Letter
Geomorphological mapping using object-based analysis and ASTER
DEM in the Paraba do Sul Valley, Brazil
1.
Introduction
Geomorphological maps are important tools for natural resources assessment, urban
and environmental planning and management, as well as the forecast and control of
natural disasters. In the particular case of Brazil, there is a lack of such maps at scales
greater than 1:1000000. The few available geomorphological maps were produced by
the RADAMBRASIL project survey (DNPM 1983), designed to assess natural
resources in the Brazilian Amazon by means of side-looking airborne radar (SLAR)
during the 1970s and 1980s. This scarcity of mapping products can be overcome by
recent advances in remote sensing and digital images processing, which enable both
the speed-up of mapping tasks and the extraction of progressively finer levels of detail.
In this context, expert systems are particularly important because they employ
advanced strategies for knowledge representation, such as semantic networks and
object-based approaches and facilitate the use of fuzzy logic to handle uncertainties
(Moore 2000, Moore et al. 2003). Definiens, an expert system for the classification of
remote sensing images, adopts such strategies (Benz et al. 2004), and has been
employed in some initiatives for semi-automatic geomorphological mapping
(Asselen and Seijmonsbergen 2006, Schneevoigt et al. 2008). This letter aims at
developing a methodology for semi-automatic geomorphological mapping using a
three-level hierarchical semantic network. In contrast to previous articles on this
matter, this work uses texture variables and an ASTER-derived DEM.
*Corresponding author. Email: almeida@dsr.inpe.br
International Journal of Remote Sensing
ISSN 0143-1161 print/ISSN 1366-5901 online # 2009 Taylor & Francis
http://www.tandf.co.uk/journals
DOI: 10.1080/01431160903111044
6614
F. F. Camargo et al.
2.
Methodology
2.1
Study area
The municipality of Sao Jose dos Campos, with a total area of 1098.6 km2, is located in
the Paraba do Sul Valley, Sao Paulo State, southeastern Brazil (figure 1). This valley is
situated in a Precambrian mountain range, which is the most prominent relief feature in
eastern South America. The lithology of this region is composed of gneisses (Archaen/
mid-Proterozoic), sin- and post-tectonic granite suites (Upper Proterozoic), sedimentary rocks (late Tertiary) and quaternary deposits. The geomorphological units consist
of ridges, mountains, alluvial plains, terraces and sedimentary hills (DNPM 1983).
2.
The data used comprised ASTER/Terra visible and near infrared sensor subsystem
(VNIR) images (bands 3N and 3B of 31 August 2004, with a nominal spatial resolution
of 15 m; processing level: L1B; WRS-2 path: 219; WRS-2 row: 076; ID:
AST_L1B.003:2071432223); 109 global positioning system (GPS) points (with orthometric heights); 1:10 000 vector files of the street network and drainage; and 10 m
interval contour lines and scattered elevation data points. Initially, the images were
oriented using Toutins model (Toutin 2004), available in the software Geomatica and
ground control points extracted from the vector files. Epipolar images were then
generated and a DEM (with 15 m spatial resolution) was obtained by means of stereocorrelation and parallax estimation. The 109 GPS points were supplied by FUNCATE
foundation, INPE (National Institute for Space Research) and IEAv (Institute for
Advanced Studies) research institutes and these points were used as the elevation
reference in the DEM accuracy validation. Two kinds of statistics were obtained
based on elevation discrepancies () between the DEM and the GPS elevations: root
0 or H1 : 0,
mean square error (RMSE) and Students t test (H0 :
where H0 is
Figure 1. Study area: (a) Brazil and Sao Paulo State (in black) and (b) shaded relief of the
municipality of Sao Jose dos Campos.
6615
Multiresolution segmentation
As previously stated, three levels of segmentation were generated: level 1 contains the
water bodies, level 2 accounts for the geomorphological units and level 3 represents the
drainage density. The drainage density was obtained from the ratio of the area of water
streams to the area of each segment at level 3. The area of streams was estimated from the
rasterized drainage vector file in binary form. Due to the interdependence of segment
(object) boundaries between levels, levels 1 and 3 were initially generated in a separate
database (project) since a very small scale factor (segmentation parameter to define the
objects size) was used in the segmentation of water bodies. In a subsequent stage, the
classification results of these two levels were imported into the final semantic network as
shape files, and hence they did not affect the segmentation of level 2. The attributes and
parameters used for segmentation (table 1) were chosen by an interactive trial and error
process. In the particular case of level 2, however, thresholds as defined in Asselen and
Seijmonsbergen (2006) were applied to the DEM. The segmentation parameters chosen
were those that could better identify the borders of the main classes and were also able to
generate objects with a suitable size for statistical treatment and exploratory analysis.
For further information on the parameter calculations, please refer to Benz et al. (2004).
2.4
Attributes
Band 3N digital
number
2. Geomorphological DEM elevation
units
3. Drainage density
Shaded relief
Scale
factor
30
0.8
0.2
0.5
0.5
50
0.7
0.3
0.3
0.7
200
0.1
0.9
0.5
0.5
6616
F. F. Camargo et al.
Table 2. Attributes used to identify the geomorphological units and
lakes/dams.
