 18 and 37 GHz allow sequential mapping of SWE distribution. In
dry snow, the profile of the snow grain sizes also influences the microwave remote
sensing signal. A more serious problem for measurement of SWE is that the snow
depths are often too large for existing passive microwave remote sensing retrieval
algorithms to provide accurate estimates. The signal becomes asymptotic and insen-
sitive to snow water equivalence at and above about 0.5 m at 37 GHz (a commonly
used frequency). Moreover, the spatial resolution of the current passive microwave
satellite sensors is too coarse to provide useful information at regional and drainage
basin scales for hydrological investigations and applications. Active microwave sen-
sors (radars), on the other hand, are sensitive to many snow parameters such as snow
density, depth, grain size, free liquid water content, and snow-pack structures that
are useful for hydrologic applications.
Active microwave sensors, especially Synthetic Aperture Radar (SAR), have
been used to estimate snow properties, such as snow wetness and snow water equiv-
alency (SWE), and to discriminate snow with other surfaces. They acquire image
measurements of the Earth day or night, in all weather, through cloud cover, smoke
and haze, and with high resolution especially useful in monitoring geophysical prop-
erties in alpine regions. This chapter summarizes the recent progresses in the tech-
nical developments for estimation of snow and vegetation properties using active
microwave instruments, mainly synthetic aperture radar (SAR).

3.1.1 Current Available Satellite SAR Data

There are several Spaceborne SAR instruments available, in past, current and near
future, from different space agencies over the world. The major sensor parameters
are summarized in Table 3.1.
ERS-1 and ERS-2 SAR instrument is one of the sensors on the first and second
European Remote Sensing Satellites that were launched in July 1991 and 1995 by
the European Space Agency. They are circling the Earth every 100 min. In 3 days,
it can cover the entire planet with exact repeat coverage at 35 days. They have the
polar orbit at 780 and 785 km for ERS-1 and ERS-2, respectively. Both ERS-1 and
ERS-2 SAR transmit and receive the microwave signals with vertical polarization
at frequency 5.3 GHz and a fixed incidence angle of 23◦ . Its primarily applications
oriented towards ocean and ice monitoring including those of sea state, sea sur-
face winds, sea surface temperature, ocean circulation and sea and ice level. Due
to its all-weather and high resolution microwave imaging capability, its applications
have been found over land from the monitoring of crops, tropical deforestation to
flooding monitoring. The commonly used products of ERS-1/2 SAR in geophysical
3 Active Microwave Remote Sensing Systems and Applications to Snow Monitoring 21

Table 3.1 Characteristics of current available spaceborne SAR


Sensor Frequency Polarization Incidence Pixel Available time
in GHz in degree resolution frame
in m
ERS-1/2 5.3 VV Fixed at 23◦ 3.8–12.5 Since 1991/1995
ASAR 5.33 Dual-polarization Varying with 3.8–150 Since 2002
mode
RADARSAT-1 5.3 HH Varying with 10–100 Since 1997
mode
SIR-C/X-SAR 1.25 and Fully Varying with 6–27 April and
5.3/9.6 polarimetric/VV data takes October, 1994
JERS-1 1.27 HH Fixed at 35◦ 18 1994–1997
PALSAR 1.27 Fully polarimetric, Varying with 10–100 Since December
dual-polarization, mode 2005
and HH
RADARSAT-2 5.3 Fully polarimetric, Varying with 3–100 2007
dual-polarization, mode
and HH
TerraSAR-X 9.6 VV or dual- Varying with 1–16 2007
polarization mode

applications are the single look complex product (SLC), the precision image (PRI),
and the geocoded product that produced from PRI data after Earth ellipsoid and
terrain corrections. The more detailed information about ERS-1 and ERS-2 SAR
instrument and data products can be found at http://www.earth.esa.int/ers.
In March 2002, the European Space Agency launched Envisat, an advanced
polar-orbiting Earth observation satellite which provides measurements of the
atmosphere, ocean, land, and ice. An Advanced Synthetic Aperture Radar (ASAR),
operating at C-band, ASAR ensures continuity with the image mode (SAR) and the
wave mode of the ERS-1/2 AMI. It enhanced capabilities in comparison with ERS
SAR include coverage (a ScanSAR mode), range of incidence angles, polariza-
tion (dual-polarizations), and modes of operation (with nine different modes). The
resulting improvements in image and wave mode beam elevation steerage allow
the selection of different swaths, providing the swath coverage of over 400 km wide
using ScanSAR techniques. In alternating polarization mode, transmit and receive
polarization can be selected allowing scenes to be imaged simultaneously in two
polarizations.
RADARSAT-1, launched in November 1995, is an operational radar satellite sys-
tem developed by Canadian Space Agency (CSA) to monitor environmental change
and the planet’s natural resources. RADARSAT-1 has the sun-synchronous orbit at
798 km altitude with 6 p.m. ascending node and 6 a.m. descending node. It trans-
mits and receives the microwave signals with horizontal polarization (HH) at C-band
(5.3 GHz). It has seven beam modes, 35 beam positions and ScanSAR modes for a
wide range of imaging options, varying resolutions from 8 to 100 m, swath widths of
22 J. Shi

50–500 km, and incidence angles from 10◦ to 59◦ , depending on the mode selection.
The more detailed information about RADARSAT-1 instrument and data products
can be found at http://www.space.gc.ca/asc/eng/satellites/radarsat1.
ALOS PALSAR is the Phased Array type L-band Synthetic Aperture Radar
(PALSAR) on the Advanced Land Observing Satellite (ALOS) that has been
launched on January 24, 2006 and is a follow on the Japanese Earth Resources
Satellite-1 (JERS-1). It provides higher performance than the JERS-1’s SAR for
land observations. PALSAR has the ScanSAR mode with the swath width of 250–
350 km depending on the number of scans in addition to the fine resolution in a
conventional mode. This swath is three to five times wider than conventional SAR
images. It has also different polarization selections from single, dual, to fully polari-
metric observations and improved orbit controlling capability that allows the more
frequent repeat-pass interferometer applications.
SIR-C/X-SAR is the Shuttle Imaging Radar-C and X-Band Synthetic Aperture
Radar. It was a cooperative experimental SAR mission between the National Aero-
nautics and Space Administration (NASA), the German Space Agency (DARA),
and the Italian Space Agency (ASI) (Evans et al., 1997). The SIR-C/X-SAR sys-
tem was flown onboard NASA’s Space Shuttle on two 10-day missions in April and
October 1994. SIR-C provides radar polarimetric digital images simultaneously at
two wavelengths, L-band (24 cm), and C-band (5.6 cm). These polarimetric data
allows derivation of the complete scattering matrix on a pixel-by-pixel basis and
more detailed information about the geometric structure and dielectric property of a
target (Van Zyl et al., 1987; Evans et al., 1988). X-band (3 cm) with VV polarization
results in a three-frequency capability. The resolution of this system ranged between
10 and 40 m, and it collected image swaths between 15 and 90 km wide.
In addition, there have been two new SAR satellites launched in 2007.
RADARSAT-2 is the follow on RADARSAT-1 system co-funded by CSA and
MacDonald Dettwiler (MDA). It has been launched in 2007 with 7 years mis-
sion duration and will support all RADARSAT-1 imaging modes. It has the same
orbit, repeat cycle and ground track as RADARSAT-1. Its improvements in com-
parison with RADARSAT-1 include the higher resolution (up to 3 m), routine
left-looking and right-looking mode for more frequent revisit, selective polariza-
tions with fully-polarimetric imaging modes, on-board GPS receivers to improve
the real-time position knowledge, and higher downlink power. The more detailed
information about RADARSAT-2 instrument and data products can be found at
http://www.radarsat2.info/rs2 satellite.
TerraSAR-X is a new German radar satellite with 514 km sun-synchronous orbit
at 98◦ Inclination and has been launched in June, 2007 with a 5 year lifetime. It car-
ries a high frequency X-band (9.65 GHz) SAR sensor that can be operated in three
different basic modes – the ScanSAR, Stripmap and Spotlight mode at a varying
geometrical resolution between 1 and 16 m. TerraSAR-X will provide single or dual
polarization data. On an experimental basis additionally quad polarization and along
track interferometry are possible.
3 Active Microwave Remote Sensing Systems and Applications to Snow Monitoring 23

3.1.2 SAR Measurement Properties

Generally, SAR provides three types of measurements that can be used in study
geophysical properties depending on the sensor properties: (1) backscattering mea-
surements at the given frequency and polarization such as ERS1/2, ASAR, and
RADARSAT-1; (2) the measurements of the polarization properties by a fully
polarimetric instrument (SIR-C, PALSAR, RADARSAT-2, and TerraSAR-X with
selected mode); and (3) interferometry from the repeat-pass measurements.
Backscattering SAR imagery provides high resolution digital images but with
only fixed polarization state of the antenna. With only one or a few intensity mea-
surements per pixel, the applications to monitor snow properties have to rely mainly
on the radiometric properties such as snow classification and estimations of snow
wetness and SWE.
Rather than just measuring amplitude, an imaging radar polarimeter measures the
amplitude and relative phase for every polarization state. These complex measure-
ments lead to nine independent real elements of the Stokes’ matrix, which describes
how the scattering mechanisms in each pixel transform the illuminating electromag-
netic wave back to the receiving antenna. The polarization feature derived from the
Stokes’ matrix, is the radar cross section as a function of the antenna polarization
state and is useful in interpreting the scattering mechanisms within a resolution cell.
The first element of the matrix is the total power or span. The nine cross products
of the pixel scattering matrix can be obtained by the linear combinations from the
span and the other Stokes’ matrix elements. In addition to the polarization signa-
ture, many other features describing the scattering mechanisms within a pixel can
be derived from the Stokes’ matrix. These include: the pixel intensity synthesized
from the polarization signature for fixed antenna polarization states, the polarization
phase difference of the scattering matrix elements, the coefficient of variation, the
enhancement factor, the scattering mechanism and the degree of polarization. Thus,
polarimetric measurements provide much more information per pixel than the single
fixed antenna SAR imagery, and have been shown to be effective in classification
of terrain. These measurements have great opportunity to minimize the topographic
effects on radar images.
In addition to single, multi-polarization, and fully polarimetric radar observa-
tions, the repeat-pass interferometric radar measurement also provides the useful
information in analyses of land surface geophysical properties. Its techniques for
topographic mapping of surfaces promise the high-resolution digital elevation mod-
els; but they also permits inference of changes in the surface over the orbit repeat
cycle from the correlation properties of the radar echoes. Measurements of inter-
ferometer correlation describe processes occurring on the time scales of the orbit
repeat time and size scales on the order of a radar wavelength, such as vegetation
growth, glacier motion, permafrost freezing and thawing, and soil moisture induced
effects.
Furthermore, SAR images have two distinct special characteristics: image
speckle and a more complicated geometric mapping in contrast to images from
optical sensors. The SAR image generation involves a coherent processing carried
24 J. Shi

out on the received signal: fading causes on SAR imagery a grainy appearance
referred to as speckle (Ulaby et al., 1982, 1986). This is because a SAR reso-
lution cell, that contains many different scatterers, is very large when compared
to the wavelength of the illuminating electromagnetic wave. The returned radar
echo results from the coherent summation of all the returns from the amplitudes
and phases of the single scatterers. Commonly, it requires processing a multi-look
image or applying a speckle filter to reduce its effects and to improve the image
quality (Ulaby et al., 1982, 1986; Lee et al., 1998). Furthermore, SAR mapping is
mainly an integration of backscattered signals having the same Doppler frequency
as along-track measurements and the same distance as across-track measurements.
The characteristic radar measurements are range – the distance from the sensor to
an object point, and time – the position of the sensor along its flight path where the
data are collected. They define the two-dimensional SAR image space. The total
received power from a resolution cell is proportional to the radar cross section, and
the rest of factors in the radar equation are generally assumed constant. The radar
cross section is usually a function of polarization, frequency, viewing geometry,
and illuminated area, in addition to factors such as the dielectric properties and
geometric structure of the targets. During the data processing, the backscattering
coefficient σo is usually presented in terms of a normalized radar cross section to
eliminate the dependence on illumination area.
σ
σo = (3.1)
A
where A is illumination area. For a flat area, both illumination area and incidence
angle of a pixel can be reasonably estimated so that the backscattering coefficient
can be obtained.
Topographic effects on radar images can be considered as two aspects: (1) the
effects on received radar power, which result from a great variation in illuminated
area and incidence angle for a pixel resolution. It can be described as the variation of
the received power from an inclined surface compared to the received power from
a horizontal surface. This variation is a function of the relative orientation of the
surfaces with respect to the illuminating source and their position relative to the sen-
sor. (2) The geometric distortions on image coordinates that are mainly dependent
upon the distance between an imaged pixel and the sensor. The effects of rugged
terrain on both the received power result from the change in illuminating area and
the geometric distortion can be corrected for a resolution cell if a digital elevation
model (DEM) with the comparable resolution and the sensor location information
are available. They are available only with the geocoded data products.

3.2 Snow Mapping with SAR

For climatological and hydrological investigations, the area of snow cover is one
of the important parameters. For instance, the snow line on glaciers is an impor-
tant quantity for hydrological applications and mass balance studies. Visible and
3 Active Microwave Remote Sensing Systems and Applications to Snow Monitoring 25

near-infrared sensors have been used extensively to measure these quantities, but
are hampered by cloud cover, which can be pervasive in some regions. In partic-
ular, snow cover must be measured on a timely basis to be useful for operational
hydrology, and the opportunities for obtaining suitable data from these sensors can
be infrequent. Microwave remote sensing is a methodology that is less influenced
by cloud conditions, depending on frequency. Especially, Synthetic Aperture Radar
(SAR) provides high resolution measurements that are comparable to the scales of
the topographical variation in mountain areas and more suitable for mapping snow
cover than passive microwave instruments.
In the microwave part of the electromagnetic spectrum, ice is almost transpar-
ent, and the radar penetration depth, depending on the frequency, can reach tens
of meters for dry snow. The major scattering source is the snow–ground inter-
face, and it is difficult to discriminate dry snow cover from bare surfaces or short
vegetation with backscattering measurements from a single-polarization radar mea-
surements. Earlier investigators found no significant difference in backscattering
from dry snow and snow-free surfaces at either C-band (Rott et al., 1993) or X-band
(Fily et al., 1995).
However, there is a large dielectric contrast between the solid and liquid phases
of water at microwave frequencies. As snow starts melting, even a small amount of
liquid water reduces the penetration depth of the radar signal, and thereby changes
the dominant scattering source from the snow–ground interface to the snow volume
and the air–snow interface. The backscattering coefficients decrease substantially at
C-band and X-band (Stiles and Ulaby, 1980; Nagler and Rott, 2000; Shi and Dozier,
1995). As the snow continues melting and the surface ages, backscattering increases
and surface scattering at the air–snow interface becomes the dominate scattering
source, as Rott et al. (1993) observed in ERS-1 data over the Ötztal, Austria, a
snow-covered alpine glacier region. Surface scattering also dominates the signals
from bare rock, soil, and glaciers, but these surfaces are usually rougher than wet
snow and therefore have stronger backscattering signals. However, the backscatter-
ing from wet snow cover may have the similar intensities as that from very smooth
bare soils, which can result in a difficulty for separation, as Haefner et al. (1993)
showed in ERS-1 data from the Swiss Alps.
Since the radiometric quality of SAR images in an alpine region is dependent
on flight and imaging parameters (e.g., flight altitude, radar elevation angle) and the
topography of the imaged area, the representation of the target materials is likely
to be inaccurate. This variation in radar backscatter that is unrelated to the surface
cover type is particularly evident for high relief surfaces where a large variation of
slope and aspect creates a great variation of local incidence angles and illuminated
areas. For example, rock surfaces with greater incidence angle could have lower
power return than that from snow or glacier covered areas with smaller incidence
angle. The effect of topography, therefore, is another major problem we are facing
in using radar to map snow covered areas, especially in alpine regions.
The study (Van Zyl et al., 1993) indicates that the topographic effects on radio-
metric properties measured from spaceborne SAR can be explained by variation in
imaged pixel area and in local incidence angle. When topographic information of
26 J. Shi

the study area is available, both spaceborne and airborne SAR image data can be
radiometrically calibrated. The remaining problem for target discrimination is the
effect of local incidence angle on the received power. If topographic information is
not available, we need to consider the effects of variations in both local incidence
angle and imaged pixel area for satellite SAR measurements.
Therefore, in attempting to use SAR data to map seasonal snow cover over re-
mote and inaccessible areas, we are faced with two major problems:
• Compensation must be made for the effects of rugged terrain.
• Snow must be distinguished from other surface covers.
The currently available techniques for mapping snow cover with SAR imagery can
be summarized mainly into three catalogues based on the usage of the SAR mea-
surement properties.

3.2.1 Multi-temporal Single-polarization Techniques

Previously, a single-polarization SAR imager provides high-resolution images but


with only one available measurement per pixel at a fixed polarization state. With
only one intensity measurement per pixel, we have to rely mainly on radiometric
properties to distinguish snow covered area from other targets, such as bare ground
or vegetation. A practical technique to mapping wet snow cover, that has been de-
veloped recently, is the multi-temporal methods. Nagler and Rott (2000) used four
ERS-1 repeat-pass images, two from ascending and two from descending orbits to
reduce the effect of layover. They compared backscattering ratios of an image with
wet snow and a reference image (either snow-free or dry snow) to generate a map
of wet snow cover, based on change detection. They found that wet snow cover
can be identified by using the time series measurements. The radar observations
have shown that the backscattering from wet snow cover could be reduced 3–4 dB
at C-band in comparison with eight dry snow cover before melting or bare surface
after melting. In addition, they also found that the requirement for accurate local
incidence angle was relaxed when they used multi-temporal ERS-1 data. (Rott and
Nagler, 1993) since the classification algorithm used the ratio measurements of wet
snow cover image to the reference image in which the part of terrain effects – pixel
illumination area has been cancelled out if two temporal images are co-registrated
accurately.
For wet snow mapping in forested regions, Luojus et al. (2006) demonstrated a
two step technique to map wet snow cover fraction using multi-temporal ERS-2
measurements for boreal forest region in Finland during the snow melting sea-
son. The first step is the forest canopy compensation. This is done by nonlin-
early fitting ERS-2 measurements with a semi-empirical forest backscattering model
(Pulliainen et al., 2003) with the forest stem volume information to estimate the
backscattering signals at ground surface and the volume backscattering signals, the
two-way transmittivity of the forest canopy. The second step is the employment
of linear interpolation algorithm that uses the reference images and the estimated
3 Active Microwave Remote Sensing Systems and Applications to Snow Monitoring 27

backscattering signals at ground surface to calculate the fraction of wet snow cov-
ered area – SCA.
σsurf
o −σo
ground,ref
SCA = 100% · o (3.2)
σsnow,ref − σground,ref
o

where σosurf is the estimated backscattering coefficient at the ground surface,


σoground,ref is the reference signal from the snow-free ground, and σosnow,ref is the ref-
erence signal from the wet snow ground. One reference image describes the signal
at the fully wet snow cover situation and the other describes the snow-free at the
end of the snow melting season. In addition to two reference images, this method
requires the knowledge of the forest stem volume distribution of the study area and
suitable to sub image scale applications (many pixels averaged together). This is be-
cause the uncertainty produced by radar speckle and the use of forest compensation
in which the requested forest stem information is more reliable at coarse resolution
than that at ERS-2’s pixel resolution.

3.2.2 Multi-frequency and Multi-polarization Techniques

Shi and Dozier (1997) evaluated the characteristics of the backscattering, polariza-
tion, and frequency ratios of the targets of the study site near Mammoth Mountain
in the Sierra Nevada, U.S using multi-frequency and polarization SIR-C/X-SAR’s
measurements. They developed two type of supervised classifiers based on clas-
sification tree technique. The first type of the classifier was developed by using
intensity measurements, polarization properties, and frequency ratios. It can map
dry snow and discriminate dry from wet snow, but it requires topographic informa-
tion for radiometric terrain correction and to reduce effects of local incidence angle.
It is about 79% as accurate as a TM binary classification, but it suffers the same
shortcoming – it underestimates total snow cover in regions of mixed pixels, espe-
cially forested regions. Its performance on the two data-takes where the snow was
dry showed that only a few pixels were misclassified as wet snow.
The second type of classifier was developed based on polarization properties
and backscattering ratios between different frequencies. Since these measurements
can be obtained correctly without radiometric terrain calibration, the classifier does
not require topographic information and can be used to map wet snow. Similarly,
Shi et al. (1994) also developed a method with multi-polarization C-band airborne
SAR to map wet snow and glacier ice without a DEM, using only measurements
of the polarization properties. Its accuracy is 77% when compared with TM binary
classification, but both underestimate total snow cover.

3.2.3 Repeat-pass Interferometric Technique

Except the technique of the first type of classifier as in the multi-frequency and po-
larization techniques, all above methods are restricted to mapping wet snow-cover
28 J. Shi

since it is difficult to discriminate dry snow cover with bare ground and short vege-
tation. In the study of using SIR-C/X-SAR data to map snow cover (Shi and Dozier,
1997), it was found that wet snow cover had very similar backscattering intensity
and polarization characteristics to smooth bare surface at C-band and X-band. For
instance, the backscattering from wet snow-cover is very similar to smooth dry soil,
alluvial surfaces, and relatively rough water surfaces. At a large drainage basin or
regional scale, where many different targets are within a scene, those techniques
might not be reliable. For similar reasons, change detection measurement might be
also unreliable since the similar change in backscattering could be caused by differ-
ent natural environment changes. In order to develop a large-scale snow mapping
technique other measurement are required to discriminate between snow and other
targets.
Interferometric radar techniques for topographic mapping of surfaces promise the
high-resolution digital elevation models; but they also permits inference of changes
in the surface over the orbit repeat cycle from the correlation properties of the radar
echoes. Measurements of interferometer correlation describe processes occurring on
the time scales of the orbit repeat time and size scales on the order of a radar wave-
length, such as vegetation growth, glacier motion, permafrost freezing and thawing,
and soil moisture induced effects. The coherence measurement between two repeat-
passes, therefore, provides a useful measurement in addition to backscattering in-
tensities in each scene and their changes between two passes, and makes it possible
to develop an algorithm for mapping both dry and wet snow covers over large area.
Strozzi et al. (1999) demonstrated that coherence measurements could provide
the separation between wet snow cover and bare ground in the cases where the
backscatter discrimination failed from analyses of ERS tandem data in Switzerland.
The low coherence observed over wet snow cover is mainly caused by the rapid
change in scattering properties and geometry as the result of wet snow meta-
morphism due to the movement of free liquid water content, ice grain growth,
displacements of adjacent scatterers, and formation of density heterogeneities
(layering, ice-lenses, etc.), which all result in a significant decorrelation. On the
other hand, the high coherence is regularly observed over no-forested snow free
areas. For forested areas, it can be easily separated with wet snow cover due to their
huge difference in backscattering intensity even if its coherence is generally low.
They provide the physical bases for separating wet snow with other surfaces. For us-
ing C-band measurements, however, it requires the short temporal scale (a few days)
between the two repeat-pass measurements in order to avoid the significant temporal
decorrelation in other surface targets such as bare or short vegetation surfaces.
Shi et al. (1997) evaluated the L-band coherence measurements between two
repeat-pass SIR-C image data from its first mission in April (with snow) and sec-
ond mission in October (without snow), 1994. This measurement indicates that the
ground is completely undisturbed between viewings the signals will be highly cor-
related. Otherwise, the decorrelation will occur. They evaluated the coherence mea-
surements of L-band VV polarization for five land cover types snow, lake, bare
surface, short vegetation, and forest. It was found that lake, snow and some forest
areas had very low coherence between two data-takes. For the lake, the decorrelation
between two data-takes is mainly due to the low S/N ratio and the changes of the
3 Active Microwave Remote Sensing Systems and Applications to Snow Monitoring 29

lake surface roughness characteristics, because of change in wind conditions (speed


and direction). For the dry snow case, the dominant scattering at L-band is from the
snow-ground interface. In addition to the change of the dielectric contrast from the
air-ground to snow-ground, existing dry snow cover will result in large decorrela-
tion. This is due to change in local incidence angle when radar signal passes through
snow layer will cause a spatial baseline decorrelation. The radar echoes will be also
expected to be close to complete decorrelation when measuring correlation between
wet snow-cover and bare ground image passes. This is because the radar signal in
snow-covered pass can only penetrate a few centimeters so that the radar senses two
different targets. The coherence measurements from forest can have large dynamic
range with very low values – similar to those from snow and lake, especially from
dense forest. On the other hand, the coherence measurements from the bare sur-
face are significantly higher than those from lake. For the bare surface, a change of
soil moisture will result in a decorrelation. However, the amount of decorrelation
is expected to be smaller because radar senses a same target with a same scattering
mechanism (only change in magnitude). The short vegetation (mainly sagebrush and
grass in this study area) has very similar coherence measurement to the bare surface
mainly due to the dominant scattering source is from the ground surface at L-band.
Therefore, the coherence measurements between one snow covered scene and one
without snow provide a very good separation between snow cover and bare surface
as well as short vegetation. These two targets are most difficult to discriminate with
snow cover. Thus, the correlation measurement provides a significant information
to map snow covered area.
A pixel-based decision tree classifier (DTC) was developed based on above char-
acteristics. While the coherence measurements to discriminate forest, open water,
and snow with short vegetation and bare ground, snow can be separated with forest
and open water by using backscattering measurements. In order to verify the clas-
sification results, a cloud-free Landsat Thematic Mapper (TM) scene on April 14,
1994 (the SIR-C/X-SAR data-take on April 13, 1994) was acquired. A snow map
was generated by the TM data and then projected (or co-registered) to the slant
range presentation of the SAR images by using the Space Shuttle ephemeris data
and ground control points. In the classification of the TM scene, there are only four
target categories: lake, snow, forest, and range land that includes both bare ground
and short vegetation. Figure 3.1 shows the SAR classification map (right) and the

Fig. 3.1 Comparison of SAR (right) and TM (left) derived classification maps. Black – Forest and
lake, gray – bare surface and short vegetation, white – snow
30 J. Shi

TM classification map (left). The comparison estimation of this two results indi-
cated that at 86% accuracy can be obtained for snow cover area under consideration
of the TM classification map as the ground truth.

3.3 Estimates of Snow Wetness from C-Band SAR

3.3.1 Relationships Between Snow Wetness and C-Band SAR


Measurements

Monitoring spatial and temporal changes of liquid water content in snow is impor-
tant for hydrological modeling, because the presence of liquid water shows that a
particular area of the basin can contribute immediately to runoff. When snow starts
melting, there is a great change in scattering mechanism in comparison with the
dry snow cover. The radar can not “see” through snow-pack since the penetration
depth is about a radar wavelength (a few centimeters) for typical wet snow cover at
C-band.
Radar backscattering measurements from wet snow are affected by two sets of
parameters: (1) sensor parameters, which include the frequency, polarization, and
viewing geometry, and (2) snow parameters, which include snow density, liquid
water content, particle sizes and shapes of ice and water inclusions, type of the cor-
relation function, and surface roughness. It has been understood that the backscat-
tering from wet snow cover is mainly controlled by snow volume backscattering and
the surface backscattering at air–snow interface.
• The volume scattering is inversely correlated to snow wetness. The liquid water
content mainly causes an increase of snow permittivity because of the high
dielectric contrast between ice and water. This results in a significant (1) de-
crease in transmission at the air–snow interface and (2) high dielectric loss, which
greatly increases the absorption coefficient.
• The surface scattering is proportional to snow wetness. As the liquid water con-
tent increases, the reflectivity at air–snow interface increases greatly so does the
surface backscattering.
The effects of snow wetness on two major scattering signals are in opposite direc-
tions and the actual relationships between radar measurements and snow wetness
depend on which scattering component is dominant scattering source. If the vol-
ume scattering is dominant scattering source, the radar measurements will show a
negative correlation to snow wetness. However, the radar measurements will have a
positive correlation to snow wetness if the surface scattering is dominant scattering
source. The relationship between backscattering and snow wetness, however, can
be either positive or negative, depending on snow characteristics and surface rough-
ness and on incidence angle. Generally, as snow wetness increases the backscatter-
ing decreases rapidly for low wetness (<3%) conditions because volume scattering
albedo decreases. When wetness is low, the dielectric contrast between air and snow
3 Active Microwave Remote Sensing Systems and Applications to Snow Monitoring 31

is small and volume scattering dominates, so backscattering is not sensitive to sur-


face roughness. As snow wetness further increases, backscattering becomes sensi-
tive to surface roughness. This is because the surface scattering component becomes
dominating, resulted from rapidly increasing surface scattering component and de-
creasing volume scattering component. This complexity of the relationship between
the backscattering and snow wetness makes it unrealistic to develop an empirical
relation between the radar signal and field measurements.

3.3.2 Snow Wetness Algorithm Using C-Band Polarimetric SAR


Measurements

An algorithm has been developed for snow wetness retrieval using C-band polari-
metric SAR imagery (Shi and Dozier, 1995). This algorithm is developed using
a database, which covers the most possible wet snow physical properties includ-
ing the wide ranges of snow wetness, density, particle size and surface roughness,
simulated by the first-order scattering model with both surface and volume scatter-
ing components. The major developments in this algorithm were: (1) a simplified
surface backscattering model, that describes the relationships between the different
polarization measurements for the conditions of most seasonal wet snow covers,
to minimize the surface roughness effects with multi-polarization measurements;
and (2) the property of the volume scattering ratio in co-polarizations which is only
a function of snow permittivity and incidence angle to minimize the volume scat-
tering albedo effects on estimation of snow wetness. The more details on how the
algorithm was developed can be found in (Shi and Dozier, 1995). The final inverse
model for estimating snow wetness by using three C-band measurements σtvv , σthh ,
and σtvvhh is given as

M1 [avx Re[αvv αhh ] × (avhx DRS − DT S ) + bvx M2 ]
 
∗ bvx M2
= M2 avx Re[αvv αhh ]+ − DTV |αvv |2 (3.3)
avhx DRS − DT S

with

M1 = σtvvhh − DTV (θi , εs ) × σtvv , M2 = σtvv + σthh − DT S (θi , εs ) × σtvvhh


DTV (θi , εs ) + DT H (θi , εs ) |αvv (θi , εs )|2 + |αhh (θi , εs )|2
DT S = , DRS = ∗ (θ , ε )]
DTV (θi , εs ) × DT H (θi , εs ) Re[αvv (θi , εs )αhh i s

where σtvvhh = Re[Stvv Sthh∗ ] is the real part of the cross product of VV and HH
complex scattering elements, αvv and αhh are the polarization amplitudes (Ulaby
et al., 1982). avx (θi ), avhx (θi ), and bvhx (θi ) are coefficients derived from statistical
analyses and depend only on incidence angle. They are given in (Shi and Dozier,
1995). DTV (θi , εs ) and DT H (θi , εs ) denote the volume backscattering ratios for
Re[VVHH∗ ] to VV and Re[VVHH∗ ] to HH which are only depending on the local
32 J. Shi

incidence angle and the dielectric constant of a wet snowpack. This concept is from
the first-order volume backscattering model for an inhomogeneous dielectric half
space medium:
3  
σvpp = ω T2pp · exp −2s2 (k1 cos(θi ) − k2 cos(θi ))2 (3.4)
4
Tvv and Thh are power transmission coefficients for a plane interface for vertical and
horizontal polarization. s is the standard deviation of the random surface height.
The loss factor exp[−2s2 (k1 cos(θi ) − k2 cos(θi ))2 ] is the rough surface effect on
the power transmission coefficient (Ulaby et al., 1986). ω is the volume scattering
albedo, which depends on snow density, wetness, particle size, size variation and
shape. Under the assumption of spherical grains or randomly orientated particles,
the volume scattering albedo is independent of the polarization. Therefore, the ratio
for the first-order volume backscattering signals of VV and HH polarizations can be
represented as a function of the local incidence angle and dielectric constant:

σvvvhh Re [Tvvhh (θi , εs )]2 σvvvhh Re [Tvvhh (θi , εs )]2


DTH (θi , εs ) = = , DTV (θi , εs ) = =
σvhh 2 (θ , ε )
Thh i s σv vvh Tvv2 (θi , εs )

The algorithm derived above requires no information about the volume scattering
albedo or the surface roughness parameter. With known local incidence angle, it
involves only the calculation of snowpack permittivity, which can be directly re-
lated to snow wetness. This algorithm is applicable to the situations of incidence
angle from 25◦ to 70◦ , and the snow surface roughness – rms height <0.7 cm and
correlation length <25 cm.
Figure 3.2 shows the comparison between the field snow wetness measurements
(x-axis) and the derived snow wetness from C-band polarimetric SAR data of SIR-C
(black squares) and AIRSAR (white triangles) at the study site around Mammoth
Mountain, California during the first SIR-C/X-SAR mission in April, 1994. Those
included eight snow pits and two transects. The SAR-inferred snow wetness val-
ues were obtained from an average value of 3 × 3 windows around the snow pit
locations. Snow wetness measurements from transects were also averaged to 100 m
scales and the inferred snow wetness from SAR was determined from a mean value
of 4 × 2 windows along the transect measurements. Most SAR-derived snow wet-
ness agreed well with estimates of snow wetness. The standard deviation of absolute
error was about 1.3% by volume which gives 2.5% error bars at 95% confidence in-
terval for absolute error. The algorithm performed well on both local and regional
scales and provided a quantitative estimate of spatial distribution of snow wetness
at the top snow layer.
As we noticed that the above algorithm requires the C-band polarimetric SAR
measurements and can not be directly applied to ASAR image data since it does
not provide the measurement of Re[Svv Shh ∗ ]. In addition, the pair of co-polarization
measurements VV and HH have little difference at small incidence angle, which
results in VV and HH measurements that are almost identical regardless of snow
wetness and surface roughness conditions. Therefore, the measurements of VV and
3 Active Microwave Remote Sensing Systems and Applications to Snow Monitoring 33

10

8
Estimated
6

2 Sirc
Airsar

0
0 2 4 6 8 10
Measured snow wetness in %
Fig. 3.2 Comparison between the field measurements and the derived snow wetness using C-band
SIR-C’s and NASA/JPL AIRSAR’s polarimetric SAR data that are represented as black squares
and white triangles in the plot

HH only provide the opportunity for estimation of snow wetness where the pixels
with the moderate to large incidence. Further studies are needed to modify the above
algorithm for using only two (ASAR) polarization measurements.

3.4 Estimates of Snow Water Equivalence from SAR

Snow water equivalence (SWE), the product of snow density and depth, is the most
important parameter for hydrological study because it represents the amount of
water potentially available for runoff. Measurement of the amount of water stored
in the snowpack and forecasting the rate of melt are thus essential for management
of water supply and flood control systems. Because of rough, irregular topography,
these attributes exhibit large spatial variability over alpine drainage basins, making
it impractical to gather enough in situ measurements. Snow density and depth are
generally not highly correlated, so they must be considered as independent vari-
ables in field surveys of the spatial distribution of snow and in the study of snow’s
microwave backscattering properties.
In studies of active microwave remote sensing from SAR on retrieval SWE,
Bernier et al. (1998) related the multi-temporal C-band backscattering measure-
ments over frozen agriculture fields to thermal resistance in the snowpack. The rela-
tionship between thermal resistance and SWE was found and used to estimate SWE.
However, it is commonly known that dry snow is almost transparent at C-band. The
snowpack only contributes a small amount of signal so that the sensitivity of C-band
measurements to SWE is poor. The more studies may be needed to demonstrate the
physics of this technique and its capability to be applied to other areas. This chapter
34 J. Shi

mainly summarizes the other two types of techniques that employ different SAR
measurement properties and with strong physical principles:
1. SAR backscattering technique with Multi-frequency (L, C, and X bands) and
dual-polarization (VV and HH) measurements (Shi and Dozier, 2000a, b).
2. Repeat-pass SAR interferometric technique: The differential phase shift mea-
surements from repeat-pass SAR interferometry can be directly related to SWE
(Rott et al., 2004) or its changes (Guneriussen et al., 2001).

3.4.1 Estimation of SWE with Multi-frequency and Polarization


SAR Backscattering Measurements

During the past years, theoretical modeling of snowpack microwave backscatter has
achieved advanced significantly with respect to understanding snowpack extinction
properties and the interactions of the microwave signal with the snowpack vol-
ume and surfaces. The dense media theory considers a snowpack as consisting
of random, discrete, spherical particles of a single size and permittivity (Tsang
et al., 1985), accounting for coherent scattering (near-field effect) within the snow-
pack, and has been extended to consider multi-size and multi-permittivity randomly
distributed spherical particle systems and systems where the particles are not ran-
domly distributed but may have a tendency to form clusters and bonds (Tsang,
1992; Ding et al., 1994). This approach is appropriate for application to the snow-
pack microstructure because snow metamorphic processes can give rise to clusters
or aggregations of ice grains. There has also been significant improvement in the
modeling of surface scattering characteristics. The Integral Equation Model (IEM)
(Fung, 1994) and the more recent Advanced Integral Equation Model (AIEM) allow
characterization of a wider range of the surface roughness conditions than past mod-
els. Validations of AIEM (Wu et al., 2001; Li et al., 2002; Chen et al., 2003) have
demonstrated significant improvement in modeling surface scattering and emission
for microwave remote sensing of land surfaces. These efforts have established a
fundamentally-improved understanding of the effects of snow physical parameters
and underlying surface dielectric and roughness properties on the microwave mea-
surements of snow-covered terrain, making it possible to characterize the microwave
backscatter behaviors more accurately.
Radar backscattering coefficient measurements at a given incidence angle θi
over seasonal snow covered terrain can, generally, be expressed as a four compo-
nent model:

σ pq
t
( f ) = σ pq
a
( f ) + σ pq
v
( f ) + σ pq
gv
( f ) + Tp · Tq · L p · Lq · σ pq
g
(f) (3.5)

The total backscattering signals consist of the surface backscattering from the in-
terfaces at the air–snow interface and at the snow–ground interface, direct volume
backscattering from snow-pack, the interaction term between snow volume and the
snow–ground interface, and are represented by the superscript, t, a, g, v, and gv. The
3 Active Microwave Remote Sensing Systems and Applications to Snow Monitoring 35

f is the radar frequency. The subscript p and q represents polarization status of the
observed radar signals. T is the power transmittivity at the air–snow interface. σg pq
is the surface backscattering signal at snow–ground interface. Lp = exp[−τp / cos θr ]
is the snow-pack attenuation factor. τ = κe d is the optical thickness, which is the
product of snow extinction coefficient – κe and snow depth – d. To carry out a for-
ward simulation of a dry snow covered terrain Eq. (3.5) requires a total of 13 surface
and snow parameters. The six snowpack parameters include snow depth, density, ice
particle size, size variation, stickiness, and temperature. In addition to the dielectric
constant of the ground, the six surface roughness parameters include RMS height,
correlation length, and the two-parameter correlation functions at the air–snow and
the snow–ground interfaces. The importance of each scattering component depends
on the sensor’s frequency, polarization, incidence geometry, and snow properties.
Based on the frequency dependence of SAR measurements to snow and ground
properties, Shi and Dozier (2000a, b) developed a multi-frequency (L, C, and X
bands) and dual-polarization (VV and HH) technique to estimate SWE and applied
SIR-C/X-SAR image data. This technique uses L-band measurements to estimate
snow density and the underground dielectric and roughness properties. The rela-
tionship between underground backscattering signals at C-band and X-band can
be estimated with the dielectric and roughness properties estimated from L-band
measurements. Then, using C-band and X-band measurements with the minimized
effects of the underlying backscattering signals estimate snow depth and ice particle
size. This technique requires all 3 SIR-C/X-SAR frequency measurements.

3.4.1.1 Estimation of Snow Density with L-Band Dual-polarization SAR

Effects of Snow on L-Band SAR Measurements

Volume scattering and extinction from dry snow are all very small at L-band. At mi-
crowave frequencies, the absorption coefficient (the imaginary part of the dielectric
constant) of ice is small, and snow grains are also small compared to an incident
L-band wavelength (24 cm). During the past two decades, little attention has been
paid to microwave interactions of snowpack at L-band (1.25 GHz) frequencies or
lower, while much work has been done at C-band (5.5 GHz) or higher frequencies
(Rott and Mätzler, 1987; Ulaby et al., 1982; Stiles and Ulaby, 1980; Ulaby and
Stiles, 1980; Ulaby et al., 1984; Kendra et al., 1998). Assuming there are no large
structures including the melting/frozen draining fingers and channels in the snow-
pack which may occur during snow melting season, snow grains can not generate
significant volume scattering at L-band since snow grains are much smaller than
incident L-band wavelengths. Under this condition, we can simplify the backscat-
tering model by considering a dry and homogeneous snowpack over a bare soil or
rock surface and Eq. (3.5) becomes

σqp
t
(k0 , θi ) = Tq (θi ) · Tp (θi ) · σqp
g
(k1 , θr ) (3.6)
36 J. Shi

θi , θr , and k1 represent the radar wave incidence angle at the air–snow interface, co-
sine propagation angle in snowpack, and radar wave propagation number in snow-
pack for the coherent component, respectively. It indicates that the L-band radar
measurements over season dry snow cover are affected by two sets of properties:
(1) snow properties through the power transmittivity at the air–snow interface T
which is a function of snow density, polarization, and incidence angle and (2) sur-
face backscattering properties at snow–ground interface through surface parameters
including the dielectric constant and surface roughness (roughness is generally de-
scribed by an autocorrelation function and standard deviation of surface roughness
height) for a random rough surface with no orientation of features.
When the electromagnetic wave passes through the snowpack, versus directly
striking the ground, the following differences occur:
• Because of refraction within the snow, the incidence angle at the snow–ground
interface is smaller.
• The incident wavelength at the snow–ground interface is shorter because the
snow is dielectrically thicker than air.
• The snow layer reduces the dielectric contrast at the snow–ground interface,
which in turn reduces the reflectivity at snow–ground interface.
• The power loss at the air–snow interface reduces the total energy incident on the
snow–ground interface.
The first two factors result in a change of the sensor observing parameters. Note that
the dielectric contrast εg /εs should be used instead of εg and that k1 should be used
g
instead of k0 when calculating the backscattering σqp at the snow–ground interface.
√ √
The shift in wavenumber k1 = k0 εs (or wavelength λ 0 = λ1 εs ) is a function of
snow density. For the range of snow densities considered −100 to 550 kg m−3 –
the L-band propagation wavelength in snow ranges from 21 to 16 cm, compared to
24 cm in air. Because the surface roughness effect depends on its size relative to
the incident radar wavelength, the shortening of the incident wavelength for higher
snow densities will result in the soil surface appearing rougher than it would if the
snow were absent. This causes an increase in the surface backscattering signal and
is especially strong for a nearly smooth surface. However, this effect caused by
the wavelength shift becomes smaller when the surface is rougher or the incident
frequency is higher.

Snell’s law specifies the change in incidence angle: sin(θi ) = sin(θr ) εs . The
incidence angle at the snow–ground interface depends only on the incidence angle
at the snow surface and the dielectric constant of snow, not on its thickness, thus
it is a function of snow density. For a given incidence angle at the snow surface,
a greater snow density causes a greater change in the incidence angle at the snow–
ground interface. For a given snow density, however, a larger incidence angle at the
snow surface results in a greater change in the refractive angle in the snow layer.
Therefore, a greater increase in the backscattered power at larger incidence angles
is expected than that at smaller incidence angles.
Furthermore, the effects of snow density at different co-polarizations HH and VV
shows a smaller increase in VV polarization than in HH polarization for the same
3 Active Microwave Remote Sensing Systems and Applications to Snow Monitoring 37

snow density and ground surface properties. The backscattering of VV polarization


as a function of incidence angle declines more slowly than that of HH polarization
when the surfaces are not too rough. Therefore, the changes in VV polarization are
smaller than that of HH polarization. However, for a very rough surface, the angular
dependence is smaller than that for a smooth surface, and the difference between
VV and HH polarization is smaller.
While a dry snowpack does not absorb or scatter the radar signal at low frequen-
cies, it nevertheless affects the magnitude of the backscattering from the underlying
rock or soil and the relationship between HH and VV polarization. The magnitude of
the effect depends on the radar incidence angle, snow density, roughness and dielec-
tric properties of the soil. Snow is more likely to enhance the backscattering mag-
nitude of a smooth soil than a rough soil. These factors enable development of an
algorithm for inferring snow density using L-band SAR measurements. At C-band
or higher frequencies, however, snow density affects the magnitude of the volume
scattering and the surface scattering properties at the snow–ground interface.

Snow Density Algorithm with L-band Dual-polarization SAR Measurements

Based on the above understanding of snow density effects on radar backscattering


at L-band, Shi and Dozier (2000a) develop an algorithm to estimate snow density
by characterizing the dependence of the surface backscattering on both the inci-
dence angle and the wavelength. It was done by establishing a backscattering σ g hh
and σ g vv database using the IEM model (Fung, 1994) over a wide range of in-
cidence angles, dielectric and roughness conditions, and incidence wavenumbers,
corresponding to a range of snow densities from 100 to 550 kg m−3 . Then, the re-
lationship of HH and VV backscattering signatures with the wide range of surface
dielectric and roughness conditions at each incidence angle and wavenumber were
characterized by using regression analysis to find coefficients to parameterize this
relationship:
  g
g  g   g
log10 σhh + σvv = a(θr , k1) + b(θr , k1) log10 σhh + σvv
g
+ c(θr , k1) log10 σhh
g
g
2
σhh σhh
+ d(θr , k1 ) log10 g + e(θr , k1 ) log10 g
σvv σvv
(3.7)

It represents a relationship between the surface backscattering coefficients σ g hh and


σ g vv at a given incidence angle and wavenumber for a wide range of random rough
surfaces. The form of the relationship minimizes its sensitivity to the surface di-
electric and roughness properties, while maximizing its sensitivity to the incidence
angle and wavenumber. The coefficients a, b, c, d and e in Eq. (3.7) depend only on
incidence angle and wavenumber at the ground surface. They are given in (Shi and
Dozier, 2000a).
38 J. Shi

By placing Eq. (3.6) in Eq. (3.7), the algorithm for estimation of snow density by
using only σhh
t and σ t SAR measurements can be derived:
vv
 
σhht σvvt
log10 +
Thh (θi , εs ) Tvv (θi , εs )
 
σhh
t
σvv
t
= a(θr , k1 ) + b(θr , k1 ) log10 +
Thh2 (θ , ε )
i s Tvv2 (θi , εs )
   t 2 
σt σ Tvv (θi , εs )
+c(θr , k1 ) log10 2 hh + d(θr , k1 ) log10 hh
Thh (θi , εs ) σvvt T 2 (θ , ε )
hh i s
 2
σhh
t T 2 (θ , ε )
vv i s
+e(θr , k1 ) log10 (3.8)
σvv
t T 2 (θ , ε )
hh i s

In Eq. (3.8), Tpp depends on the polarization pp, the incidence angle θi at the air–
snow interface, and the dielectric constant of snowpack εs . εs is the only unknown;
θi can be calculated from a combination of the spaceborne orbital data and a digital
elevation model. Therefore, for a given L-band SAR measurements of σ t hh and
σ t vv , we can estimate εs numerically by varying the coefficients of a, b, c, d, and e
to find the root of Eq. (3.8). It does not require a priori knowledge of the dielectric
and roughness properties of the soil under the snow. Furthermore, snow density
can be estimated from Looyenga’s semiempirical dielectric formula (Looyenga,
1965), which provides a good fit to Polder and van Santen’s physical formula
(Mätzler, 1996).
εs = 1.0 + 1.5995ρs + 1.861ρs3 (3.9)
Three spatial distributed snow density maps were derived by this technique us-
ing three SIR-C/X-SAR’s L-band data-takes from its first mission in April 1994
at Mammoth Mountain study site. The three data-takes were acquired in the early
morning around 6 a.m. while the snow pit measurements taken at 11 a.m. showed
no signs of liquid water in the snow, even at 2,850 m elevation. Thus the imaged
snowpacks were dry. Using the Space Shuttle orbital geometry data and a digital
elevation model, the terrain radiometric calibration factor and local incidence angle
images corresponding to the SIR-C image data were derived. The terrain radiomet-
ric correction factor is sin θ0 / sin θi , where θ0 is the incidence angle used in the
initial SAR data processing under a flat surface assumption and θi is the actual lo-
cal incidence angle (Van Zyl et al., 1993). To reduce the effect of image speckle
on estimation of snow density, the Stokes matrixes were first averaged to azimuth
(25.1 m) and slant range (13.3 m) pixel spacing to form multi-look imagery. Then,
the terrain correction was made. The snow-free pixels were also masked out based
on the snow classification results derived from SIR-C/X-SAR’s image data (Shi and
Dozier, 1997).
Figure 3.3 shows the comparison of the field measurements (x-axis) with SIR-
C’s L-band image data derived snow density (y-axis). The RMSEs are 42 kg m−3
3 Active Microwave Remote Sensing Systems and Applications to Snow Monitoring 39

0.4
Estimated

0.3

0.2

0.2 0.3 0.4


Measured snow density
Fig. 3.3 Compares of the field measurements (x-axis) with SIR-C/X-SAR’s L-band image data
(y-axis) derived snow density

and 13% for absolute and relative errors, respectively. It should be noticed that the
estimated snow density should represent the mean value from the snowpack’s top
and bottom layers. This is because the backscattering signal is mainly controlled by
two factors: (1) the snow density at the top layer near the surface mainly affects the
power transmissivity at the air–snow interface; (2) snow density at the bottom layer
controls the incidence angle and wavenumber at the snow–ground interface.
It should be noticed that the snow density algorithm described above is based on
a radiative transfer model concept – Eq. (3.6) with an assumption of the wavelength
shift when radar wave propagates in snowpack without regarding how thick of snow
is. It is understandable if snow depth is too thin, it may not result in the notable wave-
length shift. Unfortunately, the current used technique does not allow the evaluation
of this issue. The studies of the well controlled laboratory experiment or Monte
Carlo Model simulations are needed for further quantitative identification on what
fraction of snow thickness to radar wavelength will result in the wavelength shift.

3.4.1.2 Technique on Estimation of Snow Water Equivalency


with C- and X-Bands

Relationships Between Snow Water Equivalence and SAR Measurements

Field experiments using ground scatterometer data have shown different rela-
tionships between radar backscattering and SWE. For example, Ulaby and Stiles
(1980) showed that backscattering at 8.2 and 17.0 GHz had a positive relation
with SWE. Similarly, a positive relationship was also observed by an experiment
over a smooth subsurface at 5.3 and 9.5 GHz (Kendra et al., 1998). However, this
40 J. Shi

positive relationship existed only over a frozen subsurface, with no correlation over
the unfrozen subsurface (Bernier and Fortin, 1998). On the other hand, negative
relationships have been observed at similar frequencies, 5.3 and 9.6 GHz (Strozzi,
1999). In addition, Rott and Mätzler (1987) observed no significant difference be-
tween snow-free and dry snow covered regions at 10.4 GHz. Each field experiment
represented particular snow and ground conditions. The existence of both positive
and negative relationships between radar backscattering and snow water equiva-
lence indicates that this relationship is quite complex. The confusion may result
from varying combinations of snow and ground properties, because backscattered
power received by the radar over dry snow depends not only on the total snow mass
but also on the snow’s density, grain size, structure, and stratification, along with the
dielectric and roughness properties of the underlying surface. Understanding this
relationship is essential to the development of a reliable algorithm for estimating
snow water equivalence.
In order to understand the general relationship between the radar measurements
at the frequency range between C-band and Ku-band and snow depth or SWE for
dry snow cover, for simplicity, we deploy a simple backscattering model “cloud
model” in which the surface scattering term from air–snow interface (the first term
in Eq. (3.5)) and the interaction term between the ground surface and the snow
volume (the third term in Eq. (3.5)) are ignored. Two remaining scattering terms –
the direct volume and ground surface backscattering components (the second and
fourth terms in Eq. (3.5)) are dominant backscattering signals. The former can be
described as a function of the optical thickness (τ – a product of the snow extinction
coefficient κe and snow depth d) and the volume scattering albedo (ω – a function
of the snow density, particle size, size variation, stickness, and temperature). It is
given in terms of the first-order backscattering model as


3 −2τ (f)
σpp
v
(f, θi ) = T2pp (θi , εs ) ω (f) cos(θr ) 1 − exp (3.10)
4 cos(θr )

In the direct volume backscattering signals, the τ is positively related to SWE. The
underground scattering component, however, is the product of the backscattering
signals from underground surface and snow-pack attenuation properties through the
optical thickness τ, which is negatively related to this scattering component. In other
words, the effects of SWE in these two major scattering components are in an op-
posite way. It results in the major confusion in describing the relationship between
SWE and radar observations.
To further demonstrate the general relationship, we take the first-order deriva-
tive of “cloud model” with respect to snow depth d that represents the slope of the
backscattering curves in response to snow depth; it can be written as

∂ σpp
t (θ )
i 2 2κ e 3
= Tpp exp(−2κ e d/µr ) · ω µr − σpp
g0
(3.11)
∂d µr 4

The ground surface backscattering signal has a great impact on the relationship
between the backscattering and snow depth. The sign of the bracketed term in
3 Active Microwave Remote Sensing Systems and Applications to Snow Monitoring 41

Eq. (3.11) determines whether the relationship between the measured backscattering
signal and snow depth is positive or negative. If σ g0 pp > 3/4ω µr , the slope is nega-
tive. Physically, this means that the attenuated subsurface scattering signal, after it
passes through the snowpack, is larger than the backscattering signal generated by
the snowpack. It is possible for the bracketed term to be zero, in which case there is
no correlation between backscattering and snow depth. In this case, the attenuated
amount of the subsurface scattering signal is the same as that of the backscattering
signal generated by snowpack. Even for the same snowpack, a positive relationship
to snow depth may be observed over a smooth surface but a negative relationship
will be observed over a rough surface.
Moreover, the sensitivity of the backscattering signal to snow depth depends not
only on the snow properties, but also on the incidence angle, and the magnitude of
the subsurface backscattering. From Eq. (3.11), we can also see that as the snow
depth increases, the change in backscattering will decrease. Therefore, we expect
that the change in the backscattering measurements is more sensitive to shallow
snow than to thick snow. The extinction coefficient and volume scattering albedo
are positively correlated to the sensitivity of backscattering to snow depth. For a
given snow depth, the larger these parameters are, the greater changes will be ex-
pected. Similarly, the sensitivity is also proportional to the angle of incidence, which
indicates that a large incidence should be more effective than a small incidence for
the purpose of monitoring snow depth. Note that the parameters, such as the ground
dielectric and roughness properties in the ground–surface and snow–volume interac-
tion terms, also affect the relationships. The combinations of all parameters control
the overall result.
Therefore, the backscattering signal from a seasonal natural snow cover and the
signal’s relationship to snow depth is affected by three sets of parameters:
1. Sensor parameters, which include the frequency, polarization, and incidence
geometry
2. Snowpack parameters including snow density, particle size and size variation,
free liquid water content, characteristics of particle spatial distribution (sticki-
ness), and stratification
3. Subsurface parameters that include the dielectric and roughness properties at the
snow–ground interface
These complex relationships make it implausible to characterize the parameters
from the limited field experiment measurements and to derive an empirical model
for estimating snow depth or water equivalence from SAR measurements. To es-
timate snow depth and thereby water equivalence, we must separate the varying
backscattering signals of the subsurface or minimize the effect of the backscatter-
ing signal generated by the snow–ground interface. It is clear that estimating the
snow depth requires a physically based inversion model that considers all important
scattering terms and a technique to separate the backscattering components of the
snowpack itself from those of the air–snow and snow–ground interfaces.
42 J. Shi

Snow Depth Algorithm with C-band and X-band SAR Measurements

As shown in (Shi and Dozier, 2000a), the snow density ρs , ground dielectric con-
stant εg , and surface RMS height s can be estimated by L-band SAR measurements.
From these parameters, we can obtain εs , θr , k1 , T pp , and the reflectivity Rpp at
the snow–ground interface. The remaining unknowns in Eq. (3.5) are the surface
g
backscattering components σpp ( f ) from the snow–ground interface and σa pp ( f )
from the air–snow interface, the extinction coefficients κe ( f ), the volume scatter-
ing albedo ω( f ), and the snow depth d. Through analyses of the simulated data,
the techniques for estimating snow depth and ice particle size using SIR-C C-band
and X-SAR measurements were developed. They were based on the following ap-
proaches (Shi and Dozier, 2000b):
Reducing the effect of surface backscattering signal from air–snow interface. In
considerations of each backscattering component from dry snow cover in Eq. (3.5)
for the SWE retrieval, the backscattering signals from the air–snow interface are
generally considered as the “noise” since they are typically small in comparison
with the other scattering contributions and have no information on SWE. Its effect
on retrieval snow properties can be reduced by using the signal generated by the
estimated snow density and typical roughness parameters.
Developing semi-empirical models to characterize the snow–ground interac-
tion terms. These models represent the snow–ground interaction components more
realistically than formulas developed under the assumption of independent scat-
tering. Natural surfaces in alpine regions are quite rough. Therefore, a significant
contribution to the snow–ground interactions from non-coherent components is ex-
pected. The semi-empirical models are developed explicitly in terms of snowpack
volume scattering albedo, optical thickness, ground reflectivity, and surface RMS
height. They are easy to implement in the inverse algorithm.
Developing semi-empirical model to characterize the relationships between the
ground surface backscattering components at C-band and X-band. With L-band
SAR estimates of the ground dielectric constant and RMS height, the relationships
between the ground surface backscatterings at C-band and X-band can be well char-
acterized. Thus, the unknowns in the ground surface backscattering formulation can
be reduced to one.
Parameterizing the relationships between snowpack extinction properties at
C-band and X-band. Because the extinction properties are highly correlated, we
developed analytical forms for the extinction relationships at C-band and X-band.
Using these relationships, the number of unknowns in backscattering components
from the snowpack can be reduced to two: the volume scattering albedo ω and the
optical thickness τ .
From the developments in the above, the number of unknowns has been reduced
to three in Eq. (3.5), that is τ, ω and the surface component: σsvv (X). Thus, these
three unknowns can be solved numerically at each pixel by using Eq. (3.5) with three
SAR measurements: σt vv (C), σt hh (C), and σt vv (X).
To estimate snow depth d, the separation between the extinction coefficient κe
and d in the estimated τ is needed. In other words, we need to determine the absolute
3 Active Microwave Remote Sensing Systems and Applications to Snow Monitoring 43

value of κe from the estimated parameters, i.e., τ( f ), ω( f ) and snow density ρs at


each pixel. This can be done by first estimating the absorption coefficient κa (X):

τ (X) [1 − ω (X)]
κa (X) = 1.334 + 1.2182 log(Vs ) − 3.4217 log (3.12)
τ (C) [1 − ω (C)]

where Vs is the volume fraction of ice and can be derived from the estimated snow
density at each pixel by using L-band measurements. τ [1 − ω ] = κa · d = τa is the
absorption part of the optical thickness at each frequency after removing scatter-
ing effects. ττ(X)[1−ω (X)] κa (X)
(C)[1−ω (C)] = κa (C) gives snowpack temperature information (Shi and
Dozier, 2000b). Then, snow depth can be estimated by

τ (X) [1 − ω (X)]
d= (3.13)
κa (X)

Figure 3.4 shows the comparison of the field measurements with the depths esti-
mated from the SIR-C/X-SAR image data at Mammoth Mountain study site. The
algorithm inferred the overall trend of the snow depths, with an RMSE of 34 cm.
During the SIR-C/X-SAR overflights, 19 snow pits and the grain size measure-
ments with detailed vertical profiles at depth intervals of 5–10 cm were obtained in
the field measurements. Each individual ice particle size was measured by taking
average values from three cross diameters with direction differences of 60◦ . There
were more than 200 grains measured in each snow pit. By taking one step further,
a formula was also developed for estimating the optically equivalent particle size,
defined as a particle size at which the extinction properties from a natural snow
volume equal those obtained from an ideal snow volume with a uniform distribu-
tion of ice particles (Shi et al., 1993). This optically equivalent particle size can be

300

250
Estimated

200

150

100

50

50 100 150 200 250 300


Measured snow depth in cm
Fig. 3.4 Compares of the field measurements (x-axis) with SIR-C/X-SAR’s C- and X-band image
data (y-axis) derived snow depth
44 J. Shi

2.0

1.5

Estimated
1.0

0.5

0.5 1.0 1.5 2.0


Measured grain size in mm
Fig. 3.5 Compares of the field measurements of snow particle radius with SIR-C/X-SAR’s C- and
X-band image data derived grain size

represented by the weighted mean size with respect to the scattering properties for
a natural snowpack and can be estimated by

1
0.01κs (X) 3
r̄s = (3.14)
2Vs S f (2.8332 + 6.6143Vs )

where r̄s is in cm. κs (X) can be obtained from the estimated κs (X) and ω (X) at
each pixel.
Figure 3.5 compares the snow particle radius between the ground measurements
and those estimated from the SIR-C/X-SAR image data. The estimated snow parti-
cle sizes were obtained from a mean value of 3 × 3 window at corresponding snow
pit locations on the SAR images. The algorithm performed quite well. It inferred
the overall trend of the snow particle sizes, with an RMSE of 0.27 mm.
The current technique described above for estimation of SWE requires five mea-
surements: L-VV and L-HH to estimate snow density and ground dielectric and
roughness properties, plus C-VV, C-HH, and X-VV to estimate snow depth and
grain size. The sensitivity analysis indicated that the C-band SAR measurements
were affected mainly by the ground surface properties. The parts of the signal that
comes from a typical snowpack at C-band are about 30% and 15% for HH and
VV polarization, respectively. The C-band measurements are expected mainly sen-
sitive to under-ground surface condition. Estimation of snow depth using C-band
SAR measurements, therefore, requires an accurate technique to estimate the ground
backscattering component. At X-band it about 60% of the signal comes from the
snowpack itself. Thus, we expect that the measurement is much more sensitive to
snowpack and that the requirement for estimation of the ground backscattering com-
ponent is less severe for radar measurements at X-band or higher.
3 Active Microwave Remote Sensing Systems and Applications to Snow Monitoring 45

3.4.2 Estimations of SWE and Its Change with SAR Interferometry

From last section, it can be seen that retrieving SWE from SAR backscattering mea-
surements requires very accurate inversion models and is a very complex process.
The other promising technique for SWE estimation is to use the repeat-pass SAR
interferometric measurements. It has been shown that the interferometric phase shift
resulting from differences in the propagation due to constant changes in snow prop-
erties offers a direct method for estimating SWE (Rott et al., 2004) or relative change
in SWE (Guneriussen et al., 2001). This technique has been applied to low frequency
interferometric SAR measurements at C-band (Guneriussen et al., 2001; Rott et al.,
2004) and L-band (Rott et al., 2004).
Considering a layer of dry snow, the radar return is dominated by scattering signal
from snow/ground interface at low frequencies (C-band or lower). The repeat-pass
interferometric phase Φ consists of following contributions (Rott et al., 2004):

φ = φ f lat + φtopo + φatm + φsnow + φnoise (3.15)

Φflat and Φtopo are the phase differences due to changes of the relative distance be-
tween satellite and target for flat earth and for topography, respectively. Φatm results
from changes in atmospheric propagation, and Φnoise is phase noise. Φsnow is the
two-way propagation difference in the snow-pack relative to air and is resulted from
the refraction of radar wave in dry snowpack. When snowpack volume scattering
is neglected, the snow phase term Φsnow for an uniform layer of snow with the
depth – d can be written as (Guneriussen et al., 2001):


φsnow = −2ki · d · cos θi − εs − sin2 θi (3.16)

For ERS SAR with an incidence angle θi = 23◦ , the phase shift due to a change of
SWE can be approximated by a linear relation (Guneriussen et al., 2001):

φsnow = −2ki · 0.87 · ∆SW E (3.17)

It indicates that at the ERS wavelength one fringe is equivalent to 32.5 mm SWE,
and for L-band (λ = 24 cm, θi = 23◦ ) one fringe corresponds to SWE = 138 mm.
The coherence is determined by several factors and given by Hoen and Zebker
(2000):
γtotal = γthermal · γsur f ace · γvolume · γtemporal (3.18)
γthermal depends on the signal-to-noise ratio. Its contribution to decorrelation of dry
snow covered ground is usually small since the signals from dry snow ground cov-
ered are generally strong. For a dry snow cover, the wave number shifts in slant
range and in vertical direction have to be considered, which affects mainly on
the terms of γsur f ace and γvolume . Their calculations are given by Hoen and Zebker
(2000). The temporal decorrelation is a major effect on in coherence of repeat pass
SAR data over dry snow covered areas. It is mainly resulted from changes of the
46 J. Shi

snowpack properties due to snow fall or snow drift. They result in changes of the
snowpack structure and the roughness of the snow-surface at sub-pixel scale, and
consequently change the propagation path in the snowpack of the radar return from
each surface element. Rott et al. (2004) shows the term of the temporal correlation
for dry snow cover can be expressed as
 

 
γtemporal  = exp −2 · ki2 · σz2 cos θi − εs − sin2 θi (3.19)

where σz is the standard deviation of the geometric path length through the snow-
pack. It takes into account of differential phase delay due to non-uniform snow
accumulation or erosion.
Guneriussen et al. (2001) applied this technique using ERS-1/2 tandem data and
found that the repeat-pass interferometric phase measurements were very sensitive
to small changes in snow properties that could introduce a significant error in DEM
estimation even in the case of high degree of coherence. To avoid the temporal
decorrelation effects, Rott et al. (2004) used ERS repeat pass data without snow-
fall between data acquisitions and derived SWE map over their study site using this
technique with the required corrections for all other phase components in Eq. (3.14)
and the reference points with zero and known snow accumulation. Both studied have
shown that the interferometric phase shift from repeat-pass SAR measurements pro-
vides a physically based means for mapping SWE of dry snow. Temporal decorre-
lation due to differential phase shifts at sub-pixel scale caused by snow fall or wind
re-distribution is the major limitation for application of this method. In C-band SAR
data, these effects often result to complete decorrelation within a few days. L-band
is preferable for this technique on mapping SWE than C-band because it is less
affected by temporal decorrelation, better coherence and larger measurement range.

3.5 Need for Future Spaceborne System for Monitoring Snow


Properties

The current available spaceborne C-band imaging radar systems are either single po-
larization (ERS-2 with VV polarization and RADARSAT-1 with HH polarization) or
dual-polarization ASAR (HV/VV or VH/HH). For monitoring snow properties with
backscattering measurements, these current systems are only capable to identify the
extent of wet snow. They are not capable of measuring SWE or of quantifying snow
wetness. For interferometric measurements, the current and near future satellite sys-
tems, however, have the repeat orbit at least 24 days or longer. The presented SWE
estimation technique in Section 3.4.2 requires the acquisitions within a few days
and is only applicable to ERS 1/2 tandem data. For using interferometric technique
to estimate SWE, the low frequency SAR instruments (L-band) is required with a
quick repeat time interval. The recent launched PALSAR system has 48 days re-
peat pass and does not match this requirement. However, it provides the unique
tool for mapping both dry and wet snow cover with the combined interferometric
3 Active Microwave Remote Sensing Systems and Applications to Snow Monitoring 47

coherence and backscattering measurements as presented in Section 3.2.3 but do


not offer capabilities to estimate SWE and snow wetness over the range of Earth’s
environments.
Taking into account the capabilities and deficiencies of present remotely sensed
snow cover information, its is obvious that improvements are necessary in particular
regarding the observation of extent and physical properties of snowpacks – SWE and
wetness. Sufficiently high spatial resolution is needed to account for the small scale
heterogeneity of snowpack related to topography and land cover.
For using backscattering signals, two critical capabilities of the radar instrument
are required for monitoring snow water equivalence (SWE) globally. First, the radar
signal can penetrate the natural snowpacks so that the snow depth information can be
obtained from the measurements. Secondly, the radar signal from snowpack itself
has to be large enough to provide the first-order information for snowpack itself.
At the frequencies at or lower than C-band, snowpack does not generate the sig-
nificant volume scattering signals and the attenuation of the ground signals also
weak (Shi and Dozier, 2000b). At Ku-band (13–19 GHz), the penetration depths
are ranged from 0.5 (extreme large grain) to 10 m (extreme small grain). The stud-
ies have shown the one-way penetration depth of dry snow from experimental data
which at 17 GHz ranges from about 4 m for coarse grained Antarctic firn (Rott et al.,
1993) to 6 m for fine-grained Alpine winter snow (Mätzler, 1987). Secondly, the
direct volume scattering component contributes about 60% radar measurement sig-
nals in the co-polarization for a typical dry snow condition at X-band 9.25 GHz
(Shi and Dozier, 2000b). The numerical simulations (Shi et al., 2003; Shi, 2004)
have indicated that the direct volume contribution will be significantly increased in
both co-polarized and cross-polarized signals at Ku-band 17 GHz. The direct vol-
ume scattering becomes the dominant scattering source. These numbers indicate that
17 GHz is a good choice for sensing volume properties of dry snow and firn, because
at this frequency the snow-pack provides a clear signal but is optically transparent
enough to achieve sufficient penetration. Recent studies of the global snow cover
with NSCAT, operating at 14 GHz also demonstrated the sensitivity of Ku-band to
sense dry and wet snow cover (Nghiem and Tsai, 2001). Therefore, we expect that
X and Ku-band instruments have capability to penetrate the most of natural dry
snow cover globally and provide much more sensitive to snowpack itself than those
by C-band. For future satellite mission on monitoring SWE over range of Earth’s
environments, the optimal sensor configuration and corresponding algorithm devel-
opment are needed.

References

Bernier M, Fortin JP (1998) The potential of time series of C-band SAR data to monitor dry and
shallow snow cover. IEEE Trans. Geosci. Remote Sens. 36:226–243
Chen KS, Wu TD, Tsang L, Li Q, Shi J, Fung AK (2003) The emission of rough surfaces calculated
by the integral equation method with a comparison to a three-dimensional moment method
simulations. IEEE Trans. Geosci. Remote Sens. 41(1):90–101
48 J. Shi

Ding K, Zurk LM, Tsang L (1994) Pair distribution functions and attenuation rates for sticky
particles in dense media. J. Electromagnet. Waves Appl. 8:1585–1604
Evans D, Farr TG, van Zyl JJ, Zebker HA (1988) Radar polarimetry: analysis tools and applica-
tions. IEEE Trans. Geosci. Remote Sens. 26:774–789
Evans D, Plaut J, Stofan E (1997) Overview of the spaceborne imaging radar-c/x-band synthetic
aperture radar (SIR-C/X-SAR) missions. Remote Sens. Environ. 59(2):135–140
Fung AK (1994) Microwave scattering and emission models and their applications. Artech House,
Norwood, MA
Guneriussen T, Høgda KA, Johnsen H, Lauknes I (2001) InSAR for estimation of changes in snow
water equivalent of dry snow. IEEE Trans. Geosci. Remote Sens. 39(10):2101–2108
Haefner H, Holecz F, Meier E, Nusch D, Piesbergen J (1993) Capabilities and limitations of ERS-1
SAR data for snow cover determination in mountainous regions. Proc. Second ERS-1 Sympo-
sium, ESA PS, pp 353–359
Hoen EW, Zebker HA (2000) Penetration depths inferred from interferometric volume decor-
relation observed over the Greenland ice sheet. IEEE Trans. Geosci. Remote Sens.
38(6):2571–2583
Kendra JR, Sarabandi K, Ulaby FT (1998) Radar measurements of snow: experiment and analysis.
IEEE Trans. Geosci. Remote Sens. 36:864–879
Lee JS, Papathanassiou KP, Ainsworth TL, Grunes MR, Reigber A (1998) A new technique
for noise filtering of SAR interferometric phase images. IEEE Trans. Geosci. Remote Sens.
36(5):1456–1465
Li Q, Shi J, Chen KS (2002) A generalized power law spectrum and its applications to the
backscattering of soil surface on the integral equation model. IEEE Trans. Geosci. Remote
Sens. 40(2):271–280
Looyenga H (1965) Dielectric constant of heterogeneous mixtures. Physica 21:401–406
Luojus K, Pulliainen J, Metsamaki S, Hallikainen M (2006) Accuracy assessment of SAR data-
based snow-covered area estimation method. IEEE Trans. Geosci. Remote Sens. 44(2):277–287
Mätzler C (1987) Applications of the interaction of microwaves with the natural snow cover. Re-
mote Sens. Rev. 2:259–387
Mätzler C (1996) Microwave permittivity of dry snow. IEEE Trans. Geosci. Remote Sens.
34:573–581
Nagler T, Rott H (2000) Retrieval of wet snow by means of multitemporal SAR data. IEEE Trans.
Geosci. Remote Sens. 38(2):754–765
Nghiem SV, Tsai WY (2001) Global snow cover monitoring with spaceborne Ku-band Scattero-
meter. IEEE Trans. Geosci. Remote Sens. 39(10):2118–2134
Pulliainen J, Engdahl M, Hallikainen M (2003) Feasibility of multitemporal interferometric
SAR data for stand-level estimation of boreal forest stem volume. Remote Sens. Environ.
85:397–409
Rott H, Mätzler C (1987) Possibilities and limits of synthetic aperture radar for snow and glacier
surveying. Annals Glaciology 9:195–199
Rott H, Nagler T (1993) Capabilities of ERS-1 SAR for snow and glacier monitoring in alpine
areas. Proc. Second ERS-1 Symposium, pp 1–6
Rott H, Mätzler C, Strobl D, Bruzzi S, Lenhart KB (1988) Study on SAR land applications for
snow and glacier monitoring. Technical Report 6618/85/F/FL(SC), European Space Agency
Rott H, Nagler T, Scheiber R (2004) Snow mass retrieval by means of SAR interferometry Proc.
FRINGE 2003 Workshop, ESA SP-550, article-46, Frascati, Italy
Shi J (2004) Estimation of snow water equivalence with two Ku-band dual polarization radar. Proc.
IGRASS’04, IEEE No. 04CH37612C
Shi J, Dozier J (1995) Inferring snow wetness using C-band data from SIR-C’s polarimetric syn-
thetic aperture radar. IEEE Trans. Geosci. Remote Sens. 33(4):905–914
Shi J, Dozier J (1997) Mapping seasonal snow with SIR-C/X-SAR in mountainous areas. Remote
Sens. Environ. 59(2):294–307
Shi J, Dozier J (2000a) Estimation of snow water equivalence using SIR-C/X-SAR, part I: inferring
snow density and subsurface properties. IEEE Trans. Geosci. Remote Sens. 38(6):2465–2474
3 Active Microwave Remote Sensing Systems and Applications to Snow Monitoring 49

Shi J, Dozier J (2000b) Estimation of snow water equivalence using SIR-C/X-SAR, part II: infer-
ring snow depth and particle size. IEEE Trans. Geosci. Remote Sens. 38(6):2475–2488
Shi J, Jiang L (2004) Numerical simulations on estimation of snow wetness with dual-frequency
and polarization radar. Proc. SPIE, Honolulu 5654:149–156
Shi J, Davis RE, Dozier J (1993) Stereological determination of dry snow parameters for discrete
microwave modeling. Annals Glaciology 17:295–299
Shi J, Dozier J, Rott H (1994) Snow mapping in alpine regions with synthetic aperture radar. IEEE
Trans. Geosci. Remote Sens. 32(1):152–158
Shi J, Hensley S, Dozier J (1997) Mapping snow cover with repeat pass synthetic aperture radar.
Proc. IGARSS ’97, IEEE No. 97CH36042, vol II, pp 628–630
Shi J, Yueh S, Cline D (2003) On estimation of snow water equivalence using L-band and Ku-band
radar. Proc. IGRASS’03, IEEE No. 03CH37477C
Stiles WH, Ulaby FT (1980) The active and passive microwave response to snow parameters, 1,
wetness. J. Geophys. Res. 85:1037–1044
Strozzi T, Mätzler C (1998) Backscattering measurements of Alpine snowcovers at 5.3 and 35 GHz.
IEEE Trans. Geosci. Remote Sens. 36(3):838–848
Strozzi T, Wiesmann A, Mätzler C (1997) Active microwave signatures of snowcovers at 5.3 and
35 GHz. Radio Sci. 32(29):479–495
Strozzi T, Wegmuller U, Matzler C (1999) Mapping wet snow covers with SAR interferometry.
Int. J. Remote Sens. 12(20):2395–2403
Tsang L (1992) Dense media radiative transfer theory for dense discrete random media with parti-
cles of multiple sizes and permittivities. Prog. Electromagnetic Res. 6(5):181–225
Tsang L, Kong JA, Shin RT (1985) Theory of microwave remote sensing. Wiley, New York
Ulaby FT, Stiles WH (1980) The active and passive microwave response to snow parameters: 2,
water equivalent of dry snow. J. Geophys. Res. 85:1045–1049
Ulaby FT, Stiles WH (1981) The active and passive microwave response to snow parameters: 2,
water equivalent of dry snow. J. Geophys. Res. 85(C2):1045–1049
Ulaby FT, Moore RK, Fung AK (1982) Microwave remote sensing: active and passive, 2, radar
remote sensing and surface scattering and emission theory. Addison-Wesley, Reading, MA
Ulaby FT, Stiles WH, Abdelrazik M (1984) Snowcover influence on backscattering from terrain.
IEEE Trans. Geosci. Remote Sens. GE-22:126–132
Ulaby FT, Moore RJ, Fung AK (1986) Microwave remote sensing, active and passive. vol III.
Artech House, Durham, MA.
Van Zyl JJ, Zebker HA, Elachi C (1987) Image radar polarization signatures: theory and observa-
tions. Radio Sci. 22(4):529–543
Van Zyl JJ, Chapman BD, Dubois P, Shi J (1993) The effect of topography on SAR calibration.
IEEE Trans. Geosci. Remote Sens. 31(5):1036–1043
Wu TD, Chen KS, Shi J, Fung AK (2001) A transition model for the reflection coefficient in surface
scattering. IEEE Trans. Geosci. Remote Sens. 39(9):2040–2050
Chapter 4
Multi-angular Thermal Infrared Observations
of Terrestrial Vegetation

Massimo Menenti, Li Jia, and Zhao-Liang Li

Abstract This chapter reviews the experimental evidence on the anisotropy of


emittance by the soilvegetation system and describes the interpretation of this sig-
nal in terms of the thermal heterogeneity and geometry of the canopy space. Ob-
servations of the dependence of exitance on view angle by means of ground-based
goniometers, airborne and space-borne imaging radiometers are reviewed first to
conclude that under most conditions a two-components, i.e., soil and foliage, model
of observed Top Of Canopy (TOC) brightness temperature is adequate to inter-
pret observations. Particularly, airborne observations by means of the Airborne
Multi-angle TIR/VNIR Imaging System (AMTIS) and space-borne observations
by means of the Along Track Scanning Radiometers (ATSR-s) are described and
examples presented. Modeling approaches to describe radiative transfer in the soil–
vegetation–atmosphere system, with emphasis on the thermal infrared region, are
reviewed. Given the dependence of observed TOC brightness temperature on leaf-
level radiation and heat balance, energy and water transfer in the soil–vegetation–
atmosphere system must be included to construct a realistic model of exitance by
soil–vegetation systems. A detailed modeling approach of radiation, heat and wa-
ter transfer is first described then applied to generate realistic, multi-angular image
data of terrestrial landscapes. Finally, a generic algorithm to retrieve soil and foliage
component temperatures from Top Of Atmosphere (TOA) radiometric data is de-
scribed. Column water vapor and aerosols optical depth are estimated first, to obtain
TOC radiometric data from the TOA multi-angular and multi-spectral observations.

