V

O
L

1
9
,

N
O
.

4
,

P
A
G
E
S

2
1
1

2
8
6

(
2
0
1
1
)

J
O
U
R
N
A
L

O
F

N
E
A
R

I
N
F
R
A
R
E
D

S
P
E
C
T
R
O
S
C
O
P
Y

I
M

P
U
B
L
I
C
A
T
I
O
N
S
volume 19 / number 4 / 2011
ISSN 0967-0335
IN THIS ISSUE:
review: near infrared microscopy and processed animal by-products / the importance of
balanced data sets / determination of azithromycin and decladinosylazithromycin
MlHSoIutions
for PhermeceuticeIs
BOHl prov|des NlP So|ut|ons for Pharmaoeut|oa| lndustry support|ng PAT and QbD approaoh from
warehouse to na| produot re|ease test. BOHls rugged FT-NlP speotrometer, Ohemometr|o software and
NlP g|oba| network experts ensure the h|ghest qua||ty resu|ts.
Mater|a| |dent|cat|on supported by NlPF|ex N-500 w|th a robust F|ber Opt|os and numerous proven
a|gor|thms and oomprehens|ve speotra| ||brar|es.
Automated tab|et ana|ys|s at-||ne for phys|oa| propert|es and oontent un|form|ty test by the NlP
Oheokmaster.
F|na| tab|ets re|ease poss|b|e by NlPF|ex N-500 w|th the ded|oated measur|ng so||d transm|ttanoe oe||.
Easy and powerfu| ca||brat|on oreated by the Oa||brat|on W|zard funot|on |n the NlPOa| Ohemometr|os
software.
Regu|at|on Comp||ance w|th the GMP requ|rements, 21 OFP Part 11 ru|es and |nternat|ona| pharma-
oopoe|as.
Get up to date w|th the |atest teohno|ogy and nd out more on:
www.buch|.com > NIRSo|ut|ons
OueIity in your hends www.buch|.com
BOHl |aborteohn|k AG
9230 F|aw||/Sw|tzer|and
T +41 71 394 63 63
F +41 71 394 65 65
on w
w
w
.buchi.com
/nirw
in
C
o
d
e
: O
O
B
O
7
JOURNAL OF NEAR INFRARED SPECTROSCOPY
Editor-in-chief
Graeme D. Batten
Sea Spec Pty Ltd
PO Box 487, Woolgoolga, NSW 2456, Australia. Phone / Fax +61-266-562288. E-mail: thebattens@bigpond.com
(Adjunct Professor at Charles Sturt University, Honorary Professor at Sydney University)
Regional editors
Subject editors
Editorial advisory board
Asia
Sumio Kawano
National Food Research Institute, 2-1-12 Kannondai, Tsukuba 305-8642, Japan.
E-mail: kawano@affrc.go.jp
Europe
Gerard Downey
Ashtown Food Research Centre, Ashtown, Dublin 15, Ireland.
E-mail: gerard.downey@teagasc.ie
North America
Franklin E. Barton, II
Light Light Solutions, LLC, PO Box 81486, Athens, GA 30608-1486, USA.
E-mail: barton@lightlightsolutions.com
Founding editor
Anthony M.C. Davies
75 Intwood Road, Cringleford, Norwich, Norfolk NR4 6AA, UK.
E-mail: td@nnirc.co.uk
Chemistry/spectroscopy
Heinz W. Siesler
Department of Physical Chemistry, University of Duisberg-Essen, PO Box
103674, D-45117 Essen, Germany. E-mail: hw.siesler@uni-due.de
Chemometrics
Tom Fearn
Department of Statistical Science, University College London, Gower Street,
London WC1E 6BT, UK. E-mail: tom@stats.ucl.ac.uk
Clinical/medical
Karl H. Norris
Consultant, 11204 Montgomery Road, Beltsville, MD 20705, USA.
E-mail: knnirs@gmail.com
Textile bres
Joe Montalvo
USDA, ARS, SRRC, PO Box 19687, New Orleans, LA 70179, USA. E-mail:
montalvo@srrc.ars.usda.gov
Food
Tomas Isaksson
Department of Chemistry, Biotechnology and Food Science UMB, Norwegian
University of Life Sciences, PO Box 5036, N-1432 s, Norway.
E-mail: tomas.isaksson@umb.no
Pharmaceutical
Carl Anderson
Graduate School of Pharmaceutical Sciences, Duquesne University, Pittsburgh,
PA 15282, USA. E-mail: andersonca@duq.edu
Process control
Lois Weyer
ATK Elkton, 55 Thiokol Rd, Elkton, MD, USA.
E-mail: lweyer2@verizon.net
Editorial assistant: Lyn Batten
Submission of articles
Authors intending to submit an article for publication should consult the Guidelines for Authors available from the Publishers or Regional Editors. For the fastest
decision, manuscripts should be submitted via our online system at http://www.impublications.com/authors/. They may also be sent to the Editor-in-Chief or to
a Regional or Subject Editor convenient to the author.
D. Bertrand (Nantes, France)
P. Dardenne (Libramont, Belgium)
E.W. Ciurczak (Laurel, MD, USA)
M. Ferrari (LAquila, Italy)
P. Flinn (Victoria, Australia)
A. Garrido-Varo (Crdoba, Spain)
R. Giangiacomo (Lodi, Italy)
D.W. Hopkins (Battle Creek, MI, USA)
M. Iwamoto (Ibaraki, Japan)
M. Manley (Stellenbosch, RSA)
H. Mark (Suffern, NY, USA)
H. Martens (s, Norway)
I. Murray (Aberdeen, Scotland)
T. Ns (s, Norway)
B.G. Osborne (North Ryde, Australia)
Y. Ozaki (Sanda, Japan)
N. Szydlowski-Zanier (Rueil-Malmaison, France)
J.S. Shenk (Port Matilda, PA, USA)
P.C. Williams (Nanaimo, Canada)
This disk symbol on the rst page of an article indicates the author(s)s willingness to make available the data. If you want it, you should contact the authors
with a request for the data and the undertaking that it will be used for research purposes and not used in any publication without their agreement.
An asterisk against an authors name on the rst page of an article indicates the author to whom correspondence should be addressed.
JOURNAL
OF
NEAR
INFRARED
SPECTROSCOPY
Contents
Review: The potential of near infrared microscopy to detect, identify and quantify processed animal
by-products 211
Zengling Yang, Lujia Han, Juan Antonio Fernndez Pierna, Pierre Dardenne and Vincent Baeten
The importance of balanced data sets for partial least squares discriminant analysis: classication
problems using hyperspectral imaging data 233
Susanne W. Lindstrm, Paul Geladi, Oskar Jonsson and Fredrik Pettersson
Development and validation of a multivariate calibration method for determining interfacial tension of
transformers insulating oils by near infrared spectroscopy 243
Mariana S. Godinho, Anselmo E. Oliveira and Marcelo M. Sena
Soil carbon determination in a Mediterranean vertisol by visible and near infrared reectance
spectroscopy 253
Jos M. Fontn, Luis Lpez-Bellido, Juan Garca-Olmo and Rafael J. Lpez-Bellido
Quantitative calibration models for the determination of azithromycin and decladinosylazithromycin in
azithromycin injection powder by using diffuse reectance near infrared spectroscopy 265
Ji-Xiong Dong, Wen-Bo Zhou, Yan-Chun Feng, Dan-Qing Song and Chang-Qin Hu
A feasibility study on using near infrared spectroscopy to classify strawcoal blends 277
Cheng He, Zengling Yang, Guangqun Huang, Longjian Chen and Lujia Han
Erratum: Classication of viable and non-viable spinach (Spinacia oleracea L.) seeds by single seed
near infrared spectroscopy and extended canonical variates analysis 285
Merete Halkjaer Olesen, Nisha Shetty, Rene Gislum and Birte Boelt
ISSN 0967-0335
IM Publications LLP, 6 Charlton Mill, Charlton, Chichester, West Sussex PO18 0HY, UK.
Tel: +44-(0)1243-811334
Fax: +44-(0)1243-811711
Web: www.impublications.com
Aims and scope
The Journal of Near Infrared Spectroscopy aims to publish original research papers, short
communications, review articles and letters concerned with near infrared spectroscopy and technology,
its application, new instrumentation and the use of chemometric and data handling techniques within
NIR. In this way it intends to serve as a forum for the exchange of information on all aspects of NIR
spectroscopy and technology.
The Journal of Near Infrared Spectroscopy will accept contributions from all areas of application
where near infrared spectroscopy is in use, such as: Agriculture, Chemical Industry, Food, Life
Sciences, Process Control, Pharmaceuticals, Textiles and Polymers. Topics for papers may include:
Chemometrics, Calibrations, Diffuse Reection, NIR Imaging, On-Line Use, Fibre Optics, Sampling,
Spectroscopy, Instrumentation and Remote Sensing.
IM Publications LLP 2011
All rights reserved. Apart from any fair dealing for the purposes of research or private study, or criticism
or review, as permitted under the UKs Copyright, Designs and Patents Act 1988, this publication may
be reproduced, stored or transmitted, in any form or by any means, only with the prior permission in
writing of the publishers, or in the case of reprographic reproduction, in accordance with the terms of
licences issued by the Copyright Licensing Agency. Enquiries concerning reproduction outside those
terms should be sent to the publishers at the address above.
Subscription details
Journal of Near Infrared Spectroscopy is published bimonthly. The subscription rates are 295.00
(delivery addresses in UK), 410.00 (delivery in continental Europe) or US$520.00 (delivery addresses
outside Europe) for 2011. Subscriptions can only be accepted for a full calendar year, and should be
ordered from the publishers, IM Publications LLP, 6 Charlton Mill, Charlton, Chichester, West Sussex
PO18 0HY, UK, Tel: +44(0)1243811334, Fax: +44(0)1243811711, e-mail: subs@impublications.co.uk,
or from your subscription agent.
The subscription rates include airmail delivery to all countries except the UK, where normal inland
postage is used. A reduced rate subscription package is available to subscribers taking both Journal
of Near Infrared Spectroscopy and NIR news; this only applies when both publications are ordered
and paid for with a single transaction. An electronic version of Journal of Near Infrared Spectroscopy
is also available on CD-ROM and on the world wide web. Further details are available from the
publishers at the address above.
Abstracting
Journal of Near Infrared Spectroscopy is abstracted in Chemical Abstracts, Science Citation Index
Expanded (SCIE), Research Alert, Chemistry Citation Index (CCI), Current Contents/Physical,
Chemical & Earth Sciences (CC/PC&ES), Current Contents/Agriculture, Biology & Environmental
Sciences (CC/AB&ES).
Advertising
Advertising is accepted in Journal of Near Infrared Spectroscopy, further details can be obtained from
the Advertisement Manager, IM Publications LLP, 6 Charlton Mill, Charlton, Chichester, West Sussex
PO18 0HY, UK, Tel: +44 (0)1243-811334, Fax: +44 (0)1243-811711, e-mail: ian@impublications.co.uk.
Printed in the UK by Latimer Trend & Company Ltd, Plymouth.
JOURNAL
OF
NEAR
INFRARED
SPECTROSCOPY
211
ISSN: 0967-0335 IM Publications LLP 2011
doi: 10.1255/jnirs.935 All rights reserved
Feedstuffs are composed of feed materials derived from prod-
ucts of vegetal, animal or mineral origin. They are used in
their natural state or have been processed. In the case of
feed materials of animal origin, the process of transforming
animal by-products into valuable products [for example, meat
and bone meal (MBM) and fat] is called rendering. In EC
Regulation 1774/2002,
1
these products are called processed
animal proteins (PAPs), but only PAP material from category
3, which consists of animal by-products that are t for human
consumption, can be used to feed farm animals. Under this
denition, PAPs do not include blood products, milk or milk-
based products, colostrum, gelatine, hydrolysed proteins and
dicalcium phosphate. Of all the materials listed in category 1
or category 2, MBM is one of the most important PAPs among
the processed animal by-products derived from the processing
because of its quantity and market value. MBM is dened in
Commission Directive 98/67/EC as the product obtained by
heating, drying and grinding whole or parts of warm-blooded
land animals from which the fat might have been partially
extracted or physically removed. The product should be
substantially free of hooves, horn, bristle, hair and feathers,
as well as digestive tract content.
2
For more than 10 years,
MBM has been the focal point of the control laboratories and
management bodies responsible for addressing the bovine
Review
The potential of near infrared microscopy
to detect, identify and quantify processed
animal by-products
Zengling Yang,
a,b
Lujia Han,
a,b,*
Juan Antonio Fernndez Pierna,
c
Pierre Dardenne
c
and Vincent Baeten
c,*
a
College of Engineering, China Agricultural University, Beijing 100083, PR China. E-mail: hanlj@cau.edu.cn
b
State Key Laboratory of Animal Nutrition, Beijing 100091, PR China
c
Food and Feed Quality Unit, Valorisation of Agricultural Products Department, Walloon Agricultural Research Centre CRA-W, Henseval
Building, Chausse de Namur 24, 5030 Gembloux, Belgium. E-mail: baeten@cra.wallonie.be
Epidemiological studies have indicated that the most likely pathway of the infection of cattle with bovine spongiform encephalopathy
(BSE) is feed-borne contamination with animal proteins. The enforcement of the ban on meat and bone meal in feed for farmed animals,
including the species-to-species ban, is considered an important measure to prevent the spread of BSE. This review summarises more
than ten years of work on species identication, quantication, comparison with optical microscopy, combination with polymerase chain
reaction and detection based on chemometric decision rules and discriminant models. In the rst part of the review, a summary is given
of existing methods for the detection, identication and quantication of processed by-products of animal origin. In the second part, the
possibilities offered by the near infrared microscopy technique for detecting, identifying and quantifying processed by- products of ani-
mal origin is reviewed. Work needed on enforcing the feed ban in terms of both routine and ofcial control measures is also discussed.
Keywords: feed safety, near infrared microscopy NIR microscopy, bovine spongiform encephalopathy, BSE, processed animal proteins, PAPs,
detection, identication, quantication
Introduction
Z. Yang et al., J. Near Infrared Spectrosc. 19, 211231 (2011)
Received: 16 March 2011 Revised: 22 August 2011 Accepted: 25 August 2011 Publication: 26 September 2011
212 The Potential of NIR Microscopy to Detect, Identify and Quantify Processed Animal By-Products
spongiform encephalopathy (BSE) epidemic. Epidemiological
studies have shown that the most likely pathway of the infec-
tion of cattle with BSE is feed-borne contamination with animal
proteins.
3,4
The earliest case of BSE occurred in 1984 in the UK, with
further BSE outbreaks occurring in the late 1990s.
5
The rst
BSE case in Asia was reported by the Japanese Ministry
of Agriculture, Forestry and Fisheries (MAFF) in 2001. The
Organisation Internationale des Epizooties (OIE) reported that
there were about 25 countries where cattle were infected with
BSE. To date, there have been more than 190,000 reported
cases of BSE in Asia, Europe and North America (http://www.
oie.int).
For this reason, strict legislation has been enacted in Europe
and elsewhere in the world to prevent MBM from entering
ruminant feed.
3
The use of mammalian meat and bone meal
(MMBM) to feed ruminants was banned throughout Europe in
1994; this was the rst time that MMBM had been banned as a
feed ingredient.
6
The discovery of BSE cases in animals born
after the rst ban, however, forced European authorities to
extend and reinforce the partial feed ban. In 2000, the ban on
PAPs in relation to farmed animals was extended for a dened
period. Later, a permanent ban was introduced relating to the
use of protein derived from mammals to feed ruminants and
the prohibition was extended for farmed animals. Derogation
was applied to the use of milk, milk-based products, colostrum,
eggs, egg products, gelatine derived from non- ruminants
and hydrolysed proteins derived from parts of non-ruminants
and from ruminant hides and skins.
7
Derogation was also
applied to the feeding of non-ruminant farmed animals with
sh meal, dicalcium phosphate and tricalcium phosphate, as
well as with blood products derived from non-ruminants.
8

Subsequently, the use of sh meal for young farmed ruminant
species was authorised only for the production of milk substi-
tutes.
9
The MAFF in Japan introduced a complete feed ban
whereby all animal proteins, apart from milk protein, were
prohibited in feed for ruminants, swine and poultry. Later, this
ban was amended to allow protein derived from swine and
poultry to be used in feed for monogastric animals on condi-
tion that these proteins were produced in mills where no rumi-
nant materials were handled.
10
To date, there have been no
BSE cases in China, but the countrys Ministry of Agriculture
has nevertheless enacted regulations to ensure the safety of
the feed chain. These regulations state that: (1) PAPs, espe-
cially those derived from ruminants, are strictly prohibited in
feed for ruminants;
11
(2) the importation of PAPs from coun-
tries with BSE cases is prohibited; and (3) the regulations are
enforced to manage feed products of animal origin, to prevent
contamination during production processes and to ensure the
source and animal species of raw materials.
12
According to
these regulations, MBM can still be used to feed monogastric
animals in China.
As discussed earlier, the European policy was driven
mainly by the goal to eradicate BSE in the European area.
This explains why, after the partial ban in 1994, the European
Commission then reinforced the ban by introducing an
extended and permanent ban in 2001. The problem at that
time was not so much the voluntary adulteration of the feed-
stuffs but rather accidental cross-contamination. About 90%
of mills in Europe share facilities for producing both ruminant
and non-ruminant feedstuffs, so cross-contamination at feed
mills was highly probable.
13
Unintentional cross-contamina-
tion of ruminant diets with feed intended only for monogastric
and poultry species was therefore thought to be the primary
cause of the persistence of the BSE epidemic after the 1994
partial feed ban had been implemented.
13
Later, the Animal
By-Products Regulation included text to regulate the use of
animal by- products to feed the same species, thus preventing
cannibalism.
1
To implement this, methods were needed for
discriminating, at species level, the animal by-products found
in compound feed. In order to avoid cross-contamination, it
was essential to adopt appropriate strategies (including suit-
able analytical methods) during the production, transportation
and storage of all raw materials and feedstuffs.
The European Food Safety Authority recently published a
Scientic Opinion about the quantitative risk assessment of
BSE in relation to a small amount of MBM. It concluded that
the current global limit of PAPs detection in feed is still consid-
ered to be 0.1%, but it recommended continuing the develop-
ment of analytical methods to improve this limit of detection
(LOD) with regard to animal proteins in feed.
14
Various analytical methods have been proposed to ensure
effective control of the feed bans. In the next section of this
paper, a summary is given of the existing methods for the
detection, identification and quantification of processed
by-products of animal origin. These methods include optical
microscopy (OM), polymerase chain reaction (PCR), immuno-
logical techniques, near infrared (NIR) reectance spectros-
copy, near infrared (NIR) microscopy and NIR imaging. The
paper focuses particularly on the possibilities offered by the
NIR microscopy technique for detecting, identifying and quan-
tifying processed by-products of animal origin. It summarises
more than 10 years of work and discusses further studies
required to implement the NIR microscopy method as part of
routine and ofcial control measures.
Overview of existing methods for the
detection, identication and quantication of
animal proteins in feed
Since the BSE crisis and the suspected link between PAPs
and BSE, efforts have been made to develop existing and new
methods of detecting the presence of animal products in
compound feeds. This section summarises the main features
of these methods and their main pros and cons. There are
more details in recent reviews.
13,15,16
Optical microscopy
Currently, MBM is detected mainly by OM. This involves the
visual identication of specic morphological (for example,
feathers in the detection of avian products) and histological
features (for example, lacunae and canaliculae in the iden-
tication of bones). In order to perform an OM analysis, the
Z. Yang et al., J. Near Infrared Spectrosc. 19, 211231 (2011) 213
sample is ground (usually at 1 mm or 2 mm) and the particles
are spread on microscope slides and observed. There are
several sample preparation protocols, including the treat-
ment of sample fractions with specic reagents for detecting
specic particles of animal origin. These protocols allow, for
example, the detection of muscle fibres, bones, feather or
blood. According to the protocol, the analysis is performed on
the sample either as it is or on the sediment obtained after
a specic solvent has been used to extract the dense frac-
tion of the sample (for example, in observation of bone and
scales).
15,17
Figure 1 shows some specic morphological and
histological features of animal proteins.
In Europe, the OM method is the only ofcial method for
determining constituents of animal origin in feedstuffs. It has
been validated through several European inter- laboratory
studies.
17,18
The method (including all the protocols for
analysing specic particles) is described in Annex VI of the
Commission Regulation 152/2009 laying down the sampling
and analysis methods for the ofcial control of feed.
19
This
regulation has been in force since 26 August 2009 and replaced
EC Directive 126/2003.
20
In practice, the detection of illicit
ingredients of animal origin is driven mainly by the detection
of bones and scales, which is performed on the sediment
fraction obtained after decantation of the ground samples in
tetrachloroethylene. In doing so, the dense fraction (i.e. parti-
cles with a density higher than 1.62) is obtained and the parti-
cles from this fraction are observed. For detecting bones, a
trained analyst looks for specic bone features of lacunae and
canaliculae. A lacuna is a cavity with a diameter of 1050 m
where the osteocyte was located before being removed by the
rendering process.
As described in the Commission Regulation 152/2009, the
sensitivity of the method depends on the nature of the constit-
uents of animal origin. The LOD has not yet been assessed;
various studies have shown, however, that the LOD is less
than 0.1% when bones are present in PAPs included in the
feed.
21,22
The LOD of the method when there are only muscle
particles is higher than 0.1% (it is thought that the LOD for this
specic detection is about 1%).
16,17
The LOD of the OM method
is mainly affected by the nature of the constituents of animal
origin and by the skills of the operator.
The ability of OM to identify the species origin of particles
is limited. As explained earlier, sh material can be discrimi-
nated by detecting the presence of scales or bones with
specic features at the level of the particles and lacunae (the
shape and distribution of the lacunae varies, depending on the
animal source). This discrimination is not possible, however, in
the observation of muscle bres, for example. Nevertheless, it
has been reported that particles from specic parts of bones
from terrestrial animal are seldom confused with sh parti-
cles.
23
The discrimination among other animals is limited to
the detection of avian material by the presence of feathers.
The method described in 152/2009 (see Point 7 in Annex VI)
also includes a protocol for quantification. Recent inter-
laboratory studies,
24
however, have demonstrated an inability
to reproduce this quantitative protocol, thus preventing its use
in ofcial controls.
Polymerase chain reaction
Much work has been done on developing polymerase chain
reaction (PCR)-based methods for the detection and identi-
fication of DNA in animal origin. These methods are based
on genetic amplification, which is considered as one of the
most efcient ways of detecting specic DNA targets. The most
popular protocols for amplifying DNA are those using PCR. PCR
methods have a high forensic value because they are based
on researching specic DNA sequences characteristic of the
taxonomy level (for example, mammal, ruminant) or the species
level (for example, bovine, pig). Among the PCR methods
proposed, most did not transfer successfully to another labo-
ratory or pass evaluation by an independent laboratory. It is
essential to use multi-copy targets (for example, from mito-
chondrial DNA) and targets of reduced length (ideally, about
or fewer than 100 bp because animal proteins are subjected to
an aggressive rendering process which causes degradation of
DNA.
25,26
Several PCR methods have passed in-house valida-
tion successfully, however, and were positively evaluated in a
study conducted by the European Joint Research Centre.
27,28