Geomorphological units
Ridges/mountains/high hills
Alluvial plains
Alluvial intermountain plains
Attributes
High drainage density (level 3)
Mean elevation from 572 to 2100 m
Mean slope from 4.9 to 34.9
Mean elevation from 560 to 688 m
Mean entropy from 1.2 to 2.2
Mean slope from 2.5 to 12
Mean entropy from -0.01 to 1.5
Mean slope from 1.4 to 10
Elevation range from 10 to 12 m
Mean elevation from 678 to 687 m
Mean slope from 2.88 to 4.75
Existence of water bodies (level 1)
6617
Alluvial plains
Subclasses
Sedimentary low hills
(correctly classified)
Commission errors: ridges/
mountains/high hills
Alluvial plains (correctly
classified)
Attributes
Relative borderb to class ridges/
mountains/high hills (01)
Relative border to class ridges/
mountains/high hills (01)
Mean difference of accumulated
flow in relation to sedimentary low
hills (8502050)a
Relative border to class ridges/
mountains/high hills (00.5)
Relative border to class ridges/
mountains/high hills (0.40.64)
Mean difference of accumulated
flow in relation to sedimentary
low hills (8502050)
Variance of horizontal
curvature (16.3517)
The mean difference of an attribute average value for the concerned image object in relation to
the mean value of the same attribute for all image objects at the same segmentation level assigned
to a defined class (Definiens 2006). In this case, the attribute refers to the accumulated flow.
b
Relative border refers to the length of the shared border of neighbouring image objects.
The validation results for the ASTER/Terra DEM obtained an RMSE of 9.38 m, and
hence are consistent with similar work (e.g. Toutin 2008). On the other hand,
Students t test null hypothesis (H0) of no difference between the DEM and GPS
elevations was rejected. This is not unexpected because the DEM is influenced by the
height of objects that extend above the surface, such as buildings and trees. Figure 3
shows the DEM, the study area boundary superimposed on it (black line), areas with
no data (black circles) and the 109 GPS points (red dots).
For the classification results, there was confusion between the classes sedimentary low
hills and alluvial plains due to the similarity of their attributes. In this way, subclasses
were used for identifying commission errors (figure 2) and improving the classification
accuracy (see 2.4). Figures 4 and 5 present the geomorphological units obtained from
visual interpretation, which were used as the reference map, and from semi-automatic
classification, respectively. The details of figure 5 show examples of errors between the
classes ridges/mountains/high hills and sedimentary low hills (upper inset box) and
between alluvial plains and sedimentary low hills (lower inset box).
6618
F. F. Camargo et al.
Figure 3. DEM, study area boundary (black line), no data areas (black circles) and GPS
points (red dots).
Figure 4.
6619
4341
44
11
00
0.99
0.99
37
622
49
00
0.88
0.88
18
00
38
00
284
00
00
02
0.83
1.00
0.84
1.00
0.964
0.890
6620
F. F. Camargo et al.
Table 4 indicates that the class sedimentary low hills was confused with both
ridges/mountains/high hills and alluvial plains. This can be ascribed to the diversity
of hill types, ranging from mildly flat to moderately dissected surfaces. On the other
hand, the confusion between ridges/mountains/high hills and alluvial plains was of
reduced extent and limited to the contacts between these units. The Kappa index value
is regarded as indicating strong agreement (Landis and Koch 1977) and the producer
and user accuracies suggest a substantial agreement between the reference map and
the classification result.
4.
Conclusions
This case study suggests the object-based method is effective for semi-automatic
geomorphological mapping. Such mapping would not have been possible in conventional digital image processing platforms, which operate almost exclusively with
statistical data and only with one level of segmentation. An outstanding advantage
of this method is the possibility of replicating the hierarchical semantic network to
other areas with similar geomorphological characteristics, once the Definiens platform allows the fine tuning of fuzzy membership functions and their respective
thresholds, so as to fit the model to the peculiarities of particular landforms.
Acknowledgements
The authors wish to thank the reviewers for their valuable contributions to this paper.
References
ASSELEN, S. and SEIJMONSBERGEN, A.C., 2006, Expert-driven semi-automated geomorphological
mapping for a mountainous area using a laser DTM. Geomorphology, 78, pp. 309320.
BENZ, U.C., HOFMANN, P., WILLHAUCK, G., LINGENFELDER, I. and HEYNEN, M., 2004, Multiresolution, object-oriented fuzzy analysis of remote sensing data for GIS-ready information. ISPRS Journal of Photogrammetry and Remote Sensing, 58, pp. 239258.
DEFINIENS, 2006, Definiens Professional 5: Reference Book (Munich, Germany: Definiens).
DNPM, 1983, Folhas SF 23/24 (Rio de Janeiro/Vitoria): Geologia, Geomorfologia, Pedologia,
Vegetacao e uso Potencial da Terra (Rio de Janeiro/RJ: DNPM).
LANDIS, J. and KOCH, G., 1977, The measurement of observer agreement for categorical data.
Biometrics, 33, pp. 159174.
MOORE, T., 2000, Geospatial expert systems. In Geocomputation, S. OPENSHAW and R.J.
Abrahart (Eds), pp. 127159 (London: Taylor & Francis).
MOORE, A.B., MORRIS, K.P., BLACKWELL, G.K., JONES, A.R. and SIMS, P.C., 2003, Using
geomorphological rules to classify photogrammetrically-derived digital elevation models. International Journal of Remote Sensing, 24, pp. 26132626.
SCHNEEVOIGT, N.N., VAN DER LINDER, S., THAMM, H.-P. and SCHROTT, L., 2008, Detecting
Alpine landforms from remotely sensed imagery. A pilot study in the Bavarian Alps.
Geomorphology, 93, pp. 104119.
TOUTIN, T., 2004, Geometric processing of remote sensing images: models, algorithms and
methods. International Journal of Remote Sensing, 25, pp. 18931924.
TOUTIN, T., 2008, ASTER DEMs for geomatic and geoscientific applications: a review.
International Journal of Remote Sensing, 29, pp. 18551875.
VERSTAPPEN, H.T. and VAN ZUIDAM, R.A., 1991, The ITC System of Geomorphologic Survey: a
Basis for the Evaluation of Natural Resources and Hazards (Enschede, The
Netherlands: ITC).