M. Menenti
TRIO/LSIIT, University Louis Pasteur (ULP), Strasbourg, France
Istituto per i Sistemi Agricoli e Forestali del Mediterraneo (ISAFOM), Naples, Italy
m.menenti@isafom.cnr.it
L. Jia
Alterra, Wageningen University and Research Centre, The Netherlands
Z.-L. Li
TRIO/LSIIT, University Louis Pasteur (ULP), Strasbourg, France
Institute of Geographic Sciences and Natural Resources Research, Beijing, China

S. Liang (ed.), Advances in Land Remote Sensing, 51–93. 51


c Springer Science + Business Media B.V., 2008
52 M. Menenti et al.

Then vegetation fractional cover and soil and foliage component temperatures are
determined by inverting a simple two-components mixture model. The accuracy of
all elements of the algorithm is evaluated by using a combination of actual mea-
surements and synthetic radiometric data. Although applicable to multi-angular ra-
diometric data irrespective of spatial resolution, the approach presented would be
particularly relevant if space-borne observations with a footprint of 100 × 100 m or
better would be available. Observing systems, presently at the design stage, with
this capability are briefly described.

4.1 Introduction

4.1.1 Vegetation–Atmosphere Exchanges of Energy and Water

The exchange of energy between the land surface and the atmosphere and within
terrestrial vegetation canopies is a significant determinant of processes in the at-
mospheric boundary layer and in terrestrial ecosystems. In these processes it is
crucial to determine accurately the partitioning of available energy into sensible
heat flux density H (heating or cooling of the surface) and latent heat flux density
λ E (evaporation from surface) over a wide range of spatial and temporal scales.
Observation and modeling of turbulent heat fluxes at the land surface has been a
very active area of research at least since the work of Bowen (1926) on the relative
magnitude of heat transfer over dry and wet surfaces (Monteith, 1965; Feddes, 1971;
Verma et al., 1976; Hall et al., 1979; Price, 1982; De Bruin and Jacobs, 1989;
Beljaars and Holtslag, 1991; Lhomme et al., 1994; Sellers et al., 1995, 1996). Most
conventional techniques that employ point measurements to estimate the compo-
nents of energy balance are representative only of local scales and cannot be ex-
tended to large areas because of the heterogeneity of the land surface, of the dynamic
nature and of the spatial distribution of heat transfer. Remote sensing is one of the
few techniques to provide representative measurements, e.g., surface temperature
and albedo, at regional and global scales.
Methods using remote sensing techniques to estimate heat exchange at the land-
atmosphere interface fall into two main categories: (1) use surface radiometric tem-
perature Trad to calculate H then obtain λ E as the residual of the energy balance
equation (Blad and Rosenberg, 1976; Seguin et al., 1989; Hatfield et al., 1984);
(2) use Trad to estimate the Crop Water Stress Index or the evaporative fraction (the
ratio of evapotranspiration to the available energy) (Jackson et al., 1981; Menenti
and Choudhury, 1993; Moran et al., 1994). The former category can be further
subdivided into single-source, dual-source and multisource models corresponding
with a single-, dual- or multi-layer schematization of the surface respectively. Suc-
cessful estimations of heat fluxes have been achieved over horizontal homogeneous
surfaces, such as a surface fully covered by vegetation, open water and bare soil
(Jackson, 1985; Huband and Monteith, 1986; Choudhury et al., 1986). Large devi-
ations from these conditions occur at partial canopies which are geometrically and
4 Multi-angular Thermal Infrared Observations of Terrestrial Vegetation 53

thermally heterogeneous. Recent years have seen increasing evidence of specific


difficulties inherent to the heterogeneous nature of terrestrial vegetation. For in-
stance, in many semi-arid environments where the surfaces are partially covered by
vegetation, both the soil surface and cooler foliage determine the heat exchanges.
This leads to the challenge of relating the separate contributions from these elements
to the turbulent transport of heat across the land-atmosphere interface.
Exchange of water and CO2 between land surface and atmosphere determines
to a significant extent the dynamics of the Convective Boundary Layer (CBL)
(Mahfouf et al., 1987; Segal et al., 1988; Hatfield, 1991; Xue and Shukla, 1993).
Over homogeneous land surfaces the controlling factor is the partition of net ra-
diation into sensible, latent and soil heat flux. The partition of net radiation is
determined by the presence and functioning of vegetation and by available soil mois-
ture. Heterogeneous land surfaces compound the complexity of these processes,
since the spatial pattern of land surface properties determines CBL motion at small
length scales (Hatfield, 1991; Wang and Mitsuta, 1992). These studies brought the
attention of a wide scientific community to the need for significant improvements in
models of such land surface processes and of the interactions of land surfaces with
the atmosphere (e.g., Avissar and Pielke, 1989; Pielke et al., 1991). High resolution
atmospheric models may be used to interpret observations in complex landscapes as
for example done by Peng and Cheng (1993) and Yan (1999), in the framework of
the Hei He basin International Field Experiment (HEIFE).
Observations of the anisotropic emittance of land cover provide unique ac-
cess to the thermal heterogeneity of soil–vegetation systems. All land surfaces
are anisotropic and in the thermal-infrared domain, the directional variation of
emitted fluxes (described by the so-called brightness temperature) is mainly de-
termined by the distribution of temperature and emissivity between the elements
of the canopy, and by the structure of the vegetation (see Balick et al., 1987;
Kimes and Kirchner, 1983). Similar to the solar domain, the distribution of shad-
owed and illuminated parts, as well as the amount of soil and vegetation observable
from a particular direction, are the main drivers of the anisotropy models that have
been developed to describe the directional variations in thermal infrared spectral
domain (e.g., Norman and Chen, 1990; Ottermann, 1990; Smith and Goltz, 1995;
François et al., 1997).
Determination of the soil and foliage component temperature requires inverse
modeling of observed emittance, which is a challenge (Jia et al., 2003; Jia, 2004;
François et al., 1997; Li et al., 2001; François, 2002; Liu, 2002), given the strength
of the signals and the accuracy which can be achieved. On the other hand it provides
the only opportunity to overcome a major shortcoming of current parameterizations
of heat fluxes at the land–atmosphere interface (Jia, 2004).
Observations of foliage and soil temperature provide also a reliable indicator
of crop water stress, either directly as soil-foliage temperature difference or indi-
rectly through modeling of soil evaporation and plant transpiration (François, 2002;
Luquet et al., 2003).
54 M. Menenti et al.

4.1.2 Photosynthesis: Light Use Efficiency of Sun-lit


and Shadowed Leaves

Observations of leaf and soil temperature are also quite relevant for understanding
and modeling carbon exchange between terrestrial vegetation and the atmosphere.
The rate of photosynthesis depends on many factors, including carbon dioxide
concentration in the atmosphere, leaf temperature, or mineral deficiencies (in par-
ticular nitrogen) in the soil. The nitrogen content of leaves is strongly related to their
chlorophyll content (Field and Mooney, 1986).
Foliage temperature. Autotrophic respiration is the process by which some of the
chemical energy stored by photosynthesis is used by the plants themselves to grow
and develop. This process is critical to the carbon cycle because it results in the rapid
release of a large fraction of the carbon initially stored through photosynthesis back
to the atmosphere. Autotrophic respiration depends on foliage temperature, growth
rates and total biomass, as well as on the biochemical composition of the products
formed in the plants (Medlyn et al., 2002).
Soil temperature. Heterotrophic respiration is the process by which some of the
carbon stored in organic soil components is released. The soil carbon reservoir can
be very large compared to the above-ground biomass. Understanding the fluxes of
carbon to (senescence, mortality) and from (respiration or mineralization) this soil
reservoir becomes a major issue when closing the carbon cycle at the local scale.
Heterotrophic respiration is very dependent on soil temperature, and the availability
of water and nutrients, particularly nitrogen. Apart from nitrogen fertilization or
deposition, symbiotic fixation of atmospheric nitrogen, and leakage or volatilization,
the nitrogen cycle is intimately linked to the carbon cycle within the soil via the
biotic activity. Evaluation of heterotrophic respiration is a major challenge in the
description and modeling of NEP (Valentini et al., 2000; Matteucci et al., 2000).
Due to the strong temperature dependence (Fig. 4.1) of both leaf photosynthe-
sis and soil respiration, observations of foliage and soil temperature are useful to

7
Soil respiration [umol CO2 m−2 s−1 ]

1.6 6

1.4 An
An(Normalisedto 1 at 25ºC)

5 Col., Q10=2.21
1.2
4
1
Betula pendula
0.8 3
Quercus petraea
0.6
2
Acer pseudoplatanus
0.4
1 Wal., Q10=2.16
0.2
Gossypium hirsutum Mdm., Q10=2.09
0 0
0 5 10 15 20 25 30 35 40 −5 0 5 10 15 20
Leaf temperature ( ⬚C) Soil temperature at 5 cm depth [⬚C]

Fig. 4.1 Carbon fluxes and component soil and foliage temperatures: (left) relative net
photosynthesis, An (normalized to the observed value at 25 ◦C), vs. foliage temperature
(Medlyn et al., 2002); (right) soil respiration vs. soil temperature. (After Matteucci et al., 2000.)
4 Multi-angular Thermal Infrared Observations of Terrestrial Vegetation 55

understand and model these processes. As regards field experiments, the use of
very high resolution TIR images is a widely used solution to obtain direct mea-
surements of foliage and soil temperature (see Section 4.5). As regards observa-
tions from space, the only feasible solution is multi-angular observations of Top Of
Canopy (TOC) brightness temperature and subsequent inverse modeling to deter-
mine foliage and soil temperature from observed anisotropy of emittance. A mini-
mum of two spectral bands is required to establish the brightness temperature of the
surface, taking atmospheric influences into account.

4.2 Nature of Anisotropic Emittance of Land Surface

4.2.1 Anisotropic Emissivity of Leaves and Soils

The observed anisotropy of land emittance is (see Section 4.3) caused by a combi-
nation of two different processes: the inherent anisotropic emissivity of terrestrial
materials and the thermal heterogeneity of complex, three-dimensional (3D) land
targets. The latter is due to the interaction of radiative and convective energy trans-
fer with the 3D structure of land targets These processes are briefly reviewed in this
section.
The accuracy of radiometric measurements of land surface temperature depends
significantly on accurate knowledge of land emissivity. A 1% uncertainty on surface
emissivity can cause about 0.6◦ error on land surface in temperature (Becker, 1987).
An angular variation of emissivity has been observed by a number of scientists either
in the field or in the laboratory (Becker et al., 1985; Labed and Stoll, 1991; Xu
et al., 2000). For example, when sea surface wind produces wavelets, an angular
variation of surface emissivity is observed (Masuda et al., 1988; François and Ottlè,
1994), and an angular variation of the “effective” surface emissivity has also been
observed with satellite data (Nerry et al., 1998; Petitcolin et al., 2002).
Many efforts have been devoted to measure the directional emissivity of soil,
leaves and other natural surfaces. Commonly, emissivity of natural surfaces de-
creases with increased zenith angle of observation. Becker et al. (1985) measured
in the thermal infrared band the bi-directional reflectivity (BDR) of different types
of bare soils including quartz sands, agricultural soil, and well-calibrated powders,
a large variability of the BDR for different samples with no apparent systematic be-
havior apart from the backscattering peak was observed and the roughness has large
impact on the BDR which means that the same material with different grain sizes
may exhibit different angular distributions.
As an example, Fig. 4.2 gives the measured angular variation of emissivity for
water, sand, clay, slime, and gravel (Sobrino and Cuenca, 1999).
56 M. Menenti et al.

Fig. 4.2 Angular variation of surface emissivity for several natural surfaces. (After Sobrino and
Cuenca, 1999, Fig. 3.4.)

Fig. 4.3 Anisotropy change resulting from changing the single-scattering reflectance from 0.3 to
0.2 in the Hapke model. The plot is in the 30◦ azimuth plane with the angle of incidence at 32◦ .
Positive zenith values represent backscattering. (After Snyder et al., 1997.)

To describe the anisotropic emissivity of sands and soils, Snyder et al. (1997)
defined a so-called anisotropy factor, a:
πf
a= (4.1)
1−ε
where f is the bi-directional reflectance distribution function (BRDF) of the surface,
with the same meaning of the bidirectional reflectivity, ε the emissivity.
To illustrate the magnitude of anisotropy, a simulation with Hapke BRDF model
was done by varying single-scattering reflectance and zenith angle (Fig. 4.3).
4 Multi-angular Thermal Infrared Observations of Terrestrial Vegetation 57

This figure shows that the difference between the two curves for different single-
scattering reflectance is relatively small with respect to the overall variation. The
simulation by Snyder et al. (1997) demonstrates that for typical materials, over a re-
alistic range of azimuth and incident and reflected zenith angles, the upper bounds of
the impact of this difference is 4.8% (RMS) and 9.5% (maximum error) respectively.
Efforts to evaluate the anisotropy of foliage emissivity have been limited, because
it is very close to unity and has small angular variation. Therefore, for most stud-
ies, leaves are assumed to be Lambertian. The ray-tracing method has been applied
to estimate the directional reflectance of canopy in VIS/NIR. An extension of this
method to the TIR region may be useful to simulate the angular variation of canopy
thermal infrared reflectance.

4.2.2 The 3D Structure of Vegetation Canopies

The architecture of most vegetation canopies leads to a complex three-dimensional


distribution of absorbed radiant energy and, therefore, of the local balance of energy
within the canopy space (Fig. 4.4). On the one hand, within the canopy space the
surface temperature of foliage and soil varies significantly. On the other hand, the

20
11 April, 13:00, s50
Frequency (%)

15

10

0
10 15 20 25 30 35
Brightness temperature (⬚C)
20
11 April, 10:30, s50
Frequency (%)

15

10

0
10 15 20 25 30 35
Brightness temperature (⬚C)

Fig. 4.4 Thermal anisotropy of soil–vegetation system resulting from canopy geometry and vari-
ability of absorbed solar radiation by different components: wheat leaves (blue) are significantly
colder than soil (red). (After Jia, 2004.)
58 M. Menenti et al.

nadir view off - nadir view

θv

cool vegetation

warm soil

Fig. 4.5 Illustration of observed TIR radiance as a function of canopy geometry, foliage and soil
component temperatures and the zenith view angle θv . The light gray bar indicates the fraction
of foliage in the instantaneous field of view (IFOV) of the sensor, the dark gray bar indicates the
fraction of soil in the IFOV. (Jia 2004.)

vertical distribution of foliage temperature is also variable with the solar elevation,
the density of leaves and the angle distribution of leaves. The thermal heterogeneity
within a vegetation canopy leads to the fact that Trad measured by thermal infrared
(TIR) sensors is a function of canopy geometry, vertical distribution of foliage tem-
perature Tf , soil temperature Ts , sensor view angle (θv , ϕv ) and incoming radiation
(Kimes, 1980; François et al., 1997) (Fig. 4.5).

4.2.3 Radiation and Convection in the Canopy Space

For incomplete canopies, a frequent case in nature, the interaction between the
canopy and the atmosphere becomes complex due to the canopy geometry in terms
of the size and spacing between plants, the leaf density and the leaf angle distri-
bution. Figure 4.6 illustrates how the elements of a sparse canopy are interacting
with their environment. The complex canopy geometry determines the distribution
of absorbed solar radiation in the canopy, thereafter inducing spatial variability of
sources and sinks of heat and water vapor in the canopy space. A large spacing
between plants or lower leaf density, for instance, makes the exposed soil to play
an important role in the land–atmosphere interaction. Canopy geometry has also
influence on the airflow in the canopy space and the boundary layer resistance of
leaves and soil, thus changing the source/sink strength. The interaction between
thermodynamic and dynamic processes will lead to thermal heterogeneity, which
will in turn give rise to the anisotropy in the exitance of canopy. (See, e.g., Albertson
et al., 2001; Wilson et al., 2002)
4 Multi-angular Thermal Infrared Observations of Terrestrial Vegetation 59

Horizontal
Divergence

Canopy Height NET SENSIBLE WATER


RADIATION Horizontal
HEAT VAPOR
Divergence

Water Vapor Water Vapor


Storage terms in volume:
∑ Temp. change of crop Sensible Heat
Sensible Heat dz ∑ Temp. change of moist air
dy ∑ Absolute humidity change

dx Soil surface
SOIL HEAT

Horizontal
Divergence

Fig. 4.6 Schematic illustration of the 3D structure of a vegetation canopy and of the interactions
between canopy elements and the canopy environment.

4.2.4 Thermal Heterogeneity of Vegetation Canopies

The radiative and thermal state of elements in a soil-vegetation canopy system is


strongly dependent on its geometric structure (radiative transfer) and on its environ-
mental situation (convection and conduction processes). The interactions between
the land surface and the atmosphere consist of the interactions between foliage and
soil surface, foliage and air in the canopy space, soil surface and air in the canopy
space, and between the soil surface and deeper soil layers, i.e., root zone. A realistic
model requires describing the processes involved at each spatial point. However, it
may not be possible or necessary to do so. Adequate simplification is necessary to
redefine the canopy, which should retain the dominant aspects of 3D radiative, heat
and mass transfer (see Section 4.3).
Commonly, the soil temperature is much higher than foliage temperature because
of the different thermal properties between soil and foliage. Thus, to simulate energy
exchange between canopy and natural environment, vegetative canopy is commonly
represented by two-source model: the sensible heat flux between soil and air, and
the one between foliage and air. As the penetration of downwelling solar short-
wave radiation and long-wave sky radiation, the foliage temperature changes with
the depth of canopy. The temperature profile of vegetative canopy with the height
of canopy has been observed by several scientists, e.g., by Paw et al. (1989) who
measured the variation of leaf temperature with height within the canopy space by
means of a Teletemp Infrared Thermometers (Fig. 4.7).
The temperature profile of canopy layers result from the energy balance for each
layer. As seen from Fig. 4.7 both foliage at top layers and at bottom layer have
60 M. Menenti et al.

200

150
HEIGHT(cm)

110

50

150
HEIGHT(cm)

200

50

0
19 20 21 22 23 24 25 26 27 28 29 30 31
TEMPERATURE(⬚C)

Fig. 4.7 Leaf temperature profile in the canopy, at 1,200–1,300 h, for all days. The top graph
shows each individual leaf temperature measurement. The bottom figure shows a 10-point moving
average of leaf temperature. (After Paw et al., 1989, Fig. 7.)

higher temperature. The higher temperature of top layer is due to the relatively large
incident short-wave radiation which may decrease with depth within the canopy.
The bottom layer of vegetative canopy will receive more long-wave radiation than
the middle canopy layer from the soil which temperature is usually higher than
foliage, with convective cooling being less efficient than at the top of the canopy.
Lei (2004) made a detailed temperature measurements of sunlit and shaded
leaves of Blackbrush and sunlit soil from March through December 2001 at the
Clark Mountain (roughly 35.7◦ N and 115.5◦ W; elevation 1.475 m). Figure 4.8
shows the temperature difference between sunlit and shaded foliage and the
temperature difference between sunlit soil and foliage.

4.3 Observed Magnitude of Anisotropy

4.3.1 Ground Observations of Tb (θ ) at TOC

The anisotropy in canopy exitance implies that brightness temperature (Tb0 ) at the
Top Of the Canopy (TOC) changes with view zenith angle θv and azimuth angle ϕv
as shown by field measurements over a range of canopies, especially over sparse
4 Multi-angular Thermal Infrared Observations of Terrestrial Vegetation 61

4 20

Soil-leaf temperature difference (⬚C)


Spring
Leaf temperature difference (⬚C)
Spring
Summer Summer
Winter Winter
3 15

2 10

1 5

0 0
6 9 12 15 18 6 9 12 15 18
Hour of day Hour of day

Fig. 4.8 Temperature difference between: sunlit and shaded leaves (left); sunlit soil surface and
sunlit leaves (right); Clark Mountain of southeastern California. (After Lei, 2004, Figs. 4 and 5.)

Fig. 4.9 Observation of TOC brightness temperature Tb0 at different view angles. The circles rep-
resent the footprints of IFOV at the Top Of Canopy (TOC) and at the bottom of the canopy for
different view angles. The components in the volume between TOC and the bottom are observed
by a radiometer located above the canopy. (Jia, 2004.)

canopies. The dependence of observed Top Of Canopy (TOC) brightness temper-


ature Tb0 (θv , ϕv ) on the view angle is best documented by ground measurements
with a goniometer-mounted radiometer (Fig. 4.9, After Jia, 2004). Tb0 (θv , ϕv ) is the
temperature measured by a radiometer at zenith angle θv and azimuth angle ϕv and
is simply derived from the radiance Rλ measured by a radiometer by inverting the
Planck’s function. Therefore, the observation of Tb0 (θv , ϕv ) and the observation of
TIR radiance are equivalent. In this chapter, for simplification the symbol “R” de-
notes only TIR radiance and the subscript TIR will be neglected. TIR radiance of
62 M. Menenti et al.

a target is usually referred to as “exitance”, i.e., the sum of emitted and reflected
TIR radiance by the target concerned. Nielsen et al. (1984) have shown that it is
common to have large (up to 20K or more) differences between sun-lit soil and
shadowed leaf surfaces, particularly when the top soil is dry. [127] found differ-
ences between bare soil and air temperature as large as 27◦ C. For a soybean canopy
with 35% ground cover, the soil temperature exceeded the canopy temperature by
11◦ C and was 15◦ C higher than air temperature (Kimes 1980). Usually, Tb0 (θv , ϕv )
is measured by a radiometer in a specific spectral range (centered at some wave-
length) and in a particular direction (θv , ϕv ), within an instantaneous field-of-view
(IFOV) ΩIFOV . The portions of canopy components with different surface temper-
atures in the IFOV will change with the view angle (Fig. 4.9). As a consequence,
strong anisotropy in exitance, i.e., a significant variation in Tb0 (θv , ϕv ) with the di-
rection of observation, can be observed over thermally heterogeneous systems like
sparse canopies. For instance, Qualls and Yates (2001) observed in a cotton field that
the difference in Tb0 (θv , ϕv ) between the 0◦ (mixture of vegetation and soil) and the
80◦ (vegetation only) zenith view angles was 16.2◦ C at noon, while the difference
was only 0.9◦ C in the early morning. Lagouarde et al. (1995) observed a difference
of up to 3.5K for a corn canopy and 1.5K for grass (20 cm high) with a view zenith
angle between 0◦ and 60◦ around solar noon.
A goniometer (Fig. 4.10a) was designed specifically for canopy directional
Tb0 (θv , ϕv ) measurements (Zhang et al., 2002; Jia, 2004; Li et al., 2004). Two arms
are connected perpendicularly to each other with the longer one being fixed onto a

(a) (b) N

0
330 30

300 60

90
270

(c) 0 60
330 30
40
Target 300 60
20

270 90

240 120

210 150
180

Fig. 4.10 Goniometric system used to measure Tb0 (θv , ϕv ): (a) overview of the goniometer system
and of the footprint of the TIR radiometer; (b) zenithal and azimuth sampling scheme; (c) detail of
the azimuth sampling scheme. (Liu, 2002; Jia, 2004.)
4 Multi-angular Thermal Infrared Observations of Terrestrial Vegetation 63

circular track, which is set up on the ground, and the shorter one is kept horizontal
on top of which radiometers can be mounted. The longer arm can move along
the track to change the azimuth position. At each specific azimuth position ϕv the
longer arm sways over a range of zenith angles θv (maximum 60◦ ). Such move-
ments are designed and performed to measure the TIR radiance of the same target
on the ground (Fig. 4.10b) within a desired range of azimuth and zenith angles.
The diameter of the footprint, however, increases with increasing θv because of
slant viewing. A 10◦ interval is normally used for the zenith angle change, and 15◦
interval for azimuth angle change (Fig. 4.10c).
Two radiometers were used (Jia 2004) to measure Tb0 (θv , ϕv ). One radiometer
was set up on the top of the short horizontal arm to measure the radiance of the
canopy at each azimuth and zenith angle. Distance to the target was the same at
any position of the arm so that the radiometer footprint included the same target at
all positions as shown in Fig. 4.10b. The other radiometer was mounted on a mast
observing continuously the canopy at nadir. The second radiometer provided the
continuous measurements needed to correct for the temporal change in the mea-
surements of Tb0 (θv , ϕv ) during a complete goniometer scan. The latter usually took
about 20 min during which the surface temperature may change significantly due
to the variation of the solar radiation and windspeed. Due to technical problems,
different radiometers were used during the experiment (Table 4.1).
A thermal camera (AGEMA THV 900 LW), mounted on top of another goniome-
ter, was used to obtain images of surface temperature of the wheat crop Tb0 (θv , ϕv )
for prescribed azimuth and zenith directions. The AGEMA thermal camera has a
scanning HgCdTe detector and a Stirling cooler with the single channel covering
the spectral range between 8–12 µm, the frame rate is 15 Hz for 136 × 272 pixels,
and the nominal sensitivity is 80 mK at 30◦ C. The camera was equipped with a lens
having a FOV of 5◦ × 10◦ .
Figure 4.11 shows the change in Tb0 from nadir to off-nadir view zenith angles
at each view azimuth angle at different hours during the two selected days: (a) 11
April and (b) 21 April. Only the measurements across two perpendicular planes are
shown: one in the N–S direction along the canopy rows and one in the E–W direction
across the row. At each azimuth direction, measurements were made twice, e.g., at
both 0◦ and 90◦ azimuth the first observation started from θv = +60◦ through nadir
to θv = −60◦ (denoted as “go” in Fig. 4.11) and the second observation went back

Table 4.1 The characteristics of the radiometers used to measure Tb0 (θv , ϕv ) during the field
campaign of QRSLSP (Jia 2004)
Instrument Wavelength (µm) FOV (◦ )
Radiometer 1 8–11, 10.4–14 4.7
Radiometer 2 8–11, 10.4–14 8.6
Radiometer 3 8–11, 10.4–14 8.6
Radiometer (single channel) 8–14 15
Raytek radiometer 8–14 7
64 M. Menenti et al.

(a)
4 4 4
11 April 0-180 go 11 April 0-180 go 11 April 0-180 go
Tb0 (nadir) - Tb0 (off-nadir) (⬚C)

Tb0 (nadir) - Tb0 (offnadir) (⬚C )


11 : 00 0-180 back

Tb0 (nadir) Tb0 (offnadir) (⬚C)


10:30 0-180 back 0-180 back 11:30
90-270 go
90-270 go 90-270 go
3 3 3 90-270 back
90-270 back 90-270 back

2 2 2

1 1 1

0 0 0

−60 −40 −20 0 20 40 60 −60 −40 −20 0 20 40 60 −60 −40 −20 0 20 40 60


Zenith view angle (⬚) Zenith view angle (⬚) Zenith view angle (⬚)

4 4 4
11 April 0-180 go 11 April 0-180 go 11 April 0-180 go
Tb0 (nadir) - Tb0 (off-nadir) (⬚C)

Tb0 (nadir) - Tb0 (off-nadir) (⬚C)


0-180 back
Tb0 (nadir) Tb0 (offnadir) (⬚C)

12 : 00 12:30 0-180 back 13:00 0-180 back


90-270 go 90-270 go 90-270 go
3 90-270 back 3 90-270 back 3 90-270 back

2 2 2

1 1 1

0 0 0

−60 −40 −20 0 20 40 60 −60 −40 −20 0 20 40 60 −60 −40 −20 0 20 40 60

Zenith view angle (⬚) Zenith view angle (⬚) Zenith view angle (⬚)

4 4 4
11 April 0-180 go 11 April 0-180 go
Tb0 (nadir) - Tb0 (off-nadir) (⬚C)

Tb0 (nadir) - Tb0 (off-nadir) (⬚C)

11 April 0-180 go
Tb0 (nadir) - Tb0 (off-nadir) (⬚C)
13:30 0-180 back 14:00 0-180 back
14:30 0-180 back
90-270 go 90-270 go
3 3 3 90-270 go
90-270 back 90-270 back 90-270 back

2 2 2

1 1 1

0 0 0

−60 −40 −20 0 20 40 60 −60 −40 −20 0 20 40 60 −60 −40 −20 0 20 40 60

Zenith view angle (⬚) Zenith view angle (⬚) Zenith view angle (⬚)

4 4
11 April 0-180 go 11 April 0-180 go
Tb0 (nadir) - Tb0 (off-nadir) (⬚C)

15:00 0-180 back 16:00 0-180 back


T (nadir) - T (off-nadir) (⬚C)

3 90-270 go 3 90-270 go
90-270 back 90-270 back

2
2
b0

1
1
0
b0

0
−1

−60 −40 −20 0 20 40 60 −60 −40 −20 0 20 40 60

Zenith view angle (⬚) Zenith view angle (⬚)

Fig. 4.11 Measurements of the directional variation of TOC brightness temperature difference Tb0
(nadir) – Tb0 (off – nadir) with zenith view angle: (a) on 11 April and (b) on 21 April at different
times of the day at the QRSLSP site (Liu et al. 2002). Tb0 (nadir) is Tb0 at θv = 0◦ , Tb0 (off – nadir)
is Tb0 measured at θv = 0◦ . The positive zenith view angle correspond to the azimuth 0◦ and 90◦
planes, the negative zenith view angle correspond to the 180◦ and 270◦ azimuth planes. (Jia 2004.)

from view zenith θv = −60◦ through nadir to θv = +60◦ (denoted as “back” in


Fig. 4.9). All the measurements shown in Fig. 4.11 have been corrected for temporal
change taking into account the measurements provided by the second nadir looking
radiometer.
Changes in Tb0 with θv are significant and show different trends at different hours
during a day. Around noon, the observed near-nadir Tb0 is always higher than that
at off-nadir positions. On the contrary, in the early morning and the late afternoon,
4 Multi-angular Thermal Infrared Observations of Terrestrial Vegetation 65

near-nadir Tb0 tends to be lower than the off-nadir values. Maximum difference
between near-nadir Tb0 and those at off-nadir is about 2.8◦ C on 11 April, while
4.4◦ C is observed on 21 April, both around noon time.
The row structure of the winter wheat also plays an important role in determining
the angular change of Tb0 . Such structure effects are evident when comparing the
shape of the curves in the along-row direction (the plane from 0◦ to 180◦ ) and the
curves in the across-row direction (the plane from 90◦ to 270◦ ) in Fig. 4.11. The lat-
ter shows a steeper slope, particularly in the position opposite to the sun, e.g., at
270◦ plane (negative zenith view angle) at 10:30, 11:00 and 11:30 on 11 April 2001
when the sun was located between 90◦ and 180◦ planes, and asymmetric than the
former.
For the proper interpretation of multi-angular measurements the geometry of ob-
servations needs to be taken into account, because of significant differences in foot-
print size, position and shape at different view angles. The change in the diameter of
the radiometer footprint when the radiometer observes the target at different zenith
view angles implies significant differences in the canopy elements captured by ob-
servations.

4.3.2 Airborne Observations

To our knowledge the only airborne imaging system having the capability to perform
multi-angular thermal infrared observations is the Airborne Multi-angle TIR/VNIR
Imaging System (AMTIS) developed by Wang et al. (2002). This is a proto-
type three-channel multi-angle imaging system. It provides high-resolution data
in visible/near infrared and thermal infrared spectral bands for use in deriving bi-
directional reflectance factors. The precision of the viewing angle is determined by
the pointing precision of the AMTIS and the knowledge of the attitude of the air-
plane (pitch, roll and heading angles).
The AMTIS consists of one visible CCD camera, one near-infrared CCD cam-
era, one thermal video system (TVS), camera bench, swing assembly, motor driver
and controller, exposing synchronization signal generator, data grabber, real-time
display and recorder. Two CCD cameras and the TVS are mounted on the camera
bench. A stepper motor rotates the camera bench.
The instantaneous field of view (IFOV) of each pixel has been fixed at 0.3 mrad
for the VNIR bands, 1 mrad for the TIR band. The total field of view of each camera
is about 20◦ . The altitude range is roughly from 500 m to 10,000 m which thus trans-
lates into a spatial resolution ranging from 0.15 m to 3 m for VNIR, 0.5 m to 10 m for
TIR. Normally, AMTIS is operated at an altitude of 3,000 m with a ground speed
of 250 km/h, and swing along-track through discrete nine programmable viewing
angles within a range of ±45◦ . A step motor through a gearbox drives the cam-
eras bench. The accuracy of the pitch angle is 0.3◦ . The maximum swing speed is
200 steps/s, or 60◦ /s. Normally, the swing angular velocity is about 10◦ per second
when AMTIS takes photograph, then swings back at the angular velocity of about
30◦ per second.
66 M. Menenti et al.

4.3.3 Space Observations of Tb (θ ) at TOA and TOC

The measurements of thermal emission at top of atmosphere include the contri-


butions of atmosphere (through atmospheric upwelling radiation and scattering of
surface thermal emission) and those of the surface.
To perform nearly simultaneous observations of Tb0 (θv , ϕv ) from space at mul-
tiple view angles two technical solutions may be used: (a) multiple optics or (b)
accurate along-track pointing. The Along Track Scanning Radiometer 1 (ATSR-1)
flown on ERS-1 was based on solution (a). ATSR-1 was followed by ATSR-2 in
1995 and the Advanced ATSR (AATSR) on ENVISAT in 2001. The characteristics
of these systems (Mutlow et al., 1994) are summarized below.
All ATSR sensors acquire dual-view angle data (approximately 0◦ and 53◦ at
surface) in four channels for ATSR-1 and seven channels for ATSR-2 and AATSR.
The nominal noise equivalent temperature difference (NEδT) of ATSR-2 for IRT
channels is 0.04K. The use of the along track scanning technique (Fig. 4.13) makes it
possible to observe the same point on the earth’s surface at two view angles through
two different atmospheric path within a short period of time. The first view is at a
view angle of 55◦ (approximately 53◦ at the earth surface) along the direction of the
orbit track when the satellite is flying toward the target point, which is referred to as
forward observation in this Chapter. Within 21/2 min the nadir (0◦ ) view observation
is made over the same point, which will be referred to as nadir observation later on.
Figure 4.12 shows the viewing geometry of ATSR-2 observations. The swath width
of ATSR-2 is 500 km, which provides 555 pixels across the nadir (0◦ zenith angle)
swath and 371 pixels across the forward (55◦ zenith angle) swath. The nominal pixel
size of ATSR-2 is 1 × 1 km at the centre of the nadir swath and 1.5 × 2 km at the
centre of the forward swath.
The Along-Track Scanning Radiometer (ATSR) instruments are imaging ra-
diometers which are currently the only observing system able to provide from

6
Frequency (%)

0
−6 −4 −2 0 2 4 6 8 10 12
Tb(0) - Tb(45)

Fig. 4.12 Difference of surface brightness temperature Tb between nadir view and forward view
(45◦ ); observations by AMTIS in Shunyi of Beijing on 11 April 2001.
4 Multi-angular Thermal Infrared Observations of Terrestrial Vegetation 67

ATSR
Instrument

Sub-satellite Track

55⬚

Nadir view swath


(555 nadir pixels
1 km resolution)
Flight direction
Forward view swath

(371 along track pixels


1.5 km x 2 km resolution)

Fig. 4.13 Illustration of Along Track Scanning Radiometer (ATSR) observation. (After Mutlow
et al., 1999.)

space quasi-simultaneous bi-angular radiance measurements of the earth’s surface


in the TIR and SWIR spectrum regions (in addition to VIS/NIR channels). ATSR-
1 onboard the first European Remote Sensing satellite (ERS-1) was launched in
July 1991 and operated until June 1996. ATSR-1 had four channels – one short
wave infrared (SWIR) channel located at 1.6 µm and three TIR channels centred at
3.7 µm, 11 µm and 12 µm. ATSR-1 was designed particularly for providing data over
the sea. ATSR-2 onboard the ERS-2 satellite was launched in April 1995 and is cur-
rently providing data both over land and over sea. In addition to one SWIR channel
and three TIR channels as on ATSR-1, ATSR-2 and AATSR have three narrow-
band visible-near infrared channels in the blue, green and red spectrum located
at 0.55 µm, 0.67 µm and 0.87 µm respectively for vegetation monitoring. Table 4.2
gives information on channel spectral characteristics. (see the World Wide Web site
at www.atsr.rl.ac.uk/software.html for details).
The ATSR-2 data analysed below pertain to images acquired on June 6th 1999
over Spain at 10:30 a.m. local solar time. The major part of this image is over a culti-
vated zone, the rest being bare soils. This image is considered cloud-free, inasmuch
as ATSR-2 data do not indicate the presence of clouds.
The four images have been corrected for atmospheric effects: channel 1 (12 µm)
nadir and forward views; channel 2 (11 µm) nadir and forward views. Since ATSR
has a limited dynamical range of brightness temperatures, and radiometric saturation
is supposed to occur at 312◦ K, all pixels having Tb ≥ 312◦ K have been masked. At-
mospheric corrections have been performed using the split-window method given
68 M. Menenti et al.

Table 4.2 Central wavelength and bandwidth of ATSR-2 spectral channels


Channel No. Central wavelength Full width at half maximum
(µm) (µm)
1 12.0 11.60–12.50
2 11.0 10.52–11.33
3 3.7 3.47–3.90
4 1.6 1.575–1.642
5 0.87 0.853–0.875
6 0.65 0.647–0.669
7 0.55 0.543–0.565

45 TOC(nd)-TOA11(nd)
TOC(nd)-TOA12(nd)
40
TOC(fw)-TOA11(fw)
35 TOC(fw)-TOA12(fw)

30
Frequency (%)

25
20
15
10
5
0
−2 0 2 4 6 8 10 12 14
Tb(TOC) - Tb(TOA)

Fig. 4.14 Atmospheric impact on surface brightness temperature for two TIR channels at both
nadir and forward views. Data are from ATSR-2 on June 6th 1999 over Spain. (nd = nadir;
fw = forward.)

by Eq. (4.7) (Li et al., 2003). The difference in brightness temperature at a given
wavelength and a given view angle between the TOC and TOA levels (Fig. 4.14)
indicates the magnitude and the dependence on view angle of atmospheric effects.
The temperature at TOC is always higher than the temperature at the TOA level and
this difference is larger at the 45◦ view angle because of the larger atmospheric opti-
cal depth. For a given view angle, the difference is larger for channel 1 (12 µm) than
for channel 2 (11 µm) due to the stronger atmospheric absorption in channel 1. The
surface anisotropy signature is observable after atmospheric correction (Fig. 4.15).
Most pixels show a lower temperature in the forward image than in the nadir im-
age. The histogram peaks around 3.5 K at TOC level and around 5.0 K and 5.2 K at
TOA level for channels 2 and 1, respectively. This figure shows also that the angular
variation observed at TOA comes from both the angular variation at TOC and the
4 Multi-angular Thermal Infrared Observations of Terrestrial Vegetation 69

TOC(nd) - TOC(fw)
25
TOA11(nd) - TOA11(fw)
TOA12(nd) - TOA12(fw)
20
Frequency (%)

15

10

0
−5 0 5 10 15

Tn - Tf

Fig. 4.15 Difference in surface brightness temperature between nadir and forward view at TOC
and TOA. Data are from ATSR-2 on June 6th 1999 over Spain

angular dependency of the atmospheric effects, and the effects of atmosphere gen-
erally increases the TOA anisotropy signature as compared to the angular variation
observed at TOC.
The experimental evidence reviewed in this section leads to conclude that the
thermal heterogeneity of soil–vegetation systems is very significant (see, e.g., Figs.
4.4, 4.8 and 4.11). On the other hand, the same observations suggest that soil–
foliage temperature differences are much larger than the difference between sunlit
and shadow elements, both foliage and soil. In other words, available data suggest
that a 2-components conceptual model of soil–canopy systems may be adequate in
most cases, with 4 components (sunlit, shadow; foliage, soil) necessary under ex-
treme radiative forcing. This conclusion leads to the modeling approach presented
below.

4.4 Modeling the Anisotropic Exitance of Soil-Canopy Systems

Several models have been proposed and developed to describe and handle the
anistropic exitance of soil–canopy systems. They can be loosely classified in three
categories:
1. Simple geometric (deterministic) model of the system: this approach applies
to structured vegetation (row crops, tree lines, patches) inasmuch geometry is
known and the system can be described with a small number of known parame-
ters (e.g., dimensions, soil emissivity, vegetation emissivity, ...). (Sutherland and
70 M. Menenti et al.

Bartholic, 1977; Kimes et al., 1981; Kimes, 1983; Sobrino and Caselles, 1990;
Caselles et al., 1992). Attempts to incorporate a coupling with the atmospheric ra-
diation (down welling) have been scarce (Colton 1996). Such an approach is quite
useful for sensitivity studies or to assess the feasibility of the retrieval coupled to
the atmospheric correction requirements. Except when the geometry is accurately
known, or can be well inferred from other measurements (also from satellite), use-
fulness of this type of model is however limited since it cannot deal with the physical
processes within the system, and model inversion is very sensitive to uncertainties
in parameters.
2. Radiative transfer within the canopy: this approach applies to dense (or
less dense) systems that can be described statistically and using biome char-
acteristics. Models in this domain solve radiative transfer in the canopy with
atmosphere and soil as boundary conditions, assuming plant type, distribu-
tion, plant architecture, LAI (total, horizontally/vertically projected), LIDF, etc.
(Ottermann, 1990; Ottermann et al., 1992, 1995, 1999; Balick et al., 1987; McGuire
et al., 1989; Kimes, 1981; Norman and Chen, 1990; Smith and Goltz, 1995; Smith
et al., 1996). Soil temperature, leaf temperature, temperature gradient within the
canopy may either be assumed, or simultaneously solved for.
Observed TIR anisotropy may reveal whether there exists a temperature gradient
within the canopy or not. However, interpretation of directional radiance implies
that all parameters of the system are known or can be accurately retrieved from
other measurements (from satellite in the visible, NIR and SWIR domains).
Since the fluxes within the canopy are coupled to the flux above the canopy,
the micrometeorological parameters have to be known. It turns out that the surface
TIR directional effect is quite sensitive to ambient conditions. Thus, for a given
biome, the TIR emitted radiance may reverse the sign of its angular variation with
zenith angle (i.e., decrease or increase with increasing zenith angle), or even show
no variation at all.
This category of models may not be best for inverse modeling of observations.
Nevertheless, such models are extremely useful for: (1) for evaluating the order of
magnitude of the angular effect that can be expected and (2) for comparing what is
observed with outputs of models. It is worth noting that the modeled anisotropy is in
no case large, no more than a few K only if large (>60◦ ), zenith angle can be used.
The Geometric-Optics method was initially proposed to model the radiative transfer
through conifer canopy in near infrared and visible domains (Li and Strahler 1985).
After taking into account the component emittances at the pixel scale, the geometric-
optics method has been extended to the thermal infrared domain to model the
angular variation of thermal emission from vegetative canopies at the local scale
(Su et al. 2003). The core idea behind geometric optics method is that, a vegetation
canopy is assumed to be represented by different component elements with different
shapes (such as cone, elliptical, or sphere), the parameters for these shapes (such
as height, radius, and apex angle) are specified, the spatial distribution for these
canopies in pixel plane is previously determined (random, regular, or row), and
the position of solar and sensor is used as input. After accounting for the mutual
shading between canopies, the fractions of four components (sunlit foliage, sunlit
4 Multi-angular Thermal Infrared Observations of Terrestrial Vegetation 71

soil, shaded foliage and shaded soil) are computed through geometrical optical rules.
Then the contributions of each component to the reflected or emitted radiance are
combined to compute the reflected or emitted radiation at top of canopy level or
pixel level. Multi-scattering between different components is also included.
Some improvements on the original geometric optic method have been imple-
mented to take into account the spatial heterogeneity including clumping index of
vegetative canopy which describe the non-random distribution of canopy elements
including leaves, branches, and the non-random spatial distribution of vegetative
canopies in pixel scale (Chen and Leblanc 1997).
A limitation of the geometric optic method is that the mutual shading, gap distrib-
ution inside vegetative canopy, multi-reflection, and scattering cannot be accounted
for. To overcome this shortcoming, a so-called hybrid geometric-optical radiative
transfer method was proposed to improve the original geometric-optical method. By
using this geometric-optical radiative transfer method, Yu et al. (2004) simulated the
directional brightness temperature over a maize canopy with a row structure. The
gap probability between rows was computed with geometric-optical rules, while
the gap inside rows was calculated with radiative transfer theory. Row and canopy
structure parameters including row width, canopy height and width between rows
are needed to initialize their model. The comparison between the simulations with
their model and in situ measurements showed that, the angular variation of bright-
ness temperature can be precisely captured with this type of model.
Radiosity (Gerstl et al., 1991; Gerstl and Borel, 1992) is a computational al-
gorithm to describe the scattering of light between ideally diffuse (Lambertian) sur-
faces. Although this method has been devoted to modeling bi-directional reflectance
distribution function (BRDF) in VIS/NIR optical remote sensing for a long time
(Borel et al., 1991), there are few reports on its application to thermal infrared.
3. Inhomogeneous thick vegetation layer: that can be statistically described by an
angle dependent “gap fraction” or “gap frequency” (Nilson, 1971). This approach
represents an intermediate situation between 1 and 2. It allows the directional TIR
radiance to be described as a linear combination of the foliage and soil radiance con-
tributions, weighted by the gap fraction. Inverting directional TIR radiance may be
used to retrieve vegetation temperature and soil temperature, assuming, for instance,
an angle dependent canopy emissivity. A detailed and comprehensive discussion of
direct and inverse modeling is found in (François et al., 1997).
The advantage of this type of models is that the gap fraction can be phenomeno-
logically correlated with measurements in the visible-NIR domain through appro-
priate vegetation indices (Baret et al., 1995). Hence, combination of measurements
in the solar reflected domain and in the thermal infrared domain may be of great
value, if not yet enough, to solve the problem (i.e., retrieve soil and foliage temper-
atures).
A further step can be made by exploiting detailed BRDF (including hot spot
as much as possible) measurements from satellites in the visible – NIR domain to
assess, e.g., LIDF type, that would help improving the gap fraction estimate.
72 M. Menenti et al.

4.4.1 Detailed 3D Models of Radiative Transfer in Vegetation


Canopies

Realistic simulation of the canopy structure and remote sensing scenario has been
developed to model the radiative transfer of downwelling solar radiance and at-
mospheric radiance reflected by vegetative canopy and soil (Kimes and Kirchner,
1982) and of thermal emission from leaves and soil by Guillevic et al. (2002). The
so-called DART (Discrete Anisotropic Radiative Transfer) model has been devel-
oped (Guillevic et al., 2002; Gastellu-Etchegorry et al., 2004; CESBIO, 2006) to
simulate pixel scene and radiation energy budget in this scene. The whole scene
is specified by the locations of different objects (including tree, lake, soil, lake, and
building) and combined by a 3-D matrix of parallelepiped cells of various sizes. The
“atmosphere” part is simulated with “air” cells, while the “terrestrial” part is made
up of “air” cells. Cells are associated with 4 types of elements (soil, vegetation, wall,
water), with or without relief (DEM, Digital Elevation Model). More details of this
model can be found in Guillevic et al. (2002) and Gastellu-Etchegorry et al. (2004).
A realistic representation (Fig. 4.16) of canopy geometry and of the spatial het-
erogeneity of radiative and convective processes can be constructed using a discrete
3D grid (Welles and Norman, 1991). The grid points are the basic unit in the 3D
model dealing with the radiation, heat and water vapor fluxes calculations. The
components of radiative environment, such as the direct beam radiation from the
sun, the diffuse radiation from the sky and soil, and the diffuse radiation scattered
by foliage at any grid point are calculated. This description of geometry may also
be used to model convection of heat and water vapor. One should note that since we
have assumed wind speed, air temperature and vapor pressure are homogeneously
distributed along the horizontal directions, the grid points at the same vertical level
have identical values of these variables wherever they are located in the horizon-
tal directions.

4.4.2 Simpler Models: Linear vs. Non-linear, 2 vs. 4 Components


Mixture Models

Typically, the pixel or IFOV (Instantaneous Field Of View) of an imaging radiometer


includes both foliage and soil where soil is commonly much warmer than foliage.
The experimental evidence reviewed in Section 4.3 suggests that the thermal het-
erogeneity of a soil-canopy system may be represented under most environmental
conditions using foliage and soil components only. Under extreme radiative forcing
the difference between sunlit and shaded canopy elements becomes larger. Simple
linear and non-linear models have been proposed to simulate radiance or brightness
temperature measured at TOC.
A 2-component mixture model can be used to simulate the directional distribu-
tion of thermal radiation as (Ottermann et al., 1992, 1995):
4 Multi-angular Thermal Infrared Observations of Terrestrial Vegetation 73

Reference height
in atmospheric surface layer

TOC

F
Canopy space
R, u , S, C
Soil surface

Root zone

Fig. 4.16 Simplified schematic illustration of interactions between points (either containing foliage
or soil) in a three-dimensional canopy (soil + vegetation) and in the atmosphere just above the
TOC, with all possible physical, chemical and biophysical processes included. TOC is represented
by “green plane” while soil surface is represented by “orange plane”. The green blocks represent
sub-canopies. Points are symbolized by symbol “•” interaction between points are represented by
“lines”. The symbol implies that the interactions are 3D (vertical and horizontal exchanges).
Each point is characterized by absorbed radiation flux density R, windspeed u, concentration C
of a scalar (i.e., temperature, moisture, CO2 , etc.), source/sink S of a scalar. F represents the flux
density of a scalar between points. (Jia, 2004.)

Bλ (Tb0 ) = f εv Bλ (Tv ) + (1 − f )εg Bλ (Tg ) (4.2)

where, Tb0 is the brightness temperature at TOC. εv and εg are emissivities of fo-
liage and soil, respectively. f is the fractional coverage of vegetation. Tv and Tg
are temperatures of foliage and soil, respectively. Bλ denotes the Planck function.
Previous study showed that the directional thermal radiation is not sensitive to the
uncertainty of soil and leaf emissivities (François et al., 1997). This model implies
that both the multiple reflection between vegetation and soil, and the temperature
differences between sunlit and shaded canopy elements are neglected.
It should be noted that Eq. (4.2) is a linear representation of directional thermal
radiation at top of canopy without taking into account the atmospheric downwelling
radiation and the temperature differences between sunlit and shaded elements inside
canopy. As indicated, under extreme radiative forcing, the sunlit elements are much
warmer than the shaded ones. The latter implies that accurate modeling of TOC
emittance requires taking into account leaf level radiation balance and 3D canopy
74 M. Menenti et al.

structure, either by explicit modeling of leaf level radiative processes or by parame-


terization. Thus, the sunlit and shaded elements need to be treated differently as,
4
Bλ (Tb0 ) = ∑ fi εi Bλ (Ti ) (4.3)
i=1

Where: fi , εi , and Ti are the fraction, emissivity, temperature of component element


i, respectively. Totally, four elements, namely sunlit foliage, shaded foliage, sunlit
soil and shaded soil are included. The equations for a four component system are
derived in a similar way as shown here for two components (Jia, 2004).
The Eqs. (4.2) and (4.3) do not account for the multiple interactions between
canopy elements and soil which is the so-called cavity effect (Sutherland and
Bartholic, 1977). To simulate such a cavity effect, a non-linear item is introduced as
(Li et al., 2001; Menenti et al., 2001):
B(Tb0 (θ )) = P(θ )εs B(Ts ) + [1 − P(θ )] εv B(Tv ) + (1 − Ph )εv B(Tv )P(θ )(1 − εs )
+α (1 − εv )εv B(Tv ) [1 − P(θ )] + β (1 − εv )εs B(Ts ) [1 − P(θ )]
+(1 − εc )Ratm↓ (4.4)

where P(θ ) is the ground fractional cover viewed at angle P(θ ) = 1 − f (θ ), εs and
Ts are the soil emissivity and temperature, εv and Tv are leaf emissivity and tem-
perature, Ph is the hemispheric gap frequency defined as the ratio of the radiation
travelling through the canopy and reaching the soil to the incident radiation into
the canopy over the hemisphere, α and β are respectively the probability of the
radiation emitted by a leaf and reflected by other leaves in the canopy and the prob-
ability of the radiation emitted by soil and reflected by the leaves above it, εc is the
canopy emissivity and Ratm↓ is the downward hemispheric atmospheric radiance di-
vided by π . The first term represents the proportion of the soil radiation that reaches
the top of the canopy in the direction θ. The second term is the upward emitted
radiation from the vegetation in the direction. The third term represents the down-
ward radiation emitted by the vegetation and reflected by the soil and subsequently
traveling upwards through the vegetation in the view direction (vegetation–soil in-
teraction). The fourth term is the contribution of the radiation emitted by the leaves
towards other leaves into the canopy and reflected by these leaves towards outside
the canopy in the view direction (vegetation–vegetation interaction). The fifth term
is the contribution of the radiation emitted by soil towards the leaves and reflected by
these leaves towards outside the canopy in the view direction (soil–vegetation inter-
action), The last term is the downward hemispheric atmospheric radiance reflected
by the canopy system. If we define the effective emissivity of soil and vegetation,
Es and Ev , as:

Es = εs + (1 − Ph )εv (1 − εs )B(Tv )/B(Ts ) (4.5a)


Ev = εv + α (1 − εv )εv + β (1 − εv )εs B(Ts )/B(Tv ) (4.5b)

Eq. (4.4) is reduced to Eq. (4.2) with effective emissivity instead of actual emissivity.
4 Multi-angular Thermal Infrared Observations of Terrestrial Vegetation 75

4.5 Modeling of Observations and Sensor Design

4.5.1 Modeling Approach

The relationship between canopy geometry, leaf and soil properties, radiative and
convective processes is rather complex (see previous Sections) leading to significant
thermal heterogeneity, while radiometric measurements capture the overall effect of
such heterogeneity only.
To understand the nature and information conveyed by radiometric measure-
ments it is useful, therefore, to apply the modeling approach outlined in Section 4.4
to model observations and, particularly, their dependence on canopy (e.g., LAI) and
soil (e.g., water content) properties Jia et al., 2005). Simulation of radiometric data
or full images can be done considering realistic canopy and environmental condi-
tions by combining a soil–vegetation–atmosphere transfer (SVAT) model and at-
mospheric radiative transfer (RT) model (Jia, 2004).
Some TOC biophysical variables needed as input to the SVAT model are sim-
ulated using the RT models PROSPECT and GeoSAIL (Verhoef and Bach, 2003).
The TOC Tb (θ) was simulated using the SVAT model Cupid. Radiative transfer
in the atmosphere is simulated using the RT model MODTRAN 4.0 taking into ac-
count the sensor specifications, particularly spectral coverage, spectral sampling and
channel spectral response. The simulated images of directional radiance in two TIR
channels (11 and 12 µm) at top-of-atmosphere (TOA) are obtained by adding the
atmospheric effects to the top-of-canopy radiance (or images).
The SVAT model Cupid is a one-dimensional, multi-layer model that simu-
lates various plant-environment interactions (Norman, 1979; Norman and Camp-
bell, 1983; Wilson et al., 2003). The essential processes simulated in the Cupid
model are divided as above-ground processes and below-ground processes stratified
by canopy layers (see Fig. 4.16).
Above-ground processes are dominated by vegetation including the transport of
energy, mass and momentum between plants and their environment. Above-ground
canopy is stratified by horizontal layers by identical increment of leaf area index
and each horizontal vegetation layer is stratified by leaf angle class. Interactions
between leaves in each individual leaf angle class in each horizontal layer and their
local environment are first formulated. Collective effect of all the leaves in each hor-
izontal layer is integrated to obtain the response of the layer. Canopy-level responses
are simulated by numerical integration over all canopy layers where soil layers are
also included.
The below-ground processes describe the transport of heat and mass between the
roots and their soil environment and between the soil layers defined by identical
increment of depth. Unlike many other SVAT models that take soil surface as lower
boundary, Cupid has been developed to have plant root zone as the lower boundary.
The soil lower boundary conditions consist of soil temperature and soil water con-
tent near the bottom of the root zone. All the processes occurring at the soil surface
layer are simulated rather than input.
76 M. Menenti et al.

Leaf optical (reflectance, transmittance and emissivity) and physiological


properties (photosynthetic rate, respiration rate and stomatal conductance) and
leaf position and orientation arrangement determine energy and mass exchanges
at leaf level, while canopy structure (leaf density distribution, total leaf area, and
canopy height) determines the distribution of absorbed solar radiation in canopy
layers thereafter determines mass and heat transfer between layers.
The forcing of all the processes described in Cupid model is ambient atmospheric
conditions above canopy and soil boundary conditions at the bottom of the root
zone. Ambient atmospheric conditions are defined by solar radiation, air temper-
ature, humidity, wind speed, and precipitation measured at some reference height
above the canopy. Equations describing heat and mass throughout the entire canopy
(leaf energy budget for all leaves and the vertical flux-gradient equations of soil
and vegetation layers) are solved simultaneously. Among others canopy layer tem-
perature profile (including soil surface layer) and thermal radiation flux profile are
obtained from the solution of Cupid which can then be used to simulate canopy di-
rectional radiance in the concerned wavelength. This is done by assuming that the
contributions of various leaf layers and of the soil layer are appropriately weighted
by the fraction of each layer viewed by the radiometer in a particular view direction.
An overview of the simulation procedure is given in Fig. 4.17.

4.5.2 Generation of Synthetic Multi-angular Images

The at-sensor radiance is a combination of the surface-emitted radiance and the at-
mospheric contributions. The anisotropy of at-sensor radiance is due to atmospheric
scattering, absorption and emission in addition to the inherent anisotropy of TOC
radiance. The most significant interaction of TIR radiance with the atmosphere is
attributed to atmospheric absorption primarily due to ozone and water vapor in the
atmosphere. The optical path length is greater in the off-nadir views where the wa-
ter vapor path increases, thus contributing higher upwelling atmospheric radiance,
while atmospheric transmittance is smaller.
To demonstrate the approach described above a simulated data has been gen-
erated (Fig. 4.17; Jia et al., 2005) for three European sites with a heterogeneous
and detailed spatial structure to represent major, rather different biomes, such as
evergreen forest, deciduous forest, savannah, semi-grasslands and agricultural land.
Result presented here relate to two sites briefly described below:
1. Alpine Foreland site. The land surface is classified by 22 land use classes with
grassland, mixed forest, shrubs, coniferous forest, deciduous forest, and various
crops as dominant land cover types. Four dates were selected for image simu-
lation: 22/042009, 17/06/2009, 19/07/2008, 18/09/2008. The central coordinates
of these images are about 47.9◦ N and 11.1◦ E.
4 Multi-angular Thermal Infrared Observations of Terrestrial Vegetation 77

Broadband emissivity
of leaf and soil

Other inputs Cupid model Sensor geometry Atmospheric


profiles

- canopy temperature profile


(including soil surface layer)
- thermal radiation flux profile
MODTRAN

Canopy RT
model
τ(λ,θ, ϕ), Ratm↑(λ,θ, ϕ), Ratm↓(λ)

εc(λ ,θ, ϕ) Trad(λ ,θ, ϕ) Tb0(λ ,θ, ϕ)TOC

Atmospheric RT Sensor channel response


model function

Tb(11, θ, ϕ) TOA
Tb(12, θ, ϕ) TOA

SPECTRA images of
Spatial model Tb(11, q, j) TOA
Tb(12, q, j) TOA

Fig. 4.17 The major steps and applied models in the TIR image simulation. (Jia et al., 2005.)

2. Boreal forest site. Boreal forest site is located in Sodankyla of Finland with typ-
ical boreal forest covered mainly by deciduous forest (ca. 42%) and coniferous
forest (ca. 40%) and dark litter (ca. 12%). The soil in Sodankyla site is mainly
sandy type.
After obtaining the atmospheric path radiance and transmittance using MODTRAN
4.0 in combining the atmospheric profiles at the two sites, the TOA TIR images
were generated by applying these variables to the TOC TIR images. Figure 4.18
shows the images at the Sodankyla site as an example.
78 M. Menenti et al.

Fig. 4.18 Synthetic multi-angular images of TOC Tb (θ ); Boreal forest at Sodankyla, Finland:
θ = 0◦ (left) and θ = +60◦ (right).

4.6 Retrieval Algorithms

4.6.1 Introduction

An operational algorithm was described by Jia et al. (2003) to retrieve soil and fo-
liage component temperatures over heterogeneous land surface based on the analy-
sis of bi-angular multi-spectral observations made by ATSR-2. This algorithm is a
good illustration of the synergistic use of multiangular observations in the VNIR-
SWIR region and in the TIR region to retrieve simultaneously both land surface
and atmospheric variables (see also Verhoef and Menenti, 1998). This algorithm is
described in this section. On the basis of the radiative transfer theory in a canopy,
a model is developed to infer the two component temperatures using six channels
of ATSR-2. Four visible, near-infrared and short wave infrared channels are used
to estimate the fractional vegetation cover within a pixel. A split-window method
is developed to eliminate the atmospheric effects on the two thermal channels. An
advanced method using all four visible, near-infrared and short wave channel mea-
surements at two view angles is developed to perform atmospheric corrections in
those channels allowing simultaneous retrieval of aerosol optical depth and land
surface bi-directional reflectance (Fig. 4.19).
This general approach has been applied to image data acquired with the airborne
AMTIS (see Section 4.3). The algorithm was applied to retrieve foliage and soil
temperatures of a winter wheat canopy (Liu, 2002). since detailed ground measure-
ments of component temperatures and of directional emittance were only available
for this land cover type. Validation could, therefore, be done only for retrievals over
winter wheat. All other land cover types were left out of consideration (light blue
area in Fig. 4.20).
4 Multi-angular Thermal Infrared Observations of Terrestrial Vegetation 79

(TOA) ATSR data :Tb (nadir), Tb (forward), ρ (nadir),

Cloud

Retrieval of Aerosol Optical Depth Retrieval of Water Vapor

Atmospheric correction Atmospheric correction


for VIS / NIR / SWIR channels for TIR channels

Fractional vegetation cover (BOA) Surface brightness temperature


ff (nadir), ff (forward) Tb0 (nadir), Tb0 (forward)

Spatial Smoothing

Inversion model

Component temperatures
Tf , Ts

Fig. 4.19 Scheme of the operational algorithm for retrieval of Tf and Ts from ATSR multi-spectral
and dual-angular measurements. Tb is the brightness surface temperature at TOA measured by TIR
channels of ATSR-2; ρ is reflectance at TOA measured by the VIS/NIR/SWIR channels of ATSR-
2. (After Jia et al., 2003.)

4.6.2 Retrieval of Tb (θ ) at TOC (Top of Canopy) from Tb (θ )


at TOA (Top of Atmosphere)

Since the satellite measures the TOA brightness temperature (T ), and the inversion
of separate soil and vegetation temperature model needs the TOC brightness tem-
perature (Tb0 ), atmospheric corrections have to be performed. Moreover, since di-
rectional ground radiance (equivalent to ground brightness temperature) is likely to
result from radiometric 3-D heterogeneity of the surface or the surface cover, neither
a kinetic surface temperature nor an emissivity can be simply and uniquely defined
80 M. Menenti et al.

Fig. 4.20 Foliage (left) and soil (right) component temperatures determined from AMTIS multi-
angular measurements of exitance at 4,200 m height; Shunyi experiment, China, April 2001. (After
Liu, 2002.)

(Becker and Li, 1995) or even a value or angular behaviour assumed. Any separation
between temperature and emissivity would rely on a priori assumptions or “defin-
itions” of such variables. Thus, the first step in looking for directional effects is to
consider the ground radiance as a whole. The consequence of that is that accurate
atmospheric corrections have to be applied to TOA radiances (or brightness tem-
peratures) in a given channel, preferably less affected by atmospheric perturbations,
or, to say things differently, a common Split-Window method (SW) is not adequate
(nor is a double angle method with (A)ATSR nadir/forward views). Indeed, a SW
or similar regression algorithm for atmospheric corrections is designed to give the
surface kinetic temperature, based on several assumptions regarding spectral and
angular emissivity.
Single channel atmospheric correction is “just” inverting the radiative transfer
equation integrated over the channel (i) bandwidth:

Bi (Ti ) = Bi (Tb0i )τi + Ratm


i ↑ (4.6)

in order to get the ground radiance Bi (Tb0i ), where Tb0i is the ground brightness
temperature, Ti is the TOA measured brightness temperature, τi is the atmospheric
transmission and Ratmi ↑ is the atmospheric upward radiance. All quantities refer to
a particular view direction, defined by zenith angle θ at ground level. The accuracy
on Bi (Tb0i (θ)) is determined by that on the atmospheric quantities, which in turn
depends on the goodness of the radiative transfer code and the description of the
vertical structure of the atmosphere. The second condition is by far the most impor-
tant source of error. There is no simple parameterisation allowing for determination
of both τi and Ratm
i ↑. Atmospheric PTU profiles are needed.
Alternatively, if the ground brightness temperature is, or can be assumed,
independent of the channels used to measure it, the Split Window (SW )
(Sobrino et al., 1994; Becker and Li, 1995) method can be used to get Tb0 . Following
the procedure developed by Becker and Li (1990, 1995) a general SW algorithm
is derived for ATSR-2 nadir and forward views using simulated radiometric data
(Li et al., 2003):
4 Multi-angular Thermal Infrared Observations of Terrestrial Vegetation 81

Tb0 (θ ) = [a(θ ) + b(θ )W ] + [c(θ ) + d(θ )W ]T11 (θ )


+[e(θ ) + f (θ )W ][T11 (θ ) − T12 (θ )] (4.7)

where θ is view angle, W is the total column water vapor in the atmosphere. For
the large range of surface parameters and atmospheric conditions (W ≤ 4.5 g/cm2 ,
air temperature at surface Ta , 272 K ≤ Ta ≤ 311 K and −5 K ≤ Tg − Ta ≤ 15 K), the
coefficients a- f have been generated for ATSR-2 nadir and forward views. They
are a = −4.89, b = 3.74, c = 1.0205, d = −0.0151, e = 0.916, f = 0.509 the rms
residual retrieval error σ = 0.10K for the nadir image, and a = −14.41, b = 8.51, c =
1.0582, d = −0.0343, e = 0.565, f = 0.857σ = 0.24K for the forward view.

4.6.3 Retrieval of Water Vapor from ATSR-2 Split-window Channel


Data over Land

Water vapor content in the atmosphere is required to improve the accuracy of the
remotely sensed surface parameters (Sobrino et al., 1994, Francois and Ottle, 1996).
Nowadays, a number of different satellite approaches have been proposed and de-
veloped over the past two decades to measure atmospheric water vapor. According
to the wavelength used, these approaches may be grouped into three categories:
Near-infrared techniques (Frouin et al., 1990, Kaufman and Gao, 1992); Passive
microwave techniques (Prabhakara et al., 1985, Alishouse et al., 1990, Schulz et al.,
1993); and Thermal infrared techniques (Chesters et al., 1983, Susskind et al., 1984,
Kleespies and McMillin, 1990, Jedlovec, 1990, Iwasaki, 1994, Ottlé et al., 1997,
Sobrino et al., 1999).
Because the near-infrared technique is based on detecting the absorption by water
vapor of the reflected solar radiation as it is transferred down to the surface and up
through the atmosphere, use of this technique needs to have at least one channel in
the water absorption band (0.94 µm), and one nearby channel in the atmospheric
windows (0.86 µm, 1.05 µm and 1.24 µm). Since ATSR-2 (Along-Track Scanner
Radiometer) on board ERS-2 (European Remote Sensing) has only four channels in
the visible and near infrared domain (0.55 µm, 0.65 µm, 0.87 µm, 1.60 µm) and three
channels in the thermal infrared domain (3.7 µm, 11 µm and 12 µm), no channel in
the water absorption band is available, the near-infrared technique cannot be applied
to ATSR2 data, the only one applicable technique is thermal infrared technique.
Up to now, there have been several attempts to derive water vapor using two split-
window channels (11 µm and 12 µm). For instance, Kleespies and McMillin (1990)
proposed a method based on the ratio of split-window channel brightness tem-
perature differences assuming that the atmosphere and surface emissivities in the
split-window channels are invariant. Jedlovec (1990) proposed an extension of this
concept and showed that the water vapor content can be derived using the ratio of
the spatial variance of the channel brightness temperature. On the basis of these
methods, Iwasaki (1994) developed a new algorithm to reduce the non-linear ef-
fect of air temperature and unresolved cloud effect on the estimation of water vapor
82 M. Menenti et al.

content using the split-window data. Sobrino et al. (1994) improved Jedlovec (1990)
method by the use of Split-Window Covariance-Variance Ratio (SWCVR). It has
been shown that all these split-window methods are sensitive to instrument noise
and are difficult to be applied to satellite data such as AVHRR in an operational
manner (Sobrino et al., 1994, 1999). Under the condition that the atmosphere and
directional surface emissivity are constant or the effects of their spatial variations
are not larger than the combined effects of both instrument noise over the N neigh-
boring pixels, Li et al. (2003) presented a new algorithm to determine quantitatively
column water vapor content (W) directly from (A)ATSR Split–Window radiance
measurements using the following formulae:
(a) For ATSR2 nadir view:
W = 13.73 − 13.662τ j /τi (4.8)
where the subscripts i, j denote respectively channel 11 µm and channel 12 µm
of ATSR.
(b) For ATSR2 forward view (θ ∼ = 53◦ ):
W = 10.02 − 9.971τ j /τi (4.9)
with N
∑ (Ti,k − T i )(T j,k − T j )
τj εi k=1
= R ji with R ji = , (4.10)
τi εj N
∑ (Ti,k − T i )2
k=1
in which the subscript k denotes pixel k, Ti,k and T j,k are the brightness temper-
atures of pixel k in channels i and j measured at satellite level, respectively, T i
and T j are the mean (or the median) brightness temperatures of the N neighbor-
ing pixels considered, respectively.
This method was developed and applied to several ATSR2 data sets. The water
vapor contents retrieved using ATSR2 data from SGP’97 (USA), Barrax (Spain) and
Cabauw (The Netherlands) are in good agreement with those measured by the quasi-
simultaneous radiosonde. The mean and the standard deviation of their difference
are respectively 0.04 and 0.22 g/cm2 . It is shown that water vapor content derived
from ATSR2 data using the proposed algorithm is accurate enough in most cases for
surface temperature determination with split-window technique using ATSR2 data
and for atmospheric corrections in visible and near-infrared channels of ATSR2.
A more detailed description of this method and its applicability can be found in Li
et al. (2003).