These methods allow species-specic detection of sh, cattle,
pigs, poultry and sheep. This detection can be done either on
the sample as it is or on the sediment. Some methods have
been accredited in line with ISO 17025 and are used in addi-
tion to the ofcial method (i.e. OM). Today, some non-European
countries use PCR methods for ofcial control measures. When
the compound feed includes authorised animal by-products (for
example, fat, milk or egg by-products), the detection by PCR of
animal DNA is confused in assessing the presence of forbidden
materials (for example, pig, bovine or avian).
Terrestrial bones Fish bones Muscle fibres Feather
Figure 1. Specic morphological and histological features of animal proteins.
214 The Potential of NIR Microscopy to Detect, Identify and Quantify Processed Animal By-Products
As the DNA content in PAPs will depend on the source
and composition of the material used to produce it and the
rendering process applied, PCR methods cannot be used for
quantication purposes.
16
Immunological methods
There are several immunochemical techniques. In the detec-
tion of PAPs, however, the most common techniques used in
developing detection methods are enzyme-linked immuno-
sorbent assay (ELISA) and lateral ow immunoassays (dip
sticks).
16
Anseld et al.
29
developed an immunoassay method
based on the use of antibodies against a thermostable antigen.
It is able to detect the presence of ruminant and porcine
proteins in feedstuffs and has an LOD of about 0.1%. It has not
yet been validated through an inter-laboratory study.
15,18,29,30

ELISA kits have been developed that show promising results
for detecting the species origin of PAPs.
16,31
There are several
lateral flow dipstick-based methods that have the poten-
tial to detect PAPs.
31,32
They enable PAPs from ruminants
or mammals to be detected. The LOD of these methods is
about 0.51% and limitation has been reported when some of
these methods were used for detecting ruminant PAPs in pig
material.
33,34
These methods are fast, easy to implement and
suitable for identifying specic PAPs.
Near infrared reectance spectroscopy
NIR spectroscopy is the most widely used non-destructive
method in the feed industry to determine qualitative param-
eters of feed ingredients and feedstuffs. The high sample
throughput of the method, its capacity to determine a large
range of parameters
35
(from major constituents to criteria
such as digestibility) in a single analysis and the possibility of
building a network of spectrometers make this technique very
attractive for the feed sector. The fact that it can also be used
online in a feed production plant adds to its attraction. The
principle of this method is based on the absorption of light
with near infrared wavelengths by the molecules that make up
the sample. Cozzolino and co-researchers explored the possi-
bility of using NIR spectra for the characterisation of fish-
meal samples.
36
The absorption bands at 1490 nm (6711 cm
1
)
and 1944 nm (5144 cm
1
) were associated with water content,
whereas the bands at 2060 nm (4854 cm
1
) and 21682180 nm
(46134587 cm
1
) showed a high correlation with crude protein.
Fat content was related to absorption regions around 1700
1730 nm (58825780 cm
1
) and 23002310 nm (43484329 cm
1
).
Several papers have shown the potential of NIR spectros-
copy to detect PAPs of terrestrial origin in feedstuffs
3740
and
in fish meal.
4143
The NIR technique has the advantage of
being able to analyse both ground and unground samples.
It is a fast and cost-effective technique that requires a low
level of expertise.
44
The largest study conducted on this topic
at the European level was done within the framework of the
STRATFEED project.
45,46
In this EC project (www.stratfeed.cra.
wallonie.be), a network of ve spectrometers was set up and
global equations were developed. The results from this project
indicated that NIR spectroscopy could provide the feed industry
with a fast screening method for detecting the contamina-
tion of compound feed with processed animal by-products.
STRATFEED also set the LOD in the network at 11.5%, which
is not sufciently accurate to be used as evidence in cases of
fraud or accidental contamination. Nevertheless, this limit is
interesting for the auto-control performed at feed mill level.
Recent studies have reported that NIR spectroscopy can
be used to identify and quantify the animal species (poultry
by-products, pig, cattle, ruminant and non-ruminant) in PAPs.
The models developed enable the unequivocal classication
of poultry by-product meal and pork meal.
47,48
Until now, no
NIR method has been validated by an inter-laboratory study
following international guidelines.
Near infrared microscopy: principles and
instrumentation
With NIR spectroscopy, the detection of PAPs in feedstuffs is
based on a single NIR spectrum corresponding to the mean
of several scans obtained from a representative portion of
the sample. It means that the spectral information from
the specific absorption of the radiation by PAPs is diluted
by the absorption of the radiation of all the feed ingredients.
Techniques able to collect the spectrum of individual particles
from samples would be required to detect specic particles of
determined origin. In this instance, PAPs detection would not
be performed through the analysis of a single spectrum, but
through the analysis of hundreds or thousands of spectra from
individual particles. This can be done using an NIR micro-
scope. This hyphenated instrument includes a classical NIR
spectrometer coupled with an optical microscope in which the
optics have been adapted to NIR radiation. NIR microscopes
allow the spectra to be collected from extremely small sample
areas (typically, 50 m 50 m or less, depending on the instru-
ment and the conguration). Usually, the instrument includes
a charge coupled device camera and a viewing system that
magnies the visible light image of the sample, allowing the
user to visualise it and to position an infrared beam on the
sample area of interest using a motorised stage. Using the
microscope pointer, the infrared beam is focused on each point
of interest and the near infrared spectrum is collected.
49
The principle of identifying PAPs by NIR microscopy is
similar to the European official method. Instead of visibly
observing the morphological and histological features of parti-
cles, however, infrared absorbances of the compounds making
up the particle are used. OM is based on optics, visual detec-
tors (i.e. the eyes of the analyst) and a sophisticated trained
expert system (i.e. the brain of the analyst). NIR microscopy
is based on optics, NIR detectors and mathematical models
(i.e. calibration equations). In fact, NIR microscopy combines
the analytical advantages of OM and NIRS. NIR microscopy
requires a low level of expertise, it is a non-destructive method
and has a detection limit lower than 0.1%. Moreover, it has a
low level of false negatives and a high repeatability and it is
independent of the feed matrix used.
50
NIR microscopy-based
methods make the assumption that each particle is made up
of a single feed ingredient. Figure 2(a) shows a commercially
Z. Yang et al., J. Near Infrared Spectrosc. 19, 211231 (2011) 215
available instrument from the PerkinElmer Company. Figure
2(b) shows the visible image of a sample displayed on a refer-
ence surface (spectralon). The spectra in Figure 2(c) were
collected from different particles in reectance mode, derived
from ve animal origins (sh meal, poultry MBM, pork MBM,
cattle MBM and sheep MBM). The spectrum characteris-
tics obtained by NIR microscopy correspond to those of NIR
spectroscopy. The rich absorption around 58004000 cm
1
is
related to crude protein and fat. Examples of various kinds of
animal by-product particles (i.e. muscle, feathers, hair, horn,
teeth, bones, scales, blood products, milk, milk-based prod-
ucts, colostrum, gelatin, hydrolysed proteins, dicalcium phos-
phate, fat and egg by-products) have been described in the
literature.
51
Evolution of the near infrared
microscopy methodology
Detection of processed animal proteins
Several feasibility studies based on the application of NIR
microscopy to detect PAPs have been reported over the last
ten years. These applications are summarised in Table 1.
Detection by discriminant equations
In 1999, Piraux and Dardenne published the rst research
paper demonstrating the potential of NIR microscopy for
detecting unauthorised animal ingredients in compound
feed.
52
Their work involved collecting 1740 particles from 56
raw materials (allowed at that time) for feeding ruminants,
including sh meal, peas, manioc, wheat, blood meal, rape-
extracted oil cake, corn, maize gluten feed, maize germ oil
cake, soybean, ax, alfalfa and milk by-products. A total of
1291 particles from 43 animal meal products (forbidden at
that time), including MBM, meat meal, ground bones, feather
meal and poultry by-products were obtained. An NIR Perkin
Elmer microscope was used to scan the particles in reect-
ance mode with an aperture size of 50 m 50 m. The spectra
were collected from 1112 nm to 2500 nm and resulted from
an average of 100 scans. These spectra were pre-processed
using standard normal variate and detrend (SNVD), as well
as rst derivative as the mathematical treatment. An articial
neural network model was used to discriminate the allowed
particles from the forbidden ones. The model was applied
to an independent validation set containing 1872 particles.
The results of the discriminant analysis indicated that it was
possible to detect MBM particles in feedstuffs with a success
rate greater than 99% by NIR microscopy (i.e. a total of false
negative and false positive results of less than 1%). The results
showed that it was possible to detect forbidden animal parti-
cles in a ground compound feedstuff with an overall error rate
of 0.64% (the average value of the rate of allowed particles
misclassied as forbidden particles and the rate of forbidden
particles misclassied as allowed particles). The study also
showed that if the MBM proportion in a compound feedstuff
was low, it was necessary to scan a large set of particles to
nd at least one MBM particle. For example, for a feedstuff
(b) Visible image of the sample
(a) NIR microscopy
(c) NIR spectra of the PAP particles
0
0.3
0.6
0.9
1.2
7800 6800 5800 4800
Wavenumber cm
-1
A
b
s
4000
1
2
3
4
5
1 Poultry MBM
2 Cattle MBM
3 Sheep MBM
4 Pork MBM
5 Fishmeal
1 2
3
4
5
Figure 2. (a) PerkinElmer FT-NIR microscope; (b) visible image; and (c) NIR spectra of the PAPs particles.
216 The Potential of NIR Microscopy to Detect, Identify and Quantify Processed Animal By-Products
S
t
u
d
y

p
u
r
p
o
s
e
s
T
r
a
i
n
i
n
g
s
e
t
V
a
l
i
d
a
t
i
o
n
s
e
t
M
o
d
e
l
;

m
e
t
h
o
d
S
u
c
c
e
s
s
r
a
t
e
R
a
w
/
s
e
d
i
m
e
n
t

f
r
a
c
t
i
o
n
I
m
p
o
r
t
a
n
c
e

a
n
d
/
o
r
w
e
a
k
n
e
s
s
R
e
f
.
T
o

d
e
t
e
c
t
u
n
a
u
t
h
o
r
i
s
e
d

a
n
i
m
a
l
i
n
g
r
e
d
i
e
n
t
s

i
n

c
o
m
p
o
u
n
d

f
e
e
d
1
7
4
0

p
a
r
t
i
c
l
e
s

(
7
8
0

f
o
r

c
a
l
i
b
r
a
t
i
o
n

a
n
d

9
6
0

f
o
r

v
a
l
i
d
a
t
i
o
n
)

f
r
o
m

5
6

a
l
l
o
w
e
d

r
a
w

m
a
t
e
r
i
a
l
s

(

s
h

m
e
a
l
,

p
e
a
s
,

m
a
n
i
o
c
,

w
h
e
a
t
,

b
l
o
o
d

m
e
a
l
,

r
a
p
e
-
e
x
t
r
a
c
t
e
d

o
i
l

c
a
k
e
,

c
o
r
n
,

m
a
i
z
e

g
l
u
t
e
n

f
e
e
d
,

m
a
i
z
e

g
e
r
m

o
i
l

c
a
k
e
,

s
o
y
b
e
a
n
,

a
x
,

a
l
f
a
l
f
a

a
n
d

m
i
l
k

b
y
-
p
r
o
d
u
c
t
s
)
;
1
2
9
1

p
a
r
t
i
c
l
e
s

(
3
7
9

f
o
r

c
a
l
i
b
r
a
t
i
o
n

a
n
d

9
1
2

f
o
r

v
a
l
i
d
a
t
i
o
n
)

f
r
o
m

4
3

f
o
r
b
i
d
d
e
n

a
n
i
m
a
l

m
e
a
l

p
r
o
d
u
c
t
s

(
M
B
M
,

m
e
a
t

m
e
a
l
,

g
r
o
u
n
d

b
o
n
e
s
,

f
e
a
t
h
e
r

m
e
a
l

a
n
d

p
o
u
l
t
r
y

b
y
-
p
r
o
d
u
c
t
s
)
D
i
s
c
r
i
m
i
n
a
n
t

e
q
u
a
t
i
o
n
;

A
N
N
>
9
9
%

f
o
r

b
o
t
h

a
l
l
o
w
e
d

a
n
d

f
o
r
-
b
i
d
d
e
n

p
a
r
t
i
c
l
e
s

(
v
a
l
i
d
a
t
i
o
n
)
R
a
w
T
h
i
s

s
t
u
d
y

f
o
r

t
h
e

r
s
t

t
i
m
e

r
e
p
o
r
t
e
d

t
h
e

N
I
R

m
i
c
r
o
s
-
c
o
p
y

m
e
t
h
o
d

f
o
r

f
e
e
d

i
n
s
p
e
c
t
i
o
n
;
T
h
e
r
e

a
r
e

n
o

a
d
u
l
t
e
r
a
t
e
d

s
a
m
p
l
e
s

t
o

v
a
l
i
d
a
t
e

t
h
e

d
i
s
c
r
i
m
i
n
a
n
t

e
q
u
a
t
i
o
n
;
M
u
c
h

m
o
r
e

r
e
s
t
r
i
c
t
i
v
e

s
t
a
n
d
a
r
d

o
f

d
e
t
e
c
t
i
n
g

r
u
l
e
s

w
i
l
l

b
e

r
e
q
u
i
r
e
d

t
o

s
t
r
e
n
g
t
h
e
n

c
o
n

d
e
n
c
e

i
n

t
h
e

c
l
a
s
s
i

c
a
t
i
o
n
5
2
T
o

d
e
t
e
c
t
a
n
i
m
a
l
i
n
g
r
e
d
i
e
n
t
s

i
n

c
o
m
p
o
u
n
d

f
e
e
d
7
4
9
2

v
e
g
e
t
a
l

p
a
r
t
i
c
l
e
s
;
2
4
8
4

a
n
i
m
a
l

p
a
r
t
i
c
l
e
s
4
8

v
a
l
i
d
a
t
i
o
n

s
a
m
p
l
e
s

w
e
r
e

p
r
e
p
a
r
e
d

u
s
i
n
g

2
4

c
o
m
p
o
u
n
d

f
e
e
d
s

s
p
i
k
e
d

a
t

v
a
r
i
o
u
s

p
e
r
c
e
n
t
a
g
e
s

(
0
.
5

8
%
,

w
/
w

i
n

0
.
5
%

i
n
t
e
r
v
a
l
s
)

w
i
t
h

e
i
g
h
t

M
B
M

p
r
o
d
u
c
t
s
D
i
s
c
r
i
m
i
n
a
n
t

e
q
u
a
t
i
o
n
;

P
L
S
D
A
9
5
%

a
n
d

9
4
.
3
%

f
o
r

t
h
e

v
e
g
e
t
a
l

a
n
d

a
n
i
m
a
l

p
a
r
t
i
c
l
e
s

(
c
a
l
i
b
r
a
t
i
o
n
)
o
n
e

s
a
m
p
l
e

w
a
s

d
e
t
e
c
t
e
d

a
s
n
e
g
a
t
i
v
e

(
v
a
l
i
d
a
-
t
i
o
n
)
R
a
w
T
h
e

d
e
t
e
c
t
i
o
n

r
e
s
u
l
t

d
e
p
e
n
d
s

o
n

t
h
e

d
e
g
r
e
e

o
f

h
o
m
o
g
e
n
e
i
t
y

o
f

t
h
e

s
a
m
p
l
e

a
n
d

t
h
e

n
u
m
b
e
r

o
f

p
a
r
t
i
-
c
l
e
s

a
n
a
l
y
s
e
d
;
I
n

o
r
d
e
r

t
o

s
t
r
e
n
g
t
h
e
n

c
o
n
-

d
e
n
c
e

i
n

t
h
e

c
l
a
s
s
i

c
a
-
t
i
o
n
,

t
h
e

r
e
s
u
l
t
s

o
f
d
i
f
f
e
r
e
n
t

e
q
u
a
t
i
o
n
s

t
o

d
e
c
i
d
e

i
n

w
h
i
c
h

g
r
o
u
p

a

p
a
r
t
i
c
l
e

s
p
e
c
t
r
u
m

s
h
o
u
l
d

b
e

t
a
k
e
n

i
n
t
o

a
c
c
o
u
n
t
5
1
T
o

d
e
t
e
c
t

M
B
M

i
n

a
n
i
m
a
l

f
e
e
d
s
2
2
2
9

a
n
i
m
a
l

p
a
r
t
i
c
l
e
s

(
i
.
e
.

p
i
g
s
,

p
o
u
l
t
r
y
,

s
h
e
e
p
,

c
o
w
s

a
n
d

a

m
i
x
t
u
r
e

o
f

t
h
e
s
e
)
;
1
5
5
6

p
l
a
n
t
-
b
a
s
e
d

f
e
e
d
s
p
a
r
t
i
c
l
e
s

(
b
a
r
l
e
y
,

m
a
i
z
e
,
s
o
y
b
e
a
n
,

w
h
e
a
t
,

e
t
c
.
)
(
1
)

1
8

c
o
m
p
o
u
n
d

f
e
e
d
s
,

i
n
c
l
u
d
i
n
g

1
0

M
B
M
-
f
r
e
e

f
e
e
d
s
t
u
f
f
s

f
o
r

l
i
v
e
s
t
o
c
k

a
n
d

e
i
g
h
t

p
e
t

f
o
o
d
s

c
o
n
-
t
a
i
n
i
n
g

2
6

3
5
%

M
B
M
(
2
)

v
e

c
o
m
p
o
u
n
d

f
e
e
d
s

c
o
n
t
a
i
n
i
n
g

M
B
M

(
1
%
,

0
.
5
%
,

0
.
0
2
%
,

0
.
2
5
%
)

a
n
d

f
o
u
r

c
o
n
t
a
i
n
i
n
g
d
i
f
f
e
r
e
n
t

b
l
o
o
d

m
e
a
l

p
e
r
c
e
n
t
a
g
e
s

(
f
r
o
m

0
.
0
2
%

t
o

0
.
3
%
)
D
i
s
c
r
i
m
i
n
a
n
t

e
q
u
a
t
i
o
n
;

P
L
S
D
A
(
1
)

F
o
u
r

f
a
l
s
e

p
o
s
i
t
i
v
e

s
a
m
p
l
e
s

(
m
a
y
b
e

i
n
c
l
u
d
e

a
n
i
m
a
l

f
a
t

o
r

b
l
o
o
d
)
(
2
)

V
a
l
i
d
a
t
i
o
n

r
e
s
u
l
t

n
e
e
d

t
o

b
e

c
o
n

r
m
e
d

b
y

r
e
p
e
a
t
a
b
i
l
i
t
y
R
a
w
T
h
i
s

p
a
p
e
r

d
i
d

n
o
t

r
e
p
o
r
t

r
e
p
e
a
t
a
b
i
l
i
t
y

d
a
t
a
;
C
o
n
s
i
d
e
r
a
t
i
o
n

a
b
o
u
t

t
h
e

p
o
s
s
i
b
i
l
i
t
y

t
o

d
e
t
e
c
t

a
t

a

l
e
v
e
l

o
f

0
.
0
2
%

b
y

a
n
a
l
y
s
i
n
g

o
n
l
y

5
0
0

p
a
r
t
i
c
l
e
s
5
5
T
a
b
l
e

1
.