4.6.4 Retrieval of Aerosol Optical Depth from ATSR-2 Data


for Atmospheric Corrections

Atmospheric perturbations (mainly due to absorption and scattering processes)


are responsible for substantial modifications of the surface spectral reflectance
4 Multi-angular Thermal Infrared Observations of Terrestrial Vegetation 83

measured by satellite instruments. It is therefore necessary to correct the at-


mospheric effects to retrieve the surface reflectance. Methods of atmospheric
corrections are generally concerned with the estimation of the atmospheric effects
associated with molecular absorption, molecular and aerosol scattering. Current
methods for the estimation of the atmospheric effects employ a radiative transfer
model (Vermote et al., 1997, Beck et al., 1999) whose inputs are generally the
vertically integrated gaseous contents, aerosol optical properties and geometric
conditions.
If ρi∗ is the reflectance measured in channel i at the top of atmosphere (TOA),
from radiative transfer theory, the surface reflectance in channel i, ρi , can be ex-
pressed as (Rahman and Dedieu, 1994; Vermote et al., 1997)

ρiac (θs , θυ , ∆φ )
ρi (θs , θυ , ∆φ ) = (4.11)
1 + Si ρiac (θs , θυ , ∆φ )

with 
ρi∗ (θs , θυ , ∆φ ) tgi (θs , θυ )−ρi (θs ,θυ ,∆φ )
a

ρiac =
ti (θs )ti (θυ )
where θs and θυ are solar and viewing zenith angles, respectively. ∆φ is the relative
azimuth between sun and satellite direction. Si is the spherical albedo of atmosphere
in channel i. tgi is the total gaseous transmission in channel i associated with gaseous
absorption along the sun–target–sensor atmospheric path. ρia (θs , θυ , ∆φ ) is the at-
mospheric reflectance. ti (θs ) and ti (θυ ) are the total atmospheric scattering trans-
mittance along the sun-target and target-sensor atmospheric paths, respectively.
In general, the independent measurements of atmospheric composition and
aerosol optical properties are not available; it is therefore desirable to derive them
directly from satellite data. The most important gases in atmospheric corrections in
visible and near infrared channels are water vapor and ozone. Water vapor content
in the atmosphere may be derived from the two split-window channel measure-
ments as shown above, and ozone content is taken from climatological data. As
for the determination of the aerosol optical properties, if the surface reflectance
may be considered isotropic, then the difference in surface reflectance retrieved
from multi-angle directions using Eq. (4.1) may be used to derive the atmospheric
optical thickness if aerosol type is assumed. However, most land surfaces are far
from Lambertian (Hapke, 1981) With multi-angle measurements, it is impera-
tive to consider non-Lambertian reflectances. Several multi-look aerosol retrieval
schemes for ATSR-2 have been proposed (Flowerdew and Haigh, 1997; Mackay
and Steven, 1998; North et al., 1999). The iteration of a two step-process proposed
by North et al. (1999) can be used. The first step is to derive using Eq. (4.1) eight
land surface reflectances ρi (θs , θv , ∆φ ) from the TOA reflectance ρ ∗ made at four
channels (0.55 µm, 0.65 µm, 0.87 µm, 1.60 µm) and two view angles (nadir and for-
ward views), agiven
 an initial estimate of the atmospheric aerosol and optical depth
at 550 nm τ550 . The second step is to fit land surface bi-directional reflectance
84 M. Menenti et al.

model to eight retrieved surface reflectances by the minimization of the error metric
function
2 4
E= ∑ ∑ [ρi (θs , θv , ∆φ ) − ρim (θs , θv , ∆φ )]2 (4.12)
θv =1 i=1

where ρim (θs , θv , ∆φ ) is the land reflectance in channel i predicted by the reflectance
model.
Since there are maximum eight land surface reflectance measurements, land sur-
face bi-directional model must have maximum seven free model parameters so
that there is at least 1◦ of freedom available a  for atmospheric parameter retrieval,
for instance, the aerosol optical depth τ550 . Considering the land surface to be
composed of opaque facets, each with Lambertian reflectance ωi , and separating
parameters relating to the wavelength invariant three-dimensional structure of the
surface from wavelength dependent parameters describing the component spectra,
North et al. (1999) developed a seven free parameter (P nadir, P forward, γ and
ωi (i = 1, 4)) model as
γωi
ρi (θs , θv , ∆φ ) = (1 − Di )P(θs , θv , ∆φ )ωi + [Di + (1 − γ )ωi (1 − Di )]
1 − (1 − γ )ωi
(4.13)
where Di is the incident diffuse fraction which excludes the radiation scattered close
to the solar beam direction. Di can be estimated by radiative transfer model for solar
direction and aerosol optical depth. P(θs , θv , ∆φ ) is the geometric parameter depen-
dent only on view and illumination directions, γ denotes the mean hemispherically
integrated probability of escape of light without further interaction, after a scattering
event at the land surface.
In case where there are no four channels available, an alternative scheme can be
used to retrieve the aerosol optical depth by assuming that the functional shape of
the bidirectional effects is invariant with respect to the wavelength within the visible
and near-infrared region (Flowerdew and Haigh, 1997; Mackay and Steven, 1998),
namely:
ρi (θs , θυ 1 , ∆φ ) ρ j (θs , θυ 1 , ∆φ )
= (4.14)
ρi (θs , θυ 2 , ∆φ ) ρ j (θs , θυ 2 , ∆φ )
This relationship gives a constraint for atmospheric correction by forcing the re-
trieved bidirectional reflectance to have a consistent angular variation, even thought
the magnitude of the reflectance may vary greatly.
The aerosol optical thickness is therefore obtained through the minimization of
the error metric function
n n  
ρi (θs , θυ 1 , ∆φ ) ρ j (θs , θυ 1 , ∆φ ) 2
E =∑∑ − (4.15)
i=1 j>i ρi (θs , θυ 2 , ∆φ ) ρ j (θs , θυ 2 , ∆φ )

where n is the total number of channels available, i and j are channel numbers. The
details of these methods can be found in North et al. (1999).
4 Multi-angular Thermal Infrared Observations of Terrestrial Vegetation 85

4.6.5 Retrieval of Vegetation Fractional Cover

To retrieve soil and foliage temperatures from Eqs. (4.2) or (4.4), the viewing angle-
dependent vegetation fraction needs to be determined independently from additional
observations, such as the TOC reflectances in the VNIR and SWIR spectral channels
of ATSR-2 and AATSR. The approach presented here is an example, while for a
review on retrieval of fractional cover see, e.g., Weiss et al. (1999). The approach to
estimate the fractional vegetation cover uses visible and near infrared data.
A stepwise multiple linear regression were applied in this study to estimate the
fractional vegetation cover ff (θv ) using TOC reflectances ρi (θs , θv , ∆ϕ ). The step-
wise multiple linear regression is written
n
ff (θv ) = a0 (θs , θv ) + ∑ ai (θs , θv )ρi (θs , θv , ∆ϕ ) (4.16)
i=1

where n is the number of channels used.


The model OSCAR (Verhoef, 1998) has been applied to generate surface re-
flectances for the viewing and illumination conditions applying to the ATSR-
observations used in this study and to an ensemble of canopy and atmospheric con-
ditions. This synthetic data base, where f f is known for any given canopy condition,
has been used to determine the coefficients ai in Eq. 4.16. A different set of coeffi-
cients is obtained for each viewing geometry (see also Jia, 2004).

4.7 Conclusions

Vegetation fractional cover. The regression model Eq. (4.16) works rather well when
vegetation is green, i.e., corn and alfalfa, while slightly larger errors appear for
vegetation with some fraction of brown leaves, like barley at this time of the year
(DOY = 179). Table 4.3 gives the summary of the model performance represented
by RMSE and Absolute Difference (AD) for each case and for the total data set.

Table 4.3 The performance of the regression equation for estimating fractional vegetation cover
Eq. (4.15) as presented by RMSE and AD at nadir and forward 53◦ for each vegetation type. “Total”
indicates the overall performance over the entire simulation dataset
Vegetation θv = 0◦ θv = 53◦

RMSE AD RMSE AD

Corn 0.0339 0.025 0.1038 0.103


Barely 0.1086 0.109 0.0745 0.073
Alfalfa 0.0327 0.030 0.0515 0.049
Total 0.0456 0.046 0.0785 0.076
86 M. Menenti et al.

Table 4.4 RMSE between retrieval and simulation of Tf and Ts for corn, barley and alfalfa crops
Corn Barley Alfalfa
RMSE of Tf 0.381905 1.164064 0.590805
RMSE of Ts 0.669429 0.867213 0.772624

Retrieval of soil and foliage component temperatures. The validation of the re-
trieved component temperatures using ATSR-2 bi-angular measurements is chal-
lenging due to the difficulty of obtaining observations of temperatures of soil and
foliage in situ at ATSR-2 spatial resolution (i.e., 1.5 × 2.5 km for the forward view.
As done to evaluate the algorithm for ff (θv ), the same method, i.e., using syn-
thetic radiometric data generated with detailed modeling of radiative transfer in the
soil–vegetation–atmosphere system and inversion using a simplified algorithm (suit-
able for operational processing of actual data), can be used to evaluate the retrievals
of Tf . The reference Tf and Ts were simulated using the complete model CUPID
and the TOC radiance was calculated (for each image pixel) with the procedure
outlined in Section 4.5. The Tb0 (θv , ϕv ) at θv = 0◦ and θv = 53◦ obtained in this
way were then treated as observations in Eq. 4.4 to retrieve Tf and Ts . The compar-
isons between the retrievals and the simulations of Tf and Ts for the three crops are
encouraging (Table 4.4).

4.8 Conclusions: Limitations of Current Systems


and Perspectives

Over the last 25 years measurements of emittance by goniometer-mounted radiome-


ters has been the main source of observations to document and understand the
anisotropic emittance of vegetation canopies. This experimental body of knowl-
edge led to the development of detailed models of the complex processes which
determine the significant thermal heterogeneity of the canopy space. The latter is of
particular relevance towards a better understanding of the relation between canopy
architecture and foliage–atmosphere exchanges of radiation, water and carbon.
On the other hand, the capability of goniometer-mounted radiometers is severely
limited both as regards the size of targets and sampling of the spatial variability.
Actual exploitation of the unique information conveyed by anisotropic emittance re-
quires imaging radiometers to sample sufficiently large terrestrial targets at a spatial
resolution sufficiently large to integrate emitted radiance over an area representative
of the target, but sufficiently small to provide independent observations of different
land cover types.
The ATSR series of instruments is the only current observing system providing
directional (at two view angles) observations of emittance and these observations
have been used to demonstrate the relevance of directional measurements in the
4 Multi-angular Thermal Infrared Observations of Terrestrial Vegetation 87

TIR region. (see, e.g., Jia et al., 2001; Menenti et al., 2001; Hurk et al., 2002). On
the other hand, the large footprint of (A)ATSR and even more the large difference in
footprint between the nadir and forward view, make validation of data products very
challenging and restrict applications to large scale modeling. For example, the dif-
ference between foliage and soil temperature might be used as a measure of drought
or crop water stress, but given the low resolution is difficult to relate to the spatial
scale of crop and water management practices.
The European Space Agency had been studying a high spatial resolution mission
for multi-angular land observations (SPECTRA, see Menenti et al., 2005), but this
project has been abandoned. The Italian (ASI) and Canadian (CSA) Space Agencies
are developing an Earth Observation mission with similar characteristics (Galeazzi
et al.,2006) with expected launch in 2010. The latter would provide simultaneous
VNIR – SWIR hyper-spectral and multi-spectral observations at high spatial reso-
lution and is being designed specifically for land observations.
In general terms, multi-angular measurements of TIR emittance over land pro-
vide access to radiative and convective processes in the canopy space which concur
to determine the response of terrestrial vegetation to environmental forcing.

Acknowledgements Work by the Authors summarized in this chapter has been funded by
the European Space Agency (Contracts 13177/98NL/GD; 15164/01/NL/SF; 17169/03/NL/GS;
17179/03/NL/GS), the European Commission (Contracts FP6 GMES no. 502057, FP5 MIND
EVK2-CT-2002-00158), the Netherlands Users Support Program (GO-2 Contract SRON EO-
049.) and the Italian Space Agency (ASI Contract ASI I/R/073/01). The Shunyi field campaign
was part of the “Quantitative of Remote Sensing theory and application for Land Surface Pa-
rameters (QRSLSP)” project funded by China’s Special Funds for Major State Basic Research
(project No. G2000077900, led by Prof. X. Li). Dr. Qiang Liu is thanked for the data processing
of goniometer and AMTIS measurements.

References

Albertson J, Katul G, Wiberg P (2001) Relative importance of local and regional controls on cou-
pled water, carbon, and energy fluxes. Adv. Water Resour. 24:1103–1118
Alishouse JC, Snyder SA, Vongsathorn J, Ferraro RR (1990) Determination of oceanic total pre-
cipitable water from the SSM/I. IEEE Trans. Geosci. Remote Sens. 28:811–816
Avissar R, Pielke RA (1989) A parameterization of heterogeneous land-surface for atmospheric
numerical models and its impact on regional meteorology. Month. Weather Rev. 117:2113–
2136
Balick LK, Hutchison BA, Smith JA, McGuire MJ (1987) Directional thermal exitance distribu-
tions of a deciduous forest in summer. IEEE Trans. Geosci. Remote Sens. 25:410–412
Baret F, Clevers JPW, Stevens MD (1995) The robustness of canopy gap fraction estimates from
the red and near infrared reflectance: a comparison between approaches for sugar beet canopies.
Remote Sens. Environ. 54:141–151
Beck A, Anderson GP, Acharya, PK, et al. (1999) MODTRAN4 User’s Manual, Air Force Re-
search Laboratory, Hanscom AFB, MA
Becker F, Li Z-L (1990) Towards a local split window method over land surfaces. Int. J. Remote
Sens. 11(3):369–393
88 M. Menenti et al.

Becker F, Li Z-L (1995) Surface temperature and emissivity at various scales: definition, measure-
ment and related problems. Remote Sens. Rev. 12:225–253
Becker F (1987) The impact of spectral emissivity on the measurement of land surface temperature
from a satellite. Int. J. Remote Sens. 8:1509–1522
Becker F, Ramanantsizehena P, Stoll M-P (1985) Angular variation of the bidirectional reflectance
of bare soils in the thermal infrared band. Appl. Opt. 24(3):363–375
Beljaars ACM, Holtslag AAM (1991) Flux parameterization over land surface for atmospheric
models. J. Appl. Meteorol. 30:327–341
Blad BL, Rosenberg NJ (1976) Evaluation of resistance and mass transport evapotranspiration
models requiring canopy temperature data. Agron. J. 68:764–769
Borel CC, Gerstl SA, Powers BJ (1991) The radiosity method in optical remote sensing of struc-
tured 3-D surfaces. Remote Sens. Environ. 36:13–44
Bowen IS (1926) The ratio of heat losses by conduction and by evaporation from any water surface.
Phys. Rev. 27:779–787
Caselles V, Sobrino JA, Coll C (1992) A physical model for interpreting the land surface tempera-
ture obtained by remote sensors over incomplete canopies. Remote Sens. Environ. 39:203–211
CESBIO, 2006, DART User Manual v1.1.
Chen JM, Leblanc S (1997) A 4-scale bidirectional reflection model based on canopy architecture.
IEEE Trans. Geosci. Remote Sens. 35:1316–1337
Chesters D, Uccellini LW, Robinson P (1983) Low-level water vapor fields from the VISSR at-
mospheric sounder (VAS) split-window channels. J. Appl. Meteorol. 22:725–743
Choudhury BJ, Reginato RJ, Idso SB (1986) An analysis of infrared temperature observations over
wheat and calculation of the latent heat flux. Agric. Forest Meteorol. 37:75–88
Colton AL (1996) Effective thermal parameters for a heterogeneous land surface. Remote Sens.
Environ. 57:143–160
De Bruin HAR, Jacobs CMJ (1989) Forest and regional scale processes. Phil. Trans. R. Met Soc.
B (324):393–406
Feddes RA (1971) Water, heat and crop growth. Ph.D. thesis, Wageningen, Agricultural University,
The Netherlands. H. Veenman, Zonen, Wageningen, The Netherlands, 184pp
Field CB, Mooney HA (1986) The photosynthesis – nitrogen relationship in wild plants. In: TJ
Givnish (ed) On the economy of plant form and function. Cambridge University Press, New
York, pp 25–55
Flowerdew RJ, Haigh JD (1997) Retrieving land surface reflectances using the ATSR-2: a theoret-
ical study. J. Geophys. Res. 102(D14):17163–17171
Francois C (2002) The potential of directional radiometric temperatures for monitoring soil and
leaf temperature and soil moisture status. Remote Sens. Environ. 80:122–133
François C, Ottlè C (1994) Estimation of the angular variation of the sea surface emissivity with
the ATSR/ERS-1 data, Remote Sensing of Environment, 48:302–308
Francois C, Ottle C (1996) Atmospheric corrections in the thermal infrared: global and water va-
por dependent split-window algorithms-applications to ATSR and AVHRR data. IEEE Trans.
Geosci. Remote Sens. 34(2):457–470
Francois C, Ottle C, Prevot L (1997) Analytical parameterization of canopy directional emissivity
and directional radiance in thermal infrared. Application on the retrieval of soil and foliage
temperatures using two directional measurements. Int. J. Remote Sens. 18(12):2587–2612
Frouin R, Deschamps P-Y, Lecomte P (1990) Determination from space of atmospheric total water
vapor amounts by differential absorption near 940 nm: theory and airborne verification. J. Appl.
Meteorol. 29:448–460
Galeazzi C, Gamache M, Hollinger A, De Cosmo V, Menenti M, Staenz K, Beck J, Bergeron
M, Ananasso C, Loizzo R, Maszkiewicz M (2006) ASI-CSA joint development phase for a
hyperspectral mission. In: I. Reusen and J. Cools (eds.) Imaging Spectroscopy: Innovation in
Environmental Research. EARSEL, Paris:1–11
Gastellu-Etchegorry JP, Martin E, Gascon F (2004) DART: a 3D model for simulating satellite
images and studying surface radiation budget. Int. J. Remote Sens. 25(2):73–96
4 Multi-angular Thermal Infrared Observations of Terrestrial Vegetation 89

Gerstl SAW, Borel CC (1992) Principles of the radiosity method versus radiative transfer for
canopy reflectance modelling. IEEE Trans. Geosci. Remote Sens. 30(2):271–275
Gerstl SAW, Borel CC, Powers BJ (1991) The radiosity method in optical remote sensing of struc-
tured 3D surfaces. Remote Sens. Environ. 36:13–44
Guillevic P, Gastellu-Etchegorry JP, Demarty J, Prévot L (2002)Thermal infrared radiative trans-
fer within three-dimensional vegetation. J. Geophys. Res. (Atmos.) 108(D8):ACL 6-1, CiteID
4248, doi:10.1029/2002JD002247
Hall AE, Canell GH, Lawton HW (1979) Agriculture in semi-arid environments. Springer, Berlin–
Heidelberg, New York, 340pp
Hapke B (1981) Bidirectional reflectance spectroscopy. 1. Theory. J. Geophys. Res. 86:3039–3054
Hatfield JL, Reginato RJ, Idso SB (1984) Evaluation of canopy temperature – evapotranspiration
models over various crops. Agric. Meteorol. 32:41–53
Hatfield JL (1991) Development of agricultural crops and practices adapted to arid land conditions,
1991. Proc. SCS Multiconference on AI simulation: Agricultural Crops and Practices Adapted
to Arid Lands. University of New Mexico Press, Albuquerque, pp 39–60
Huband NDS, Monteith JL (1986) Radiative surface temperature and energy balance of a wheat
canopy. Bound.-Lay. Meteorol. 36:1–17
Hurk BJJM, Jia L, Jacobs C, Menenti M, Li Z-L (2002) Assimilation of land surface temperature
data from ATSR in an NWP environment – a case study. Int. J. Remote Sens. 23(24):5193–5210
Iwasaki H (1994) Estimation of precipitable water over land using the split-window data from the
NOAA satellite. J. Meteorol. Soc. Jpn. 72(2):223–233
Jedlovec GJ (1990) Precipitable water estimation from high-resolution split window radiance mea-
surements. J. Appl. Meteorol. 29:863–876
Jia L, Menenti M, Su ZB, Djepa V, Li Z-L, Wang J (2000) Modeling sensible heat flux using
estimates of soil and vegetation temperatures: the HEIFE and IMGRASS experiments. In:
M Beniston, M Verstraete (eds), Satellite remote sensing and climate simulations. Kluwer,
Dordrecht, The Netherlands
Jia L, Menenti M, Su ZB, Djepa V, Li Z-L, Wang J (2001) Modeling sensible heat flux using
estimates of soil and foliage temperatures: the HEIFE and IMGRASS experiments, in “Remote
sensing and climate modeling: Synergies and Limitations”, M Beniston and M Verstraete (eds)
Kluwer, Dordrecht, 23–49
Jia L, Li Z-L, Menenti M, Su Z-B, Verhoef W, Wan Z-M (2003) A practical algorithm to infer
soil and foliage component temperatures from bi-angular ATSR-2 data. Int. J. Remote Sens.
24(23):4739–4760
Jia L (2004) Modeling heat exchanges at the land-atmosphere interface using multi-angular thermal
infrared measurements. Wageningen University, The Netherlands, ISBN 90-8504-041-8, 199pp
Jia L, Li Z-L, Menenti M (2005) Simulation of directional thermal infrared radiance images using
soil-vegetation-atmosphere heat and radiation transfer models. Proc. 9th Int. Symp. on Physical
Measurements and Signatures in Remote Sens. (ISPMSRS), ISSN 1682–1750 ISPRS, XXXVI
(7/W20):331–333
Jackson RD, Idso SB (1975) Surface albedo and desertification. Science 189:1012–1013
Jackson RD, Idso SB, Reginato RJ, Pinter PJ (1981) Canopy temperature as a crop water stress
indicator. Water Resour. Res. 17(4):1133–1138
Jackson RD (1985) Evaluating evapotranspiration at local and regional scales. Proc. IEEE
73:1086–1095
Kaufman YJ, Gao B-C (1992) Remote sensing of water vapor in the near IR from EOS/MODIS.
IEEE Trans. Geosci. Remote Sens. 30(5):871–884
Kimes DS (1980) Effects of vegetation canopy structure on remotely sensed canopy temperatures.
Remote Sens. Environ. 10:165–174
Kimes DS (1981) Remote sensing of temperature profiles in vegetation canopies using multiple
view angles and inversion techniques. IEEE Trans. Geosci. Remote Sens. 19(2):85–90
Kimes DS (1983) Remote sensing of row crop structure and component temperatures using direc-
tional radiometric temperatures and inversion techniques. Remote Sens. Environ. 13:33–55
90 M. Menenti et al.

Kimes DS, Smith JA, Link LE (1981) Thermal IR exitance model of a plant canopy. Appl. Opt.
20(4):623–632
Kimes DS, Kirchner JA (1982) Radiative transfer model for heterogeneous 3-D scenes. Appl. Opt.
21(22):4119–4129
Kimes DS, Kirchner JA (1983) Directional radiometric measurements of row-crop temperatures.
Int. J. Remote Sens. 4(2):299–311
Kleespies TJ, McMillin LM (1990) Retrieval of precipitable water from observations in the split
window over varying surface temperature. J. Appl. Meteorol. 29:851–862
Labed J, Stoll MP (1991) Spatial variability of land surface emissivity in the thermal infrared band:
spectral signature and effective surface temperature. Remote Sens. Environ. 38:1–17
Lagouarde JP, Kerr YH, Brunet Y (1995) An experimental study of angular effects on surface
temperature for various plant canopies and bare soils. Agric. Forest Meteorol. 77:167–190
Lei SA (2004) Leaf temperatures of Blackbrush: relationships with air and soil surface tempera-
tures. USDA Forest Service Proc. RMRS-P-31
Li Z-L, Jia L, Su Z, Wan Z, Zhang R-H (2003) A new approach for retrieving precipitable water
vapour from ATSR2 split-window channel data over land area. Int. J. Remote Sens. 24:5095–
5117
Li X-W, Strahler AH (1985) Geometric-optical modeling of a conifer forest canopy. IEEE Trans.
Geosci. Remote Sens. GE-23(5):705–721
Li Z-L, Stoll MP, Zhang RH, Jia L, Su Z (2001) On the separate retrieval of soil and vegetation
temperatures from ATSR2 data. Science in China, Series D 44(2):97–111
Li Z-L, Zhang R-H, Sun X-M, Su H-B, Tang X-Z, Zhu Z-L, Sobrino JA (2004) Experiment sys-
tem for the study of the directional thermal emission of natural surfaces. Int. J. Remote Sens.
25(1):195–204
Liu Q (2002) Study on Component Temperature Inversion Algorithm and the Scale Structure for
Remote Sensing Pixel. Ph.D. dissertation of Chinese Academy of Sciences, 111p
Liu QH, Li XW, Chen LF (2002). Field campaign for quantitative remote sensing. In: Beijing.
Proc. IEEE 2002 Int. Geosci. Remote Sens. Symp. (IGARSS’02) vol VI, pp 3133–3135
Lhomme J-P, Monteny B, Chehbouni A, Troufleau D (1994) Determination of sensible heat flux
over Sahelian fallow savannah using infra-red thermometry. Agric. Forest Meteorol. 68:93–105
Luquet D, Bégué A, Vidal A, Clouvel A, Dauzat J, Olioso A, Gu X-F, Tao Y (2003) Using multidi-
rectional thermography to characterize water status of cotton. Remote Sens. Environ. 84:411–
421
Mackay G, Steven MD (1998) An atmospheric correction procedure for the ATSR-2 visible and
near infrared land surface data. Int. J. Remote Sens. 19(15):2949–2968
Mahfouf JF, Richard E, Mascart P (1987) The influence of soil and vegetation on the development
of mesoscale circulations. J. Clim. Appl. Meteor. 26:1483–1495
Masuda K, Takashima T, Takayama Y (1988) Emissivity of pure and sea waters for the model sea
surface in the infrared window regions. Remote Sens. Environ. 313–329
Matteucci G, Dore S, Stivanello S, Rebmann C, Buchmann N (2000) Soil respiration in beech
and spruce forests in Europe: trends, controlling factors, annual budgets and implications for
the ecosystem carbon balance. In: ED Schulze (ed), Carbon and Nitrogen Cycling in European
Forest Ecosystems, vol 142. Springer, Berlin, pp 217–236
McGuire MJ, Smith JA, Balick LK, Hutchison BA (1989) Modeling directional thermal radiance
from a forest canopy. Remote Sens. Environ. 27:169–186
Medlyn BE, Dreyer E, Ellsworth DS, Forstreuter M, Harley PC, Kirshbaum MUF, Le Roux X,
Montpied P, Strassemeyer J, Walcroft A, Wang K, Loustau D (2002) Temperature response of
parameters of a biochemically-based model of photosynthesis. II. A review of experimental
data. Plant, Cell Environ. 25:1167–1179
Menenti M, Choudhury BJ (1993) Parameterization of land surface evapotranspiration using a
location-dependent potential evapotranspiration and surface temperature range. In: HJ Bolle, et
al. (eds), Exchange processes at the land surface for a range of space and time scales, IAHS,
Oxfordshire, vol 212, pp 561–568
4 Multi-angular Thermal Infrared Observations of Terrestrial Vegetation 91

Menenti M, Jia L, Li Z-L, Djepa V, Wang J, Stoll MP, Su ZB, Rast M (2001) Estimation of soil and
vegetation temperatures from directional thermal infrared observations: the HEIHE, SGP’97,
IMGRASS experiments. J. Geophys. Res. 106:11997–12010
Menenti M, Rast M, Baret F, Hurk B, Knorr W, Mauser W, Miller J, Moreno J, Schaepman M,
Verstraete M (2005) Understanding vegetation response to climate variability from space: re-
cent advances towards the SPECTRA Mission. Riv. Ital. Teleril. Special Issue Hyperspectral
Observation of Terrestrial Environments. 33/34:193–206
Monteith JL (1965) Evaporation and environment. In: GE Fogg (ed), The state and movement of
water in living organisms, Sympos. Soc. Exp. Biol. Academic, New York, 19:205–234
Moran MS, Clarke TR, Inoue Y, Vidal A (1994) Estimating crop water deficit using the relation be-
tween surface air temperature and spectral vegetation index. Remote Sens. Environ. 49(3):246
263
Mutlow CT, Závody AM, Barton IJ, Llewellyn-Jones DT (1994) Sea surface temperature measure-
ments by the along-track scanning radiometer on the ERS 1 satellite: early results. J. Geophys.
Res. 99(C11):22575–22588
Mutlow C, Murray J, Bailey P, Birks A, Smith D (1999) ATSR-1/2 User Guide, http://www.atsr.rl.
ac.uk/documentation/docs/userguide/atsr user guide rev 3.pdf
Nerry F, Labed J, Stoll M-P (1988) Emissivity signatures in the thermal IR band for remote sensing:
calibration procedure and method of measurement. Appl. Opt. 27(4):758–764
Nielsen DC, Clawson KL, Blad BL (1984) Effect of solar azimuth and infrared thermometer view
direction on measured soybean canopy temperature. Agron. J. 76:607–610
Nilson T (1971) A theoretical analysis of the frequency of gaps in plant stands. Agric. Meteorol.
8:25–38
Norman JM (1979) Modeling the complete crop canopy. In: BJ Barfield, JF Gerber (eds) Modifi-
cation of the Aerial Environment of Plants. ASAE Monogr. Am. Soc. Agric. Engr., St. Joseph,
MI, pp 249–277
Norman JM, Campbell G (1983) Application of a plant-environment model to problems in irriga-
tion. In: D Hillel (ed) Advanced in irrigation, vol 2. Academic, New York, pp 155–188
Norman JM, Chen JL (1990) Thermal emissivity and infrared temperatures dependence on plant
canopy architecture and view angle. In: Proc. Int. Geosci. Remote Sens. Symp. (IGARSS’90,
IEEE Geosci. Remote Sens. Soc., New York), pp 1747–1750
North PRJ, Briggs SA, Plummer SE, Settle JJ (1999) Retrieval of land surface bi-directional re-
flectance and aerosol opacity from ATSR-2 multiangle imagery. IEEE Trans. Geosci. Remote
Sens. 37(1):526–537
Otterman J (1990) Inferring parameters for canopies non uniform in azimuth by model inversion.
Remote Sens. Environ. 33:41–53
Otterman J, Brakke TW, Susskind J (1992) A model for inferring canopy and underlying soil
temperatures from multi-directional measurements. Bound.-Lay. Meteorol. 61:81–97
Otterman J, Susskind J, Brakke T, Kimes D, Pielke R, Lee TJ (1995) Inferring the thermal-infrared
hemispheric emission from a sparsely-vegetation surface by directional measurement. Bound.-
Lay. Meteorol. 74:163–180
Ottermann J, Brakke TW, Fuchs M, Lakshmi V, Cadeddu M (1999) Longwave emission from a
plant/soil surface as a function of the view direction: dependence on the canopy architecture.
Int. J. Remote Sens. 20(11):2195–2201
Ottlé C, Outalha S, Francois C, Le Maguer S (1997) Estimation of total atmospheric water vapor
content from split-window radiance measurements. Remote Sens. Environ. 61:410–418
Paw U KT, Ustin SL, Zhang CA (1989) Anisotropy of thermal infrared exitance in sunflower
canopies. Agric. Forest Meteorol. 48:45–58
Peng XD, Cheng LS (1993) Numerical simulations of the PBL mean structure and fluxed in HEIHE
region. Proc. Int. Symp. HEIFE. Kyoto University, Kyoto, Japan, pp 437–442
Petitcolin F, Nerry F, Stoll MP (2002) Mapping temperature independent spectral indice of emis-
sivity and directional emissivity in AVHRR channels 4 and 5. Int. J. Remote Sens. 23(17):3473–
3491
92 M. Menenti et al.

Pielke RA, Dalu G, Snook JS, et al. (1991) Nonlinear influence of mesoscale land use on weather
and climate. J. Climate 4:1053–1069
Prabhakara C, Chang HD, Chang ATC (1985) Remote sensing of precipitable water over the oceans
from NIMBUS 7 microwave measurements. J. Appl. Meteorol. 21:59–68
Price JC (1982) On the use of satellite data to infer surface fluxes at meteorological scales. J. Appl.
Meteorol. 21:1111–1122
Qualls RJ, Yates DN (2001) Directional radiometric temperature profiles within a grass canopy.
Adv. Water Resour. 24:145–155
Rahman H, Dedieu G (1994) SMAC: a simplified method for the atmospheric correction of satellite
measurements in the solar spectrum. Int. J. Remote Sens. 15:123–143
Schulz J, Schluessel P, Grassl H (1993) Water vapor in the atmospheric boundary layer over oceans
from SSM/I measurements. Int. J. Remote Sens. 14:2773–2789
Segal R, Avissar M, McCumber C, Pielke RA (1988) Evaluations of vegetation effects on the
generation and modification of mesoscale circulations. J. Atmos. Sci. 45:2268–2292
Seguin B, Assad E, Freteaud JP, Imbernon J, Kerr YH, Lagouarde JP (1989) Use of meteorological
satellite for water balance monitoring in Sahelian regions. Int. J. Remote Sens. 10:1001–1017
Sellers PJ, Heiser MD, Hall FG, Goetz SJ, Strebel DE, Verma SB, Desjardins RL, Schuepp PM,
McPherson JI (1995) Effects of spatial variability in topography, vegetation cover and soil
moisture on area-averaged surface fluxes: A case study using the FIFE 1989 data. J. Geophys.
Res. Vol. 100 (D12): 25607–25630
Sellers PJ, Randall DA, Collatz GJ, Berry JA, Field CB, Dazlich DA, Zhang C, Colleli GD,
Bounoua L (1996) A revised land surface parameterization (SIB2) for atmospheric GCMS.
Part 1: model formulation. J. Climate 9(4):676–705
Smith JA, Goltz SM (1995) A thermal exitance and energy balance model for forest canopies.
IEEE Trans. Geosci. Remote Sens. 32(5):1060–1066
Smith JA, Chauchan NS, Ballard J (1996) Remote sensing of land surface temperature: the direc-
tional viewing effect. IEEE Trans. Geosci. Remote Sens. 13(4):2146–2148
Snyder WC, Wan Z-M, Zhang Y, Feng Y-Z (1997) Thermal Infrared (3–14 µ m) bidirectional
reflectance measurements of sands and soils. Remote Sens. Environ. 60(1):101–109
Sobrino JA, Caselles V (1990) Thermal infrared model for interpreting the directional radiometric
temperature of a vegetative surface. Remote Sens. Environ. 33:193–199
Sobrino J, Li Z-L, Stoll MP, and Becker F, 1994. Improvements in the split-window technique
for land surface temperature determination. IEEE Trans. Geosc. and Rem. Sens. Vol. 32(2):
243–253
Sobrino J, Li Z-L, Stoll MP, Becker F (1994) Improvements in the split-window technique for land
surface temperature determination. IEEE Trans. Geosci. Remote Sens. 32(2):243–253
Sobrino JA, Raissouni N, Simarro J, Nerry F, Petitcolin F (1999) Atmospheric water vapor content
over land surfaces derived from AVHRR data: application to the Iberian Peninsula. IEEE Trans.
Geosci. Remote Sens. 37(3):1425–1434
Sobrino JC, Cuenca J (1999) Angular variation of thermal infrared emissivity for some natural
surfaces from experimental measurements. Appl. Opt. 38(18):3931–3936
Su H-B, Zhang R-H, Tang X-Z, Sun X-M, Zhu Z-L, Li Z-L (2003) An alternative method to
compute the component fractions in the geometrical optical model: visual computing method.
IEEE Trans. Geosci. Remote Sens. 41(3):719–724
Susskind J, Rosenfield J, Renter D, et al. (1984) Remote sensing of weather and climate parameters
from HIRS2/MSU on TIROS-N. J. Geophys. Res. 89:4677–4697
Sutherland RA, Bartholic JF (1977) Significance of vegetation in interpreting thermal radiation
from a terrestrial surface. J. Appl. Meteorol. 18:759–763
Valentini R, Dolman AJ, Ciais P, Schulze ED, Freibauer A, Schimel DS, Heimann M (2000) Ac-
counting for carbon sinks in the biosphere, European perspective. CARBOEUROPE European
Office. Max-Planck Institute for Biogeochemistry, P.O. Box 100164. 07700 Jena, Germany, 17p
Verhoef W, Menenti M (1998) Final Report of Spatial and Spectral Scales of Spaceborne Imaging
Spectro-radiometers (SASSSIS). Contract Report 12072/96/NL/CN, ESA. Final Report, NLR-
CR-98213, National Aerospace Lab. NLR, Amsterdam, The Netherlands, 314pp
4 Multi-angular Thermal Infrared Observations of Terrestrial Vegetation 93

Verhoef W (1998) Theory of radiative transfer models applied in optical remote sensing of vegeta-
tion canopies. Ph.D. thesis, Wageningen University, The Netherlands, 311pp
Verhoef W, Bach H (2003) Simulation of hyperspectral and directional radiance images using cou-
pled biophysical and atmospheric radiative transfer models. Remote Sens. Environ. 87:23–41
Verma SB, Rosenberg NJ, Blad BL, Baradas NW (1976) Resistance-energy balance method for
predicting evapotranspiration, determination of boundary layer resistance, and evaluation of
error effects. Agron. J. 68:776–782
Vermote EF, Tanre D, Deuze JL, Herman M, Morcrette JJ (1997) Second simulation of the satellite
signal in the solar spectrum, 6S: an overview. IEEE Trans. Geosci. Remote Sens. 35:675–686
Wang J, Mitsuta Y (1992) An observation study of turbulent study of turbulent structure and trans-
fer characteristics in Heihe oasis. J. Meteor. Soc. Jpn. 70:1147–1154
Wang J, Yan F, Xiao J, Wei T, Wang J (2002) Development of an airborne multi-angle TIR/VNIR
imaging system. IEEE 2002 Int. Geosci. Remote Sens. Symp. (IGARSS’02), vol IV, pp 2245–
2248
Weiss M, Baret F, Myneni R, Pragnère A, Knyazikhin Y (1999) Investigation of a model inversion
technique for the estimation of crop characteristics from spectral and directional reflectance
data. Agronomie 20:3–22
Welles JM, Norman JM (1991) Photon transport in discontinuous canopies: a weighted random
approach. In: RB Myneni, J Ross (eds), Photon-vegetation interactions. applications in optical
remote sensing and plant ecology. Springer, Heidelberg, Germany, pp 389–414
Wilson K, Goldstein A, Falge E, Aubinet M, Baldocchi D, Berbigier P, Bernhofer C, Ceulemans R,
Dolman H, Field C, Grelle A, Ibrom A, Law BE, Kowalski A, Meyers T, Moncrieff J,
Monson R, Oechel W, Tenhunen J, Valentini R,Verma S (2002) Energy balance closure at
FLUXNET sites. Agric. Forest Meteorol. 113(1–4):223–243
Wilson TB, Norman JM, Bland WL, Kucharik CJ (2003) Evaluation of the importance of La-
grangian canopy turbulence formulations in a soil–plant–atmosphere model. Agric. Forest Me-
teorol. 115:51–69
Xu X, Chen L, Zhuang J (2000) The passive measurements of object’s directional emissivity in
laboratory. Science in China 43:55–61
Xue Y, Shukla J (1993) The influence of land surface properties on Sahel climate. Part I: desertifi-
cation. J. Climate 6:2232–2245
Yan Y (1999) Numerical simulation of land surface process on heterogeneous surface. Ph.D. dis-
sertation. Lanzhou Institute of Plateau Atmospheric Physics. Chinese Academy of Sciences,
Lanzhou, China, 90pp
Yu T, Gu X-F, Tian G-L, Legrand M, Baret F, Hanocq J-F, Bosseno R, Zhang Y (2004) Modeling
directional brightness temperature over a maize canopy in row structure. IEEE Trans. Geosci.
Remote Sens. 42(10):2290–2304
Zhang R-H, Sun X-M, Su H-B, Zhu Z-L, Tang X-Z (2002) A new design for measuring direc-
tional radiant temperature and data analysis. Proc. IEEE 2002 Int. Geosci. Remote Sens. Symp.
(IGARSS’02), vol V, pp 2768–2770
Chapter 5
Terrestrial Applications of Multiangle Remote
Sensing

Mark J. Chopping

Abstract Multiangle remote sensing is an emerging technology that enables


important applications of terrestrial (land) remote sensing, in ecology and land
cover mapping as well as in a variety of disciplines in the Earth Sciences. Advances
have been realized in three major areas: the measurement or characterization of
canopy structure and surface roughness; the separation of the contributions of the
upper canopy and the background in forest and shrub-dominated environments; and
improvements in the accuracy of classifications of land cover in environments as
dissimilar as deserts and ice sheets. The focus of the chapter is on land surface ap-
plications of solar wavelength multiangular data acquired from the air and space; it
avoids discussion of methods for retrieval of essentially radiometric quantities such
as shortwave albedo, bidirectional and hemispheric reflectance factors (Chapter 9),
retrieval of land surface temperature (Chapter 4), theoretical radiative transfer
modeling studies, and those based uniquely on field measurements. The chapter
introduces existing instruments providing multiangle data; a review of work per-
formed under the broad headings Empirical and Synergistic Approaches, Radiative
Transfer, The RAMI Exercise, Canopy Openness, Clumping Index, Structural Scat-
tering Index, Geometric-Optical and Hybrid Models, Direction and Wavelength,
The Background in Canopy Reflectance Modeling, Land Cover and Community
Type Mapping, Snow and Ice, and Dust Emissions; and finally, brief discussions
under the headings Near-Simultaneous and Accumulated Sampling, Angular Sam-
pling, and Scale and Multiangle Observation. The emphasis is on the added value
that existing solar wavelength multiangular data can provide to applications. There
has been a wide array of approaches, all of which have resulted in studies demon-
strating progress and many of which show the advantages possible over the use
of purely nadir-spectral techniques, particularly for accessing measures of canopy
structure.

Mark J. Chopping
Department of Earth and Environmental Studies, Montclair State University, USA
chopping@pegasus.montclair.edu

S. Liang (ed.), Advances in Land Remote Sensing, 95–144. 95


c Springer Science + Business Media B.V., 2008
96 M.J. Chopping

5.1 Introduction

The exploitation of data obtained by observing the Earth’s surface at different


viewing and illumination angles in the solar wavelengths is a rapidly growing field.
An excellent summary of work in this field is given in Liang et al. (2000) with a
recent update by Diner et al. (2005). Multiangle remote sensing seeks to exploit
the reflectance anisotropy of the planet’s terrestrial surface that is described by the
bidirectional reflectance distribution function (BRDF) that quantifies the angular
distribution of spectral radiance that is scattered by a surface (Nicodemus et al.,
1977). The BRDF is thus a fundamental description of the terrestrial surface from
the photon’s point of view. A key concept is that multiangle measurement provides
a unique way to infer surface information within the observed element (“pixel”) and
is thus useful for mapping land surface parameters over large areas. Remote sensing
instruments acquiring imagery at varying angles explicitly are designed to collect
multiple looks over as short a timeframe as possible and – in the case of those flown
on satellites – taking into account the constraints imposed by the orbits of their
platforms. Multiangle data sets may also be constructed by accumulation of obser-
vations over periods from days to weeks, following the longstanding practice of
filtering images from many swaths to construct quasi-cloud-free composites. A lim-
iting factor of this method is the temporal displacement of the measurements, since
either the atmosphere or the surface, or both, may vary over short timescales (hours
to days). The BRDF gives the reflectance of a target as a function of illumination
geometry and viewing geometry; it depends on wavelength and is determined by
the structural and optical properties of the surface, such as shadow-casting, multiple
scattering, mutual shadowing, transmission, reflection, and absorption by surface el-
ements, facet orientation distribution and facet (e.g., leaf, soil particle) density. The
BRDF of a surface (unit: sr−1 ) is defined for all illumination and viewing directions
over the upper hemisphere and is not itself measurable as the light incident on the
surface is partly diffuse and the measurements involve finite solid angles (Liang,
2004; Bacour and Bréon, 2005). Researchers have sometimes ignored the cor-
rect nomenclature and taken multiangle observations to provide a (usually sparse)
sampling that is approximately bidirectional, accepting the assumption that the
size/distance ratio of the illuminating source and the observing sensor is zero. How-
ever there are now efforts under way to ensure that the terminology in use is more
accurate and consistent (Martonchik et al., 2000; Schaepman-Strub et al., 2006).
The radiometric data products available from multiangle instruments include
bidirectional reflectance factor (BRF; BRDF normalized by the equivalent re-
flectance of a Lambertian, or diffuse scattering surface) and/or hemispherical-
directional reflectance factor (HDRF; the single integral of BRDF on the incoming
directions; i.e., including direct and diffuse illumination), directional-hemispherical
reflectance (black-sky albedo; single integral of BRDF on the outgoing directions)
and bihemispherical reflectance (white-sky albedo; the double integral of BRDF
over all viewing and illumination directions). These quantities must be derived via
modeling since infinitesimal elements of solid angle do not include measurable
amounts of radiant flux, and small light sources and unlimited sensor fields-of-view
5 Terrestrial Applications of Multiangle Remote Sensing 97

do not exist; all measurable quantities of reflectance are therefore performed in the
conical or hemispherical domain (Schaepman-Strub et al., 2006). The most accurate
description of the spectro-directional radiance measurements made by remote sens-
ing instruments is therefore hemispherical-conical reflectance factor, even though
solar illumination is usually highly directional, especially under clear skies.
The most important physical phenomena underlying the BRDF are surface scat-
tering (geometric effects, shadowing and shadow-hiding; Camacho-de Coca et al.,
2004); and volume scattering by facets (leaves, soil particles). It is an intrinsic sur-
face property. Multiangle remote sensing seeks to exploit fundamental aspects of the
remote sensing problem and leads to important improvements in the accuracy with
which vegetation and other surface parameters can be retrieved. The information re-
turned is unique in that the data are sensitive to both the optical (spectral reflectance,
transmittance, and absorption) and structural properties of surfaces (e.g., canopy ar-
chitecture: horizontal and vertical heterogeneity), that is, from both the angular and
spectral domains (Diner et al., 1999, 2005). Since most multiangle work pursues
the one-dimensional (1-D) approach that seeks to derive “sub-pixel” information via
variation in radiance with solar and/or viewing angles and in which each “pixel” is
treated independently, the range of scales over which multiangle sensing can play a
role is large (the term “pixel” is used here since in remote sensing it is commonly un-
derstood to correspond to the area under the sensor ground-projected field-of-view,
even though this usage is incorrect: “pixel” is a contraction of “picture element”).
Much work has gone into developing algorithms for retrieving well-understood
physical parameters with straightforward interpretations via 1-D model inversions
(Chopping et al., 2003; Widlowski et al., 2004; Koetz et al., 2005a), although re-
searchers must grapple with the problem of non-uniqueness of solutions (different
combinations of model parameters or state variables can result in similar patterns
of observations), which can make this difficult (Weiss et al., 2000; Gobron et al.,
2000; Combal et al., 2002). Nevertheless, using simulations of top-of-atmosphere
radiance observations in 201 spectral bands and seven look angles, Verhoef (2005)
demonstrated that:
• Multi-angular hyperspectral measurements could in some cases provide dramat-
ically improved retrievals of canopy structure, leaf properties, soil moisture, and
radiative variables (e.g., fraction of photosynthetically-active radiation, albedo)
than monodirectional measurements.
• The information content of the spectral and angular domains is highly comple-
mentary, i.e., the additional information from multiangle remote sensing is not
redundant.
• For remotely estimating canopy architecture variables, multiangle information is
more important than hyperspectral information (Fig. 5.1).
Although the results above were obtained in the absence of real-world constraints,
it can be asserted from first principles that canopy leaf foliar chemistry and ecosys-
tem function are better assessed using hyperspectral methods, while canopy physi-
cal structure (in the sense of physiognomy, or architecture) is better assessed using
multiangle methods (Schaepman, 2006). These two approaches are not mutually
98 M.J. Chopping

30
Retrieval accuracy (dB)

Hyperspectral Monodir.
Hyperspectral Multidir.
Multispectral Multidir.
20 Multispectral Monodir.

10

Cd _b
Cs _g
Al R

Ca
Cw _b
Ca

N_
Cd _g
fC

Cs _b
N_

ze

fd e
fA ver

Di
LA

LI a

fB

Cv
SM

Cw _g

fb
LI

ho Fb
be

en
D

ar
DF

ss
PA
o

ta
_g

m
b

_b

t
b

g
m

%
b
%
do

s
Geo-biophysical variable

Fig. 5.1 Relative retrieval accuracies expressed in dB above the a priori uncertainty levels, for low
sensor noise (0.1 W/m2 /µm/sr) and four mission types. The parameters estimated were fractional
cover (fCover); fraction of absorbed PAR (fAPAR); albedo; soil moisture (SM%); green leaf chloro-
phyll (Cab g), water (Cw g), dry matter (Cdm g), senescence pigments (Cs g) and the N parameter
representing leaf mesophyll structure (CN g); brown leaf parameters (the same but identified with
the suffix b), canopy leaf area index (LAI), leaf inclination distribution function (LIDF in 2 pa-
rameters); a hot spot parameter (hot), the fraction of brown leaves (fB), dissociation factor (Diss),
crown cover fraction (Cv%), and tree shape factor (zeta); and the fractions of dense vegetation
and bare soil in the immediate environment of the observation (fDENS and fBAR), accounting for
adjacency effects. (Courtesy of Wout Verhoef).

exclusive but to date no studies have demonstrated the joint benefits of a spectro-
directional system with regular and broad geographic coverage; however, case stud-
ies over limited regions have demonstrated the considerable potential afforded by
exploiting both domains simultaneously (Garcı́a-Haro et al., 2006).

5.2 Major Multiangle Instruments

Under a restricted definition for Earth Observation, including only instruments that
provide near-simultaneous acquisition of >2 angular looks of the same area on
the Earth’s surface, and excluding the thermal or longer wavelengths, only three
such instruments have ever been placed in orbit: the NASA Multiangle Imaging
Spectro-Radiometer (MISR) flown on the NASA Earth Observing System Terra;
the French Space Agency’s (CNES) POLarization and Directionality of the Earth’s
Reflectance (POLDER), flown initially on the Japanese ADEOS satellite series and
currently on the CNES Parasol platform; and Sira (UK)’s Compact High Resolu-
tion Imaging Spectrometer (CHRIS), flown on the European Space Agency (ESA)
Proba-1 satellite. Terra was launched in December 1999, Proba in 2001, and Parasol
in 2004. Unprecedented efforts have also been made in the last 15 years properly
to treat data from across-track scanning instruments such as the NOAA Advanced
Very High Resolution Radiometer (AVHRR) and the MODerate resolution Imaging
Spectroradiometer (MODIS) flying on NASA’s Terra and Aqua EOS satellites as
5 Terrestrial Applications of Multiangle Remote Sensing 99

intrinsically dependent on surface BRDF. Although for the latter instruments multi-
angular data sets are acquired sequentially rather than near-simultaneously, work
with these accumulating, or “sequential” systems is also making important contri-
butions to multiangle remote sensing. For this reason, reference is made here to
MODIS as a “multiangle” instrument, even though it was not designed to acquire
multi-angle imagery and the variation in viewing and illumination angles is an ac-
cident of the across-track acquisition method. Indeed, under this paradigm many
other moderate resolution sensor systems, including those on geostationary plat-
forms, could also be deemed to be “multiangle”. It is preferable to think of these
systems in this way rather than to insist on the narrower definition because angular
artifacts in data from these instruments might otherwise be overlooked.
This chapter will concentrate on applications using data from orbiting instru-
ments providing more than two views; this excludes Japan Aerospace Exploration
Agency’s ASTER on Terra, which exploits dual views via a “geometric strategy”
rather than the “radiometric strategy” necessary with data from sensors with a
larger instantaneous field-of-view (IFOV) and the Advanced Along-track Scanning
Radiometer (AATSR), an important European instrument whose dual views have
proven valuable for aerosol characterization (King et al., 1999). The geometric strat-
egy involves real or apparent differences in the shape or location of observed objects,
resulting from changes in perspective (for targets with 3-D structure), stereoscopic
parallax (displacement dependent upon distance from the observer), or actual mo-
tion of the target during the time interval between views; the radiometric approach
refers to changes in the brightness, color, contrast, or other radiance-related quanti-
ties as a function of view and/or illumination angles (Diner et al., 2005, Table 5.1).

5.2.1 MISR on Terra

The Multi-angle Imaging SpectroRadiometer (MISR) was designed and built at


NASA’s Jet Propulsion Laboratory (JPL), in Pasadena, California and flies on
NASA’s Terra satellite. MISR measures Earth’s spectral radiance in four spectral
bands (blue, green, red and near-infrared), at each of nine look angles in the forward
and aft directions along the flight path (Fig. 5.2). One camera views at nadir (close
to 0◦ view zenith angle), while the others view at ±26.1◦ , ±45.6◦ , ±60.0◦ , and
±70.5◦ (Fig. 5.3). Spatial samples are acquired every 275 m. In the global mode, all
nine cameras report the red band and the nadir camera reports all other bands in at
this resolution; for the off-nadir cameras the other bands are downsampled to 1.1 km
via contiguous averaging of four cross-track by four along-track line samples within
the instrument before data transmission. Over a period of 7 min, a ∼400 km wide
swath of Earth comes into view at all nine angles. Global coverage is achieved every
9 days, with repeat coverage between 2 and 9 days depending on latitude (the repeat
time is shorter at higher latitudes). In local mode, all four spectral bands are retained
at the acquisition resolution of 275 m in all nine cameras. Special attention has
Table 5.1 Major satellite instruments with multiangle capability (Adapted from Diner et al., 2005.)
100

MISR (NASA) (CNES) POLDER (ESA) CHRIS/Proba MODIS (NASA) CERES (NASA) AATSR (ESA)

Description Multiangle pushbroom Wide Field-of-View Tiltable Across-track Scanners Circular scanner
imager pushbroom radiometer pushbroom imaging scanner
spectrometer
Launch Date December 1999, on EOS August 1996, on ADEOS October 2001, on December 1999, December 1999, March 1, 2002, on
Terra 1; December 02, on Proba-1 on EOS Terra on EOS Terra ENVISAT
ADEOS II; December 04,
on Parasol
Angular coverage 0◦ , 26◦ , 46◦ , 60◦ , 70◦ Comprehensive (over Five looks in the 0–55◦ Limb to limb Two looks, 55◦
forward and backward of several days) along-track continuously forward and close
nadir in along-track direction to nadir in the
direction along-track
direction.
Multiangle strategies Radiometric, geometric Radiometric Radiometric, Radiometric Radiometric Radiometric
employed geometric
Spectral coverage 446, 558, 672, 866 nm 444.5, 444.9, 492.2, Programmable up 36 bands: Shortwave: 0.555, 0.659,
564.5, 670.2, 763.3, to 63 bands, 0.4–14.5 µm 0.3–5 µm Thermal 0.865, and 1.61
763.1, 907.7, 860.8 nm 0.4–1.0 µm window: 8–12 µm 0.555, 0.659,
Bandwidth: Total: 0.3–200 µm 0.865, and
<11 nm 1.61 µm (VIS-IR);
3.7, 10.8, and
123.7, 10.8, and
12 µm (TIR)
Spatial resolution 275 m or 1.1 km, 6 × 7 km nominal, 6 × 6 18–34 m 250 m − 1 km at 20 km at nadir 1 km at nadir
depending on channel after processing, depending on the nadir, depending
2,400 km swath node selected, on channel
13 km swath
Scene dimensions Pole-to-pole × 400 km Pole-to-pole 13 × 13 km Pole-to-pole Pole-to-pole × Pole-to-pole
2, 400 × 1, 800 km approximately × 2, 300 km limb-to-limb × 512 km
Global coverage in 9 days 1 day NA 2 days 1 day 6 days
M.J. Chopping
5 Terrestrial Applications of Multiangle Remote Sensing 101

Fig. 5.2 Computer-generated image showing the Terra spacecraft and MISR instrument operation.
The direction of flight is toward the lower left. The locations imaged by the 9 MISR cameras, each
with 4 spectral bands, are illustrated here with translucent surfaces (Courtesy of Shigeru Suzuki
and Eric M. De Jong, Solar System Visualization Project. NASA/JPL-Caltech, image P-49081.)

Fig. 5.3 MISR images of Zambia and Botswana, Africa, acquired on August 25, 2000. The left
image is a “true” color composite from the red, green, and blue vertical-viewing (nadir) camera.
The middle image combines data from the green, red, and near-infrared bands. The right image
contains red band data only but as a composite of imagery from the nadir, 70.5◦ forward, and
70.5◦ aftward cameras. The color variations in the multi-angle composite arise because light is
reflected at different angles. (Courtesy of the MISR Science Team.)

been paid to providing highly accurate absolute calibration, using on-board hard-
ware consisting of deployable solar diffuser plates and several types of photodiodes
(Diner et al., 1998; Bruegge et al., 2002). The land surface reflectance products cur-
rently generated at the NASA Langley Atmospheric Sciences Data Center include
the spectral hemispherical–directional reflectance factors (HDRF) at the nine MISR
view angles and the associated bihemispherical reflectances (BHR). The HDRF and
the BHR characterize the surface reflectance under direct and diffuse illumination.
102 M.J. Chopping

An algorithm for the generation of vegetation green leaf area index (LAI) and the
fraction of photosynthetically active radiation absorbed by vegetation (fPAR) from
MISR BHR and bidirectional reflectance factor (BRF) was implemented for oper-
ational processing (Knyazikhin et al., 1998c). Unlike single-angle approaches, the
MISR LAI/fPAR algorithm does not rely on prescribed vegetation type maps as it
uses the multiangular information to constrain the retrievals.

5.2.2 MODIS on Terra

The MODerate resolution Imaging Spectroradiometer (MODIS) instrument flying


on NASA’s Earth Observing System (EOS) Terra and Aqua satellites provides 12
bit imagery in 36 spectral bands ranging in wavelength from 0.4 to 14.4 µm. Two
bands (red and NIR) are imaged at a nominal resolution of 250 m at nadir, with five
bands at 500 m, and the remaining 29 bands at 1 km. MODIS is an across-track scan-
ning radiometer providing observation along a scanline that observes the surface at
zenith angles of ±55◦ . At the EOS orbit of 705 km, it achieves a 2,330 km swath and
provides global coverage every 1–2 days. MODIS-AM was placed into orbit with
the launch of Terra on December 18, 1999; the second instrument, MODIS-PM, is
flying on the Aqua spacecraft launched on May 4, 2002. Thanks to the very wide
swath, MODIS provides multiangular measurements of spectral radiance from the
same location on the surface only through the accumulation of looks over a period
of several days (sometimes called the “sequential” method); it is not capable of near-
simultaneous multiangle observation. While global coverage is provided in a period
of 48 h or less, cloud cover means than data from several overpasses are almost al-
ways required in order to map clear-to-surface surface-leaving spectral radiances.
The across-track scanning method provides angular observations closer to the so-
lar principal plane (the plane intersected by the sensor, target, and the Sun) than
along-track instruments at mid-latitudes, although this is highly dependent on lati-
tude. The major disadvantage of this method is that the surface may change during
the accumulation period; however the variation in the signal owing to accumulated
sampling is generally far lower than that owing to the BRDF. Since MODIS data are
highly dependent on viewing and illumination geometry and surface BRDF, major
efforts have been expended to properly account for BRDF effects, and to exploit the
directional signal to retrieve global land albedo (see Chapter 9).

5.2.3 CHRIS on Proba-1

The CHRIS sensor was developed by Sira Electro-Optics (UK). Since April 2006
Sira space and imaging operations have been a part of Surrey Satellite Technol-
ogy Ltd., Guildford, UK, a spin-off company of the University of Surrey. CHRIS
produces imagery in up to 62 spectral channels in the range 415–1,050 nm with a
5 Terrestrial Applications of Multiangle Remote Sensing 103

Fig. 5.4 (a) Artist’s impression of the Proba satellite (b) CHRIS instrument multi-angle acquisition
via tilting of the platform (Courtesy of ESA and Mike Barnsley, University of Wales at Swansea.)

spectral resolution of 5–12 nm. It is remotely configurable in terms of both spectral


channels and spatial resolution (European Space Agency, 1999), with a nominal
ground resolution of between 17 and 34 m. The nominal swath is 13 km. CHRIS
was launched on the Proba-1 satellite (Verhaert, Belgium) on October 21, 2001
(Fig. 5.4). Proba-1 is a small platform, weighing approximately 100 kg and measur-
ing approximately 60 × 60 × 80 cm. Thanks to its four reaction wheels the platform
is highly maneuverable: along-track pointing allows a given site to be imaged five
times during a single overpass, while across-track pointing ensures that the revisit
time for a site of interest is potentially less than a week (Barnsley et al., 2004). Proba
was not placed into its intended orbit because of an anomaly with the final stage of
the Indian Polar Satellite Launch Vehicle (PSLV), resulting in a much lower and
more elliptical orbit than intended. This caused a number of operational problems
that necessitated on-ground and on-board software changes and delayed the start of
the science program. It took more than 12 months to rewrite the software required
for precise orbital navigation, so that the first useable overlapping hyperspectral im-
agery was available to most users only in early 2003 (albeit often with four partially
overlapping looks, rather than the five intended).
Since then, further improvements have been made so that five overlapping im-
ages are generally available and the overlap area is larger. Both the satellite and the
CHRIS instrument continue to function well at the time of writing (April 2006),
although striping is evident in some images owing to detector anomalies; however,
third-party algorithms have been written to correct for this. The CHRIS/Proba sys-
tem has several unique features: it employs tilting of the platform to gain a series
of angular views in the along-track direction; the platform nods in order to acquire
images; and the CHRIS sensor is programmable for spatial resolution and spec-
tral coverage. There are well over 60 groups around the world engaged in research
with CHRIS in categories including calibration/validation, land and environment,
hydrology, atmosphere, agriculture, forestry, hazards, renewable resources, coastal
zones, topographic mapping, and geology (European Space Agency, 2006). It is
104 M.J. Chopping

worth noting that not all projects seek to exploit the multiangle feature of the instru-
ment; many projects are focusing on the exploitation of the hyperspectral capability.

5.2.4 POLDER on ADEOS / Parasol

The Polarization and Directionality of the Earth’s Reflectance (POLDER) in-


strument developed by France was initially launched on the Japanese ADEOS-1
(ADvanced Earth Observation Satellite) platform in August 1996, with the satellite
failing in June 1997 (a solar panel failure terminated operations after 8 months of
continuous acquisition); a second identical instrument, was launched on ADEOS-2
in December 2002, with that satellite failing in October 2003 (another solar panel
failure terminated operations with 7 months of continuous acquisition). This ver-
sion of POLDER provided a swath of ∼2, 400 km with a footprint of 6 × 7 km2 . The
instrument was a two-dimensional CCD detector array with wide field of view tele-
centric optics and a rotating wheel carrying spectral and polarization filters. There
were 15 filters and polarizers in the visible and the near infrared range providing
nine spectral bands, three of which (443, 670 and 865 nm) were associated with
polarized filters. Three bands were dedicated to the observation of ocean color (443,
490, 565 nm) but POLDER-1 was designed to provide dedicated measurements
of clouds and atmospheric aerosols. Its applications in terrestrial remote sensing
have therefore taken a secondary role, although they are not insignificant (Maignan
et al., 2004; Bréon et al., 2002; Camacho-de Coca et al., 2002). While POLDER-1
acquired spectral radiance data at multiple angles for the same location on a near-
simultaneous basis in the along-track direction, its full angular sampling potential
was only realized after a number of successive orbits has been accomplished, typ-
ically within 15 days. This means that POLDER-1 was a hybrid system and has
some features in common with accumulating instruments.
The third POLDER instrument – hereafter called POLDER-2 as it differs from
the first version, POLDER-1 – is being flown on the CNES Parasol platform, which
was launched from Europe’s spaceport in Kourou, French Guiana in December 2004
(Fig. 5.5). Parasol is a micro-satellite that flies in formation with the NASA EOS
“A Train”. The POLDER-2 instrument has provided near-continuous acquisition
since March 2005. It observes in eight spectral bands from 440 (blue) to 910 nm
(NIR) using a bi-dimensional CCD matrix providing a spatial resolution of ∼6 km;
three of the bands are polarized: 440, 670, 865 nm. As with the POLDER instru-
ment on the ADEOS-2 platform, the Parasol-borne POLDER-2 provides its most
dense angular sampling via accumulation over a period of a few days. Over succes-
sive days, the instrument is at different places in the sky, which provides another
set of angular measurements: after a few days, the directional space within 60◦
of view zenith angles is fully covered with a good sampling of the principal and
perpendicular planes, including the hot spot around which the solar and viewing
directions coincide.
5 Terrestrial Applications of Multiangle Remote Sensing 105

(a) (c)
50
45
40

[%]
35
30
Ref_443
25 Ref_565
Ref_670
20
−80 −60 −40 −20 0 20
Phase Angle

(b) (d)
10

8
6
[%]

4
2 RefPol_443
0 RefPol_370
RefPol_865
−2
−80 −60 −40 −20 0 20
Phase Angle

Fig. 5.5 POLDER-2 images over desert target in the Sahara (a) 3 color composites in total light
443–670–865 nm (b) 3 color composites in polarized light 443–670–865 nm (c) directional signa-
tures in total light (d) directional signatures in polarized light (Courtesy of F-M Bréon, Laboratoire
des Sciences du Climat et de l’Environnement.)

5.3 Multiangle Applications: Vegetation Canopies

Satellite remote sensing would ideally be able to provide information on both the
structure and function of ecosystems and the types of plant communities from which
they are formed, at regular intervals, and for the entire terrestrial surface. While re-
mote sensing of vegetation function is an established fundamental and very impor-
tant application of remote sensing – spectral measures have proved very effective in
mapping vegetation productivity – less has been achieved with respect to remotely
measuring vegetation structure. The optical properties of vegetation have allowed
the exploitation of reflected light in the solar spectral wavelengths via spectral vege-
tation indices, spectral unmixing techniques and in the temporal domain, phenology
metrics, that have proved useful in obtaining global maps of vegetation parameters,
such as the MODIS Vegetation Continuous Fields (VCF) % Tree Cover product
(Hansen et al., 2002). In spite of these successes, it is difficult to access canopy
structure information using single-angle sensing since structural surface properties
are largely confounded in nadir-spectral measures.
106 M.J. Chopping

Canopy structure is defined as the vertical and spatial distribution, orientation


and density of foliage and its supporting structures. Vegetation structure determines
the pattern of light attenuation and the distribution of photosynthesis, respiration,
transpiration, and nutrient cycling in the canopy (Widlowski et al., 2004). Net pri-
mary production of woody vegetation is always dependent on vegetation structure,
since the maximum rate of carbon uptake is constrained by total leaf area, and the
rate of carbon loss is constrained by total woody biomass (Widlowski et al., 2004).
Measures of canopy structure are required in order to estimate biomass, to assess
geographically shifting abundances of woody plant material (shrubs and trees), to
estimate carbon emissions to the atmosphere from wildfires, and to examine where
forest regrowth and secondary succession are occurring, and at what rates (Liu and
Kafatos, 2007). Canopy architecture and spatial heterogeneity are important not
only in forests but also affect the available carbon storage, vegetation dynamics and
productivity of grasslands undergoing woody encroachment, shrublands, and savan-
nas: different degrees of structural heterogeneity lead to a wide range of estimates
for these environments (Asner, 2000; Chopping et al., 2003). It is also necessary for
mapping fire fuel loads; and for estimating surface roughness, with the implications
this has for fluxes between the surface and the atmosphere.
The surface BRDF is dependent on the structural as well as optical characteris-
tics of surfaces and it should thus be possible to exploit angular measurements to
retrieve structural canopy parameters. Although this is not straightforward, in the
last few years multiangle reflectance data from satellites have been used to develop
a variety of approaches (Pinty et al., 2002; Bacour et al., 2002; Gao et al., 2003;
Hu et al., 2003; Widlowski et al., 2004; Jenkins et al., 2004; Nolin, 2004; Koetz
et al., 2005b; Chen et al., 2005; Strahler et al., 2005; Diner et al., 2005; Kimes et al.,
2006; Laurent et al., 2005; Bach et al., 2005; Chopping et al., 2006c; Disney et al.,
2006; Garcı́a-Haro et al., 2006; Heiskanen, 2006; Chopping et al., 2007). The
approaches taken are very diverse and include 3-D radiative transfer modeling,
geometric-optical canopy reflectance modeling, exploitation of the structural
information available in the parameters returned via inversion of physical and
semi-empirical models, purely empirical methods, and heuristic and data mining
algorithms. All have demonstrated some degree of success.

5.3.1 Empirical and Synergistic Approaches

For some years Earth observation scientists have been considering the range and
importance of synergies that may be realized between multiangle instruments and
other types of sensor. The focus has been on synthetic aperture radar (SAR) and –
more recently – lidar, both active systems, although good synergies with high res-
olution stereoscopic imagers can also be found for some applications (e.g., cloud
top heights). Disney et al. (2006) demonstrated that a combined structural and ra-
diometric modeling approach provides a flexible and powerful method for simu-
lating the remotely sensed signal of a forest canopy in the solar and microwave
5 Terrestrial Applications of Multiangle Remote Sensing 107

domains, useful for exploring the impact of canopy structure on the resulting sig-
nal and for combined retrievals of forest structural parameters. They developed a
detailed 3-D structural model of a conifer forest canopy in order to simulate the
observed reflectance (optical) and backscatter (microwave) signals and showed that
it is feasible to model forest canopy scattering using detailed 3-D models of tree
structure (including the location and orientation of individual needles). A structural
growth model of Scots pine was modified to simulate observed growth stages of
a Scots pine canopy from age 5 to 50 years. Model canopies were generated by
adding needles to tree structural models and these were used to drive optical and
microwave models of canopy scattering. Simulated canopy radiometric response
was compared with airborne hyperspectral reflectance data from the HyMAP in-
strument (Integrated Spectronics, Australia) and airborne synthetic aperture radar
backscatter data. Model simulations agreed well with observations, particularly at
solar wavelengths. The choice of needle shape and phyllotaxy was shown to have
a significant impact on multiple scattering behavior at the branch scale. In the mi-
crowave domain, simulated backscatter values agreed reasonably well with obser-
vations at L-band and less so at X-band. L-band simulated backscatter significantly
underestimated observed backscatter at younger canopy ages, probably as a result
of inappropriate modeling of the soil/understory background.
Successful and convincing empirical applications of multiangle reflectance data
have been rare until recently, when the exploration of synergies with lidar data be-
came feasible. Researchers at the NASA Goddard Space Flight Center, Greenbelt,
Maryland, performed work in which forest structural data from the airborne Laser
Vegetation Imaging Sensor (LVIS) were used with multivariate regression and
neural networks to allow the estimation of forest vertical structure from multiangle
reflectance data from the Airborne Multi-angle Imaging SpectroRadiometer (Air-
MISR) (Kimes et al., 2006). The AirMISR data were acquired on 28 August 2003
over a 7,000 ha experimental forest near Howland, Maine, consisting of small plan-
tations, multi-generation clearings and large natural forest stands (predominantly
boreal-northern hardwood transitional forest with spruce-hemlock-fir, aspen-birch,
and hemlock-hardwood associations). Twenty-eight AirMISR multiangle spectral
radiance values were selected from the 36 possible combinations (nine angles and
four spectral bands). Data acquired at nadir viewing and at ±26.1◦ , ±45.6◦ , and
±60.0◦ in the flying direction were used and resampled into 15 × 15 m pixels. The
LVIS data were acquired 3 days later, with a total of ∼160, 000 LVIS shots collected
in the study area. The LVIS data set provided a relatively direct measure of forest
vertical structure at a fine scale (20 m diameter footprints). For each laser shot, the
following were recorded: location (long, lat), surface height (m), and the heights of
the 25%, 50%, 75% and 100% of waveform energy data (m). These data were grid-
ded into 15×15m pixels using inverse-distance weighted average of the four nearest
shots. The best model accurately predicted the maximum canopy height as measured
by LVIS using AirMISR data with a root mean square error (RMSE) of 0.92 m and a
R2 = 0.89 (Fig. 5.6). This result is all the more remarkable because these high accu-
racies were achieved over a study site with an elaborate patchwork of forest commu-
nities with exceptional diversity in forest structure. The study concluded that models
108 M.J. Chopping

25
a
Predicated LVIS H100 Values (m)
20

15

10

5
5 10 15 20 25
True LVIS H100 Values (m)
25
b
Predicted Undisturbed Stand
True/Predicted Forest Height (m)

True Undisturbed Stand


20
Predicted Selective Cut
True Selective Cut

15

10

0
0 25 50 75 100
LVIS Energy Quartile Heights (%)

Fig. 5.6 (a) LVIS forest canopy heights predicted by AirMISR versus heights from LVIS for
17 × 17 pixel windows. The model was applied to 1,000 random test points. (b) Predicted versus
true LVIS energy for canopy height (H100) and quartile heights (H75, H50, H25) for the logged
and unlogged areas. The predictive models used only AirMISR data (Reprinted from Kimes et al.,
2006. Copyright, Elsevier. With permission.)

using multiangular data are capable of accurately predicting the vertical structure of
forest canopies. Further work using data from the Geoscience Laser Altimeter Sys-
tem (GLAS) on the Ice, Cloud, and Land Elevation Satellite (IceSAT) platform to
train MISR-derived estimates of canopy and extend them continuously over large
areas has provided encouraging but somewhat uneven results. This is most likely
to be at least partly owing to the azimuthal orientation of MISR data away from
the principal plane – that sometimes can be advantageous, e.g., for the retrieval of
albedo and fPAR (B. Pinty, personal communication) – since the AirMISR data were
acquired close to this plane (K.J. Ranson, personal communication).
5 Terrestrial Applications of Multiangle Remote Sensing 109

Fig. 5.7 Predicting stand basal area with AirMISR (non-linear multiple regression) using an arti-
ficial neural network with 5 nodes. Training R2 = 0.68, RMS Error = 8 m2 /ha (Courtesy of Rob
Braswell and Julian Jenkins, Institute for the Study of Earth, Oceans and Space (EOS), University
of New Hampshire.)

Another study using independent field biometric measurements to train multian-


gle data was carried out over the Bartlett Experimental Forest in New Hampshire.
An artificial neural network with five nodes was used to map basal stand area over
an area of 19.1 km2 . A high correlation was obtained between optical multiangle
reflectance signature measured with AirMISR and field observations of stand basal
area, with an RMSE of 8 m2 /ha (Jenkins et al., 2004; Fig. 5.7).
These studies show that multiangle data can be used to leverage lidar or sur-
face assets to provide wall-to-wall coverage using purely empirical methods, rather
than explicit modeling of light scattering. Further advances might be expected using
lidar synergistically with multiangle data in modeling approaches as well as in an
empirical training mode (Hese et al., 2005; see Geometric-Optical and Hybrid Mod-
els, below).

5.3.2 Radiative Transfer

Three-dimensional radiative transfer (RT) modeling is used as part of MISR surface


product retrievals. A look-up table (LUT) is used to rapidly determine the matching
modeled reflectances and the associated values of leaf area index (LAI) and fraction
110 M.J. Chopping

of photosynthetically active radiation (fPAR): all necessary radiative transfer para-


meters are pre-computed and stored in a Canopy Architecture Radiative Transfer
(CART) file. The MISR LAI retrieval algorithm does not depend on a particular
canopy radiation model as the elements of the CART are components of various
forms of the energy conservation law. The MISR leaf area index (LAI) product has
been shown to be accurate to within 0.5 LAI in herbaceous vegetation and savannas
and is an overestimate by about 1.0 in broadleaf forests (Diner et al., 2005). When
using single angle data, uncertainty in biome classification can lead to large errors.
However, the use of multiangle data minimizes the impact of biome misidentifi-
cation on LAI retrievals; that is, despite errors in biome classification, the use of
multiangle information leads to a similar error rate – even when the biome is mis-
classified – as is obtained with single angle approaches and a prescribed biome map.
The multiangle method therefore permits LAI retrieval without the use of a pre-
scribed biome map. In a case study in Africa, 80% of the LAI values were retrieved
using an incorrect biome type – but with a probability of about 70%, uncertainties
in LAI retrievals due to biome misclassification did not exceed uncertainties in the
observations (Hu et al., 2003).
A parameter closely related to LAI, the canopy “recollision probability” (defined
as the probability with which a photon scattered in the canopy interacts with a phy-
toelement again) has been shown to describe canopy spectral absorption and scatter-
ing; this is a new and developing area of multiangle imaging research. Knyazikhin
et al. (1998a, b) proposed that – to a good approximation – the amount of radi-
ation absorbed by a canopy should depend only on the wavelength and a canopy
structural parameter (p), which is wavelength independent. Knowing the recolli-
sion probability value of a canopy, the scattering coefficient of the canopy at any
wavelength can be predicted from the leaf scattering coefficient at the same wave-
length. Knyazikhin et al. (1998a, b) also introduced a similar parameter (pt ) relating
canopy transmittances at two different wavelengths to the leaf scattering coefficients
at these wavelengths. Given the absorption (p value) and transmission (pt value), to-
tal reflectance (the upward scattered part of the incident radiation) is also known. It
is recognized that LAI may not depend on only these parameters alone but may also
vary with leaf orientation and the degree of foliage clumping. A recent study by
Smolander and Stenberg (2005) confirmed that the spectral absorption and scatter-
ing of structurally simple uniform canopies can be well described by the canopy p
value, which furthermore showed close relationship with the LAI but insensitivity
to solar zenith angle.

5.3.3 The Radiation Transfer Model Intercomparison (RAMI)


Exercise

Radiative transfer model accuracy and robustness are key issues for improving
remotely sensed surface variables that may depend heavily on the angular distrib-
ution of samples. The Radiation Transfer Model Intercomparison (RAMI) exercise
5 Terrestrial Applications of Multiangle Remote Sensing 111

(Pinty et al., 2001, 2004a) has been ongoing since 1999 with the most recent and
third phase completed in 2005. RAMI provides a framework within which spec-
tral bidirectional reflectance models designed to simulate the transfer of radiation
within plant canopies and over bare soil surfaces can be benchmarked. This is use-
ful because the interpretation of remote sensing data generated by Earth Observing
satellites hinges on the exploitation of such models: it is clearly very important that
they are reliable, accurate, and fit for purpose. RAMI approaches this endeavor by
providing a series of test cases of different degrees of complexity, broadly divided
into homogeneous and heterogeneous cases (Fig. 5.8).
Participants perform simulations based on these scenarios and deliver model-
ing results to the Institute for Environment and Sustainability at the European
Commission Joint Research Center in Ispra, Italy, that coordinates the exercise.
The subsequent publication of the findings provides an objective way to evaluate
the performance and limits of applicability of the models. In addition, new models
may be developed and existing ones improved by comparing their output with the
available simulation results on the RAMI web pages (http://rami-benchmark.jrc.it/).
The overall benefit to the remote sensing community is a demonstration of model
maturity and a better understanding of how the models may be used in the inter-
pretation of remote sensing data. The latest results, that include 5 1-D models and

Fig. 5.8 Selected scenarios in the Phase 3 RAMI set of radiative transfer modeling exercises: (a)
homogeneous turbid cases in the solar domain: leaves treated as a turbid medium (b) homogeneous
discrete cases in the solar domain: leaves represented as a large number of non overlapping disc-
shaped objects (c) heterogeneous turbid cases in the solar domain: large number of non overlapping
spherical objects with crowns represented as a turbid medium (d) discrete floating spheres in the
solar domain: large number of non overlapping spherical objects representing the individual plant
crowns that contain randomly distributed finite size disc-shaped scatterers, with the orientation of
the scatterers following a uniform distribution function (e) Heterogeneous Scene-Based Experi-
ments: large number of disc-shaped scatterers contained within a series of non-overlapping spher-
ical and cylindrical volumes representing identically sized plant crowns (f) Boreal Birch Stand
Scene: large number of randomly located non-overlapping tree-like entities of differing sizes with
individual objects represented by ellipsoidal crowns with randomly distributed finite sized foliage.
See: http://rami-benchmark.jrc.it/HTML/RAMI3/RAMI3.php
112 M.J. Chopping

13 3-D models, are reported in Widlowski et al. (2006). It should be noted that
model accuracy is an important but not the only limiting factor in the application
of reflectance models in retrieving surface information useful to applications using
multiangle remote sensing data: inversion algorithms (Chapters 6–8), model para-
digm and suitability for purpose, noise and contamination in data sets, and angular
and spectral sampling also play important roles.

5.3.4 Canopy Openness

One approach to estimating canopy openness makes use of parametric fits of red
wavelength MISR or AirMISR data to the Rahman-Pinty-Verstraete (RPV) model
(Rahman et al., 1993) or the modified version of this (MRPV, Engelsen et al.,
1996). The MRPV model parameters from inversion against MISR data are included
in the operational MISR Land product at level 2. The RPV and MRPV models
are based on a consideration of the main aspects of BRDF shapes and are three-
parameter models. The first parameter is a factor describing the overall amplitude
of reflectance; the second describes the steepness of the BRDF bowl or bell shape;
and the third controls the relative importance of back- and forward-scattering via
a Henyey-Greenstein function (for the RPV model) or an exponential function in
scattering angle (for the MRPV model). The equations for the MRPV model are
given in Eq. (5.1):

BRFRPV = ρs (θs , θv , φ ; ρ0 , b, k) = ρ0 M F H (5.1)

where:

M = [cos θs cos θv (cos θs + cos θv )]k−1 (5.2)


F = exp(−b cos(g)) (5.3)
H = 1 + [(1 − ρ )/(1 + G)] (5.4)

where ρ is the average measured reflectance factor and

G = [tan2 θs + tan2 θv , −2 tan θs tan θv cos φ ]1/2 (5.5)


cos g = cos θs cos θv + sin θs sin θv cos φ (5.6)

The first parameter, ρ0 , is a factor describing the overall amplitude of reflectance;


the second, b, controls the relative importance of back- and forward-scattering via a
Henyey-Greenstein function in the original RPV formulation, replaced by an expo-
nential term in the MRPV model Eq. (5.2) and in the MISR BRDF/albedo product.
The rationale for this modification is that except for the hotspot term the coefficients
of the MRPV model can be rapidly solved by linear least-squares by taking loga-
rithms of the expression; but note that a recently-developed algorithm allows effi-
cient, reliable and accurate inversion of the original RPV model and additionally
5 Terrestrial Applications of Multiangle Remote Sensing 113

Fig. 5.9 Typical BRF anisotropy in the red spectral domain for radiatively homogeneous (left) and
heterogeneous (right) vegetation canopies. The BRFs of the heterogeneous surface were generated
using a 3-D RT model and typified by a bell-shape (k > 1). The BRFs of the homogeneous surface
covers were generated using a 3D RT model and a 1D IPA approach, and are generally bowl-shaped
(k < 1) (Reproduced from Widlowski et al., 2001. Copyright, AGU.)

provides a measure of uncertainty on the retrieved parameters (Lavergne et al.,


2007). The modified Minnaert function parameter, k, is used to quantify the degree
by which the observed bidirectional reflectance factor data resemble a bowl- or bell-
shaped pattern in azimuthal planes (Fig. 5.9). The presence of a very sparse or very
dense and quasi-homogeneous vegetation layer within the sensor IFOV is character-
ized by a bowl-shaped reflectance anisotropy shape (k < 1), while vertically elon-
gated foliage clumps of medium-to-high densities uniquely generate bell-shaped
shapes (k > 1) (Widlowski et al., 2001, 2004; Pinty et al., 2002; Fig. 5.9). This
broad relationship holds well for snow-covered areas, with homogeneous bare or
non-forested areas exhibiting a bowl-shaped pattern (k < 1), indicating that sub-
pixel homogeneity (sparse or extremely dense vegetation cover) will result in k < 1.
A dependency on vegetation density was demonstrated for the first time in Gobron
et al. (2002).
The k parameter has a limited physical meaning (Nolin, 2004); in order to trans-
late these relationships to measurable canopy structural parameters, it is possible
to interpret k values as a ratio between mean effective scene height and the ratio of
mean tree density and mean nearest-tree distance (Widlowski et al., 2004). Effective
scene height is defined as the mean height of all structures within the IFOV weighted
by their fractional surface coverage. The relationship can be seen in kred results
obtained by these authors using radiation transfer simulations performed with an
illumination zenith angle of 30◦ , a soil reflectance of 0.126, a bark reflectance
of 0.251 and a leaf reflectance (transmittance) of 0.018 (0.021) (Fig. 5.10). The
data points generated differ only because of their structural characteristics (stem
and foliage density). Researchers using multiangle data from the European CHRIS
instrument found that the results of RPV model inversion enabled discrimination
114 M.J. Chopping

Fig. 5.10 The organization of the kred values in a plane defined by the “mean effective scene
height”, i.e., the height (m) of all structures within the IFOV weighted by their fractional surface
coverage and – on a log scale – the ratio between average tree density (stems/m2 ) and the mean
nearest tree distance (m). The data points show the characterizations of ∼200 structurally diverse
boreal forest scenes (Reproduced from Widlowski et al., 2004. Copyright, Springer. With kind
permission of Springer Science and Business Media.)

between different surface types based on the their inherent reflectance anisotropy
and that the k parameter was successfully linked to lidar measurements representing
the 3-D structure of the canopy (Koetz et al., 2005a).

5.3.5 Clumping Index

Foliage clumping is an important forest structural canopy attribute: it affects both


the gap fraction for the same LAI, radiation interception and distribution within the
canopy, which in turn affects photosynthesis. It can be quantified in a Clumping
5 Terrestrial Applications of Multiangle Remote Sensing 115

Legend
0.0 - 0.3
0.3 - 0.4
0.4 - 0.5
0.5 - 0.6
0.6 - 0.65
0.65 - 0.7
0.7 - 0.8
0.8 - 1.0
No Value
Water
Excluded

Fig. 5.11 Global clumping index map derived from POLDER 1 data using the normalized dif-
ference between interpolated hotspot and darkspot NIR reflectance and applied to vegetated land
cover. Clumping increases with decreasing values of the index (Reprinted from Chen et al., 2005.
Copyright, Elsevier. With permission.)

Index (CI) based on a vegetation dispersion parameter Ω obtained via a modification


to Beer’s law. A physical modeling approach to obtaining CI has been taken by
two collaborating groups in Canada at the University of Toronto and the Canada
Centre for Remote Sensing, Ottawa (Chen et al., 2005). These authors have produced
the first global maps of foliage clumping with multiangle POLDER data, assisted
by a geometrical optical model (Fig. 5.11). They used the Normalized Difference
between Hotspot and Darkspot (NDHD), an angular index that can be obtained from
multiangle remote sensing that characterizes reflectance anisotropy and has been
successfully related to ground measurements of CI (Chen et al., 2003). Using the
relative magnitude of the darkspot to the hotspot minimizes the dependence of the
index on foliage optical properties that are important determinants of bidirectional
reflectance. This work built on previous research using data from POLDER (Lacaze
et al., 2002).

5.3.6 Structural Scattering Index

A semi-empirical approach applied to MODIS data is the use of the Structural Scat-
tering Index (SSI) developed to exploit the kernel weights derived by inversion of
the LiSparse-RossThick kernel-driven BRDF model (Gao et al., 2003). The MODIS
BRDF/Albedo (MOD43) product is generated on a regular 16-day cycle using this
model. The model depends on three parameters (kernel weights) describing the in-
teraction of light with the surface. These parameters are used in a forward version
of the model to reconstruct the surface anisotropic effects and to correct MODIS
reflectances to a common view geometry (the MOD43B4 Nadir BRDF-Adjusted
116 M.J. Chopping

Reflectances – or NBAR – product). The model can also be used to compute


integrated black-sky albedo at some solar zenith angle as well as white-sky albedo
MOD43B3. The model is a linear, semi-empirical, kernel-driven (LiSK) model for-
mulated as a superposition of weighted kernels Eq. (5.7):

R(θs , θv , ϕ ) = fiso + ( fgeo × kgeo (θs , θv , ϕ )) + ( fvol × kvol (θs , θv , ϕ )) (5.7)

where R(θs , θv , ϕ ) is the modeled bidirectional reflectance for the solar zenith,
view zenith, and relative azimuth angles θs , θv , and ϕ , respectively. The model ker-
nels kgeo and kvol are analytical functions of the solar and viewing angles derived via
simplifying physical terms (Roujean et al., 1992; Wanner et al., 1995): kgeo accounts
for geometric scattering and shadowing, while kvol accounts for volume scattering
from a discrete medium of randomly located facets, assuming the single scatter-
ing approximation, an isotropic facet distribution function, and an optically thick
medium. The kernel weights fiso , fgeo and fvol are theoretically dependent on the
three dimensional structure and optical properties of a canopy and its background;
fiso should represent the bidirectional reflectance viewing a surface at nadir with the
overhead Sun, while fgeo and fvol measure the relative contributions of the geomet-
ric/shadowing and volume scattering components. Models of this kind are also of-
ten referred to generically as Li-Ross models, since the terms describing anisotropic
scattering are obtained from functions derived by Xiaowen Li and Juhan Ross (see
Wanner et al., 1995).
The SSI is calculated as ln( fvol nir / fgeo red ), where fvol nir is the volume scat-
tering kernel weight in the NIR band and fgeo red is the geometric kernel weight
in the red band. It is based on the principle that surfaces with a higher (lower)
and a relatively homogeneous (heterogeneous) vegetation cover will exhibit BRDFs
that exhibit a behavior closer to (further from) that of an ideal turbid medium and
will therefore generally have a higher (lower) volume scattering kernel weights in
the near-infrared (NIR) wavelengths; while in the red wavelengths sparser, more
clumped vegetation will exhibit higher kernel weights (Fig. 5.12). Since SSI de-
pends on the vegetation cover fraction as well as vertical structure, a relative struc-
tural scattering index has been defined that partially removes the effects of cover
by estimating a linear relationship between SSI and a spectral vegetation index.
This approach has been found to be useful in exploiting canopy structure as trans-
lated through a kernel-driven model and should provide superior results for mapping
land cover.

5.3.7 Geometric-optical and Hybrid Models

The use of discrete object models such as geometric-optical (GO) and hybrid
geometric-optical/radiative transfer (GORT) models represents another approach to
exploiting multiangle data (Strahler et al., 2005). Geometric-optical (GO) models
treat the surface as an assemblage of discrete, identical, and large objects defined by
5 Terrestrial Applications of Multiangle Remote Sensing 117

4
Structural Scattering Index

−1
0 0.2 0.4 0.6 0.8 1
Nadir View NDVI (SZN = 30)

All Samples Evergreen Needleleaf Forest


Open Shrubland (Jan-Mar) Open Shrubland (Apr-Oct)
Barren/Sparsely Veg (Nov-Mar) Barren/Sparsely Veg (Apr-Oct)

Fig. 5.12 Structural Scattering Index (SSI) vs nadir view NDVI from MODIS. The solid trend line
represents the change of SSI as the function of nadir view NDVI. Samples above the trend line
show the stronger volumetric scattering; those below show the stronger geometric scattering. SSI
is also dependent on temporal changes in cover types (Reprinted from Gao et al., 2003. Copyright,
Elsevier. With permission.)

geometric primitives (e.g., spheroids, cones, or cylinders) placed in a Poisson distri-


bution above an underlying surface (Li and Strahler, 1985, 1992) which is frequently
considered Lambertian (Scarth and Phinn, 2000). The remotely sensed observation
is modeled as a linear combination of contributions from viewed sunlit and shaded
crown and background components, with each contribution a product of component
reflectance and the fraction of the sensor’s field of view occupied by the component
Eq. (5.8):
R = C × k C + G × kG + T × kT + Z × kZ (5.8)

where C, G, T , and Z are the (assumed Lambertian) spectral reflectances or “com-


ponent signatures” of the sunlit crown and background and shaded crown and back-
ground, respectively; and the kn are the component fractions.
GO models vary considerably in complexity, with some including terms that
allow for volumetric scattering within crowns rather than simple signatures (Ni and
Li, 2000; Chen et al., 2003, 2005); and in the way the background contribution is
represented (Ni and Li, 2000; Chopping et al., 2005). They hold the potential for ob-
taining upper canopy information such as stand density, openness, mean object size
and canopy height, and crown morphology, which are useful in multiple disciplines
(ecological modeling, forestry, hydrology, radiation budget; Garcı́a-Haro et al.,
2006). GO models have been used in the calculation of Clumping Index (see above);
118 M.J. Chopping

some results from applying a simple GO model to estimating woody plant cover in
semi-arid environments in the southwestern USA are provided in the sections enti-
tled The Background in Canopy Reflectance Modeling.
Using lidar data obtained over forest in the boreal ecosystem-atmosphere study
(BOREAS) sites in central Canada, Ni-Meister (2005) showed that lidar waveforms
can be simulated accurately and with precision using a modified version of the
Geometric-Optical/Radiative Transfer (GORT) canopy reflectance model (Li et al.,
1995), driven with canopy structure inputs. This model is dependent on canopy
geometry parameters such as tree size, shape, and tree density; and on the spec-
tral reflectance properties of leaves and the BRDF of the background. The GORT
model was designed for use with passive remote sensing data and is appropriate
for the interpretation of multiangle data. Since it can be used to relate canopy pa-
rameters in both the active and passive domains it can be used as a bridge between
large footprint lidar and moderate resolution multiangle data. This is a particularly
promising direction.

5.3.8 Direction and Wavelength

The future will certainly see greater use made of the directional signal in many
wavelengths as the advantages of spectro-directional remote sensing become more
apparent (Garcı́a-Haro et al., 2006; Schaepman, 2006). However, most use is still
made of red and NIR band data, mainly because it is in these regions that the dom-
inant signatures in both spectral and angular domains are see; partly because these
have long been used in remote sensing of vegetation; partly because MODIS and
MISR both provide moderate resolution data (250–275 m) in these bands; and partly
because interactions with the canopy are better understood. As seen above, sev-
eral groups working with MISR data have adopted modeling approaches in which
only the red band reflectance data are used. This is predicated on the following
arguments:
• The single scattering regime is dominant in the red spectral domain.
• It exploits the reflectance and absorption contrasts between vertically clumped
vegetation and the background.
• The linear mixture assumption underlying geometric-optical models is more
valid for the red than near infra-red wavelengths.
• BRDF model inversion experiments using numerical methods show that there are
generally fewer problems such as trapping at local minima in the red compared
to the near-infrared (Gemmell, 2000).
• It keeps the modeling problem (relatively) simple.
On the other hand, there have been attempts to investigate and exploit the spectral
and directional domains simultaneously (Weiss et al., 2000; Gobron et al., 2000,
2002; Bacour et al., 2002; Casa and Jones, 2005; Bach et al., 2005; Garcı́a-Haro
et al., 2006). Weiss et al. (2000) addressed the question of a potentially optimal set of
5 Terrestrial Applications of Multiangle Remote Sensing 119

spectral bands and directions for retrieval of biophysical parameters using the Scat-
tering by Abitrarily Inclined Leaves (SAIL) radiative transfer model and the PAR-
CINOPY nested radiosity/turbid medium model code. These were used to simulate
multiangle remote sensing data in the principal and perpendicular planes (32 direc-
tions) for a solar zenith angle of 45◦ and for a wide range of canopy conditions.
They found that for estimating the primary variables LAI (leaf area index) and Cab
(chlorophyll a + b content), the optimal number of bands is approximately six and
the optimal number of directions for these bands is between four and seven, with six
directions or less required for reliable retrievals of the primary variables plus frac-
tional vegetation cover and fraction of absorbed photosynthetically active radiation
(fAPAR). Additional directions were found to contain redundant information and
so their use would merely induce noise in the retrieval process. The ideal directions
would be located in the principal plane close to the hot spot direction and in the
forward scattering direction for large zenith angles. Note that these results assume
idealized conditions (e.g., no cloud cover).
Bacour et al. (2002) sought to utilize both the directional and the spectral infor-
mation of the images acquired in 16 flights by the AirPOLDER instrument (eight
bands in the visible-NIR) over the Alpilles test site (January–October 1997), as
part of the Alpilles-ReSeDA campaign. Estimation of biophysical variables was
executed by inversion of three one-dimensional radiative transfer models, SAIL,
KUUSK and IAPI, coupled with the PROSPECT leaf optical properties model. They
focused on the capability of model inversion to retrieve the leaf area index (LAI) of
wheat, maize, sunflower and alfalfa crops. A quasi-Newton inversion algorithm was
used to estimate the Cab , LAI, the mean leaf inclination angle (θl ), the hot spot para-
meter (sl ) and a multiplicative soil parameter (asoil ). The three models were shown
to accurately estimate the LAI compared to planimeter measurements but tended
to underestimate LAI for values above 2.3. Compensation effects between LAI and
Cab emerged from the spatial analyses of these variables, although the uncertainties
were low. Casa and Jones (2005) investigated an unconventional approach for the
estimation of leaf area index (LAI) and leaf angle distribution (LAD), based inver-
sion of a canopy ray tracing model against multiangle data from a ground-based
multispectral camera. The model was developed using the Persistence of Vision
Raytracer (POVRAY) and inversion was carried out using a look-up-table approach.
Tests using an extensive data set gathered on a potato crop during experimental tri-
als carried out at Viterbo (Italy) over 3 years showed that LAI was estimated with a
RMSE varying from 0.29 to 0.75 in the different years.
Bach et al. (2005) used data from the CHRIS instrument on the Proba platform to
invert the four-stream Soil-Leaf-Canopy radiative transfer model SLC, an extension
of the canopy reflectance models SAILH and GeoSAIL, producing maps of LAI,
leaf chlorophyll, and leaf inclination distribution for agricultural fields in the Upper
Rhine Valley test site along the German/French border for July 2003 (Fig. 5.13).
The input parameters include structural and physiological information on the
canopy, soil optical properties and the observation geometry. The Hapke model is
used to model a non-Lambertian soil BRDF, including its variation with moisture
content. The canopy itself is modeled with a two-layer version of the model SAILH
120 M.J. Chopping

>2.0 >35
2.0-2.5 35-40 Angle LIDF a
2.5.3.0 40-45
3.0-3.5 45-50
50-55 56⬚ -0.30
3.5-4.0 55-60
4.0-4.5 52⬚ -0.20
60-65 49⬚ -0.10
4.5-5.0 65-70
5.0-5.5 70-75
5.5-6.0

Fig. 5.13 LAI, chlorophyll per leaf area (µg/cm2 ) and average leaf angle maps for a 3 × 4 km
area of maize, retrieved by adjusting the SLC model against data from CHRIS for the Upper Rhine
Valley test-site along the German/French border (July 2003)

and transmittance of green and brown leaves is calculated using the PROSPECT
sub-model. The leaf angle distribution is approximated by parameters that de-
scribe the average leaf slope and the “bimodality” of the distribution. The spatial
distributions of all three retrieved parameters delineated fields and the additional
information obtained from the directional data showed that canopy structure was
crop-specific but also changed with phenological development. The field pattern in
the average leaf slope map (Fig. 5.13c) values was thought to result from different
varieties of maize. When the average leaf angle was retrieved for the same site
2 weeks after the initial CHRIS acquisition, the overall average angle retrieved
changed by about 10◦ to a more vertical distribution. This can be interpreted as mat-
uration of maize where leaves become more vertical with development. The study
showed that crop parameters retrieved from multiangle CHRIS data can provide
input parameters essential for crop growth models.
Garcı́a-Haro et al. (2006) developed an approach called directional spectral
mixture analysis (DISMA) for retrieving vegetation parameters with the focus on
fractional cover and leaf area index. This seeks to combine a consideration of the
spectral signatures of soil and vegetation components with an analytical approxi-
mation of the radiative transfer equation, resulting in a fast, invertible model suit-
able for use with discontinuous canopies. Data from the AirPOLDER and HyMap
instruments were used to test the model and its inversion using a lookup table
(LUT). Retrievals of LAI corresponded well to ground measurements of LAI, with
an RMSE of 0.5–0.6 and an R2 of the fitting of around 0.92. The spatial distribution
matched that obtained by inversion of the physically-based New Advanced Discrete
Model (NADIM) radiative transfer code, also known as the Semi-discrete model
(Gobron et al., 1997) (Fig. 5.14).
5 Terrestrial Applications of Multiangle Remote Sensing 121

(a) (b)

0.0 0.5 1.0 1.5 2.0 2.5 3.0 >3.5

Fig. 5.14 (a) Green LAI obtained via DISMA using AirPOLDER (b) the corresponding results
from inversion of the physically-based NADIM radiative transfer code. The maps shows a 5 × 5 km
area 28 km from Albacete, Spain; 1% and 7% of the pixels exceed the saturation value of 3.5 (From
Garcı́a-Haro et al., 2006. Copyright, IEEE. With permission.)

5.3.9 The Background in Canopy Reflectance Modeling

Many researchers have recognized that for successful modeling of heterogeneous,


clumped, or non-closed canopies it is important to account adequately for the
contribution of the background of soil and understory (Gemmell, 2000; Ni and
Li, 2000; Chen et al., 2005; Koetz et al., 2005b; Bach et al., 2005; Chopping
et al., 2006a). In all canopies except closed – especially where the understory is
heterogeneous – attempts to exploit canopy reflectance models can be confounded
by spatial variation in background reflectance magnitude and anisotropy. In the last
few years modelers have moved from the assumption of a Lambertian background,
through imposing a single, static background BRDF, to attempts to estimate a spa-
tially dynamic background BRDF. Recent work has shown that the soil/understory
reflectance can be obtained for both coniferous and deciduous forests using MISR
data, with the retrieved values following seasonal trajectories similar to those of
adjacent grasslands, a partial validation of the approach (Canisius and Chen, 2007).
Recent progress in dynamically estimating the anisotropic soil-understory con-
tribution for discontinuous open shrub canopies in desert grasslands exhibiting a
wide range of canopy-background configurations – including young, small honey
mesquite shrubs over a dark grassland matrix and older, larger ones over bright,
sandy soils – has been made using the simplified geometric-optical (SGM) model
incorporating a kernel weighting approach with MISR data. Estimating the back-
ground contribution from MISR-derived LiSparse-RossThin model kernel weights
while fixing number shrub density, shape and height at typical values and using
numerical optimization to retrieve mean shrub radius has allowed the mapping
of fractional shrub cover at landscape scales in the United States Department of
122 M.J. Chopping

Agriculture (USDA), Agricultural Research Service (ARS) Jornada Experimental


Range in southern New Mexico. A root mean square error of 0.03 in fractional
cover with respect to values estimated from high resolution IKONOS panchromatic
imagery was obtained (measured fractional cover range: 0.03–0.19); however only a
moderate proportion of the variation in the measured data was explained by the esti-
mated data set, with an overall R2 of 0.19 (Chopping et al., 2006b). When the same
algorithm was used to predict fractional shrub cover for pastures a few kilometers
from the study area for which independent estimates were available (from image
segmentation on a 0.6 m QuickBird panchromatic image), the distributions exhib-
ited an improved spatial correlation (Fig. 5.15) and an R2 of 0.47 was obtained with
two model variants (Fig. 5.16).
When the algorithm was applied with MISR data over much larger areas – for
the Jornada Experimental Range (∼783 km2 ) and the Sevilleta National Wildlife
Refuge (∼1, 000 km2 ) and their environs – the maps include trees on the San
Andres mountains, in other elevated areas (e.g., Summerford Mountain in the Jor-
nada), and in the riparian environments of the Rio Grande (southwest quadrant of
the Jornada map and center of the Sevilleta map), in addition to shrubs. The re-
sulting distributions compare well with those of trees in the MODIS Vegetation

Fig. 5.15 (a) QuickBird shrub map for pasture 12 in the Jornada, red = shrub, white = background
(b) shrub cover aggregated to 250 m cells; brighter = greater shrub cover (c) retrieved using MISR
red band data to invert the SGM GO model (d) MISR/SGM shrub cover map for pastures 8/9 in the
Jornada (e) the corresponding QuickBird map (From Chopping et al., 2006c. Copyright, American
Geophysical Union.)
5 Terrestrial Applications of Multiangle Remote Sensing 123
Fractional Shrub Cover (MISR/SGM)

Fractional Shrub Cover (MISR/SGM)


0.30 y = 0.4358x + 0.0208 0.30 y = 0.5427x + 0.0777
R2 = 0.4676 R2 = 0.4824
++
+ +
+ + +
+ +
0.20 0.20 ++ ++
++ + +++
+
++ + +
++ +
+
+
+++
+ ++ +
+ + + +++ + +
++ + + ++++ ++++ ++
++ +
+ + + +
+ + + + + + +++
0.10 + + + + 0.10 + ++ +
++ +
+++ + +++ + + + + ++ ++
++
+ ++ ++ +
+ ++ +++++ ++++
++ +
++ + ++
+ + + + + +
+ ++ ++
0.00 + 0.00
0.00 0.10 0.20 0.30 0.00 0.10 0.20 0.30
Fractional Shrub Cover (QB-NIH) Fractional Shrub Cover (QB-NIH)

Fig. 5.16 Retrieved vs. measured shrub cover (a) SGM (b) SGM with sunlit crown unweighted
by the fraction viewed (i.e., kC = 1.00) (Reproduced from Chopping et al., 2006c. Copyright,
American Geophysical Union.)

Continuous Fields (VCF) percent tree cover maps (Fig. 5.17). Note that the MISR-
derived map includes all large woody plants: it is based on exploiting the canopy
structure information encapsulated in the multiangle reflectance data (Chopping,
2006). Subsequent inversions of the SGM that allowed the crown shape para-
meter to vary indicated that it is possible to obtain regional maps of canopy height
as well as crown cover, allowing estimates of aboveground woody biomass. Re-
trievals of cover, canopy height, and biomass showed good matches with US For-
est Service maps, with coefficients of determination of 0.78, 0.69, and 0.81, and
absolute mean errors of 0.10, 2.2 m, and 10.1 Mg/ha, respectively, after filtering for
high model fitting error, the effects of topographic shading, and a small number of
outliers (Chopping et al., 2007; http://csam.montclair.edu/∼chopping/wood/).

5.3.10 Land Cover and Community Type Mapping

Many studies have shown that there is much potential for improving the accu-
racy of land cover classification if patterns in the angular domain as well as the
spectral domain can be exploited (Abuelgasim et al., 1996; Hyman and Barnsley,
1997; Bicheron et al., 1997; Sandmeier and Deering, 1999). Chopping et al. (2002)
showed that the kernel weights of Li-Ross LiSK models obtained by adjusting a
LiSparse-RossThin variant against accumulated multiangle data from the NOAA
AVHRR provide contingency tests for community types in semi-arid environments
in Inner Mongolia Autonomous Region (IMAR; 10 types) and New Mexico (NM;
19 types) that are superior to those obtained using maximum-value compositing
using the Normalized Difference Vegetation Index (NDVI) as the criterion (denom-
inated MVC), and in particular perform better in the worst case: the kernel weights
124 M.J. Chopping

Fig. 5.17 Woody plant fractional crown cover map obtained by adjusting the SGM geometric-
optical model against MISR red band data (a) in the Sevilleta National Wildlife Refuge in central
New Mexico (b) % tree cover from the corresponding MODIS VCF product (c) in the Jornada
Experimental Range in southern New Mexico, dotted line is the tree-shrub boundary from the VCF
map (d) % tree cover from the corresponding VCF map product. Solid lines indicate the boundaries
of the Sevilleta and fencelines in the Jornada. (Reproduced from Chopping et al., 2006c. Copyright
2006 American Geophysical Union.)

provided minimum reliabilities of 83% and 69% for IMAR and NM experiments,
respectively, compared to only 11% and 26%, respectively, with the MVC data set.
These performances were reflected in Kappa Index values of 0.93 and 0.91 for the
IMAR and NM spectro-directional data sets against 0.74 and 0.46 for the MVC data
sets, respectively (the Kappa Index is a means to test two data sets to determine the
extent to which their similarities or differences are due to chance). When directional
5 Terrestrial Applications of Multiangle Remote Sensing 125

information was incorporated into the signatures, errors of omission or commission


fell from ∼24% and ∼51% obtained with MVC to only ∼6% and ∼8% for the
IMAR and NM experiments, respectively.
More robust cross-validated tests of classifications of multiangle remote sens-
ing metrics from MISR and nadir-spectral data from the An (nadir) camera have
demonstrated that using multi-angular data and anisotropy patterns raises the over-
all classification accuracy importantly. Su et al. (2007) studied maximum likelihood
and support vector machine (SVM) algorithms for mapping community types in
the Jornada Experimental Range and the Sevilleta National Wildlife Refuge, New
Mexico (19 classes). Half of the samples were randomly selected as the training
set and the other half retained as the testing set. A total of 66 classifications were
performed with various combinations of data sets: the ρ0 , k, and b parameters of
the MRPV model (see Canopy Openness, above); the isotropic, geometric and vol-
ume scattering kernel weights of a Li-Ross BRDF model; the structural scattering
index; and MISR surface reflectance estimates. Using multiangle data raised the
overall classification accuracy from 45.4% obtained with nadir observations only
to 60.9%, and with surface anisotropy patterns derived from MRPV and LiSparse-
RossThin BRDF models (separately) an overall accuracy of 67.5% can be obtained
with a maximum likelihood classifier. Using a non-parametric SVM algorithm the
classification accuracy was increased to 76.7%. Note that the classes in these exper-
iments are community types that often differ very subtly in terms of their spectral
signatures, rather than broad land cover types (Fig. 5.18).

Fig. 5.18 Community type maps for the Sevilleta National Wildlife Refuge. (a) LTER vegetation
map (b) MISR maximum-likelihood classification (c) MISR support vector machine classification
(Reproduced from Su et al., 2007. Copyright, Elsevier. With permission.)
126 M.J. Chopping

5.4 Other Applications

Most applications – or prototype applications – of multiangle remote sensing of land


have been concerned with vegetation canopies. However, there are other important
Earth Science applications that are often tangentially related to vegetation, includ-
ing estimating land albedo and the Earth’s radiant energy budget (see Chapter 9);
snow and ice; and mapping dust emission sources to simulate the likelihood of large
emissions.

5.4.1 Snow and Ice

Multiangle remote sensing data are able to provide important additional information
on Earth surfaces covered with snow and ice. For example, differentiating between
clouds and snow or ice surfaces using spaceborne detectors is difficult because the
surface may often be as bright and as cold as the overlying clouds, and because po-
lar atmospheric temperature inversions sometimes mean that clouds can be warmer
than the underlying surface. Mega-sastrugi ice fields in East Antarctica, with dune-
like features as high as 4 m and separated by 2–5 km – a result of unusual snow
accumulation and redistribution processes influenced by the prevailing winds and
climate conditions – appear more like cloud formations: they exhibit a rippled ap-
pearance. However, the Angular Signature Cloud Mask (ASCM), a MISR product
(Di Girolamo and Wilson, 2003) is able to detect clouds over snow and ice as well as
over ocean and land (Fig. 5.19). MISR imagery indicated that these mega-sastrugi
were stationary surface features between 2002 and 2004.

(a) (b)

708 backward viewing camera confidence high low low high


cloud surface

Fig. 5.19 (a) December 16, 2004 MISR image over Antarctica, showing sastrugi (b) color-
coded image showing the Angular Signature Cloud Mask results (Courtesy of the MISR Team,
NASA/JPL/Caltech, and L. Di Girolamo and M.J. Wilson, University of Illinois.)
5 Terrestrial Applications of Multiangle Remote Sensing 127

Hendriks and Pellikka (2004) presented a study of the multiangular reflectance


of a glacier surface. They used multiple ALTM (Optech, Ontario, Canada) digital
metric camera frames acquired from the air on August 12, 2003, to examine the
angular spectral reflectance response of snow, firn, and ice surfaces on the Hintere-
isferner glacier in Austria. The camera was operated in color infrared (CIR) mode,
resulting in images in three spectral bands: green (510–600 nm), red (600–720 nm)
and near-infrared (720–800 nm); the pixel size varied between 0.25–0.30 m as a re-
sult of topography. They found important differences in the angular signatures of the
three surface types and compared their results to MISR imagery (local mode: nom-
inal 275 m footprints at all angles and in all four bands). Their results showed that
MISR data reveal the backscattering of dirty ice, firn and old snow; BRF increases
going from the nadir image in the backward direction and at the viewing camera
Ca (60◦ ) backwards, BRF is 30% higher relative to the value in the nadir viewing
camera (An) (Fig. 5.20).
Multiangle views provide unique sub-pixel resolution information about the ice
sheet surface that can be used to improve the characterization of climate and ice dy-
namics processes. Nolin et al. (2002) had previously shown that MISR data can
be used as a proxy for surface roughness. They developed a normalized differ-
ence angular index (NDAI) using a combination of forward and backward scat-
tered radiation in the MISR red band at the 60◦ fore and aft viewing angles of the
MISR instrument. A positive (negative) NDAI value indicates that backward (for-
ward) scattering exceeds forward (backward) scattering and that the surface is rough

0.52

0.47

0.42 Firn/Old snow

0.37
BRF

0.32

0.27
Dirty ice
0.22

0.17
Cf Bf Af An Aa Ba Ca
0.12
60 45.6 26.1 0 26.1 45.6 60
View angle (degrees)

Fig. 5.20 MISR BRFs for dirty ice and firn/old snow. The green, red and NIR bands are shown
by the green, red and blue lines, respectively, for the Hintereisferner glacier in Austria, August 14,
2003 (Courtesy of J.P.M. Hendriks and P. Pellikka.)
128 M.J. Chopping

Fig. 5.21 Ice sheet surface classes for 2000–2005 derived from ISODATA classification of summer
NIR albedo and NDAI from MISR images acquired in the vicinity of the Jakobshavn glacier and
its upland drainage basin. Contour lines of elevation are at 250 m intervals (Reprinted from Nolin
and Payne, 2006. Copyright, Elsevier. With permission.)

(smooth). The NDAI was shown to be highly correlated with surface roughness de-
rived from an airborne laser altimeter. Nolin and Payne (2006) used the ISODATA
clustering technique with NIR albedo and NDAI values from surface hemispherical-
directional reflectance for consecutive MISR summer images from 2000–2006 over
an area in the vicinity of the Jakobshaven glacier in Greenland. The classified maps
(Fig. 5.21) demonstrated good spatial and temporal consistency for seven ice sheet
classes for all 6 years; moreover, the classes roughly correspond with glacier facies
mapped previously by other researchers. The classes differ in albedo, roughness (in-
cluding the presence of crevasses), wetness, and the age of the snow at the surface.
Nolin (2004) performed a study using data from MISR to demonstrate how
the angular pattern of reflectance from vegetation over snow can provide infor-
mation on forest cover density. This is important in snow studies as vegetation
structure and density affect the dynamics of snow accumulation and ablation and
affects the ability to estimate snow-covered area accurately from satellite-based
sensors. The study area was located in north-central Colorado. MISR red band
level 1B2 (top-of-atmosphere radiometrically and geometrically calibrated spectral
radiances) data from 15 February 2002 were converted to top-of-atmosphere bidi-
rectional reflectance factors – no atmospheric correction was applied. The Rahman–
Pinty–Verstraete (RPV) semi-empirical parametric model was successfully used to
simulate the angular patterns of reflectance. The model’s k parameter was used to
characterize the angular signatures of selected pixels. In the RPV model this para-
meter is used to quantify the degree by which the observed bi-directional reflectance
5 Terrestrial Applications of Multiangle Remote Sensing 129

2
71%
1.5
93%
1 99%

0.5
Normalized BRF

−0.5

−1

−1.5

−2

−2.5
−80 −60 −40 −20 0 20 40 80 60
MISR camera angle
Fig. 5.22 MISR angular signatures for three forest cover densities of fir–spruce forest. Note the
change in shape with increasing tree cover; this is captured in the RPV model’s k parameter
(Reprinted from Nolin, 2004. Copyright, Wiley. With permission.)

factor data resemble a bowl- or bell-shaped pattern. The results showed distinct pat-
terns in the retrieved k parameter values, with a marked dependence on density and
cover type. Non-forested areas exhibited a bowl-shaped pattern (k < 1.0) of re-
flectance versus viewing angle. Low-density deciduous forests also exhibited this
bowl-shaped reflectance pattern, changing as density increases. Other forest cover
types show transitional patterns between bowl and bell shapes and distinct bell-
shaped patterns (k > 1.0) for higher densities (Fig. 5.22). The relationship between
k and density does not hold for forest cover densities that approach 100%. For a
density of 99%, the fir – spruce forest cover type has a distinct bowl shape and a k
value of only 0.69. This is in agreement with previous work indicating that sub-pixel
homogeneity (whether because of sparse vegetation cover or extremely dense vege-
tation cover) will result in k < 1.0. This study indicated from a qualitative standpoint
that multiangle reflectance data captures information on forest cover density at the
sub-pixel scale.

5.4.2 Dust Emissions

Laurent et al. (2005) used POLDER multiangle data with a surface roughness
parameter estimated from the Roujean kernel-driven model (Roujean et al., 1992).
They used the geometric kernel weight normalized by the diffuse (isotropic) scatter-
ing kernel weight (i.e., geo/iso) to simulate dust emissions from deserts in East Asia
130 M.J. Chopping

(N. China and Mongolia). The approach was derived from work previously accom-
plished by Roujean et al. (1997) and employed an empirical relationship established
by Marticorena et al. (2004). The composite surface roughness map they developed
describes the spatial variability of the erosion threshold (10 m wind velocities) on
dust emission frequency. The map was found to be consistent with geomorphologic
interpretations from Landsat imagery and with soil properties described in the liter-
ature. The retrieved roughness lengths are in agreement with the roughness lengths
experimentally determined over similar surface types in other deserts of the world.
The authors computed dust emission frequencies for 3 years (1997–1998–1999)
by combining the 10 m erosion threshold wind velocities, the European Centre for
Medium-Range Weather Forecasts (ECMWF) surface wind fields, the snow depth
and the soil moisture computed using the Food and Agriculture Organization of
the United Nations (FAO) soil texture profiles, and ECMWF meteorological data.
The simulated frequencies of significant dust emissions were compared to the fre-
quencies of occurrence of Total Ozone Mapping Spectrometer (TOMS) Absorbing
Aerosol Index (AAI) higher than 0.7. Both the location and the relative intensity of
the highest dust emission frequencies identified from the simulations were in agree-
ment with the observations (Fig. 5.23).

0.8
Simulated dust event frequency

0.6

0.4

0.2

0 0.2 0.4 0.6 0.8 1


Frequency of TOMS AAI > 0.7

Fig. 5.23 Monthly frequencies of significant simulated dust emissions (flux > 10−10 g cm−2 s−1 )
as a function of the monthly frequencies of TOMS AAI > 0.7 over the Taklimakan desert for the
3 years 1997, 1998, and 1999. Small dots represent individual data; circles represent the averaged
frequency of simulated dust emissions for classes (5% width) of frequency of TOMS Absorbing
Aerosol Index (AAI) > 0.7; the solid line represent the linear fit of the averaged data (without
accounting for the last class which is not representative) (Reproduced from Laurent et al., 2005.
Copyright, American Geophysical Union.)
5 Terrestrial Applications of Multiangle Remote Sensing 131

5.5 Some Considerations in Multiangle Acquisition

5.5.1 Near-simultaneous and Accumulated Sampling

Near-simultaneous acquisition of multiple observations of the same area on the


Earth’s surface is possible by viewing in the along-track direction. From low Earth
orbit, multiple looks of the same area can be acquired over a time span of less than
10 minutes. This strategy may be contrasted with accumulated sampling, in which
repeat multiangular coverage of a particular area is obtained with a wide cross-track
swath and observations are accumulated during multiple orbits. Accumulation of
multiangle data in this manner takes many days. This may degrade the angular data
set to some degree, depending on how rapidly surface conditions change and – to
some extent – on changes in the atmosphere, since it is more difficult to correct ex-
treme conditions. For along-track systems, there will be delay of only a few minutes
between the first and last acquisitions, whether the instrument employs multiple
sensors (MISR), is tilted (CHRIS/Proba), has a conical scan (AATSR), or has a
very wide field-of-view (POLDER). This lag is unavoidable unless a constellation
of geostationary platforms is available – an unlikely scenario for the near future.
Most Earth observing instruments for which the effects of the surface BRDF have
been studied and – to a far lesser extent – exploited are across-track instruments
on polar-orbiting instruments such as the AVHRR and MODIS. These systems are
designed to view large swaths, covering the globe at least once per 24 h period.
They are intrinsically off-nadir viewing devices but to be used as multiangle in-
struments, observations of the same land areas must be accumulated from multiple
overpasses. In addition, sensors such as Spinning Enhanced Visible and Infra-Red
Imager (SEVIRI) on Meteosat Second Generation (MSG) geostationary satellites
provide angular sampling through diurnal variation in illumination angles. The mo-
tivation for BRDF studies with accumulating instruments has been twofold: the need
to adjust observations to a common Sun-target-sensor geometry for consistency (a
much-touted but actually rather rare quality in Earth Observation in general); and to
obtain more accurate estimates of terrestrial albedo than could possibly be obtained
from a single view, since sunlight is never scattered in a perfectly diffuse manner at
the surface.
In the case of the MODIS BRDF/Albedo algorithm, an additional product is
Nadir-equivalent BRDF Adjusted Reflectance (NBAR), defined as the best estimate
of reflectance at nadir viewing at the mean solar zenith angle of the observations
used to invert the Li-Ross BRDF model (in this case LiSparseMODIS-RossThick).
NBAR was quickly found to provide superior performance in applications such as
land cover mapping (Zhang et al., 2000). The possibility of obtaining additional
surface information by exploiting the directional signal encapsulated in these data
was something of an afterthought until the 2000s; however, in the last 5 years this
has changed, with several studies addressing the retrieval of vegetation structural
parameters such as LAI and fractional vegetation coverage (Camacho-de Coca et al.,
2002; Chen et al., 2005).
132 M.J. Chopping

Accumulating instruments have one major advantage over those providing near-
simultaneous multiangular acquisitions: they provide greater spatial and temporal
coverage. This is particularly important for global applications where cloud and
cloud shadow contamination present a problem for instruments with revisit periods
of less than a few days. However, near-simultaneous multiangle imaging provides
some important advantages over accumulated sampling:
• It greatly helps to solve the aerosol scattering problem over land and especially
over bright surfaces – including deserts – where spectral information alone cur-
rently appears to be inappropriate.
• Directional data sets acquired by accumulation may afford serious limitations for
applications under those conditions where either the atmosphere or the surface –
or both – change more rapidly than the accumulation period. This advantage is
more applicable in some circumstances than in others and is probably more de-
pendent on the development of the surface rather than that of the atmosphere.
For example, changes owing to snowfall, snowmelt, flood, and fire may not be
resolvable by accumulated sampling as they happen much more rapidly than the
sampling rate. Near-instantaneous imaging is therefore the only way to assess
surface directional properties under these circumstances. In other cases, for ex-
ample, where changes in surface conditions are slow and/or sparse and intermit-
tent, accumulated sampling is more acceptable.
• There is the potential to assess information on canopy structure and retrieve LAI
and fPAR with greater accuracy (Hu et al., 2003).

5.5.2 Angular Sampling

The angular sampling of both multiangle imaging and accumulating instruments


is determined by sensor design (field-of-view/swath), the platform on which the
sensor is flown, and by the constraints imposed by the orbit into which the platform
is injected. The first and last of these are straightforward. However, that the platform
itself can impact the angular sampling regime of a sensor has a meaning restricted
to satellites that are able to re-orient themselves in order to acquire multiple looks.
Currently this definition applies to only one system: CHRIS on the Proba-1 satellite.
In the future, increasing use of constellations of small, agile satellites such as Proba
might be expected to provide greater flexibility in angular sampling. One such sys-
tem concept developed at NASA/GSFC, dubbed Leonardo-BRDF has not yet been
pursued (Wiscombe, 2000). The angular sampling provided depends on whether
the instrument is an across-track scanner (using a revolving or oscillating mirror;
e.g., AVHRR, MODIS, the forthcoming US National Polar-orbiting Operational
Environmental Satellite System (NPOESS) Visible/Infrared Imager/Radiometer
Suite, or VIIRS); an along-track scanner ((A)ATSR(−2)); multiple along-track
moderate swath CCD arrays (MISR); or an along-track wide swath CCD array
(POLDER/Parasol). Important issues with respect to angular sampling include:
5 Terrestrial Applications of Multiangle Remote Sensing 133

• The range of viewing and illumination zenith angles


• The azimuthal coverage of the samples (how well the hemisphere is covered,
particularly with respect to the principal and perpendicular planes)
• The angular sampling density, particularly around the hot spot (rarely achieved)
• The spatial resolution, determined by the size of the solid angle of the individual
observing element, which is somewhat large for POLDER and much smaller for
MISR and CHRIS
• The interval over which the sampling is acquired (minutes for true multiangle
imagers such as MISR, CHRIS, and POLDER; up to weeks for accumulators)
• The frequency with which a given sampling may be repeated
Since most instruments are flown on polar-orbiting platforms with sun-synchronous
orbits having orbital inclinations only somewhat greater than 90◦ and acquisitions
are made either side of local noon, across-track scanners tend to provide a sampling
that is closer to the principal plane – where variation in radiance with viewing and
illumination zenith is greatest – for mid-latitudes. Along-track imagers on platforms
with sun-synchronous orbits such as MISR provide a sampling that is often far from
the principal plane but approaches it with increasing latitude; and for any given
latitude the sampling remains consistent. Owing to the Proba-1 orbit and the narrow
field-of-view, CHRIS provides a less regular sampling from site to site and it is only
occasionally able to image a target area from directly overhead view; Proba-1 has
to be tilted at some – usually small – angle in the across-track direction so that the
target area is viewed. The platform acquires images of the target when the zenith
angle of the platform is one of the following: ±55◦ , ±36◦ and 0◦ . This means
that the angles at which images are acquired vary from site-to-site, depending on
their positions with respect to the orbital track. The pattern with which any site is
accessible to the platform varies at roughly 8-day intervals, but with some change
in sampling because the orbit does not repeat exactly (Barnsley et al., 2004). The
POLDER design allows the most comprehensive sampling of the Earth’s radiation
field from space to date. The very wide field-of-view (2, 400 × 1, 800 km2 ) allows
observation of the same target under many different angular configurations (be-
tween 10 and 15 observations for each passage of the satellite), with view zenith
angles up to ∼60◦ for the full azimuth range (Laurent et al., 2005).

5.5.3 Scale and Multiangle Observation

The spatial scale at which multiangle observations over land are made has an im-
portant impact on the kinds of applications in which they might be used, as well as
on data transmission rates and geographic coverage. Progress in remote sensing has
often been measured in terms of increasing the sampling rate in the spatial domain,
so that the smaller the ground sampling distance and the higher the spatial resolution,
the better (NASA, 2000). Indeed, this is often postulated to be the major limitation of
moderate resolution remote sensing devices: while wall-to-wall coverage is provided,
heterogeneity in surface features cannot be resolved, limiting many applications,
134 M.J. Chopping

including cover type classification and the assessment of small-scale but widespread
fragmentation in a variety of ecosystems (Phinn et al., 1996). However, this para-
digm is flawed in certain respects: although high resolution imagers show the details
of land features, this comes at the price of high data volumes monetary expense,
and restricted coverage. Although multi-angle observation can only provide statisti-
cal metrics on sub-pixel structure, this is often sufficient to address the problem of
interest. Moreover, data volumes are smaller and global coverage is achievable. For
example, an hypothetical 10 m resolution single-angle imager generates ten times
more data than a 100 m multiangle imager having ten angles, for the same swath; it
can therefore cover a swath ten times wider for the same data volume.
Note too that with multiangle remote sensing a small ground-projected instanta-
neous field-of-view (GIFOV) can be a disadvantage; specifically, for the radiomet-
ric, or 1-D, approach that seeks to derive “sub-pixel” information via variation in
radiance with Sun and/or viewing angles and in which each “pixel” is treated in-
dependently. In this approach, if the GIFOV is small in relation to surface features
such as trees or shrubs, there is a higher probability that an observation will include a
sampling of elements that is unrepresentative of the surface as a whole; furthermore,
it may also include important contributions from partial features at the extremi-
ties. These may result in noise-like fluctuations and adjacency effects in the angular
signal that will impact on models that treat the surface as either a semi-infinite,
homogenous, plane–parallel medium, or as a set of identical objects distributed spa-
tially in a Poisson distribution (e.g., geometric-optical models). In both cases, an
unrepresentative sampling of features and/or a large relative contribution from par-
tial features at the edge of the GIFOV will obscure the angular signal. If the GIFOV
is large in relation to surface features then the sampling of surface elements will
be more representative and the periphery will make a relatively small contribution.
Other considerations are that a large GIFOV may result in greater proportions of ob-
servations with unavoidable cloud and cloud shadow contamination and with mix-
tures of too many surface types. On the other hand, a very small GIFOV may also
suffer from multiple scattering to/from adjacent footprints (“pixel cross talking”).
Clearly, for any given landscape there is an optimum sampling resolution: if the
GIFOV is much larger than the typical length scales then spatial trends in BRDF and
parameters derived from model inversions will be less well-defined. There is also a
sampling resolution that is the global optimum; i.e., over all surface types of interest.
Following Pinty et al. (2002, 2004), Widlowski et al. (2005) addressed the question
of the degree to which 1-D radiative transfer models can explain as well as describe
the reflectance fields of 3-D forest targets of varying composition and complexity,
over a range of spatial sampling scales (sensor ground sampling resolutions). Ex-
plaining these reflectance fields requires that the model’s internal parameters (state
variables) match and are consistent with those of the 3-D target; it is not sufficient
that the modeled anisotropy matches that of the target. Both conditions must hold if
any model is to provide surface parameters on inversion which have meaning and
utility in remote sensing applications; if they are not, there is considerable doubt
on the reliability of surface parameters retrieved. The study tackled this problem
through addressing two important questions:
5 Terrestrial Applications of Multiangle Remote Sensing 135

• Can 1-D models represent the anisotropy of a reflectance field generated by a


more complex 3-D model at multiple spatial resolutions? and
• Will the inversion of a 1-D model (at low spatial resolution where they are more
appropriate) lead to systematic (and significant) biases in the retrieval of surface
properties?
These questions were pursued by using data sets for a level Scots pine landscape
simulated by the Rayspread 3-D Monte Carlo raytracing model, using allometric
relationships for a range of canopy parameter configurations in the red and near
infra-red wavelengths. A 1-D model known as the Semi-discrete model (Gobron
et al., 1997) was used to generate a large series of lookup tables for the comparisons.
The study examined the impact of changing sensor GIFOV and foliage clumping on
the match between the 3-D and 1-D models, as well as differences in interception,
absorption and transmittance of solar radiation through and by the canopy arising
from the 1-D and 3-D modeling. The conclusions were that 1-D models can provide
good matches to reflectance fields generated by 3-D models in terms of both mag-
nitude and directionality but that it does not follow that the 1-D model’s internal
parameters and radiation budget match those of the 3-D target. It was shown that
differences in the shape of the reflectance anisotropy between a 3-D target and its
1-D homologue – featuring identical spectral and structural properties with the ex-
ception of foliage clumping – lead to the retrieval of BRF-equivalent 1-D solutions
with parameter values that diverge from those of the 3-D target. Part of this problem
is understood as the problem of lack of uniqueness of solutions. A key finding is that
for conifer forest canopies, an observation spatial scale of less than 100 m is more
likely to introduce large discrepancies between the reflectance fields generated by
3-D and 1-D models for a wide range of conditions.

5.6 Conclusions

This chapter has reviewed recent work towards exploiting solar wavelength re-
mote sensing data acquired at multiple viewing angles from the air and space
in applications in forestry, ecology, land cover mapping and land cover change,
agriculture, hydrology, and glaciology, with important implications for enhancing
ecological, crop growth, biogeochemical, hydrologic, and energy balance models.
Even when excluding the numerous studies concerned with goniometric measure-
ments, multiangle observations in the thermal wavelengths, and model development
and simulation work, there has clearly been a great deal of activity over the last
few years. Advances have been seen in a diversity of approaches, with notable gains
for synergistic methods that seek to use multiangle data together with other kinds
of remote sensing data and/or high resolution inventory data. The synergistic use
of multiangle data with those from lidar instruments has provided an especially
promising direction for the future, especially if the capability of models such as
GORT to link multiangle and lidar data can be further developed. While canopy re-
flectance models have become more accurate, often incorporating a wider range of
136 M.J. Chopping

spectral measures, the hindrances to the use of geometric-optical models that have
prevented widespread adoption in applications are being removed through methods
for isolating the contribution of the soil/understory background from that of the up-
per canopy. Studies directed at assessing the gains from adopting the multiangle
approach over nadir-spectral sensing, as well as obtaining meaningful and consis-
tent metrics that characterize surface and canopy conditions, have universally found
multiangle data to be valuable.
Although efforts to exploit the additional, unique information available through
multiangle remote sensing have been ongoing for many years, progress towards
applications is accelerating, with high quality data sets available from MISR and
MODIS on NASA’s EOS satellites, a third POLDER in orbit on Parasol, innovative
new experimental systems such as CHRIS on Proba, greater demonstrated potential
for satisfying multiple user groups in diverse disciplines (e.g., atmospheric science,
ecology, glaciology, land management, and climate modeling), the realization of
important synergies with active instruments, and a greater number of researchers
engaged. This is borne out by the tenfold increase in the number of peer-reviewed
publications using data from MISR, POLDER or CHRIS in the 10 years to 2005
(Fig. 5.24). Canopy reflectance modeling work with a variety of model types and
improved reference data from high resolution imaging and lidar continues to shed
new light on the constraints to robust retrievals of biophysical parameters and on
important issues such as optimal scales of observation. Efforts continue to be made
to engage remote sensing scientists with user groups (e.g., Chopping and Diner,
2005), although this is an ongoing task. Continued improvements in knowledge and
understanding, together with greater experience with both near-simultaneous and

70
POLDER-1 + POLDER-2 + PARASOL
60 MISR

CHRIS
Number of publications

50

40

30

20

10

0
1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005

Year

Fig. 5.24 Growth in peer-reviewed publications focused on exploitation of POLDER, MISR, and
CHRIS, 1990–2005 (Courtesy of David J. Diner, MISR scientist, NASA/JPL.)
5 Terrestrial Applications of Multiangle Remote Sensing 137

accumulating multiangle instruments, will drive the design of future Earth observa-
tion systems. The examples presented in this chapter provide a good, albeit not ex-
haustive, indication of the range and depth of work that has been accomplished over
recent years in just one area of multiangle remote sensing (terrestrial applications
in the solar wavelengths). They reflect the progress achieved in the ten years since
the NASA-sponsored Workshop on Multi-angular Remote Sensing for Environmen-
tal Applications (Privette et al., 1997) and point the way to further opportunities for
making better use of the unique information provided by multiangle data over land.

Acknowledgements Special thanks are owed to David J. Diner for his very valuable sugges-
tions on the manuscript as well as the many others who helped me in compiling this review;
and in particular Daniel Kimes, Alan Strahler, Bernard Pinty, Jean-Louis Roujean, François-Marie
Bréon, David Jupp, Jon Ranson, Jing Chen, John V. Martonchik, Anne Nolin, Rob Braswell, Julian
Jenkins, Mike Barnsley, Wout Verhoef, Wenge Ni-Meister, Andres Kuusk, Soeren Hese, Hamlyn
Jones, Raffaele Casa, F. Javier Garcı́a Haro, Johan Hendriks, Petri Pellikka, Janne Heiskanen,
Larry Di Girolamo, Mathias Disney, Wolfgang Lucht, Sampo Smolander, Gunar Fedosejevs,
Mike Cutter, Gabriela Schaepman, and Michael Schaepman. I thank also the participants in the
NASA/MISR Workshop on Multiangle Remote Sensing in Ecological Modeling not mentioned
above. I must also acknowledge the many people not named here who sent me their recent and
often unpublished work, and especially those whose work I could not incorporate here: I am deeply
grateful. Any errors or omissions are uniquely mine.

Glossary

1-D One-dimensional
3-D Three-dimensional
AAI Absorbing Aerosol Index
AATSR Advanced Along-Track Scanning Radiometer
ADEOS Advanced Earth Observation Satellite (Japan)
ARS Agricultural Research Service (USDA)
ASCM Angular Signature Cloud Mask
AVHRR Advanced Very High Resolution Radiometer (NOAA)
BHR Bihemispherical Reflectance
BOREAS Boreal Ecosystem-Atmosphere Study (NASA)
BRDF Bidirectional Reflectance Distribution Function
BRF Bidirectional Reflectance Factor
Cab Chlorophyll a + b content (leaves in a given canopy)
CART Canopy Architecture Radiative Transfer (MISR LAI algorithm)
CCD Charge Coupled Device
CHRIS Compact High Resolution Imaging Spectrometer
CI Clumping Index
CIR Color Infrared
DISMA Directional Spectral Mixture Analysis
ECMWF European Center for Medium range Weather Forecasting
EOS Earth Observing System (NASA)
138 M.J. Chopping

FAO Food and Agriculture Organization of the United Nations


fAPAR fPAR Fraction of Absorbed Photosynthetically-Active Radiation
GIFOV Ground-projected (sensor) Instantaneous Field-Of-View
GLAS Geoscience Laser Altimeter System
GO Geometric-Optical (model)
GORT Geometic-Optical/Radiative Transfer (model)
HDRF Hemispherical-Directional Reflectance Factor
ICESat Ice, Cloud, and Land Elevation Satellite (NASA EOS)
IFOV (sensor) Instantaneous Field-Of-View
ISRO Indian Space Research Organization
LAD Leaf Angle Distribution
LAI Leaf Area Index (one-sided leaf area per unit ground area)
Lidar Light Detection and Ranging (sensor)
LiSK Linear, Semi-empirical, Kernel-driven
LVIS Laser Vegetation Imaging Sensor
MISR Multiangle Imaging Spectro-Radiometer (NASA/JPL EOS)
MODIS MODerate resolution Imaging Spectroradiometer (NASA EOS)
MRPV Modified Rahman-Pinty-Verstraete (model)
MSG Meteosat Second Generation
MVC Maximum-Value Compositing (on NDVI)
NADIM New Advanced Discrete Model (model)
NBAR Nadir BRDF-Adjusted Reflectances (EOS MOD43B4 product)
NDAI Normalized Difference Angular Index
NDVI Normalized Difference Vegetation Index
NIR Near Infra-Red (sometimes written near-infrared)
NDHD Normalized Difference between Hotspot and Darkspot
NPOESS National Polar-orbiting Operational Environmental Satellite System
POLDER Polarization and Directionality of the Earth’s Reflectance
POVRAY Persistence of Vision Raytracer
PSLV Polar Satellite Launch Vehicle (ISRO; India)
RAMI Radiation Transfer Model Intercomparison (RAMI) Exercise
RMSE Root Mean Square Error
RPV Rahman-Pinty-Verstraete (model)
RT Radiative Transfer
SAIL Scattering by Abitrarily Inclined Leaves (model)
SAR Synthetic Aperture Radar
SEVIRI Spinning Enhanced Visible and Infra-Red Imager
SGM Simplified Geometric-optical Model
SSI Structural Scattering Index
TOMS Total Ozone Mapping Spectrometer
UK United Kingdom of Great Britain and Northern Ireland
USDA United States Department of Agriculture
VCF Vegetation Continuous Fields (MODIS product)
VIIRS Visible/Infrared Imager/Radiometer Suite (NPOESS)
5 Terrestrial Applications of Multiangle Remote Sensing 139

References

Abuelgasim AA, Gopal S, Irons JR, Strahler AH (1996) Classification of ASAS multi-angle
and multispectral measurements using artificial neural networks. Remote Sens. Environ.
57(2):79–87
Asner GP (2000) Contributions of multi-view angle remote sensing to land surface and biogeo-
chemical research. Remote Sens. Rev. 18:137–162
Bach H, Begiebing S, Verhoef W (2005) Analyses of hyperspectral and directional CHRIS data for
monitoring of agricultural crops using the canopy reflectance model SLC. Proc. 9th Int. Symp.
on Physical Measurements and Signature in Remote Sens. (ISPMSRS 2005), vol. I, Int. Soc.
Photogrammetry and Remote Sens., 53–55
Bacour C, Bréon F-M (2005) Variability of biome reflectance directional signatures as seen by
POLDER. Remote Sens. Environ. 98:80–95
Bacour C, Jacquemoud S, Leroy M, Hautecoeur O, Weiss M, Prévot L, Bruguier N, Chauki H
(2002) Reliability of the estimation of vegetation characteristics by inversion of three canopy
reflectance models on airborne POLDER data. Agronomie 22:555–565
Barnsley MJ, Settle JJ, Cutter MA, Lobb DR, Teston F (2004) The PROBA/CHRIS mission: a low-
cost smallsat for hyperspectral multiangle observations of the Earth surface and atmosphere.
IEEE Trans. Geosci. Remote Sens. 42(7):1512–1520
Bicheron P, Leroy M, Hautecoeur O, Bréon F-M (1997) Enhanced discrimination of boreal forest
covers with directional reflectances from the airborne POLDER instrument. J. Geophys. Res.
102(D24):29517–29528
Bréon F-M, Maignan F, Leroy M, Grant I (2002) Analysis of hot spot directional signatures mea-
sured from space. J. Geophys. Res. 107(D16):10.1029/2001JD001094
Bruegge CJ, Chrien NL, Ando RR, Diner DJ, Abdou WA, Helmlinger MC, Pilorz SH, Thome
KJ (2002) Early validation of the Multi-angle Imaging SpectroRadiometer (MISR) radiometric
scale. IEEE Trans. Geosci. Remote Sens. 40(7):1477–1492
Camacho-de Coca F, Garcı́a-Haro FJ, Martı́nez B, Meliá J (2002) An operational strategy for
retrieval of vegetation products from SEVIRI and AVHRR-3 data. Proc. LSA SAF Training
Workshop, Ed. EUMETSAT, Darmstadt, Lisbon, July 8–10, pp 1–10
Camacho-de Coca F, Bréon FM, Leroy M, Garcia-Haro FJ (2004), Airborne measurement of hot
spot reflectance signatures. Remote Sens. Environ. 90:63–75
Canisius F, and Chen JM (2007) Retrieving vegetation background reflectance from Multi-angle
Imaging SpectroRadiometer (MISR) data. Remote Sens. Environ. 107(1–2):312–321
Casa R, Jones HG (2005) LAI retrieval from multiangular image classification and inversion of a
ray tracing model. Remote Sens. Environ. 98:414–428
Chen JM, Liu J, Leblanc SG, Lacaze R, Roujean J-L (2003) Multi-angular optical remote sensing
for assessing vegetation structure and carbon absorption. Remote Sens. Environ. 84:516–525
Chen JM, Menges CH, Leblanc SG (2005) Global mapping of foliage clumping index using multi-
angular satellite data. Remote Sens. Environ. 97:447–457
Chopping MJ (2006) Progress in retrieving canopy structure parameters from NASA multi-angle
remote sensing. Proc. Int. Geosci. Remote Sens. Symp. 2006 and 27th Canadian Remote Sens.
Symp., Denver, CO, July 31–August 4, 2006
Chopping M, Diner D (2005) Workshop on Ecological Modeling using NASA Multi-angle Remote
Sensing. The Earth Observer, NASA/GSFC, November–December 2005, 17(6):32–34
Chopping MJ, Rango A, Ritchie, JC (2002) Improved semi-arid community type differentiation
with the NOAA AVHRR via exploitation of the directional signal. IEEE Trans. Geosci. Remote
Sens. 40(5):1132–1149
Chopping MJ, Rango A, Havstad KM, Schiebe FR, Ritchie JC, Schmugge TJ, French A, Su L,
McKee L, Davis RM (2003) Canopy attributes of Chihuahuan Desert grassland and tran-
sition communities derived from multi-angular airborne imagery. Remote Sens. Environ.
85(3):339–354
140 M.J. Chopping

Chopping M, Martonchik JVM, Rango A, Peters DPC, Su L, Laliberte A (2005) Geometric-optical


modeling of desert grassland canopy structure with MISR. Proc. 9th Int. Symp. on Physical
Measurements and Signature in Remote Sens. (ISPMSRS 2005), vol. I, Int. Soc. Photogram-
metry and Remote Sens. 141–143
Chopping M, Su L, Rango A, Martonchik JV, Peters DPC, Laliberte A (2006a) Remote sensing
of woody shrub cover in desert grasslands using MISR with a geometric-optical canopy re-
flectance model, ISPMSRS’05. Special Issue of Remote Sens. Environ.
Chopping M, Su L, Laliberte A, Rango A, Peters DPC, Kollikathara N (2006b) Mapping shrub
abundance in desert grasslands using geometric-optical modeling and multiangle remote sens-
ing with CHRIS/Proba. Remote Sens. Environ. 104(1):62–73
Chopping, MJ, Su L, Laliberte A, Rango A, Peters DPC, Martonchik JV (2006c) Mapping woody
plant cover in desert grasslands using canopy reflectance modeling and MISR data. Geophys.
Res. Lett., 33, L17402, doi:10.1029/2006GL027148
Chopping M, Moisen G, Su L, Laliberte A, Rango A, Martonchik JV, and Peters DPC (2007) Large
area mapping of southwestern forest crown cover, canopy height, and biomass using the NASA
Multiangle Imaging SpectroRadiometer, Remote Sens. Environ, in press.
Combal B, Baret F, Weiss M (2002) Improving canopy variables estimation from remote sens-
ing data by exploiting ancillary information. Case study on sugar beet canopies. Agronomie
22:205–215
Di Girolamo L, Wilson MJ (2003) A first look at band-differenced angular signatures for cloud
detection from MISR. IEEE Trans. Geosci. Remote Sens. 41(7):1730–1734
Diner DJ, Beckert JC, Reilly TH, Bruegge CJ, Conel JE, Kahn RA, Martonchik JV, Ackerman TP,
Davies R, Gerstl SAW, Gordon HR, Muller J-P, Myneni RB, Sellers PJ, Pinty B, Verstraete MM
(1998) Multi-angle Imaging SpectroRadiometer (MISR) instrument description and experiment
overview. IEEE Trans. Geosci. Remote Sens. 36(4):1072–1087
Diner DJ, Asner GP, Davies R, Knyazikhin Y, Muller J-P, Nolin AW, Pinty B, Schaaf CB, Stroeve J
(1999) New directions in Earth observing: scientific applications of multi-angle remote sensing.
Bull. Am. Meteorol. Soc. 80(11):2209–2229
Diner DJ, Braswell BH, Davies R, Gobron N, Hu J, Jin Y, Kahn RA, Knyazikhin Y, Loeb N, Muller
J-P, Nolin AW, Pinty B, Schaaf CB, Seiz G, Stroeve J (2005) The value of multiangle measure-
ments for retrieving structurally and radiatively consistent properties of clouds, aerosols, and
surfaces. Remote Sens. Environ. 97:495–518
Disney M, Lewis P, Saich P (2006) 3D modelling of forest canopy structure for remote sensing
simulations in the optical and microwave domains. Remote Sens. Environ. 100:114–132
Engelsen O, Pinty B, Verstraete MM, Martonchik JV (1996) Parametric bidirectional reflectance
factor models: evaluation improvements and applications. Report EUR16426EN European
Commission Joint Research Centre Space Applications Institute, Ispra, Italy, 120pp
European Space Agency (1999) Exploitation of CHRIS data from the Proba mission for science
and applications. Experimenters’ Handbook Issue 4: Baseline Programme, Scientific Campaign
Unit, ESA-ESTEC
European Space Agency (2006) Achievements of the Proba mission: projects and publications at:
http://earth.esa.int/pub/ESA DOC/PROBA/ProjectsandPublications.html
Gao F, Schaaf CB, Strahler AH, Jin Y, Li X (2003) Detecting vegetation structure using a kernel-
based BRDF model. Remote Sens. Environ. 86:198–205
Garcı́a-Haro FJ, Camacho-de Coca F, Meliá J (2006) A directional spectral mixture analysis
method. Application to multiangular airborne measurements. IEEE Trans. Geosci. Remote
Sens. 44:365–377
Gemmell F (2000) Testing the utility of multi-angle spectral data for reducing the effects of
background spectral variations in forest reflectance model inversion. Remote Sens. Environ.
72:46–63
Gobron N, Pinty B, Verstraete MM, Govaerts Y (1997) A semi-discrete model for the scattering of
light by vegetation. J. Geophys. Res. 102:9431–9446
5 Terrestrial Applications of Multiangle Remote Sensing 141

Gobron N, Pinty B, Verstraete MM, Martonchik JV, Knyazikhin Y, Diner DJ (2000) Potential of
multiangular spectral measurements to characterize land surfaces: conceptual approach and
exploratory application. J. Geophys. Res. 105(D13):17539–17549, 10.1029/2000JD900154
Gobron N, Pinty B, Verstraete MM, Widlowski J-L, Diner DJ (2002) Uniqueness of multiangular
measurements part 2: joint retrieval of vegetation structure and photosynthetic activity from
MISR. IEEE Trans. Geosci. Remote Sens. 40:1574–1592
Hansen MC, DeFries RS, Townshend JRG, Sohlberg RA, Dimiceli C, Carroll M (2002) Towards an
operational MODIS continuous field of percent tree cover algorithm: examples using AVHRR
and MODIS data. Remote Sens. Environ. 83:303–319
Heiskanen J (2006) Tree cover and height estimation in the Fennoscandian tundra–taiga transition
zone using multiangular MISR data. Remote Sens. Environ. doi:10.1016/j.rse.2006.03.015
Hendriks JPM, Pellikka P (2004) Multiangular reflectance of a glacier surface. Presentation at the
2004 Joint Assembly (CGU, AGU, SEG & EEGS), Montréal, Canada, May 17–23, 2004
Hese S, Lucht W, Schmullius C, Barnsley M, Dubayah R, Knorr D, Neumann K, Riedel T, Schröter
K (2005) Global biomass mapping for an improved understanding of the CO2 balance – the
Earth observation mission Carbon-3D. Remote Sens. Environ. 94:94–104
Hu J, Tan B, Shabanov N, Crean KA, Martonchik JV, Diner DJ, Knyazikhin Y, Myneni RB (2003)
Performance of the MISR LAI and FPAR algorithm: a case study in Africa. Remote Sens.
Environ. 88:324–340
Hyman AH, and Barnsley MJ (1997) On the potential for land cover mapping from multiple view
angle (MVA) remotely-sensed images. In: Observations and Interactions: Proc. 23rd Annu.
Conf. Exhibition Remote Sens. Soc., 135–140
Jenkins J, Ollinger S, Braswell R, Martin M, Plourde L, Smith M-L, et al. (2004) Detecting patterns
of canopy structure and carbon uptake with multi-angle remote sensing. Presentation at MISR
Science Team Meeting, Pasadena, CA, December 9, 2004
Kimes DS, Ranson KJ, Sun G, Blair JB (2006) Predicting lidar measured forest vertical structure
from multi-angle spectral data. Remote Sens. Environ. 100(4):503–511
King MD, Kaufman YJ, Tanré D, Nakajima T (1999) Remote sensing of tropospheric aerosols
from space: past, present, and future. Bull. Am. Meteorol. Soc. 80(11):2229–2260
Knyazikhin Y, Kranigk J, Myneni RB, Panferov O, Gravenhorst G (1998a). Influence of small-
scale structure on radiative transfer and photosynthesis in vegetation cover. J. Geophys. Res.,
D: Atmospheres 103:6133–6144
Knyazikhin Y, Martonchik JV, Myneni RB, Diner DJ, Running SW (1998b) Synergistic algorithm
for estimating vegetation canopy leaf area index and fraction of absorbed photosynthetically
active radiation from MODIS and MISR data. J. Geophys. Res. 103(D24):32257–32276
Knyazikhin Y, Martonchik JV, Diner DJ, Myneni RB, Verstraete M, Pinty B, Gobron N (1998c)
Estimation of vegetation canopy leaf area index and fraction of absorbed photosynthetically ac-
tive radiation from atmosphere-corrected MISR data. J. Geophys. Res. 103(D24):32239–32256
Koetz B, Kneubühler M, Widlowski J-L, Morsdorf F, Schaepman M, Itten K (2005a) Assess-
ment of canopy structure and heterogeneity from multi-angular CHRIS-PROBA data. 9th Int.
Symp.on Physical Measurements and Signatures in Remote Sens. (ISPMSRS), Beijing, China,
2005, 76–78
Koetz B, Kneubühler M, Schopfer J, Schaepman M, Itten K (2005b) Influence of changing back-
ground on CHRIS/Proba data over an heterogeneous canopy. Proc. EARSeL 4th Workshop on
Imaging Spectroscopy, Warsaw, Poland, 1–9
Lacaze R, Chen JM, Roujean J-L, Leblanc SG (2002) Retrieval of vegetation clumping index using
hot spot signatures measured by POLDER instrument. Remote Sens. Environ. 79:84–95
Laurent B, Marticorena B, Bergametti G, Chazette P, Maignan F, Schmechtig C (2005) Simulation
of the mineral dust emission frequencies from desert areas of China and Mongolia using an
aerodynamic roughness length map derived from the POLDER/ADEOS 1 surface products. J.
Geophys. Res. 110(D18):D18S04, doi:10.1029/2004JD005013
Lavergne T, Kaminski T, Pinty B, Taberner M, Gobron N, Verstraete M, Vossbeck M, Widlowski
J-L, Giering R (2007) Application to MISR land products of an RPV model inversion package
using adjoint and Hessian codes. Remote Sens. Environ. 107(1–2):362–375
142 M.J. Chopping

Li X, Strahler AH (1985) Geometric-optical modeling of a conifer forest canopy. IEEE Trans.