D
e
t
e
c
t
i
o
n

o
f

p
r
o
c
e
s
s
e
d

a
n
i
m
a
l

p
r
o
t
e
i
n
s

b
y

n
e
a
r

i
n
f
r
a
r
e
d

m
i
c
r
o
s
c
o
p
y
.
Z. Yang et al., J. Near Infrared Spectrosc. 19, 211231 (2011) 217
S
t
u
d
y

p
u
r
p
o
s
e
s
T
r
a
i
n
i
n
g
s
e
t
V
a
l
i
d
a
t
i
o
n
s
e
t
M
o
d
e
m
e
t
h
o
d
S
u
c
c
e
s
s
r
a
t
e
R
a
w
/
s
e
d
i
m
e
n
t

f
r
a
c
t
i
o
n
I
m
p
o
r
t
a
n
c
e

a
n
d
/
o
r
w
e
a
k
n
e
s
s
R
e
f
.
T
o

e
s
t
a
b
l
i
s
h

t
h
e

s
u
i
t
a
b
i
l
i
t
y

o
f

N
I
R

m
i
c
r
o
s
c
o
p
y

f
o
r

d
e
t
e
c
t
i
o
n

o
f

M
B
M

b
y

d
e
t
e
r
m
i
n
i
n
g

t
h
e

s
e
n
s
i
t
i
v
i
t
y

a
n
d

s
p
e
c
i

c
i
t
y
4
7
1
6

p
a
r
t
i
c
l
e
s

o
f

a
n
i
m
a
l

o
r
i
g
i
n
;
3
2
8
6

p
a
r
t
i
c
l
e
s

f
r
o
m

c
o
m
-
p
o
u
n
d

f
e
e
d
1
5

s
a
m
p
l
e
s
,

b
o
t
h

n
o
n
-
a
d
u
l
t
e
r
a
t
e
d

a
n
d

a
d
u
l
-
t
e
r
a
t
e
d

w
i
t
h

M
B
M

a
t

0
.
0
5
%
,

0
.
1
%
,

0
.
5
%

a
n
d

1
%
,

w
e
r
e

u
s
e
d

t
o

s
t
u
d
y

t
h
e

r
e
p
e
a
t
a
b
i
l
i
t
y
.
4
8

s
p
i
k
e
d

s
a
m
p
l
e
s
,

o
b
t
a
i
n
e
d

b
y

m
i
x
i
n
g

2
4

f
e
e
d
s
t
u
f
f
s

w
i
t
h

e
i
g
h
t

M
B
M

p
r
o
d
u
c
t
s

a
t

d
i
f
-
f
e
r
e
n
t

l
e
v
e
l
s

(
0
.
5

8
%
,

w
/
w

i
n

0
.
5
%

i
n
t
e
r
v
a
l
s
)
,

a
n
d

2
4

c
o
m
p
o
u
n
d

f
e
e
d
s

s
a
m
p
l
e
s

w
e
r
e

u
s
e
d

t
o

s
t
u
d
y

t
h
e

r
e
l
i
a
b
i
l
i
t
y

o
f

t
h
e

m
e
t
h
o
d
(
1
)

D
i
s
c
r
i
m
i
n
a
n
t

e
q
u
a
t
i
o
n
;

P
L
S
D
A
(
2
)

D
e
c
i
s
i
o
n

r
u
l
e
;

V
i
s
u
a
l

b
s
e
r
v
a
t
i
o
n
S
u
c
c
e
s
s

r
a
t
e

i
d
e
n
t
i
f
y
i
n
g

a
n
i
m
a
l

p
a
r
t
i
c
l
e
s

w
a
s

9
7
.
5
%

b
y

P
L
S
D
A

e
q
u
a
t
i
o
n
,

a
n
d

9
6
.
6
%

b
y

d
e
c
i
s
i
o
n

r
u
l
e

(
c
a
l
i
b
r
a
t
i
o
n
)
n
o

f
a
l
s
e

p
o
s
i
t
i
v
e

o
r

f
a
l
s
e

n
e
g
a
t
i
v
e

r
e
s
u
l
t
s

(
v
a
l
i
d
a
t
i
o
n
)
S
e
d
i
m
e
n
t
T
h
e

r
e
p
e
a
t
a
b
i
l
i
t
y

a
n
d

r
e
l
i
-
a
b
i
l
i
t
y

o
f

t
h
i
s

d
i
s
c
r
i
m
i
n
a
n
t

m
o
d
e
l

w
e
r
e

v
a
l
i
d
a
t
e
d
;
F
u
r
t
h
e
r

r
e
s
e
a
r
c
h

f
o
r

d
e
c
i
-
s
i
o
n

r
u
l
e
s

w
i
l
l

b
e

n
e
e
d
e
d

t
o

d
i
s
t
i
n
g
u
i
s
h

m
o
r
e

e
a
s
i
l
y

a
n
d

i
m
p
r
o
v
e

t
h
e

a
c
c
u
r
a
c
y
5
4
T
o

e
s
t
a
b
l
i
s
h

t
h
e

L
O
D

a
n
d

r
e
p
e
a
t
a
b
i
l
i
t
y

o
f


N
I
R

m
i
c
r
o
s
c
o
p
y
d
i
s
c
r
i
m
i
n
a
n
t

m
o
d
e
l
N
A
A

s
a
m
p
l
e

s
p
i
k
e
d

a
t

0
.
1
%

w
i
t
h

M
B
M

w
a
s

a
n
a
l
y
s
e
d

1
0

t
i
m
e
s
D
i
s
c
r
i
m
i
n
a
n
t

e
q
u
a
t
i
o
n
;

P
L
S
D
A
N
o

f
a
l
s
e

n
e
g
a
t
i
v
e

r
e
s
u
l
t
s

(
v
a
l
i
d
a
t
i
o
n
)
S
e
d
i
m
e
n
t
T
h
e

L
O
D

a
n
d

r
e
p
e
a
t
a
b
i
l
i
t
y

o
f

t
h
e

m
e
t
h
o
d

w
e
r
e

v
a
l
i
-
d
a
t
e
d
5
0
,

5
1
T
o

d
e
v
e
l
o
p

a
p
p
r
o
p
r
i
a
t
e

d
e
c
i
s
i
o
n

r
u
l
e
s

a
n
d
d
e
m
o
n
-
s
t
r
a
t
e

t
h
e

t
r
a
n
s
f
e
r
a
b
i
l
i
t
y
N
A
2
0

s
a
m
p
l
e
s

w
i
t
h
d
i
f
f
e
r
e
n
t

M
B
M
p
e
r
c
e
n
t
a
g
e
s

(
0

8
%
)

w
e
r
e

a
n
a
l
y
s
e
d
D
e
c
i
s
i
o
n

r
u
l
e
;
v
i
s
u
a
l
o
b
s
e
r
v
a
t
i
o
n
N
o

f
a
l
s
e

p
o
s
i
t
i
v
e

o
r

f
a
l
s
e

n
e
g
a
t
i
v
e

r
e
s
u
l
t
s

(
v
a
l
i
d
a
t
i
o
n
)
S
e
d
i
m
e
n
t
A
p
p
r
o
p
r
i
a
t
e

d
e
c
i
s
i
o
n

r
u
l
e
s

w
e
r
e

d
e
v
e
l
o
p
e
d

a
n
d
v
a
l
i
d
a
t
e
d
;
T
h
e

t
r
a
n
s
f
e
r
a
b
i
l
i
t
y

o
f
d
e
c
i
s
i
o
n

r
u
l
e
s

w
a
s
d
e
m
o
n
s
t
r
a
t
e
d

b
y

t
h
r
e
e

l
a
b
s
;
F
u
r
t
h
e
r

w
o
r
k

s
h
o
u
l
d

b
e

l
a
u
n
c
h
e
d

t
o

s
p
e
c
i
e
s

i
d
e
n
t
i
-

c
a
t
i
o
n

a
n
d

o
n

q
u
a
n
t
i

c
a
-
t
i
o
n

o
f

P
A
P
s
5
6
*
N
o
t

a
v
a
i
l
a
b
l
e
.
T
a
b
l
e

1

(
c
o
n
t
i
n
u
e
d
)
.
D
e
t
e
c
t
i
o
n

o
f

p
r
o
c
e
s
s
e
d

a
n
i
m
a
l

p
r
o
t
e
i
n
s

b
y

n
e
a
r

i
n
f
r
a
r
e
d

m
i
c
r
o
s
c
o
p
y
.
218 The Potential of NIR Microscopy to Detect, Identify and Quantify Processed Animal By-Products
sample containing 0.5% MBM, in order to observe at least
one MBM particle with a probability of 95%, about 1000 parti-
cles need to be analysed.
52,53
This study demonstrated for the
rst time that the NIR microscopy technique was suitable to
address issues related to feed inspection. However, additional
works on adulterated samples to validate the discriminant
equations, as well as on the rules to decide in which group a
particle spectrum should be placed, had to be performed.
Based on this preliminary research, the method has been
developed and validated in-house within the framework of the
European STRATFEED project. In this project, two protocols
were developed (Figure 3). For the analysis of the raw fraction
(non-sediment), the detection was based on the presence of
various particles (for example, muscle and feathers, as well as
those found in the sediment fraction) of animal origin present
in the sample. For the analysis of the sediment fraction,
however, detection was based only on the presence of specic
animal particles (for example, bone, scales, cartilage).
The spectra of particles were collected from 1112 nm
to 2500 nm in reflectance mode with an aperture size
of 50 m 50 m. Then a large spectral library was constructed,
including more than 20,000 spectra from feed ingredients and
compound feed. Based on the results of a principal component
analysis, vegetal and animal particle subspaces were dened
and then the animal particle sub-spaces were divided into
sh particles and terrestrial particles. Finally, the terrestrial
particle sub-space was sub-divided into poultry particles and
mammal particles. Various supervised methods, including
partial least squares discriminant analysis (PLSDA), articial
neural networks and soft independent modelling of class
analogy (SIMCA), were also tested in order to build the discri-
minant equations.
50,51,53,54
The SIMCA method constructs
a principal component analysis model for each individual
group of interest and the allocation of new samples is based
on the computing of distances to each model. The PLSDA
method constructs a regression model between the spectral
information and a matrix of dummy variables representing
the different groups of interest. The articial neural network
technique processes information in a similar way as the
human brain does. A network is composed of a large number
of highly interconnected processing elements (neurons)
working in parallel to solve a specic problem. It is a learning
algorithm, i.e. before using a network for prediction it must
be trained with known data. This is necessary to ensure that
the articial network provides useful results. While learning,
the artificial network compares its output with observed
(known) output values of learning data.
Near infrared microscopy analysis on the raw fraction
The detection of animal particles in the sample (as it was)
was achieved by analysing 7492 vegetal particles and 2484
animal particles in reectance mode with an aperture size
of 50 m 50 m. The samples were selected to cover the full
diversity of feed ingredients used in formulating feedstuffs. In
order to construct the discriminant equation, the PLSDA algo-
rithm was used. Up to 95% of the vegetal feed ingredient parti-
cles and 94.3% of the animal particles were correctly classi-
ed. In order to test the mathematical models constructed, 48
validation samples were prepared using 24 compound feeds
(destined for feeding cattle, goats, poultry, pigs and rabbits)
spiked at various percentages (0.58%, w/w in 0.5% inter-
vals) with eight MBM products. A randomised factorial design
was used. Between 141 and 710 particles from the validation
samples were analysed. Of the 48 validation samples, 47 were
shown to be positive for the presence of MBM. Only one sample
containing 0.5% MBM was erroneously detected as negative.
51

The detection ability depends on the degree of homogeneity of
the samples and on the number of particles analysed. In this
study, a new strategy was proposed to strengthen condence
in the classification results. Considering the results from
different equations, it is decided in which group a particle
spectrum should be placed. If a particle belongs to the animal
group, it should be correctly discriminated by both the vegetal
vs animal equation and the vegetal vs terrestrial animal (or
sh) equation simultaneously.
De la Roza-Delgado et al. have also worked on the detec-
tion of animal particles in animal feeds.
55
They used 2229
particle spectra of animal origin (i.e. pigs, poultry, sheep, cows
and a mixture of these) and 1556 spectra from plant-based
feeds (barley, maize, soybean, wheat etc.) in order to establish
discriminant equations by using PLSDA as the chemometric
tool. The best results were obtained by applying SNVD and rst
derivative as pre-treatments. They used 18 compound feeds,
including 10 MBM-free feedstuffs for livestock and eight pet
foods containing 2635% MBM, to validate the discriminant
equation. No false positive or false negative results were
Sieve
(>250m)
Sedimentation
Tetrachloroethylene
Sieve
(>250m)
NIRM
Analysis
Grinding
(1-2mm)
NIRM
Analysis
Samples
Protocol 1
On the raw fraction
Protocol 2
On the sediment fraction
Figure 3. Schematic presentation of the two NIR microscopy protocols developed and validated in the STRATFEED project.
Z. Yang et al., J. Near Infrared Spectrosc. 19, 211231 (2011) 219
detected. They reported that the protocol developed could
be used to detect the presence of MBM at concentrations as
low as 0.02%, with 100% correct classication determined
by the analysis of a validation set, including ve compound
feeds containing 1% MBM (1%, 0.5%, 0.02%, 0.25%) and
four containing different blood meal percentages (from 0.02%
to 0.3%). However, their paper did not report a repeatability
study of the proposed protocol that would allow them to draw
this conclusion. The protocol is based on the analysis of 500
individual particles, meaning that if the sample is perfectly
homogeneous (in terms of particle distribution, density and
size) the probability of detecting at least one animal particle
in a sample adulterated at 0.02% is 0.09%. More than 15,000
particles would have to be analysed if a 95% level of condence
were to be reached with samples spiked at 0.02%.
Near infrared microscopy analysis on the sediment
fraction
The protocol based on the analysis of samples as they are
(non-sedimented samples) has demonstrated the potential
of NIR microscopy to detect MBM at levels as low as 0.5%. In
order to reduce the measuring time and the LOD, the addi-
tion of a sedimentation step before NIR microscopy analysis
was proposed.
50,51,54
The protocol developed is fairly simple
and very similar to the one used in the European official
OM method (Annex VI in EC Regulation 152/2009). Sediment
analysed by the ofcial method is based on the principle of
concentration of the bone fraction using a high-density solvent
(i.e. tetrachloroethylene). In order to perform the discrimina-
tion, Baeten et al. constructed a PLSDA discriminant model
with 8002 spectra scanned from the sediment of 27 meals
of animal origin (including cattle, pig, sheep and poultry) and
seven MBM-free compound feeds.
54
Some 97.5% of the animal
particles were classied as animal, with a 95% condence
interval (twice the standard deviation). The misclassication
ratio for animal particles was 0.064%. Two non-animal parti-
cles were wrongly classied in the animal class (meaning a
ratio of 0.061% misclassification). Some 15 samples, both
non-adulterated and adulterated with MBM at 0.05%, 0.1%,
0.5% and 1%, were used to study the repeatability. All samples
containing MBM were correctly classied as positive and
all samples free of MBM were correctly classied as nega-
tive. Some 48 spiked samples, obtained by mixing 24 feed-
stuffs with eight MBM products at different levels (0.5%, 1%,
1.5%, 2%, 2.5%, 3%, 3.5%, 4%, 4.5%, 5%, 5.5%, 6%, 6.5%, 7%,
7.5% and 8%) and 24 compound feeds samples were used to
study the reliability of the method, including a sedimentation
step on a wide range of MBM-free feeds. As a result, the 48
spiked samples were correctly detected as positive and the
24 blank feeds were correctly detected as negative. The LOD
and repeatability of this discriminant model were validated.
A sample spiked at 0.1% with MBM was analysed 10 times.
All the replicates were detected as positive in terms of the
presence of bones in the sample.
51
The analysis of the sedi-
ment fraction by NIR microscopy is an attractive and powerful
method for detecting low levels of MBM in feedstuffs. With
a protocol focused on the sediment part of the sample, NIR
microscopy can be used to detect MBM in feed at levels as low
as 0.05%.
54
In feed control, a crucial aspect is the transferability of the
developed methods from the developing laboratory to another
laboratory. Within the framework of the STRATFEED project, it
was decided to analyse blind samples in two laboratories that
had not participated in developing the protocol. The samples
were randomly analysed and the results showed that the NIR
microscopy method based on the sediment fraction could be
easily transferred to another laboratory and still return a false
negative value of less than 5%.
51
Detection by decision rules based on visual
observation
Decision rules were developed based on absorbance values
[log(1/R)] at different wavelengths of the spectrum from sedi-
ment particles in order to identify their animal origin. Baeten et
al. proposed a decision rule using absorbance values at three
wavelengths (i.e. 1944 nm, 2060 nm and 2148 nm) from the
rst (d1) and second (d2) derivative spectra to decide whether
particle i belonged to the animal group or not.
51
The wave-
lengths were chosen as representative of the MBM spectra.
The absorption bands at 1944 nm, 2060 nm, and 2148 nm
were associated with water, protein and starch, respectively.
A particle i belonged to the animal group if d1(i,2148) > 0.001,
d1(i,1944) > 0.001, d1(i,1944) < 0.003 and d2(i,2060) < 0.
54
von Holst et al. proposed a modication of these decision
rules for analysing particle spectra from the sediment.
56

The decision rules are based on three criteria: (1) the pres-
ence of maxima in the 19201960 nm, 20302070 nm and
21502200 nm regions; (2) the presence of minima in the
20102030 nm (y1), 20702150 nm (y2) and 22102250 nm (y3)
regions; and (3) fullling the formula [(absorbance around
y1 + absorbance around y3)/2] > absorbance around y2. This
last criterion depends to a great extent, however, on the slope
of the spectrum and this can lead to mathematically classi-
fying as animal a sample that visually is not. In order to solve
this problem, the third criterion can be generalised (even for
non-sediment samples) using the following equation: {absorb-
ance [line segment (y1y3)]
wavelength defined by y2
} > absorbance
(y2) (Figure 4). Based on the decision rules, 20 samples with
different MBM percentages (08%) were analysed and no false
positive or false negative results were observed.
von Holst et al. reported also the results of a transferability
study in which these decision criteria based on the visual
inspection of the NIR spectrum were applied to measure-
ments obtained by three independent laboratories.
56
The three
laboratories [Walloon Agricultural Research Centre (CRA-W),
Institute for Health and Consumer Protection (IHCP), and
Institute for Reference Materials and Measurements (IRMM)]
involved in the transferability study were equipped with a
Fourier transform (FT) near infrared microscope from the
same company (PerkinElmer), but the model and specica-
tion were different; at CRA-W a Spectrum IdentiCheck FT-NIR
system and Auto Image microscopy, at IHCP a FT infrared
220 The Potential of NIR Microscopy to Detect, Identify and Quantify Processed Animal By-Products
Spectrum 2000 system and Auto Image microscopy and at
IRMM a Spectrum One NTS system and Spotlight micros-
copy. Each laboratory analysed the same 20 samples with and
without different levels of MBM (08%). Staff at CRA-W and
IHCP analysed 3500 particle spectra, respectively, and staff at
IRMM scanned about 5500 spectra. All positive and negative
samples were correctly identied at each of the three labora-
tories based on the proposed decision rules.
56
In summary, NIR microscopy is an interesting alternative
method with many advantages, such as being an objective,
reliable, sensitive and highly selective technique. However,
the detection result greatly depends on the homogeneity of
the samples and the number of particles analysed. A large
number of particles need to be analysed when the content of
animal ingredients in an adulterated sample is too low.
Species identication of processed animal
proteins
Several studies have been carried out to assess the potential of
NIR microscopy for species identication (Table 2). This issue
is essential in order to have suitable methods that allow the
lifting of the total ban and permit intra-species recycling (for
example, use of pig meal for feeding poultry). These studies
include analyses on both raw and sediment fractions.
Near infrared microscopy analysis of the raw
fraction
Baeten et al. were the rst to study, on a large scale, the poten-
tial of NIR microscopy methodology for identifying the species
origin of an animal particle.
51,53
In order to achieve this goal,
7492 vegetal particle spectra and 2484 animal particle spectra
(including 1595 terrestrial animal particle spectra, 889 sh
meal particle spectra, 505 poultry by-product particle spectra
and 1090 bovine and pig meal particle spectra) were used to
construct species identication equations using the PLSDA
algorithm (Table 3). Correct classication rates of 95.0%, 94.3%,
95.5%, 90.0%, 84.1% and 90.1% were obtained, respectively, for
the vegetal, animal, terrestrial animal, sh, poultry by-product
and pig/bovine particles. If particle i belonged to the terres-
trial animal group, it should be discriminated correctly by
Equations (1), (2) and (4). If particle i belonged to the sh meal
group, it should be discriminated correctly by Equations (1),
(2), (4) and (5). If particle i belonged to the poultry by-product
or the bovine and pig MBM groups, it should be discrimi-
nated correctly by Equations (1), (2), (3) and (4). This strategy
strengthens condence in the classication.
Ten sh meal samples spiked with MBM at various levels (3%,
6% and 9%) were used to detect terrestrial animal by-products
in sh meal by analysing about 600 particles of each sample.
All the spiked samples were shown to be positive in terms of
the presence of terrestrial animal by-products.
51
However, the
discriminating power of NIR microscopy for the detection of
traces of PAPs at higher taxonomic levels should be proved by
further research.
Prez-Marn et al. developed a methodology using only a
library of animal meal by-products for identifying terrestrial
MBM and sh meal.
57
The training database comprised 11,727
spectra of particles coming from pure terrestrial meal (eight
bovine, four ovine, 40 porcine, and 41 avian samples) and 5843
spectra of particles of 65 shmeal samples. The method was
based on a two-stage strategy. In the rst stage, animal parti-
cles were identied by using either the k-nearest-neighbour
(kNN) method or SIMCA. In the second stage, kNN was used to
discriminate between terrestrial and sh particles. kNN with
second derivative spectra and ve neighbours correctly classi-
ed 98.5% when using cross-validation. Some 20 experimental
mixtures of sh meal spiked with various bovine meals (2.0
16%) were used to validate the method. About 200 particles
per sample were analysed by NIR microscopy. The presence
of terrestrial animal by-products was correctly detected in
20 external validation samples. Furthermore, 20 commercial
compound feeds were used to validate the constructed models.
These samples included pure compound feeds, compound
feeds adulterated with 2.5% or 3.0% shmeal and compound
feeds adulterated with various terrestrial MBM (14.329.4%).
The models worked perfectly for the pure compound feeds;
however in the case of adulterated samples, although all the
terrestrial MBM was detected, more than half of the identi-
fied animal particles were wrongly classified as fish. This
result indicated that the accuracy of identifying sh meal and
terrestrial MBM needs to be improved.
Fumire et al. proposed combining NIR microscopy and
real-time PCR to authenticate, at species level, the presence
of animal particles (see the section Combining near infrared
microscopy and other analytical technologies).
58,59
To imple-
ment this, an identication of the animal particles was done
initially using NIR microscopy. After this analysis, each suspect
particle was individually and carefully transferred in a vial for
1850 1900 1950 2000 2050 2100 2150 2200 2250 2300
0.5
0.55
0.6
0.65
0.7
Wavelength
l
o
g

(
1
/
R
)
y1
y3
y2
Figure 4. Spectral conditions to be fullled by a spectrum from
a particle of animal origin (adapted from von Holst et al.
56
).
Z. Yang et al., J. Near Infrared Spectrosc. 19, 211231 (2011) 221
T
a
b
l
e

2
.

S
p
e
c
i
e
s

i
d
e
n
t
i

c
a
t
i
o
n

o
f

p
r
o
c
e
s
s
e
d

a
n
i
m
a
l

p
r
o
t
e
i
n
s

b
y

n
e
a
r

i
n
f
r
a
r
e
d

m
i
c
r
o
s
c
o
p
y
.
S
t
u
d
y

p
u
r
p
o
s
e
T
r
a
i
n
i
n
g

s
e
t
V
a
l
i
d
a
t
i
o
n

s
e
t
M
o
d
e
l
;

M
e
t
h
o
d
S
u
c
c
e
s
s

r
a
t
e
R
a
w
/
s
e
d
i
m
e
n
t

f
r
a
c
t
i
o
n
I
m
p
o
r
t
a
n
c
e

a
n
d
/
o
r

w
e
a
k
n
e
s
s
e
s
R
e
f
.
T
o

d
e
m
o
n
s
t
r
a
t
e

t
h
e

p
o
t
e
n
t
i
a
l

o
f

N
I
R

m
i
c
r
o
s
c
o
p
y

f
o
r

s
p
e
c
i
e
s

i
d
e
n
-
t
i

c
a
t
i
o
n

o
f

P
A
P
s
7
4
9
2

v
e
g
e
t
a
l

p
a
r
t
i
c
l
e
s

a
n
d

2
4
8
4

a
n
i
m
a
l

p
a
r
t
i
c
l
e
s

(
i
n
c
l
u
d
-
i
n
g

1
5
9
5

t
e
r
r
e
s
t
r
i
a
l

a
n
i
m
a
l

p
a
r
t
i
c
l
e
s
,

8
8
9

s
h
m
e
a
l

p
a
r
t
i
c
l
e
s
,

5
0
5

p
o
u
l
t
r
y

b
y
-
p
r
o
d
u
c
t

p
a
r
t
i
c
l
e
s

a
n
d

1
0
9
0

b
o
v
i
n
e

a
n
d

p
i
g

p
a
r
t
i
c
l
e
s
)
1
0

s
h
m
e
a
l

s
a
m
-
p
l
e
s

s
p
i
k
e
d

w
i
t
h

M
B
M

a
t

v
a
r
i
o
u
s

l
e
v
e
l
s

(
3
%
,

6
%

a
n
d

9
%
)
D
i
s
c
r
i
m
i
n
a
n
t

e
q
u
a
t
i
o
n
s
;

P
L
S
D
A
9
5
.
5
%
,

9
0
.
0
%
,

8
4
.
1
%

a
n
d

9
0
.
1
%

f
o
r

t
h
e
,

t
e
r
r
e
s
t
r
i
a
l

a
n
i
m
a
l
,

s
h
,

p
o
u
l
t
r
y

b
y
-
p
r
o
d
u
c
t

a
n
d

p
i
g
/
b
o
v
i
n
e

p
a
r
t
i
c
l
e
s

(
c
a
l
i
b
r
a
t
i
o
n
)
n
o

f
a
l
s
e

n
e
g
a
t
i
v
e

r
e
s
u
l
t
s

(
v
a
l
i
d
a
t
i
o
n
)
R
a
w
H
i
e
r
a
r
c
h
y

e
q
u
a
t
i
o
n
s

w
e
r
e

e
s
t
a
b
l
i
s
h
e
d
,

w
h
i
c
h

s
t
r
e
n
g
t
h
e
n
s

c
o
n

d
e
n
c
e

i
n

t
h
e

c
l
a
s
s
i

c
a
t
i
o
n
;
T
h
e

s
p
i
k
e
d

s
a
m
p
l
e
s

n
e
e
d
e
d

t
o

h
a
v
e

a

g
o
o
d

l
e
v
e
l

o
f

h
o
m
o
g
e
n
e
i
t
y
;
T
h
e

d
i
s
c
r
i
m
i
n
a
t
i
o
n

p
o
w
e
r

o
f

N
I
R

m
i
c
r
o
s
c
o
p
y

f
o
r

t
h
e

d
e
t
e
c
t
i
o
n

o
f

t
r
a
c
e
s

o
f

P
A
P
s

a
t

a

h
i
g
h
e
r

t
a
x
o
n
o
m
i
c

l
e
v
e
r

s
h
o
u
l
d

b
e

v
a
l
i
d
a
t
e
d

b
y

f
u
r
t
h
e
r

r
e
s
e
a
r
c
h
e
s
5
1
T
o

d
i
s
c
r
i
m
i
n
a
t
e

t
e
r
r
e
s
t
r
i
a
l

M
B
M

f
r
o
m

s
h
m
e
a
l
1
1
,
7
2
7

s
p
e
c
t
r
a

f
r
o
m

t
e
r
r
e
s
t
r
i
a
l

m
e
a
l
s

a
n
d

5
8
4
3

f
r
o
m

s
h
m
e
a
l
2
0

e
x
p
e
r
i
m
e
n
t
a
l

m
i
x
t
u
r
e
s

o
f

s
h
-
m
e
a
l

s
p
i
k
e
d

w
i
t
h

v
a
r
i
o
u
s

b
o
v
i
n
e

m
e
a
l
s

a
t

v
a
r
i
o
u
s

l
e
v
e
l
s

(
2
.
0

1
6
%
)
;
2
0

c
o
m
m
e
r
c
i
a
l

c
o
m
p
o
u
n
d

f
e
e
d
s
,

i
n
c
l
u
d
i
n
g

p
u
r
e

c
o
m
p
o
u
n
d

f
e
e
d
s
,

c
o
m
p
o
u
n
d

f
e
e
d
s

a
d
u
l
t
e
r
a
t
e
d

w
i
t
h

2
.
5
%

o
r

3
.
0
%

s
h
m
e
a
l
,

a
n
d

c
o
m
p
o
u
n
d

f
e
e
d
s

a
d
u
l
t
e
r
a
t
e
d

w
i
t
h

v
a
r
i
o
u
s

t
e
r
-
r
e
s
t
r
i
a
l

M
B
M

a
t

d
i
f
f
e
r
e
n
t

l
e
v
e
l
s

(
1
4
.
3

2
9
.
4
%
)
T
w
o
-
s
t
e
p

s
t
r
a
t
e
g
y
:
(
1
)

a
n
i
m
a
l

p
a
r
t
i
c
l
e
s

a
r
e

i
d
e
n
t
i

e
d

u
s
i
n
g

a

g
l
o
b
a
l

(
S
I
M
C
A
)

o
r

a

l
o
c
a
l

(
k
N
N
)

d
i
s
-
t
a
n
c
e

m
e
a
s
u
r
e
;
(
2
)

t
e
r
r
e
s
t
r
i
a
l

a
n
d

s
h

p
a
r
t
i
c
l
e
s

a
r
e

d
i
s
c
r
i
m
i
n
a
t
e
d

b
y

k
-
n
e
a
r
e
s
t
-
n
e
i
g
h
-
b
o
u
r
s

(
k
N
N
)
9
8
.
5
%

(
c
a
l
i
b
r
a
t
i
o
n
)
N
o

f
a
l
s
e

n
e
g
a
t
i
v
e

r
e
s
u
l
t
s

(
v
a
l
i
d
a
t
i
o
n
,

f
o
r

m
i
x
t
u
r
e
s

o
f

s
h
-
m
e
a
l

s
p
i
k
e
d

w
i
t
h

b
o
v
i
n
e

m
e
a
l
s
)
M
o
r
e

t
h
a
n

h
a
l
f

o
f

t
h
e

i
d
e
n
t
i

e
d

a
n
i
m
a
l

p
a
r
t
i
c
l
e
s

w
e
r
e

w
r
o
n
g
l
y

c
l
a
s
s
i

e
d

a
s

s
h

(
v
a
l
i
d
a
t
i
o
n

f
o
r

m
i
x
t
u
r
e
s

o
f

c
o
m
p
o
u
n
d

f
e
e
d
s

s
p
i
k
e
d

w
i
t
h

t
e
r
r
e
s
t
r
i
a
l
R
a
w
T
h
i
s

s
t
u
d
y

o
n
l
y

u
s
e

a

l
i
b
r
a
r
y

o
f

a
n
i
m
a
l

m
e
a
l

b
y
-
p
r
o
d
u
c
t
s

f
o
r

i
d
e
n
t
i
-
f
y
i
n
g

t
e
r
r
e
s
t
r
i
a
l

M
B
M

a
n
d

s
h
m
e
a
l
;
T
h
i
s

r
e
s
u
l
t

i
n
d
i
c
a
t
e
s

t
h
e

a
c
c
u
r
a
c
y

o
f

i
d
e
n
-
t
i
f
y
i
n
g

s
h
m
e
a
l

a
n
d

t
e
r
r
e
s
t
r
i
a
l

M
B
M

n
e
e
d

t
o

b
e

i
m
p
r
o
v
e
d
5
7
222 The Potential of NIR Microscopy to Detect, Identify and Quantify Processed Animal By-Products
S
t
u
d
y

p
u
r
p
o
s
e
T
r
a
i
n
i
n
g
s
e
t
V
a
l
i
d
a
t
i
o
n

s
e
t
M
o
d
e
l
;

M
e
t
h
o
d
S
u
c
c
e
s
s

r
a
t
e
R
a
w
/
s
e
d
i
m
e
n
t

f
r
a
c
t
i
o
n
I
m
p
o
r
t
a
n
c
e

a
n
d
/
o
r

w
e
a
k
n
e
s
s
e
s
R
e
f
.
T
o

d
i
s
c
r
i
m
i
n
a
t
e

c
a
t
t
l
e

f
r
o
m

p
i
g

p
a
r
t
i
c
l
e
s
7
3
8

p
a
r
t
i
c
l
e
s

(
2
7
8

c
a
t
t
l
e

a
n
d

4
6
0

p
i
g

p
a
r
t
i
c
l
e
s
)
N
A
D
i
s
c
r
i
m
i
n
a
n
t

e
q
u
a
t
i
o
n
s
;
P
L
S
D
A

a
n
d

S
V
M
9
7
.
3
%

f
o
r

S
V
M

a
n
d

9
9
%

f
o
r

P
L
S
D
A

(
c
a
l
i
b
r
a
t
i
o
n
)
R
a
w
W
i
t
h

t
h
e

h
e
l
p

o
f

P
C
R
,

t
h
e

s
p
e
c
i
e
s
-
s
p
e
c
i

c

N
I
R

m
i
c
r
o
s
c
o
p
y

s
p
e
c
t
r
a
l

d
a
t
a
b
a
s
e

c
a
n

b
e

e
s
t
a
b
l
i
s
h
e
d
;
T
h
e

c
o
m
b
i
n
a
t
i
o
n

o
f

N
I
R

m
i
c
r
o
s
c
o
p
y

a
n
d

P
C
R

b
e
n
e

t
s

f
r
o
m

t
h
e

a
d
v
a
n
t
a
g
e
s
,

a
n
d

o
v
e
r
-
c
o
m
e
s

t
h
e

l
i
m
i
t
a
t
i
o
n
s

o
f

b
o
t
h

m
e
t
h
o
d
s
5
8
,
5
9
T
o

i
d
e
n
t
i
f
y

m
a
m
m
a
l
i
a
n

M
B
M

a
n
d

s
h
m
e
a
l
1
3
8
0

p
a
r
t
i
c
l
e
s

w
i
t
h

4
2
0

s
h
m
e
a
l

p
a
r
t
i
c
l
e
s
T
h
e

v
a
l
i
d
a
t
i
o
n

s
e
t

i
n
c
l
u
d
e
d

6
3
0

p
a
r
-
t
i
c
l
e
s

w
i
t
h

9
0

s
h
-
m
e
a
l

p
a
r
t
i
c
l
e
D
i
s
c
r
i
m
i
n
a
n
t

e
q
u
a
t
i
o
n
s
;
S
V
M
1
0
0
%

(
c
a
l
i
b
r
a
t
i
o
n
)
9
5
%

(
l
e
a
v
e
-
o
n
e
-
o
u
t

c
r
o
s
s

v
a
l
i
d
a
t
i
o
n
)
9
5
.
5
%

(
v
a
l
i
d
a
t
i
o
n
)
S
e
d
i
m
e
n
t
F
u
r
t
h
e
r

d
e
v
e
l
o
p
m
e
n
t
s

w
i
l
l

f
o
c
u
s

o
n

e
x
t
e
n
d
i
n
g

l
i
b
r
a
r
i
e
s

w
i
t
h

m
o
r
e

s
h

a
n
d

M
M
B
M

p
a
r
t
i
c
l
e
s
;
T
h
i
s

r
e
s
u
l
t

i
n
d
i
c
a
t
e
s

t
h
e

a
c
c
u
r
a
c
y

o
f

i
d
e
n
t
i
f
y
i
n
g

s
h
m
e
a
l

a
n
d

m
a
m
m
a
l
i
a
n

M
B
M

n
e
e
d

t
o

b
e

i
m
p
r
o
v
e
d
6
0
N
A
,

n
o
t

a
v
a
i
l
a
b
l
e
.
T
a
b
l
e

2
.

S
p
e
c
i
e
s

i
d
e
n
t
i

c
a
t
i
o
n

o
f

p
r
o
c
e
s
s
e
d

a
n
i
m
a
l

p
r
o
t
e
i
n
s

b
y

n
e
a
r

i
n
f
r
a
r
e
d

m
i
c
r
o
s
c
o
p
y
.
Z. Yang et al., J. Near Infrared Spectrosc. 19, 211231 (2011) 223
PCR analysis, and its origin (animal or vegetal) was predicted
using existing discrimination models. The NIR microscopy
spectra of 738 particles (278 cattle and 460 pig particles) were
obtained. After the NIR microscopy analysis, the MBM parti-
cles were recovered and analysed by real-time PCR to conrm
their species origin. Only the spectra corresponding to parti-
cles giving a clear PCR result were implemented in the spec-
tral databases used to develop discriminant models. Support
vector machines (SVMs) and PLSDA models were constructed
using the training dataset in order to discriminate between
cattle and pig particles. When the models were applied to
an independent dataset, both procedures correctly classied
more than 97% of the samples. SVM was correct for 97.3% of
the samples and PLSDA was correct for 99% of the samples.
The combination of NIR microscopy and PCR benets from the
advantages, and overcomes the limitations, of both methods.
The combination of the PCR with the NIR microscopy technique
can overcome the interference with DNA present in some
allowed animal proteins in ruminant nutrition (for example,
milk) or other animal products (for example, fat, blood).
Near infrared microscopy analysis on the sediment
fraction
De la Haba et al. investigated the ability of the NIR microscopy
method to discriminate bone particles of sh origin and those
of terrestrial animal origin in the sediment fraction.
60
A total of
2010 spectra from sediment particles were collected, including
510 sh meal particle spectra. The spectra were split into two
sub-sets, a calibration set and a validation set. The calibra-
tion set included 1380 particle spectra with 420 fish meal
particle spectra and the validation set included 630 particle
spectra with 90 sh meal particle spectra. Two discriminant
equations [sh vs non-sh and mammalian MBM (MMBM)
vs non-MMBM] were established using the SVM algorithm.
61

For the calibration set, the correct classication success rate
was 100%. For the leave-one-out cross-validation and for the
validation set, success rates of 95% and 95.5% were obtained,
respectively. The SVM models were applied to other samples,
such as pure sh meal containing 0.1% MMBM, compound
feed containing 0.1% sh meal and 2% MMBM and compound
feed containing 0.1% MMBM and 1% sh meal. In all cases,
the SVM equation was able to detect the presence of fish
meal and MMBM. For three validation samples, the numbers
of analysed particles were 344, 332 and 493, respectively, and
the number of detected mammalian MBM particles were 152,
5 and 30, respectively. Compared with the MMBM theoretical
content, the percentage of MMBM particles detected was too
high.
As previously stated, species identication by NIR micro-
scopy is feasible, especially for terrestrial vs shmeal, the
premise for this conclusion being that an authentic species-
specic MIRM spectral database has been established.
Obtaining authentic samples of different origin is an impor-
tant and difficult issue. As explained, the confirmation of
species of the samples to include has to be done by comple-
mentary techniques, such as PCR. Most papers focus on the
E
q
u
a
t
i
o
n
G
r
o
u
p
s
N

D
i
s
c
r
i
m
i
n
a
t
i
o
n
V
e
g
e
t
a
l
A
n
i
m
a
l
T
e
r
r
e
s
t
r
i
a
l

a
n
i
m
a
l
F
i
s
h
P
o
u
l
t
r
y
B
o
v
i
n
e

a
n
d

p
i
g
1
V
e
g
e
t
a
l
/
a
n
i
m
a
l

2
T
e
r
r
e
s
t
r
i
a
l

a
n
i
m
a
l
/

s
h

3
(
B
o
v
i
n
e

a
n
d

p
i
g
)
/
p
o
u
l
t
r
y

4
V
e
g
e
t
a
l
/
t
e
r
r
e
s
t
r
i
a
l

a
n
i
m
a
l

5
(
V
e
g
e
t
a
l

a
n
d

t
e
r
r
e
s
t
r
i
a
l
)
/

s
h

N
u
m
b
e
r

e
q
u
a
t
i
o
n
s

u
s
e
d
2
3
3
4
4
4
D
a
t
a

a
r
e

t
a
k
e
n

f
r
o
m

B
a
e
t
e
n

e
t

a
l
.
5
0

a
n
d

D
e

L
a

R
o
s
a

D
e
l
g
a
d
o

e
t

a
l
.
5
5
T
a
b
l
e

3
.

D
e
s
c
r
i
p
t
i
o
n

o
f

t
h
e

e
q
u
a
t
i
o
n
s

c
o
n
s
t
r
u
c
t
e
d

f
o
r

s
p
e
c
i
e
s

i
d
e
n
t
i

c
a
t
i
o
n
.
224 The Potential of NIR Microscopy to Detect, Identify and Quantify Processed Animal By-Products
discrimination of terrestrial from shmeal. The discrimination
power of NIR microscopy for the identification of PAPs at
higher taxonomic levels (for example, mammalian vs non-
mammalian, or ruminant vs non-ruminant) should be demon-
strated or validated by further research.
Quantication of processed animal proteins
Only feasibility studies on the application of NIR microscopy to
quantify PAPs have been reported over the last few years. NIR
microscopy applications in quantication of PAPs are summa-
rised in Table 4.
Analysis on the raw fraction
The rst attempt to use NIR microscopy to quantify the pres-
ence of PAPs was made by Piraux and Dardenne.
52
These
authors applied a regression model between the proportion of
MBM in compound feed and the area proportion of MBM parti-
cles. Image analysis was used to measure the area proportion
of MBM particles. Non-adulterated compound feeds and MBM
mixed in at weight proportions of 0%, 2%, 4%, 6%, 8% and 10%
were used to build the calibration model. Three independent
validation samples containing 2%, 6% and 6% MBM were
prepared by another laboratory and used to test the model
constructed. Using NIR microscopy, the estimated results
were 2.02%, 4.57% and 3.36%, respectively. The results,
although they were not sufciently accurate for quantitative
control, were promising. In this study, quantication based on
the area of PAPs, does not take into account the difference
in density of the different types of particles. Further develop-
ments or new methods are required to improve the accuracy
of quantication.
Baeten et al. conducted a large study to assess the poten-
tial of NIR microscopy for the quantication of PAPs.
51
In this
study, the performance of the method in quantifying MBM in
sh meal was tested. To do this, the raw fraction of several
sh meals was analysed and the quantication results were
obtained by calculating the percentage of particles classied
in the animal group over the total number of particles classi-
ed in one of the groups. For 10 sh meal samples spiked with
MBM at various levels (3%, 6% and 9%), there was an R
2
of 0.65
between the reference values and the percentage estimated
by NIR microscopy. The results showed the potential of the
method for the quantication of MBM in shmeal. However,
the accuracy of the method needs to be drastically improved.
Prez-Marn et al. developed a method using a library of
animal meal by-products only for quantifying terrestrial MBM
in sh meal.
57
The method was based on a two-stage strategy.
In the rst stage, animal particles were indentied using a
global or local distance measure. In the second stage, kNN
was used to discriminate terrestrial and sh particles. The
proportion of terrestrial MBM in fish meal was calculated
by the ratio of the number of terrestrial particles to the total
number of particles. Then 20 experimental mixtures of sh
meal spiked with various bovine meals (2.516%) were used
to validate the method. About 200 particles per sample were
analysed by NIR microscopy. The percentage of particles clas-
sied as terrestrial showed broad agreement with the true
sample composition, but the accuracy was too low to be used
as a criterion for quantication (an R
2
of 0.42 between the
reference values and the percentage estimated by NIR micro-
scopy).
More recently, Abbas et al. proposed the key parameters
needed to develop a quantitative NIR microscopy approach in
order to meet the new European legislation requirements.
62

The proposed parameters included a no-sample prepara-
tion, an optimisation of parameters, the use of both the gross
(>250 m) and ne (<250 m) fractions of samples, the use of
the transmission mode to analyse the ne raw fraction and
the reection mode for the gross raw fraction. Further work is
needed to develop an accurate quantication method based on
these recommendations.
All these results show that the proposed protocols based on
the NIR microscopy technique for quantication can be used,
at best, as a semi-quantitative method. For accurate quanti-
cation, however, there is still much work to do.
Analysis on the sediment fraction
The possibility of quantifying PAPs in a compound feed by
analysing the sediment fraction has been also investigated.
51,54

In order to obtain the sediment fraction, the protocol used in
EC Regulation 152/2009, and also valid for OM, was used.
Some 48 validation samples were prepared using 24 compound
feeds (destined for cattle, goats, poultry, pigs and rabbits)
spiked at different percentages (0.5% to 8%, w/w in 0.5% inter-
vals) of eight PAPs from different sources in term of species
origin and rendering process. The estimated percentage of
animal ingredients in spiked compound feeds was calculated
as follows: animal ingredients in the sample estimated by
NIR microscopy (%) = [weight bones in the sample estimated
by NIR microscopy (g)/% bones in the animal feed ingredient
used to spike the sample (= f factor)].
For 48 samples spiked with eight PAPs, the coefcient of
determination (R
2
) between reference and estimated values
was about 0.57. The R
2
for individual PAPs ranged from 0.64 to
0.97. These values show that the quantication of PAPs using
the NIR microscopy protocol based on the sediment fraction
is highly dependent on the PAPs source, as observed in the
quantication protocol based on OM.
51
It should be stressed
that usually the percentage of bones in the PAPs (i.e. f factor)
is not known for unknown contamination samples.
In order to quantify by NIR microscopy the banned PAPs in
the sediment fraction of feedstuffs, two protocols to obtain the
sediment were tested. The rst uses only tetrachloroethylene
to prepare a sediment fraction that has a density higher than
1.62 g mL
1
.
The second method (also called the French method) uses
two solvents, tetrachloroethylene and tetrabromoethane, to
prepare a sediment fraction with a density between 1.62 and
2.2 units. Fifteen samples adulterated with PAPs at levels of
0%, 0.05%, 0.1%, 0.5% and 1% were used to study the quan-
tication. The percentage of animal ingredients in the feeds,
calculated by NIR microscopy, ranged from 0.186% to 1.508%
Z. Yang et al., J. Near Infrared Spectrosc. 19, 211231 (2011) 225
S
t
u
d
y

p
u
r
p
o
s
e
T
r
a
i
n
i
n
g

s
e
t
V
a
l
i
d
a
t
i
o
n

s
e
t
M
o
d
e
l
;

M
e
t
h
o
d
P
e
r
f
o
r
m
a
n
c
e
R
a
w
/
s
e
d
i
m
e
n
t

f
r
a
c
t
i
o
n
I
m
p
o
r
t
a
n
c
e

a
n
d
/
o
r

w
e
a
k
n
e
s
s
e
s
R
e
f
.
T
o

d
e
m
o
n
s
t
r
a
t
e

t
h
e

p
o
t
e
n
t
i
a
l

o
f

N
I
R

m
i
c
r
o
s
c
o
p
y

f
o
r

q
u
a
n
t
i
f
y
i
n
g

M
B
M

c
o
n
t
e
n
t
s

i
n

c
o
m
p
o
u
n
d

f
e
e
d
s
7
8
0

p
a
r
t
i
c
l
e
s

f
r
o
m

5
6

a
l
l
o
w
e
d

r
a
w

m
a
t
e
r
i
a
l
s
;

3
7
9

p
a
r
t
i
c
l
e
s

f
r
o
m

4
3

f
o
r
b
i
d
d
e
n

a
n
i
m
a
l

m
e
a
l

p
r
o
d
u
c
t
s
1
0

c
o
m
p
o
u
n
d

f
e
e
d

a
d
u
l
t
e
r
a
t
e
d

w
i
t
h

d
i
f
-
f
e
r
e
n
t

M
B
M

a
t

l
e
v
e
l
s

o
f

0
%
,

2
%
.