Geosci. Remote Sens. 23:705–721
Li X, Strahler AH (1992) Geometric-optical bidirectional reflectance modeling of the discrete
crown vegetation canopy: effect of crown shape and mutual shadowing. IEEE Trans. Geosci.
Remote Sens. 30(2):276–291
Li X, Strahler AH, Woodcock CE (1995) A hybrid geometric optical-radiative transfer approach
for modeling albedo and directional reflectance of discontinuous canopies. IEEE Trans. Geosci.
Remote Sens. 33:466–480
Liang S (2004) Quantitative Remote Sensing of Land Surfaces. Wiley, Hoboken, NJ, 534 pp
Liang S, Strahler AH, Barnsley MJ, Borel CC, Gerstl SAW, Diner DJ, Prata AJ, Walthall CL (2000)
Multiangle remote sensing: past, present and future. In:, Liang S and Strahler AH (eds), The
Special Issue on Land Surface Bidirectional Reflectance Distribution Function (BRDF); Recent
Advances and Future Prospects, Remote Sens. Rev. 18:83–102
Liu X, Kafatos M (2007) MISR multi-angular spectral remote sensing for temperate forest mapping
at 1.1-km resolution. International Journal of Remote Sensing 28(2):459–464
Maignan F, Bréon F-M, and Lacaze R (2004) Bidirectional reflectance of Earth targets: evaluation
of analytical models using a large set of spaceborne measurements with emphasis on the hot
spot. Remote Sens. Environ. 90:210–220
Marticorena B, Chazette P, Bergametti G, Dulac F, Legrand M (2004) Mapping the aerodynamic
roughness length of desert surfaces from the POLDER/ADEOS bi-directional reflectance prod-
uct. Int. J. Remote Sens. 25:603–626
Martonchik JV, Bruegge CJ, Strahler A (2000) A review of reflectance nomenclature used in re-
mote sensing. Remote Sens. Rev. 19:9–20
NASA (2000) NASA Earth Science Enterprise Research Strategy for 2000–2010
Ni W, Li X (2000) A coupled vegetation – soil bidirectional reflectance model for a semi-arid
landscape. Remote Sens. Environ. 74:113–124
Ni-Meister W (2005) 3D vegetation structure extraction from lidar remote sensing. Proc. 9th Int.
Symp. on Physical Measurements and Signature in Remote Sens. (ISPMSRS 2005), vol. I, Int.
Soc. Photogrammetry and Remote Sens. 47–49
Nicodemus FE, Richmond JC, Hsia JJ, Ginsberg IW, Limperis T (1977) Geometrical Considera-
tions and Nomenclature for Reflectance. In: National Bureau of Standards, US Department of
Commerce, Washington, DC
Nolin AW (2004) Towards retrieval of forest cover density over snow from the Multi-angle Imaging
SpectroRadiometer (MISR). Hydrol. Process. 18:3623–3636
Nolin AW, Payne M (2006) Classification of glacier zones in Western Greenland using albedo and
surface roughness from the Multi-angle Imaging SpectroRadiometer (MISR). Remote Sens.
Environ. 107(1–2):264–275
Nolin AW, Fetterer FM, Scambos TA (2002) Surface roughness characterizations of sea ice and ice
sheets: case studies with MISR data. IEEE Trans. Geosci. Remote Sens. 40(7):1605–1615
Phinn S, Franklin J, Hope A, Stow D, Huenneke L (1996) Biomass distribution mapping using
airborne digital video imagery and spatial statistics in a semi-arid environment. J. Environ.
Manage. 47:139–164
Pinty B, Gobron N, Widlowski J-L, Gerstl SAW, Verstraete MM, Antunes M, Bacour C, Gascon F,
Gastellu J-P, Goel N, Jacquemoud S, North P, Qin W, Thompson R (2001) Radiation Transfer
Model Intercomparison (RAMI) exercise. J. Geophys. Res. 106:11937–11956
Pinty B, Widlowski J-L, Gobron N, Verstraete MM, Diner DJ (2002) Uniqueness of multiangular
measurements – Part I: an indicator of subpixel surface heterogeneity from MISR. IEEE Trans.
Geosci. Remote Sens. 40(7):1560–1573
Pinty B, Widlowski J-L, Taberner M, Gobron N, Verstraete MM, Disney M, Gascon F, Gastellu
J-P, Jiang L, Kuusk A, Lewis P, Li X, Ni-Meister W, Nilson T, North P, Qin W, Su L, Tang S,
Thompson R, Verhoef W, Wang H, Wang J, Yan G, Zang H (2004a) RAdiation transfer
Model Intercomparison (RAMI) exercise: results from the second phase. J. Geophys. Res. 109,
D06210 101029/2003JD004252
5 Terrestrial Applications of Multiangle Remote Sensing 143

Pinty B, Gobron N, Widlowski J-L, Lavergne T, Verstraete MM (2004b) Synergy between 1-D
and 3-D radiation transfer models to retrieve vegetation canopy properties from remote sensing
data. J. Geophys. Res. 109, D21205
Privette JL, Deering D, Wickland D (1997) Report on the workshop on multiangular remote sens-
ing for environmental applications. NASA Technical Memorandum 113202, 54 pp.
Rahman H, Pinty B, Verstraete MM (1993) Coupled surface-atmosphere reflectance (CSAR)
model. 2. Semiempirical surface model usable with NOAA Advanced Very High Resolution
Radiometer data. J. Geophys. Res. 98:20791–20801
Roujean J-L, Leroy M, Deschamps P-Y (1992) A bidirectional reflectance model of the Earth’s
surface for the correction of remote sensing data. J. Geophys. Res. 97:20455–20468
Roujean J-L, Tanré D, Bréon F-M, Deuzé J-L (1997) Retrieval of land surface parameters from air-
borne POLDER bidirectional reflectance distribution function during HAPEX-Sahel. J. Geo-
phys. Res. 102:11201–11218
Sandmeier S, Deering DW (1999) Structure analysis and classification of boreal forests using air-
borne hyperspectral BRDF data from ASAS. Remote Sens. Environ. 69(3):281–295
Scarth P, Phinn S (2000) Determining forest structural attributes using an inverted geometric-
optical model in mixed eucalypt forests, Southeast Queensland, Australia. Remote Sens. Envi-
ron. 71:141–157
Schaepman M (2006) Spectrodirectional remote sensing: from pixels to processes. Int. J. Appl.
Earth Obs. Geoinformation 6(3–4):271–282
Schaepman-Strub G, Schaepman M, Painter T, Dangel S, Martonchik J (2006) Reflectance
quantities in optical remote sensing – definitions and case studies. Remote Sens. Environ.
103(1):27–42
Smolander S, Stenberg P (2005) Simple parameterizations of the radiation budget of uniform
broadleaved and coniferous canopies. Remote Sens. Environ. 94(3):355–363
Strahler AH, Jupp DLB, Woodcock CE, Li X (2005) The discrete-object scene model and its
application in remote sensing. Proc. 9th International Symp. on Physical Measurements and
Signatures in Remote Sens. 1:166–168
Su L, Chopping MJ, Rango A, Martonchik JV, Peters DPC (2007) Support vector machines for
recognition of semi-arid vegetation types using MISR multi-angle imagery. Remote Sens.
Environ. 107(1–2):299–311
Verhoef W (2005) Earth observation model sensitivity analysis to assess mission performances
in terms of geo-biophysical variable retrieval accuracies. Proc. 9th Int. Symp. on Physical
Measurements and Signature in Remote Sens. (ISPMSRS 2005), Beijing, October 17–19, 2005,
vol. I, Int. Soc. Photogrammetry and Remote Sens. 324–327
Wanner W, Li X, Strahler AH (1995) On the derivation of kernels for kernel-driven models of
bidirectional reflectance. J. Geophys. Res. 100:21077–21090
Weiss M, Baret F, Myneni R, Pragnere A, Knyazikhin Y (2000) Investigation of a model inversion
technique for the estimation of crop characteristics from spectral and directional reflectance
data. Agronomie 20:3–22
Widlowski J-L, Pinty B, Gobron N, Verstraete MM, Davis AB (2001) Characterization of
surface heterogeneity detected at the MISR/TERRA subpixel scale. Geophys. Res. Lett.
28(24):4639–4642
Widlowski J-L, Pinty B, Gobron N, Verstraete M, Diner DJ, Davis B (2004) Canopy structure pa-
rameters derived from multi-angular remote sensing data for terrestrial carbon studies. Climatic
Change 67:403–415
Widlowski J-L, Pinty B, Lavergne T, Verstraete MM, Gobron N (2005) Using 1-D models to
interpret the reflectance anisotropy of 3-D canopy targets: issues and caveats. IEEE Trans.
Geosci. Remote Sens. 43:2008–2017, doi10.1109/TGRS.2005.853718
Widlowski J-L, Taberner M, Pinty B, Bruniquel-Pinel V, Disney M, Fernandes R, Gastellu-
Etchegorry J-P, Gobron N, Kuusk A, Lavergne T, Leblanc S, Lewis P, Martin E, Mottus M,
North PJR, Qin W, Robustelli M, Rochdi N, Ruiloba R, Soler C, Thompson R, Verhoef W,
Verstraete MM, Xie D (2006) The third RAdiation transfer Model Intercomparison
(RAMI) exercise: documenting progress in canopy reflectance modelling. J. Geophys. Res.
112(D9):D09111. doi:10.1029/2006JD007821
144 M.J. Chopping

Wiscombe W (2000) Formation Flying, Leonardo, Balloon-ERB and Triana: Forays to the Frontier,
NASA/NIAC workshop presentation, November 7, 2000, online at http://www.niac.usra.
edu/files/library/meetings/misc/nov00 wks/wwiscombe.pdf, last access 06/14/06
Zhang X, Schaaf CB, Gao F, Friedl MA, Strahler AH, Hodges JCF (2000) Mapping land cover and
green vegetation abundance using MODIS-like data: a case study of New England. Proc. Int.
Geosci. Remote Sens. Symp. 2000: vol 5, pp 2005–2007
Chapter 6
Modeling the Spectral Signature of Forests:
Application of Remote Sensing Models
to Coniferous Canopies

Pauline Stenberg, Matti Mõttus, and Miina Rautiainen

Abstract The vertical and horizontal structure of forest canopies is one of the most
important driving factors of various ecosystem processes and has received increas-
ing attention during the past 20 years and served as an impetus for earth observation
missions. In the remote sensing community, the variables which describe canopy
structure are called biophysical variables, and are directly coupled with the fun-
damental physical problem behind remote sensing: radiative transfer in vegetation.
There are basically three different approaches to interpreting biophysical variables
from remotely sensed data: (1) empirical, (2) physically based, and (3) various com-
binations of them. The physical approach builds upon an understanding of the phys-
ical laws governing the transfer of solar radiation in vegetative canopies, and for-
mulates it mathematically by canopy reflectance models which relate the spectral
signal to biophysical properties of the vegetation. In this chapter, we will first out-
line the basic principles and existing physically based model types for simulating
the spectral signature of forests. After this, the focus is on the specific issues related
to applying these models to the complex 3D structure of coniferous canopies.

6.1 Introduction

The assessment of many fundamental ecological questions at global scale is possible


only through remote sensing, since integrated analyses of the biosphere, atmosphere
and hydrosphere require simultaneous measurements over large areas.

Pauline Stenberg
Department of Forest Resource Management, University of Helsinki, Finland
Pauline.Stenberg@helsinki.fi
Miina Rautiainen
Department of Forest Resource Management, University of Helsinki, Finland
Matti Mõttus
Tartu Observatory, Tõravere, Tartumaa, Estonia

S. Liang (ed.), Advances in Land Remote Sensing, 147–171. 147


c Springer Science + Business Media B.V., 2008
148 P. Stenberg et al.

The three-dimensional (3D) structure of forest canopies is amongst the most im-
portant driving factors of various physiological and ecological processes. By an-
alyzing the 3D structure of canopies, it is possible to detect growth patterns and
phenological cycles of forests, as well as other changes such as insect or fire out-
breaks and damages, illegal deforestation, and environmental stresses caused by
drought or pollution. In the remote sensing community, the variables which de-
scribe canopy structure are called biophysical variables. Biophysical variables are
coupled with the radiative transfer problem (Chandrasekhar, 1960) in vegetation
(Ross, 1981) and can also be defined as state variables of the radiative transfer prob-
lem, in other words, the smallest set of variables which are needed to fully describe
the physical state of the system at a given scale (Verstraete et al., 1996). However,
mere description of the structure of the canopy through biophysical variables is not
the only goal: the variables should be applicable for end-users in, for example, water
and carbon cycle or climate modeling.
It is obvious that measuring biophysical variables in situ is both laborious and
time-consuming in forests – and impossible at global scale within a short time frame.
Space-born monitoring is thus required, and algorithms for interpreting these vari-
ables from remotely sensed data need to be developed. Most common remotely
sensed biophysical variables of forests include leaf area index (LAI), fraction of
photosynthetically active radiation (400–700 nm) absorbed by vegetation (fPAR)
and fraction of canopy cover (fCover). LAI and fCover are geometrical variables
which are related to canopy gap fraction, i.e., the fraction of ground seen in a given
direction (Nilson, 1977) – canopy gap fraction is in fact determined by LAI and
its spatial distribution and leaf inclination distribution. fPAR, on the other hand, is
an outcome of radiative transfer in vegetation, and the opposite of reflectance from
vegetation. In addition, there are biophysical variables which are not geometric, but
instead influence the spectral properties of scattering elements. An example of such
variables is the chlorophyll content of green leaves (which also is closely connected
to the nitrogen content of foliage). However, there is evidence that no clear distinc-
tion between foliar biochemistry and LAI can be made in practical remote sensing
(Yoder and Pettigrew-Crosby, 1995).
There are basically three different approaches: (1) statistical (empirical) and (2)
physically based, and (3) various combinations of them (e.g., neural networks), to
assess biophysical variables from spectral signals provided by optical satellite im-
ages. In the empirical approaches, commonly used in, for example, regional or na-
tional forest inventories, the vegetation characteristics of interest are estimated based
on statistical relationships (regressions) obtained by collecting training data on the
spectral signatures of a variety of objects. These methods are limited to a specified
viewing and illumination configuration, and require large sets of reliable ground
truth data. The physical approach, in contrast, builds upon an understanding of the
physical laws governing the transfer of solar radiation in vegetative canopies and
formulating it mathematically by reflectance models. Reflectance models, in turn,
relate biophysical properties of the vegetation, or sets of canopy and stand parame-
ters, to the spectral signal. Physically based methods are, at least in theory, more
robust since they are not limited to a single configuration or vegetation biome type.
6 Modeling the Spectral Signature of Forests 149

Physically based methods have progressively become more and more attractive
for the assessment or monitoring of biophysical characteristics of vegetation. They
are the only methods which can fully take into use the versatile hyperspectral and
multiangular information provided by the modern satellite sensors, and they are
better suited for many current large scale applications than the empirical techniques.
Several aspects must be considered when physically based models are developed
for operational monitoring purposes. To begin with, the sensitivity of the model
to the biophysical variable to be retrieved should be maximized. Simultaneously,
undesired effects, for example from the atmosphere or terrain topography, should be
eliminated as efficiently as possible using other available models and data. Finally,
when considering the optimal model, the larger the area (e.g., the whole Earth), the
more the model should rely on theory and the less on local data bases for input as the
sets are usually not compatible from one region or country to another. In addition,
we should take into account technical questions such as computation time required
by the model.
Canopy reflectance models were originally developed for agricultural crops, fol-
lowed by broad-leaved forests, and even though presently there exist a number
of reflectance models designed to be applicable also to conifers, some missing
“difficult-to-model” properties and insufficient empirical data on some key para-
meters have still limited their optimal use. Several recent studies (Rautiainen and
Stenberg, 2005a; Smolander and Stenberg, 2005) give quantitative support to the
hypothesis put forward a long time ago (Norman and Jarvis, 1975; J. Ross, 1981–
2002, personal communication), that a major reason for the distinct spectral signa-
ture of coniferous forests lies in their hierarchical grouped structure, which governs
the processes of multiple scattering within the canopy, as well as canopy absorption.
Moreover, modeling tools are emerging by which grouping at different scales can be
accounted for in canopy reflectance calculations. The effect of grouping on canopy
PAR absorption and photosynthesis has been a well studied subject in forest pro-
duction ecology (Oker-Blom, 1986; Leverenz and Hinckley, 1990; Wang and Jarvis,
1990; Oker-Blom et al., 1991; Nilson, 1992; Cescatti, 1997a) but only more recently,
after becoming aware of the specific problems related to interpretation of satellite
images over the boreal zone, has modeling the radiation regime of coniferous forests
received increasing attention in the remote sensing community. The boreal zone
spreads through Fennoscandia, Siberia, Alaska and Canada, and hosts a multitude
of coniferous tree species adapted to the cold and drought climate conditions of
the region. The forests are typically rather open with dense crowns (consisting of up
to a few million needles per crown) and an abundant green understory and moss or
lichen layer. The documented complex structure of the forests is further complicated
by the fact that acquiring ground observations from many parts of the boreal zone is
especially difficult due to the remoteness and climate of the region. For application
of forest reflectance models in the boreal zone, remaining problems of high priority
are thus how to model efficiently and sufficiently realistically (1) the hierarchic 3D
canopy structure and (2) the contribution to the remotely sensed spectral signal from
the background, typically composed of mixed green understory, and (3) how to sepa-
rate the signals of the forest canopy layer and the understory from each other for cor-
rect interpretation of tree canopy biophysical variables from optical satellite images.
150 P. Stenberg et al.

In this chapter, we will first outline the basic principles and current status of
existing models for simulating the spectral signature of forests. After this, we will
focus on the specific issues related to applying these models in the boreal zone,
including how to collect the required input data. Finally, more practical issues, such
as application of ground truth data, model scaling and validation, will be addressed.

6.2 Modeling the Spectral Signature of Forests

The basic premises for optical remote sensing of vegetation are that the solar ra-
diation received by a remotely located sensor (e.g., on a satellite) upon interaction
with the vegetation canopy carries in it the signature of the canopy, and that this
spectral signature can be deciphered to obtain the information of interest (Goel,
1989). Satellite-borne sensors measure the mean intensities of radiation at differ-
ent wavelengths emanating from a target on Earth. Correct interpretation of these
measurements to yield the biophysical variables of interest requires an accurate
specification of the relationships between these variables and the canopy leaving
radiation field. This specification is quantified by canopy reflectance models. The
models are parameterized using mathematical descriptions of canopy structure to-
gether with optical properties of the plant elements and the underlying surface to
produce spectral signatures of canopy leaving radiation. In addition, the spectral
and angular properties of the incoming radiation and the receiving sensor need to be
specified (Table 6.1). Finally, the spectral signature depends on the resolution of the
measurement (i.e., pixel size), which needs to be considered in defining the spatial
scale of the model.

Table 6.1 Variables affecting the spectral signature of a forest (excluding terrain topography and
atmosphere)
Sensor Illumination Forest tree layer Understory and soil
canopy

Zenith and azimuth Angles of incidence Macroscale structure Geometrical structure


viewing angles and azimuth (distribution, size (amount, distribution,
and shape of tree size and shape of
crowns) understory plants)
Wavelength band Wavelength Microscale structure Spectral properties
(i.e., spectral (distribution, size of understory plants
sensitivity) and shape of leaves,
needles, shoots and
branches)
Resolution Other structural Soil optical
elements (e.g., properties (influenced
distribution, size and by e.g., soil moisture
shape of tree trunks and texture)
and branches)
Spectral properties
of all the canopy
elements
6 Modeling the Spectral Signature of Forests 151

The mathematical descriptions of the process of radiative transfer in vegetation


canopies cover a range of different approaches, starting from the traditional turbid
medium approach to geometric-optical and hybrid models, the main difference be-
ing the degree of detail at which canopy structure at different hierarchical levels is
described (Table 6.1).
The turbid medium approximation, introduced to modeling the radiation regime
inside vegetation by Ross (1981), is based on the classical radiative transfer theory
(Chandrasekhar, 1960). The radiative transfer theory has been successfully applied
to predicting the distribution of shortwave (optical and infrared) photons in several
media like atmospheres, stellar dust or water. Radiative transfer theory is, in princi-
ple, the law of conservation of energy written out for beams of radiation traveling
in all possible directions in a volume filled with optically active, i.e., absorbing and
scattering, material. It is thus a function of five coordinates: three spatial coordi-
nates determining the elementary volume for which the energy conservation law is
applied, and two variables defining the direction of the radiation beam. In the clas-
sical radiative transfer theory, the fate of a photon is determined by the probability
of being absorbed or scattered while traveling a unit distance (these probabilities
are called the absorption and scattering cross-sections, respectively). Additionally,
interactions between different rays for each point are described by the probability
distribution of directions the photon can be scattered in, measured relative to the
photon’s traveling direction before scattering. This is called the photon scattering
phase function. The equation of radiative transfer merely states that the number of
photons exiting each elementary volume equals the number of photons entering the
volume minus the number of absorbed photons, written out in a correct mathemati-
cal form. More generally, a source term has to be added for radiating media, e.g., in-
side stars or for the thermal infrared radiation in atmosphere. In the optical region
under consideration in this chapter, no radiation is usually emitted by the medium.
The difference between the classical radiative transfer theory and the turbid
medium approach lies in the treatment of the directional properties of the medium:
in the classical radiative transfer theory, the scattering phase function and the ab-
sorption and scattering cross-sections are assumed to be independent of the direction
of photon travel. This assumption does not hold for the transfer of radiation inside
a vegetation canopy. Thus, in the turbid medium approximation, plant cover is de-
scribed as consisting of geometric elements which usually have a preferred orienta-
tion described, in turn, by a directional attribute. Now, the absorption and scattering
cross-sections become functions of the direction of photon travel inside the canopy
as well as of the actual shape of the elementary units (phytoelements) constituting
the canopy. For example, in a canopy of preferably horizontal flat leaves, the prob-
ability of hitting a leaf while traveling a distance unit is much larger for a photon
moving in the vertical direction than for one moving in the horizontal direction.
Consequently, the interaction cross-section (sum of absorption and scattering cross-
sections) in such a canopy is larger in the vertical direction than in the horizontal.
Orientation of plant elements is described by statistical distributions. For example,
the orientation of leaves is described by the direction (inclination and azimuth) of
their normal and that of stems, branches and needles by the direction of their axis.
152 P. Stenberg et al.

A number of different distributions suitable for describing the distribution of leaf


normals of actual canopies are given in, e.g., Weiss et al. (2004). Cross-sections can be
calculated from these distributions by integrating over all possible orientation angles.
Other than implying that the scattering phase function and the cross-sections de-
pend on direction within the canopy, the turbid media approach leaves the classical
radiative transfer equations intact: canopy elements are still of infinitesimal size
and do not cast shadows, i.e., the “far-field approximation” still holds. Although
anisotropy (or dependence on photon travel direction) makes finding a solution to
the radiative transfer equation more difficult and yields some traditional methods
unpractical, a plethora of techniques is still available for finding an approximate
solution. The basic equations for the radiation field can be found, for example, in
the book by Ross (1981) or various other sources (e.g., Myneni et al., 1989).
The crudest (and fastest) way of solving the radiative transfer problem for a
plane-parallel turbid media (or in “slab geometry”) is by using the two-stream
(Kubelka-Munk) equations. Only two directions, up and down, are used to describe
the radiation field inside a horizontally homogeneous and infinite medium and the
solutions can be found analytically. In its simplest form, the two stream model as-
sumes isotropic distribution of both up- and downward traveling radiation, i.e., the
intensity does not depend on view angle inside both hemispheres. This condition
is clearly violated if direct solar radiation is present. To overcome the problem, the
contribution of the direct beam is separated, thus creating an additional source for
the remaining more isotropic diffuse radiation field. This technique is almost uni-
versally employed in all algorithms for solving the radiative transfer problem in
plant canopies. This is because besides its extremely anisotropic character, direct
illumination gives rise to the so-called hot-spot phenomenon (see discussion later in
this chapter) that needs to be accounted for if measurements are taken near the back-
scattering direction. The solutions of the two-stream equation are analytical and thus
extremely computer-efficient. It should also be noted that the two-stream equations
cease to be approximations for a (theoretical) canopy of infinitesimally small hori-
zontal ideally scattering leaves but, instead, give the exact analytical solution as the
reflected and scattered radiation field inside such a canopy is isotropic.
A classical example of an application of the two-stream approximation is the
SAIL (Scattering and Arbitrarily Inclined Leaves) model (Verhoef, 1984, 1985).
Besides the two fluxes, the model calculates intensities for two other directions: the
solar direction (for reasons discussed above) and the view direction. This enables to
take into account anisotropy in the distribution of reflected radiation. The absorption
and scattering coefficients, and the scattering phase function are calculated from a
distribution function of leaf inclination angles. Despite its age, the model is still used
for remote sensing purposes, although sometimes modified to fit the particular needs
or encapsulated in other models that include a higher level description of canopy
structure (Kuusk, 1995; Kuusk, 2001; Andrieu et al., 1997). The SAIL model is
reasonably accurate, and can take the competition from much more sophisticated
3D models, in the case of relatively homogeneous canopy types (crops, grasslands)
that have a small number of structural parameters (Jacquemoud et al., 1995; Koetz
et al., 2005; Meroni et al., 2004; Andrieu et al., 1997; Weiss and Baret, 1999; Zarco-
Tejada et al., 2003).
6 Modeling the Spectral Signature of Forests 153

If the media cannot be considered homogeneous in the horizontal directions, as is


the case for many natural vegetation canopies, the radiative transfer equation can be
solved in three dimensions (Myneni, 1990; Knyazikhin and Marshak, 1991). This
approach is useful for highly structured scenes where the two-stream model cannot
be adequately parameterized to include canopy heterogeneity. Once such a scene
is generated based on measured data or a forest structure model, the canopy is di-
vided into cells and the equation (or, after discretization, a set of coupled integro-
differential equations) is solved numerically using the discrete ordinates iterative
approach on the grid (Myneni, 1990; Gastellu-Etchegorry et al., 1996; Knyazikhin
et al., 1998a). This method is based on the calculation of radiation intensities in
discretized directions. In each iteration, the contribution of radiation scattered to
every discrete direction from all other directions is calculated for each canopy cell
using a pre-calculated scattering matrix until the intensities converge to a solution.
Naturally, this method, besides its mathematical complexity (which is further in-
creased by using various convergence acceleration techniques), also requires a real-
istic description of the scene and reasonable cell-size to exclude shadowing effects.
The method is computer-intensive but the results can be considered accurate as this
method is based on the physical principles of radiative transfer.
Another common method used for calculating radiation reflected by a canopy, for
which a complete 3D description of the scene is given, is the Monte Carlo method.
As the radiation field above plant canopy is composed of small contributions by
individual photons, the angular and spatial variation in the intensity of reflected
radiation can be viewed as a distribution function describing the possible exit direc-
tions and locations of light quanta. Monte Carlo methods (or ray tracing algorithms)
are based on random sampling of this distribution. If a sufficient number of photons
are traced, the counts of photons exiting in each direction are relatively close to the
actual directional distribution of reflected radiation, although some numerical noise
is inevitable. Although in studies of radiative transfer in canopies, the term “Monte
Carlo” is commonly used to denote tracing the trajectories of photons in a realis-
tic detailed 3D canopy scene (Gerstl et al., 1986; Ross and Marshak, 1988; North,
1996; Chelle and Andrieu, 1998; Govaerts and Verstraete, 1998; Thompson and
Goel, 1998; Disney et al., 2006), Monte Carlo calculations can also be used to trace
photons in a 3D turbid media with varying optical properties, like the geometric-
optical approximation described below (Gerard and North, 1997; Garcı́a-Haro and
Sommer, 2002). An overview of the Monte Carlo modeling approach is given by
Disney et al. (2000).
If a 3D description of a scene is given, the reflected radiation field can also be
calculated using the radiosity principle (Borel et al., 1991; Gerstl and Borel, 1992;
Goel et al., 1991; Chelle and Andrieu, 1998; Garcı́a-Haro et al., 1999; Qin et al.,
2002; Soler et al., 2003). Radiation field inside the canopy is calculated using view
factors for individual elements: if the illumination conditions of a particular canopy
element are given, its brightness as viewed from the positions of all other elements
can be calculated based on its geometry and scattering properties. The brightness of
an object is a well-defined radiometric quantity that can be used to calculate the par-
tial flux contributed by the object. Thus, if the magnitude and angular distribution
154 P. Stenberg et al.

of radiation field incident on a canopy element can be calculated from brightness


of all other canopy elements and the intensity of the incident unscattered radiation,
the brightness of the element as viewed from any point inside or above the canopy
can be calculated. It is an iterative process complicated by the large number and
mutual shading of elements present in a natural canopy. From the computing per-
spective, the most time-consuming task is calculating these view factors or solving
the geometrical relations between each pair of foliage elements. Once these fac-
tors are known, calculations for different wavelengths for which only the scattering
properties of elements are different, but not the geometry of the canopy as a whole
are relatively fast.
In the classical radiative transfer approach, the optical properties of the medium
are assumed to change continuously (for the 3D case, this assumption is valid inside
each cell). This is naturally not the case for actual vegetation canopies. The Monte
Carlo and radiosity methods that describe the canopy as consisting of separate solid
objects remove this restriction and take into account the mutual shading of elements
(remove the far-field approximation). The assumption of discontinuities in canopy
optical properties, or the existence of voids between scatterers, can also be inserted
into the radiative transfer equation. The foliage are a density can be described using
the statistical distribution of an indicator function, in other words, a function of the
spatial coordinates that equals unity only if a scattering element is present at the
point described by the coordinates. This leads to the stochastic radiative transfer
equation in plant canopies (Menzhulin and Anisimov, 1991; Shabanov et al., 2000).
Besides the mean reflecting properties of a canopy, this theory allows to calculate
its statistical moments.
Often the terms deterministic (“non-stochastic”) and stochastic are used as syn-
onyms to 1D and 3D models, respectively. However, with the exception of Monte
Carlo models, most of the existing canopy radiation/reflectance models, in fact use a
non-stochastic approach in computing the spectral signatures. That is, even though
the canopy structure may be described using statistical distributions, the canopy
radiation field is solved for a mean realization (with average characteristics) of
canopy structure. A truly stochastic approach, in contrast, would be to evaluate the
3D canopy radiation field for all possible statistical realizations of the canopy, and
then average the corresponding radiation fields to obtain the ensemble average sig-
nature (Shabanov et al., 2000). The stochastic and non-stochastic approaches can
result in different relationships between mean characteristics of canopy structure
and canopy-leaving radiation.
Sometimes, a more abstract description of the 3D structure of a natural forest
canopy is required. The exact locations of all canopy elements are rarely known
or reliably predicted for a forest canopy. Therefore, instead of describing the struc-
ture of a canopy by specifying the locations of the smallest canopy elements (e.g.,
leaves, needles or shoots), the canopy can be first divided into larger subunits. An
obvious choice for the crudest first-level division would be the tree crown. The
canopy can then be viewed as a congregation of geometrical tree crown envelopes
with gaps between them. The locations of the tree crowns are described by a sta-
tistical distribution, thus accounting for mutual shadowing and the distribution of
6 Modeling the Spectral Signature of Forests 155

between-crown gaps. Depending on the accuracy by which the processes of radia-


tive transfer are described in these models, the models are commonly categorized as
geometric-optical or hybrid models, the term “hybrid” being a shortcut for “hybrid
radiative transfer/geometric-optical model”.
In the simplest form, a geometric-optical canopy reflectance model considers
just the effect of the geometrical structure of a forest on the remotely sensed image
(Li and Strahler, 1985, 1986, 1992; Welles and Norman, 1991; Chen and Leblanc,
1997). The image is supposed to consist of (sub-pixel-sized) regions of differ-
ent brightness: directly illuminated tree crowns and ground, and shadowed tree
crowns and ground. The fractions of these components are calculated using given
illumination and view angles and mutual shadowing derived from geometrical
considerations. These types of models can accurately account for only first-order
scattered radiation, i.e., photons that reach the sensor having interacted only once
with the geometric shapes comprising the canopy. Diffuse radiation is included
using correction factors. This makes it difficult to use these models for different
wavelengths where the optical properties of canopy elements are not similar. The
share of first-order scattering can be calculated straightforwardly: this reflectance
component, besides depending on canopy geometry, which is the same for all wave-
lengths, is a linear function of leaf single scattering albedo (leaf reflectance plus
transmittance). Diffuse photons, on the other hand, interact with leaves (or other
canopy elements) several times and the possible number of interaction is itself a
function of leaf albedo at the specified wavelength. This makes canopy diffuse
reflectance strongly nonlinear with leaf albedo, and results in highly wavelength-
specific diffuse radiation correction factors.
Including the diffuse radiation field into the geometric-optical model leads to
a hybrid radiative transfer/geometric-optical model (Nilson and Peterson, 1991;
Li and Strahler, 1995; Ni et al., 1999; Atzberger, 2000; Chen and Leblanc, 2001;
Huemmrich, 2001; Kuusk and Nilson, 2000; Peddle et al., 2004). The methods
to include diffuse radiation can range from exact solutions (similar to those used
in solving the 3D radiative transfer) to tracing the photons in the canopy com-
prised of crown envelopes seen as large chunks of turbid media (similar to the
Monte Carlo approach). However, to maintain the high efficiency achieved by de-
lineating the canopy into abstract crowns, a simple and fast approximation is often
used. Now, the problem of finding reflected radiation intensity is divided into two
sub-problems: (1) finding the first-order scattering component using a geometrical
figure filled with absorbing and scattering foliage elements illuminated by a beam
of direct radiation, taking into account mutual shadowing by other semi-transparent
crowns, and (2) calculating the share of diffuse radiation using a simpler (e.g., two-
stream) solution of the radiative transfer equation. Naturally, in calculating both
components one has to consider the effect of a partially reflecting ground surface or
undergrowth. Joining the two sub-models is not an easy task: the multiple-scattering
component has to be parameterized to include the effect of canopy structure so that
conservation of energy and correct partitioning between first- and higher order
scattering are maintained.
156 P. Stenberg et al.

Hybrid models are an efficient and flexible tool for describing the radiation
regime of complex forests. The effect of different crown shapes can be included
by using various geometrical shapes: ellipsoids and cones of different proportions
are the most commonly used. Also, higher-order clumping can be introduced by
further dividing the canopy into abstract objects and adding their contributions to
total reflectance (e.g., Smolander and Stenberg, 2005). However, most models still
assume that the crown envelopes are uniformly filled with scatterers, although from
biological considerations and biometrical measurements it is known not to be true.
Using more realistic foliage distributions would considerably affect the transparency
of the crowns by introducing more clumping at both branch and whorl scale or
by locating photosynthesizing material close to the crown edge. While accounting
for higher-order structure usually leads to an increase in predicted transmittance if
the higher-order objects are distributed randomly, a regular distribution of foliage
clumps that has been suggested for natural canopies (Cescatti, 1997b) may diminish
or even reverse the effect.
The choice of the model to be used for solving a problem involving calculation
of canopy reflectance or radiation balance clearly depends on many factors. The
amount of required computer resources, manpower and time is clearly different for
the different approaches. Also, the optimal choice depends on the object under in-
vestigation: for relatively homogeneous canopies (crops, grasslands), a two-stream
model give very good results. Sometimes, if canopy reflectance calculations form
only a small contribution to a larger problem under investigation that depends on
many variables with possibly high uncertainties, the use of a simpler model is justi-
fied. This is clearly not the case for highly structured vegetation covers like shrub-
lands or boreal coniferous forests where the canopy upper surface is not flat, and
mutual shading and between-crown transmittance between tree crowns have to be
taken into account. Also, the amount of a priori knowledge can be a limiting fac-
tor when constructing a detailed 3D description of a canopy. In this case, a less
detailed approach with a smaller number of input parameters might be preferred,
e.g., a hybrid geometric-optical/radiative transfer model.
Another characteristic of the reflecting properties of a medium that contains
finites objects filling a three-dimensional volume is the hot spot phenomenon. “Hot-
spot” is a bright area in a remotely sensed image opposite to the source of illumi-
nation caused by a lack of shadows in the exact backscattering direction. The width
of this reflectance peak (or the size of the brighter area in an image) depends on the
geometric properties of the reflecting medium. Based on their working principles,
geometric-optical models take this phenomenon into account, but only partly. When
looking from the direction of illumination, no shaded crowns can be seen. This re-
sults in an increase in the predicted reflectance. Yet, the hot-spot phenomenon is also
produced at a finer level (e.g., leaves), which in turn leads to a distinct anisotropy
of the brightness of tree crowns. In more sophisticated reflectance models, this leaf-
level hot-spot is added to the wider hot spot created by crown structure. Also, a
hot-spot correction can be added to models based on the radiative transfer equation
(e.g., Gerstl et al., 1986; Marshak, 1989; Verstraete et al., 1990; Jupp and Strahler,
1991; Kuusk, 1991). The correction factor is semi-empirical due to the complex
6 Modeling the Spectral Signature of Forests 157

structure of vegetation canopies. The phenomenon itself is difficult to measure ac-


curately since it tends to be in the shadow of the sensor. Due to common view angle
configurations, it is also not registered by most remote sensing instruments.
Apart from the choice of radiation transfer model and availability of adequate
computing resources, quality of both structural and spectral input data has a large
impact on model performance. Today, good-quality data of either sorts are scarce as
they are difficult to measure and have a high natural variance. Moreover, as develop-
ing accurate radiative transfer models for vegetation canopies is a work in progress,
a ready-made solution may very likely not yet exist. However, the models presented
here (or, a quite up-to-date performance comparison can be found from RAMI web-
site (http://rami-benchmark.jrc.it) (Pinty et al., 2004b)) cover a large number of ap-
proaches which, if adequate input is provided in terms of both measurement data
and dedication, can be applied to a wide variety of problems. More information on
the subject can be found, in addition to the works referred to in this section, from
previously published reviews and textbooks (e.g., Myneni et al., 1991; Liang and
Strahler, 2000).
Canopy reflectance models can be constructed in many ways; some can be run in
the forward mode, some both in the forward and inverse modes. In model inversion
(discussed thoroughly in another chapter of this volume), radiation measurements
are converted into variables of interest characterizing the target. In remote sens-
ing of forests, invertible models can be used to infer biophysical variables from
reflectances (or back-scattering) registered by the remote air-or satellite-borne sen-
sors. Commonly used methods for the inverse estimation of vegetation characteris-
tics from satellite images include (1) comparing the observed signal to a database of
previously computed spectral signals for a wide selection of different canopies and
choosing the closest matches (look-up tables, LUT) (Knyazikhin et al., 1998b), or
(2) iteratively optimizing model input parameters to match the observed signal as
closely as possible using different optimization routines (Kuusk and Nilson, 2000).
Needless to say, the goodness of the estimates depends crucially on how realistic the
model is. Another basic requirement for successful estimation is of course that the
vegetation characteristic in question has a detectable influence on the spectral signal,
and that measurement errors can be corrected for. Even so, a remaining limitation is
the well known fact that the inversion problem is ill-posed: no unique solution exists
but different combinations of input parameters produce the same spectral signal. To
be able to solve the inverse problem, in other words, to reduce the array of possible
solutions to one solution, we need to acquire information from outside the problem
itself. Technical or mathematical advances do not remove the underlying ambiguity
of the ill-posed nature of the inversion problem in remote sensing. Another central
problem in inversion is scaling: the forest reflectance models assume that the given
forest structure continues infinitely. However, often in the case coarse and medium
resolution satellite images, the vegetation (forested area) may cover only part of
the pixel. At best, thus, the “averaged solutions” to the inversion problem may be
accurate at larger, often regional, scales but not at small scales.
158 P. Stenberg et al.

6.3 Structural and Spectral Characteristics of Coniferous


Forests

Northern hemisphere boreal forests, forming the largest unbroken, circumpolar for-
est zone in the world, are dominated by coniferous tree species. The radiation regime
of coniferous forests is known to differ in many respects from that of other vege-
tation, e.g., broadleaved forests and agricultural crops (Norman and Jarvis 1975;
Oker-Blom et al., 1991; Nilson, 1992). From the perspective of optical remote
sensing, a specific problem encountered in the boreal forest zone is the poor per-
formance of commonly used spectral vegetation indices (e.g., the NDVI) for the
estimation of biophysical variables (such as LAI). One reason for the generally ob-
served relative insensitivity of these indices to changes in LAI is apparently caused
by an abundant presence of mixed green understory in boreal coniferous forests
(Chen and Cihlar, 1996; Eklundh et al., 2001; Stenberg et al., 2004; Peltoniemi
et al., 2005; Rautiainen et al., 2007). The influence of understory on the spectral
signal naturally poses a problem also for the correct image interpretation using phys-
ically based models. In recent years, yet another phenomenon specific to coniferous
forests has become a subject of increasing scientific interest by the optical remote
sensing community. From empirical observations, it has long been known that conif-
erous forests have a lower reflectance, especially in the near-infrared (NIR), than
broadleaved forests, but the physical background to this behavior is still partly un-
explored. In the following, we will discuss how some conifer-specific characteristics
affect the spectral signature of coniferous forests.
There are several structural and spectral attributes which are specific to conif-
erous forests and require modifications to the radiative transfer models formulated
for their broadleaved counterparts (Fig. 6.1). The non-flat, three-dimensional struc-
ture of conifer needles of varying, species specific shape, first of all, requires proper

Needles

Shoots
Tree crowns,
internal shoot
distribution patttern

Fig. 6.1 Hierarchic structure of coniferous crowns: A three-layer scheme


6 Modeling the Spectral Signature of Forests 159

(re-)definition of some of the key variables and concepts of these models: e.g., the
leaf orientation, the mean projection of unit foliage area (“G-function”), the leaf
scattering phase function, and the leaf area index (Stenberg, 2006). These vari-
ables and distributions were all originally defined and derived under the assump-
tion of flat leaves and, so, for example, equations given for the mean projection
of planar leaf area (e.g., Warren Wilson 1967) are not directly applicable to nee-
dles but must be derived with consideration given to the needle shape (Oker-Blom
and Kellomäki, 1982; Lang, 1991; Chen and Black, 1992). A more serious theo-
retical problem for modeling however arises from that in many coniferous species
needles are closely grouped together as shoots. The resulting small-scale variation
in needle area density cannot readily be represented in the formulation of the ra-
diative transfer equation based on the concept of an “elementary volume”, which
should be small enough that essentially no mutual shading between the elements
exists but large enough for statistical laws to apply (Ross, 1981). It has been pro-
posed, therefore, that the shoot should be treated as the basic structural element
of coniferous canopies – an approach that has actually long been used in models
of canopy light interception and photosynthesis (e.g., Oker-Blom and Kellomäki,
1983; Nilson and Ross, 1997; Cescatti, 1997a), but has not yet been fully imple-
mented in forest reflectance models due to the lack of data and models describ-
ing the scattering properties of shoots (but see Smolander and Stenberg, 2003).
More data are available on another key parameter entering the shoot based models,
namely the shoot silhouette to total area ratio (STAR) (Oker-Blom and Smolander,
1988), which is conceptually analogous to the G-function defined for flat leaves
(Nilson, 1971), but includes a clumping coefficient accounting for the mutual shad-
ing of needles in the shoot. The clumping or mutual shading of needles in shoots
acts to decrease the interaction cross section area of a given amount of total needle
area, i.e., the extinction coefficient, and this effect can in radiative transfer models
be parameterized (quantified) using the STAR. The decrease in shoot single scat-
tering albedo, as compared to that of a single needle, has also been shown to be
closely related to STAR, which thus can be used to modify the scattering proper-
ties of an elementary volume containing shoots (Smolander and Stenberg, 2003).
However, theoretical models and, above all, empirical data on the shoot scattering
phase function for different species are still needed for correct parameterization of
coniferous canopy reflectance models. In some current models (Knyazikhin et al.,
1998b; Kuusk and Nilson, 2000), shoot level grouping is accounted for in quanti-
fying (i.e., reducing) the extinction coefficient (interaction cross section area of the
elementary foliated volume), but its effect on the volume scattering phase function
has not yet been fully implemented. It should be noted that when shoots (instead
of single needles, or needle surface area elements) are treated as the basic elements,
the optical properties (transmittance and reflectance) of single needles no longer suf-
fice to describe the scattering properties of the elementary volume in the radiative
transfer equation.
Also at higher levels of organization, coniferous forests display a distinct grouped
pattern (Fig. 6.1). The canopies are typically formed of dense, narrow and deep tree
crowns. Crown shape, volume, and density have a considerable effect on the total
160 P. Stenberg et al.

canopy reflectance, as well as on the bidirectional reflectance factor (BRF) of conifer


stands (Rautiainen et al., 2004). The same properties, together with the stand density
(number of trees per hectare), also determine how much of the background (forest
floor) is visible to the satellite instruments. Because northern coniferous forests are
often rather open, with an abundant green understory, the spectral signal is a com-
plex mixture of (overstory) canopy and background reflectance.

6.4 Effect of Clumping on the Spectral Signal: p-Theory


as an Example

The high degree of grouping (or clumping) of foliage at different hierarchical scales
in coniferous forest stands (Fig. 6.1) is considered to be the most important rea-
son behind the different spectral signature of these forests, in particular the low
reflectance in NIR, as compared to broadleaved forests (Nilson et al., 1999; Rauti-
ainen and Stenberg, 2005a). Incorporating the effects of grouping in canopy re-
flectance models poses a true challenge as very detailed canopy descriptions cannot
readily be integrated into models operating at large spatial scales – this is typically
the case in remote sensing of vegetation. Thus, especially for large scale applica-
tions, it is important to search for some key parameters that could capture the most
essential structural features of a forest stand. One such canopy structural parameter,
proposed to govern canopy absorption and scattering, is the spectrally invariant (i.e.,
wavelength independent) “p-parameter” introduced by Knyazikhin et al. (1998b).
The “p-theory”, in short, predicts that the amount of radiation scattered by a canopy
(bounded underneath by a black surface) should depend only on the wavelength and
the spectrally invariant parameter (p), which can be interpreted as the probability
that a photon scattered from a leaf (or needle) in the canopy will interact within
the canopy again – the “recollision probability” (Smolander and Stenberg, 2005;
Rautiainen and Stenberg, 2005a). The usefulness of this parameter in practical ap-
plications depends on whether and how well p can be related to (or derived from)
other commonly available forest stand data. Nonetheless, it is a powerful modeling
tool because it links canopy absorption (α C ) and scattering (ωC ) at any wavelength
(λ ) to the phytoelement (leaf or needle) scattering coefficient (ωL ) at the consid-
ered wavelength, while simultaneously preserving the law of energy conservation
(Panferov et al., 2001; Wang et al., 2003). The relationship between canopy and leaf
scattering coefficients is described by the simple equation:

ωL (λ ) − pωL (λ )
ωC (λ ) = (6.1)
1 − pωL (λ )

It follows from Eq. (6.1) that, at any given wavelength, the canopy spectral scattering
coefficient (ωC ) decreases in a nonlinear fashion with increasing canopy aggregation
or grouping (quantified by the p parameter). More importantly, in any given canopy
(fixed p), the relationship between ωC and ωL is also nonlinear, implying that canopy
6 Modeling the Spectral Signature of Forests 161

structure (or grouping, specifically) does not only affect the magnitude of the canopy
leaving radiation but also modifies its spectral distribution.
Simulation studies and empirical measurements have provided support for the
validity and usefulness of the p-theory. Equation (6.1) was shown to precisely pre-
dict the absorption and scattering in structurally homogeneous canopies simulated
with a Monte Carlo model (Smolander and Stenberg, 2003; Smolander and Sten-
berg, 2005), and to hold true also in heterogeneous canopies (Mõttus et al., 2007)
simulated using the hybrid FRT model (Kuusk and Nilson, 2000).
Another spectral invariant parameter (pt ) has been proposed to control canopy
transmission, i.e., the part of the scattered radiation that exits the canopy down-
wards (Knyazikhin et al., 1998a, b; Panferov et al., 2001; Shabanov et al., 2003).
This structural parameter so far lacks a clear physical interpretation, as has been
given for the other spectral invariant – the recollision probability (p), but if it can be
formulated and shown to be valid (both theoretically and empirically), the two para-
meters offer a simple and effective tool for parameterization of the canopy radiation
budget. Namely, given the absorption (p value) and transmission (pt value), total re-
flectance (the upward scattered part of the incident radiation) is also known (because
they all sum up to one). Thus, the spectral invariants p and pt would allow calcu-
lation of all components, i.e., spectral absorptance, transmittance and reflectance
of the canopy shortwave radiation budget for any given wavelength knowing the
leaf (or needle) scattering coefficient at the same wavelength. If the relationships
between p and canopy structural parameters such as LAI are known, the spectral
signature of the canopy can be predicted in terms of LAI or, conversely, inverse
estimation LAI can be done based on measured canopy reflectance.
The relationship described by Eq. (6.1), linking together canopy scattering co-
efficients at a specific wavelength to the leaf albedo at the same wavelength, can
actually be applied at different hierarchical levels and provides a tool for scaling
grouping effects. In the simulation studies by Smolander and Stenberg (2003, 2005),
it was found that Eq. (6.1) could be used to scale from needle to shoot scattering by
replacing p by the “recollision probability within a shoot” (psh ). Moreover, it was
shown that the canopy level recollision probability (p) could be decomposed into
(1) the probability (psh ) that the new interaction occurs within the same shoot where
the first (former) interaction took place, and (2) the probability (pc ) that interaction
occurs with another shoot in the canopy, according to the formula:

p = psh + (1 − psh )pc (6.2)

Equation (6.2), which in a similar manner could be further developed (decomposed)


to account, e.g., for grouping at the crown level, shows how the whole canopy p
value is affected by grouping at different spatial scales. For a given LAI, the p
value increases with the degree of grouping present in the spatial dispersion of the
leaf area. Thus, for example, considering a broadleaved and coniferous canopy with
similar LAI and macro-scale structure (Table 6.1), the coniferous canopy would
have higher p or, equivalently, a smaller escape probability for the radiation incident
162 P. Stenberg et al.

on the canopy. This explains, at least partly, the observed lower canopy reflectance
(albedo) of coniferous forests as compared to broadleaved forests.
Two shortcomings of the p-theory for practical applications should however be
noted: (1) It describes only canopy scattering, i.e., the contribution from background
reflectance must be modelled separately and (2) it is not able to describe the angular
distribution of scattered radiation. To solve these problems, the p-theory should be
combined with other physically-based reflectance modeling concepts.
With increasing p, the scattered part of the radiation intercepted by the foliated
canopy (the crowns), and thus also canopy reflectance, decreases. However, the ef-
fect of groping on whole stand reflectance is not as straightforward. To demonstrate
the combined effect grouping has on canopy reflectance, on one hand, and on the
contribution of understory reflectance, on the other hand, we use the simple parame-
terization model, PARAS (Rautiainen and Stenberg, 2005a). In this semi-physical
model, forest BRF is calculated as a sum of the ground and canopy components:
ωL − pωL
BRF = cg f (θ1 )cg f (θ2 )ρground + f (θ1 , θ2 )i0 (θ2 ) (6.3)
1 − pωL
The parameters of Eq. (6.3) are defined as follows: θ1 and θ2 are the viewing and
illumination zenith angles, cgf denotes the canopy gap fraction in the directions of
view and illumination (Sun), ρground is the BRF of the ground, f is the canopy scat-
tering phase function, i0 (θ2 ) is canopy interceptance or the fraction of the incoming
radiation interacting with the canopy, and ωL is needle (or leaf) scattering coeffi-
cient. With increasing degree of grouping (larger p), canopy interceptance (i0 ) si-
multaneously decreases while the canopy gap fractions (cgf) increase and, thus, the
contribution from ground (understory) reflectance increases. Especially grouping at
larger scale (between crowns) may more importantly influence the total stand re-
flectance through its effect on increasing the contribution from the background than
through its effect on the canopy contribution.
Incorporating the effect of shoot scale clumping in the PARAS canopy re-
flectance model was shown to produce realistic reflectance values in the near-
infrared (NIR) of coniferous forests. This can be seen as a major improvement since
the low NIR reflectance observed in coniferous areas is one of the main anomalies
that models have not been able to account for. The results give support to the hy-
pothesis that in coniferous canopies large part of the clumping occurs at the shoot
level and, thus, that the incorporation of shoot structural and spectral properties into
current forest reflectance models will significantly improve their performance.

6.5 Scaling of Canopy Reflectance

In the previous section, the term “scaling” was introduced to describe an application
of recollision probability at different canopy grouping levels, i.e., the dimensionless
scattering characteristics calculated at one level were applied at another, larger scale.
This approach is similar to using downscaled models for predicting the behavior
6 Modeling the Spectral Signature of Forests 163

of the full-scale object. The procedure can be applied repeatedly as long as the
character of the process (in our case, interaction of photons with canopy elements)
remains unchanged. In optical remote sensing, the limit is determined by the ratio
of photon wavelength to the size of the scatterers, i.e., radiative transfer is still
determined by geometric optics, and refraction can be ignored.
In the current section, “scaling” has a different meaning that is more familiar to
the remote sensing community. Here, we use it to describe the spatial variation of
canopies that has to be taken into account when comparing sensors with different
view configurations (i.e., sensors that produce images with different pixel sizes).
Generally, models can produce only point estimates of canopy reflectance, although
most of them have a method to account for within-plot variability (statistical distri-
bution of tree locations, etc.). This works well as long as the canopy is horizontally
homogeneous (which is what these models actually presume) and its structure does
not vary across a relatively large area, substantially exceeding the dimensions of
a pixel in the image produced by the sensor. Natural forest canopies, however, al-
most never possess such a property: they include clusters of high tree concentrations
and canopy openings, due to harvesting or windfall. As the reflectance process is
strongly nonlinear, this heterogeneity is difficult to take into account in the spatially
averaged pixels of remote sensing instruments.
The scale at which model results can be related to the reflected signal in a
straightforward way depends on the size of the structural units of the canopy. For
example, models that utilize an assumption of a statistical distribution of tree loca-
tions predict an average signal produced by such a canopy and cannot be applied
to model distinct patterns created by a particular configuration of individual trees
or the variance of intensity in the image of a single crown. They were not designed
for these purposes as overly detailed patterns are commonly not required for remote
sensing of larger areas. Another reason is that biophysical variables (LAI, fCover)
are defined for a larger canopy area and cannot be used to describe a single tree.
When moving to larger resolutions, the models may also fail due to the non-
linearity of the radiative transfer problem and a highly varying or discontinuous
distribution of tree locations. Canopy spectral properties are a function of spatial
resolution (Tian et al., 2000; Pinty et al., 2004a), and especially in regions with a
fragmented forest area, the signal will always be a mixture of reflectance from many
different ground (soil/understory) and forest canopy compartments. This may lead
to a situation where a model cannot be used as such (without alterations) to model
signals by sensors with different spatial resolutions (Tian et al., 2002).
The scaling problem affects the retrieval of different biophysical variables with
varying severity. While some variables can be considered almost scale-independent
(fCover, fPAR), predicting others (e.g., LAI) requires a careful consideration of spa-
tial variation (Weiss et al., 2000). Algorithms for a correct treatment of the scaling
problem of vegetation reflectance modeling are scarce although a physically-based
theory for scaling was developed by Tian et al. (2002), for example. The incorpo-
ration of scaling algorithms directly in radiation transfer models is so far an almost
unexploited subject (Widen, 2004).
164 P. Stenberg et al.

6.6 Use of Stand Inventory Data as Model Input

Ground truth, or ground observations, is required for both developing and testing
physically based forest reflectance models and their inversion. There are two gen-
eral paths for acquiring this information: by conducting measurements or by using
existing data bases (e.g., land cover maps, regional or national forest inventories,
spectral data banks of common plant species). However, a general problem with all
these data sources may be their age when compared to the more recently obtained
remotely sensed image.
Stand inventory data (which typically consists of information on tree species,
tree height and diameter, and stand density or basal area) form a common input
set required by most 3D forest reflectance models. However, stand inventories do
not provide all the parameters typically included in physically based reflectance
models, e.g., crown shape, leaf area density and orientation, nor do they include
the biophysical variables of interest in this chapter (LAI, fPAR, fCover). Testing
reflectance models should be done with a carefully measured, comprehensive input
data set which includes all the input required by the model and which does not rely
on other models (i.e., is not generated by other models). However, when operational
inversion for obtaining biophysical variables (LAI, fPAR, or fCover in this case)
is considered, such comprehensive data sets are not available, and other, let us say
less appropriate, sets need to be utilized. A typical source would be regional or
national forest inventory data bases which have recently become more suitable for
the purpose in the boreal zone of North America and Europe. Traditionally forest
mensuration science dealt with determining the volume of stands and logs and then
studying stand growth and yield. Nowadays, however, the scope has widened and
regional and national forest inventories have become more efficient environmental
monitoring tools and are conducted to ensure the sustainable use of forests. Thus,
they include more stand variables than tree height and diameter and form a better
interface for forest reflectance models. Nevertheless, if forest reflectance models are
run through a routine forest management data base (e.g., Rautiainen, 2005), there
are several structural input parameters, such as crown shape, woody area index or
shoot size, in the models which can hardly be obtained from these data bases. Thus,
alternative solutions for obtaining the structural information need to be considered.
A possibility worth exploring further would be to create realistic input data on
these additional structural parameters from basic stand variables through regres-
sion models and allometric relationships. Such models exist for relating different
biomass components of a tree or forest (e.g., Marklund, 1988) but parameters de-
scribing crown architecture have been of less interest in forest mensuration. In the
context of radiative transfer modeling, one of the most central parameters not avail-
able from routine stand inventory data sets is crown shape which here serves as an
example. Crown shape determines the limits of integration over the crown envelope
and thus the scattering volume. Therefore, crown shape needs to be predicted from
routine stand variables such as tree height and breast height diameter. This can be
done by, for example, developing a crown shape model based on extensive measure-
ments and then using the model to predict crown shape for a stand where it has not
6 Modeling the Spectral Signature of Forests 165

been measured (e.g., Baldwin and Peterson, 1997; Rautiainen and Stenberg, 2005b;
Mõttus et al., 2006). Alternatively, crown shape can be assumed to have a constant
geometric shape (e.g., a cone or an ellipsoid) for a given tree species.

6.7 Concluding Remarks

Remote sensing mapping techniques have traditionally been statistically oriented


but physically based methods have progressively become more and more attrac-
tive because they are better suited for many current large scale applications related
to global mapping of vegetation (e.g., Knyazikhin et al., 1998b). Implementation of
physically based remote sensing methodology has become feasible thanks to the de-
velopment of fast computer hardware and high resolution and multi-angular satellite
instruments. The rapid technological and methodological development of satellite
derived products offers a range of new possibilities for the mapping of terrestrial
biophysical parameters such as vegetation type, forest cover (fCover) and leaf area
index (LAI) from remotely sensed data (Myneni et al., 1997; Chen et al., 2002).
Remote sensing based methodologies are especially important, and in fact the only
feasible alternative, in regions where reliable field information is not available, is
difficult to reach, or is too expensive. This, on the other hand, makes the impor-
tance of validation an even more crucial part of the development process aiming at
producing trustworthy information righteously required by the end-users.
Several research networks operate today on the refinement of remote sensing
mapping methods and on validation of satellite derived biophysical products, such as
LAI, fCover and fPAR, that are presently routinely generated by a range of sensors
(Knyazikhin et al., 1998b; Morisette et al., 2006).
For critical assessment of the accuracy of remotely sensed estimates of biophysi-
cal variables, there should be available a set of statistically representative (sufficient)
and independent reference data with known accuracy. Also, special care should be
taken that similar (standardized) measurement methodologies and designs are used
to produce reference data of the variable of interest (Morisette et al., 2006). Global
validation networks have a very important mission in providing these validation data
needed for the further development of the remote sensing methodology.
Physically based remote sensing methods still have room for improvement, and
we have claimed in this chapter that there is especially a need to further develop re-
flectance models designed for coniferous forests, accounting for their complex 3D
canopy structure and the small-scale variation in needle area density. The widely dis-
cussed global trends of greening and vegetation status require accurate algorithms
for boreal forests. An efficient use of physically based forest reflectance models in
the boreal coniferous forest zone would require that the model input data include
variables such as crown shape and canopy cover, and spatial pattern of trees, to rep-
resent their clumped structure. In combination with standard forest inventory data,
these key structural parameters would form a very valuable data base. It seems rea-
sonable to believe that with better representation of the 3D canopy structure, the
166 P. Stenberg et al.

accuracy of remotely sensed estimates of canopy biophysical characteristics would


improve considerably. Accordingly, future tasks of high priority include means to
create realistic input data on crown shape, tree pattern and canopy cover from basic
stand variables, for example through regression models and allometric relationships.
In boreal forests, another limitation to successful application of canopy reflectance
models is a typically large proportion of non-contrasting background reflectance in
the visible and near-infrared (NIR) part of the spectra due to abundant mixed green
understory. Stand inventory data do not include spectral properties of the forests,
and there is a need for more data on the angular reflectance properties of common
forest species, including not only understory vegetation but tree components, such
as needles, shoots and bark.
Use of the p-theory is for application in large scale remote sensing mapping
methods is another question of large current interest. Can a simple parameteriza-
tion model using the spectral invariants be built to mimic canopy reflectance with
sufficient accuracy? It has already been shown, and put into practice in the MODIS
LAI/FPAR algorithm, that the spectral invariants provide a powerful calculation tool
in reflectance models and preserve the law of energy conservation. However, if one
could derive specific relationships between the spectral invariants (p and pt ) and
basic canopy structural parameters such as LAI and canopy cover, it would provide
an effective tool by which the effect of the clumping of foliage at different hierar-
chical levels of the forest stand structure could be incorporated in forest reflectance
models.

References

Andrieu B, Baret F, Jacquemoud S, Malthus T, Steven M (1997) Evaluation of an improved version


of SAIL model for simulating bidirectional reflectance of sugar beet canopies. Remote Sens.
Environ. 60:247–257
Atzberger C (2000) Development of an invertible forest reflectance model: The INFOR-Model.
In: Buchroithner (ed), A decade of Trans-European Remote Sensing Cooperation, Proc. 20th
EARSeL Symposium, 14–16 June 2000, Dresden, pp 39–44
Baldwin VC, Peterson KD (1997) Predicting the crown shape of loblolly pine trees. Can. J. For.
Res. 27:102–107
Borel CC, Gerstl SAW, Powers BJ (1991) The radiosity method in optical remote sensing of struc-
tured 3-D surfaces. Remote Sens. Environ. 36:13–44
Cescatti A (1997a) Modelling the radiative transfer in discontinuous canopies of assymmetric
crowns. I. Model structure and algorithms. Ecol. Model. 101:263–274
Cescatti A (1997b) Modelling the radiative transfer in discontinuous canopies of assymmetric
crowns. II. Model testing and application in a Norway spruce stand. Ecol. Model. 101:275–
284
Chandrasekhar S (1960) Radiative Transfer. Oxford University Press, London, UK
Chelle M, Andrieu B (1998) The nested radiosity model for the distribution of light within plant
canopies. Ecol. Model. 111:75–91
Chen J, Pavlic G, Brown L, Cihlar J, Leblanc SG, White HP, Hall RJ, Peddle DR, King DJ,
Trofymow JA, Swift E, van der Sanden J, Pellikka PKE (2002) Derivation and validation of
Canada-wide coarse-resolution leaf area index maps using high-resolution satellite imagery
and ground measurements. Remote Sens. Environ. 80:165–184
6 Modeling the Spectral Signature of Forests 167

Chen JM, Black TA (1992) Defining leaf area index for non-flat leaves. Plant Cell Environ. 15:
421–429
Chen JM, Cihlar J (1996) Retrieving leaf area index of boreal conifer forests using Landsat TM
images. Remote Sens. Environ. 55:153–162
Chen JM, Leblanc SG (1997) A four-scale bidirectional reflectance model based on canopy archi-
tecture. IEEE Trans. Geosci. Remote Sens. 35:1316–1337
Chen JM, Leblanc SG (2001) Multiple-scattering scheme useful for geometric optical modeling.
IEEE Trans. Geosci. Remote Sens. 39:1061–1071
Disney MI, Lewis P, North PRJ (2000) Monte Carlo ray tracing in optical canopy reflectance
modelling. Remote Sens. Rev. 18:163–196
Disney MI, Lewis P, Saich P (2006) 3D modelling of forest canopy structure for remote sensing
simulations in the optical and microwave domains. Remote Sens. Environ. 100:114–132
Eklundh L, Harrie L, Kuusk A (2001) Investigating relationships between Landsat ETM + sensor
data and leaf area index in boreal conifer forest. Remote Sens. Environ. 78:239–251
Garcı́a-Haro FJ, Sommer S (2002) A fast canopy reflectance model to simulate realistic remote
sensing scenarios. Remote Sens. Environ. 81:205–227
Garcı́a-Haro FJ, Gilabert MA, Meliá J (1999) A radiosity model for heterogeneous canopies in
remote sensing. J. Geophys. Res. 104:12159–12175
Gastellu-Etchegorry J, Demarez V, Pinel, V, Zagolski F (1996) Modeling radiative transfer in het-
erogeneous 3-D vegetation canopies. Remote Sens. Environ. 58:131–156
Gerard FF, North PRJ (1997) Analyzing the effect of structural variability and canopy gaps on
forest BRDF using a geometric-optical model. Remote Sens. Environ. 62:46–62
Gerstl SAW, Borel CC (1992) Principles of the radiosity method versus radiative transfer for
canopy reflectance modeling. IEEE Trans. Geosci. Remote Sens. 30:271–275
Gerstl SAW, Simmer C, Powers BJ (1986) The canopy hot-spot as crop identifier. In: MSJ Damen,
G Sicco Smit, HT Verstappen (eds), Remote Sensing for Resources Development and Environ-
mental Management. Proc. 7th International Symposium, Enschede, 25–29 August 1986. AA
Balkema, Rotterdam
Goel NS (1989) Inversion of canopy reflectance models for estimation of biophysical parameters
from reflectance data. In: G Asrar (ed), Theory and Applications of Optical Remote Sensing.
Wiley, New York, pp 205–251
Goel NS, Rozehnal I, Thompson RL (1991) A computer-graphics based model for scattering from
objects of arbitrary shapes in the optical-region. Remote Sens. Environ. 36:73–104
Govaerts YM, Verstraete MM (1998) Raytran: a Monte Carlo ray-tracing model to compute
light scattering in three-dimensional heterogeneous media. IEEE Trans. Geosci. Remote Sens.
36:493–505
Huemmrich KF (2001) The GeoSail model: a simple addition to the SAIL model to describe dis-
continuous canopy reflectance. Remote Sens. Environ. 75:423–431
Jacquemoud S, Baret F, Andrieu B, Danson FM, Jaggard K (1995) Extraction of vegetation bio-
physical parameters by inversion of the PROSPECT + SAIL models on sugar beet canopy
reflectance data. Application to TM and AVIRIS sensors. Remote Sens. Environ. 52:163–172
Jupp DLB, Strahler AH (1991) A hotspot model for leaf canopies. Remote Sens. Environ. 38:
193–210
Knyazikhin Y, Marshak A (1991) Fundamental equations of radiative transfer in leaf canopies, and
iterative methods for their solution. In: RB Myneni, J Ross (eds), Photon–Vegetation Interac-
tions, Springer, Berlin-Heidelberg, pp 9–43
Knyazikhin Y, Kranigk J, Myneni RB, Panfyorov O, Gravenhorst G (1998a) Influence of small-
scale structure on radiative transfer and photosynthesis in vegetation cover. J. Geophys. Res.
103(DD6):6133–6144
Knyazikhin Y, Martonchik JV, Myneni RB, Diner DJ, Running SW (1998b) Synergistic algorithm
for estimating vegetation canopy leaf area index and fraction of absorbed photosynthetically
active radiation from MODIS and MISR data. J. Geophys. Res. 103(D24):32257–32276
Koetz B, Baret F, Poilvé H, Hill J (2005) Use of coupled canopy structure dynamic and radia-
tive transfer models to estimate biophysical canopy characteristics. Remote Sens. Environ. 95:
115–124
168 P. Stenberg et al.

Kuusk A (1991) The hot spot effect in plant canopy reflectance. In: RB Myneni, J Ross (eds),
Photon–Vegetation Interactions, Springer, Berlin-Heidelberg, pp 139–159
Kuusk A (1995) A fast, invertible canopy reflectance model. Remote Sens. Environ. 51:342–350
Kuusk A (2001) A two-layer canopy reflectance model. J. Quant. Spectrosc. Radiat. Transfer
71:1–9
Kuusk A, Nilson T (2000) A directional multispectral forest reflectance model. Remote Sens.
Environ. 72:244–252
Lang ARG (1991) Application of some of Cauchy’s theorems to estimation of surface areas
of leaves, needles and branches of plants, and light transmittance. Agric. For. Meteorol. 55:
191–212
Leverenz JW, Hinckley TM (1990) Shoot structure, leaf area index and productivity of evergreen
conifer stands. Tree Physiology. 6:135–149
Li X, Strahler AH (1985) Geometric-optical modeling of a conifer forest canopy. IEEE Trans.
Geosci. Remote Sens. GE-23:705–721
Li X, Strahler AH (1986) Geometric-optical bidirectional reflectance modeling of a conifer forest
canopy. IEEE Trans. Geosci. Remote Sens. GE-24:906–919
Li X, Strahler AH (1992) Geometric-optical bidirectional reflectance modeling of the discrete
crown vegetation canopy: effect of crown shape and mutual shadowing. IEEE Trans. Geosci.
Remote Sens. 30:276–292
Li X, Strahler AH (1995) A hybrid geometric-optical radiative transfer approach for modeling
albedo and directional reflectance of discontinuous canopies. IEEE Trans. Geosci. Remote
Sens. 33:466–480
Liang S, Strahler AH (guest editors) 2000. Land surface bi-directional reflectance distribution func-
tion (BRDF): Recent advances and future prospects. Special issue of Remote Sensing Reviews.
18:83–511
Marklund L (1988) Biomassafunktioner för tall, gran och björk i Sverige. Report 4 (in Swedish).
Department of Forest Survey, Swedish University of Agricultural Sciences, Umeå, Sweden.
73 pp
Marshak AL (1989) The effect of the hot spot on the transport equation in plant canopies. J. Quant.
Spectrosc. Radiat. Transfer L.2:615–630
Menzhulin GV, Anisimov OA (1991) Principles of statistical phytoactinometry. In: RB Myneni, J
Ross, (eds), Photon–Vegetation Interactions, Springer, Berlin-Heidelberg, pp 111–137
Meroni M, Colombo R, Panigada C (2004) Inversion of a radiative transfer model with hyperspec-
tral observations for LAI mapping in poplar plantations. Remote Sens. Environ. 92:195–206
Morisette JT, Baret F, Privette JL, Myneni RB, Nickeson J, Garrigues S, Shabanov N, Weiss M,
Fernandes R, Leblanc S, Kalacska M, Sánchez-Azofeifa GA, Chubey M, Rivard B, Stenberg
P, Rautiainen M, Voipio P, Manninen T, Pilant A, Lewis T, Iiames J, Colombo R, Meroni M,
Busetto L, Cohen W, Turner D, Warner ED, Petersen GW, Seufert G, Cook R (2006) Validation
of global moderate resolution LAI Products: a framework proposed within the CEOS Land
Product Validation subgroup. IEEE Trans. Geosci. Remote Sens. 44(7):1804–1817
Mõttus M, Stenberg P, Rautiainen M (2007) Photon recollision probability in heterogeneous
forest canopies: Compatibility with a hybrid GO model. J. Geophys. Res. 112:D03104,
doi:10.1029/2006JD007445
Mõttus M, Sulev M, Lang M (2006) Estimation of crown volume for a geometric radiation model
from detailed measurements of tree structure. Ecological Modelling 198(3–4):506–514.
Myneni RB (1990) Modeling radiative transfer and photosynthesis in three-dimensional vegetation
canopies. Agric. For. Meteorol. 55:323–344
Myneni RB, Ross J, Asrar G (1989) A review on the theory of photon transport in leaf canopies.
Agric. For. Meteorol. 45:1–153
Myneni RB, Ross J (eds) (1991) Photon–Vegetation Interactions, Springer, Berlin-Heidelberg,
565 pp
Myneni RB, Keeling CD, Tucker CJ, Asrar G, Nemani RR (1997) Increased plant growth in the
northern high latitudes from 1981 to 1991 Nature. 386 (6626):698–702
6 Modeling the Spectral Signature of Forests 169

Ni W, Li X, Woodcock C, Caetano M, Strahler A (1999) An analytical hybrid GORT model for


bidirectional reflectance over discontinuous plant canopies. IEEE Trans. Geosci. Remote Sens.
37:987–999
Nilson T (1971) A theoretical analysis of the frequency of gaps in plant stands. Agric. Meteorol.
8:25–38
Nilson T (1977) A theory of radiation penetration into nonhomogeneous plant canopies. In: The
Penetration of Solar Radiation into Plant Canopies, Acad. Sci. ESSR Report, Tartu, 5–70 (in
Russian)
Nilson T (1992) Radiative transfer in nonhomogeneous plant canopies. In: G Stanhill (ed), Ad-
vances in Bioclimatology 1. Springer, Berlin-Heidelberg, pp 59–88
Nilson T, Peterson U (1991) A forest canopy reflectance model and a test case. Remote Sens.
Environ. 37:131–142
Nilson T, Ross J (1997) Modeling radiative transfer through forest canopies: implications for
canopy photosynthesis and remote sensing. In: HL Gholz, K Nakane, H Shimoda (eds),
The Use of Remote Sensing in the Modeling of Forest Productivity. Kluwer, Dordrecht, The
Netherlands, pp 23–60
Nilson T, Anniste J, Lang M, Praks J (1999) Determination of needle area indices of coniferous for-
est canopies in the NOPEX region by ground-based optical measurements and satellite images.
Agric. For. Meteorol. 98–99:449–462
Norman J, Jarvis P (1975) Photosynthesis in Sitka spruce (Picea sitchensis (Bong.) Carr.). V. Ra-
diation penetration theory and a test case. J. Appl. Ecol. 12(3):839–877
North PRJ (1996) Three-dimensional forest light interaction model using a Monte Carlo method.
IEEE Trans. Geosci. Remote Sens. 34:946–956
Oker-Blom P (1986) Photosynthetic radiation regime and canopy structure in modeled forest
stands. Acta For. Fenn. 197:1–44
Oker-Blom P, Kellomäki S (1982) Effect of angular distribution of foliage on light absorption and
photosynthesis in the plant canopy: theoretical computations. Agric. Meteorol. 26:105–116
Oker-Blom P, Kellomäki S (1983) Effect of grouping of foliage on the within-stand and within-
crown light regime: comparison of random and grouping canopy models. Agric. Meteorol.
28:143–155
Oker-Blom P, Smolander H (1988) The ratio of shoot silhouette area to total needle area in Scots
pine. For. Sci. 34:894–906
Oker-Blom P, Lappi J, Smolander H (1991) Radiation regime and photosynthesis of coniferous
stands. In: RB Myneni, J Ross (eds), Photon-Vegetation Interactions: Applications in Optical
Remote Sensing and Plant Ecology. Springer, New York, pp 469–499
Panferov O, Knyazikhin Y, Myneni RB, Szarzynski J, Engwald S, Schnitzler KG, Gravenhorst G
(2001) The role of canopy structure in the spectral variation of transmission and absorption of
solar radiation in vegetation canopies. IEEE Trans. Geosci. Remote Sens. 39:241–253
Peddle DR, Johnson RL, Cihlar J, Latifovic R (2004) Large area forest classification and biophysi-
cal parameter estimation using the 5-scale canopy reflectance model in multiple-forward-mode.
Remote Sens. Environ. 89:252–263
Peltoniemi JI, Kaasalainen S, Näränen J, Rautiainen M, Stenberg P, Smolander H, Smolander S,
Voipio P (2005) BRDF measurement of understorey vegetation in pine forests: dwarf shrubs,
lichen and moss. Remote Sens. Environ. 94:343–354
Pinty B, Gobron N, Widlowski J-L, Lavergne T, Verstraete MM (2004a) Synergy between 1-D
and 3-D radiation transfer models to retrieve vegetation canopy properties from remote sensing
data. J. Geophys. Res. 109, D21205, doi:10.1029/2004JD005214
Pinty B, Widlowski JL, Taberner M, Gobron N, Verstraete MM, Disney M, Gascon F, Gastellu
JP, Jiang L, Kuusk A, Lewis P, Li X, Ni-Meister W, Nilson T, North P, Qin W, Su L, Tang
S, Thompson R, Verhoef W, Wang H, Wang J, Yan G, Zang H (2004b) Radiation transfer
model intercomparison (RAMI) exercise: results from the second phase. J. Geophys. Res. 109,
D06210, doi:10.1029/2003JD004252
Qin W, Gerstl SAW, Deering DW, Goel NS (2002) Characterizing leaf geometry for grass and crop
canopies from hotspot observations: a simulation study. Remote Sens. Environ. 880:100–113
170 P. Stenberg et al.