4
%
,

6
%
,

8
%

a
n
d

1
0
%
;
t
h
r
e
e

c
o
m
p
o
u
n
d

f
e
e
d

a
d
u
l
t
e
r
a
t
e
d

w
i
t
h

2
%
,

6
%

a
n
d

6
%

M
B
M
.
R
e
g
r
e
s
s
i
o
n

m
o
d
e
l

b
e
t
w
e
e
n

t
h
e

p
r
o
-
p
o
r
t
i
o
n

o
f

M
B
M

i
n

c
o
m
p
o
u
n
d

f
e
e
d

a
n
d

t
h
e

a
r
e
a

p
r
o
p
o
r
t
i
o
n

o
f

M
B
M

p
a
r
t
i
c
l
e
s
y

=

0
.
6
1
x
,

R
2

=

0
.
8
6
(
c
a
l
i
b
r
a
t
i
o
n
)
2
.
0
2
%
,

4
.
5
7
%

a
n
d

3
.
3
6
%

f
o
r

2
%
,

6
%

a
n
d

6
%
,

r
e
s
p
e
c
t
i
v
e
l
y
.

(
v
a
l
i
d
a
t
i
o
n
)
R
a
w
F
u
r
t
h
e
r

a
n
a
l
y
s
i
s

a
r
e

r
e
q
u
i
r
e
d

t
o

i
m
p
r
o
v
e

t
h
e

a
c
c
u
r
a
c
y
;
Q
u
a
n
t
i

c
a
t
i
o
n

b
a
s
e
d

o
n

t
h
e

a
r
e
a
,

d
o
e
s

n
o
t

c
o
n
s
i
d
e
r

t
h
e

d
i
f
f
e
r
e
n
t

d
e
n
s
i
t
y

o
f

d
i
f
f
e
r
e
n
t

p
a
r
t
i
c
l
e
s
5
2
T
o

q
u
a
n
t
i
f
y

M
B
M

c
o
n
t
e
n
t

i
n

s
h
-
m
e
a
l
N
A
1
0

s
h

m
e
a
l

s
a
m
p
l
e
s

s
p
i
k
e
d

w
i
t
h

M
B
M

a
t

v
a
r
i
o
u
s

l
e
v
e
l
s

(
3
%
,

6
%

a
n
d

9
%
)
C
a
l
c
u
l
a
t
i
o
n

o
f

t
h
e

p
e
r
c
e
n
t
a
g
e

o
f

p
a
r
-
t
i
c
l
e
s

c
l
a
s
s
i

e
d

i
n

t
h
e

a
n
i
m
a
l

g
r
o
u
p

o
v
e
r

t
h
e

t
o
t
a
l

n
u
m
-
b
e
r

o
f

p
a
r
t
i
c
l
e
s

c
l
a
s
s
i

e
d
R
2

=

0
.
6
5
(
v
a
l
i
d
a
t
i
o
n
)
R
a
w
T
h
i
s

m
e
t
h
o
d

o
f

q
u
a
n
-
t
i

c
a
t
i
o
n

c
a
n

s
p
o
t

t
h
e

t
r
e
n
d
s

o
f

M
B
M

c
o
n
-
t
e
n
t
s

i
n

s
h
m
e
a
l
,

b
u
t

t
h
e

a
c
c
u
r
a
c
y

n
e
e
d

t
o

b
e

i
m
p
r
o
v
e
d
;
Q
u
a
n
t
i
t
a
t
i
v
e

a
n
a
l
y
s
i
s

w
a
s

p
e
r
f
o
r
m
e
d

b
y

c
a
l
c
u
l
a
t
i
n
g

r
a
t
i
o
s

o
f

n
u
m
b
e
r
s

o
f

p
a
r
t
i
c
l
e
s
.

T
h
e

s
i
z
e
s

a
n
d

d
e
n
s
i
-
t
i
e
s

o
f

p
a
r
t
i
c
l
e
s

f
r
o
m

d
i
f
f
e
r
e
n
t

i
n
g
r
e
d
i
e
n
t
s

a
r
e

n
o
t

c
o
n
s
i
d
e
r
e
d
5
1
T
o

q
u
a
n
t
i
f
y

t
e
r
r
e
s
t
r
i
a
l

M
B
M

c
o
n
t
e
n
t

i
n

s
h
-
m
e
a
l
1
1
,
7
2
7

s
p
e
c
t
r
a

o
f

p
u
r
e

t
e
r
r
e
s
t
r
i
a
l

m
e
a
l
s

a
n
d

5
8
4
3

o
f

s
h
m
e
a
l
2
0

e
x
p
e
r
i
m
e
n
t
a
l

m
i
x
-
t
u
r
e
s

o
f

s
h

m
e
a
l

s
p
i
k
e
d

w
i
t
h

v
a
r
i
o
u
s

b
o
v
i
n
e

m
e
a
l
s

a
t

v
a
r
i
o
u
s

l
e
v
e
l
s

(
2
.
0

1
6
%
)
R
a
t
i
o

o
f

t
h
e

n
u
m
-
b
e
r

o
f

t
e
r
r
e
s
t
r
i
a
l

p
a
r
t
i
c
l
e
s

t
o

t
h
e

t
o
t
a
l

n
u
m
b
e
r

o
f

p
a
r
t
i
c
l
e
s
R
2

=

0
.
4
2
(
v
a
l
i
d
a
t
i
o
n
)
R
a
w
T
h
e

a
c
c
u
r
a
c
y

o
f

i
d
e
n
-
t
i
f
y
i
n
g

s
h

m
e
a
l

a
n
d

t
e
r
r
e
s
t
r
i
a
l

M
B
M

n
e
e
d

t
o

b
e

i
m
p
r
o
v
e
d
;
Q
u
a
n
t
i
t
a
t
i
v
e

a
n
a
l
y
s
i
s

w
a
s

p
e
r
f
o
r
m
e
d

b
y

c
a
l
c
u
l
a
t
i
n
g

r
a
t
i
o
s

o
f

n
u
m
b
e
r
s

o
f

p
a
r
t
i
c
l
e
s
.

T
h
e

s
i
z
e
s

a
n
d

d
e
n
s
i
-
t
i
e
s

o
f

p
a
r
t
i
c
l
e
s

f
r
o
m

d
i
f
f
e
r
e
n
t

i
n
g
r
e
d
i
e
n
t
s

a
r
e

n
o
t

c
o
n
s
i
d
e
r
e
d
5
7
T
a
b
l
e

4
.

Q
u
a
n
t
i

c
a
t
i
o
n

o
f

p
r
o
c
e
s
s
e
d

a
n
i
m
a
l

p
r
o
t
e
i
n
s

b
y

n
e
a
r

i
n
f
r
a
r
e
d

m
i
c
r
o
s
c
o
p
y
.
226 The Potential of NIR Microscopy to Detect, Identify and Quantify Processed Animal By-Products
for the one-solvent method and from 0.089% to 1.018% for the
two-solvent method. High R
2
values of 0.93 and 0.99 between
the reference values and the percentage estimated by NIR
microscopy were obtained for the one-solvent and two-solvent
methods, respectively. Compared with the true PAP content,
the two-solvent method gave a more precise result than the
one-solvent method,
54
probably due to the fact that, in the two-
solvent method, the sediment percentage was highly corre-
lated to the percentage of MBM (R
2
= 0.97). However, when
the two-solvent method was used to prepare the sediment
fraction for OM analysis, bone fragments might be overlooked
and a higher number of false negative results might appear.
15
Compared with quantification based on the sediment
fraction, quantification based on the raw fraction is more
promising. However, the published papers do not consider the
size and the density of the particles. These variables should be
taken into account in further studies.
Comparison of the near infrared microscopy
and optical microscopy methods
Compared with OM, the NIR microscopy method does not
require an experienced analyst to conduct the analyses. When
using NIR microscopy, the subjective judgment of the analyst is
replaced by a particle-specic spectrum that can be identied
by calibration equations or decision rules. Baeten et al. used
17 samples to compare the results of NIR microscopy with
those from OM.
54
Table 5 summarises the results of these anal-
yses. All the samples were shown to be positive. For the bone
weight in the sediment obtained by the OM and NIR micros-
copy methods, the correlative analyses were conducted and the
determination coefcient R
2
was 0.87. In addition, the Student
t-test ( = 0.05) showed that there was no signicant difference
between the two results. The NIR microscopy method does not
underestimate or overestimate bone content in the sediment
fraction. These results demonstrate that the NIR microscopy
method can give results equivalent to those obtained by OM,
the ofcial European Union method. Compared to OM, the NIR
microscopy method does not depend on the skills of the analyst
and the whole process can be automated.
Further developments in the
near infrared microscopy
methodology
Combining near infrared microscopy and
other analytical technologies
As the NIR microscopy methodology is non-destructive, the
particles classied as being of animal origin can be selected and
subjected to another analytical procedure. An analytical proce-
dure combining NIR microscopy and PCR protocols was recently
proposed.
59,63
This involved spreading animal PAP particles on
a spectralon plate and analysing them by NIR microscopy to
identify their plant or animal origin and to give a rough idea of
S
t
u
d
y

p
u
r
p
o
s
e
T
r
a
i
n
i
n
g

s
e
t
V
a
l
i
d
a
t
i
o
n

s
e
t
M
o
d
e
l
;

M
e
t
h
o
d
P
e
r
f
o
r
m
a
n
c
e
R
a
w
/
s
e
d
i
m
e
n
t

f
r
a
c
t
i
o
n
I
m
p
o
r
t
a
n
c
e

a
n
d
/
o
r

w
e
a
k
n
e
s
s
e
s
R
e
f
.
T
o

q
u
a
n
t
i
f
y

M
B
M

c
o
n
t
e
n
t
s

i
n

c
o
m
-
p
o
u
n
d

f
e
e
d
s
N
A
4
8

v
a
l
i
d
a
t
i
o
n

s
a
m
p
l
e
s

w
e
r
e

p
r
e
p
a
r
e
d

u
s
i
n
g

2
4

c
o
m
p
o
u
n
d

f
e
e
d
s

(
d
e
s
t
i
n
e
d

f
o
r

c
a
t
t
l
e
,

g
o
a
t
s
,

p
o
u
l
t
r
y
,

p
i
g
s

a
n
d

r
a
b
b
i
t
s
)

s
p
i
k
e
d

a
t

d
i
f
f
e
r
e
n
t

p
e
r
c
e
n
t
a
g
e
s

(
0
.
5

8
%
,

w
/
w

i
n

0
.
5
%

i
n
t
e
r
v
a
l
s
)

o
f

e
i
g
h
t

P
A
P
s

f
r
o
m

d
i
f
f
e
r
e
n
t

s
o
u
r
c
e
s

i
n

t
e
r
m

o
f

s
p
e
c
i
e
s

o
r
i
g
i
n

a
n
d

r
e
n
d
e
r
i
n
g

p
r
o
c
e
s
s
A
n
i
m
a
l

i
n
g
r
e
d
i
-
e
n
t
s

i
n

t
h
e

s
a
m
p
l
e

e
s
t
i
m
a
t
e
d

b
y

N
I
R

m
i
c
r
o
s
c
o
p
y

(
%
)

=

[
w
e
i
g
h
t

b
o
n
e
s

i
n

t
h
e

s
a
m
p
l
e

e
s
t
i
m
a
t
e
d

b
y

N
I
R

m
i
c
r
o
s
c
o
p
y

(
g
)
/
%

b
o
n
e
s

i
n

t
h
e

a
n
i
m
a
l

f
e
e
d

i
n
g
r
e
d
i
e
n
t

u
s
e
d

t
o

s
p
i
k
e

t
h
e

s
a
m
p
l
e

(
=

f

f
a
c
t
o
r
)
]
R
2

=

0
.
5
7
(
v
a
l
i
d
a
t
i
o
n
)
S
e
d
i
m
e
n
t
T
h
e

q
u
a
n
t
i

c
a
t
i
o
n

o
f

P
A
P
s

u
s
i
n
g

t
h
e

N
I
R

m
i
c
r
o
s
c
o
p
y

p
r
o
t
o
c
o
l

b
a
s
e
d

o
n

t
h
e

s
e
d
i
m
e
n
t

f
r
a
c
t
i
o
n

i
s

h
i
g
h
l
y

d
e
p
e
n
d
e
n
t

o
n

t
h
e

P
A
P
s

s
o
u
r
c
e
;
T
h
e

f

f
a
c
t
o
r

(
i
.
e
.

t
h
e

p
e
r
c
e
n
t
a
g
e

o
f

b
o
n
e
s

i
n

t
h
e

P
A
P
s
)

i
s

d
i
f

c
u
l
t

t
o

o
b
t
a
i
n

f
o
r

u
n
k
n
o
w
n

s
a
m
p
l
e
s
5
1
,
5
4
N
A
,

n
o
t

a
v
a
i
l
a
b
l
e
.
T
a
b
l
e

4

(
c
o
n
t
i
n
u
e
d
)
.