Rautiainen M (2005) Retrieval of leaf area index for a coniferous forest by inverting a forest re-
flectance model. Remote Sens. Environ. 99:295–303
Rautiainen M, Suomalainen J, Mõttus M, Stenberg P, Voipio P, Peltoniemi J, Manninen, T (2007)
Coupling forest canopy and understory reflectance in the Arctic latitudes of Finland. Remote
Sens. Environ. 110:332–343
Rautiainen M, Stenberg P (2005a) Application of photon recollision probability in coniferous
canopy reflectance simulations. Remote Sens. Environ. 96:98–107
Rautiainen M, Stenberg P (2005b) Simplified tree crown model using standard forest mensuration
data for Scots pine. Agric. For. Meteorol. 128:123–129
Rautiainen M, Stenberg P, Nilson T, Kuusk A (2004) The effect of crown shape on the reflectance
of coniferous stands. Remote Sens. Environ. 89:41–52
Ross J (1981) The Radiation Regime and Architecture of Plant Stands. Junk Publishers, The
Hague, 391 pp
Ross KJ, Marshak AL (1988) Calculation of canopy bidirectional reflectance using the Monte
Carlo method. Remote Sens. Environ. 24:213–225
Shabanov NV, Knyazikhin Y, Baret F, Myneni RB (2000) Stochastic modeling of radiation regime
in discontinuous vegetation canopies. Remote Sens. Environ. 74:125–144
Shabanov NV, Wang Y, Buermann W, Dong J, Hoffman S, Smith GR, Tian Y, Knyazikhin Y, My-
neni RB (2003) Effect of foliage spatial heterogeneity in the MODIS LAI and FPAR algorithm
over broadleaf forests. Remote Sens. Environ. 85:410–423
Smolander S, Stenberg P (2003) A method to account for shoot scale clumping in coniferous
canopy reflectance models. Remote Sens. Environ. 88:363–373
Smolander S, Stenberg P (2005) Simple parameterizations for the radiation budget of uniform
broadleaved and coniferous canopies. Remote Sens. Environ. 94:355–363
Soler C, Sillion FX, Blaise F, Dereffye P (2003) An efficient instantiation algorithm for simulating
radiant energy transfer in plant models. ACM Trans. Graphics. 22:204–233
Stenberg P (2006) A note on the G-function for needle leaf canopies. Agric. For. Meteorol.
Stenberg P, Rautiainen M, Manninen T, Voipio P, Smolander H (2004) Reduced simple ratio better
than NDVI for estimating LAI in Finnish pine and spruce stands. Silva Fennica. 38(1):3–14
Thompson RL, Goel NS (1998) Two models for rapidly calculating bidirectional reflectance: Pho-
ton spread (ps) model and statistical photon spread (sps) model. Remote Sensing Reviews
16:157–207
Tian Y, Zhang Y, Knyazikhin Y, Myneni RB, Glassy JM, Dedieu G, Running SW (2000) Proto-
typing of MODIS LAI and FPAR algorithm with LASUR and LANDSAT data. IEEE Trans.
Geosci. Remote Sens. 38:2387–2401
Tian Y, Wang Y, Zhang Y, Knyazikhin Y, Bogaert J, Myneni RB (2002) Radiative transfer based
scaling of LAI retrievals from reflectance data of different resolutions. Remote Sens. Environ.
84:143–159
Verhoef W (1984) Light-scattering by leaf layers with application to canopy reflectance modeling –
the SAIL model. Remote Sens. Environ. 16:125–141
Verhoef W (1985) Earth observation modeling based on layer scattering matrices. Remote Sens.
Environ. 17:165–178
Verstraete MM, Pinty B, Dickinson RE (1990) A physical model of the bidirectional reflectance of
vegetation canopies. Part 1: theory. J. Geophys. Res. 95:11765–11775
Verstraete M, Pinty B, Myneni R (1996) Potential and limitations of information extraction on the
terrestrial biosphere from satellite remote sensing. Remote Sens. Environ. 58:201–214
Wang Y, Buermann W, Stenberg P, Smolander H, Häme T, Tian Y, Hu J, Knyazikhin Y, Myneni
RB (2003) A new parameterization of canopy spectral response to incident solar radiation: case
study with hyperspectral data from pine dominant forest. Remote Sens. Environ. 85:304–315
Wang YP, Jarvis PG (1990) Influence of crown structural properties on PAR absorption, photosyn-
thesis, and transpiration in Sitka spruce: application of a model (MAESTRO). Tree Physiology.
7:297–316
Warren Wilson J (1967) Stand structure and light penetration. III. Sunlit foliage area. J. Appl. Ecol.
4(1):159–165
6 Modeling the Spectral Signature of Forests 171

Weiss M, Baret F (1999) Evaluation of canopy biophysical variable retrieval performances from
the accumulation of large swath satellite data. Remote Sens. Environ. 70:293–306
Weiss M, Baret F, Myneni RB, Pragnèrea A, Knyazikhin Y (2000) Investigation of a model in-
version technique to estimate canopy biophysical variables from spectral and directional re-
flectance data. Agronomie. 20:3–22
Weiss M, Baret F, Smith GJ, Jonckheere I, Coppin P (2004) Review of methods for in situ leaf
area index (LAI) determination. Part II. Estimation of LAI, errors and sampling. Agric. For.
Meteorol. 121:37–53
Welles JM, Norman JM (1991) Photon transport in discontinuous canopies: a weighted ran-
dom approach. In: RB Myneni, J Ross (eds), Photon-Vegetation Interaction, Springer, Berlin-
Heidelberg, pp 389–414
Widen N (2004) Assessing the accuracy of land surface characteristics estimated from multi-
angular remotely sensed data. Int. J. Remote Sens. 25:1105–1117
Yoder B, Pettigrew-Crosby R (1995) Predicting nitrogen and chlorophyll content and concentra-
tions from reflectance spectra (400–2500 nm) at leaf and canopy scales. Remote Sens. Environ.
53:199–211
Zarco-Tejada PJ, Rueda CA, Ustin SL (2003) Water content estimation in vegetation with MODIS
reflectance data and model inversion methods. Remote Sens. Environ. 85:109–124
Chapter 7
Estimating Canopy Characteristics
from Remote Sensing Observations:
Review of Methods and Associated Problems

Frédéric Baret and Samuel Buis

Abstract This article describes the methods and problems associated to the esti-
mation of canopy characteristics from remote sensing observations. It is illustrated
over the solar spectral domain, with emphasis on LAI estimation using currently
available algorithms developed for moderate resolution sensors. The principles of
algorithms are first presented, distinguishing between canopy biophysical and ra-
diometric data driven approaches that may use either radiative transfer models or
experimental observations. Advantages and drawback are discussed with due atten-
tion to the operational character of the algorithms. Then the under-determination
and ill-posedness nature of the inverse problem is described and illustrated. Finally,
ways to improve the retrieval performances are presented, including the use of prior
information, the exploitation of spatial and temporal constraints, and the interest in
using holistic approaches based on the coupling of radiative transfer processes at
several scales or levels. A conclusion is eventually proposed, discussing the three
main components of retrieval approaches: retrieval techniques, radiative transfer
models, and the exploitation of observations and ancillary information.

7.1 Introduction

Many applications require an exhaustive description of the spatial domain of interest


that may cover a large range of scales: from the very local one corresponding to pre-
cision agriculture where cultural practices are adapted to the within field variability,
through environmental management generally approached at the landscape scale, up
to biogeochemical cycling and vegetation dynamics investigated at national, conti-
nental and global scales. Most of these applications are using our knowledge on the

F. Baret and S. Buis


UMR1114, INRA-CSE, 84 914 Avignon, France
baret@avignon.inra.fr

S. Liang (ed.), Advances in Land Remote Sensing, 173–201. 173


c Springer Science + Business Media B.V., 2008
174 F. Baret, S. Buis

main physical, chemical and biological processes involved such as energy balance,
evapotranspiration, photosynthesis and respiration. This knowledge is encapsulated
into a variety of surface process models. However, to account for the spatial het-
erogeneity observed at all scales, dedicated imaging systems are required to get a
distributed description of surface characteristics within the domain of interest. By
its capacity to cover exhaustively large space areas, remote sensing provides a very
pertinent answer to those requirements. However, remote sensing observations sam-
ple the radiation field reflected or emitted by the surface, and thus do not provide
directly the biophysical characteristics required by the models for describing some
state variables of the surface. An intermediate step is therefore necessary to trans-
form the remote sensing measurements into estimates of the surface biophysical
characteristics.
Many methods have been proposed to retrieve surface characteristics from re-
mote sensing observations. They span from simple empirical ones with calibration
over experimental data sets, up to more complex ones based on the use of radia-
tive transfer models. Radiative transfer models summarize our knowledge on the
physical processes involved in the photon transport within vegetation canopies or
atmosphere, and simulate the radiation field reflected or emitted by the surface for
given observational configuration, once the vegetation and the background as well
as possibly the atmosphere are specified. Retrieving canopy characteristics from the
radiation field as sampled by the sensor aboard satellite needs to “invert” the radia-
tive transfer model.
This article aims at presenting the state of the art in the estimation of surface
characteristics from remote sensing observations. Although this is a very general
problem in remote sensing, it will be illustrated by examples taken in the solar do-
main (400–2,500nm), with emphasis put on the current operational algorithms that
are mainly used for medium resolution sensors such as MODIS, MERIS, AVHRR,
VEGETATION, POLDER and SEAWIFS. Among the possible canopy characteris-
tics accessible from remote sensing in the reflective solar domain, we will focus on
leaf area index (LAI), defined as half the developed area of green elements per unit
horizontal soil (Stenberg, 2006). As a matter of fact, LAI is one of the key canopy
state biophysical variables required by many process models to describe energy and
mass exchanges in the soil/plant/atmosphere system.

7.2 Principles of Biophysical Variable Retrieval Algorithms

Remote sensing data result from radiative transfer processes within canopies that
depend on canopy variables, and observational configuration (wavelength, view and
illumination directions). Canopy variables include the variables of interest for the
applications such as LAI, and the other variables that are not of direct use for the
applications but that influence the radiative transfer, such as soil background prop-
erties. The causal relationship between the variables of interest and remote sens-
ing data corresponds to the forward (or direct) problem (Fig. 7.1). They could be
7 Estimating Canopy Characteristics from Remote Sensing Observations 175

Other variables

Radiative Forward problem


transfer

Variables of Observation Remote sensing


interest configuration data

Retrieval
Inverse problem algorithm

Prior information

Fig. 7.1 Forward (solid lines) and inverse (dashed lines) problems in remote sensing

either described through empirical relationships calibrated over experiments or us-


ing radiative transfer models based on a more or less close approximation of the
actual physical processes. Conversely, retrieving the variables of interest from re-
mote sensing measurements corresponds to the inverse problem, i.e., developing al-
gorithms to estimate the variables of interest from remote sensing data as observed
in a given configuration. Prior information on the type of surface and on the dis-
tribution of the variables of interest can also be included in the retrieval process to
improve the performances as we will see later.
A panoply of retrieval techniques currently used have been reviewed in the early
1990s by several authors (Asrar et al., 1989; Goel, 1989; Pinty and Verstraete, 1991)
and more recently by Kimes et al. (2000) and Liang (2004). They can be split into
two main approaches (Fig 7.2) depending if the emphasis is put on remote sens-
ing data (radiometric data) or on the variables of interest to be estimated (canopy
biophysical variables).

7.2.1 Canopy Biophysical Variables-driven Approach

The approach requires first to calibrate the inverse model: a parametric model rep-
resenting the inverse model is adjusted over a learning data set (Fig. 7.2, left). It
mainly consists in adjusting the parameters to fit a response surface between ref-
lectance values and the corresponding canopy variables of interest (LAI in this ex-
ample). Once calibrated, the parametric model is run to compute the variables of
interest from the observed reflectance values. The learning data set can be gener-
ated either using simulations of radiative transfer models, or based on concurrent
experimental measurements of the variables of interest and reflectance data.
176 F. Baret, S. Buis

Biophysical variables Radiometric data


driven approach driven approach

Learning data set


RT RT
LAI process Reflectance LAI Reflectance
process


Inverse
LAI* Reflectance
Model
Measurement

Parameters

Fig. 7.2 The two main approaches used to estimate canopy characteristics from remote sensing
data for LAI estimation. On the left side the approach focusing on the biophysical variables show-
ing the calibration of the inverse model. Once the inverse model is calibrated it can be applied
using the measured reflectance as input. On the right side, the approach focusing on radiometric
data showing the solution search process leading to the estimated LAI value, LAI ∗ . “∆” represents
the cost function to be minimized over the biophysical variables (left) or over the radiometric
data (right)

7.2.1.1 Calibration over Experimental Data Sets

This was the first approach historically used, the reflectance in few bands being
generally combined into vegetation indices (VI) designed to minimise the influence
of confounding factors such as soil reflectance and atmospheric effects (Baret and
Guyot, 1991). The relationships between VIs and canopy variables are calibrated
over experimental observations (Asrar et al., 1984; Huete, 1988; Wiegand et al.,
1990; Wiegand et al., 1992; Richardson et al., 1992). Recently Chen et al. (2002)
used simple VIs to derive LAI estimates from AVHRR and VEGETATION across
Canada. This was extended at the global scale by Deng et al. (2006). In agreement
with several observations, these authors found that the relationships vary from one
cover type to another as illustrated by Fig. 7.3. The development of such empir-
ical transfer functions is limited by the difficulty to get a training data base that
represents the whole range of possible conditions encountered over the targeted sur-
faces, i.e., combinations of geometrical configurations, type of vegetation and states
including variability in development stage and stress level, and type of background
and state (roughness, moisture). Measurement errors associated both to the variables
of interest and to radiometric data may also propagate into uncertainties and bi-
ases in the algorithm and should be explicitly accounted for Fernandes and Leblanc
(2005) and Huang et al. (2006). Further, since ground measurements having a foot-
print ranging from few meters to few decametres, specific sampling designs should
be developed to represent the sensor pixel. This task is obviously more difficult for
medium and coarse resolution sensors as outlined by Morisette et al. (2006). Higher
spatial resolution observations could be used to extend the local ground measure-
ments to the actual pixel size of medium or coarse resolution sensors.
7 Estimating Canopy Characteristics from Remote Sensing Observations 177

10

Coniferous
8 Deciduous
Mixed
Others

LAI 6

0
0 5 10
RSR VEGETATION

Fig. 7.3 Empirical relationships between LAI values as a function of the simple ratio vegetation
index (RSR) computed for VEGETATION data for four types of canopies (After Chen et al., 2002)

7.2.1.2 Calibration over Radiative Transfer Model Simulations

To avoid limitations associated to the empirical nature of the training data base, ra-
diative transfer models could be used alternatively to generate a training data base.
Radiative transfer models can be used to create a data base covering a wide range
of situations and configurations. Several authors have therefore proposed replac-
ing actual observations by numerical experiments based on radiative transfer model
simulations to calibrate empirical relationships (Sellers, 1985; Baret and Guyot,
1991; Rondeaux et al., 1996; Leprieur et al., 1994; Banari et al., 1996; Huete et al.,
1997; Verstraete and Pinty, 1996). Based on these principles, operational algorithms
developed for medium resolution sensors are currently used: MGVI for MERIS
(Gobron et al., 2000) further extended to other sensors, MODIS back-up algorithm
based on NDVI (Knyazikhin, 1999), POLDER algorithm based on DVI computed
from bidirectional reflectance factor (BRF) measurements normalized to a standard
geometrical configuration (Roujean and Lacaze, 2002). Nevertheless, although quite
often effective, VIs are intrinsically limited by the empiricism of their design and the
small number of bands concurrently used (generally 2–3). This might not be a major
problem for fAPAR and fCover variables that are relatively simple to estimate, but
would be more difficult for variables such as LAI or chlorophyll content (Cab ) show-
ing higher level of non linearity with reflectance measurements (Weiss et al., 2000).
The efficient interpolation capacity of neural network (NNT) can be exploited to
adjust surface responses (Leshno et al., 1993). Several authors have proposed such
an approach since the beginning of the 1990s (Smith, 1992; Smith, 1993; Atkinson
and Tatnall, 1997; Kimes et al., 1998; Abuelgasim et al., 1998; Gong et al., 1999;
178 F. Baret, S. Buis

Danson et al., 2003). Neural networks were compared with a specific implementa-
tion of multiple regression, the projection pursuit regression, and were concluded
to achieve very similar performances (Fang and Liang, 2005). Baret et al. (1995)
demonstrated that NNT used with individual bands were performing better than
classical approaches based on vegetation indices especially when calibrated with
radiative transfer model simulations rather than with experimental observations.
Weiss et al. (2002a) validated such techniques over a range of crops for estimating
the main canopy biophysical parameters LAI and fCover from airborne POLDER
instrument. Recently, several authors developed operational products for medium
resolution sensors, starting from top of canopy level: Lacaze (2005) for POLDER,
Bacour et al. (2006) for MERIS, and Baret et al. (2007) for VEGETATION instru-
ments. Baret et al. (2006b) proposed an operational algorithm from the MERIS top
of atmosphere data by coupling an atmospheric radiative transfer model to the sur-
face one, exploiting explicitly 13 over the 15 bands of MERIS.
Several ways may be used to build a data set for training empirical relationships
depending on the performances targeted. Evaluation of the performances of an al-
gorithm is generally achieved by computing the Root Mean Square Error (RMSE)
value over a test data base made of representative cases. Best performances will
therefore be obtained when the variables in the training data base are distributed
similarly to those in the testing one, i.e., close to the actual distribution of the vari-
ables: the coefficients of the empirical transfer function will be optimized for these
conditions, and uncertainties will be minimal for the most frequent cases. Although
achieving poorer performances in term of RMSE, a more even distribution of the
uncertainties may be alternatively obtained using uniform distributions of the vari-
ables. Note that, for a given number of cases simulated in the training data base, the
density of cases that populate the space of canopy realization may rapidly decrease
as a function of the number of required variables. Experimental plans may be used
in this situation as proposed by Bacour et al. (2002b), in order to focus on the first
order effects and interactions. Additionally, Baret et al. (2006b) proposed to steam-
line the data base in the reflectance space by retaining the cases that belong both to
the simulated and actual remote sensing measurements spaces (Fig. 7.4). This al-
lows discarding cases that were simulated but not actually observed. Conversely, it
allows also identifying cases which are observed but not simulated. This is achieved
by first compiling a large data base of reflectance measurements that should be rep-
resentative of the possible situations available. Then the reflectance mismatch is

Simulations Actual measurements


Cases not represented
Cases not represented
in the simulated database
in the measured database
Training database: selection
of cases in the intersection space

Fig. 7.4 Streamlining the simulated training data set by comparison to actual measurements. The
intersection between the space of simulated radiometric data (in dark gray) with that of the actual
measurements (in light gray) is used as the training data base
7 Estimating Canopy Characteristics from Remote Sensing Observations 179

computed for each case in the simulated data base: it is the minimum RMSE value
computed between the reflectance in the simulated data base and the ensemble of ac-
tual measurements. A threshold corresponding to the uncertainties in the radiometric
measurements is then used to decide whether a simulated case is rejected from the
training data base. Additional criterions could be used to streamline the training data
base, based on the expected consistency between several products such as LAI and
fAPAR as proposed by Bacour et al. (2006).
Although the use of radiative transfer models appears very appealing, this ap-
proach is however limited by several aspects. The first one is the capacity of the
models to get a faithful description of the radiative transfer in canopies. Up to now,
most radiative transfer models used are computer efficient ones allowing populating
large training data base within few hours/days with a single regular computer. They
generally correspond to simple description of canopy architecture which may not
represent the actual one, particularly regarding the clumped nature of many vegeta-
tion types. This leads to model uncertainties that may dominate all other sources of
uncertainties for some of the vegetation types. Recent advances in modeling more
complex canopy architecture (e.g., Soler et al., 2001; Lewis et al., 2004) offer great
potential for improvement. However, the second limitation will probably counter-
balance these advancements: building a realistic training data set requires a fair de-
scription of the distribution and co-distribution of the corresponding architectural
variables to define the actual space of canopy realization. For the simplest radia-
tive transfer models (e.g., Verhoef, 1984; Kuusk, 1995; Gobron et al., 1997) at least
three architectural variables are required (LAI, leaf angle distribution function and
size of the leaves relative to canopy height), the distribution of which being very
poorly known. This is even more difficult when using more complex and realistic
architectural description that requires more variables.
Note that in these approaches based on radiative transfer model simulations, ra-
diometric measurements uncertainties have to be added to the simulations when
building up the training data base. This allows more robustness within the training
process and thus improved retrieval performances. Accounting for these uncertain-
ties is also critical when large differences exist among bands used or when these
uncertainties are strongly correlated.
Canopy biophysical variables driven approaches present the advantage of being
very flexible. For example, estimates of biophysical variables from one sensor could
be used to constitute the training data base for another sensor. This could be applied
over high spatial resolution products that are aggregated to coarser spatial resolu-
tion to generate an appropriate training data base. This could also apply to generate
consistent products between sensors.

7.2.2 Radiometric Data-driven Approach

While the previous approach was focusing on minimizing the distance between the
variables retrieved from the inverse model and those from the training data set,
the alternative approach is based on finding the best match between the measured
180 F. Baret, S. Buis

reflectance values and those either simulated by a radiative transfer model or stored
within a database made of experimental observations. No proper calibration step
is required in this approach. However, several ingredients of these techniques are
difficult to evaluate (uncertainties, parameters of the search algorithm) and need
generally some tuning over a “prototyping” data set. The performances of the ap-
proach will both depend on the minimization algorithm itself and on the level of
ill-posedness of the inverse problem as a function of measurement configuration
and model and measurement uncertainties. Several minimization techniques have
been used: classical iterative optimization, simulated annealing (Bacour, 2001), ge-
netic algorithms (Fang et al., 2003; Renders and Flasse, 1996), look up tables and
Monte Carlo Markov Chains (Zhang et al., 2005). However, classical iterative op-
timization techniques (OPT) and look up tables (LUT) have been the most widely
used and will be described with more details below.

7.2.2.1 Iterative Optimisation (OPT)

This classical technique consists in updating the values of the unknown input bio-
physical canopy radiative transfer model variables until the simulated reflectance
closely fit the corresponding measurements (Goel and Deering, 1985; Kuusk, 1991a
and 1991b; Goel, 1984a and b; Pinty et al., 1990; Jacquemoud et al., 1995; Privette
et al., 1996; Bicheron and Leroy, 1999; Combal et al., 2000; Bacour et al., 2002a;
Combal et al., 2002). A good review on optimization methods used in remote sens-
ing for land applications can be found in Bacour (2001). The goodness of fit between
measured and simulated reflectance spectra is quantified by a cost function (J) that
may account explicitly for measurements and model uncertainties. The cost function
may be theoretically derived from the maximum likelihood (Tarentola, 1987). When
no prior information is available and when uncertainties associated to each config-
uration used are assumed independent and gaussian, J is assessed using norm L2,
i.e., sum over the N observational configurations of the square of the difference be-
tween the measured reflectance values (R) and those simulated (R̂), weighed by the
variance (σ2 ) associated to both reflectance measurements and model uncertainties:
N
(Rn − R̂n )2
J= ∑ σn2
(7.1)
n=1

However, because of the difficulty to provide an estimate of σ2 , several approxima-


tions have been used as shown in Bacour (2001). It spans from the simple ones such
as norm L1 to norm L2 with no weighing of the configurations, up to more complex
based on some modeling of the variance term (Table 7.1).
The main limitation of OPT techniques is twofold. (1) Firstly, the algorithm
might converge to a local minimum of the cost function that could be far away
from the global one expected to correspond to the actual solution. This can be partly
avoided by using a range of initial solutions, coupled with constraints on the range
of variation of the variables to be estimated. The use of a priori information in the
7 Estimating Canopy Characteristics from Remote Sensing Observations 181

Table 7.1 The cost functions (J) used in several studies dealing with radiative transfer model in-
version for canopy biophysical variables retrieval. N is the number of configurations (bands and
directions); R̂n and Rn being respectively the simulated and measured reflectance values for con-
figuration n. θν and φ are the zenith and relative azimuth view angles
Cost function References
N
Rn −R̂n
J= ∑ Rn Gao and Lesht, 1997)
n=1
N  

J = ∑  RnR−nR̂n  (Qiu et al., 1998)
n=1
N  2
Rn −R̂n
J= ∑ Rn +R̂n
(Gobron et al., 1997)
n=1
N
J = ∑ (Rn − R̂n )2 (Goel and Thompson, 1984; Pinty et al., 1990;
n=1 Privette et al., 1996; Braswell et al., 1996; Jacque-
moud et al., 2000; Combal et al., 2002)
N  2
Rn −R̂n
J= ∑ Rn (Nilson and Kuusk, 1989; Kuusk, 1991a and b;
n=1 Bicheron and Leroy, 1999; Weiss et al., 2000)
N  2
Rn −R̂n cos(θν ·sin(φ ))+1
J = ∑ ωn Rn ; ωn = 2 (Bacour et al., 2002a)
n=1

cost function generally improves the convexity of the error surface, which is critical
as we will see later (Combal et al., 2002). The descent algorithm may also limit
the trapping in a local minimum by reducing the rate of descent. However, a com-
promise has to be chosen between rapid convergence achieved with large descent
rate, and limiting the probability of falling in a local minimum achieved with a slow
descent rate. Further, the optimization algorithm may sometimes lack of robustness
due to numerical problems occurring generally with very small values of J. The
criterion used to stop the iterations is in addition not always easy to adjust, requir-
ing some preliminary tests (Bonnans et al., 2006). (2) Secondly, the OPT algorithm
requires large computer resources because of its iterative nature. However, there
are ways to speed up the process by limiting the number of model runs for each
iteration using the adjoint model that provides an analytical expression of the gradi-
ent of the cost function (Lauvernet et al., 2007). Nevertheless, OPT techniques are
still difficult to use routinely and exhaustively over large images, although image
segmentation may help reducing significantly the number of pixels to process, the
optimization process being restricted to a limited set of representative pixels. Note
that these techniques allow getting some estimates of the uncertainties associated
to the solution under some assumptions. However, the distribution of the solution
will be here always unimodal, conversely to what could be achieved with the other
radiometric driven approaches.
The main advantage of iterative optimization methods is their flexibility, allow-
ing retrieving canopy characteristics from several observational configurations. It is
even possible to invert radiative transfer models concurrently over several pixels.
This opens great potentials for exploiting additional temporal or spatial constraints
as we will see later.
182 F. Baret, S. Buis

7.2.2.2 Look Up Tables

This is conceptually the simplest technique, although its implementation is not triv-
ial (Weiss et al., 2000). It is the basis of the MODIS and MISR LAI and fAPAR
products (Knyazikhin et al., 1999). Firstly a large data base (the Look Up Table,
LUT) is generated, consisting of sets of input variables of the canopy radiative trans-
fer model used. Then, the corresponding reflectance values are simulated. The LUT
can alternatively be based on experimental observations, although this requires a
very good sampling of the space of canopy realization. Once the LUT has been
generated, finding the solution for a given set of reflectance measurements consists
in selecting the closest cases in the reflectance table according to a cost function,
and then extracting the corresponding set of canopy biophysical variables. Note that
the distribution of the solution could be obtained by accounting for the uncertain-
ties associated to the reflectance values as discussed by Knyazikhin et al. (1998a
and b).
This technique overcomes some of the limitations of iterative optimization tech-
niques. As a matter of fact, the search for the solution is global here, leading to
the true minimum if the space of canopy realisation is sufficiently well sampled.
Note that for generating the LUT, the space of canopy realization has to be sampled
to represent the surface response, i.e., with better sampling where the sensitivity
of reflectance to canopy characteristics is the higher (Weiss et al., 2000; Combal
et al., 2002). This is different from the sampling of the training data base required
in canopy biophysical variables driven approaches.
The implementation of a LUT technique in algorithmic operational chains is very
efficient because the radiative transfer model is run off-line. However, LUT tech-
niques require a fixed number of inputs unless having very large tables that could
be more difficult to manipulate. In addition, the way the solution is defined is not
always based on solid theoretical background. The cases selected as possible solu-
tions are either defined as a fraction of the initial population of cases (after tests
and trials) such as in Weiss et al. (2000) or Combal et al. (2002). It can be also
defined by a threshold corresponding to measurement and model uncertainties as in
Knyazikhin et al. (1998a and b).

7.2.2.3 Bayesian Methods: Importance Sampling and MCMC

Alternative methods are available which are based on statistical backgrounds: Monte-
Carlo Markov Chains (MCMC) and Importance Sampling (IS) (Makowski D.,
J. Hiller, et al., 2006). These two Bayesian methods approximate the posterior
distribution, i.e., the distribution of the variables when the reflectance measurement
is known. Although very little attention has been paid to these techniques at the
exception of Zhang et al. (2005) who used with success the MCMC Metropolis-
Hastings algorithm with MODIS data. However, Metropolis-Hastings algorithm is
an iterative process that might not be well suited for operational applications at
7 Estimating Canopy Characteristics from Remote Sensing Observations 183

large scale, similarly to OPT methods. Conversely, IS methods that do not require
multiple iterations might be efficient for this purpose and need to be properly eval-
uated for remote sensing applications.

7.3 The Under-determined and Ill-posed Nature of the Inverse


Problem in Remote Sensing

7.3.1 Under-Determination of the Inverse Problem

Estimating biophysical variables from remote sensing measurements is often an


under-determined problem: the number of unknowns is generally larger than the
number of independent radiometric information remotely sampled by sensors. In
the case of a simple canopy radiative transfer model such as SAIL (Verhoef, 2002),
canopy reflectance at the top of canopy (ρ toc ) for a given illumination and view
geometry (θ s, θ v, ϕ ) is simulated (Eq. (7.2)) using three variables describing
canopy structure that do not depend on wavelength (LAI, average leaf angle (ALA)
and hot spot parameter (hot) as modelled by Kuusk (1995)), and leaf reflectance
(refl) and transmittance (tran) as well as soil reflectance (Rs) that obviously depend
on wavelength (λ ).

ρtoc (λ, θs, θv, ϕ)


= CAN(LAI, ALA, hot, re f l(λ ), tran(λ ), Rs(λ, θs, θv, ϕ), θs, θv, ϕ) (7.2)

Several studies report that canopy (and soil) bidirectional reflectance distribution
function (BRDF) could be decomposed using empirical or semi-empirical orthog-
onal functions with generally 2–4 kernels (Lucht, 1998; Bréon et al., 2002; Weiss
et al., 2002). Therefore, 7–9 characteristics (3 canopy structure, 2 leaf properties
[refl, tran] input variables and the 2–4 terms describing soil BRDF, Rs(λ, θs, θv, ϕ)
have to be estimated out of a maximum of 4 independent information derived from
BRDF measurements in a single band. Retrieval of canopy characteristics from
BRDF measurements in a single band is therefore not possible without introducing
other information in the system, particularly when soil background plays a signifi-
cant role, i.e., for low to medium LAI values.
Similar observations are made when considering the reflectance spectral vari-
ation: leaf spectral properties may be described by a dedicated model such as
PROSPECT (Jacquemoud and Baret, 1990) requiring at least 5 input variables: mes-
ophyll structure parameter (N), chlorophyll (Cab ), dry matter (Cdm ), brown pigment
(Cbp ) and water (Cw ) contents:

[re f l(λ ), tran(λ )] = LEAF(N, Cab , Cdm , Cbp , Cw , λ ) (7.3)


184 F. Baret, S. Buis

Soil reflectance Rs(λ , θ s, θ v, ϕ ) may be described by a model such as that pro-


posed by Jacquemoud et al. (1992) and derived from that of Hapke (1981). It
requires a single scattering albedo ω (λ ) that varies with wavelength and soil com-
position, between 1 to 4 phase function coefficients (αi ), and a roughness parameter
(r). According to Price (1990), soil spectral variation, may be approximated as a
linear combination of 2–10 end-members. This is assumed to apply similarly to the
spectral variation of the single scattering albedo with weigh w j and end members
ω j (λ ):
ω (λ ) = ∑ w j · ω j (λ ) (7.4)
j

The whole soil spectral and directional reflectance field could subsequently be sim-
ulated with at least five parameters:

Rs(λ , θ s, θ v, ϕ ) = SOIL([w j ], [αi ], r, λ , θ s, θ v, ϕ ) (7.5)

Consequently, the whole spectral and directional top of canopy reflectance field
could therefore be modelled by coupling together the soil, leaf and canopy re-
flectance models, which leads to at least 13 input variables. These 13 unknowns
have to be estimated from the information content in remote sensing measurements.
Most of currently available sensors for which operational biophysical products are
available have a relatively small number of configurations: from two for AVHRR
(red and near infrared bands), to 15 bands for MERIS (VIS and NIR) and MODIS
(VIS, NIR, SWIR) with several bands dedicated to particular atmosphere, cloud,
snow/ice, or ocean characteristics. In the case of multidirectional sensors, the
number of configurations may be larger as in the case of MISR (36 configura-
tions = 9 cameras ×4 bands), or POLDER (84 configurations = 14 directions ×6
bands). However, the actual dimensionality of remote sensing measurements is
much smaller than the number of available configurations considering the relatively
high level redundancy between bands (Price, 1994; Price, 1990; Liu et al., 2002;
Green and Boardman, 2001) and directions (Zhang et al., 2002a and b; Weiss et al.,
2002b). Although further investigation is required to better quantify the actual
dimensionality of remote sensing observations, it is clear that retrieval of surface
characteristics from reflectance measurements is an under-determined problem in
many cases. Improving retrieval performances will require introducing ancillary
information and constraints in the system.

7.3.2 Evidence of the Ill-posed Problem

A problem is well posed if and only if its solution exists, is unique, and depends
continuously on the data (Garabedian, 1964). Several authors have reported that the
inverse problem in remote sensing is ill-posed (Knyazikhin et al., 1999; Combal
et al., 2001; Baret et al., 2000) because of its under-determination and uncertainties
attached to models and measurements. In addition, models may incorporate sets of
7 Estimating Canopy Characteristics from Remote Sensing Observations 185

0.3 80

Cab (µg.cm-2)
Reflectance

0.2 60

0.1 40

0 20
400 600 800 0 0.5 1 1.5 2
Wavelength (nm) LAI

Fig. 7.5 Actual reflectance measurements (left plot, solid lines representing the mean and stan-
dard deviations) and the corresponding closer simulations achieved with a simple turbid medium
radiative transfer model (the series of dots). On the right, the input LAI and Cab (the “+” symbols)
variables used to simulate the reflectance spectra shown on the left plot. The actual LAI and Cab
measurements are displayed with their associated confidence interval (bold line corresponding to
1 standard deviation). Data acquired over a sugar beet experiment conducted in 1990

variables that appear always in combinations such as products between variables. In


these conditions, very similar reflectance spectra simulated by a radiative transfer
model (Fig. 7.5, left) may correspond to a wide range of solutions (Fig. 7.5, right).
In the case illustrated by Fig. 7.5, high correlation is found between LAI and those
leaf chlorophyll content estimated values. This compensation between variables was
sometimes termed “ambiguity” (Baret et al., 1999) or “equi-finality” (Shoshany,
1991; Teillet et al., 1997). This may also indicate that the product LAI · Cab should
be used in place of individual estimates of LAI and Cab . Although not appearing
formally in the radiative transfer model, this product is physically meaningful from
the radiative transfer processes perspective and corresponds to the actual optical
thickness of the medium (Weiss et al., 2000).
Measurement and model uncertainties may also induce instability in the solu-
tion of the inverse problem. This is particularly true for well developed canopies,
where a small variation in the measured reflectance can translate into large varia-
tion of variables such as LAI, for which reflectance “saturates”, i.e., is very little
sensitive to LAI variation. A proper sensitivity analysis should help quantifying in-
teractions between input variables. A complementary sensitivity analysis conducted
over the cost function could also help evaluating the identifiability of the solution,
i.e., if output variables could be accurately retrieved from a given set of observations
(Salteli, 2004).
Regularization techniques are thus necessary to obtain a stable and reliable so-
lution of the ill-posed inverse problem. This could be achieved both by using prior
information on the distribution of the variables, and by exploiting some constraints
on the variables. These two issues will be investigated separately in the following.
186 F. Baret, S. Buis

7.4 Improving the Retrieval Performances

7.4.1 Using Prior Information

If no remote sensing measurement is available, the best estimates of the variables


would come from the prior information on their distribution (Fig. 7.6d), capital-
izing, all the knowledge coming from bibliography, past experiments or experts.
Conversely, when a radiative transfer model is available along with remote sensing
measurements, the variables can be estimated by inverting the RT model without
using any prior information. This will be illustrated using a simple example: esti-
mating LAI from NDVI vegetation index. In this case the RT model consists in an
analytical relationship as proposed by Baret and Guyot (1991):

NDV I = NDV I∞ + (NDV Is -NDV I∞ ) · e−K·LAI (7.6)

1 1 1
a b c
0.8 0.8 0.8

0.6 0.6 0.6


NDVI
NDVI

NDVI

0.4 0.4 0.4

0.2 0.2 0.2 Estimates from RS


Remote Sensing Radiative Transfer measurements and RT model
measurements (NDVI) model: NDVI=f(LAI)+e without prior information
0 0 0
0 0.02 0.04 0 2 4 6 8 0 2 4 6 8
PDF LAI LAI
0.05 1 0.05
d Prior information on LAI
e f NDVI estimates
with PI
0.04 0.8 0.04
Prior Information
0.03 0.6 0.03
NDVI
PDF

PDF

NDVI estimates
0.02 0.4 0.02 without PI

0.01 0.2 Estimates from RS 0.01


measurements and RT model
with prior information
0 0 0
0 2 4 6 8 0 2 4 6 8 0 2 4 6 8
LAI LAI LAI

Fig. 7.6 Estimation of canopy variables by combining remote sensing measurements, radiative
transfer model and prior information. All these pieces of information are represented by their prob-
ability distribution function (PDF): (a) PDF of remote sensing measurements in the simple case of
NDVI; (b) PDF of RT model simulations (NDV I = f(LAI)) accounting for model uncertainties; (c)
PDF of LAI as retrieved from RT model and NDVI measurement and their associated uncertain-
ties, without using prior information; (d) PDF of LAI used as prior information; (e) Computation
of LAI PDF as estimated from NDVI measurements and RT model, using prior information on LAI;
(f) PDF of the solution (posterior distribution) when using only prior information (idem as plot d),
using RT model and NDVI measurements and their associated uncertainties only, and using all the
information available (RT model and NDVI measurements and their associated uncertainties and
prior information). The three contour plots (b, c, e) are coded from white to black for zero to max
PDF values with the same gray scale. Very simple assumptions on uncertainties models and values
are used here just for illustration
7 Estimating Canopy Characteristics from Remote Sensing Observations 187

with NDV Is and NDV I∞ being respectively the bare soil and asymptotic values
of NDVI, and K an extinction coefficient (K = 0.8). However, uncertainties are
associated both with remote sensing measurements (Fig. 7.6a NDV Ir = ℵ(0.8, 0.1)
where ℵ(x, σ2 ) means a Gaussian distribution with mean x and variance σ2 ) and
the RT model (Fig. 7.6b RT model represented by Eq. (7.2) with a Gaussian noise
ℵ(0, 0.1)). Accounting for these uncertainties in the form of the corresponding
probability distribution function (PDF) allows deriving the PDF of the estimated
variable (Fig. 7.6c). The small sensitivity of NDVI to LAI as compared to mea-
surement and model uncertainties induce a relatively broad PDF for the larger
LAI values (Figs. 7.6c, f). This corresponds to an ill-posed problem, where a wide
range of possible solutions match very similar measurements. The combination of
RT model, remote sensing measurements and prior information on the variables
(here LAI = ℵ(2, 1.5) allows getting more reliable solutions accounting for all the
sources of information available in an optimal way (Figs. 7.6e, f).
The example provided above for a measurement value of NDV I = 0.8 could be
extended to the whole range of NDVI values. It shows that the mode of the dis-
tribution of the solution corresponding to the maximum likelihood (maximum of
the PDF) strongly depends on the type of input information used (Fig. 7.7, left).
When only prior information is used, the mode stays constant and obviously inde-
pendent from measurements. When RT model and measurements are used with their
uncertainties, the LAI mode is generally close to the values obtained without con-
sidering uncertainties, assuming perfect model and measurements. However, over
the saturation domain corresponding to NDVI values higher than 0.85, accounting
for the uncertainties provides lower modal values because of the non linearity of the

6 3
Mode RT model
Standard deviation
5 2.5
RT model and RT model and
measurement measurement
uncertainties uncertainties
4 2
RT model and
LAI

LAI

3 measurement 1.5
uncertainties and
prior information Prior information
2 1
Prior information
RT model and
measurement
1 0.5 uncertainties and
prior information

0 0
0 0.5 1 0 0.5 1
NDVI NDVI

Fig. 7.7 Mode (plot on the left) of the distribution of the solution (LAI) of the inverse problem
as a function of the measured value (NDVI). The mode corresponds to the maximum PDF value,
i.e., the maximum likelihood. Four estimates are displayed: using only prior information; using RT
model (LAI = RT −1 (NDV I)) assumed to be perfect with perfect measurements (no uncertainities
accounted for); using RT model and measurements with their associated uncertainities; using RT
model and measurements with their associated uncertainities and prior information. On the right,
the standard deviation of the distribution of the solution is also displayed for the several cases. The
case with perfect RT model and measurements is not displayed here because its standard deviation
is null by definition
188 F. Baret, S. Buis

model. When prior information is used in addition to RT model and remote sensing
measurements, differences of LAI mode are marginal over the domain where NDVI
is sensitive enough to LAI. Conversely, over the saturation domain, LAI modal val-
ues are always lower (closer to prior information value) than those observed when
not using prior information which would lead to a bias. However, the interest of
using the prior information is clearly demonstrated when considering the standard
deviation of the distribution of the solutions (Fig. 7.7, right).
Introducing prior information in the inversion process provides a very significant
reduction of the variability of the posterior distribution. This is obviously more im-
portant for the larger NDVI values corresponding to the saturation domain: in this
case, very large scattering of the retrieved LAI values is expected when no prior
information is used. Although the maximum likelihood is often used as “the solu-
tion”, the variability within the posterior distribution as represented by its standard
deviation appears to be very informative and useful.
The theory behind this Bayesian approach has been extensively described by
Tarantola (2005). When restricting the solution as that maximizing the likelihood,
i.e., corresponding to the maximum of the PDF, a general formulation of the cost
function may be derived under Gaussian distribution assumption:

J = (R − R̂)t ·W −1 · (R − R̂) + (V̂ −Vp )t ·C−1 · (V̂ −Vp ) (7.7)


     
Radiometricinformation Prior information

where V̂ is the vector of the input biophysical variables estimates, R corresponds to


the vector of remote sensing measurements of dimension N (the number of bands
and directions used), R̂ is the vector of the simulated reflectance corresponding to
the solution V̂ (the vector of canopy biophysical variables) and Vp the vector of
prior values of biophysical variables. Matrices W and C are the covariance matrices
characterizing respectively the radiometric and model uncertainties, and that of the
prior information. Note that the first part of this equation corresponds to the distance
between the measured and the simulated radiometric data. It simplifies into Eq. (7.1)
if the covariance terms of matrix W are assumed to be zero, i.e., measurement and
model uncertainties are independent between configurations. The second part of
Eq. (7.7) corresponds to the distance between the values of the estimated variables
and those of the prior information. Very few studies are currently based on this
formulation of the cost function where prior information is explicitly used (Combal
et al., 2002).
Implementing the cost function as expressed by Eq. (7.7) requires some reason-
able estimates of covariance matrices W and C as well as prior values Vp . The terms
of W should reflect both measurement and model uncertainties. While some rough
estimates of measurement uncertainties could be derived from the sensor specifica-
tion, model uncertainties are far more difficult to estimate. Further, they may de-
pend significantly on the type of situation considered, such as low or high vegeta-
tion amount. Even more difficult to estimate, are the covariance terms in W : mea-
surement and model uncertainties may have important structure that translates into
high covariance terms which are however very poorly known. When using simul-
7 Estimating Canopy Characteristics from Remote Sensing Observations 189

taneously a large number of configurations as in the case of hyperspectral observa-


tions, these covariance terms will be very important to account for: they will allow
weighing properly the several configurations used. The difficulty to estimate the
covariance terms explains why a small number of configurations is often selected
when a larger number is available as in the case of hyperspectral and/or directional
observations.
Retrieval approaches should be used within well defined and if possible restricted
domains. Larger domains will generally degrade retrieval performances since the
prior information will be looser defined, similarly to the covariance matrices char-
acterizing uncertainties. However, splitting the whole domain into a set of sub-
domains may introduce problems due to misclassification and attribution errors as
observed by Lotsch et al. (2003), and artefacts at the limit between classes translat-
ing into more chaotic spatial or temporal variation of the solution.
The way prior information is introduced in the inversion process depends on the
inversion technique used. The cost function represented by Eq. (7.7) is used within
iterative optimization and LUTs. Bayesian methods include the a priori distribution
through the use of the Bayes theorem to estimate the a posteriori distribution. For
biophysical variables driven approaches the training data base should reflect the
actual knowledge on the distribution of the variables. Note that the difficulty in
defining explicitly the covariance terms in the uncertainties on remote sensing inputs
(RT model and measurements) for the radiometric data driven approaches remains
in the biophysical variables driven approaches for the generation of the training data
base. However, implicit introduction of these terms may be achieved when using a
training data base made from actual satellite measurements as suggested by Bacour
et al. (2006).

7.4.2 Using Additional Constraints

7.4.2.1 Coupling Models

The radiative transfer in each element of the soil/leaf/canopy/atmosphere system is


strongly coupled to the radiative transfer in the whole system. The simple example
given previously to demonstrate the under-determined nature of the inverse problem
in remote sensing shows that top of canopy reflectance could be written as:
ρ toc (λ , θ s, θ v, ϕ ) = CAN(LAI, ALA, hot, LEAF(N, Cab , Cdm , Cbp , Cw ),
SOIL ([w j ], [αi ], r, λ , θ s, θ v, ϕ ) , θ s, θ v, ϕ ) (7.8)

The same applies when retrieving some characteristics of the system from top of
atmosphere reflectance (ρ toa ) measurements as usually achieved by sensors aboard
satellite:

ρ toa (λ , θ s, θ v, ϕ ) = ATM(ρ toc (λ , θ s, θ v, ϕ ), τ550 , Ä, Patm , Cwv , C03 , λ , θ s, θ v, ϕ )


(7.9)
190 F. Baret, S. Buis

where ATM represents an atmospheric RT model such as 6S (Vermote et al., 1997)


or MODTRAN (Berk et al., 1998), τ550 , Ä, being respectively the aerosol optical
thickness at 550 nm and the Angström coefficient, Patm is the atmospheric pressure,
Cwv is the water vapor content and C03 the ozone content.
Retrieval of characteristics of some element of the system without solving (im-
plicitly or explicitly) the whole system will therefore be sub-optimal as demon-
strated below.
Let consider retrieving leaf biophysical properties [N, Cab , Cdm , Cbp , Cw ] from
top of canopy remote sensing observations in B wavebands using a decoupled sys-
tem and an iterative optimization technique. For sake of simplicity, soil reflectance
will be assumed to be known. Estimates of leaf properties could be achieved in
two steps. First, estimate the variables [LAI, ALA, hot, re f l(λ ), tran(λ )] from the
reflectance in each of the B bands. A cost function accounting for the reflectance
in the B bands should be minimized with the constraint that [LAI, ALA, hot] does
not vary with wavelength. The number of unknowns in the system will therefore be
(3 + 2 · B) corresponding to the 3 canopy structure variables and the 2 (reflectance
and transmittance) leaf optical properties time the B bands. The second step of the
process consists in estimating leaf biophysical properties [N, Cab , Cdm , Cbp , Cw ]
from the retrieved leaf reflectance and transmittance in the B bands. The variables
[N, Cab , Cdm , Cbp , Cw ] are tuned by minimizing a cost function accounting for
leaf reflectance and transmittance in the B bands. Obviously, increasing the number
of bands will not improve the underdetermined nature of the problem because the
number of unknowns in the first step of the process will grow twice faster. In addi-
tion, since no biophysical constraints are set on the spectral variation of leaf optical
properties, canopy structure variables derived from the first step may express larger
and unrealistic range of variation. The proper way to solve this type of problem is to
minimize a cost function accounting for canopy reflectance over the B wavebands
based on the coupled leaf and canopy models. In this case, the number of unknowns
will be eight (the three canopy structure variables and the five leaf characteristics)
which is independent from the number of wavebands used. This allows limiting the
under-determined nature of the problem by increasing the spectral sampling.
Most of the retrieval approaches from top of canopy radiometric observations are
now using implicitly or explicitly coupled models as shown in Table 7.2. However,
although offering great potentials as demonstrated recently (Baret, 2006b), the use
of coupled atmosphere/surface models is still not very well developed because each
sub-problem was handled by different communities.

7.4.2.2 Spatial Constraints

Up to now, most retrieval algorithms are applied to independent pixels, neglecting


the possible spatial structure as observed on most images. However, some authors at-
tempted to exploit these very obvious patterns at high spatial resolution. The “object
retrieval” approach proposed by Atzberger (2004) is based on the use of covariance
between variables as observed over a limited cluster of pixels representing the same
7 Estimating Canopy Characteristics from Remote Sensing Observations 191

Table 7.2 Synthesis of the several algorithms currently used operationally to retrieve canopy bio-
physical variables. 1: (Lacaze, 2004); 2: (Knyazikhin et al., 1999); 3: (Gobron et al., 1999); 4:
(Weiss et al., 2002; Baret et al., 2007); 5: (Chen et al., 2002; Deng et al., 2006); 6: (Bacour et al.,
2006)
RT models Inversion prior
# Algorithm uncertainties
leaf soil Canopy Atmosphere technique information

PROSPECT
POLDER PRICE Kuusk Some variables fixed
1 N, Cab, TOC NNT measurements
LAI, fAPAR 2 abundances LAI, ALA, hot Range of variation
(Cw,Cdm, Cs)

Hapke measurements
MODIS/MISR prescribed for DISORD specific values for 6
2 3 typical + TOC LUT prescribed at
LAI, fAPAR each biome 6 biomes biomes
understorey 20%
(1)
MERIS PROSPECT
5 typical soil Gobron TOA Range of variation
3 MGVI N, Cab, Parametric not specified
unique BRDF LAI, ALA, hot (MODTRAN) (uniform distribution)
fAPAR (Cw,Cdm, Cs)

(2)
VEGETATION brightness SAIL model and
PROSPECT
CYCLOPES parameter LAI, ALA, hot, measurements approximation of actual
4 N, Cab, TOC NNT
LAI, fAPAR, &reference vCover prescribed at 4% distribution
Cw,Cdm, Cbp
fCover spectra (relative)

Empirical relations for specific biomes


VEGETATION-
using TM sensor and the corresponding ground Specific relations for
5 Canada-Global TOC Parametric not specified
measurements over some sites each biome
LAI
Prescribed BRDF model
PROSPECT
brightness 2 versions: model and
MERIS green/brown SAIL,
parameter - TOC version measurements approximation of actual
6 LAI, fAPAR, separated LAI, ALA, hot, NNT
&reference - TOA version prescribed at 4% distribution
fCover, LAIxCab N, Cab, Cdm, vCover
spectra (SMAC) (relative)
Cw, Cbp

class of object such as an agricultural field. Results show quite significant improve-
ment of the retrieval performances for LAI, Cab and Cw , presumably because of a
better handling of the possible compensation between LAI and ALA in the retrieval
process as suggested by Atzberger (2004) and outlined by Jacquemoud (1993).
Other approaches based on models with random effects (Faivre and Fischer,
1997) may be also very attractive, although rarely used within the land remote
sensing community. They allow characterizing a population by their two first statis-
tical moments (mean and variance). In the case of remote sensing applications, this
could be applied over a cluster of P pixels belonging to the same class of surface as
in the “object retrieval” approach of Atzberger (2004). The inversion process could
be achieved by tuning both the mean and variance values of each input variable over
the P pixels using iterative optimization techniques. The individual values of each
pixel could be derived from the estimated mean and variance values of the vari-
ables and the departure between the actual radiometric measurements of the pixels
and the mean values over the object. The under-determination of the problem could
significantly decrease with this approach: the number of unknowns to estimate is
independent on the number of pixels considered in the cluster and is just twice the
number of variables to estimate (mean and variance).
Although quite promising, these methods need further evaluation, and probably
adaptation before being accepted and used by the remote sensing community. Note
that only statistical distributions are used for both methods presented, although addi-
tional geo-statistical constraints could be exploited particularly for the higher spatial
resolutions, based on variograms (Garrigues et al., 2006).
192 F. Baret, S. Buis

7.4.2.3 Temporal Constraints

The dynamics of canopies results from elementary processes under the control of
climate, soil and the genetic characteristics of the plants that incrementally change
canopy structure and optical properties of the elements. Very brutal and chaotic time
course are therefore not expected, at the exception of accidents such as fire, flooding,
harvesting, or lodging. The smooth character of the dynamics of canopy variables
may be exploited as additional constraint in the retrieval process. Kötz et al. (2005)
proposed using a semi-empirical model of canopy structure dynamics to improve
remote sensing estimates of LAI over maize crops. Results show a significant im-
provement of estimates, particularly for the larger LAI values where saturation of re-
flectance is known to be a problem. This approach requires a semi-empirical model
of canopy structure dynamics (here LAI) describing the whole growth cycle with
few parameters. In the case of the model used by Kötz et al. (2005) five parameters
are needed. In this case, the under-determined nature of the inverse problem will
decrease only if more than five dates of remote sensing observations are available
and well distributed over the growth cycle. However, because the parameters of the
model of LAI dynamics have some biological meaning, prior information on them
could be accumulated and efficiently exploited.
More recently, Lauvernet et al. (2007) proposed a “multitemporal patch” in-
version scheme to account for both spatial and temporal constraints. Reflectance
data are here considered observed from top of atmosphere. Atmosphere/canopy/
leaf/soil RT models are thus coupled to simulate top of atmosphere reflectance from
the set of input variables as stated by Eqs. (7.8) and (7.9). Spatial and temporal
constraints are based on the assumption that the atmosphere is considered stable
over a limited area (typically few kilometres) but varies from date to date, and
that surface characteristics vary only marginally over a limited temporal window
(typically ±7 days) but may strongly change from pixel to pixel. This has obviously
important consequences on the under-determined nature of the inverse problem as
demonstrated hereafter. The atmosphere characteristics [Patm , Cwv , C03 ], except the
aerosol ones [τ550 , Ä], are assumed to be known from independent observations
such as meteorological estimation or dedicated sensors or algorithms. The observa-
tional configuration [λ, θ s, θ v, ϕ ] is also known at the time of image acquisition.
Soil reflectance was simply approximated as lambertian, with reflectance propor-
tional to a reference soil spectra according to a brightness parameter Bs (Bacour
et al., 2006). The brightness parameter is assumed to vary both from date to date
and pixels to pixels, without any constraints. The forward model resulting from
nesting the RT models presented previously could be written as a function of the
ten unknowns [N, Cab , Cdm , Cbp , Cw , LAI, ALA, hot, Bs, τ550 , Ä] with nA = 2
atmosphere variables, nC = 8 canopy and leaf variables and ns = 1 soil variables.
Let consider d dates of observation available over a limited temporal window during
which the canopy variables are about constant, and a spatial window of p pixels
for which the atmosphere is considered homogeneous. The number of unknowns,
N(p, d) in the case of concurrent inversion of an ensemble of d dates and p pixels
using the spatial and temporal constraints described above is therefore:
7 Estimating Canopy Characteristics from Remote Sensing Observations 193

N(p, d) = d · nA + p · nC + d · p · nS (7.10)

Inverting the nested radiative transfer models concurrently over an ensemble of d


dates and p pixels will significantly reduce the total number of unknowns. (N(p, d))
as compared to p times d independent instantaneous pixel inversions (p.d.N(1, 1)).
Figure 7.8 shows that the number of unknowns to be estimated within the same
inversion process for p pixels and d dates as compared to p.d single pixel and
single date inversions (N(p, d)/(p.d.N(1, 1))) decreases significantly up to about
10 pixels. However, the main advantage over “ensemble” inversion is reached when
applying concurrently the inversion process to several dates. Using two dates and
more than 10 pixels allows dividing by almost 2 the number of unknowns. Note that
these results concern only the number of unknowns, and is therefore applicable to
any observational configuration characterized by a set of bands and directions.
Results on the performances achieved demonstrate the interest of the approach
for the estimation of most of the variables, particularly for the aerosol characteristics
and for LAI, LAI × Cab and ALA canopy characteristics. However, again, this new
approach was only demonstrated over RT simulations, and its interest should be
verified over experiments with actual remote sensing data and the corresponding
ground truth.

0.9
d=1

0.8

0.7
N(p,d) / (p.d.N(1,1 ))

0.6

0.5
d=2

0.4
d=3
0.3
d=4
d=5
0.2 d=6

0.1

0
0 5 10 15 20 25 30 35 40 45 50
Number of pixels (p)

Fig. 7.8 Ratio, N(p, d)/(p.d.N(1, 1)), between the number of unknowns when applying the inver-
sion process concurrently to p pixels and d dates with that of single pixel and date inversion as a
function of the number of dates and pixels considered
194 F. Baret, S. Buis

7.5 Conclusion

This overview of retrieval approaches is based on methods currently used, while


alternative ways to solve the problem and hopefully improve the accuracy and ro-
bustness of estimates were briefly introduced. Several ingredients of the algorithms
were identified apart from the retrieval techniques themselves: radiative transfer
models, observations, additional information and constraints. We will briefly sum-
marize the conclusions for each of these ingredients in the following.

7.5.1 Retrieval Techniques

The several techniques investigated have been classified as radiometric variables or


biophysical variables driven approaches. However, both types of methods could be
either derived from actual measurements or based on radiative transfer model simu-
lations. The best approaches are obviously the ones that will be trained over data sets
that are as close as possible to the evaluation data set. For this very reason, canopy
biophysical variables trained over empirical data sets would be ideal. In addition,
canopy biophysical variables driven approaches present the advantage of being very
computer efficient once trained, allowing easy implementation within operational
processing chains. However, because of the difficulty of getting a large enough
training data set representing the actual distribution of cases (observational config-
uration, type of canopies and state, background properties, eventually atmosphere
characteristics), training data base made of radiative transfer model simulations is
preferred. These hybrid techniques as termed by Liang (2004) require however the
radiative transfer models to be well adapted to the type of canopy they target, and
their adequacy to be quantified to properly input model uncertainties. In addition,
the structure of uncertainties on the radiometric variables and distribution and co-
distribution of the input biophysical variables should be also known. An alternative
approach currently not yet explored would consist in bridging the two retrieval ap-
proaches: actual sensor measurements are used to build the training data base allow-
ing to keep all the structure of measurement uncertainties. This data base should be
representative of the cases investigated, which might be possible by specific spatial
and temporal sampling schemes as proposed by Baret et al. (2006a) in the case of
global observations. The corresponding best estimates of canopy biophysical vari-
ables could be derived from inversion methods such as iterative optimization tech-
niques for which all the information available should be exploited: fusion of all
currently available sensors observations, prior information and spatial and temporal
constraints.
As a matter of fact, most radiometric variables driven approaches are very flex-
ible and could easily ingest data from several sensors, bands and directions, at the
expense of computer requirements which make these methods more difficult for an
operational use. Conversely, canopy biophysical driven approaches are not as flex-
ible as radiometric driven approaches: they are generally tuned for a limited set of
7 Estimating Canopy Characteristics from Remote Sensing Observations 195

observational conditions: using other configurations would require a specific train-


ing or a dramatic enlargement of the training data base.
Retrieval methods will be more efficient when applied to a limited set of surface
types as compared to a very generic (global) solution. Approaches based on a classi-
fication would thus allow closer adaptation to each class of both the radiative transfer
model and prior information. However, attribution errors may significantly alter the
performances. Using a continuous classification (Hansen et al., 2002; Hansen and
DeFries, 2004; Schwartz and Zimmermann, 2005) will probably limit this source
of uncertainties and avoid getting artefacts when two consecutive pixels will jump
from one class to another.
Biophysical variables estimates are generally integrated within other process
models such as hydrology or biogeochemical cycling along with other ground obser-
vations. Quantification of the associated uncertainties is therefore required to prop-
erly merge these several sources of information. Current available products did not
provide quantitative evaluation of the confidence interval around the solution, but
are limited to qualitative indices. Bayesian approaches provide a direct access to the
distribution of the solution of the inverse problem and may be very useful for esti-
mating the uncertainties. Current operational algorithms need further developments
to fully satisfy this important user requirement.

7.5.2 Radiative Transfer Models

Performances of methods based on radiative transfer models are largely depending


on the realism of the simulations. Radiative transfer models are based on a set of
assumptions, particularly regarding the description of canopy architecture. A more
realistic description of canopy architecture will require additional input variables
and will be probably more demanding in computer resources. Knowledge of prior
distribution and co-distribution of these additional canopy structure variables will
constitute a limitation. Further, using such more realistic radiative transfer model
requiring a larger number of unknowns will not necessarily improve the retrieval
performances because the under-determination of the problem will be even more
limiting. A compromise should therefore be found between the realism of the de-
scription of canopy structure, and its complexity.
Particular attention should be paid on the definition of the variables used in the
radiative transfer model that should match the one required for the application. For
example, the original LAI definition (Stenberg, 2006) may be altered depending
on the way and scale at which leaf clumping is accounted for (Chen and Leblanc,
1997). Great caution should be also paid when comparing retrievals with ground
measurements or inter-comparing several products.
As demonstrated here, holistic approaches based on the coupling of canopy, leaf
and soil models are optimal for best performances. Eventually, coupled surface and
atmosphere models would certainly help solving in an elegant way the retrieval of
surface variables from top of the atmosphere observations.
196 F. Baret, S. Buis

7.5.3 Observations and Ancillary Information

The observational configuration is an important element that drives the accuracy


of canopy biophysical variables estimation. It depends obviously on the variables
targeted. For the time being, sufficient maturity is achieved for the estimation of
LAI, fAPAR, the cover fraction, chlorophyll and water contents variables to imple-
ment operational algorithms for delivering the corresponding products to the user
community. The interest of multidirectional and hyperspectral observations is still
to be rigorously demonstrated for these variables by comparison over actual ground
measurements.
Frequent observations are required to monitor the dynamics of the vegetation that
conveys a large amount of information on the functioning of the surface. With the
hopefully venue of systems capable of high revisit frequency with high spatial reso-
lution, new retrieval methods should be developed to exploit the temporal and spatial
dimensions in addition to the more classical spectral and in a lesser way directional
ones. This would allow benefiting from the spatial and temporal constraints and con-
sequently reduce the number of unknowns to be retrieved. Ultimately, this approach
will converge towards direct assimilation of top of atmosphere radiances into sur-
face process models. However, the research community is not mature enough on the
coupling between radiative transfer models and canopy process models. Radiative
transfer model inversion had still to mature and improve the accuracy of surface
variables estimation before jumping towards radiance data assimilation.
Knowledge and management of uncertainties is one of the critical issues for the
retrieval algorithms. If measurement uncertainties coming from the sensor are rel-
atively well known, their structure (covariance between bands and directions for
example) is poorly documented. This is even worse when considering model uncer-
tainties that may change dramatically from place (and time) to place (and time) with
presumably specific features (covariance between configurations).
The other critical issue is the lack of prior information on the distribution of
most land surface attributes. However, this could be accumulated from the numer-
ous experiments organized in support of satellite images. A mechanism should thus
be developed to capitalize on the information gathered within the remote sensing
research community as well as other communities working with ecosystems. Note
that getting high spatial resolution data will considerably ease the characterization
of prior distribution of the variables, provided that each pixel could be properly
classified.
Any retrieval algorithm should be properly validated before delivering its prod-
ucts to the user community according to consensus protocols (Morisette et al.,
2006). This process will not only provide a way to characterize the associated uncer-
tainties, it will be also critical for improving the algorithms. A short feedback loop
should therefore be set-up between algorithm prototyping and validation. When re-
trieval algorithms are based on radiative transfer modeling, this will implicitly merge
observations and model to improve robustness and accuracy of the products at the
expense of a decrease in the desired independency between the validation and cali-
bration processes.
7 Estimating Canopy Characteristics from Remote Sensing Observations 197

References

Abuelgasim AA, Gopal S, Strahler AH (1998) Forward and inverse modelling of canopy direc-
tional reflectance using a neural network. Int. J. Remote Sens. 19(3):453–471
Asrar G, Fuchs M, Kanemasu ET, Hatfield JL (1984) Estimating absorbed photosynthetic radiation
and leaf area index from spectral reflectance in wheat. AJ, 76:300, 306
Asrar G, Myeni RB, Kanemasu ET (1989) Estimation of plant canopy attributes from spec-
tral reflectance measurements. Theory and Application of Optical Remote Sensing. J. Wiley,
Indianapolis: 252–296
Atkinson PM, Tatnall ARL (1997). Neural network in remote sensing. Int. J. Remote Sens.
18(4):699–709
Atzberger C (2004) Object-based retrieval of biophysical canopy variables using artificial neural
nets and radiative transfer models. Remote Sens. Environ. 93:53–67
Bacour C (2001) Contribtion à la détermination des paramètres biophysiques des couverts végétaux
par inversion de modèles de réflectance: analyse de sensibilité et configurations optimales. PhD
Thesis, Université Paris 7 – Denis Diderot, Paris (France), 206 pp
Bacour C, et al. (2002a) Reliability of the estimation of vegetation characteristics by inversion of
three canopy reflectance models on airborne POLDER data. Agronomie, 22:555–566
Bacour C, Jacquemoud S, Tourbier Y, Dechambre M, Frangi JP (2002b) Design and analysis of
numerical experiments to compare four canopy reflectance models. Remote Sens. Environ.
79:72–83
Bacour C, Baret F, Béal D, Weiss M, Pavageau K (2006) Neural network estimation of LAI, fA-
PAR, fCover and LAIxCab, from top of canopy MERIS reflectance data: principles and valida-
tion. Remote Sens. Environ. 105:313–325
Banari A, Huete AR, Morin D, Zagolski F (1996) Effets de la couleur et de la brillance du sol sur
les indices de végétation. Int. J. Remote Sens. 17(10):1885–1906
Baret F, Guyot G (1991) Potentials and limits of vegetation indices for LAI and APAR assessment.
Remote Sens. Environ. 35:161–173
Baret F, Clevers JGPW, Steven MD (1995) The robustness of canopy gap fraction estimates
from red and near infrared reflectances: a comparison of approaches. Remote Sens. Environ.
54:141–151
Baret F, Knyazikhin J, Weiss M, Myneni R, Pragnère A (1999) Overview of Canopy Biophysical
Variables Retrieval Techniques. ALPS’99, Meribel, France, CNES, Paris
Baret F, Weiss M, Troufleau D, Prévot L, Combal B (2000) Maximum information exploitation for
canopy characterisation by remote sensing. Remote sensing in agriculture. Aspects of Applied
Biology. Association of Appied Biologists, pp. 71–82
Baret F, et al. (2006a) Evaluation of the representativeness of networks of sites for the global valida-
tion and inter-comparison of land biophysical products. Proposition of the CEOS-BELMANIP.
IEEE Trans.Geosci. Remote Sens. 44(7:special issue on global land product validation):1794–
1803
Baret F, et al. (2006b) Algorithm Theoretical Basis Document for MERIS Top of Atmosphere Land
Products (TOA VEG), INRA-CSE, Avignon
Baret F, et al. (2007) LAI, fAPAR and fCover CYCLOPES global products derived from VEGE-
TATION. Part 1: principles of the algorithm. Remote Sens. Environ. 110:275–286
Berk A, et al. (1998) MODTRAN cloud and multiple scattering upgrades with application to
AVIRIS. Remote Sens. Environ. 65:367–375
Bicheron P, Leroy M (1999) A method of biophysical parameter retrieval at global scale by inver-
sion of a vegetation reflectance model. Remote Sens. Environ. 67:251–266
Bonnans, JF, Gilbert, JC, Lemaréchal, C, and Sagastizábal, CA (2006) Numerical Optimization.
Theoretical and Practical Aspects (Second Edition). Springer, New-York. 494pp.
Braswell BH, et al. (1996) Extracting ecological and biophysical information from AVHRR optical
data: an integrated algorithm based on inverse modeling. J. Geophys. Res. 101(D18):23335–
23348
198 F. Baret, S. Buis

Bréon FM, Maignan F, Leroy M, Grant I (2002) Analysis of hot spot directional signatures mea-
sured from space. J. Geophys. Res. 107(D16):AAC 1 1–15
Chen JM, Leblanc SG (1997) A four-scale bidirectional reflectance model based on canopy archi-
tecture. IEEE Trans. Geosci. Remote Sens. 35:1316–1337
Chen JM, et al. (2002) Derivation and validation of Canada wide coarse resolution leaf area index
maps using high resolution satellite imagery and ground masurements. Remote Sens. Environ.
80:165–184
Combal B, Ochshepkov SL, Sinyuk A, Isaka H (2000) Statistical framework of the inverse problem
in the retrieval of vegetation parameters. Agronomie 20:65–77
Combal B, Baret F, Weiss M (2001) Improving canopy variables estimation from remote sens-
ing data by exploiting ancillary information. Case study on sugar beet canopies. Agronomie
22:2–15
Combal B, et al. (2002) Retrieval of canopy biophysical variables from bi-directional reflectance
data. Using prior information to solve the ill-posed inverse problem. Remote Sens. Environ.
84:1–15
Danson FM, Rowland CS, Baret F (2003) Training a neural network with a canopy reflectance
model to estimate crop leaf area index. Int. J. Remote Sens. 24(23):4891–4905
Deng F, Chen JM, Chen M, Pisek J (2006) Algorithm for global leaf area index retrieval using
satellite imagery. IEEE Trans. Geosci. Remote Sens. 44(8):2219–2229
Faivre R, Fischer A (1997) Predicting crop reflectances using satellite data observing mixed pixels.
J. Agric., Biol. Environ. Stat. 2:87–107
Fang H, Liang S (2005) A hybrid inversion method for mapping leaf area index from MODIS
data: experiments and application to broadleaf and needleleaf canopies. Remote Sens. Environ.
94:405–424
Fang H, Liang S, Kuusk A (2003) Retrieving leaf area index using a genetic algorithm with a
canopy radiative transfer model. Remote Sens. Environ. 85:257–270
Fernandes R, Leblanc SG (2005) Parametric (modified least squares) and non parametric (Theil-
Sen) linear regressions for predicting biophysical parameters in the presence of measurements
errors. Remote Sens. Environ. 95:303–316
Gao W, Lesht BM (1997) Model inversion of satellite-measured reflectance obtaining surface bio-
physical and bidirectional characteristics of grassland. Remote Sens. Environ. 59:461–471
Garabedian P (1964) Partial differential equations. Wiley, New York, 672pp.
Garrigues S, Allard D, Baret F, Weiss M (2006) Quantifying spatial heterogeneity at the landscape
scale using variogram models. Remote Sens. Environ. 103:81–96
Gobron N, Pinty B, Verstraete MM, Govaerts Y (1997) A semi-discrete model for the scattering of
light by vegetation. J. Geophys. Res.
Gobron N, Pinty B, Verstraete M, Govaerts Y (1999) Level 2 vegetation index algorithm theoretical
Basis. Joint Research Center, Ispra, Italy
Gobron N, Pinty B, Verstraete M, Widlowski JL (2000) Advanced vegetation indices optimized for
upcoming sensors: design, performances and applications. IEEE Trans. Geosci. Remote Sens.
38(6):2489–2505
Goel NS (1989) Inversion of canopy reflectance models for estimation of biophysical parameters
from reflectance data. In: G. Asrar (Editor), Theory and Applications of Optical Remote Sens-
ing. Wiley, New York, pp. 205–251
Goel NS, Deering DW (1985) Evaluation of a canopy reflectance model for LAI estimation through
its inversion. IEEE Trans. Geosci. Remote Sens. GE-23(5):674–684
Goel NS, Thompson RL (1984) Inversion of vegetation canopy reflectance models forestimating
agronomic variables. III. Estimation using only canopy reflectance data as illustrated by the
SUIT model. Remote Sens. Environ. 15:223–236
Goel NS, Strebel DE, Thompson RL (1984) Inversion of vegetation canopy reflectance models for
estimating agronomic variables. II. Use of angle transforms and error analysis as illustrated by
the SUIT model. Remote Sens. Environ. 14:77–111
Gong P, Wang DX, Liang S (1999) Inverting a canopy reflectance model using a neural network.
Int. J. Remote Sens. 20(1):111–122
7 Estimating Canopy Characteristics from Remote Sensing Observations 199

Green RO, Boardman J (2001) Exploration of the relationship between information content and
signal to noise ratio and spatial resolution in AVIRIS spectral data. In: RO Green (ed), AVIRIS
Workshop, Pasadena, CA
Hansen MC, et al. (2002) Towards an operational MODIS continuous firld percent tree coveralgo-
rithm: examples using AVHRR and MODIS data. Remote Sens. Environ. 83:309–319
Hansen MC, DeFries RS (2004) Detecting long-term global forest change using continuous fields
of tree-cover maps from 8-km Advanced Very High Resolution Radiometer (AVHRR) data for
the years 1982–99. Ecosystems 7:695–716
Hapke B (1981) Bidirectional reflectance spectroscopy. 1. Theory. J. Geophys. Res. 86:3039–3054
Huang D, et al. (2006) The importance of measurement error for deriving accurate reference leaf
area index maps and validation of the MODIS LAI products. IEEE Trans. Geosci. Remote Sens.
44(7):1866–1871
Huete AR (1988) A soil adjusted vegetation index (SAVI). Remote Sens. Environ. 25:295–309
Huete AR, Liu HQ, Batchily K, van Leeuwen W (1997) A comparison of vegetation indices over
a global set of TM images for EOS-MODIS. Remote Sens. Environ. 59:440–451
Jacquemoud S (1993) Inversion of the PROSPECT+SAIL canopy reflectance model from AVIRIS
equivalent spectra: theoretical study. Remote Sens. Environ. 44:1–25
Jacquemoud S, Baret F (1990) PROSPECT: a model of leaf optical properties spectra. Remote
Sens. Environ. 34:75–91
Jacquemoud S, Baret F, Hanocq JF (1992) Modeling spectral and directional soil reflectance.
Remote Sens. Environ. 41:123–132
Jacquemoud S, Baret F, Andrieu B, Danson M, Jaggard K (1995) Extraction of vegetation biophys-
ical parameters by inversion of the PROSPECT+SAIL model on sugar beet canopy reflectance
data. Application to TM and AVIRIS sensors. Remote Sens. Environ. 52:163–172
Jacquemoud S, Bacour C, Poilvé H, Frangi JP (2000) Comparison of four radiative transfer mod-
els to simulate plant canopies reflectance: direct and inverse mode. Remote Sens. Environ.
74:471–481
Kimes DS, Nelson RF, Manry MT, Fung AK (1998) Attributes of neural networks for extracting
continuous vegetation variables from optical and radar measurements. Int. J. Remote Sens.
19(14):2639–2663
Kimes DS, Knyazikhin Y, Privette JL, Abuelgasim AA, Gao F (2000) Inversion methods for
physically-based models. Remote Sens. Rev. 18:381–439
Knyazikhin Y, et al. (1998a) Estimation of vegetation canopy leaf area index and fraction of ab-
sorbed photosynthetically active radiation from atmosphere-corrected MISR data. J. Geophys.
Res. 103(D24):32239–32256
Knyazikhin Y, Martonchik JV, Myneni RB, Diner DJ, Running SW (1998b) Synergistic algorithm
for estimating vegetation canopy leaf area index and fraction of absorbed photosynthetically
active radiation from MODIS and MISR data. J. Geophys. Res.103(D24):32257–32275
Knyazikhin Y, et al. (1999) MODIS leaf area index (LAi) and fraction of photosynthetically active
radiation absorbed by vegetation (FPAR) product (MOD15) algorithm theoretical basis docu-
ment. http://eospso.gsfc.nasa.gov/atbd/modistables.html
Kötz B, Baret F, Poilvé H, Hill J (2005) Use of coupled canopy structure dynamic and radia-
tive transfer models to estimate biophysical canopy characteristics. Remote Sens. Environ.
95:115–124
Kuusk A (1991a) Determination of vegetation canopy parameters from optical measurements. Re-
mote Sens. Environ. 37:207–218
Kuusk A (1991b) The inversion of the Nilson-Kuusk canopy reflectance model, a test case. Int.
Geosci. Remote Sens. Symp. (IGARSS’91), Helsinki, Finland, pp 1547–1550
Kuusk A (1995) A fast invertible canopy reflectance model. Remote Sens. Environ. 51:342–350
Lacaze R (2004) Land Surface Level3 Algorithm: The inversion of a vegetation model by a neural
network. short report. http://smsc.cnes.fr/POLDER/SCIEPROD/lsp2algol3.htm
Lacaze R (2005) POLDER-2 Land surface level 3 products. User manual and product validation,
Medias/CNES, Toulouse
200 F. Baret, S. Buis

Lauvernet C, Baret F, Hascoët L, Buis S, Le Dimet FX (2008) Multitemporal-patch ensemble inver-


sion of coupled surface-atmosphere radiative transfer models for land surface characterization.
Remote Sens. Environ. (in press)
Leprieur C, Verstraete MM, Pinty B (1994) Evaluation of the performance of various vegetation
indices to retrieve vegetation cover from AVHRR data. Remote Sens. Rev. 10:265–284
Leshno M, Ya Lin V, Pinkus A, Shocken S (1993) Multilayer feedforward networks with non
polynomial activation function can approximate any function. Neural Networks 6:861–867
Lewis P, et al. (2004) Calibration of and L-system model of winter wheat for remote sensing
modelling and inversion. In: C. Godin (ed), 4th Int. Workshop on Functional-Structural Plant
Models, Montpellier, France, pp 257–261
Liang S (2004) Quantitative Remote Sensing of Land Surfaces. Wiley, Hoboken, NJ
Liu W, et al. (2002) Relating soil surface moisture to reflectance. Remote Sens. Environ. 81:238–246
Lotsch A, Tian Y, Friedl MA, Myneni RB (2003) Land cover mapping in support of LAI and fPAR
retrievals from EOS-MODIS and MISR: classification methods and sensitivities to errors. Int.J.
Remote Sens. 24(10):1997–2016
Lucht W (1998) Expected retrieval accuracies of bidirectional reflectance and albedo from EOS-
MODIS and MISR angular sampling. J. Geophys. Res. 103(D8):8763–8778
Makowski, D, Hillier, J, Wallach, D, Andrieu, B Jeuffroy, MH (2006) Parameter estimation for crop
models. In D. Wallach, D. Makowski and J.W Jones (eds.), Working with dynamic crop models:
evaluation, analysis, parameterization, and applications, pp. 101–149. Elsevier, Amsterdam
Morisette J, et al. (2006) Validation of global moderate resolution LAI Products: a framework
proposed within the CEOS Land Product Validation subgroup. IEEE Trans. Geosci. Remote
Sens. 44(7):1804–1817
Nilson T, Kuusk A (1989) A reflectance model for homogeneous plant canopy and its inversion.
Remote Sens. Environ. 27:157–167
Pinty B, Verstraete MM (1991) Extracting information on surface properties from bidirectional
reflectance measurements. J. Geophys. Res. 96:2865–2874
Pinty B, Verstraete MM, Dickinson RE (1990) A physical model of the bidirectional reflectance of
vegetation canopies. 2. Inversion and validation. J. Geophys. Res. 95(D8):11767–11775
Price J (1994) How unique are spectral signatures? Remote Sens. Environ. 49:181–186
Price JC (1990) On the information content of soil reflectance spectra. Remote Sens. Environ.
33:113–121
Privette JL, Emery WJ, Schimel DS (1996) Inversion of a vegetation reflectance model with NOAA
AVHRR data. Remote Sens. Environ. 58:187–200
Qiu J, Gao W, Lesht BM (1998) Inverting optical reflectance to estimate surface properties of
vegetation canopies. Int. J. Remote Sens. 19(4):641–656
Renders J-M, Flasse SP (1996) Hybrid methods using genetic algorithms for global optimization.
IEEE Trans. Systems, Man, and Cybernetics 26:243–258
Richardson AJ, Wiegand CL, Wanjura DF, Dusek D, Steiner JL (1992) Multisite analyses of
spectral-biophysical data for sorghum. RSE, 41:71:82
Rondeaux G, Steven MD, Baret F (1996) Optimization of soil adjusted vegetation indices. Remote
Sens. Environ. 55:95–107
Roujean JL, Lacaze R (2002) Global mapping of vegetation parameters from POLDER multian-
gular measurements for studies of surface-atmosphere interactions: a pragmatic method and
validtion. J. Geophys. Res. 107(D12):ACL 6 1–14
Saltelli, A, Tarantola, S, Campolongo, F, Ratto, M (2004) Sensitivity Analysis in practice. A guide
to assessing scientific models. Wiley & Sons, Chichester. 217pp.
Schwartz MD, Zimmermann G (2005) A new GLM-based method for mapping tree cover contin-
uous fields using regional MODIS reflectance data. Remote Sens. Environ. 95:428–443
Sellers PJ (1985) Canopy reflectance photosynthesis and transpiration. Int. J. Remote Sens.
3:1335–1372
Shoshany M (1991) The equifinality of bidirectional distribution functions of various microstruc-
tures. Int. J. Remote Sens. 12(11):2267–2281
7 Estimating Canopy Characteristics from Remote Sensing Observations 201

Smith JA (1992) LAI inversion from optical reflectance using neural network trained with multiple
scattering model. In: WR (ed), IGARS’92. IEEE, South Shore Harbour Resort and Conference
Center, Houston, TX, pp 757–759
Smith JA (1993) LAI inversion using backpropagation neural network trained with multiple scat-
tering model. IEEE Trans. Geosci. Remote Sens. 31(5):1102–1106
Soler CFS, Blaise F, de Reffye P (2001) A physiological plant growth simulation engine based on
accurate radiant energy transfer. 4116, INRIA, Montbonnot-Saint-Martin, France
Stenberg P (2006) A note on the G-function for needle leaf canopies. Agric. For. Meteorol. 136
(1–2):76
Tarantola A (2005) Inverse problem theory problem and methods for model parameter estimation.
Soc. Ind. Appl. Math., 358 pp
Tarentola A (1987) Inverse problem theory. Methods for data fitting and model parameter estima-
tion. Elsevier, New York
Teillet PM, Gauthier RP, Staenz K, Fournier RA (1997) BRDF equifinality studies in the context
of forest canopies. In: G Guyot, T Phulpin (eds), Physical Measurements and Signatures in
Remote Sensing, Balkema, Rotterdam, Courchevel, France, pp 163–170
Verhoef W (1984) Light scattering by leaf layers with application to canopy reflectance modeling:
the SAIL model. Remote Sens. Environ. 16:125–141
Verhoef, W (2002) Improved modeling of multiple scattering in leaf canopies: The model SAIL++.
In J. A. Sobrino (ed.), 1st Int. Symp. on Recent Advances in Quantitative Remote Sensing,
pp. 11–20, Universitat de Valencia, Spain
Vermote EF, Tanré D, Deuzé JL, Herman M, Mocrette JJ (1997) Second simulation of the satellite
signal in the solar spectrum, 6S: an overview. IEEE Trans. Geosci. Remote Sens. 35(3):675–
686
Verstraete M, Pinty B (1996) Designing optimal spectral indexes for remote sensing applications.
IEEE Trans. Geosci. Remote Sens. 34(5):1254–1265
Weiss M, Baret F, Myneni R, Pragnère A, Knyazikhin Y (2000) Investigation of a model inversion
technique for the estimation of crop charcteristics from spectral and directional reflectance data.
Agronomie 20:3–22
Weiss M, et al. (2002a) Validation of neural net techniques to estimate canopy biophysical variables
from remote sensing data. Agronomie 22:547–554
Weiss M, et al. (2002b) Evaluation of kernel-driven BRDF models for the normalization of
Alpilles/ReSeDA POLDER data. Agronomie 22:531–536
Wiegand CL, Gerberman AH, Gallo KP, Blad BL, Dusek D (1990) Multisite analyses of spectral-
biophysical data for corn. Remote Sens. Environ. 33:1–16
Wiegand CL, et al. (1992) Multisite analysis of spectral biophysical data for wheat. Remote Sens.
Environ. 42(1):1–22
Zhang Y, Shabanov N, Knyazikhin Y, Myneni RB (2002a) Assessing the information content of
multiangle satellite data for mapping biomes I. Theory. Remote Sens. Environ. 80:435–446
Zhang Y, Tian Y, Myneni RB, Knyazikhin Y, Woodcock CE (2002b) Assessing the information
content of multiangle satellite data for mapping biomes I. Statistical analysis. Remote Sens.
Environ. 80:418–434
Zhang Q, et al. (2005) Estimating light absorption by chlorophyll, leaf and canopy in a decidu-
ous broadleaf forest using MODIS data and a radiative transfer model. Remote Sens. Environ.
99(3):357
Chapter 8
Knowledge Database and Inversion

Jindi Wang and Xiaowen Li

Abstract Physical remote sensing models usually need dozens of parameters to


have reasonable degree of precision. With limited information provided by remotely
sensed observations, it is an ill-posed problem to estimate parameters through model
inversion. Many researches have developed inversion models and algorithms. We
developed a priori knowledge based inversion strategy and algorithm. The spectrum
knowledge database of typical land surface objects has been established to provide
the prior knowledge of model parameters. Some approaches are presented in this
chapter, which include the uncertainty and sensitivity matrix for analysis of obser-
vation data and parameters, the model inversion method supported by the knowledge
database, the scaling correction on estimated parameters. Some study directions in
model inversion, such as how to accumulate and use spatial and temporal change
knowledge, how to validate the parameter inversion results, are also discussed.