Q
u
a
n
t
i

c
a
t
i
o
n

o
f

p
r
o
c
e
s
s
e
d

a
n
i
m
a
l

p
r
o
t
e
i
n
s

b
y

n
e
a
r

i
n
f
r
a
r
e
d

m
i
c
r
o
s
c
o
p
y
.
Z. Yang et al., J. Near Infrared Spectrosc. 19, 211231 (2011) 227
species origin. The particles classied as animal particles were
then recovered and put individually into wells of the PCR plate.
Real-time PCR protocols adapted for single-particle analysis
and targeting short mitochondrial DNA fragments identied the
species origin of these particles. The combination of NIR micro-
scopy and PCR in a single procedure is an elegant solution that
benets from the advantages, and overcomes the limitations, of
both methods. Furthermore, this combination can provide a new
strategy to establish authentic species-specic NIR microscopy
spectral databases. It is a time-consuming procedure, however;
the limitation of this combination of methods is essentially the
number of particles that need to be manually analysed by NIR
microscopy and isolated prior to analysis of their DNA.
Mapping and imaging
For analysing samples spiked with PAPs at levels as low as
0.1%, at least 3000 particles should be scanned. This would
take several days to scan, which is unrealistic. Using a
mapping system coupled with NIR microscopy or new tech-
niques such as hyperspectral imaging make it possible to
overcome this limitation.
61,64,65,66
By using the NIR microscopy
mapping option, a quantity of the feed sample is mapped by
scanning an area that is divided into equally spaced points in
both X and Y directions, based on a scan interval adapted to
the size of the fraction (<250 m). Spectra are automatically
collected at these points. Animal origin detection can then be
done, as explained earlier, using classication equations or
decision rules. Quantication can be also done, as explained in
the section Quantication of processed animal proteins.
Hyperspectral images are three-dimensional datasets
containing light intensity measurements where two dimensions
(x and y) represent spatial distances and a third dimension ()
represents spectral variation such as wavelength. They can be
interpreted as stacks of, typically, hundreds of two-dimensional
spatial images at different wavelengths, or tens of thousands
of spectra aligned in rows and columns. Three instrumentation
approaches are used to obtain hyperspectral images, known as
point, line or plane scans depending on the orientation of the scan-
ning dimension relative to the two-dimensional spatial sample
axes. NIR hyperspectral imaging systems have been proposed for
successfully detecting MBM in feed in combination with decision
rules or chemometric discriminant equations.
49,61,64,65
The recent
success of NIR hyperspectral imaging can be attributed to a
combination of various factors: (1) a non-destructive method, (2)
digitally tuneable infrared optical lters and (3) a drastic increase
in both the speed and capacity of laboratory computing platforms.
A framework for developing and validating an NIR hyperspectral
imaging method as a standard protocol has been proposed.
66
This
involves using various criteria and tests to assess LOD, repeat-
ability and risk of cross-contamination and to validate the NIR
hyperspectral imaging method for detecting PAPs in compound
feed in line with ISO 17025.
Conclusions
The NIR microscopy method combines the analytical advan-
tages of microscopy and spectroscopy techniques. The
Theoretical MBM (%) Results from NIR microscopy
analysis (%)
Results from OM analysis (%)
0.5 0.38 0.51
1 0.75 1.10
1.5 1.48 1.46
2 4.03 4.56
2.5 3.01 3.62
3 5.99 4.06
3.5 2.63 3.62
4 6.79 5.60
4 4.79 5.50
4.5 4.03 3.32
5 5.55 3.49
5.5 4.41 3.59
6 2.96 4.84
6.5 5.40 8.99
7 13.84 10.53
7.5 7.25 7.55
8 10.29 8.28
Table 5. Results from near infrared microscopy and optical microscopy analyses.
55
228 The Potential of NIR Microscopy to Detect, Identify and Quantify Processed Animal By-Products
principle of NIR microscopy analysis is similar to the OM
method for detecting MBM in feed. Current research shows
that NIR microscopy is an objective, sensitive and highly
selective technique for detecting MBM in compound feed.
Compared with the reference method (i.e. OM), the main
advantages of NIR microscopy are that it is non-destruc-
tive, can be automated and does not require experienced
analysts.
The evolution of the NIR microscopy methodology has
been summarised in this paper, including detection based on
discriminant models and decision rules, species identication,
quantication, comparison with OM, combination with other
techniques such as PCR, transferability and further develop-
ments in the area. Since 2006, NIR microscopy has been used
for routine analysis within the framework of the activities of
the Community Reference Laboratory for Animal Proteins in
Feedstuffs (CRL-AP, www.crl.cra.wallonie.be). These analyses
are performed in line with ISO 17025.
NIR microscopy is an interesting alternative technique for
the detection of PAPs. However, the detection greatly depends
on the degree of homogeneity of the sample and on the number
of particles analysed. When the content of animal ingredients
in an adulterated sample is too low, a large number of parti-
cles should be analysed. Increasing the number of particles
analysed by unit of time is a critical issue for the development
of NIR microscopy for the detection of PAPs in feeds. The use
of NIR microscopy coupled with a mapping system or the use
of hyperspectral imaging could be promising solutions.
The European extended ban (preventing cannibalism) will
be lifted if reliable analytical methods for the species specic
identification of PAPs in feed become available. As noted
in this paper, NIR microscopy has shown potential for the
species classication of PAPs of various origins, especially the
discrimination of terrestrial and sh meal. Further research is
needed, however, with regard to the discrimination of higher
taxonomic levels, such as poultry, mammals and even rumi-
nants and pigs. Efforts also have to be made to accelerate
the combined NIR microscopy/PCR method and establish the
species-specic NIR microscopy spectral database in order
to be able to test a large number of species simultaneously.
Combining the NIR microscopy method with other protocols
that allow species-specic detection (for example, analysis of
species-specic proteins by mass spectrometry) needs to be
tested and encouraged.
The potential of the NIR microscopy method to quantify
MBM concentration in feed could not yet support the eventual
introduction of a tolerance level in the feed ban. Currently, for
the raw fraction, the quantitative analysis was performed by
simply calculating ratios of numbers of particles. The sizes
and densities of particles from different ingredients are not
the same and this could be the main reason for the difference
between theoretical value and NIR microscopy estimated value.
For the sediment fraction, the f factor (i.e. the percentage of
bones in the PAPs) is an additional constraint in setting up
a quantitative protocol because it is difcult to obtain it for
unknown samples. Compared to the quantication protocol
based on the analysis sediment fraction, the protocol based
on the raw fraction is more promising. However, the size and
density of particles have not been considered in the published
studies. Further work is needed to develop an accurate quan-
tication method based on NIR microscopy.
Acknowledgements
We thank the European Commission, through the Sixth
Framework Programme (under the Integrati ng and
Strengthening the European Research Area Specic Targeted
Project) within the SAFEED-PAP project (FOOD-CT-2006-
036221) (http://safeedpap.feedsafety.org/), for funding
this work. We also thank the nancial support provided by
Natural Science Foundation of China (Project No. 31072062),
the Program of International S&T Cooperation (Project No.
2010DFA34540) and the Convention WBI/MOST 2010-2011
(Project 6). The information contained in this paper reects
the authors views; the European Commission is not liable for
any use of the information contained therein.
References
1. European Union, Regulation (EC) No. 1774/2002 lay-
ing down health rules concerning animal by-products
not intended for human consumption, Off. J. Eur. Comm.
L273, 1 (2002).
2. European Union, Directive 98/67/EC amending
Directives 80/511/EEC, 82/475/EEC, 91/357/EEC and
Council Directive 96/25/EC and repealing Directive
92/87/EEC, Off. J. Eur. Comm. L261, 27 (1998).
3. G. Horn, Review of the Origin of BSE. Department for
Environment, Food and Rural Affairs (DEFRA), London,
UK (2001).
4. M.J. Prince, J.A. Bailey, P.R. Barrowman, K.J. Bishop, G.R.
Campbell and J.M. Wood, Bovine spongiform encepha-
lopathy, Rev. Sci. Tech. Off. Int. Epiz. 22(1), 37 (2003).
5. Anonymous, The Bovine Spongiform Encephalopathy
(BSE) Inquiry, 2000. Available at: www.bseinquiry.gov.uk
(accessed on 30 January 2003).
6. European Union, Decision 94/381/EC concerning cer-
tain protection measures with regard to bovine spongi-
form encephalopathy and the feeding of mammalian
derived protein, Off. J. Eur. Comm. L172, 23 (1994).
7. European Union, Regulation (EC) No 999/2001 laying
down rules for the prevention, control and eradication of
certain transmissible spongiform encephalopathies, Off.
J. Eur. Comm. L147, 6 (2001).
8. European Union, Regulation (EC) No 1292/2005 amend-
ing Annex IV to Regulation (EC) No 999/2001 of the
European Parliament and of the Council as regards ani-
mal nutrition, Off. J. Eur. Comm. L205, 5 (2005).
9. European Union, Regulation (EC) No 956/2008 amend-
ing Annex IV to Regulation (EC) No 999/2001 of the
Z. Yang et al., J. Near Infrared Spectrosc. 19, 211231 (2011) 229
European Parliament and of the Council laying down
rules for the prevention, control and eradication of cer-
tain transmissible spongiform encephalopathies, Off. J.
Eur. Comm. L260, 8 (2008).
10. N. Shinodo, T. Kusama, T. Yoshida, T. Sugiura, K.
Kadowaki, T. Onodera and K. Sugiura, Developing PCR
primers using a new computer program for detection of
multiple animal-derived materials in feed, J. Food Prot.
71, 2257 (2008).
11. Chinese Regulation, Prohibiting feeding of processed
animal proteins to ruminants, Ministry of Agriculture 7, 1
(2001).
12. Chinese Regulation, Rules for management of feed
products of animal origin, Ministry of Agriculture 40, 1
(2004).
13. G. Gizzi, L.W.D. van Raamsdonk, V. Baeten, I. Murray, G.
Berben, G. Brambilla and C. von Holst, An overview of
tests for animal tissues in animal feeds used in the pub-
lic health response against BSE, Rev. Sci. Tech. Off. Int.
Epiz. 22(1), 311 (2003).
14. EFSA, Scientic Opinion on the revision of the quanti-
tative risk assessment (QRA) of the BSE risk posed by
processed animal proteins (PAPs), EFSA J. 9(1), 1947
(2011).
15. L.W.D. Van Raamsdonk, C. Von Holst, V. Baeten, G.
Berben, A. Boix and J. de Jong, New developments in
the detection and identication of processed animal
proteins in feeds, Anim. Feed Sci. Technol. 133, 63 (2007).
doi: 10.1016/j.anifeedsci.2006.08.004
16. O. Fumire, P. Veys, A. Boix, C. von Holst, V. Baeten and
G. Berben, Methods of detection, species identica-
tion and quantication of processed animal proteins in
feedingstuffs, Biotechnol. Agron. Soc. Environ. 13(S), 59
(2009).
17. P. Veys and V. Baeten, New approach for the quanti-
cation of processed animal proteins in feed using light
spectroscopy, Food Additives and Contaminants 27(7),
926 (2010). doi: 10.1080/19440041003671296
18. G. Gizzi, C. von Holst, V. Baeten, G. Berben and L.W.D.
van Raamsdonk, Determination of processed animal
proteins, including meat and bone meal, in animal feed,
J. Assoc. Off. Anal. Chem. Int. 87(6), 1334 (2004).
19. European Union, Regulation (EC) No 152/2009 lay-
ing down the methods of sampling and analysis for the
ofcial control of feed, Off. J. Eur. Comm. L54, 103 (2009).
20. European Union, Directive 2003/126/EC on the
analytical method for the determination of constituents
of animal origin for the ofcial control of feedingstuffs,
Off. J. Eur. Comm. L339, 7 (2003).
21. P. Veys and V. Baeten, CRL-AP Interlaboratory Study 2006
Final Report. CRA-W, Gembloux, Belgium (2007).
22. P. Veys and V. Baeten, CRL-AP Interlaboratory Study 2007
Final Report. CRA-W, Gembloux, Belgium (2008)
23. L.W.D. van Raamsdonk, W. Hekman, J.M. Vliege, V.
Pinckaers, H. Van der Voet and S.M. van Ruth, The 2008
Dutch NRL/IAG Prociency Test for Detection of Animal
Proteins in Feed. RIKILT, Wageningen, the Netherlands
(2008). Available at: http://library.wur.nl/way/bestanden/
clc/1876397.pdf (accessed on 23 September 2011).
24. L.W.D. Van Raamsdonk, V. Pinckaers, J.M. Vliege and
S.M. van Ruth, Animal Proteins in Feed: the IAG Ring Test
2011 Report. RIKILTInstitute of Food Safety Wageningen
University & Research Centre, The Netherlands (2011).
25. O. Fumire, M. Dubois, V. Baeten, C. von Holst and G.
Berben, Effective PCR detection of animal species in
highly processed animal byproducts and compound
feeds, Anal. Bioanal. Chem. 385(6), 1045 (2006). doi:
10.1007/s00216-006-0533-z
26. D. Frezza, V. Giambra, F. Chegdani, C. Fontana, G.
Maccabiani, N. Losio, E. Faggionato, B. Chiappini, G.
Vaccari, C. von Holst, L. Lanni, S. Saccares and P.
Ajmone-Marsan, Standard and light-cycler PCR meth-
ods for animal DNA species detection in animal feed-
stuff, Innov. Food Sci. Emerg. Technol. 9(1), 18 (2008). doi:
10.1016/j.ifset.2007.04.008
27. M. Prado, G. Berben, O. Fumire, G. Van Duijn, J.
Mensinga-Kruize, S. Reaney, A. Boix and C. von Holst,
Detection of ruminant meat and bone meals in animal
feed by real-time polymerase chain reaction: result of
an interlaboratory study, J. Agric. Food Chem. 55(18),
7495 (2007). doi: 10.1021/jf0707583
28. M. Prado, O. Fumire, A. Boix, A. Marien, G. Berben and
C. von Holst, Novel approach for interlaboratory trans-
fer of real-time PCR methods: detecting bovine meat
and bone meal in feed, Anal. Bioanal. Chem. 394(5), 1423
(2009). doi: 10.1007/s00216-009-2796-7
29. M. Anseld, S.D. Reaney and R. Iackman, Production
of a sensitive immunoassay for detection of ruminant
and porcine proteins, heated to >130C at 2.7 bar, in
compound animal feedstuffs, Food Agric. Immunol. 12(4),
273 (2000). doi: 10.1080/09540100020008146
30. M. Anseld, S.D. Reaney and R. Iackman, Performance
assessment and validation of a sensitive immu-
noassay for detection of ruminant and porcine
heat stable proteins in compound animal feed-
stuffs, Food Agric. Immunol. 12(4), 285 (2000). doi:
10.1080/09540100020008155
31. A. Boix, F. Serano, S. Bellorini and C. von Holst,
Ruggedness Study of Immunoassays for Processed Animal
Proteins Detection in Feed: Neogen Reveal for Ruminant
Feed Test System. JRC-IRMM, Geel, Belgium (2006).
Available at: http://irmm.jrc.ec.europa.eu/html/activi-
ties/meat_and_bone_meal/Ruggedness_study_sep-
tember06/pdf (accessed 26 November 2006).
32. A. Boix, C. von Holst, V. Baeten, G. Berben and J.
Vancutsem, Determination of Processed Animal Proteins
(PAPS) Including Meat and Bone Meal (MBM) in Feed.
European Commission, DG Joint Research Centre,
Institute for Reference Materials and Measurements,
Geel, Belgium, p. 2440 (2004).
33. F.C. Chen, Y.H. Hsieh and R.C. Bridgman, Monoclonal
antibodies against troponin I for the detection of
230 The Potential of NIR Microscopy to Detect, Identify and Quantify Processed Animal By-Products
rendered muscle tissues in animal feedstuffs, Meat Sci.
62, 405 (2002). doi: 10.1016/S0309-1740(02)00029-3
34. S.H. Kim, T.S. Huang, T.A. Seymour, C.I. Wei, S.C. Kempf,
C.R. Bridgman, R.A. Clemens and H. An, Production
of monoclonal antibody for the detection of meat and
bone meal in animal feed, J. Agric. Food Chem. 52, 7580
(2004). doi: 10.1021/jf048789a
35. J. Weeranantanaphan, G. Downey, P. Allen and D.W. Sun,
A review of near infrared spectroscopy in muscle food
analysis: 20052010, J. Near Infrared Spectrosc. 19(2), 61
(2011). doi: 10.1255/jnirs.924
36. D. Cozzolino, A. Chree, I. Murray and J.R. Scaife, The
assessment of the chemical composition of shmeal by
near infrared reectance spectroscopy, Aquaculture
Nutr. 8, 149 (2002). doi: 10.1046/j.1365-2095.2002.00206.x
37. A. Garrido-Varo and V. Fernandez, NIRS technology for
the quantitative prediction of the percentage of meat
and bone meal added to a feed compound. A feasibil-
ity study, in Dans Report of the Workshop Identication
of Animal Ingredients in Compound Feeds, Ed. by J.S.
Jorgensen. CEMA-DG VI-SMT Program, 26 May, Lyngby,
Denmark (1998).
38. A. Garrido-Varo, M.D. Prez-Marn, J.-E. Guerrero,
A. Gmez-Cabrera, M.J. De La Haba, J. Bautista, A.
Soldado, F. Vicente, A. Martinez, B. de la Roza-Delgado
and S. Termes, Near infrared spectroscopy for enforce-
ment of European legislation concerning the use of ani-
mal by-products in animal feeds, Biotechnol. Agron. Soc.
Environ. 9(1), 3 (2005).
39. Z. Yang, L. Han, Q. Li and X. Fan, Use of discriminant
analysis on NIRS to detect meat and bone meal content
in ruminant concentrates, J. Anim. Feed Sci. 16(Suppl.
2), 442 (2007).
40. Z. Yang, L. Han, Q. Li, X. Fan and X Su, Discriminant
analysis of meat and bone meal content in ruminant feed
based on NIRS, Trans. Chinese Soc. Agric. Machinery
40(7), 124 (2009).
41. I. Murray, L.S. Aucott and I.H. Pike, Use of discriminant
analysis on visible and near infrared reectance spectra
to detect adulteration of shmeal with meat and bone
meal, J. Near Infrared Spectrosc. 9(5), 297 (2001). doi:
10.1255/jnirs.315
42. Z. Yang, L. Han, X. Liu and Q. Li, Detecting and quantify-
ing meat meal or meat and bone meal contamination
in shmeal by visible and near infrared reectance
spectra, Anim. Feed Sci. Technol. 147(4), 357 (2008). doi:
10.1016/j.anifeedsci.2008.02.005
43. Z. Yang, L. Han, X. Liu and Q. Li, Rapidly qualitative
discrimination of meat and bone meal in shmeal by vis-
ible and near infrared reectance spectroscopy, Trans.
Chinese Soc. Agric. Engin. 25(7), 308 (2009).
44. I. Murray, M.D. Prez-Marn, A. Garrido-Varo, J.E.
Guerrero, A. Puigdomenech, P. Dardenne, V. Baeten
and J. Zegers, Stratfeed near infrared instrument
network for detecting animal tissues in feedingstuffs,
in Proceedings of the 11th International Conference on
Near-Infrared Spectroscopy, 611 April 2003, Cordoba,
Spain, Ed by T. Davies and A. Garrido Varo. NIR
Publications, Chichester, UK, pp. 291294 (2004).
45. V. Baeten and P. Dardenne, The contribution of near
infrared spectroscopy to the ght against the mad cow
epidemic, NIR news 12(6), 12 (2001).
46. I. Murray, A. Garrido Varo, D. Prez Marn, J.E. Guerrero,
V. Baeten, P. Dardenne, S. Termes, J. Zegers and R.
Frankhuizen, Macroscopic near-infrared reectance
spectroscopy (WP5), in Stratfeed: Strategies and Methods
to Detect and Quantify Mammalian Tissues in Feedingstuffs.
Ofce for Ofcial Publications of the European
Communities, Brussels, Belgium (2005).
47. M.J. De La Haba, A. Garrido Varo, D.C. Prez Marn and
J.E. Guerrero, Near infrared analysis as a rst-line
screening technique for identifying animal species in
rendered animal by-product meals, J. Near Infrared
Spectrosc. 15(4), 237 (2007). doi: 10.1255/jnirs.735
48. M.J. De La Haba, A. Garrido Varo, D.C. Prez Marn and
J.-E. Guerrero, Near infrared spectroscopy calibrations
for quantifying the animal species in processed animal
proteins, J. Near Infrared Spectrosc. 17(3), 109 (2009).
doi: 10.1255/jnirs.838
49. V. Baeten and P. Dardenne, Spectroscopy: develop-
ments in instrumentation and analysis, Grasasy Aceites
53(1), 45 (2002). doi: 10.3989/gya.2002.v53.i1.289
50. V. Baeten, A. Michotte Renier, G. Sinnaeve, A. Garrido
Varo and P. Dardenne, Analysis of the sediment
fraction of feed by Near-infrared microscopy (NIR
microscopy), in Proceedings of the 11th International
Conference on Near-Infrared Spectroscopy, 611 April 2003,
Cordoba, Spain, Ed by T. Davies and A. Garrido Varo. NIR
Publications, Chichester, UK, pp. 663666 (2004).
51. V. Baeten C. von Holst, I. Fissiaux, A. Michotte Renier, I.
Murray and P. Dardenne, The near infrared microscopic
(NIR microscopy) method: combination of the advan-
tages of optical microscopy and near-infrared spectros-
copy (WP5-NIR microscopy), in Strategies and Methods
to Detect and Quantify Mammalian Tissues in Feedingstuffs.
European Commission, Brussels, Belgium (2004).
52. F. Piraux and P. Dardenne, Feed authentication by
near-infrared microscopy, in Proceedings of the 9th
International Conference on Near-Infrared Spectroscopy,
June 1999, Verona, Italy, Ed by A.M.C. Davies and R.
Giangiacomo. NIR Publications, Chichester, UK, pp.
535541 (1999).
53. V. Baeten, A. Michotte-Renier, G. Sinnaeve and P.
Dardenne, Analyses of feedingstuffs of near-infrared
microscopy (NIR microscopy): detection and quantica-
tion of meat-and-bone-meal (MBM), in Proceedings
of the Sixth Food Authenticity and Safety International
Symposium (FASIS), 2830 November 2001, Nantes, Ed
by Eurons Scientic. FASIS Organisation Committee,
Nantes, France, pp. 111 (2001).
54. V. Baeten, C. von Holst, A. Garrido Varo, J. Vancutsem, A.
Michotte-Renier and P. Dardenne, Detection of banned
Z. Yang et al., J. Near Infrared Spectrosc. 19, 211231 (2011) 231
meat and bone meal in feedstuffs by near-infrared
microscopic analysis of the dense sediment fraction,
Anal. Bioanal. Chem. 382(1), 149 (2005). doi: 10.1007/
s00216-005-3193-5
55. B. De La Roza Delgado, A. Soldado, A. Martinez
Fernandez, F. Vicente, A. Garrido Varo, D.C. Prez Marn,
M.J. De La Haba and J.E. Guerrero Ginel, Application
of near-infrared microscopy (NIR microscopy) for the
detection of meat and bone meals in animal feeds: A tool
for food and feed safety, Food Chem. 105, 1164 (2007).
doi: 10.1016/j.foodchem.2007.02.041
56. C. von Holst, V. Baeten, A. Boix, B. Slowikowski,
J.A. Fernndez Pierna, S. Tirendi and P. Dardenne,
Transferability study of a near-infrared microscopic
method for the detection of banned meat and bone meal
in feedingstuffs, Anal. Bioanal. Chem. 392, 313 (2008).
doi: 10.1007/s00216-008-2232-4
57. D.C. Prez Marn, T. Fearn, J.E. Guerrero and A. Garrido
Varo, A methodology based on NIR-microscopy for the
detection of animal protein by-products, Talanta 80, 48
(2009). doi: 10.1016/j.talanta.2009.06.026
58. O. Fumire, G. Berben, J.A. Fernndez Pierna, C. Andre,
P. Dardenne and V. Baeten, Speciation of mammalian
MBM particles with NIR microscopy and real time
PCR analyses, in Proceeding in the 12th International
Conference on Near Infrared Spectroscopy (ICNIRS): NIR
in Action, 1115 April (2005), Auckland, New Zealand, Ed by
G.R. Burling-Claridge, S.E. Holroyd and R.M. Sumner.
New Zealand Near Infrared Spectroscopy Society,
Aukland, New Zealand, pp. 129132 (2007).
59. O. Fumire, A. Marien, J.A. Fernndez Pierna, V.
Baeten and G. Berben, Development of a real-time
PCR protocol for the species origin conrmation of
isolated animal particles detected by NIR micro-
scopy, Food Addit. Contam. 27(8), 1118 (2010). doi:
10.1080/19440049.2010.481639
60. M.J. De la Haba, J.A. Fernndez Pierna, O. Fumire,
A. Garrido Varo, J.E. Guerrero, D.C. Prez Marn, P.
Dardenne and V. Baeten, Discrimination of sh bones
from other animal bones in the sedimented fraction of
compound feeds by near infrared microscopy, J. Near
Infrared Spectrosc. 15, 81 (2007). doi: 10.1255/jnirs.688
61. J.A. Fernndez Pierna, V. Baeten, A. Michotte Renier,
R.P. Cogdill and P. Dardenne, Combination of support
vector machines (SVM) and near infrared (NIR) imaging
spectroscopy for the detection of meat and bone meal
(MBM) in compound feeds, J. Chemometrics 18, 341
(2004). doi: 10.1002/cem.877
62. O. Abbas, J.A. Fernndez Pierna, A. Boix, C. von Holst,
P. Dardenne and V. Baeten, Key parameters for the
development of a NIR microscopic method for the quan-
tication of processed by-products of animal origin in
compound feedingstuffs, Anal. Bioanal. Chem. 397, 1965
(2010). doi: 10.1007/s00216-010-3706-8
63. O. Fumire, C. Andre, G. Berben, P. Dardenne and V.
Baeten, An original strategy coupling NIR microscopy
and PCR for detection and species identication of MBM
particles, Poster 26 presented during the International
symposium on Food and Feed Safety in the Context of
Prion Diseases, 1618 June 2004, Namur. Available
at: http://Stratfeed.cra.wallonie.be/page/events.php
[online] (accessed 23 September 2011)].
64. J.A. Fernndez Pierna, P. Dardenne and V. Baeten,
Screening of compound feeds using NIR hyperspectral
data, Chemometrics and Intelligent Laboratory Systems
84, 114 (2006). doi: 10.1016/j.chemolab.2006.03.012
65. D. Pavino, S. Squadrone, M. Cocchi, G. Martra, D.
Marchis and M.C. Abete, Towards a routine application
of vibrational spectroscopy to the detection of bone frag-
ments in feedingstuffs: Use and validation of a NIR scan-
ning microscopy method, Food Chem. 121, 826 (2010).
doi: 10.1016/j.foodchem.2009.12.092
66. J.A. Fernndez Pierna, P. Dardenne and V. Baeten,
In-house validation of a near infrared hyperspectral
imaging method for detecting processed animal proteins
(PAP) in compound feed, J. Near Infrared Spectrosc. 18,
121 (2010). doi: 10.1255/jnirs.872
JOURNAL
OF
NEAR
INFRARED
SPECTROSCOPY
233
ISSN: 0967-0335 IM Publications LLP 2011
doi: 10.1255/jnirs.932 All rights reserved
Near infrared (NIR)-based hyperspectral imaging
1,2
is a very
powerful analytical technique for many organic and synthetic
materials but it needs efcient algorithms for handling the
huge amount of data generated. In many situations a type
of classication is needed in which detected classes can be
identied and visualised. Partial least squares discriminant
analysis (PLS-DA) is an important classication method used
in multivariate data analysis. Many other methods based on
a variety of principles have been proposed in the literature to
solve the problem of two-class or multiple classication, for
example, linear discriminant analysis (LDA), soft independent
modelling of class analogy classication, neural network, and
numerous other methods.
37
The main sub-division in classi-
cation methods is in supervised and unsupervised classica-
tion. In supervised classication, class membership is known
for the training data. In unsupervised classication, the goal is
to nd meaningful classes in the data without prior knowledge
of how many objects there are in one class or their identity.
LDA and PLS-DA are two very similar methods used in super-
vised classication. The difference is that the rst method
The importance of balanced data sets for
partial least squares discriminant analysis:
classification problems using hyperspectral
imaging data
Susanne W. Lindstrm,
a,b,
* Paul Geladi,
c
Oskar Jonsson
b
and Fredrik Pettersson
b
a
Ume University, Applied Physics and Electronics, Ume, Sweden. E-mail: Susanne.wiklundlindstrom@chem.umu.se
b
UmBio AB, Ume, Sweden
c
Biomass Technology and Chemistry, Swedish University of Agricultural Sciences, Ume, Sweden
This study investigates the effect of imbalanced spectral data in the training set, when developing partial least squares discriminant
analysis (PLS-DA) classication models for use in future predictions. The experimental study was performed using a real hyperspectral
short-wavelength infrared image data set collected from bakery products (buns) containing contaminants (ies) but similar applications
for other insects, paper and plastic were also tested. The contaminants represent a very small proportion of the images relative to the
bun. The PLS-DA model aims at accurately detecting and classifying the contaminants and this requires a modication of the calibra-
tion data set. The paper deals with problems caused by unbalanced calibration data sets and how to remedy them. In the example it was
demonstrated that, by balancing the calibration data from 58,476 bun pixels + 279 y pixels to 279 bun + 279 y pixels, the number of true
predictions could be improved with a smaller number of PLS components used in the model. The improvement for ies increased from
65% true predictions with ten PLS components to > 99% true prediction with ve to six PLS components. The true prediction for bun went
from 100% to 99.5% with six PLS components which is an acceptable reduction. Theoretical explanations are included.
Keywords: hyperspectral imaging, PLS-DA, classication, unbalanced model, obtaining a balanced dataset
Introduction
S.W. Lindstrm et al., J. Near Infrared Spectrosc. 19, 233241 (2011)
Received: 15 March 2011 Revised: 5 August 2011 Accepted: 11 August 2011 Publication: 5 September 2011
234 The Importance of Balanced Data Sets for PLS-DA
can only be applied if there are few uncorrelated variables.
Hyperspectral images based on NIR (7802500 nm) and visible
(400780 nm) wavelength ranges consist of many and corre-
lated variables and therefore rarely full this requirement.
There are some special issues associated with PLS-DA,
commonly occurring when working with hyperspectral data.
Classication problems with almost equal sizes of class A and
class B were always presented in the traditional classication
literature and PLS-DA had no problems. One big problem for
images is the possible unequal number of pixels for each class.
If the number of pixels in one class is very much larger than
in the other class, the data set is unbalanced and the model
will fail to predict both classes well, which is demonstrated
in this paper. This is a common problem when working with
hyperspectral images and should thus be handled correctly.
A similar problem has previously been addressed for clinical
data using neural nets.
8,9
A good classication model should
be able to predict both classes well, not only one of the classes.
A diagnostic for prediction quality is necessary and, in this
paper, all results are shown as the percentage of true predic-
tions (TPs), with predictions based on a true test set.
The problem with unequal size hyperspectral data classes
when applying PLS-DA for classication is illustrated by the
use of an example from the automated inspection of bakery
products. An industrial bakery plant produces different types
of buns and cookies and there is a problem with different
contaminants falling on top of the buns, for example, ies,
paper and plastics. Sometimes, only visual inspection by a
trained operator is used as quality control. From a visual
inspection it is difficult to distinguish flies from chocolate
chips and raisins, so the example used here is based on a
real production problem. Hyperspectral imaging identication
of contaminants is proposed as a potentially more effective
alternative to visual inspection. In order to detect and identify
different contaminants, a supervised PLS-DA model was built.
The problem found with this application was that the contami-
nants class was often less than 1% of the total image area
under study, i.e. the classes were very much unbalanced in
size, which impaired the accuracy of the PLS-DA prediction
model. The result of unbalanced PLS-DA models is presented
and a solution to the problem is suggested. In this paper, a
space-lling design algorithm in latent variables was applied
to each class in order to select a sub-set of samples (pixels)
from the class with more pixels.
10
This type of sample selection
procedure makes the classes balanced in size while retaining
population properties. An additional advantage is faster model
calculation.
PLS-DA belongs to the supervised classication methods.
The method is traditionally used in classication problems
where one of the main goals of the PLS-DA algorithm is to
have excellent predictions and classications for new unknown
samples. Lately, PLS-DA has also become popular as a method
for variable selection and interpretation, for example, in the
eld of metabolomics.
11
For hyperspectral images, the total
number of samples (N) is the same as the total number of
pixels used in the training data set for modelling. For object
groups A and B the following vectors and matrices in X are
created:
X = [ X
A
;X
B
] (1)
with X
A
: (I K) spectra for the A type objects and X
B
: (J K)
spectra for the B type objects.
To perform PLS-DA, two y vectors describing class member-
ship are constructed for classes A and B:
y = [ y
A
; y
B
] (2)
with y
A
: (I 1) a vector of ones for class A and y
B
: (J 1) a vector
of zeros for class B.
PLS-DA then simply nds the best PLS regression model
between Z and v, the mean-centred versions of X and y:
v = Zb + f (3)
b : (K 1) a vector of PLS regression coefcients, and f : [I + J] 1)
or (N 1) a vector of residuals.
The vector b from Equation (3) can then be used for predicting
class membership for new pixels of unknown class member-
ship. If the predicted value of y is close to 1, then the unknown
object belongs to the A class; if it is close to 0 it belongs to
the B class. A value close to 0.5 indicates ambiguous class
membership. Values far outside the interval [0:1] would
indicate no membership to A or B. In imaging, ambiguous
class membership is expected because of mixed pixels. The
predicted conclusions are approximate because the values
never become exactly 0 or 1, even for true class members. For
this reason, it is nice to have distributions, so that a probability
of class membership can be given. Having many objects in
classes A and B allows a normal distribution around 1 or 0 to
be used. This would typically be the case in imaging, where a
class almost always comprises >1000 pixels.
PLS also includes a reduction of Z [Equation (3)]:
Z = Z
hat
+ E (4)
with Z
hat
a model of Z that resembles Z more and more as the
number of PLS components increases and E a residual that
becomes smaller for more PLS components.
Materials and methods
Data collection
Flies, plastics, paper and buns etc. were collected from a
production site of buns and cookies. The ies, plastics and
paper are classied as possible contaminants on the buns and
unwanted in production and sales. In general, the round buns
had a diameter of approximately 80 mm. Insects of 14 mm in
length and pieces of packaging material (paper, plastic) of a
few square millimetres to a few square centimetres had to be
detected and the contaminated buns had to be removed from
the production line.
Data acquisition was performed by using the UmBio
Inspector workstation (UmBio AB, Ume, Sweden) including
a line scanning short-wavelength infrared (SWIR) camera
S.W. Lindstrm et al., J. Near Infrared Spectrosc. 19, 233241 (2011) 235
(Specim Imaging Ltd AB, Oulu, Finland) recording absorb-
ance values for the wavelengths ranging from 1000 nm to
2500 nm with a 6.3 nm spectral resolution. Each line meas-
urement was performed using 1.7 ms exposure time with a
31 mm lens having a eld of view of 100 mm. The number of
wavelength bands was 239 after removal of some bad wave-
lengths at the extremes. A typical image data cube had the
size 320 pixels 320 pixels 239 wavelength bands. Each pixel
had a size of 0.3 mm 0.3 mm.
The data of an image of one bun were split into a training set
and a test set for demonstration purposes for this paper. The
models were later repeated for more buns and more contami-
nants and more robust training models were made from these.
These results are not reported here but the general nding
was the same.
Data cleaning
After acquisition, each NIR image (sample) was comprised of a
three-dimensional data matrix size 320 pixels 320 pixels 239
wavelengths, where each pixel was one spectrum. Digital
images are rectangular, but the object studied was round, so
background removal had to be applied on the data cube. At the
same time, other errors, such as bad pixels, specular reec-
tion etc. were removed. The cleaning of the image from bad
pixels and background noise was performed by using inter-
action (brushing
1,2,12
) between principal component analysis
(PCA) score plots and score images. PCA was used without
scaling or mean-centring of the data, to separate signals from
the bun from background noise. The cleaning of data was
done using the software Evince Image 2.4.0 (UmBio AB, Ume,
Sweden).
The cleaned data cube was reshaped to a two-dimensional
data matrix to X[N, K] = X[I + J, K] where N was the total number
of pixels in the training data set, J was the number of pixels
for buns and I was the number of pixels for ies in the training
data set. After data cleaning, J = 65,378 and I = 479 and the
spectral wavelengths was K = 239. In the reshaped data matrix,
the rows represented the pixels from the buns and ies, I + J,
and the variables represented the spectral wavelength, K.
Partial least squares discriminant analyis
classication modelling
The cleaned data were partitioned into two classes: class A
from buns and class B from ies. The supervised selection
of which pixels belonged to each class was performed by
interactively working with the PCA score scatterplot from the
hyperspectral image and the RGB image. Thereafter, the data
were separated into a training set and a test set by spatially
cutting the image into two pieces, making sure that each set
contained pixels of both bun and ies. The training set included
58,476 pixels from buns and 279 pixels from ies and the test
set included 6902 pixels from buns and 200 pixels from ies
(see Table 1). The PLS-DA calculations were performed as
described above using mean-centring on the X[(I + J), K] and
Y[(I + J), 2] data. For the unbalanced example, the training data
set size was X[(58,476 + 279), 239]. In the balanced example,
the MarengoTodeschini (MT) algorithm
13
was used on the
buns to select a sub-group of pixels to be included in PLS-DA
modelling, while all other pixels were excluded from the
training data set. The MT is further described in this paper
under the section on balancing methods. The MT selection
makes the two classes balanced, i.e. equal number of pixels
in both classes, X[(279 + 279), 239]. These two data sets, i.e.
unbalanced and balanced, were compared in terms of the
percentage of TPs from the test set, using both classes and
from different PLS-DA components. With PLS models varying
in number of components, the number of components in the
calibration model had to be taken into account. For robust-
ness reasons, the number of PLS components should be as
small as possible while still giving acceptable classication
results. The model statistical results were thus presented as
the number of components A, R
2
X, R
2
y and the percentage of
TPs from the test set:
R
2
X = 1 SS(E)/SS(Z) (5)
where the amount of explained variation in Z equals that in X
[see Equations (3) and (4)].
R
2
Y = 1 SS(f)/SS(v) (6)
where the amount of explained variation in v equals that in y
[see Equations (3) and (4)].
The percentage of TP = 100 (number of correct predictions)/
(number of observations in test set). The percentage of TP was
calculated for both ies and buns separately.
Balancing methods
There are different ways to achieve balanced classication
models. One way is to normalise calculations within the
algorithms internally. An example of this is weighting the
loadings of PLS models. Another example is to adjust the
number of observations that are included in the data set that
the multivariate models are calculated for. The class of lower
representation could potentially be extended by a simula-
tion of a sampling procedure, but with hyperspectral data
it can be a tricky procedure to accurately capture a realistic
distribution of values and it is more attractive to reduce the
size and complexity of the already huge data sets that are
commonly analysed in the eld. Another way to adjust the
Class
Size
Training set Test set
Unbalanced
Bread 58,476 6902
Flies 279 200
Balanced
Bread 279 6902
Flies 279 200
Table 1. Number of observations (pixels) in each class for an
unbalanced and a balanced training set and test set.
236 The Importance of Balanced Data Sets for PLS-DA
characteristics of data sets is to include the same number
of samples for all classes so that all classes are equally
contributing to the modelling procedure. To retain as much
as possible of the variation in the data set after maybe drasti-
cally reducing its size, it is important to select the samples to
include as much as possible of the variation in the candidate
set of samples. One possible way to perform a statistically
sound selection is to apply a space-lling design on latent
variables. Latent variables can be obtained by calculating a
PCA model for the samples of each class separately. From
each of these models, the same number of samples as is
included in the smallest class is selected. This is achieved
by selecting samples so that the distances between all these
are maximised and so that the selection explains as much
as possible of the variation in the candidate data set. The
MT algorithm was used to make an informative selection of
samples from each individual class PCA model. This algo-
rithm selects an optimal sub-set out of a larger set of candi-
dates by considering distances in multivariate space. The
distances between objects in the optimal sub-set are opti-
mised. An interesting image-based alternative is described
by Esbensen and Lied.
1
Results and discussion
An example of a typical bun from the production site is seen
in Figure 1. By looking at a PCA score plot from the bun
with a y, it is seen that these two classes separate into two
clusters using components 1 and 2 in the score scatterplot
(Figure 2). When calculating a PLS-DA model, a separation
Figure 1. Digital colour image of a typical bun used in the study.
The image represents 130 90 mm
2
. The bun was approxi-
mately 80 mm in diameter. (This gure is reproduced in colour
on the web)
Figure 2. PCA score scatterplot from component t1 vs t2. A separation between bun and ies can be seen. The grey dots are from the
ies and the black ones are from buns.
S.W. Lindstrm et al., J. Near Infrared Spectrosc. 19, 233241 (2011) 237
Figure 3. PLS-DA Score plot from (a) an unbalanced model and (b) a balanced model. The light grey dots are from the ies and the
black dots are from buns.
(a)
(b)
238 The Importance of Balanced Data Sets for PLS-DA
between the two classes was observed [Figures 3(a) and (b)].
However, although a visual separation was seen in the score
plot, the unbalanced model was poor in terms of predictive
ability.
Table 2 presents a model summary of the amount of
explained variation for X and Y in each component, using
balanced and unbalanced data sets. The balanced model will
very quickly reach a high explanation of both X and Y, R
2
X and
R
2
Y respectively, but the unbalanced model never seemed to
reach a high explanation of X and Y. A simple rule of thumb
is that a PLS model should reach an R
2
Y of 0.65 after fewer
than six components. This was not the case for the unbal-
anced training data.
The previous results gave some explanation to the next
result, which demonstrated how the predictive ability changed
with the increased number of components for both data sets.
The results from the test set show that the unbalanced model
predicted the test set bun with 100% true predictions; mean-
while the ies only had 65% TP after 11 PLS-DA components
[Figure 4(a)]. When using a balanced training data set and only
six components, both bun and ies have 99.5% TP, which is
a very high percentage of correct predictions [Figure 4(b)]. It
was possible to reduce both classes even more and using only
50 pixels for bun and 0 pixels for y still gave >95% TP for both
classes.
It is obvious from these results that a balanced model will
give better results in terms of predictive ability. This behav-
iour can, to some extent, be explained by the observed
result. One explanation is that the Y vector will never be
explained high enough in order to have high predictive
ability (Table 2).
Another theoretical explanation is that a PLS component
tries to maximise the covariance between t and u, the PLS
scores for X and y, respectively. Simplied, this can be shown
as cov(t,u) = corr(t,u) * [var(t) * var(u)].
14,15
With unbalanced
calibration sets, many of the PLS components will make
Var(t) large for the largest class and ignore making the size
of Corr(t,u) large. With a balanced calibration set, directions
in multivariate space that maximise Corr(t,u) will more easily
dominate and this produces a more robust PLS model using
less components. Another geometrical explanation to this
behaviour is that the PLS components are not following the
Data set A R
2
X R
2
X_CUM R
2
Y R
2
Y_CUM
Unbalanced 1 0.879 0.879 0.015 0.015
2 0.090 0.969 0.089 0.104
3 0.022 0.991 0.032 0.136
4 0.005 0.996 0.074 0.211
5 0.001 0.997 0.152 0.362
6 0.000 0.997 0.180 0.542
7 0.000 0.997 0.039 0.582
8 0.001 0.998 0.009 0.591
9 0.000 0.998 0.036 0.627
10 0.000 0.998 0.006 0.634
11 0.000 0.998 0.011 0.645
Balanced 1 0.927 0.927 0.539 0.539
2 0.058 0.985 0.210 0.748
3 0.005 0.990 0.041 0.789
4 0.006 0.996 0.028 0.817
5 0.001 0.997 0.053 0.870
6 0.001 0.998 0.014 0.883
7 0.000 0.998 0.008 0.892
8 0.000 0.998 0.007 0.899
9 0.000 0.998 0.003 0.902
10 0.000 0.999 0.003 0.905
11 0.000 0.999 0.003 0.907
A is the number of components in the model;
R
2
X is the amount of variation in X explained in each component;
R
2
X_CUM is the cumulative R
2
X;
R
2
Y is the amount of variation explained in Y for each component;
and R
2
Y_CUM is the cumulative R
2
Y.
Table 2. Model results from unbalanced and balanced training data sets.
S.W. Lindstrm et al., J. Near Infrared Spectrosc. 19, 233241 (2011) 239
direction of the variation between the two classes. Instead the
component is directed towards the direction of the highest
variation within the X data from the bun class [Figure 5(a)].
When the model is balanced, the right direction is found, i.e.
the direction between the two class centres [Figure 5(b)]. This
fact was not mentioned in the literature previously, as far as
the authors have been able to nd out.
A further experimental observation about the geometrical
explanation can be found by looking at the weights (w1) and
loadings (p1) from the first component. These two vectors
should be rather similar if the correct direction between the
two classes is found. If the vectors are not similar, it means
that the rst component contains a lot of orthogonal varia-
tion.
16
In Figures 6(a) and (b) the w1 and p1 from the unbal-
anced and the balanced PLS-DA model is seen. For the unbal-
anced model w1 and p1 are not very similar in shape, but for
the balanced model they are very similar.
Conclusions
Contrary to what one would expect, maximising the number of
members of a training class for PLS-DA classication may not
be optimal. If there is an imbalance in members between two
classes, it is better to reduce the number of members of the
larger class in order to obtain balance. This is especially the
case in hyperspectral imaging where some classes have many
thousands of members. This was shown with the example of
Figure 4. Percentage of true predictions (TPs) from bun (diamonds) and ies (squares) (a) in the unbalanced model and (b) in the bal-
anced model. The vertical axis represents the number of components in the model and the horizontal axis is the percentage of TPs.
Notice that the scales on the vertical axes are different.
(a)
(b)
240 The Importance of Balanced Data Sets for PLS-DA
Figure 5. Geometric explanation of the problem of an unbalanced training data set. (a) Unbalanced model. The PLS-DA component is
directed towards the highest variation in the data, which is within the bun class (black arrow); the dashed arrow shows the direction
required. (b) Balanced model. The component is directed between the two class centres, as it should be.
(a) (b)
Figure 6. Wavelength versus loading, p1, and weights, w1, from (a) an unbalanced PLS-DA model and (b) a balanced PLS-DA model.
(a)
(b)
S.W. Lindstrm et al., J. Near Infrared Spectrosc. 19, 233241 (2011) 241
a bun with small ies, measured with a hyperspectral SWIR
camera. The bun was represented by 99.5% of all pixels in the
cleaned image and the y pixels occupied only 0.5%, which
made these classes very unbalanced in size.
A number of observations can be made here:
1. In imaging, classes consisting of thousands of pixels can
easily be obtained;
2. Class sizes are often extremely unbalanced, especially when
looking for contamination in a bulk material;
3. Mixed pixels will always occur because pixels are square
in the spatial domain and most particles imaged do not have
straight edges parallel to the pixel edges;
4. A naively constructed unbalanced PLS-DA model proved to
be inadequate for prediction of the smallest classes using a
true test set;
5. A number of ways may be devised for balancing. These
include: (a) collecting more image data for the smallest class,
(b) articially generating class members for the smallest class,
(c) reducing the largest class to obtain balance while main-
taining its statistical properties. Quite a number of methods
with different statistical properties can be devised for carrying
out the above operations;
6. The problem of the PLS-DA model is not on the y side but
on the X-side because a PLS model tries to simultaneously
optimise correlation (between X and y) and variance in X. This
requires too many PLS components for unbalanced data with
a large variance in the largest class of X. Moreover, the amount
of within class variation will have an impact on the problem
with unbalanced data sets. If the classes have a low within
class variability, then the problem might be small. If the within
class variation is high, the unbalanced problem is likely to be
more severe.
For future research, a test on a larger classification
example, possibly a real process analytical technology appli-
cation could be tried. It is also possible to use PLS-DA with
more than two classes and many real applications would
produce this situation, for example, buns against insects,
buns against paper, buns against plastic. PLS-DA can be
used as PLS2 for this situation. A potential for many different
balancing/compensation methods is also foreseen for the
future.
The severity of the problem with unbalanced data in clas-
sication models using hyperspectral images, is likely to
be related to the amount of variation within and between
classes. In those cases where the within-class variation
is small and the between-class variation is very high, the
correct direction of the PLS-DA component might be found,
although the data is unbalanced. However, this should also
be investigated further.
Acknowledgements
P.G. acknowledges the FIELD NIRce (Botnia Atlantica,
EU-InterregIV) project for nancial support. S.W.L. acknowl-
edges the Swedish Energy Agency for nancial support.
References
1. H. Grahn and P. Geladi (Eds), Techniques and Applications
of Hyperspectral Image Analysis. John Wiley & Sons Ltd,
Chichester, UK (2007).
2. S. Sasic and Y. Ozaki (Eds), Raman, Infrared and Near-
Infrared Chemical Imaging. John Wiley & Sons Inc.,
Hoboken, NJ, USA (2010).
3. R. Brereton, Data Analysis for the Laboratory and Chemical
Plant. John Wiley & Sons Ltd, Chichester, UK (2003).
4. R. Henrion and G. Henrion, Multivariate Dataenanalyse,
Methodik und Anwendung in der Chemie und Verwandten
Gebieten. Springer Verlag, Berlin, Germany (1995).
5. K. Varmuza and R. Filzmoser, Introduction to Multivariate
Statistical Analysis in Chemometrics. CRC Press, Boca
Raton, FL, USA (2009).
6. R. Johnson and D. Wichern, Applied Multivariate
Statistical Analysis. Prentice-Hall, Englewood Cliffs, NJ,
USA (1982).
7. R. Brereton, Chemometrics for Pattern Recognition. John
Wiley & Sons Ltd, Chichester, UK (2009).
8. J.S. Lan, V.L. Berardi, B.E. Patuwo and M. Hu, A joint
investigation of misclassication treatments and imbal-
anced datasets on neural network performance, Neural
Comput. Applic. 18, 7 (2009). doi: 10.1007/s00521-009-
0239-1
9. M.A. Mazurowski, P.A. Habas, J.M. Zurada, J.Y. Lo, J.A.
Baker and G.D. Tourassi, Training neural network
classiers for medical decision making: The effects
of imbalanced datasets on classication perform-
ance, Neural Networks 21, 2 (2008). doi: 10.1016/j.neu-
net.2007.12.031
10. S. Wold, M. Josefson, J. Gottfries and A. Linusson,
The utility of multivariate design in PLS modeling, J.
Chemometr. 18, 3 (2004). doi: 10.1002/cem.861
11. S. Wiklund, E. Johansson, L. Sjstrm, E. Mellerowicz, U.
Edlund, J. Shockcor, J. Gottfries, T. Moritz and J. Trygg,
Visualization of GC/TOF-MS-based metabolomics data
for identication of biochemically interesting compounds
using OPLS class models, Anal. Chem. 80, 115 (2008).
doi: 10.1021/ac0713510
12. P. Geladi and H. Grahn, Multivariate Image Analysis. John
Wiley & Sons Ltd, Chichester, UK (1996).
13. E. Marengo and R. Todeschini, A new algorithm for opti-
mal, distance-based experimental design. Chemometr.
Intell. Lab. Syst. 16, 37 (1992). doi: 10.1016/0169-
7439(92)80076-G
14. A. Hksuldsson, Prediction Methods in Science and
Technology. Thor Publishing, Copenhagen, Denmark
(1996).
15. A. Hoskuldsson, PLS regression methods, J.
Chemometr. 2, 211 (1988). doi: 10.1002/cem.1180020306
16. J. Trygg and S. Wold, Orthogonal projections to latent
structures (O-PLS), J. Chemometr. 16, 119 (2002). doi:
10.1002/cem.695
JOURNAL
OF
NEAR
INFRARED
SPECTROSCOPY
243
ISSN: 0967-0335 IM Publications LLP 2011
doi: 10.1255/jnirs.933 All rights reserved
Transformers are critical elements for the transmission and
distribution of energy in electricity power systems. They are
expected to operate for up to 40 years and a transformers
life is actually the life of the internal insulation system, which
consists of Kraft paper immersed in insulating mineral oil.
Kraft paper is composed of cellulose (about 90%), hemi-
celluloses and lignin. Cellulose degrades slowly with the
ageing of the transformer and the polymer chains break down,
releasing degradation products into the oil.
1
Insulating oils
are composed of naphthenic, parafnic and minor amounts
of aromatic hydrocarbons, and can be degraded producing
carboxylic acids. The main transformer ageing accelera-
tors are moisture, oxygen and heat. Thermal ageing of the
Kraft paper is the critical life-limiting process, and its main
Development and validation of a
multivariate calibration method for
determining interfacial tension of
transformers insulating oils by near
infrared spectroscopy
Mariana S. Godinho,
a
Anselmo E. Oliveira
b
and Marcelo M. Sena
a,c,*
a
Mestrado em Cincias Moleculares, UnUCET, Universidade Estadual de Gois, PO Box 459, 75001-970, Anpolis, GO, Brazil.
E-mail: marcsen@ufmg.br
b
Instituto de Qumica, Universidade Federal de Gois, 74001-970, Goinia, GO, Brazil
c
Departamento de Qumica, ICEx, Universidade Federal de Minas Gerais, 31270-901, Belo Horizonte, MG, Brazil
The aim of this study was to develop and validate a multivariate calibration method based on partial least squares (PLS) and near
infrared spectroscopy for determining the interfacial tension of insulating oils used in electricity power transformers. Forty-eighty oil
samples were obtained from an electricity power company and were divided into a calibration set of 38 leaving 10 for the validation set.
The reference values were measured with a ring tensiometer and ranged from 15 dyn cm
1
to 45 dyn cm
1
. The spectra were registered
from 1280 nm to 2500 nm and the best PLS model was selected with 10 latent variables, providing a root mean square error of predic-
tion of 2 dyn cm
1
and a coefcient of determination between reference and predicted values of 0.962. This method was validated in the
context of multivariate calibration and was considered to have adequate accuracy (prediction errors between 9% and 12%), precision
and linearity. A bias test veried the absence of systematic errors in the model. A pseudo-univariate calibration curve was constructed
based on the concept of net analyte signal. The proposed method was simple, rapid, non-destructive and considered appropriate for the
routine use in a quality control laboratory of an electricity power company.
Keywords: NIR spectroscopy, PLS, power transformers, insulating oil, chemometrics, net analyte signal, NAS,
multivariate analytical validation
Introduction
M.S. Godinho, A.E. Oliveira and M.M. Sena, J. Near Infrared Spectrosc. 19, 243251 (2011)
Received: 21 July 2011 Revised: 29 August 2011 Accepted: 29 August 2011 Publication: 26 September 2011
244 Determining Interfacial Tension of Transformers Using Insulating Oils
degradation products are dissolved in the insulating oil, such
as furanic derivatives, water and gases (CO, CO
2
, CH
4
, H
2
).
2

The main furanic compounds found in transformer oils are
2-furfuraldehyde, 2-furfuryl alcohol, 2-methyl-furfural and
5-hydroxy-2-methyl-furfural.
1
Among these furans, 2-furfural-
dehyde has the highest solubility in oil and can be considered
as a cellulosic insulation degradation indicator.
3
Currently,
electricity power companies are searching for simple, rapid,
reliable and non-destructive methods to assess the condi-
tions of transformers. The existing methods do not provide
reliable results or are destructive and involve damage to the
transformer. The methods evaluated include dielectric meas-
urements,
4
determination of the degree of polymerisation of
the Kraft paper by gel permeation
5
or size exclusion chroma-
tography
6
and analysis of insulating oils by moisture determi-
nation with Karl Fischer titration,
7
dissolved gas analysis
8
and
furan determination with high-performance liquid chroma-
tography.
9
In recent years, some workers have proposed direct
and non-destructive methods for assessing the conditions
of transformers, such as the determination of the degree of
polymerisation
10
and water content
11
of insulating papers by
near infrared (NIR) spectroscopy and the determination of the
interfacial tension of insulating oils by image analysis.
12
These
direct methods have in common the need to use multivariate
calibration chemometric methods.
Brazilian legislation prescribes assays of insulating oils for
evaluating the condition of transformers, with tests for varia-
bles such as power factor, dielectric rigidity, moisture, neutral-
isation index and interfacial tension.
13
Interfacial tension is
considered the parameter most sensitive to the whole insu-
lating system degradation, since it tends to decrease over time,
mainly due to polar substances released from the paper, such
as furan compounds and water. Interfacial tension is a measure
of the force required to separate the interface between insu-
lating oil and water. For its determination, Brazilian companies
use the ofcial method
14
based on ring tensiometer measure-
ments, which is slow, destructive and has low precision. For
these reasons, the objective of our research was to develop an
alternative, simple, direct, rapid and non-destructive method
for determining interfacial tension of insulating oils by using
NIR spectroscopy and multivariate calibration with partial
least squares (PLS). Moreover, aiming at assuring the reli-
ability of the method, a robust procedure was implemented for
analytical validation, including outlier detection, employment
of the net analyte signal (NAS) and estimation of gures of
merit (FOM), such as linearity, accuracy, precision, analytical
sensitivity, limit of quantication and bias.
Multivariate analytical
validation
Detection of outliers
A crucial step in the development and validation of any multi-
variate calibration model is the detection of outliers. Outliers
can be detected based on the identication of samples with
extreme leverages, large residuals in the spectral data (X
block) or large residuals in the predicted properties (Y block).
In this study, we adopted a methodology based on the recent
literature
1517
for simultaneously carrying out these three
diagnostics. The leverage, h
i
, measures the inuence of each
sample on the model. Samples with h
i
larger than a limit value,
h
lim
, should be deleted and the model rebuilt, according to the
following equation, which holds for mean centred data.
15
h
lim
= 3(A + 1)/n
cal
(1)
where A is the number of latent variables (LVs) of the PLS
model and n
cal
is the number of calibration samples. Outliers
with high spectral data residuals are detected by comparison
of the total standard deviation, s(e), with the standard devia-
tion of each sample, s(e
i
), related to the values of absorb-
ance measured and predicted with A LVs. If a sample has
s(ei) > ns(e), where n is a constant that varies from 2 to 3, it
should be deleted from the calibration set.
17
In this work,
n = 2 was adopted, corresponding to approximately the 95%
condence level. These two tests are applied to both calibra-
tion and validation sets. Outliers with high Y residuals can be
detected by comparing the root mean square error of calibra-
tion (RMSEC) of the model with absolute errors of individual
samples. If a sample has a difference between its reference
value and its estimated value larger than a constant from two
to three times the RMSEC, it should be deleted. In this work,
this constant was set as 3. This last test is applied only to the
calibration set. A deeper description of these tests can be
found in the appropriate reference.
17
Net analyte signal and pseudo-univariate
calibration curves
Net analyte signal is dened as a vector orthogonal to the
space of interferences, which is uniquely related to the analyte
or property of interest. A NAS vector, x
^
A,I
nas
, can be estimated
for each sample, i, from the regression vector of a PLS model
with A LVs, b:
18
x
^
A,i
nas
| = b(b
T
b)
1
b
T
x
i
(2)
The norm of each NAS vector yields a scalar ns
i
, which can
be interpreted as a selective univariate analyte signal. Thus,
the concept of NAS makes it possible to express multivariate
calibration models in a simple and more interpretable way,
through the pseudo-univariate calibration graphs.
19
Firstly,
x
^
A,i
nas
vectors are estimated for the calibration samples, and
then a regression coefcient, b
^
nas
, is calculated by a linear
regression between a vector containing the scalar ns values
(ns) and the vector of reference properties (y).
b
^
nas
= (ns
T
ns)
1
ns
T
y (3)
Finally, the regression model can be expressed as:
y
^
= b
^
nas