8.1 Questions on and Possible Answers to Physical Remote


Sensing Model Inversion

One of the primary problems of remote sensing science is the retrieval of infor-
mation on land surface parameters from remotely sensed data. Since satellite re-
mote sensing deals with a complex system coupling atmosphere and land surface,
physical remote sensing models usually need several to tens parameters to describe
the relation between land surface parameters and remote sensing signals for hav-
ing reasonable degree of model precision. But, with limited information provided
by remotely sensed observations, it remains a challenge to estimate parameters
through remote sensing model inversion. In mathematics, it is an ill-posed problem

Jindi Wang and Xiaowen Li


Research Center for Remote Sensing & GIS
Beijing Normal University, Beijing, China
wangjd@bnu.edu.cn

S. Liang (ed.), Advances in Land Remote Sensing, 203–217. 203


c Springer Science + Business Media B.V., 2008
204 J. Wang, X. Li

to estimate more model parameters from less observations. In past years, many re-
searchers have developed new inversion models and algorithms (Verstraete et al.,
1996; Myneni et al., 1995). Some inversion research has developed better cost
function for physical model inversion. In addition, different kinds of mathematical
optimization algorithms have been developed (Kimes et al., 2000; Kunsk, 1991).
Verstraete et al. (1996) pointed out the main limitation and the necessary observa-
tions for obtaining successful model inversion. Kimes et al. (1991) presented the
knowledge-based expert system for inferring vegetation characteristics. Fang et al.
(2005) used a genetic algorithm to estimate leaf area index (LAI) of vegetation
canopy with radiative transfer model inversion. Neural network methods were also
used for model inversion (Smith, 1993). Liang (2004) summarized the main research
achievements in estimating land surface biophysical variables and surface radiation
budgets.
As advances in the field of multi-angular remote sensing progress, bidirectional
reflectance distribution function (BRDF) models can be inverted to estimate the
structural parameters and spectral component signatures of land surface cover types,
such as the MODIS albedo and LAI products. Some operational algorithms have
been implemented to generate products of land surface parameters, such as albedo,
LAI, fraction of photosynthetically active radiation (FPAR), net primary production
(NPP) from MODIS and MISR observations at the spatial resolution of 250 m or
1 km. The surface BRDF and albedo product from POLDER has been developed
at a pixel resolution of approximately 6 km. Some validation work on these data
products has been carried out and it has been found that there are still some uncer-
tainties on model, observation and reference data which influence the accuracy of
these products (Morisette et al., 2006; Tan et al., 2005).
However, when we want to improve the estimation accuracy of retrieved surface
parameters to meet requirements of applications at different spatial scales, the inver-
sion of physical remote sensing models is a very difficult problem that still requires
further studies from the viewpoint of both information theory (Li et al., 1998) and
the comprehensive practice of model inversion (Privette et al., 1997; Wanner et al.,
1997; Li et al., 2000b; Liang, 2004). The real physical system that couples the at-
mosphere and land surface is extremely complex and it requires many parameters to
describe it faithfully. Any physical model can only be an approximation of this real
system, and a good model will have many important parameters to capture the ma-
jor variations of the real system. However, remotely sensed observations are usually
more or less correlated. The remotely sensed signal, no matter how fine its spectral
and angular resolution, contains only limited information. Therefore, BRDF model
inversion problems, such as those in geoscience generally, are usually underdeter-
mined, making the use of a priori knowledge necessary.
As the ancient philosopher Confucius pointed out, “Our knowledge consists of
two parts – what we know, and what we know we don’t know”. In the case that
remote sensing signals contain limited but valuable information, it is important to
extract information about what we don’t know or what is uncertain, rather than to
invert all model parameters at the same time, pretending that we know nothing.
Using this principle in earlier work, we expressed a priori knowledge of model
8 Knowledge Database and Inversion 205

parameters as the best guesses for the associated uncertainties, and developed the
a priori knowledge based inversion strategy and algorithm (Li et al., 1997). The re-
sults were encouraging, and thus we tried to formalize the approach, and to establish
the a priori knowledge database of typical land surface parameters (Li et al., 2002).

8.1.1 A Priori Knowledge-based Inversion Strategy

We developed a priori knowledge based inversion strategy and algorithm (Li et al.,
1998, 2001) called, Multi-stage, Sample-direction dependent, Target-decision
(MSDT) (Li et al., 1997). There are three questions should be answered in or-
der to run the strategy. The first question is how to express the a priori knowledge
and to make it available in model inversion algorithm. The second question is how
to divide the model parameters set into subsets, in order to invert the parameters by
using the most sensitive data. The third question is how to accumulate the knowl-
edge during the inversion procedure, when we have one scene observing data or a
data set containing more scenes of continuous observations.
In our previous study (Li et al., 1998), the a priori knowledge of model para-
meters is expressed as a joint probability density, while a priori knowledge of the
model accuracy and measurement noise is expressed as a conditional joint proba-
bility density. In inversion model, the a priori probability density function (PDF) of
the observations can be defined, and can be used in the cost function based Bayesian
inference theory. An important feature of Bayesian inversion is that there is no pre-
requisite number of independent observations for a successful inversion. So long as
new observations are acquired, a priori probability density in parameter space can
be modified to obtain posterior density, allowing knowledge to be accumulated.
Taking the land surface spectral albedo inversion as an example, Li et al. (2001)
showed how the a priori knowledge significantly improves the retrieval of surface
bidirectional reflectance and spectral albedo from satellite observations. In the pa-
per, the a priori knowledge are extracted from field measurements of bidirectional
reflectance factors for various surface cover types in red and near-infrared bands.
Bidirectional reflectance and albedo are retrieved by the kernel-driven BRDF model
inversion that uses surface reflectance observations derived from orbiting satellites.
A priori knowledge is applied when noise and poor angular sampling reduce the
accuracy of model inversion. In such cases, a priori knowledge can indicate when
retrieved kernel weights or albedos are outside expected bounds, leading to a closer
examination of the data. If data are noisy, a priori knowledge can be used to smooth
the data. If the data exhibit poor angular sampling, a priori knowledge can be used
according to Bayesian inference theory to yield a posteriori estimates of the un-
known kernel weights, where Bayesian theory is applied in the data space rather
than in the parameter space. Extensive studies and simulations using 73 sets of field
observations and 395 space-borne observation sets from the POLDER instrument
demonstrates the importance of a priori information in improving inversions and
BRDF retrievals.
206 J. Wang, X. Li

The further studies introduce Tarantola’s inversion cost function in to BRDF


model inversion algorithm. Tarantola’s inversion theory has also been widely ap-
plied in geophysical inversion (Tarantola, 1987) and atmospheric remote sensing
(Rodgers, 1976), because the inverse problems in those fields are even more ill-
posed than in land surface remote sensing. Based on Tarantola’s theory, the cost
function used in land surface parameters inversion is defined as the parameters’
posterior probability density, and can be expressed as

1
pM (X) = const · exp − ( f (X) −Yobs )T CD−1 ( f (X) −Yobs )
2

−1
+ (X − Xprior )T CM (X − Xprior ) (8.1)

where Yobs is for observation data and Xprior is for a priori knowledge of parame-
ters. The principle of the cost function is very similar to Bayesian inversion. One
of its advantages is its ability to clearly express the errors of the model, the ob-
served data, and the parameters’ initial value, with co-variances matrixes CD and
CM respectively. This creates a new challenge on forming the error co-variances of
model, data and parameters, this requires a great deal of a priori knowledge.
Previous studies have demonstrated the effectiveness of the inversion model, us-
ing simulated data and field measurements (Yan et al., 2001; Wang et al., 2000). In
these studies, the parameters’ initial values can be estimated by our field measure-
ments, while the co-variance was set as the parameters’ standard deviation under the
assumption that their estimated values have a normal distribution.

8.1.2 Uncertainty and Sensitivity Matrix of Parameters for Model


Inversion

It is important to examine the uncertainty and sensitivities of the parameters in


model inversion, since we expect to estimate the values of parameters that are sen-
sitive to observations and model, while giving certain values for other parameters
that have relatively fewer uncertainties. To make the analysis of parameter uncer-
tainty and sensitivity more effective, we defined the uncertainty and sensitivity ma-
trix (USM), which is an objective expression of the prior knowledge.
In order to construct the USM, we assume that a BRDF model has N spectral
bands and K structural parameters and L spectral parameters of component materi-
als, making K + N × L in total. Since the multiangular observations have M samples,
the USM will have M × N rows and K + N × L columns. This matrix is too large to
effectively use for calculations. Because the structural parameters are independent
on the spectral band, we decomposed the matrix into a structural matrix that has
M × N rows, K columns, where the N matrix of M × L spectral parameters corre-
sponds to N bands, making a total N + 1 matrices. An element of the USM can be
expressed as (Li et al., 1997),
8 Knowledge Database and Inversion 207

BRDF( j|i)
USM[i][ j] = (8.2)
BRDF(i)

where BRDF( j|i) is the maximum difference of BRDF as a function of the jth pa-
rameter within its uncertainty, given the ith geometry of illumination and viewing;
BRDF(i) is BRDF as predicted by the model at the ith geometry, when all para-
meters at their best guess values. These best guess values of parameters are also
from the accumulation of prior knowledge, and they can be updated when further
parameter estimates are obtained during the MSDT inversion procedure.
This USM definition has three advantages: (1) the uncertainty of the initial guess
for being inverted parameter is taken into account; (2) the USM is less dependent on
the initial guess; (3) all elements of the USM have the same units, and are therefore
quantitatively comparable. This USM has thus been used widely in our recent model
inversion studies. It can also be applied for parameters sensitivity analysis for models
and data in relative studies (Li et al., 1997; Gao and Zhu, 1997; Yan et al., 2001).

8.1.3 Getting the Prior Knowledge on Typical Land Surface


Parameters

As stated above, after developing the model inversion strategy, we recognized that
the most important issue in remote sensing model inversion is accumulation of prior
knowledge regarding the model parameters. Given that we are not only concerned
with the validation of the model, but also with how to apply the inversion algorithm
for land surface parameter estimation at the scale of remote sensing images, we have
to consider establishing a very powerful prior knowledge database that assembles
the initial estimates of the parameters and also their co-variance values during the
inversion procedure. These initial parameter values and their co-variance matrix are
also necessary for supporting data assimilation algorithms.
In practice, we firstly classify the a priori knowledge into different levels: the
general knowledge about the land surface, or “global knowledge”, knowledge re-
lated to land cover type, and target-specific knowledge. The means to accumulate
knowledge at these levels may be different, but should include the following:
1. Applicable forward model(s)
2. Physical limits and probability density in model parameter space
3. Statistics of model accuracy and noise in remote sensing signals
4. Seasonal change associated with land cover types or targets
5. Confidence of the above knowledge
Note that even a single observation can change the a priori PDF of more than one
parameter significantly. Li et al. (1998) provided an example on how the accumula-
tion of knowledge is achieved in parameter space. However, the required numerical
integration in parameter space is time-consuming whenever parameters set is large,
as is the case with the retrieval of surface BRDF and albedo from satellite data.
208 J. Wang, X. Li

Our recent approaches are to establish an a priori knowledge database of the


parameters of typical land surface targets, including the spatial and temporal distri-
bution knowledge of parameters. This knowledge database also includes the forward
models, remote sensing data and a priori knowledge at both global and land cover
type related levels. Based on the knowledge base, it is expected that the accuracy of
land surface parameters estimation at the scale of remote sensing observations will
improve (Wang et al., 2003).

8.2 The Spectrum Knowledge Database of Typical Land Surface


Objects

Establishing the spectrum knowledge database of typical land surface objects is sig-
nificant for the development of quantitative remote sensing. The research on spec-
tral features of objects is the foundation of remote sensing applications. Land cover
and land use classification and image interpretation are usually based on the spec-
trum recognizing method. Many kinds of spectrum matching techniques have been
developed. However, users cannot always obtain the desired accuracy for classi-
fication and identification. One of the problems is that the spectral data of objects
measured at different scales are not comparable. It is rare to have the correct relation
between the spectrum of objects measured indoors or in the field, and the spectral
image data acquired through remote sensing observations. The available spectrum
data measured indoors and in the field do not have enough corresponding descrip-
tors to clearly determine the related spectral environmental variables. This can result
in some confusion in spectrum applications using spectra matching and other image
processing algorithms.
The second problem is that we need to integrate the physical models and mea-
surement data into a close linked system to allow models and data to be effec-
tively applied for land surface parameters estimation. When the linked system is
developed, the prior knowledge of the model parameters can be extracted from the
database and the model prediction and parameters inversion algorithm can also be
included in the system. That is our approach to convert the individual measurement
data sets to be the knowledge database with the stated goal of having effective phys-
ical model explanations and predictions.
The establishment of the spectrum knowledge database of typical land surface
objects is supported by China’s National High Technology Research and Develop-
ment Program. The spectrum knowledge database consists of the measuring spectral
data set and related environmental variables of objects at different observing scales,
the physical models set and a priori knowledge set which provide the main geo-
graphical background data for models.
8 Knowledge Database and Inversion 209

8.2.1 The Definition of Spectrum at Three Measured Scales

To avoid confusion in spectral matching for land cover type recognition, we clearly
defined and measured three kinds of spectrum for different observed objects, while
considering the application requirements. The three spectrums are the spectrums of
materials, endmembers and remote sensed pixels.
The spectrum of materials is usually measured under well controlled measure-
ment conditions in the laboratory. The measured objects can be crop leaves, mineral
samples, and water samples. They are essential spectrum data of the spectrum data-
base, and are usually taken as given variables in physical models.
The so-called spectrum of endmembers is the spectrum of components of re-
motely sensed pixel, it is the basic parameter in general remote sensing models. The
spectrum of endmembers is usually measured in the field, where the surface of the
measured object is relatively uniform, and where the measuring FOV of sensor is
less than pixel size. For example, in the scene integrated geometric optical BRDF
model on canopy reflectance, the sunlit crown, sunlit background, shaded crown and
shaded background are endmembers of the modeled forest pixel. Their spectrums
are usually the main parameters of scene integrated models, which may differ from
the spectrum of leaves and soils.
The spectrum of remotely sensed pixels comes from remotely sensed observa-
tions, and is usually the main concern of users. Note that when we consider different
modeling scales, the measured endmember spectrum and the spectrum of remotely
sensed pixels should be used according to the scale of the application (Li et al.,
2002; Wang et al., 2003).

8.2.2 Typical Land Surface Objects Spectrum Database

In the spectral data set, for every typical object the measured data includes not only
its surface spectra, but also the environmental variables of each observed object.
The variables are all physical model parameters, which can be used to predict the
spectra of a given surface object by a physical model. In the first step of the database
establishment, the main typical land surface objects include three land cover types:
the main crops growing in China, which are winter wheat, rice, maize, cotton and
cole; rocks and minerals; and bodies of water.
The collected data consist of two parts. One part is the individual spectrum data
sets measured during the last 20 years. In order to keep this data available, we com-
piled the data following our own data collection standards. For instance, some of
the data were measured in the laboratory and in standard well controlled measuring
conditions, such as for minerals.
The other part consists of new measurements. We first determined the environ-
mental variables to be measured, which are dependent on the mature physical mod-
els, and widely required parameters for remote sensing applications. We established
the technical criteria for remote sensing experiment instruments, laboratories and
210 J. Wang, X. Li

field sites, and also the technical criteria and requirements for the measurement of
object’s spectrum and environmental variables. All new data measurements follow
these criteria and regulations, ensuring that the measurements of surface spectra and
environmental variables are made together. For those objects which surface features
may change with time, such as crop vegetation canopy, we made measurements dur-
ing each of their main growth stages across their entire growing season. These data
sets with temporal changes can be used to extract the temporal change information
of the canopy surface reflectance and canopy structural parameters, which is the
most important prior knowledge for canopy reflectance model inversion and data
assimilation.

8.2.3 From Spectrum Database to Knowledge Base

To make the spectrum database applicable for quantitative remote sensing, we fur-
ther describe how to develop the spectrum database to the spectrum knowledge base.
We do this by integrating the spectrum database, the remote sensing image base, the
remote sensing model base, and the geographic background database to create the
spectrum knowledge database. The remote sensing model base is the key part of
the knowledge base. The models describe the relationship between the spectrum
reflectance measurements and related environmental variables. The spectra of ma-
terials, of endmembers, of remote sensing pixels and the measuring environment
variables can be linked by models when the scale of observation is known. The geo-
graphic background knowledge data sets provide the prior knowledge of the model
parameters for running the models.
A significant function of the knowledge base is to make surface spectra simu-
lation or prediction. Because there are too many types of land surface objects, the
amount of objects we can measure is always limited. The potential change of the
objects are infinite, we can not include all the cases for even one given object. For
example, for winter wheat, we almost cannot obtain all the spectrum for its entire
growing seasons from the different regions where it is planted. Considering the ap-
plication of spectrum database, users may occasionally require the spectra of winter
wheat for a given date from a specific place, which may not be stored in the database.
The simulation or prediction of spectrum is achieved by using physical-based
remote sensing models, and applicable empirical models on objects, such as crop
simulation models, as well as the knowledge base. Therefore, our knowledge base
developed the ability to simulate spectra. When the spectra of an object requested
by the user is not in the database, the remote sensing models can be used to simulate
the spectra, based on measured spectrum and environmental variables of similar
land surface objects saved in the knowledge base. For example, our database stored
some standard spectrum of winter wheat at its different growing stages, while the
user requires the spectrum at a specific stage in order to predict its growing state
from his remotely sensed image. The spectrum from existing spectrum and from
related structural parameters of several growing seasons will be calculated with a
8 Knowledge Database and Inversion 211

spectral prediction modular. The crop simulation model will provide information
about the growing tend, the geographic knowledge base will provide the information
about phenological and regional characteristics. The spectral database will provide
the spectrum of material and components, and also the structural parameters of a
typical related winter wheat. In this way, the modular can interpolate and extend the
spectrum in temporal and spatial scales, and also predict the spectrum at a given field
of view of sensor, a given sunlight, and under the specific atmospheric conditions.
The simulated spectra will provide the user with a reference. Spectra data extension
is an important feature of our knowledge base. It can also be available as a research
platform for studies on object surface spectra prediction and model validation.
At present, in our knowledge base, the model base includes general models, phys-
ical models, and some special application models. The geographic background data
includes the following: the DEM of 1:100,000–1:250,000 of the demonstration re-
gion, and land use map of the demonstration region, the base data of 1:4,000,000
covering the whole of China (DEM, physiognomy, vegetation, soil, geosciences,
rivers and lakes), phenological phase of typical crops, Chinese geological map
(1:5,000,000), such as Nonmetals Metals Mineral Resource Map, Mineral Resource
Map and China water resource map (1:4,000,000).
From the measured database, we can construct the prior knowledge of land sur-
face parameters by the statistical data analysis. The a priori knowledge of parame-
ters can be expressed as their mean and variance at spatial and temporal scales of
the accumulated data. The establishment of the spectrum knowledge base allows
knowledge based inversion strategy and algorithm to be used to process remotely
sensed image (Wang and Li, 2004).

8.2.4 The Internet-based Spectrum Knowledge Database Service


System

The internet-based spectrum knowledge database service system includes the soft-
ware system and its corresponding hardware environment. This system consists of
several modules, such as spectrum querying, model calculation, knowledge query-
ing, image management, application demonstrations, and system management. The
spectrum and models in our knowledge database were all imported according to the
data quality level regulations.
The system also has several demonstrations of typical field remote sensing appli-
cations, which include the following: production estimation and growth monitoring
of crops in North China; precise agriculture demonstration of winter wheat, demon-
strations of cotton growth monitoring; rock and mineral mapping using airborne
hyperspectral remotely sensed data; water quality evaluation of the Huangpu River
in Shanghai with airborne hyperspectral remotely sensed data. By employing the ex-
planations in the demonstration system, users can learn how to apply the spectrum
and other data to extract required land surface parameters from remotely sensed
images.
212 J. Wang, X. Li

8.3 Land Surface Parameters Inversion Supported


by the Knowledge Database

Based on the spectrum database, we propose our methods to estimate land surface
parameters, which make effective use of the spectrum knowledge. There are two
kinds of questions that need to be considered. One is how to abstract the spatial and
temporal distribution of the prior knowledge of parameters from database. When we
estimate vegetation parameters from remotely sensed data, the data we used can be
from a scene covering large region, so we need to have the spatial distribution prior
knowledge of parameters around the same region to support the prior knowledge-
based inversion algorithm. When we want to understand the temporal changes of
crop growth parameters during its growing seasons, we need to have the knowledge
with temporal distribution of estimated parameters, such as when we perform data
assimilation using a crop growth model.
The second kind of questions we should consider is how to introduce the infor-
mation on the spatial and temporal distribution of parameters into the inversion algo-
rithm. The model we use for the inversion should be not only suitable for processing
one special scene remotely sensed data, but also for the data that changes with time.
A dynamic model should then be developed, and a data assimilation algorithm also
should be developed. The new method should be able to perform time sequential
matching for better estimation of parameters at given time and date.

8.3.1 Extract Spatial and Temporal Distribution of Land Surface


Parameters

From the spectrum database, we can extract the spatial and temporal distributed a
priori knowledge of land surface parameters by means of statistical data analysis, as
well as model predictions. To do this, we should first understand the kinds of spatial
and temporal data that are available, and then how to make use of them.
The prior knowledge data of parameters, which are applied in inversion and data
assimilation algorithms, can be at several spatial and temporal scales.
The prior knowledge data at different spatial scales can be categorized in two
types. One type of data relate to the observed objects, but at different resolutions at
which the data is acquired. The spatial resolution of acquired data could be at tens
of centimeters, at tens of meters, or even at several kilometers. Another type of data
relate to the spatial distribution variability from the same or different types of land
surface objects. The spectrum database provides knowledge data by mainly focus-
ing on the former kinds of data. For the latter, some geographic background data
with several certain scales can be available. Looking at the example of inversion of
vegetation canopy reflectance model, the function of prior knowledge provided by
spectrum database can be the following: first, spectrum database can provide the
spectrum of materials or endmembers, such as those of leaves and soil from the
8 Knowledge Database and Inversion 213

background of crop canopy, which can be used to fix some parameter values, if a
model has too many parameters and there are only limited observations to be avail-
able for estimate the all model parameters. Secondly, the spectrum database can
provide some statistical information of parameters, such as the parameters’ mean
value and standard deviation, which can be used as initial guesses and the uncer-
tainty index of parameters to be retrieved.
For prior knowledge data at temporal scale, the spectrum database can provide
some referencable parameters for dynamic models, while retrieving the variable in-
formation of land surface parameters. Looking again at the example of crop canopy
reflectance model inversion, due to crop growth over time, the observed spectrum
will change. The dynamic change spectral observations can provide an opportu-
nity to retrieve biophysical and biochemical parameters of the crop surface. Thus,
the dynamic spectral data of continuing growing vegetation can be used to extract
some specific knowledge from the temporal sequence. This type of knowledge may
be used to compensate for the limited information from remotely sensed data. In
addition, the knowledge is useful for data assimilation. The continuous temporal
knowledge of the parameters can be applied to compare with the dynamic model
prediction, and then to provide reference feedback to adjust the model’s parameter
estimation for the data assimilation procedure. This should result in a better estima-
tion for parameters when their values change with time and with natural features,
such as the leaf area index in the crop growth model.

8.3.2 Inversion Model Based on Bayesian Network to Integrate


Physical Model and Prior Knowledge

Recent research has shown that combining the empirical formula method and the
physical model inversion into a new hybrid inversion scheme for estimating sur-
face parameters will be a promising trend (Fang and Liang, 2005; Liang, 2004).
Following previous work on hybrid inversion, we build a new hybrid inversion
scheme which uses a Bayesian Network (BN) to determine the mapping relation-
ship between simulated reflectances and their corresponding biophysical parameters
(Qu et al., 2005). As a hierarchical probability model, BN can not only be used
as a non-parameters regression model, but can also be used to deduce information
from multi-layered parameters (Marcot et al., 2001). This differs from other non-
parameters regression methods, such as Neural Network. In our approach, we focus
on the incorporation of prior knowledge derived from spectrum database and phys-
ical model into a unified framework. A simple Bayesian Network is illustrated as
Fig. 8.1.
Using the Bayesian Network to retrieve parameters, the posterior probability den-
sity distribution of A can be calculated using the observed data and their ancillary
parameters, the following Eq. (8.3) can be derived using Bayesian theorems.
214 J. Wang, X. Li

Fig. 8.1 A simple Bayesian


Network C Ancillary

A Parameter space

B Data

p(A|B = bi ,C = ck ) ∝ p(C = ck )p(A|C = ck )p(B = bi |A) (8.3)

Where p(A|C = ck ) presents the prior knowledge of interested parameters, p(C = ck )


represent the probability distribution of other variables, such as time, location, ele-
vation, land use, and other assistant information, which can affect the ancillary
parameters, p(A|C = ck ) is the probability density distribution of the parameters to
be derived after obtaining the above information. The two probability distributions
can be obtained from the spectrum database, and the quantitative influence between
them, i.e., p(A|C) can be obtained statistically by using data simulated by physical
models. By extending the Bayesian theorem into the Bayesian Network, which
uses a multifactor deducing method, the prior knowledge about the land surface
parameters can be extracted from the spectrum database, and then can be combined
with the physical model to retrieve the information on the desired parameters.
By integrating the observed data and new information into a unified framework
to infer knowledge, this new hybrid inversion scheme has shown that it can incorpo-
rate more information besides model parameters into the process of remote sensing
model inversion to retrieve biophysical and biochemical parameters. The process
of extracting knowledge from the historical database, and using it in the retrieving
information from remote sensing models and remotely sensed data is the summing-
up of prior knowledge and new information. In this process, the information from
newly obtained data can be updated and added to the existing knowledge on para-
meters.
In our preliminary research, our proposed inversion approach was used and val-
idated through retrieving the biophysical and biochemical parameters of winter
wheat. We abstracted the prior knowledge of canopy reflectance model parameters
from the spectrum database to obtain crop growth parameter distribution at different
growth stages. We developed the inversion model based on the Bayesian Network to
integrate the physical model and prior knowledge to estimated LAI and chlorophyll
(a + b) with simulated BRDF spectral data, and to estimated LAI with ETM image
(Qu et al., 2005).
8 Knowledge Database and Inversion 215

8.4 Parameters Scaling and Validation

Scaling effects are the basic issues in remote sensing research. We studied the scal-
ing effect in thermal infrared remote sensing models and the reciprocity principle
when these basic physical principles are applied in processing special problems in
remote sensing (Li et al., 2000a). When we estimate model parameters by remote
sensing model inversion, one important question is whether the parameters require a
uniform estimation value when observations are at different spatial scales. In our ex-
periment of model inversion, one parameter may have different estimated values if
the available remote sensing data are at different scales of observation. In such case,
it is still a problem on how to improve the precision of model parameters estimation.
Taking the estimation of crop planting area as example, we usually obtain different
estimated area value from AVHRR data or from TM data. And the difference may
not follow a certain formula when land cover types from the observed regions are
different. This is mainly due to a greater amount of mixed pixels in AVHRR data,
and the heterogeneity of land surface. It is very similar to the fractal problems, such
as the coastline measurement using the data from different spatial resolution. In
order to describe the difference resulting from different scales of observation, we
proposed the concept of histo-variogram, defined as the “total fractal dimension”,
and applied it to derive a formula for scaling correction on land surface parameter
estimates (Zhang et al., 2003).
We use LULC data as an example to study land use area estimation method
using down-scaling. For the crop area estimates, we defined the initial area S0 which
means the observed area at initial measured scale, stated its relation to the total
fractal dimensions (d) and coefficients ( f (S0 )), and then applied it to up-scaling
and down-scaling in crop area estimation. The estimated crop area of standard pixel
(Sδ ) at different scale (δ) is expressed as

Sδ = f (s0 )δ d−2 (8.4)

where ( f (S0 )) is the function of the initial area (S0 ) and with the same unit as S0 .
δ is for measured scale, can be expressed with fractal of standard pixel.
The relation between fractal (δ), coefficient (I) and standardized area (Sδ 0 ) is:
   
(Iδ i − Sδ 0 )/Iδ i = 1.62Dδ i + 0.03 δ0 δi − δ0 (8.5)

where I is the f (S0 ) expressed in Eq. (8.4), can be calculated by Eq. (8.4) with the
initial measured scale δi in the process of scaling-up, δ0 is for finer measured scale
than δi , Sδ 0 is the crop area we want to estimate by down-scaling method at the
measured scale δ0 .
We can obtain the estimated area when the scale of the data is changing based on
the down-scaling method. In this way, using LULC data, we obtain the result of crop
plant area estimation result by the down-scaling method, where the relative error is
less than 5% when the estimated pixel size is 16 times that of the original data.
216 J. Wang, X. Li

8.5 Summary and Discussion

The previous research we described are our main ideas on remote sensing model
inversion strategy and algorithm, base on the analysis of the special ill-posed in-
version problem in land surface parameters inversion. The available land surface
spectrum knowledge database provides useful models and a priori knowledge of
model parameters. We are presently researching the following:
1. How to use spatial and temporal change knowledge in remote sensing model
inversion? Based on our previous research, we introduce a dynamic model to ob-
tain a time consequence of model parameters. The idea on data assimilation will
then be applied to improve the precision and reliability of parameters estimation.
2. How to validate and evaluate the parameters’ inversion result when the raw data
are at different scales? This is still a major problem filled with many uncertain-
ties. The problem regards the scaling effect of models, parameters and observa-
tions.
3. How to meet users’ requirements for land surface parameters estimated by re-
mote sensing data, so as to provide input for common land model (CLM)? In this
case, the a priori knowledge of the parameters’ spatial distribution and temporal
signature will be obtained from the knowledge base, and then used to invert these
dynamic changing parameters by introducing the main idea of data assimilation.

References

Fang H, Liang S (2005) A hybrid inversion method for mapping leaf area index from MODIS
data: experiments and application to broadleaf and needleleaf canopies. Remote Sens. Environ.
94(3):405–424
Gao F, Zhu Q (1997) Process system of measured bidirectional reflectance in Changchun labora-
tory. J. Remote Sens. 1(Suppl):123–130
Kimes DS, Harrison PR, Ratcliffe PA (1991) A knowledge-based expert system for inferring vege-
tation characteristics. Int. J. Remote Sens. 12:1987–2020
Kimes DS, Knyazikhin Y, Privette J, Abuelgasim A, Gao F (2000) Inversion methods for
physically-based models. Remote Sens. Rev. 18:381–439
Kuusk A (1991) Determination of vegetation canopy parameters from optical measurements.
Remote Sens. Environ. 47:194–202
Li X, Yan G, Liu Y, Wang J, Zhu C (1997) Uncertainty and sensitivity matrix of parameters in
inversion of physical BRDF model. J. Remote Sens. 1(Suppl):113–122
Li X, Wang J, Hu B, Strahler AH (1998) On utilization of prior knowledge in inversion of remote
sensing models. Sci. China (Series D) 41(6):580–586
Li X, Wang J, Strahler AH (2000a) Scale effects and scaling-up by geometric-optical model. Sci.
China (Series E) 43(Suppl):17–22
Li X, Gao F, Wang J, Strahler AH (2000b) Estimation of the parameter error propagation in inver-
sion based BRDF observations at single sun position. Sci. China (Series E) 43(Suppl):9–16
Li X, Gao F, Wang J, Strahler AH (2001) A priori knowledge accumulation and its application to
linear BRDF model inversion. J. Geophys. Res. 106(D11):11925–11935
8 Knowledge Database and Inversion 217

Li X, Wang J, Gao F, Strahler A, Su L (2002) Accumulation of spectral BRDF knowledge for quan-
titative remote sensing of land surfaces. Proc. The First Int. Symposium on Recent Advances
in Quantitative Remote Sensing, Torrent, Valencia, Spain, 16 September 2002
Liang S (2004) Quantitative Remote Sensing of Land Surfaces. Wiley, New York
Marcot BG, Holthausen RS, Raphael MG, Rowland M, Wisdom M (2001) Using Bayesian belief
networks to evaluate fish and wildlife population viability under land management alternatives
from an environmental impact statement. Forest Ecol. Manage. 153(1–3):29–42
Morisette JT, Baret F, Privette JL, Myneni RB, et al. (2006) Validation of global moderate-
resolution LAI products: a framework proposed within the CEOS land product validation sub-
group. IEEE Trans. Geosci. Remote Sens. 44(2):1804–1817
Myneni RB, Maggion S, Iaquinta J, et al. (1995) Optical remote sensing of vegetation: modeling,
caveats and algorithms. Remote Sens. Environ. 51:169–188
Privette JL, Eck TF, Deering DW (1997) Estimating spectral albedo and nadir reflectance
through inversion of simple BRDF models with AVHRR/MODIS-like data. J. Geophys. Res.
102(D24):29529–29542
Qu Y, Wang J, Song J, Lin H (2005) A hybrid inversion scheme to estimate biophysical parameters
of winter wheat: simulation and validation. Proc. ISPMSRS’05, Beijing,, 17–19 October 2005,
pp 743–746
Rodgers CD (1976) Retrieval of atmospheric temperature and composition from remote measure-
ment of thermal radiation. Rev. Geophys. Space Phys. 14(4):609–624
Smith JA (1993) LAI inversion using a back propagation neural network trained with a multiple
scattering model. IEEE Trans. Geosci. Remote Sens. 31:1102–1106
Tan B, Hu J, Huang D, et al. (2005) Assessment of the broadleaf crop leaf area index product from
the Terra MODIS instrument. Agric. Forest Meteorol. 135:124–134
Tarantola A (1987) Inverse Problem Theory – Methods for Data Fitting and Model Parameter
Estimation. Elsevier, Amsterdam, The Netherlands, 613pp
Verstraete MM, Pinty B, Myneni RB (1996) Potential and limitation of information extraction on
the biosphere from satellite remote sensing. Remote Sens. Environ. 52:201–214
Wanner W, Strahler AH, Hu B, Lewis P, Muller J-P, Li X, Barker Schaaf CL, Barnsley MJ (1997)
Global retrieval of bidirectional reflectance and albedo over land from EOS MODIS and MISR
data: theory and algorithm. J. Geophys. Res. 102(D24):17143–17162
Wang J, Li X (2004) The spectrum knowledge base of typical objects and remote sensing inversion
of land surface parameters. J. Remote Sens. 8(Suppl):4–7 (in Chinese)
Wang J, Li X, Sun X, Liu Q (2000) Component temperatures inversion for remote sensing pixel
based on directional thermal radiation model. Science in China (Series E) 43(Suppl):41–47
Wang J, Zhang L, Zhang H, Li X, Liu S (2003) Current progress of constructing the spectrum
knowledge base of typical objects of land surfaces, ESA SPECTRA workshop in France (Italy),
November 2003
Yan G, Wang J, Li X (2001) Making use of a priori knowledge of vegetation spectrum in inversion
for canopy structure parameters. Proc. IGARSS’01, Sydney, Australia, 9–13 July 2001
Zhang H, Jiao Z, Yang H, Li X, et al. (2003) Research on scale effect of histogram. Sci. China
(Series D) 45(10):1–12
Chapter 9
Retrieval of Surface Albedo from Satellite
Sensors

Crystal Schaaf, John Martonchik, Bernard Pinty, Yves Govaerts, Feng Gao,
Alessio Lattanzio, Jicheng Liu, Alan Strahler, and Malcolm Taberner

Abstract Observations from a number of polar-orbiting and geostationary satellite


sensors are now being used to produce operational land surface albedo products
for range of modeling applications. The MODIS, MISR and Meteosat algorithms
are presented as examples of the current strategies being employed to best exploit
multi-day sequential, multi-angular instantaneous, and multi-temporal observations
and accurately specify the reflective qualities of the underlying surface. While these
retrievals represent a major advance in the remote sensing of the spatial and temporal
heterogeneity of the surface, issues such as atmospheric correction, directional-to-
hemispherical conversion, and spectral interpolation remain to confound the satel-
lite signal and introduce uncertainties and variability within and between products.
Nevertheless, the potential of using multiple products and fusing recent observations
with remotely sensed historical data must be explored as a realistic way to meet the
needs of the modeling community.

Keywords: MODIS · MISR · Meteosat · albedo · reflectance · anisotropy

Crystal Schaaf, Jicheng Liu and Alan Strahler


Department of Geography and Environment, Boston University, Boston, USA
schaaf@bu.edu
Feng Gao
Earth Resources Technology, Jessup, USA
John Martonchik
Jet Propulsion Laboratory, Pasadena, USA
Bernard Pinty and Malcolm Taberner
Global Environment Monitoring Unit, IES, EC Joint Research Centre, Ispra (VA), Italy
Yves Govaerts
EUMETSAT, Darmstadt, Germany
Alessio Lattanzio
Makalumedia GMBH, Darmstadt, Germany

S. Liang (ed.), Advances in Land Remote Sensing, 219–243. 219


c Springer Science + Business Media B.V., 2008
220 C. Schaaf et al.

9.1 Introduction

The availability of a large number of directional observations sampling the viewing


hemisphere over a particular land surface can effectively capture its surface
anisotropy and thus be used to accurately compute the surface albedo of that surface.
While numerous samples may be possible in the field or laboratory, remotely sensed
retrieval methods based on data from individual satellites usually must suffice with a
limited number of directional reflectances of the surface, and the producers of such
data sets must acknowledge that these observations may not necessarily represent
a well-distributed sampling (Privette et al., 1997). Therefore a model is usually
adopted to characterize the surface anisotropy – a model which can be inverted with
a finite set of angular samples and then be used to predict surface reflectance in
any sun-view geometry and derive surface albedo (Roujean et al., 1992; Walthall
et al., 1985; Rahman et al., 1993; Engelsen et al., 1996; Wanner et al., 1997;
Pinty et al., 2000a; Bréon et al., 2002; Maignan et al., 2004).
The acquisition of directional measurements from an individual sensor is de-
termined by its scanning configuration and the platform’s orbital characteristics
(Barnsley et al., 1994). However, cloud obscuration always reduces the number of
clear-sky observations possible. Therefore, in the case of a single field of view sen-
sor such as the MODerate Resolution Imaging Spectroradiaometer (MODIS), on
board the polar orbiting Terra and Aqua platforms, an adequate directional sam-
pling of surface reflectances can only be obtained by the accumulation of sequen-
tial observations over a specified time period. Multi-angular instruments such as the
Multiangle Imaging SpectroRadiometer (MISR) instrument (also on board the Terra
platform) obtain sufficient simultaneous directional observations to specify the sur-
face anisotropy whenever a cloud-free acquisition is possible. Geostationary sensors
(such as Meteosat) must trade numerous acquisitions under different illumination
conditions during a day for directional observations to obtain the angular informa-
tion necessary to sample the surface’s directional characteristics. Since 2000, all of
these approaches have been implemented operationally to produce robust surface
albedo fields for use in climate, hydrological, biogeochemical, and weather predic-
tion models.

9.2 Background

As a key land physical parameter controlling the surface radiation energy bud-
get (Dickinson, 1983, 1995), global surface albedo with an absolute accuracy of
0.02–0.05 is required by climate models at a range of spatial and temporal scales
(Henderson-Sellers and Wilson, 1983). Land cover-based schemes have histori-
cally been adopted in most of the land surface models and climate models for the
parameterization and specification of surface albedo (Bonan et al., 2002; Sellers
et al., 1996). Natural landscapes, however, are a collection of nested objects in a
hierarchy and various processes control the biophysical characteristics at differ-
ent spatial scales (Woodcock and Harward, 1992; Collins and Woodcock, 2000).
9 Retrieval of Surface Albedo from Satellite Sensors 221

Therefore, land surface models usually allow a sub-grid specification of land cover
proportions to account for the heterogeneity of surface properties within a grid
(Dickinson et al., 1995; Bonan et al., 2002), while the climate models are generally
implemented at coarser spatial resolutions. However, the increasing spatial resolu-
tion of modern climate models makes it necessary to examine the spatial features
of global surface albedo and the effect of spatial scales on the albedo specifica-
tion. Therefore, a consistent and accurate global albedo data set is essential to the
investigation of the sensitivity of climate to various types of forcing and to the
identification of the effects of human activities. Satellite remote sensing represents
the only efficient way to compile such consistent global albedo characterizations.
Historically, global albedo data sets have been derived from the Advanced Very
High Resolution Radiometer (AVHRR) (Csiszar and Gutman, 1999) and the Earth
Radiation Budget Experiment (ERBE) radiometer (Li and Garand, 1994). With
the advent of routine albedo products derived from MODIS (Gao et al., 2005;
Schaaf et al., 2002; Lucht et al., 2000), MISR (Martonchik, 1997; Martonchik
et al., 1998b), CERES (Clouds and the Earth’s Radiant Energy System), POLDER
(Polarization and Directionality of the Earth’s Reflectances) which is currently on
board PARASOL (Polarization & Anisotropy of Reflectances for Atmospheric Sci-
ences) coupled with Observations from a Lidar (PARASOL) (Leroy et al., 1997;
Hautecoeur and Leroy, 1998; Bicheron and Leroy, 2000; Maignan et al., 2004),
and Meteosat (Pinty et al., 2000a,b), albedo data sets with spatial resolutions of
500 m to 20 km and temporal frequencies of daily to monthly are now available. Al-
though the retrieval of albedo from these instruments represents a major advance in
sensing the spatial and temporal surface heterogeneity, issues such as atmospheric
correction, directional-to-hemispherical conversion, and spectral interpolation can
still confound the satellite signal and introduce uncertainties. Most of these satel-
lite products rely on sophisticated radiative transfer methods (Vermote et al., 1997;
Berk et al., 1998; Liang et al., 1999; Liang, 2000) and bidirectional modeling
(Roujean et al., 1992; Walthall et al., 1985; Rahman et al., 1993; Engelsen et al.,
1996; Wanner et al., 1995; Wanner et al., 1997; Martonchik et al., 1998b; Pinty
et al., 2000a, b) to obtain accurate surface quantities.
The modeling community has enthusiastically begun to utilize these global and
regional satellite albedo products as they have become available (Oleson et al., 2003;
Zhou et al., 2003; Tian et al., 2004; Roesch et al., 2004; Knorr et al., 2001; Myhre
et al., 2005a, b). With 5 or more years of data now available, interannual variations
can be explored and short-term climatologies prepared which compensate for tran-
sient cloudiness or snowcover (Moody et al., 2005; Gao et al., 2005; Barlage et al.,
2005). However, there remains the need to generate analogous surface albedo prod-
ucts prior to year 2000 and in particular over the last 25 years or so where Earth
observing systems from space (e.g., the series of weather satellites) have been ac-
quiring relevant data. Unfortunately, the design of the large majority of these global
observation systems for environmental applications has been driven solely by de-
mands in the domain of meteorology and weather forecasting and, as a consequence,
these sensors do not fulfill some basic requirements for quantitative remote sensing
applications over land, such as those related to the accurate sensor characterization,
geolocation, and calibration.
222 C. Schaaf et al.

However, despite a number of technological limitations, these historical weather


sensors constitute the only possible solution left to remote sensing scientists to as-
sess such quantities as surface albedos at a global scale over the past decades. The
required sequential accumulation of data over multiple days, i.e., for different view
conditions, as adopted for MODIS sensors for inferring flux quantities from a num-
ber of instantaneous radiance measurements, can be extended to exploit the AVHRR
series data archive (see for instance d’Entremont et al., 1999). An analogous strat-
egy of sequential accumulation (but over every single day, i.e., for different solar
illumination conditions), can be envisaged in the case of the archived measurements
collected by sensors placed on a geostationary orbit. The exploitation of the reci-
procity principle then allows the production for every sample area, of daily accumu-
lated datasets of radiances measured at different viewing angles (see for instance,
Lattanzio et al., 2006). Assuming thus that the geophysical system under investiga-
tion does not suffer from drastic changes during the period of data accumulation,
e.g., multiple hours (days) for geostationary (polar) orbiting sensors, the temporal
sampling of the radiance field for a given location can be interpreted as an angular
sampling.

9.3 MODIS Albedo and Anisotropy Algorithm

The operational MODIS albedo and anisotropy algorithm makes use of a kernel-
driven, linear model of the Bidirectional Reflectance Factors (BRFs), which relies
on the weighted sum of an isotropic parameter and two functions (or kernels) of
viewing and illumination geometry (Roujean et al., 1992) to estimate the Bidirec-
tional Reflectance Distribution Function (BRDF). One kernel is derived from ra-
diative transfer models (Ross, 1981) and the other is based on surface scattering
and geometric shadow casting theory (Li and Strahler, 1992). The kernel weights
selected are those that best fit the cloud-cleared, atmospherically corrected surface
reflectances available for each location over a 16-day period (Lucht et al., 2000;
Schaaf et al., 2002). This model combination (Ross-Thick/Li-Sparse-Reciprocal or
RTLSR) has been shown to be well suited to describing the surface anisotropy of
the variety of land covers that are distributed world-wide (Privette et al., 1997;
Lucht et al., 2000) and is similar to the kernel-driven schemes used to obtain
anisotropy and albedo information by the POLDER (Leroy and Hautecoeur, 1998;
Bicheron and Leroy, 2000; Maignan et al., 2004) satellite sensor. Once an appro-
priate anisotropy model has been retrieved, integration over all view angles results
in a Directional Hemispherical Reflectance (DHR) or a black-sky albedo at any
desired solar angle and a further integration over all illumination angles results in
a BiMemispherical Reflectance (BHR) under isotropic illumination or a white-sky
albedo. These albedo quantities are intrinsic to a specific location and are governed
by the character and structure of its land cover. They can be combined with appro-
priate optical depth information to produce an actual (blue-sky) albedo for a spe-
cific time such as would be measured at the surface by field sensors under ambient
9 Retrieval of Surface Albedo from Satellite Sensors 223

illumination. The anisotropy models can also be used to compute surface reflec-
tances at any other view or solar zenith angle desired. The spectral acquisitions can
also be combined via narrow to broadband conversion coefficients (Liang et al.,
1999; Liang, 2000) to provide broadband anisotropy information and thus broad-
band albedos similar to those routinely collected in the field with pyranometers and
commonly used in large-scale models.
The MODIS instruments on both Aqua and Terra have a 16-day repeat cycle
and provide measurements on a global basis every 1–2 days. The 16-day period has
also been chosen as an appropriate tradeoff between the availability of sufficient
angular samples and the temporal stability of surface (Wanner et al., 1997; Gao
et al., 2001). This assumption becomes tenuous during periods of strong phenolog-
ical change such as vegetation greenup, senescence, or harvesting. By overlapping
processing of the data such that retrievals are attempted every 8 days (based on
all clear observations over the past 16 days), some of the phenological variabil-
ity can be more accurately captured. Other periods of rapid change at the surface
such as ephemeral snowfall also provide challenges in retrieving appropriate sur-
face albedos. The MODIS algorithm addresses this by determining whether the
majority of the clear observations available over a 16-day period represent snow-
covered or snow-free situations and then retrieving the albedo of the majority con-
dition accordingly.
For those locations where the full anisotropic model described above can not be
confidently retrieved due to poor or insufficient input observations, a backup algo-
rithm is employed. This method (Strugnell and Lucht, 2001; Strugnell et al., 2001)
relies on a global database of archetypal anisotropic model based on a land cover
classification and historical high quality full model retrievals. This a priori data base
is then used as a first guess of the underlying anisotropy and any available observa-
tions are used to constrain the model. While considered a lower quality result, Jin
et al. (2003a, b) and Salomon et al. (2006) have found that this backup method often
performs quite well under normal situations (e.g., Fig. 9.1).
In view of the often insufficient angular sampling available, a synergistic use
of multi-sensor observations has offered the best opportunity to improve both the
coverage and the quality of global anisotropy and albedo retrievals. Terra has a de-
scending equatorial crossing time of 10:30 a.m., while Aqua is flying in an asce-
nding orbit with a 1:30 p.m. equatorial crossing time. By combining MODIS ob-
servations from both Terra and Aqua, more high-quality, cloud-free observations
(under varying solar zenith angle) are available to generate better constrained model
retrievals (see Fig. 9.1). Since the MODIS-Terra and MODIS-Aqua have similar in-
strument characterizations and utilize the same atmospheric correction algorithm,
the combination of these data is fairly straightforward. However, the calibration and
geolocation of both instruments must be continually monitored for compatibility and
the quality of the aerosol retrieval from each sensor and its effect on the respective
atmospherically-corrected surface reflectances must also be accounted for. In gen-
eral, the combined Terra and Aqua MODIS product processing stream begins with
a detailed quality check of each atmospherically corrected surface reflectance and
then assigns various penalty weights to the individual observations according to
224 C. Schaaf et al.

MODIS Surface Albedo vs. Boulder, Colorado, 2003


0.90

0.81 Ground Site (Daily at LSN)

Ground Site (16-Day Average)


0.72
TERRA (Full Inv.)
TERRA (Mag. Inv.)
0.63
AQUA + TERRA (Full Inv.)
Surface Albedo

0.54 AQUA + TERRA (Mag. Inv.)

0.45

0.36

0.27

0.18

0.09

0.00
1 17 33 49 65 81 97 113 129 145 161 177 193 209 225 241 257 273 283 305 321 337 353 369
Julian Day

Fig. 9.1 MODIS 16-day broadband albedo as compared to field data from the Boulder BSRN
tower site

the quality flag contained in each surface reflectance product (Schaaf et al., 2002).
Thus, the quantified uncertainty of the sensor-specific surface reflectances is di-
rectly integrated into the retrieval. Results from the combined Terra-MODIS and
Aqua-MODIS algorithm (Fig. 9.2) indicate that the increase in the number of ob-
servations does result in more higher quality retrievals and can decrease the use of
backup retrievals by as much as 50% (Salomon et al., 2006).
The MODIS BRDF/Albedo standard operational products (Lucht et al., 2000;
Schaaf et al., 2002; Gao et al., 2005) provide the best fit RTLSR model parameters
describing the surface anisotropy, black sky and white sky albedo quantities,
the nadir (view-angle-corrected) surface reflectance of each location, and exten-
sive quality information. The best fit RTLSR model parameters are retrieved
for the first seven spectral bands of MODIS and three additional broadbands
(0.3–0.7 µm, 0.7–5 µm, 0.3–5 µm). These anisotropy models are then used to com-
pute white sky albedo and black sky albedo at local solar noon for the same seven
spectral bands and three broadbands. The anisotropy models are also used to correct
surface reflectances for view angle effects and provide BRFs at a common nadir
view angle (Fig. 9.3). These Nadir BRDF-Adjusted Reflectances (NBAR) are com-
puted for the seven spectral bands and are used as the primary input for the MODIS
Land Cover and Land Cover Dynamics Products due to their stability and temporal
consistency (Friedl et al., 2002; Zhang et al., 2003). In addition to the standard 500 m
and 1 km tiled products in a sinusoidal projection, these same science data sets are
9 Retrieval of Surface Albedo from Satellite Sensors 225

Fig. 9.2 Quality improvement possible by combining MODIS observations from both the Terra
and Aqua platforms. Top panel shows Terra alone (green high quality, red lower quality) while
bottom panel shows Terra and Aqua (March 2006)
MODIS Reflectance (MODO9GHK) 2004-126 Nadir BRDF-Adjusted Reflectance (NBAR) 2004-126

NIR(0.1-0.4) Red (0.0-0.2) Green (0.0-0.18) NIR(0.1-0.4) Red (0.0-0.2) Green (0.0-0.18)

MODO9GHK. View Angular effects Nadir BRDF-Adjusted Reflectance (NBAR).


between two swaths Angular effect is removed
(North China Plain) (North China Plain)

Fig. 9.3 Application of the MODIS anisotropy model parameter product to correct adjoining sur-
face reflectance swaths for view angle effects by generating Nadir BRDF-Adjusted Reflectances
(NBAR) 5 May 2004
226 C. Schaaf et al.

Fig. 9.4 MODIS global white sky albedo (from Terra and Aqua) March 2006

also routinely produced at a 0.05◦ spatial resolution in a global geographic (lati-


tude/longitude) projection specifically for use by global modelers (Gao et al., 2005).
In Fig. 9.4, the global false color field of spectral white sky albedo (March 2006)
captures the seasonal variation due to vegetation phenology and snow cover extent.

9.4 MISR Albedo and Anisotropy Algorithm

The Multi-angle Imaging SpectroRadiometer (MISR) on the EOS Terra platform


consists of nine pushbroom cameras, viewing symmetrically about nadir in for-
ward to aftward directions along the spacecraft track. Image data are acquired
with nominal view zenith angles relative to the surface reference ellipsoid of
0.0◦ , 26.1◦ , 45.6◦ , 60.0◦ , and 70.5◦ in four spectral bands (446, 558, 672, and 866 nm)
and with a crosstrack ground spatial resolution of 275 m to 1.1 km and a swath width
of about 400 km (Diner et al., 1998). After these data are radiometrically calibrated,
georectified, and averaged to a uniform resolution of 1.1 km, the land data undergo
a series of processing steps, resulting in a myriad of surface parameters.
The basic land surface products currently being generated include the spectral
hemispherical-directional reflectance factor (HDRF) at the nine MISR view angles
and the associated BHR. The HDRF is a measure of the directional reflectance of the
surface under ambient atmospheric illumination (i.e., direct plus diffuse radiation).
It is the ratio of the directionally reflected radiance from the surface to the reflected
radiance from an ideal lambertian target under identical illumination conditions as
the surface. The BHR is the HDRF integrated over all reflection angles in the up-
ward hemisphere, i.e., it is the surface albedo under ambient atmospheric illumina-
tion. Related MISR surface parameters are the spectral BRFs at the nine MISR view
9 Retrieval of Surface Albedo from Satellite Sensors 227

angles and the DHR. The BRF and the DHR characterize the surface in the same
way as the HDRF and BHR, respectively, but are defined for the condition of direct
(i.e., collimated beam) illumination only. Thus, the top-of-atmosphere (TOA) MISR
radiances are first atmospherically corrected to produce the HDRF and the BHR,
surface reflectance properties as would be measured at ground level but at the MISR
spatial resolution. The HDRF and BHR then are further atmospherically corrected to
remove all diffuse illumination effects, resulting in the BRF and DHR. In addition
to these spectral surface reflectance products, the BHR and DHR, integrated over
the wavelength region of Photosynthetically Active Radiation (PAR) (400–700 nm),
are also computed. The determination of these surface products requires that the at-
mosphere be sufficiently characterized in order for the correction process to occur.
This characterization is accomplished by means of aerosol retrieval, a process per-
formed on a region 17.6 × 17.6 km in size, containing the 1.1 × 1.1 km size subre-
gions (Martonchik et al., 1998a, 2002a). After a surface BRF is determined at the
subregion scale it is fitted to the three parameter modified Rahman–Pinty–Verstraete
(MRPV) empirical model (Rahman et al., 1993; Engelsen et al., 1996), which pro-
vides a convenient representation of the surface scattering characteristics The details
of the retrieval methodologies used to derive these various surface products have
been described by Martonchik et al. (1998b).
The unique capabilities of MISR’s multiple cameras allow for a simultaneous
sampling of the surface anisotropy. By coupling the angular information with the
spectral information, the MISR observations can be exploited to capture ephemeral
effects such as springtime snow cover. On 17 April 2001 MISR observed a rural
part of Manitoba and Saskatchewan about 110 km north of the US border (Path 34,
Orbit 7083). Most of MISR’s imaging data have a resolution of 1.1 km, but all nine
cameras in the red band (672 nm) and all four bands in the nadir camera take global
data at the higher resolution of 275 m. Figure 9.5 shows two false color images of the
Canadian scene at 275 m resolution, one emphasizing spectral information and the
other, angular information. The image on the left is a multispectral color composite
in which the MISR green band (558 nm), red band, and near IR band (866 nm)
nadir view imagery are colored blue, green, and red. Here, vegetation appears red
due to its high reflectivity in the near IR band and low reflectivity in the green and
red bands. The image on the right is a multiangular color composite in which the
60◦ forward view, the nadir view, and 60◦ aftward view images are colored blue,
green and red, respectively, essentially color coding the angular signature of the
scene. Thus, for example, a region with a reflectance predominately in the nadir
direction will appear green. Prominent features in both images are the Assiniboine
and Qu’Appelle rivers, running southward and eastward, respectively.
The bidirectional reflectance factors for three sites marked by yellow arrows in
Fig. 9.5 are displayed in Fig. 9.6. The solar zenith angle is 42◦ and the azimuth
angles of the BRFs are about 32◦ from the principal plane. The northern-most site
is reddish in color in both composite images in Fig. 9.5, indicating vegetation with
substantial backscatter. This backscatter signature is more clearly shown in the BRF
plot in Fig. 9.6 The BHR (i.e., actual albedo), is 0.08 in the red band and the NDVI
is 0.49, implying moderately dense vegetation. The eastern-most site also has a
228 C. Schaaf et al.

Fig. 9.5 Two false color images (275 m resolution) of an area (240 × 175 km) in central Canada on
17 April 2001 centered on the Saskatchewan–Manitoba border. The left image is a multispectral
composite in which red (more like purple/blue in the pdf/doc versions of the images) indicates veg-
etation. The image on the right is a multiangular composite in which green indicates predominate
scattering in the nadir direction

Moderately dense vegetation (albedo = 0.08, NDVI = 0.49)


Snowy forest (albedo = 0.18, NDVI = 0.24)
Agricultural field with light snow (albedo = 0.18, NDVI = 0.13)
0.25

0.20
BRF (red band)

0.15

0.10

0.05
−80 −60 −40 −20 0 20 40 60 80
View Zenith angle (deg)

Fig. 9.6 Bidirectional reflectance factors for three sites marked by yellow arrows in Fig. 9.5

vegetative character, colored purple and red in the multiangular and multispectral
composites, respectively. However, the BRFs for this site are higher and with more
forward scattering than for the previous site, and with an increased red band BHR
of 0.18 and a lower NDVI of only 0.13. The brightening of the BRF, the increase
9 Retrieval of Surface Albedo from Satellite Sensors 229

in forward scattering, and the decrease in the NDVI is indicative of the presence of
some snow in and among small scale vegetation, probably agricultural. The last site
in Fig. 9.5 is located in Duck Mountain Provincial Park and is colored green in the
multiangular composite and red in the multispectral composite. This color combi-
nation implies vegetation but with strong scattering in the nadir direction. The BRF
plot in Fig. 9.6 shows this signature in more detail, and it is quite different than those
from the other two sites. Here, the vegetation is a forest with snow on the ground
between the trees. When viewing in or near the nadir direction, the snow is highly
visible and the BRF is at its highest. As the view angle progressively increases, the
ratio of snow cover to tree structure decreases, lowering the reflectance until, at the
extreme off-nadir view angles, the BRF is virtually all canopy with its characteristic
pronounced backscatter.
During the 7 weeks from 14 August to 29 September 2000, numerous orbits of
MISR data were analyzed and compiled for southern Africa, as part of the dry season
campaign of the Southern Africa Regional Science Initiative (SAFARI-2000), an in-
ternational effort to study linkages between land and atmospheric processes. During
this period a number of AERONET sunphotometer sites (Holben et al., 1998) were
operational over the region, providing independent determinations of aerosol opti-
cal depth which were compared to those retrieved using MISR data (Diner et al.,
2001). This validation study produced very favorable results, allowing considerable
confidence to be placed in the subsequent atmospheric correction procedures and
in the quality of the retrieved surface products. Figure 9.7 is a true color, 1.1 km
resolution mosaic of the surface DHR for southern Africa, derived from 27 orbital
swaths accumulated during this time period. The bright feature in the center is the
Makgadikgadi Pans, an extensive salt bed in Botswana.
In the interior part of southern Africa, much of the land can be classified as
savanna and grassland. Figure 9.8 shows the HDRFs in all four MISR bands for
grassland not far from Johannesburg. The grass is dried out, as can be discerned
from the monotonically increasing HDRF with wavelength. Using data from 15
August (Path 168, Orbit 3509), this particular site was positioned on the extreme
western edge of MISR’s orbital swath, providing multispectral measurements within
1◦ of the retro-solar direction (direct backscatter). The resulting hotspot, due to an
almost complete lack of shadowing within the structured surface, can be seen very
clearly in Fig. 9.8 as an enhancement of the HDRF in all bands at 49◦ view zenith
angle, which is also the solar zenith angle. The hotspot, while not common in MISR
surface retrievals, does occur for a wide range of latitudes, appearing at different
camera view angles, depending on the season.
In addition to the standard MISR products available at 275 m or 1.1 km, many are
also available in a format of monthly global maps at a spatial resolution of 0.5◦ in
both latitude and longitude. An example of this type of map is displayed in Fig. 9.9,
showing surface DHR in natural color for the month of September 2005. The in-
dividual 0.5◦ pixels are created by averaging all 1.1 km DHR values accumulated
within that pixel for that month. The white specks evident in some areas are fill pix-
els where no 1.1 km DHR values were available for the entire month, due mainly to
cloud activity.
230 C. Schaaf et al.

Fig. 9.7 MISR true color, 1.1 km resolution mosaic of the surface directional-hemispherical re-
flectance (DHR) for southern Africa (14 August–29 September 2000)

DRY GRASSLAND, SOUTH AFRICA

Blue Green Red Near-IR

0.45
0.40
0.35
0.30
HDRF

0.25
0.20
0.15
0.10
0.05
0.00
−80 −60 −40 −20 0 20 40 60 80
View Zenith angle (deg)

Fig. 9.8 Spectral and angular variation of HDRF for dried grassland on 15 August 2000 for a site
near Johannesburg. Note the hotspot at the view zenith backscatter angle of 49◦ , which is also the
solar zenith angle
9 Retrieval of Surface Albedo from Satellite Sensors 231

Fig. 9.9 Surface DHR in natural color for the month of September 2005

9.5 Meteosat Albedo and Anisotropy Algorithm

The cornerstone of the retrieval algorithm for geostationary satellites developed by


Pinty et al. (2000a) relies on the temporal sampling of geostationary satellites (data
acquired every 15 or 30 min from sunrise to sunset) as if it were an instantaneous
angular sampling of the radiance field emerging at the top of the atmosphere. The
frequency of measurements of the same Earth location is indeed a unique capability
offered by geostationary satellites that thus translates into an increasing number of
conditions or positive constraints to be satisfied by the retrieval algorithm.
The physics of the Meteosat retrieval aims at solving an inverse radiation trans-
fer problem simultaneously with respect to the lower boundary condition, i.e., the
surface bidirectional reflectance factor (BRF), and the aerosol optical thickness
(Martonchik et al., 1998b, 2002b; Pinty et al., 2000a). All other effects due, for in-
stance, to water vapour and ozone in the atmosphere are accounted for via a prescrip-
tion of gas concentrations taken from either climatology and/or weather forecast
models (re) analysis. In order to simplify the problem further, the gaseous absorp-
tion processes are treated separately from the aerosol-scattering-absorbing effects
by the specification of two distinct atmospheric layers, one for the representation of
the molecular absorption only and the other for the modeling of the coupled surface-
aerosol radiation transfer processes. The inverse algorithm is basically focusing on
the estimates of key variables, namely the aerosol load and surface scattering prop-
erties, for which the a priori knowledge is quite limited or somewhat uncertain and
the level of variability is quite high.
The mathematics of the retrievals is established in such a way that the ampli-
tude of the surface BRF is propagated to the top of the atmospheric scattering layer
while its shape is modulated by the atmospheric scattering and absorbing properties
232 C. Schaaf et al.

(Pinty et al., 2000a). This is made possible thanks to (1) the mathematical formu-
lation of the surface BRF model, namely the RPV model (Engelsen et al., 1996;
Rahman et al., 1993) which separates the amplitude from the shape of the surface
BRF and (2) the decomposition of some atmospheric functions like the upward and
downward diffuse transmission with a Fourier expansion limited to the first two
components. This approach proved to be computer efficient and accurate for model-
ing the radiance field at the top of a scattering-only atmosphere (Martonchik et al.,
2002b). In this way, the angular field of the BRF at the top of the scattering at-
mosphere can be simply expressed as a sum of contributions invoking the coupling
between the surface BRF shape and atmospheric scattering functions that can all be
pre-computed and called during the retrieval procedure. For a given set of measured
BRF values at the top of the atmosphere the latter procedure itself solves (1) a lin-
ear equation to calculate first the amplitude values of the surface BRF for the given
pre-computed scattering functions and (2) a second order cost function estimating
the closeness between the measured and the modeled BRF values at the top of the
atmosphere.
The retrieval procedure then ends up with an identification of probability distri-
bution functions of the acceptable solutions, i.e., those satisfying one or multiple
criteria depending on the number of degrees of freedom and the distributions of
uncertainties in both the observations and in the forward model. The selection of
the “most probable” solution for any given set of measured BRF amongst the set
of acceptable solutions can be performed using various criteria including the iden-
tification of the solution corresponding to the arithmetic mean of the distribution
of the amplitude of the surface BRF values. The solution retained is thus a set of
model variables and parameters describing the surface scattering problem, such as
the parameters characterizing the shape of the surface BRF and the effective aerosol
loads, associated with the selected value for the amplitude of the surface BRF. The
aerosol loads together with the surface scattering properties are given via effective
optical thickness values for a prescribed aerosol type corresponding to average stan-
dard aerosol conditions regarding their detailed properties and vertical distribution
as well. Since the retrieval strategy delivers the optimized set of the RPV model
parameters characterizing the surface BRDF, one can generate DHR or black-sky
albedo for any solar angle and/or BHR or white-sky albedo products (Pinty et al.,
2000a).
The abundance of cloudy conditions occurring during a day over the Earth disk
sections sensed by geostationary satellites motivates the implementation of a pro-
cedure screening conditions that do not correspond to clear-sky cases. Pinty et al.
(2000b) suggested adoption of an angular consistency check by which the daily top
of the atmosphere radiance series for each individual pixel is used in an attempt to
fit the MRPV model. This is based on a recursive filtering technique which identi-
fies sequentially during the day, the outliers deviating significantly from the fitted
MRPV model solutions. In the vast majority of the cases, larger (lower) radiance
values than the MRPV fitted solutions are associated with cloudy (shadowed) con-
ditions and the spatio-temporal fields of these outliers were shown to, indeed, dis-
play very consistent cloud fields and associated shadows along the course of the day
9 Retrieval of Surface Albedo from Satellite Sensors 233

(Pinty et al., 2000a). Measurements acquired at solar and view zenith angles larger
than 70◦ or corresponding to cloud conditions are rejected. A minimum of six valid
clear sky observations are necessary for the activation of the retrieval procedure.
These cloudy as well as additional undesired conditions translate into incomplete
geographic surface albedo map products. This caveat can be overcome by applying
a time composite algorithm selecting, over a given composite time period, the par-
ticular day delivering the albedo value which is the closest to the average of the
ensemble of values retrieved during that same time period. In this way, the geo-
physical values for each pixel can be delivered with all the relevant information
used to generate them in the retrieval algorithm such as, for example, the number
of observations used to perform the retrieval and estimation of the retrieval uncer-
tainties among others. Sensors on board geostationary satellites sample the scattered
solar radiance fields in a usually single, large (according to today’s standards) spec-
tral band (see Fig. 9.10), e.g., ≈ 0.4–1.1 µm for Meteosat and GMS, 0.05–0.8 µm

Fig. 9.10 Examples of sensor spectral responses on board geostationary satellites in the solar do-
main (red line). The green solid line illustrates typical reflectance of green vegetation
234 C. Schaaf et al.

for GOES, which prevents us benefiting from advanced atmospheric correction al-
gorithms based on multi-spectral information, as is possible for the MODIS and
MISR sensors. Due to the fast spatio-temporal variability of the cloud, water vapor
and aerosol fields, the crucial problem of accurately assessing/removing the unde-
sired effects induced by these atmospheric components on the measured radiances
remains to be solved on the sole basis of the available information gathered by the
sensor or other sources of information. For those components, such as ozone, that
have a non-negligible but still limited impact on the surface retrievals and whose
variability remains somewhat small, the use of climatological values is generally ac-
ceptable. In most if not all cases to be addressed, the availability of a spectrally large
single band only renders the partitioning between the surface and atmospheric con-
tributions quite difficult since the scattering and absorption effects in the atmosphere
are wavelength-dependent and coupled with spectrally variant surface properties.
To date, the constraints imposed by operational exploitation infrastructures have
not favored the processing of multi-sensor measurements assembled via data fusion
procedures. This state of affairs encourages the development of albedo retrieval al-
gorithms relying on the analysis of data (and data strings) collected by each geo-
stationary sensor in stand alone mode. In turn, this places stringent requirements on
the reliability and overall performance of the retrieval algorithm which thus, on the
basis of a single spectrally large band, must be able to identify cloud occurrence
and then solve, as well as possible, the coupled surface-atmosphere radiation trans-
fer problem. In that context, multiple sensitivity test have to be conducted in order
to optimize the crucial choices to be made such as, for instance, between the length
of the period to perform sequential data accumulation, e.g., between a few hours
and a few days, and the impacts of the assumption hindering this multi-angular data
emulation procedure, e.g., no drastic changes in the geophysical system under in-
vestigation. Meteosat data processed with this algorithm have already been used in
a variety of applications (see for instance, Pinty et al., 2000c; Knorr et al., 2001;
Myhre et al., 2005b).

9.6 Fusion of Modern and Historical Surface Albedo Products

Documenting the Earth climate and its variability requires access to reliable and
accurate long time series of environmental products. Hence, archived meteorologi-
cal satellite observations could contribute to the generation of climatic data records
providing that (1) significant efforts are devoted to the improvement of the spectral
characterization and calibration of these radiometers and (2) developing state of the
art retrieval algorithms. The first point aims at reducing and controlling as much as
possible the impact of measurements uncertainties on the accuracy of the retrievals
while the second point similarly at model uncertainties. Previous studies have, in-
deed, already demonstrated the possibility to perform post-launch improvements of
the radiometer characteristics (e.g., Govaerts, 1999). In the specific case of surface
albedo retrieval, the science context set up by the requirements in the exploitation
9 Retrieval of Surface Albedo from Satellite Sensors 235

of data acquired by modern sensors such as those on board Terra and Meteosat
Second Generation for instance has motivated a large number of studies addressing
these points. The associated improved knowledge translates, in turn, into the devel-
opment and use of better approaches for the optimal exploitation of measurements
taken by “old generation” sensors.
While active efforts are underway within the NASA community to couple
MODIS observations to the historical AVHRR archives (Pedelty et al., 2007) con-
siderable progress has already been made in exploiting historical geostationary
satellite data to establish a historical data set of global surface albedo values. This
requires the back processing and analysis of measurements assembled by the fleet
of geostationary satellites over the past 25 years or so. The quality of these retrievals
can be assessed by various means including first, the comparison of these retrieved
albedo values against those operationally generated since year 2000 from modern
and technologically advanced instruments such as MODIS and MISR (Pinty et al.,
2004) and second, the intercomparison of surface albedo products generated over
geographical regions of overlap that are, therefore, sampled simultaneously by
two adjacent sensors together placed on geostationary orbit but located at different
longitudes (Govaerts et al., 2004).
In order to conduct comparison exercises of relevance for climate model appli-
cations, surface albedo products have to be made available over large spectral re-
gions covering the energetically relevant solar domain [0.3–3.0 µm] split, whenever
possible, into its broad visible [0.3–0.7 µm] and near-infrared [0.7–3.0 µm] parts.
Achieving this step, usually called spectral conversion, requires developing appro-
priate tools to transform albedo product values, estimated over and weighted by the
spectral response of the geostationary sensor, into values representative of the de-
sired broad spectral range of interest (see for instance Liang (2000) and Govaerts
et al. (2006)). One possible solution consists in approximating a parametric expres-
sion relating the measurements from the sensor to those that would be provided
by an ideal rectangular shape sensor covering the solar domain of interest. Such
an expression can be established on the basis of (1) a large number of simulations
of top of atmosphere radiance fields of various geophysical situations that can be
expected for the region of interest (e.g., vegetation with varying density, bare soils
with different brightness, snow surfaces, coupled with a diversity of atmospheric
conditions) and (2) a multi-regression analysis fitting at best the sensor-like values
against those representative of the desired solar domain. This spectral conversion
constitutes quite a delicate step and, as a matter of fact, its reliability relies strongly
on the degree of coincidence between the distributions of the simulated and actual
conditions. Its impact on the uncertainty of the final albedo products also depends
crucially on the sensor spectral response function since the spectral conversion basi-
cally assumes that strong correlations exist between radiances taken across various
wavelengths of the solar spectrum.
Preliminary attempts to compare surface albedo products from modern sen-
sors, such as MODIS and MISR, against those generated by the retrieval algo-
rithm outlined here for geostationary satellites result into quite positive conclusions.
Figure 9.11 illustrates an example of results to be expected when comparing
236 C. Schaaf et al.

Fig. 9.11 Scatterplot (density plot) between the MODIS white sky albedo and Meteosat BHR
values retrieved between 20–65 ◦ W longitude and 40 ◦ S − 40 ◦ N latitude during the first 2 weeks of
year 2001. This plot includes the high quality flag MODIS products only and the outliers from the
two distributions have been removed (less than 5% of the total number of valid retrievals). The full,
dashed, and dotted lines feature the fit obtained using the slope of the means, the linear regression,
and the primary eigenvector, respectively

Meteosat and MODIS (considering high quality flags only) surface albedo products,
in units of white sky albedo (BHR), over a large geographical region extending from
Southern Europe and covering the entire African continent (between 20◦ W–65◦ N
longitude and 40◦ S–40◦ N latitude. This figure is built from the analysis of prod-
ucts available during the first 2-week period of year 2001 after removal of outliers
detected in large majority along the coastlines. The 10-day composite Meteosat
products have been re-mapped into the MODIS Climate Modeling Grid at a spa-
tial resolution of 0.05◦ . The MODIS and Meteosat spectral albedo products were
both converted into an ideal rectangular shape (0.4–1.1 µm) in order to (1) make
the best possible use of the available spectral information for both sensors, i.e.,
one large band in the Meteosat case and four narrow bands well distributed over this
spectral interval in the MODIS case, and (2) minimize the uncertainties associated
with the required spectral conversion. All three indicators used to characterize the
statistical differences between the MODIS and Meteosat products, i.e., the slope of
9 Retrieval of Surface Albedo from Satellite Sensors 237

the means (full line), the regression line (dashed line) and the primary eigenvector
(dotted line), show limited variability around the one-to-one line. The observed sta-
tistical differences are largely within the range of the systematic error/uncertainty
due to the calibration knowledge (about 6% for Meteosat-7) (Govaerts et al., 2004)
and/or model approximations (for instance, the decoupling of the gaseous absorp-
tion from the aerosol scattering effects).
Figure 9.12 illustrates results from a comparison between the broadband black
sky (DHR) albedo products generated, at 30◦ solar angle, by the analysis of the daily
radiances collected during the first 10-day period of May 2001 by both GOES West
(GOES-10) and GOES East (GOES-8) over the common land regions of the Earth
disk that they jointly sample (see top panel in Fig. 9.12). These results illustrate the
robustness of the albedo retrieval algorithm since its application yields differences
between data sets from adjacent sensors that are well within the range of their es-
timated uncertainty level (about 10–15% for the GOES sensors) of their broadband
albedo products. Figure 9.12 confirms earlier comparison results obtained by ana-
lyzing albedo products generated by two adjacent Meteosat sensors (Govaerts et al.,
2004).
The conclusions drawn from Figs. 9.11 and 9.12 suggest the generation of his-
torical series of surface albedo products based on the exploitation of the fleet of
geostationary satellites. Figure 9.13 is a demonstration example of an output from
such an initiative which assembles broadband products (DHR at 30◦ Sun zenith an-
gle) retrieved for the first 10-day period of May 2001 from five different satellites,
namely GOES West and East, Meteosat-7 and Meteosat-5, and GMS-5. Table 9.1
provides statistics about these retrievals and the estimated uncertainties associated
with each satellite retrievals. The measurement error includes both the radiometric
uncertainties and approximations in the forward model. The estimated error on the
DHR values is then derived from the uncertainty on the retrieved surface parame-
ters. It thus looks feasible to build global albedo products for the last 25 years or
so, for those places covered by archived data. These preliminary results open new
avenues for the exploitation of geostationary archive data and prototype the fusion
of such the generated products.

9.7 Summary

The MODIS, MISR and Meteosat algorithms represent three complementary strate-
gies for characterizing land surface reflectance anisotropy and obtaining measures of
land surface albedo. Each algorithm makes use of the unique capabilities of its sen-
sor to capture the spectral, spatial, temporal, and angular information necessary to
accurately specify the reflective qualities of the underlying surface cover. With more
than 6 years of MODIS and MISR observations now available, as well as the op-
portunity to utilize the historical geosynchronous satellite record, the modeling and
data analysis communities enjoy unprecedented access to consistent, high-quality
albedo and anisotropy data of the Earth’s land surface.
238 C. Schaaf et al.

Fig. 9.12 Comparison between broadband DHR (30◦ ) values retrieved over the common geo-
graphical region covered by both GOES West (GOES-10) and GOES East (GOES-8). The top
panel corresponds to the density plot of the two DHR distributions for the period 1–10 May 2001.
The bottom panel documents the histogram of the relative differences between the two distribu-
tions. The vertical lines colored in blue feature the mean value of these differences (dash-dotted)
and one standard deviation from the mean (dashed)
9 Retrieval of Surface Albedo from Satellite Sensors 239

Fig. 9.13 Top panel: Location of operational geostationary satellites which archive data and are
used to derived products shown in the bottom panel. The circles show the 60◦ viewing angle limit.
Bottom panel: Illustration of the broadband surface albedo map derived from the application of the
geostationary satellite retrieval algorithm on measurements taken simultaneously by GOES-8/10,
Meteosat-5/-7 and GMS-5 over the period 1–10 May 2001

Table 9.1 Number of days processed during the 1–10 May 2001 period for each satellite.
< Img/day > is the mean number of measurements available per day (note that some of the GOES
images did not provide the nominal geographical coverage). <Meas. R. Err.> is the average mea-
surement relative error, i.e., including both the radiometric error and forward model uncertainty.
<DHR R. Err.> is the mean estimated DHR relative error
Satellite Nbr days <Img/day> <Meas. R. Err.> <DHR R. Err>

GOES-10 10 22.9 5.2% 12.5%


GOES-8 10 13.7 6.8% 14.4%
MET-7 10 17.3 7.4% 8.7%
MET-5 10 16.3 10.0% 10.1%
GMS-5 10 9.9 8.4% 10.5%
240 C. Schaaf et al.

Acknowledgements Authors Schaaf, Strahler, and Liu are supported by NASA (under grant
NNG04HZ14) and by their colleagues on the MODIS Science Team while Martonchik is sup-
ported by the MISR Science Team. Authors Pinty, Govaerts, Lattanzio, and Taberner are grateful to
the Japan Meteorological Agency (JMA) and the Satellite Services Group of the National Oceanic
and Atmospheric Administration (NOAA) for providing the GMS-5 and GOES-8/-10 data, respec-
tively. Their contributions would not have been possible without the support of the Global Envi-
ronment Monitoring unit of the Institute for Environment and Sustainability at the Joint Research
Centre, and EUMETSAT.

References

Barlage M, Zeng X, Mitchell K, Wei H (2005) A global 0.05-deg maximum albedo


dataset of snow-covered land based on MODIS observations. Geophys. Res. Lett. 32,
doi:10.1029/2005GL022881
Barnsley MJ, Strahler AH, Morris KP, Muller J-P (1994) Sampling the surface bidirectional re-
flectance distribution function (BRDF), 1, evaluation of current and future satellite sensors.
Remote Sens. Rev. 8:271–311
Berk A, Bernstein LS, Anderson GP, Acharya PK, Robertson DC, Chetwynd JH, Adler-Golden
SM (1998) MODTRAN cloud and multiple scattering upgrades with application to AVIRIS.
Remote Sens. Environ. 65:367–375
Bicheron P, Leroy M (2000) Bidirectional reflectance distribution function signatures of major
biomes observed from space. J. Geophys. Res. 105:26669–26681
Bonan GB, et al. (2002) The land surface climatology of the NCAR community land model cou-
pled to the NCAR community ClimateModel. J. Clim. 15:3123–3149
Bréon FM, Maignan F, Leroy M, Grant I (2002) Analysis of hot spot directional signatures mea-
sured from space. J. Geophys. Res. 107(16):4282–4296
Collins JB, Woodcock CE (2000) Geostatistical methods for analysis of multiscale variation in
spatial data. Geographical Analysis 32(1):50–63
Csiszar, I, Gutman, G (1999) Mapping global land surface albedo from NOAA/AVHRR. J. Geo-
phys. Res. 104:6215–6228
d’Entremont R, Schaaf CB, Lucht W, Strahler A (1999) Retrieval of red spectral albedo and bidi-
rectional reflectance using AVHRR HRPT and GOES satellite observations of New England.
J. Geophys. Res. 104:6229–6239
Dickinson RE (1983) Land surface processes and climate-surface albedos and energy balance.
Adv. Geophys. 25:305–353
Dickinson RE (1995) Land processes in climate models. Remote Sens. Environ. 51:27–38
Diner DJ, Beckert JC, Reilly TH, Bruegge CJ, Conel JE, Kahn RA, Martonchik JV, Ackerman TP,
Davies R, Gerstl SAW, Gordon HR, Muller J-P, Myneni RB, Sellers PJ, Pinty B, Verstraete MM
(1998) Multi-angle Imaging SpectroRadiometer (MISR) instrument description and experiment
overview. IEEE Trans. Geosci. Remote Sen. 36:1072–1087
Diner DJ, Abdou WA, Bruegge CJ, Conel JE, Crean KA, Gaitley BJ, Helmlinger MC, Kahn RA,
Martonchik JV, Pilorz SH, Holben BN (2001) MISR aerosol optical depth retrievals over south-
ern Africa during the SAFARI-2000 dry season campaign. Geophys. Res. Lett. 28:3127–3130
Engelsen O, Pinty B, Verstraete M M, Martonchik JV (1996) Parameteric bidirectional reflectance
fractor models: evaluation, improvements, and applications. Technical Report EUR 16426 EN.
EC Joint Research Centre
Friedl MA, McIver DK, Hodges JCF, Zhang X, Muchoney D, Strahler AH, Woodcock CE, Gopal
S, Schnieder A, Cooper A, Baccini A, Gao F, Schaaf C (2002) Global land cover from MODIS:
algorithms and early results. Remote Sens. Environ. 83:287–302
Gao F, Schaaf C, Strahler AH, Lucht W (2001) Using a multi-kernel least variance approach to
retrieve and evaluate albedo from limited BRDF observations. Remote Sens. Environ. 76:57–66
9 Retrieval of Surface Albedo from Satellite Sensors 241

Gao F, Schaaf C, Strahler A, Roesch A, Lucht W, Dickinson R (2005) The MODIS BRDF/Albedo
Climate Modeling Grid Products and the variability of albedo for major global vegetation types.
J. Geophys. Res. 110: D01104, doi:10.1029/2004JD00519
Govaerts Y (1999) Correction of the Meteosat-5 and -6 VIS vand relative spectral response with
Meteosat-7 characteristics. Int. J. Remote Sens. 20:3677–3682
Govaerts Y, Clerici M, Clerbaux N (2004) Operational calibration of the Meteosat radiometer VIS
band. IEEE Trans. Geosci. Remote Sens. 43:31900–31914
Govaerts Y, Lattanzio A, Pinty B, Schmertz J (2004) Consistent surface albedo re-
trieved from two adjacent geostationary satellite. Geophys. Res. Lett. 31:L15201,
doi:10.1029/2004GL020418 03
Govaerts Y, Pinty B, Taberner M, Lattanzio A (2006) Spectral conversion of surface albedo de-
rived from Meteosat first generation observations. IEEE Geosci. Remote Sens. Lett. 3:23–27,
doi:10.1109/LGRS.2005.854202
Henderson-Sellers A, Wilson MF (1983) Surface albedo data for climatic modeling. Rev. Geophys.
Space Phys. 21:1743–1778
Holben BN, Eck TF, Slutsker I, Tanre D, Buis JB, Setzer A, Vermote E, Reagan JA, Kaufman YJ,
Nakajima T, Lavenu F, Jankowiak I, Smirnov A (1998) AERONET – A federated instrument
network and data archive for aerosol characterization. Remote Sens. Environ. 66:1–16
Hautecoeur O, Leroy MM (1998) Surface bi-directional reflectance distribution function observed
at global scale by POLDER/ADEOS. Geophys. Res. Lett. 22:4197–4200
Jin Y, Schaaf CB, Woodcock CE, Gao F, Li X, Strahler AH, Lucht W, Liang S (2003a) Consis-
tency of MODIS surface BRDF/Albedo retrievals: 1. Algorithm performance. J. Geophys. Res.
108(D5):4158, doi:10.1029/2002JD002803
Jin Y, Schaaf CB, Woodcock CE, Gao F, Li X, Strahler AH, Lucht W, Liang S (2003b) Consistency
of MODIS surface BRDF/Albedo retrievals: 2. Validation. J. Geophys. Res. 108(D5):4159,
doi:10.1029/2002JD002804
Knorr W, Schnitzler KG, Govaerts Y (2001) The role of bright desert regions in shaping North
African climate. Geophys. Res. Let. 28:3489–3492
Lattanzio A, Govaerts Y, Pinty B (2006) Consistency of surface anisotropy characterization with
Meteosate observations. Adv. Space Res. doi:10.1016/j.asr.2006.02.049
Leroy M, Deuze JL, Breon FM, Hautecoeur O, Herman M, Buriez JC, Tanre D, Bouffies S,
Chazette P, Roujean J-L (1997) Retrieval of atmospheric properties and surface bidirectional
reflectances over land from POLDER/ADEOS. J. Geophys. Res. 102:17023–17037
Li X, Strahler AH (1992) Geometric-optical bidirectional reflectance modeling of the discrete
crown vegetation canopy: effect of crown shape and mutual shadowing. IEEE Trans. Geosci.
Remote Sens. 30:276–292
Li Z, Garand L (1994) Estimation of surface albedo from space: a parameterization for global
application. J. Geophys. Res. 99:8335–8350
Liang S (2000) Narrowband to broadband conversions of land surface albedo I: algorithms. Remote
Sens. Environ. 76:213–238
Liang S, Strahler AH, Walthall CW (1999) Retrieval of land surface albedo from satellite observa-
tions: a simulation study. J. Appl. Meteorol. 38:712–725
Lucht W, Schaaf CB, Strahler AH (2000) An algorithm for the retrieval of albedo from space using
semiempirical BRDF models. IEEE Trans. Geosci. Remote Sens. 38:977–998
Maignan F, Bréon FM, Lacaze R (2004) Bidirectional reflectance of Earth targets: evaluation of
analytical models using a large set of spaceborne measurements with emphasis with the hot
spot. Remote Sens. Environ. 90:210–220
Martonchik JV (1997) Determination of aerosol optical depth and land surface directional reflec-
tances using multi-angle imagery. J. Geophys. Res., 102:17015–17022.
Martonchik JV, Diner DJ, Kahn RA, Ackerman TP, Verstraete MM, Pinty B, Gordon HR (1998a)
Techniques for the retrieval of aerosol properties over land and ocean using multiangleimaging.
IEEE Trans. Geosci. Remote Sens. 36:1212–1227
Martonchik JV, Diner DJ, Pinty B, Verstraete MM, Myneni RB, Knyazikhin Y, Gordon HR (1998b)
Determination of land and ocean reflective, radiative, and biophysical properties using multi-
angle imaging. IEEE Trans. Geosci. Remote Sens. 36:1266–1281
242 C. Schaaf et al.

Martonchik JV, Diner DJ, Crean KA, Bull M (2002a) Regional aerosol retrieval results from MISR.
IEEE Trans. Geosci. Remote Sens. 40:1520–1531
Martonchik JV, Pinty B, Verstraete MM (2002b) Note on an improved model of surface BRDF-
atmospheric coupled radiation. IEEE. Trans. Geosci. Remote Sens. 40:1637–1639
Moody EG, King MD, Platnick S, Schaaf CB, Gao F (2005) Spatially complete global spectral
surface albedos: value-added datasets derived from Terra MODIS land products. IEEE Trans.
Geosci. Remote Sens. 43:144–158
Myhre G, Kvalevag MM, Schaaf CB (2005a) Radiative forcing due to anthropogenic vege-
tation change based on MODIS surface albedo data set. Geophys. Res. Lett. 32:L21410,
doi:10.1029/2005GL024004
Myhre G, Govaerts Y, Haywood JM, Berntsen TK, Lattanzio A (2005b) Radiative ef-
fect of surface albedo change from biomass burning. Geophys. Res. Lett. 32:L20812,
doi:10.1029/2005GL022897
Oleson KW, Bonan GB, Schaaf C, Gao F, Jin Y, Strahler A (2003) Assessment of global
climate model land surface albedo using MODIS data. Geophys. Res. Lett. 30(8):1443,
doi:10.1029/2002GL016749
Pedelty J, Devadiga S, Masuoka E, Brown M, Pinzon J, Tucker C, Vermote E, Prince S, Nagol J,
Justice C, Roy D, Ju J, Schaaf C, Liu J, Privette J, Pinheiro, A (2007) Generating a Long-term
Land Data Record from the AVHRR and MODIS Instruments, Proceedings, IEEE International
Geosciences and Remote Sensing Symposium (IGARSS07), Barcelona, Spain, 23–27 July,
2007
Pinty B, Roveda F, Verstraete MM, Gobron N, Govaerts Y, Martonchik J, Diner D, Kahn R (2000a)
Surface albedo retrieval from METEOSAT – Part 1: theory. J. Geophys. Res. 105:18099–18112
Pinty B, Roveda F, Verstraete MM, Gobron N, Govaerts Y, Martonchik J, Diner D, Kahn R (2000b)
Surface albedo retrieval from METEOSAT – Part 2. Appl. J. Geophys. Res. 105:18113–18134
Pinty B, Verstraete MM, Gobron N, Roveda F, Govaerts Y (2000c) Do manmade fires affect
the Earth’s surface reflectance at continental scales? Eos Trans. American Geophys. Union
81:388–389
Pinty B, Taberner M, Liang S, Govaerts Y, Martonchik JV, Lattanzio A, Schaaf CB, Verstraete MM,
Dickinson RE, Gobron N, Widlowski J-L (2004) Intercomparison of surface albedo products
from various spaceborne sensors. In: Proc. of the Workshop on Inter-Comparison of Large
Scale Optical and Infrared Sensors, ESA ESTEC, Noordwijk, The Netherlands, 12–14 October
2004. ESA ESTEC
Privette JL, Eck TF, Deering DW (1997) Estimating spectral albedo and nadir reflectance
through inversion of simple BRDF models with AVHRR/MODIS-like data. J. Geophys. Res.
102:29529–29542
Rahman H, Pinty B, Verstraete MM (1993) Coupled surface-atmosphere reflectance (CSAR)
model – 2: semiempirical surface model usable with NOAA advanced very high resolution
radiometer data. J. Geophys. Res. 98:20791–20801
Roesch A, Schaaf C, Gao F (2004) Use of Moderate-Resolution Imaging Spectroradiometer bidi-
rectional reflectance distribution function products to enhance simulated surface albedos. J.
Geophys. Res. 109(D12), doi:10.1029/2004JD004552
Ross JK(1981) The Radiation Regime and Architecture of Plant Stands,Junk W (ed), Norwell,
MA: Artech House, p. 392
Roujean J-L, Leroy M, Deschamps PY (1992) A bidirectional reflectance model of the Earth’s
surface for the correction of remote sensing data. J. Geophys. Res. 97:20455–20468
Salomon J, Schaaf CB, Strahler AH, Gao F, Jin Y (2006) Validation of the MODIS Bidirectional
Reflectance Distribution Function and albedo retrievals using combined observations from the
aqua and terra platforms. IEEE Trans. Geosci. Remote Sens. 44:No. 6
Sellers PJ, Los SO, Tucker CJ, Justice CO, Dazlich DA, Collatz CJ, Randall DA (1996) A revised
land surface parameterization (SiB2) for atmospheric GCMs, part II, the generation of global
fields of terrestrial biospheric parameters from satellite data. J. Clim. 9:706–737
9 Retrieval of Surface Albedo from Satellite Sensors 243

Schaaf CB, Gao F, Strahler AH, Lucht W, Li X, Tsang T, Strugnell NC, Zhang X, Jin Y, Muller J-P,
Lewis P, Barnsley M, Hobson P, Disney M, Roberts G, Dunderdale M, Doll C, d’Entremont
R, Hu B, Liang S, Privette JL (2002) First operational BRDF, Albedo and Nadir Reflectance
Products from MODIS. Remote Sens. Environ. 83:135–148
Strugnell N, Lucht W (2001) An algorithm to infer continental-scale albedo from AVHRR data,
land cover class and field observations of typical BRDFs. J. Climate 14:1360–1376
Strugnell N, Lucht W, Schaaf C (2001) A global albedo data set derived from AVHRR data for use
in climate simulations. Geophys. Res. Lett. 28:191–194
Tian Y, Dickinson RE, Zhou L, Myneni RB, Friedl M, Schaaf CB, Carroll M, Gao F (2004) Land
boundary conditions from MODIS data and consequences for the albedo of a climate model.
Geophys. Res. Lett. 31, doi:10.1029/2003GL019104
Vermote EF, Tanre D, Deuze JL, Herman M, Morcrette JJ (1997) Second simulation of the satellite
signal in the solar spectrum: an overview. IEEE Trans. Geosci. Remote Sens. 35:675–686
Walthall CL, Norman JM, Welles JM, Campbell G, Blad BL (1985) Simple equation to approxi-
mate the bidirectional reflectance from vegetation canopies and bare soil surfaces. Appl. Opt.
24:383–387
Wanner W, Li X, Strahler AH (1995) On the derivation of kernels for kernel-driven models of
bidirectional reflectance. J. Geophys. Res. 100:21077–21090
Wanner W, Strahler AH, Hu B, Lewis P, Muller J-P, Li X, Barker Schaaf CL, Barnsley MJ (1997)
Global retrieval of bidirectional reflectance and albedo over land from EOS MODIS and MISR
data: theory and algorithm. J. Geophys. Res. 102:17143–17162
Woodcock CE, Harward VJ (1992) Nested-Hierarchical Scene Models and image segmentation.
Int. J. Remote Sens. 13(16):3167–3187
Zhang X, Friedl MA, Schaaf CB, Strahler AH, Hodges JCF, Gao F, Reed BC, Huete A (2003)
Monitoring vegetation phenology using MODIS. Remote Sens. Environ. 84:471–475
Zhou L, Dickinson RE, Tian Y, Zeng X, Dai Y, Yang Z-L, Schaaf CB, Gao F, Jin Y, Strahler A,
Myneni RB, Yu H, Wu W, Shaikh M (2003) Comparison of seasonal and spatial variations of
albedos from Moderate-Resolution Imaging Spectroradiometer (MODIS) and Common Land
Model. J. Geophys. Res. 108(D15):4488, doi:10.1029/2002JD003326
Chapter 10
Modeling and Inversion in Thermal Infrared
Remote Sensing over Vegetated Land Surfaces

Frédéric Jacob, Thomas Schmugge, Albert Olioso, Andrew French,


Dominique Courault, Kenta Ogawa, Francois Petitcolin, Ghani Chehbouni,
Ana Pinheiro, and Jeffrey Privette

Frédéric Jacob
Formerly at Remote Sensing and Land Management Laboratory
Purpan Graduate School of Agriculture, Toulouse, France
Now at Institute of Research for the Development
Laboratory for studies on Interactions between Soils – Agrosystems – Hydrosystems
UMR LISAH SupAgro/INRA/IRD, Montpellier, France
frederic.jacob@supagro.inra.fr
Thomas Schmugge
Gerald Thomas Professor of Water Resources
College of Agriculture
New Mexico State University, Las Cruces, NM, USA
Albert Olioso and Dominique Courault
National Institute for Agronomical Research
Climate – Soil – Environment Unit
UMR CSE INRA/UAPV, Avignon, France
Andrew French
United States Department of Agriculture/Agricultural Research Service
US Arid Land Agricultural Research Center, Maricopa, AZ, USA
Kenta Ogawa
Department of Geo-system Engineering, University of Tokyo
Japan
Francois Petitcolin
ACRI-ST, Sophia Antipolis, France
Ghani Chehbouni
Institute of Research for the Development
Center for Spatial Studies of the Biosphere
UMR CESBio CNES/CNRS/UPS/IRD, Toulouse, France
Ana Pinheiro
Biospheric Sciences Branch, NASA’s GSFC, Greenbelt, MD, USA
Jeffrey Privette
NOAA’s National Climatic Data Center, Asheville, NC, USA

S. Liang (ed.), Advances in Land Remote Sensing, 245–291. 245


c Springer Science + Business Media B.V., 2008
246 F. Jacob et al.

Abstract Thermal Infra Red (TIR) Remote sensing allows spatializing various land
surface temperatures: ensemble brightness, radiometric and aerodynamic tempera-
tures, soil and vegetation temperatures optionally sunlit and shaded, and canopy
temperature profile. These are of interest for monitoring vegetated land surface
processes: heat and mass exchanges, soil respiration and vegetation physiologi-
cal activity. TIR remote sensors collect information according to spectral, direc-
tional, temporal and spatial dimensions. Inferring temperatures from measurements
relies on developing and inverting modeling tools. Simple radiative transfer equa-
tions directly link measurements and variables of interest, and can be analytically
inverted. Simulation models allow linking radiative regime to measurements. They
require indirect inversions by minimizing differences between simulations and ob-
servations, or by calibrating simple equations and inductive learning methods. In
both cases, inversion consists of solving an ill-posed problem, with several parame-
ters to be constrained from few information.
Brightness and radiometric temperatures have been inferred by inverting simu-
lation models and simple radiative transfer equations, designed for atmosphere and
land surfaces. Obtained accuracies suggest refining the use of spectral and temporal
information, rather than innovative approaches. Forthcoming challenge is recov-
ering more elaborated temperatures. Soil and vegetation components can replace
aerodynamic temperature, which retrieval seems almost impossible. They can be
inferred using multiangular measurements, via simple radiative transfer equations
previously parameterized from simulation models. Retrieving sunlit and shaded
components or canopy temperature profile requires inverting simulation models.
Then, additional difficulties are the influence of thermal regime, and the limitations
of spaceborne observations which have to be along track due to the temperature fluc-
tuations. Finally, forefront investigations focus on adequately using TIR information
with various spatial resolutions and temporal samplings, to monitor the considered
processes with adequate spatial and temporal scales.

10.1 Introduction

Using TIR remote sensing for environmental issues have been investigated the
last three decades. This is motivated by the potential of the spatialized infor-
mation for documenting the considered processes within and between the Earth
system components: cryosphere [1–2], atmosphere [3–6], oceans [7–9], and land
surfaces [10]. For the latter, TIR remote sensing is used to monitor forested ar-
eas [11–14], urban areas [15–17], and vegetated areas. We focus here on vegetated
areas, natural and cultivated. The monitored processes are related to climatology,
meteorology, hydrology and agronomy: (1) radiation, heat and water transfers at the
soil–vegetation–atmosphere interface [18–24]; (2) interactions between land surface
and atmospheric boundary layer [25]; (3) vegetation physiological processes such as
transpiration and water consumption, photosynthetic activity and CO2 uptake, vege-
tation growth and biomass production [26–39]; (4) soil processes such as respiration
10 Modeling and Inversion in Thermal Infrared Remote Sensing 247

and CO2 uptake, evapotranspiration and water depletion, spatio-temporal variability


of soil moisture [39–43]; (5) long-term dynamics of land cover [44], land surface
radiative budget [45–48], water shortage and drought [49].
TIR remote sensing allow retrieving emissivity and temperature, with various
complexity degrees presented in Section 10.2. The remotely sensed information
is collected from operational and prospective sensors, listed in Section 10.3. This
information is characterized by temporal and spatial dimensions (Section 10.3.1),
as well as by spectral and directional dimensions (Section 10.3.2). Then, inferring
emissivity and temperature consists of developing and inverting modeling tools,
by exploiting the dimensions of the collected information (Section 10.4). Based
on TIR fundamentals (Section 10.4.1), simple radiative transfer equations directly
link measurements to emissivities and temperatures of interest (Section 10.4.2),
and simulation models describe the influence of radiative regime on measurements
(Section 10.4.3). However, simple radiative transfer equations must be parameter-
ized, and simulation models require significant information. Further, inversion is
not trivial: most of simulation models are not directly invertible, and the numer-
ous parameters to be constrained from remote sensing often make inversion an ill-
posed problem (Section 10.4.4). The several solutions proposed to overcome these
difficulties are assessed using validations, intercomparisons, and sensitivity studies
(Section 10.5).
Current limitations and proposed solutions are presented with an increasing
complexity for the temperatures of interest (Section 10.6). Atmospheric pertur-
bations are corrected by inverting modeling tools for atmosphere, and surface
brightness temperature measurements are simulated using modeling tools for land
surfaces (Section 10.6.1). Surface emissivity effects are removed using simple radia-
tive transfer equations (Section 10.6.2). Reported performances suggest accuracies
rather close to requirements, though refinements are necessary. Recovering tem-
perature for the one source modeling of heat transfers is still not trivial, since the
required parameterization significantly varies in time and space (Section 10.6.3).
Recent studies suggested focusing on more elaborated temperatures: soil and veg-
etation components, optionally sunlit and shaded, and canopy temperature profile.
Their retrieval is a forthcoming challenge, with efforts on measuring, modeling and
inversion (Section 10.6.4). The paper ends with forefront investigations about space
and time issues in TIR remote sensing: monitoring land processes with adequate
spatial scales and temporal samplings, by using available remote sensing observa-
tions (Section 10.7).

10.2 Land Surface Emissivity/Temperature from TIR Remote


Sensing

This section defines the various terms considered in TIR remote sensing, which
are related to land surface emissivity and temperature. We focus on their physi-
cal definitions and various interests. The corresponding equations are detailed in
Section 10.4.
248 F. Jacob et al.

– Surface brightness temperature is equivalent to the radiance outgoing from the


target, by assuming a unity emissivity [50], and corresponds to the basic TIR
remote sensing measurement. It is recovered from at sensor measurements per-
forming atmospheric corrections. It can be assimilated, using modeling tools for
land surface, into process models such as SVAT and crop models [18, 38, 39, 43].
– Ensemble waveband emissivity is needed to derive radiometric temperature from
brightness temperature [50, 51]. It is also useful for retrieving ensemble broad-
band emissivity, a key parameter for land surface radiative budget [52–54].
– Ensemble radiometric temperature is emissivity normalized [50, 51]; and corre-
sponds to kinetic temperature for an homogeneous and isothermal surface [55]. It
is used to estimate surface energy fluxes and water status from spatial variability
indicators: the vegetation index / temperature triangle [41, 56–58]; or the albedo /
temperature diagram [23, 37, 59, 60]. It is also used for retrieving soil and vege-
tation temperatures from two source energy balance modeling [19, 24].
– Aerodynamic temperature is air temperature at the thermal roughness length [50].
It is the physical temperature to be used with one source models of surface energy
fluxes based on excess resistance [61–63]. These can be SVAT models [39, 64];
or energy balance models [22, 23, 37, 59, 60, 65, 66].
– Soil and vegetation temperatures correspond to kinetic [67] or radiometric [68]
temperatures. They are often used for two-source modeling. The latter can be
SVAT models [43, 67, 69]; or energy balance models [20, 70, 71]. Retrieving
these temperatures requires an adequate estimation of directional ensemble emis-
sivity.
– Sunlit and shaded components are refinements of soil and vegetation tempera-
tures. They can significantly differ, according to various factors which drive the
thermal regime: the water status, the solar exposure resulting from the canopy
geometry and the illumination direction. These components are of interest for
understanding canopy directional brightness and radiometric temperatures [58,
72–74].
– Canopy temperature profile, from the soil surface to the top of canopy, is the
finest temperature one can consider. Similarly to sunlit and shaded components
for soil and vegetation temperatures, this thermal regime is considered for under-
standing canopy directional brightness and radiometric temperatures, in relation
with local energy balance within the canopy [75–78].
The seek accuracies vary from one application to another,according to the sensitiv-
ities of process models. For temperature, the goal is accuracy better than 1 K [79].
For emissivity, the goal is absolute accuracy better than 0.01 [80]. Recovering both
relies on exploiting the dimensions of the TIR remotely sensed information.

10.3 Available Information from TIR Remote Sensing

The four dimensions of the remotely sensed information are temporal and spa-
tial (Section 10.3.1), and spectral and directional (Section 10.3.2). Due to orbital
10 Modeling and Inversion in Thermal Infrared Remote Sensing 249

rules and technological limitations, current spaceborne sensors cannot provide full
information over these dimensions. Further, the latter can be linked, according to
the mission objectives: a daily monitoring with sun-synchronous sensor requires a
kilometric resolution with an across track angular sampling. Exploratory missions
with airborne and ground-based sensors are under progress, for assessing the po-
tential of original remotely sensed information. Table 10.1 provides an overview of
the main operational and prospective sensors. We deal here with recent, current and
forthcoming US and EU missions.

10.3.1 Temporal and Spatial Capabilities

The temporal dimension corresponds to the time interval between consecutive ob-
servations. It is of importance for monitoring land surface temperature and related
processes: radiative and convective transfers, soil respiration and vegetation physi-
ological activity. The spatial dimension corresponds to the ground resolution of the
measurements. It is of importance for the meaning of surface temperature collected
over kilometric size pixels which include different land units. Both dimensions are
strongly correlated for current TIR spaceborne sensors: high temporal samplings for
finer monitoring correspond to coarse spatial resolutions with larger heterogeneity
effects, and reversely.
The highest temporal samplings are provided by geostationary sensors: 15–30
min with GOES Imager [81] and MSG/SEVIRI [82], corresponding to ground reso-
lutions between 2 and 4 km. Intermediate scales correspond to kilometric resolution
sensors onboard sun-synchronous platforms, providing daily nighttime and daytime
observations: NOAA/AVHRR [83], ADEOS/GLI [84], and Terra-Aqua/MODIS
[85]. A 3 day temporal sampling with a 1 km resolution has been provided by
ERS/ATSR-1 and -2, and ENVISAT/AATSR [86]. The highest spatial resolutions
are 60 and 120 m from Landsat/TM & ETM [87], and 90 m from Terra/ASTER
[88]; with 16-day temporal samplings. ASTER and Landsat/ETM missions have
limited lifetimes, with currently no follow on TIR high spatial resolution missions
from space.
Regarding current possibilities, new spaceborne sensors are demanded, to moni-
tor land processes with adequate temporal and spatial scales. Past missions IRSUTE
and SEXTET proposed 40–60 m spatial resolutions with a 1-day revisit [89, 90] and
SPECTRA proposed 50 m with 3 days [91]. MTI mission offers a 20 m resolution
with a 7-day revisit [92], but the military context restricts the data access. Airborne
prospective observations have allowed studying temporal and spatial issues, with
metric resolutions and adjustable revisits. Let us cite the airborne missions TIMS
[93], DAIS [94], MAS [95] and MASTER [96]; and the airborne-based ReSeDA
program [97–99].
250 F. Jacob et al.

Table 10.1 Nominal characteristics for operational and prospective sensors; in relation with recent,
current and forthcoming US and EU missions. VZA means View Zenith Angle, VAA means View
Azimuth Angle. Across (respectively along) track means viewing directions in a plan perpendicular
(respectively parallel) to the satellite path.

Sensor Daytime Spatial Spectral Directional


sampling resolution features features

Spaceborne

MSG 1 MIR: 3.9 µm 1 latitude-


15 mn 3 km
SEVIRI 5 TIR: 8.7, 9.7, 10.8, 12, 13.4 µm dependent VZA
GOES 10 and 12 1 MIR: 3.7 µm 1 latitude-
30 mn 2–4 km
Imager 2 TIR: 10.8, 12 µm dependent VZA
NOAA 15–17 1 MIR: 3.8 µm Across track
1 day 1 km
AVHRR / 3 2 TIR: 11, 12 µm VZA: ±55◦
Terra-Aqua 3 MIR: 3.8, 3.95, 4.1 µm Across track
1 day 1 km
MODIS 3 TIR: 8.6, 11, 12 µm VZA: ±55◦
ADEOS 1 MIR: 3.7 µm Across track
1 day 1 km
GLI 3 TIR: 8.6, 10.8, 12 µm VZA: ±40◦
ERS-ATSR 1 and 2 1 MIR: 3.7 µm Along track
3 days 1–2 km
ENVISAT-AATSR 2 TIR: 10.8, 12 µm VZA: 0, 55◦
Landsat 5–7 Close nadir
16 days 120 m 1 TIR: 11.5 µm
TM and ETM VZA
Terra Close nadir
16 days 90 m 5 TIR: 8.3, 8.6, 9.1, 10.7, 11.3 µm
ASTER VZA

Airborne

TIMS 6 TIR: 8.4, 8.8, 9.2, . . . Across track


- 1–5 m
(multispectral) . . . 9.9, 10.7, 11.7 µm VZA: ±38◦
DAIS 6 TIR: 8.7, 9.7, 10.5, . . . Across track
- 1–5 m
(multispectral) . . . 11.4, 12.0, 12.7 µm VZA: ±26◦
MAS / MASTER 10 TIR: 7.8, 8.2, 8.6, 9.1, 9.7, . . . Across track
- 1–5 m
(multispectral) . . . 10.1, 10.6, 11.3, 12.1, 12.9 µm VZA: ±40◦
MIR: [2.5–5.3] µm
SEBASS Close nadir
- 1–5 m TIR: [7.6–13.5] µm
(hyperspectral) VZA
Spectral resolution > 0.1 µm

Ground-based

Two temperature
- ∼50 cm 1 broadband over [8–13] µm Nadir VZA
Box method
Hyperspectral FTIR Optical spectral range: [2–20] µm Nadir VZA
- Few cm
BOMEM suite Spectral resolution: 1 cm−1
Goniometric VZA ∈ [0–90◦ ]
- Few cm 1 broadband over [8–13] µm
systems VAA ∈ [0–360◦ ]
10 Modeling and Inversion in Thermal Infrared Remote Sensing 251

10.3.2 Spectral and Directional Capabilities

The spectral dimension corresponds to the number and location of sensor wave-
bands within the TIR and optionally the MIR domains. The directional dimension
corresponds to the number and angular distribution of viewing directions. Both di-
mensions are used for recovering emissivities and temperatures via modeling tools.
The basic spectral configuration corresponds to TM and ETM, with 1 channel.
Richer information is provided via two channels with GOES Imager, AVHRR and
the ATSR suite; three channels with MODIS and GLI; and five channels with
SEVIRI and ASTER. Additional MIR information can be combined with TIR
information, to be used with continuous observations from geostationary sensors
(SEVIRI, GOES Imager), or day night observations from sun-synchronous sensors
(AVHRR, MODIS).
The basic directional configuration corresponds to SEVIRI, GOES Imager, TM,
ETM, and ASTER; with a single viewing direction. Richer information is collected
from across track viewing with AVHRR, MODIS, GLI; and along track viewing
with the ATSR suite. Across track viewing allows a daily monitoring, while sam-
pling the angular dynamic within a given temporal window (16 days for MODIS).
This is of interest for stable surface properties such as emissivity. For surface tem-
perature which fluctuates, capturing the angular dynamic requires almost simulta-
neous observations. This is possible with ATSR along track bi-angular observations
only, which is limited.
Future spaceborne missions will pursue current ones for long-term records: the
GOES suite [100], NPOESS/VIIRS following AVHRR and MODIS [101]. MTI
provides original information: 2 MIR/3 TIR bands, 0◦ and 50◦ along track. At
the airborne level, the spectral dimension has been investigated with multispectral
(TIMS, DAIS, MAS & MASTER) and hyperspectral (SEBASS [102]) sensors, and
the directional dimension has been assessed with video cameras (see [103] with the
ReSeDA program). At the ground level, the spectral dimension has been explored
with hyperspectral sensors (FTIR BOMEM [104]), or with broadband radiometers
[105–107], and the directional dimension has been examined with goniometric sys-
tems [58, 108, 109].
In the context of monitoring land processes, the various types of information
presented here are valuable for recovering land surface emissivity and temperature.
Using this information requires designing modeling tools and inversion methods,
either under development for prospective studies or with operational capabilities.

10.4 Developing Modeling Tools and Inversion Methods

Modeling tools aim at forwardly simulating, with different complexities, measured


brightness temperature from emissivities and temperatures of interest. Table 10.2
provides an overview of the modeling tools currently used. Based on TIR fundamen-
tals (Section 10.4.1), simple radiative transfer equations directly link measurements
252 F. Jacob et al.

Table 10.2 Listing of the modeling tools currently used, with an increasing complexity. The sec-
ond rightmost column gives the related medium, and the rightmost column gives the types of land
surface emissivity and temperature currently investigated with each tool. The modeling of at-
mospheric radiative transfer is considered here in the context of performing atmospheric correc-
tions.

Modeling Literature Related Investigated land surface


tools examples medium temperatures and emissivities

Simple radiative transfer equations

Atmospheric Eq. 10.6 • Brightness temperature


Atmosphere
radiative transfer [81, 178] (Atmospheric corrections)
Composite surface Eq. 10.7 • Ensemble emissivity and
Land surface
radiative transfer [12, 158] radiometric temperature
Split Window and Eq. 10.9 Atmosphere and • Ensemble radiometric temperature
Dual Angle [125, 126] land surface (atmospheric corrections)
Soil and vegetation Eq. 10.10
Land surface • Soil and vegetation temperatures
radiative transfer [68, 127]
Kernel-driven Eq. 10.11 • Ensemble emissivity
Land surface
radiative transfer [128, 129] • Soil and vegetation temperatures

Simulation models

• Brightness temperature
Radiative transfer MODTRAN [134] Atmosphere
(Atmospheric corrections)
• Brightness temperature
Prevot’s [139] • Ensemble emissivity
Radiative transfer Land surface
SAIL [74, 137] • Soil and vegetation temperatures
with sunlit and shaded components
Kimes’s [141] • Brightness temperature
Geometric-optics Caselles’s [143] Land surface • Soil and vegetation temperatures
Yu’s [73] with sunlit and shaded components
CUPID [147]
Geometric-optics Thermo [148] • Brightness temperature
Land surface
radiative transfer Jia’s [149] • Canopy temperature profile
DART [76, 77]
• Brightness temperature
Monte Carlo
[127, 150, 151] Land surface • Ensemble emissivity
ray tracing
• Soil and vegetation emissivities

to emissivities and temperatures of interest (Section 10.4.2), and simulation models


describe the influence of radiative regime on measurements (Section 10.4.3). Next,
inversion methods aim at backwardly retrieving emissivities and temperatures of
interest from measurements (Section 10.4.4).
10 Modeling and Inversion in Thermal Infrared Remote Sensing 253

10.4.1 Fundamentals in TIR Remote Sensing

The use of TIR remote sensing to infer the temperatures of interest involves an
aerodynamic issue for the related temperature, and a radiative issue for the other
temperatures.

10.4.1.1 Aerodynamic Issue

Aerodynamic temperature Taero is not radiative-based and cannot be remotely


sensed. It is required for one source modeling of surface energy fluxes, since it
corresponds to the value of the logarithmic-based air temperature profile Tair (z) at
thermal roughness length zoh [110]. For a negligible displacement height, sensible
heat flux H is expressed from the air temperature gradient between zoh and reference
level zre f :

Tair (zoh ) − Tair (zre f )


H= with Taero = Tair (zoh ) (10.1)
rah (zoh , zre f )

where rah (zoh , zre f ) is aerodynamic resistance for heat between zoh and zre f [111].
Due to larger resistance for heat transfers, zoh is lower than mechanical roughness
length zom [112]. The link between both is the aerodynamic kB−1 parameter [113]:

zom
kB−1 = ln (10.2)
zoh

The physical meanings of Taero and zoh are equivocal. Taero is an effective tempera-
ture for heat sources that are soil and vegetation [114]. zoh is an effective level for
which Tair = Taero . Their retrieval from remote sensing is not trivial (Section 10.6.3).
Nevertheless, Taero can be unequivocally derived from soil and vegetation tempera-
tures Tsoil and Tveg , by merging one source and two source modeling [20, 115]:

Tsoil T
veg Tair (zre f )
ra,soil + ra,veg + rah
Taero = 1 1 1
(10.3)
ra,soil + ra,veg + rah

where ra,soil (respectively ra,veg ) is aerodynamic resistance from the soil (respec-
tively vegetation) to zom , and rah is aerodynamic resistance from zom to zre f [111].

10.4.1.2 Radiative Issue

Apart from aerodynamic temperature, the land surface temperatures inferred from
TIR remote sensing are radiative-based. Then, fundamentals deal with the TIR
radiative regime within atmosphere and over land surfaces. This includes three
mechanisms which drive the wave matter interactions: emission, absorption, and
254 F. Jacob et al.

scattering. Emitted radiance L(λ , θ , T ) from a natural object at a kinetic tempera-


ture T is written:
C1 λ −5
L(λ , θ , T ) = ε (λ , θ ) B(λ , T ) = ε (λ , θ )     (10.4)
π exp TC2λ − 1

λ is the monochromatic wavelength. θ is the emission direction. B(λ , T ) is the


blackbody emitted radiance, expressed from Planck’s Law. C1 and C2 are first and
second radiative constants. Emissivity ε (λ , θ ) is the conversion factor from ther-
modynamic to radiative energy, lower than 1 for natural objects. This so-called
e-emissivity definition is linked to emission mechanisms, since it is the ratio of the
actual to the blackbody emitted radiances for the same kinetic temperature. Under
local thermodynamic equilibrium, Kirchhoff’s Law assumes emissivity and absorp-
tivity are equal. For opaque elements, emissivity is then linked to hemispherical-
directional reflectance ρ (λ j , θ ):

ε (λ j , θ ) = 1 − ρ (λ j , θ ) (10.5)

ρ (λ j , θ ) is the average of bidirectional reflectance over illumination angles [116].


This so-called r-emissivity definition is derived from Kirchhoff’s Law, and therefore
linked to reflection mechanisms. Finally, emitted radiance from a given element can
be reflected by other elements, inducing changes in radiation path, called scattering
effects.
Within the atmosphere, scattering is negligible: the radiative regime is driven by
the temperature and density of absorbers and emitters (water vapor, CO2 , O3 , . . .).
A clear atmosphere behaves as an horizontally homogeneous medium: the radiative
regime primarily depends on vertical profiles for temperature and density of ab-
sorbers and emitters (Fig. 10.1). Over heterogeneous land surfaces with structured
patterns, the radiative regime is more complex than within atmosphere: soil and
vegetation act as emitters, absorbers and scatterers for canopy and atmospheric irra-
diances. Additional effects are surface and volume scatterings (Fig. 10.2). Surface
scattering corresponds to shadowing effects for a geometric medium, with sunlit
and shaded areas. Volume scattering corresponds to reflections between soil and
vegetation: radiation is trapped within the canopy.
TIR remotely sensed measurements result from the processes discussed above.
Sensor brightness temperature is driven by vertical profiles for temperatures and
densities of atmospheric constituents. Surface brightness temperature results from
the radiative regime over a heterogeneous and non isothermal area. Then, emissivity
and kinetic temperature are equivocal: the canopy acts as an effective medium with
ensemble emissivity and radiometric temperature [50]. Besides, e- and r-emissivities
differ according to vegetation amount, since spatial averaging for e-emissivity in-
cludes emitted radiance as an additional weighting factor [50, 117]. Due to its
simpler formulation, r-emissivity is preferred [68, 71, 74, 118–120]. Further, the
measured brightness temperature results from emission, but also from absorption
and scattering of canopy and atmospheric irradiances. This induces spectral and
directional variations, driven by (1) radiative properties of soil and vegetation
10 Modeling and Inversion in Thermal Infrared Remote Sensing 255

Fig. 10.1 Atmospheric TIR radiative regime for an off nadir propagation. The key processes to be
considered for atmospheric corrections are emission and absorption by atmospheric constituents.
Within a horizontally homogeneous atmosphere, the radiative regime depends on the vertical fields
of temperature and density for emitters and absorbers. Regardless of considered layer (zi or zk ),
radiative regime is driven by atmospheric absorption (1), atmospheric emission (2), and surface
emission through atmosphere transmission (3). (Adapted from [264].)

(reflectance and emissivity), (2) surface scattering with sunlit and shaded areas, and
(3) volume scattering with the cavity effect. These three factors induce ensemble
emissivity is anisotropic, with values greater than that of vegetation as the latter
quantitatively increases [118, 121, 122].
Various modeling tools have been developed to simulate sensor and surface
brightness temperature measurements. The first way is using simple radiative trans-
fer equations for directly linking measurements to emissivities and temperatures of
interest. The second way is using simulation models for understanding the influence
of the TIR radiative regime on the measured brightness temperature.

10.4.2 Simple Radiative Transfer Equations

Simple radiative transfer equations directly link TIR measurements to emissivities


and temperatures of interest. Their advantages are linearity and simplicity, but most
of them are limited to homogeneous media by assuming turbidity and azimuthal
isotropy.
256 F. Jacob et al.

Fig. 10.2 Surface (or geometric) and volume (or volumetric) scattering. Surface scattering in-
duces shadowing effects with hotter and cooler elements. Volume scattering induces an increase of
brightness temperature by adding a component to emission. (Adapted from [129].)

Measured brightness temperature at the sensor level Tbrs is linked to surface


brightness temperature Tbs via the atmospheric radiative transfer equation:
 
B (λ j , Tbrs (θ , λ j )) = B (λ j , Tbs (θ , λ j )) τa (θ , λ j ) + B λ j , Tba↑ (θ , λ j ) (10.6)

θ is the view zenith angle. λ j is the equivalent waveband over the sensor channel
j [123]. B(λ , T ) is the blackbody emitted radiance, expressed from Planck’s Law
(Eq. 10.4). τa is the atmospheric transmittance, vertically integrated between the
surface and the sensor. B(λ j , Tba↑ ) is the atmospheric upward radiance towards the
sensor.
Surface brightness temperature is expressed as the sum of canopy emission and
scattering of atmospheric irradiance, via the composite surface radiative transfer
equation:
10 Modeling and Inversion in Thermal Infrared Remote Sensing 257

B(λ j , Tbs (θ , λ j )) = ε (λ j , θ ) B(λ j , Trad (θ )) + (1 − ε (λ j , θ )) B(λ j , Tba↓ (λ j )) (10.7)

B(λ j , Tba↓ ) is the hemispherical average of atmospheric downward radiance. ε (λ j , θ )


and Trad (θ ) are ensemble emissivity and radiometric temperature. Ensemble emis-
sivity can be expressed from emissivities of soil εsoil (λ j ) and vegetation εveg (λ j ),
with the optional inclusion of a correction term d ε for the cavity effect [124]:

ε (λ j , θ ) = Fsoil (θ ) εsoil (λ j ) + Fveg (θ ) εveg (λ j ) + 4 d ε Fveg (θ ) Fsoil (θ ) (10.8)

Fsoil (θ ), Fveg (θ ) are directional gap and cover fractions, with Fsoil (θ ) = 1 − Fveg (θ ).
Brightness temperature measured from space can be linked to emissivity and
radiometric temperature by merging Eqs. 10.6 and 10.7. Another possibility is si-
multaneously considering atmospheric and surface effects: Split Window (SW) and
Dual Angle (DA) methods directly express radiometric temperature Trad as a spec-
tral or angular difference between two brightness temperatures Tbrs at the sensor
level [125, 126]:
ε 1 + ε2
rs
Trad = Tb1 rs
+ A(Tb1 − Tb2
rs rs
) + B(Tb1 − Tb2
rs 2
) +C + D(ε1 − ε2 ) + E (10.9)
2
ε is surface emissivity. A, B, C, D, E are empirical coefficients. Indices 1 and 2 are
two spectral channels for SW method, or two view zenith angles for DA method.
The angular differencing uses variations in atmospheric transmittance between dif-
ferent paths for two view zenith angles. The spectral differencing uses variations
in atmospheric transmittance due to different water vapor absorptions for two spec-
trally close channels.
The emission term of Eq. 10.7 can be split into soil and vegetation components,
which yields the soil and vegetation radiative transfer equation [68, 119, 127]:

ε (λ j , θ ) B (λ j , Trad (θ )) = τ can (θ ) εsoil (λ j ) B(λ j , Tsoil )


+ ω (θ , εveg (λ j )) B(λ j , Tveg ) (10.10)

Tsoil and Tveg are soil and vegetation radiometric temperatures [68]. τ can (θ ) and
ω (θ , εveg (λ j )) are vegetation directional transmittance and fraction of emitted ra-
diation. The angular effects can also be described with linear kernel driven ap-
proaches, by expressing the directional emission as a linear combination of generic
shapes [128]:

ε (λ j , θ ) B (λ j , Trad (θ )) =
N
∑ βi (λ j ) Ki (Tveg , Tsoil , εveg (λ j ), εsoil (λ j ), θ , θs , ϕ − ϕs ) (10.11)
i=1

θs is the solar zenith angle. ϕ − ϕs is the relative azimuth between illumination


and viewing directions. βi,λ j are weighting coefficients. Kernels Ki describe gray
body isotropy, volume scattering, and surface scattering. Various kernel formula-
tions may be proposed, by linearizing different sets of complex equations. Kernel
driven approaches are also used to derive ensemble r-emissivity from accurate
258 F. Jacob et al.

hemispherical - directional reflectance (Eq. 10.5): [129] expressed TIR BRDF as


a linear combination of generic shapes, following previous works over the solar
domain [130–132].

10.4.3 Simulation Models

Simulation models mimic the TIR radiative regime within atmosphere and canopies,
to understand spatial, spectral and directional behaviors of brightness temperature
measurements. These models are classified here via an increasing complexity: radia-
tive transfer, geometric-optics, geometric-optics/radiative transfer, and ray tracing.
Radiative transfer models are designed for turbid media (atmosphere, homoge-
neous canopies). Assuming turbidity and azimuthal isotropy, they split the medium
into a finite layer number, and account for volume scattering between layers. For the
atmosphere, volume scattering is negligible, and each layer is described with tem-
perature and densities of absorbers and emitters. For canopies, soil and vegetation
layers are described with temperature; and with densities of absorbers, emitters and
scatterers, derived from LAI and LIDF. Brightness temperature is simulated using
the stream concept: transmittance, upward and downward radiances are computed
for each layer, and vertically integrated (see MODTRAN for atmosphere [133, 134]
and SAIL for canopy [68, 74, 135–137]. Simulations can also be probabilistic cal-
culations for photon interception, deduced from the directional gap fraction of each
layer [118, 138, 139].
Geometric-optics models are designed for structured patterns over land surfaces,
such as row crops of cotton or maize. Considering vegetation as an opaque medium,
they account for surface scattering with shadowing effects. Sunlit and shaded areas
are described via their cross sections, derived from canopy geometry (vegetation
height, row size, etc.), illumination and viewing directions, and directional gap frac-
tion within and between rows. Canopy brightness temperature is computed from the
resulting spatial distribution of temperature [73, 121,140–143].
The finest radiosity models are geometric-optics/radiative transfer models,
designed for complex land surfaces. By accounting for both volume and surface
scattering, they are appropriate to vegetation patchworks. They can conjugate a
radiative transfer and a geometric-optic module [144, 145]. They can be more com-
plex, such as 3-Dimensional mock-ups based models. This allows a finer description
of the radiative regime within canopies, but requires significant information about
the micro-scale conditions. Examples are CUPID [146, 147]; Thermo [72, 148];
Jia’s model [149], and DART [76, 77]. Further, accounting for convective and ener-
getic transfers allow understanding their influence on the radiative regime, such as
with DART-EB [78].
The finest modeling degree is Monte Carlo ray tracing, which stochastically cal-
culates photon trajectories within turbid or geometric atmosphere and canopies.
A photon is tracked from birth (emission or penetration within medium) to death
(absorption or escape from medium), with scattering based on probabilistic wave
10 Modeling and Inversion in Thermal Infrared Remote Sensing 259

matter interactions. Millions of simulations describe spectral, directional and spa-


tial behaviors. Ray tracing is used to assess the influence of multiple scattering on
spatial aggregation and angular dynamics, over heterogeneous and non isothermal
land surfaces [127, 150, 151].

10.4.4 Inversion Methods

Retrieving variables from measurements is an inverse problem. Given a set of m


measurements M for a physical system, with k known parameters K and p unknown
parameters P to be retrieved, direct F and inverse F −1 problems are written [152]:
⎡ ⎤ ⎛⎡ ⎤ ⎞ ⎡ ⎤ ⎛⎡ ⎤ ⎞
M1 P1 P1 M1
⎢ .. ⎥ ⎜⎢ . ⎥ ⎟ ⎢ . ⎥ −1 ⎜⎢ . ⎥ ⎟
⎣ . ⎦ = F ⎝⎣ .. ⎦ , [K1 · · · Kk ]⎠ ⇐⇒ ⎣ .. ⎦ = F ⎝⎣ .. ⎦ , [K1 · · · Kk ]⎠(10.12)
Mm Pp Pp Mm

Inversion is possible if there are more independent equations than unknowns


(m ≥ p). Direct inversion analytically writes the inverse problem. This is possible
for simple radiative transfer equations (Section 10.4.2), but not for most simulation
models (Section 10.4.3). For the latter, indirect inversion numerically sets parame-
ters such as simulations agree with observations [153]. It has been improved for
accuracy and rapidity, by calibrating neural networks, lookup tables, genetic algo-
rithms or regression trees [152, 154]. Inversion can be a well-posed problem, when
solving an overdetermined equation system using optimization techniques. How-
ever, it is usually an ill-posed problem, with several parameters to be constrained
from few observations. Proposed solutions use a priori information about soil and
vegetation properties, or parameter ranges [152, 155, 156].
Inversion over the TIR domain is not as developed as over the solar domain. This
results from (1) additional micrometeorological complex influences, and (2) the lack
of high resolution data. Atmospheric simulation models have been inverted cali-
brating neural networks [51], SW and DA methods (Eq. 10.9) [125, 126], or the
atmospheric radiative transfer equation (Eq. 10.6) [157, 158]. Over land surfaces,
simulation models have been assessed in the forward mode [72–74, 77, 144]. No in-
vestigation was found about their indirect inversion, but they can serve as references
for parameterizing simple radiative transfer equations which are directly invertible.
Thus, various formulations have been assessed for the soil and vegetation radiative
transfer equation (Eq. 10.10), optionally accounting for multiple scattering and non
linearities [68, 71, 118, 119, 127]. Further, inverting simple radiative transfer equa-
tions is often an ill-posed problem. For instance, inverting the composite surface
radiative transfer equation (Eq. 10.7) from N multispectral observations includes
N emissivities and radiometric temperature. Similarly, inverting the soil and veg-
etation radiative transfer equation (Eq. 10.10) or linear kernel driven approaches
(Eq. 10.11) from multiangular observations requires angular parameters: ensemble,
soil and vegetation emissivities; vegetation transmittance.
260 F. Jacob et al.

10.5 Assessing Modeling Tools and Inversion Methods

Modeling tools and inversion methods have been assessed experimentally through
validation exercises, and theoretically via sensitivity studies.
Validation exercises have been conducted over databases collected in the frame-
work of various international programs such as FIFE [159], EFEDA [160], HAPEX
[161], ReSeDA [97], JORNEX [162], FLUXNET [163], DAISEX [164], SALSA
[165], SMACEX [166]. Assessments over these various datasets allow accounting
for different biomes and climates. Some exercises were ground-based [73, 104,
145]. Most of them were airborne-based [94, 96, 103, 157, 167–169,170–174], for
assessments in actual conditions by reducing spatial heterogeneity effects. Few vali-
dations were conducted using spaceborne observations with hectometric resolutions
[175–179]; and with kilometric ones over areas almost homogeneous [180–182].
Original exercises based on classifications were designed for kilometric scale het-
erogeneities [126, 183], while new improvements for the solar domain should be
implemented over the thermal one [184]. Complementary to validations, intercom-
parisons are now feasible thanks to multisensor missions such as Terra. This allows
accounting for larger panels of environmental situations [158].
Validations and intercomparisons have also been performed using simulated
datasets. This allow considering more conditions than measured datasets, and focus-
ing on physics modeling without measurement intrinsic errors [81, 82, 126, 185].
Simulated datasets are necessary when dealing with elaborated temperatures: aero-
dynamic, soil and vegetation, sunlit and shaded components, and canopy temper-
ature profile [68, 71, 73, 76, 118]. Indeed, validating the latter using measured
datasets is not trivial, since the corresponding ground-based measurements are dif-
ficult to implement.
Additionally to validations and intercomparisons, sensitivity studies allow as-
sessing information requirements such as accuracies on remotely sensed informa-
tion, medium structural and radiative properties. Examples are (1) accuracy on
atmospheric status for retrieving brightness temperature [171, 178], (2) accuracy
on observations, atmospheric status and land use for recovering ensemble emissiv-
ity and radiometric temperature [12, 157, 158, 169, 182, 186–188], (3) accuracy on
canopy structural parameters and radiative properties for deriving soil and vegeta-
tion temperatures [68, 118, 189]. Finally, sensitivity studies of simulation models
provide valuable information about the pertinent parameters for inversion [73, 76],
with innovative approaches over the solar domain based on adjoint models (Baret
et al., this issue).

10.6 Current Capabilities and Future Directions

From the basic materials presented before, we focus now on current investigations,
via an increasing temperature complexity. Success and failures suggest future direc-
tions.
10 Modeling and Inversion in Thermal Infrared Remote Sensing 261

10.6.1 Surface Brightness Temperature

Surface brightness temperature is derived from that at the sensor level by inverting
modeling tools for atmosphere. It is simulated using modeling tools for land sur-
faces. In both cases, these tools are simple radiative transfer equations or simulation
models.

10.6.1.1 Atmospheric Radiative Regime and Related Corrections

Atmospheric corrections for the retrieval of surface brightness temperature can be


performed inverting simulation models, via the calibration of the atmospheric ra-
diative transfer equation (Eq. 10.6) for a given atmosphere [12, 81, 157, 158, 171,
172, 177, 178]. An operational context faces two challenges: reducing computation
time to process millions observations, and accurately characterizing the atmospheric
status.
To reduce by a third-order computation time for simulation models without accu-
racy degradation, [190] implemented correlated-K methods, by quickly integrating
waveband atmospheric absorption and emission. Predictor-based models accurately
compute the latter for a range of reference profiles, to next differencing current ones
and nearest predictors [191]. Multilayer computation based on water vapor contin-
uum absorption can replace simulation models, with an accuracy degradation lower
than 1 K [81]. Computation time can also be reduced via inversion by including
a range of atmospheres into the simulation set. Expressing transmittance and up-
welling radiance of Eq. 10.6 from atmosphere water vapor content and mean tem-
perature yields an accuracy degradation lower than 2 K [180]. Neural networks can
replace Eq. 10.6 considering atmospheric profiles and view zenith angle, with an
accuracy degradation lower than 0.5 K [186, 192].
The atmospheric status can be well documented using ancillary information:
measured profiles allow reaching a 1 K accuracy [171, 172, 177, 178], but mete-
orological networks are not dense enough for regional inversion. One alternative
is profile simulation from meteorological models [193, 194]. Such information is
soon available with a 3 h sampling, and a 0.25◦ latitude/longitude griding to be
re-sampled to sensor resolutions via interpolation procedures [12, 195]. The relief
influence is handled using digital elevation models, now available with decametric
resolutions and metric accuracies [196]. Also, the TIR observations to be corrected
can inform about the atmospheric status. Atmosphere absorption and emission can
be retrieved from multispectral and hyperspectral observations, using variabilities
of atmospheric properties [80, 197]. Thus, water vapor content was adjusted from
ASTER multispectral observations, such as emissivity spectrum is flat over vege-
tation or water [185]. It was also inferred from the ATSR-2 SW channels with a
0.2 g. cm−2 accuracy, using the SWVCR which relies on the spatial variability of
SW surface brightness temperatures [198].
Solar or TIR observations collected onboard the same platform also provide
coincident information about the atmospheric status. [199] expressed water vapor
262 F. Jacob et al.

content as a polynomial of MODIS near infrared radiance ratios, with a 0.4 g. cm−2
accuracy. Atmospheric sounders allow inferring profiles of temperature and water
vapor density, using Eq. 10.6 or neural networks [3, 4]. Previous sounders such as
TOVS permitted to reach a 0.4 g. cm−2 accuracy on water vapor content [200]. New
sounders such as IASI [201], with finer spectral samplings and spatial resolutions,
should provide accuracies better than 1 K and 10% for atmospheric profiles of tem-
perature and humidity.

10.6.1.2 Land Surface Radiative Regime and Related Measurements

Surface brightness temperature is simulated using simple radiative transfer equa-


tions or simulation models. The former provide easy and efficient solutions for as-
similating TIR remote sensing data into land process models. The latter are fine and
accurate solutions for understanding TIR remotely sensed measurements.
To constrain land process model parameters, surface brightness temperature can
be simulated using simple radiative transfer equations coupled with SVAT models.
[39] coupled the composite surface radiative transfer equation (Eq. 10.7) with a crop
and a one source SVAT model. The latter calculated ensemble radiometric temper-
ature by closing the surface energy budget. R-emissivity was estimated using the
SAIL TIR version of [136], documented by the crop model for vegetation struc-
tural parameters. Similarly, [67] coupled the soil and vegetation radiative transfer
equation (Eq. 10.10) with a two source SVAT model. The latter calculated soil and
vegetation temperatures by closing the energy budget for each, while setting soil
and vegetation emissivities to nominal values.
Calculating surface brightness temperature from simulation models requires in-
formation about vegetation structure (row crop, LAI, LIDF, cover fraction), soil
and vegetation radiative properties (emissivity, reflectance), and thermal regime
(canopy temperature distribution). The latter can be derived from a SVAT model,
which solves local energy budget according to meteorological conditions (solar
position, wind speed, air temperature), vegetation status (leaf stomatal resistance),
and soil moisture. Then, simulation models mimic the radiative regime using more
or less complex descriptions of the thermal regime: a unique vegetation tempera-
ture [73], soil and vegetation temperatures with optional sunlit and shaded compo-
nents [74, 137], additional vegetation layer temperatures for specific crops [145], or
canopy temperature profiles [78].
Simulation models are currently under development, verification and analysis
[73, 74, 76, 78, 144, 145]. Current investigations focus on spectral behaviors [120],
but especially on directional effects which allow normalizing multiangular obser-
vations (Fig. 10.3). For instance, [58, 72] angularly normalized water stress indices
over row structured crops. Similarly, [202] normalized across track observations
from sun-view geometry effects, for a daily monitoring at the continental scale.
10 Modeling and Inversion in Thermal Infrared Remote Sensing 263

0
80
330 30

60 34

300 40 60 33

20 32

31
270 90

30

29
240 120

28

210 150

180

Fig. 10.3 Simulating measured brightness temperature over a maize canopy in row structure, with
a resulting angular dynamic about 8 K [73]. Black star indicates the solar direction. The brightness
temperature maximum value is located in the solar direction. However, this hot spot effect is not
systematical (Section 10.6.4.)

10.6.1.3 Partial Conclusions

The various methods developed to perform atmospheric corrections are of inter-


est, since they were designed for optimizing the collected information according
to sensor configurations. Measured or simulated profiles are tributary to their rep-
resentativeness, and coincident information relies on strong assumptions. Despite
these limitations, significant progresses were made the last decades, with accura-
cies now close to 1 K. Current investigations focus on refinements rather than new
developments.
Simulating brightness temperature is ongoing for describing brightness tempera-
ture measurements, according to the various land surface behaviors: geometric like,
radiative transfer like, or both. Validation results emphasized good performances
with accuracies close to 1 K, though significant documentations are required about
thermal regime, medium structure and radiative properties. Such simulation models
will be of interest for future designs of inversion methods, conjointly to the solar
domain (Section 10.4.4).
264 F. Jacob et al.

10.6.2 Ensemble Emissivity and Radiometric Temperature

Ensemble radiometric temperature is derived by directly inverting composite surface


radiative transfer equation (Eq. 10.7), or indirectly inverting simulation models via
differencing equations (Eq. 10.9). The first way is two-step based and requires pre-
vious atmospheric corrections. The second way is one-step based by simultaneously
correcting atmosphere and surface effects. In both cases, performances depend on
characterizing these effects. Inverting Eq. 10.7 is an ill-posed problem, with N equa-
tions from channel