ns + e (4)
where e is a vector containing the residuals. This model
provides results similar to PLS.
20
M.S. Godinho, A.E. Oliveira and M.M. Sena, J. Near Infrared Spectrosc. 19, 243251 (2011) 245
Figures of merit
Analytical validation is an important step for assuring the reli-
ability of analytical methods, but the traditional regulations in
different areas remain conceived in a univariate way. With the
growing use of NIR spectroscopy methods, a need to establish
an analytical validation based on the multivariate thinking
has emerged.
21
Thus, new ways of estimating FOM have been
proposed, several of them based on the NAS. FOM such as
precision and accuracy can be estimated in a similar manner
for multivariate and univariate methods. Nevertheless, the
accuracy in multivariate calibration is usually expressed
through RMSEC and root mean square error of prediction
(RMSEP). Evaluation of linearity through the traditional calibra-
tion curves (signal versus concentration plots) for multivariate
models is not possible. Apart from the already mentioned
pseudo-calibration curves, the most used way of assessing
linearity is the verication of the random behaviour of the
residuals from the t of the reference versus predicted values.
The sensitivity (SEN) of multivariate methods is estimated as
the NAS at unit concentration. For inverse calibration methods,
such as PLS, SEN can be calculated from the inverse of the
norm of the regression vector:
SEN = ||ns|| = 1 / ||b|| (5)
where || || indicates the Euclidian norm of a vector. However,
since SEN is dependent on the particular conditions of the
method, such as analytical technique and chemical matrix,
it is hardly used for comparison purposes. A more conven-
ient FOM is the analytical sensitivity (), which is dened, by
analogy with univariate calibration,
22
as the ratio between SEN
and the instrumental noise (). can be estimated through the
pooled standard deviation of a vector containing ten replicate
spectra of the blank.
21
= SEN / (6)
The inverse of (
1
) provides an estimation of the minimum
concentration difference that is distinguished by the analyt-
ical method considering the random experimental noise as
the only source of error, regardless of the specic technique
employed.
Also, based on NAS, the limit of quantication (LOQ) can be
estimated as:
LOQ = 10 / ||ns|| (7)
Bias is a term that evaluates systematic errors and is dened
as the difference between the limiting mean (the asymptotic
value or population mean of the distribution that character-
ises the measured quantity) and the true value. Bias is calcu-
lated only in the validation set and in this work was estimated
according to ASTM:
15
( )
1
bias (8)
v
n
i
v
n =

=
ref
y y
i i
where y
i
ref
and y
^
are the property values of reference and
predicted by PLS model, respectively, and n
v
is the number of
samples in the validation set. Standard deviation of validation
errors (SVD) is also calculated and used in a t test to determine
if the validation estimates show a statistically signicant bias.
( )
bias
9
1

( )
ref
i i
v
SVD
n




=

y y
The estimated t value is then compared to the critical t value
with the adequate number of degrees of freedom, which is
equal to n
v
.
Experimental
Insulating oil samples and reference method
Forty-eight insulating oil samples were provided by Centrais
Eltricas de Gois S.A., a state electricity power company.
Each sample was obtained from an independent transformer.
These samples were collected between March and September
2008, during the local dry season, from transformers located
in the whole State of Gois in the centralwest region of Brazil.
The transformers showed a wide variety of conditions of degra-
dation, as evidenced by the observed large range of interfacial
tension values and by their lifetimes, which varied from 1 to
about 30 years. Interfacial tension of the wateroil in these
samples was measured by using a torsion Krss K8 tensiom-
eter, according to Brazilian norm NBR 6234.
14
Instrumentation
Spectra of oil samples were obtained with a FT-NIR spectrom-
eter (Perkin Elmer Spectrum 100N, Shelton, USA) and a quartz
cuvette of 1.00 cm optical path, at a temperature of 24 1C.
Each spectrum was measured in the range from 1280 nm to
2500 nm (7800 cm
1
to 4000 cm
1
) with a resolution of 4 cm
1
and
32 scans. Instrumental noise was estimated using ten repli-
cate spectra of the empty cuvette (blank). Data were handled
with MATLAB software, version 7.9 (The MathWorks, Natick,
USA), and PLS Toolbox, version 5.2 (Eigenvector Technologies,
Manson, USA). A home-made routine was also used for the
detection of outliers.
17
Results and discussion
New insulating oils present a high interfacial tension, while
degraded oils show lower values due to the presence of polar
substances and tiny particles. In this work, the experimental
values measured for oil samples varied from 15 dyn cm
1

to 45 dyn cm
1
. The reference method
14
used in this study
has low precision, providing results with only two signicant
digits. Nevertheless, this precision is considered appropriate
by the electricity power company, since the essential decision
they make based on this analysis is whether the transformer
should be discarded or retained. The minimum allowable
limits for interfacial tension are 22 dyn cm
1
for the trans-
formers rated below 69 kV, and 25 dyn cm
1
for those rated
above 69 kV.
23
For comparison purposes, the detection limits
246 Determining Interfacial Tension of Transformers Using Insulating Oils
for methods based on NIR spectra used to be about 0.1%
(m m
1
).
24
NIR spectra of the 48 insulating oil samples are shown in
Figure 1. In this gure, it is possible to assign the intense
bands in the region between 2250 nm and 2500 nm to the
combinations of CH stretching and CH bending, and the
bands between 1680 nm and 1890 nm to the rst overtone of
methyl, methylene and methyne CH stretching.
25
Weaker
bands can be assigned to combinations of CH and CC
stretching and associated to CH of aldehydes (21302200 nm),
and to the combinations of rst overtone of methyl, methylene
and methyne CH stretching (13601430 nm).
By using the KennardStone algorithm,
26
samples were
divided into sets of 38 and 10 for the calibration and valida-
tion sets, respectively. It is important to emphasise that this
algorithm was the only tool used in this work for selecting
the most representative samples for the calibration set and
considered the entire spectral region. The number of LVs
was selected by leave-one-out cross-validation and the best
PLS model was obtained using only mean centring as data
preprocessing. Other preprocessing methods, multiplicative
scatter correction, standard normal variate and rst deriva-
tive with SavitskyGolay smoothing, were also tried, but the
resulting PLS models provided errors similar or larger than
the previous model. In the following, this model was opti-
mised by the detection of outliers.
17
For the rst round, four
outliers were detected in the calibration set, two based on
their spectral (X) residuals and two based on their resid-
uals in Y block. This is depicted in Figure 2, where it is also
possible to observe that no sample showed leverage above
the critical limit. No outliers were detected in the validation
set. After deleting these four samples, the model was recon-
structed and no more outliers were detected. Table 1 shows
the changes after this process through the results for the
PLS model. With 10 LVs, the model accounted for 99.98% and
96.54% of the total data variance in X and Y blocks, respec-
tively. The use of a relatively high number of LVs in the model
was indicative of the complex chemical composition of these
oils, since the use of fewer LVs signicantly degraded the
cross-validation performance. This model provided errors
of prediction between 9% and 12% for the validation
samples (Table 2), values considered acceptable by elec-
trical energy companies. These results, RMSEP of 2 dyn cm
1

and coefcient of determination (R
2
) between reference and
predicted values of 0.962, demonstrated that the proposed
1400 1600 1800 2000 2200 2400
0
0.5
1
1.5
2
Wavelength (nm)
A
b
s
o
r
b
a
n
c
e
Figure 1. NIR spectra of the 48 insulating oil samples analysed in this work.
Model First Final
Number of calibration samples 38 34
Number of validation samples 10 10
Number of latent variables 9 10
RMSEC (dyn cm
1
) 2.3 1.6
RMSEP (dyn cm
1
) 2.1 1.9
R
2
0.935 0.962
Table 1. Results for the optimisation of the partial least squares
model through the detection of outliers.
M.S. Godinho, A.E. Oliveira and M.M. Sena, J. Near Infrared Spectrosc. 19, 243251 (2011) 247
method was superior to a previous method based on image
analysis and N-PLS, which presented an RMSEP of 3 dyn cm
1

and R
2
= 0.903.
12
Figure 3 shows the regression coefcients of the devel-
oped model in relation to wavelength. It can be noted that
the spectral regions that most contributed to this model
were in the ranges from 1380 nm to 1430 nm and from
2300 nm to 2450 nm. Several local PLS models were built
with different spectral regions and these presented results
worse than the one based on data from the whole spec-
trum, with the exception of the model limited to the range
from 1380 nm to 1430 nm, which presented similar results
(RMSEP of 2 dyn cm
1
). It is interesting to observe in Figure
1 that this spectral range corresponds to a band of lesser
intensity, which may be associated with the presence of
2-furfuraldehyde, an indicator of the degradation of the
insulating paper.
The PLS model based on the whole spectra was validated
through the estimation of FOM. The accuracy was evaluated
based on RMSEC and RMSEP (Table 1). The precision was
evaluated at two levels, repeatability and intermediate preci-
sion. For repeatability, a relative standard deviation (RSD) of
3.2% was obtained for six replicates of a sample of interme-
diate interfacial tension (30 dyn cm
1
). For intermediate preci-
sion, another six replicates were analysed on another day by a
different analyst, providing a RSD of 3.9%. Although it cannot be
considered a quantitative measure of linearity, the t of refer-
ence versus predicted interfacial tension values expresses the
average agreement of the model (Figure 4). More appropri-
ately, the verication of the random behaviour of the residuals
distribution, observed in Figure 5, conrmed the linearity of
the model. The sensitivity was calculated based on Equation
(5) as 0.046. This value was used jointly with the estimation of
Reference value
(dyn cm
1
)
Predicted
(dyn cm
1
)
Error
(%)
18 17 6
20 21 5
24 25 4
29 28 3
31 29 6
33 35 6
33 37 12
40 41 3
42 44 5
43 39 9
Table 2. Reference and predicted interfacial tension values for
validation samples.
0 0.5 1
0
0.5
1
1.5
2
2.5
3
3.5
4
Leverage
F
r
e
q
u
e
n
c
y
0 2 4 6
x 10
-4
0
1
2
3
4
5
6
7
X residuals
Y

r
e
s
i
d
u
a
l
s
(b)
(a)
Figure 2. Visualisation of the outlier detection for the calibration set. (a) Histogram of leverage values. (b) Plot of spectral residuals (X)
versus interfacial tension residuals (Y). The solid lines indicate the acceptance limits for outlier detection.
248 Determining Interfacial Tension of Transformers Using Insulating Oils
1400 1600 1800 2000 2200 2400
-200
-150
-100
-50
0
50
100
150
Wavelength (nm)
C
o
e
f
i
c
i
e
n
t
s

o
f

r
e
g
r
e
s
s
i
o
n
Figure 3. Regression vector for the developed PLS model.
15 20 25 30 35 40 45
15
20
25
30
35
40
45
Reference interfacial tension (dyn cm
-1
)
P
r
e
d
i
c
t
e
d

i
n
t
e
r
f
a
c
i
a
l

t
e
n
s
i
o
n

(
d
y
n

c
m
-
1
)
Figure 4. Plot of reference versus predicted values for the calibration (circles) and validation (triangles) samples. The solid line shows
the data t and the dashed line the ideal t (slope of 1 and zero intercept).
M.S. Godinho, A.E. Oliveira and M.M. Sena, J. Near Infrared Spectrosc. 19, 243251 (2011) 249
15 20 25 30 35 40 45
-5
-4
-3
-2
-1
0
1
2
3
4
Reference interfacial tension (dyn cm
-1
)
A
b
s
o
l
u
t
e

E
r
r
o
r

o
f

P
r
e
d
i
c
t
i
o
n

(
d
y
n

c
m
-
1
)
Figure 5. PLS residuals for the calibration (circles) and validation (triangles) samples.
0.01 0.015 0.02 0.025 0.03
15
20
25
30
35
40
45
NAS Norm
R
e
f
e
r
e
n
c
e

V
a
l
u
e
s

(
d
y
n

c
m
-
1
)
Figure 6. Pseudo-univariate calibration curve. Plot of NAS norms versus reference values for the calibration (circles) and validation
(triangles) samples.
250 Determining Interfacial Tension of Transformers Using Insulating Oils
instrumental noise (0.045) for calculating as 1.02 cm dyn
1
.
The inverse of , 1 dyn cm
1
, indicates the minimum inter-
facial tension difference that the method is able to distin-
guish considering the random instrumental noise as the only
source of error. The LOQ was estimated using Equation (7) as
10 dyn cm
1
, which is consistent with what was expected.
12
Bias
was estimated as 0.160 2.379 and a t test with 10 degrees of
freedom and 95% condence level demonstrated that there is
no signicant errors in the model (t
calc
= 0.212 < t
crit
= 2.228).
Finally, a pseudo-univariate calibration curve was obtained
(Figure 6), demonstrating a simple manner to express this
multivariate model. The values of norm of NAS are equivalent
to selective signal values obtained from each sample spec-
trum vector. The t of this curve is presented in the following
equation:
interfacial tension = 1463.7 ||ns|| + 1.0462 (10)
Conclusions
A simple, direct, rapid and non-destructive NIR method was
developed to determine the interfacial tension of insulating
oils, which is an important tool for diagnosing the conditions of
electricity power transformers. The method was validated by
multivariate calibration and was considered accurate, precise
and linear in the range from 15 dyn cm
1
to 45 dyn cm
1
. The
method satises the requirements for the routine use in the
quality control laboratory of an electricity power company,
providing errors of prediction smaller than 10%. Based on
the concept of NAS, an exact presentation of the PLS model
as a univariate calibration graph was also provided, making
it easier to interpret the results. Finally, this work leads to the
possible implementing of a non-invasive on-line assessment
of the conditions of transformer insulating systems.
Acknowledgements
The authors thank Centrais Elticas de Gois S.A. (CELG)
for providing the insulating oil samples, Dr R.J. Poppi
(IQ-UNICAMP, Campinas) for the use of the spectropho-
tometer, Dr J.W.B. Braga (IQ-UnB, Braslia) for the MATLAB
routine for outliers detection, and M.R. Blanco (quality control
laboratory of CELG, Goinia) for valuable discussions.
References
1. A.M. Emsley, X. Xiao, R.J. Heywood and M. Ali,
Degradation of cellulosic insulation in power transform-
ers. Part 2: Formation of furan products in insulating
oil, IEE Proc. Sci. Meas. Technol. 147, 110 (2000). doi:
10.1049/ip-smt:20000259
2. R. Sanghi, Chemistry behind the life of a transformer,
Resonance 8, 17 (2003). doi: 10.1007/BF02837865
3. S. Levchik, J. Scheirs, G. Camino, W. Tumiatti and M.
Avidano, Depolymerization processes in the thermal
degradation of cellulosic paper insulation in electrical
transformers, Polym. Degrad. Stab. 61, 507 (1998). doi:
10.1016/S0141-3910(97)00249-8
4. T.K . Saha and P. Purkait, Understanding the impacts of
moisture and thermal ageing on transformers insula-
tion by dielectric response and molecular weight meas-
urements, IEEE Trans. Dielect. Elect. Insul. 15, 568 (2008).
doi: 10.1109/TDEI.2008.4483479
5. T.K. Saha, M. Darveniza, Z.T. Yao, D.J.T. Hill and G. Yeung,
Investigating the effects of oxidation and thermal deg-
radation on electrical and chemical properties of power
transformers insulation, IEEE Trans. Power Del. 14, 1359
(1999). doi: 10.1109/61.796229
6. A.M. Emsley, M. Ali and R.J. Heywood, A size exclusion
chromatographic study of polymer degradation, Polymer
41, 8513 (2000). doi: 10.1016/S0032-3861(00)00243-3
7. IEC, Insulating liquidsoil impregnated paper and
pressboard determination of water by automatic
coulometric Karl Fischer titration. IEC Standard 60814
(1997).
8. S. Singh and M.N. Bandyopadhyay, Dissolved gas analy-
sis technique for incipient fault diagnosis in power trans-
formers: A bibliographic survey, IEEE Electr. Insul. Mag.
26, 41 (2010). doi: 10.1109/MEI.2010.5599978
9. A.M. Emsley, X. Xiao, R.J. Heywood and M. Ali,
Degradation of cellulosic insulation in power transform-
ers. Part 3: Effects of oxygen and water on ageing in oil,
IEE Proc. Sci. Measur. Techn. 147, 115 (2000). doi: 10.1049/
ip-smt:20000021
10. P.J. Baird, H. Herman, G.C. Stevens and P.N. Jarman,
Non-destructive measurement of the degradation of
transformer insulating paper, IEEE Trans. Dielect. Elect.
Insul. 13, 309 (2006). doi: 10.1109/TDEI.2006.1624275
11. P.J. Baird, H. Herman, G.C. Stevens and P.N. Jarman,
Spectroscopic measurement and analysis of water
and oil in transformer insulating paper, IEEE Trans.
Dielect. Elect. Insul. 13, 293 (2006). doi: 10.1109/
TDEI.2006.1624274
12. M.S. Godinho, A.E. Oliveira and M.M. Sena,
Determination of interfacial tension of insulating
oils by using image analysis and multi-way cali-
bration, Microchem. J. 96, 42 (2010). doi: 10.1016/j.
microc.2010.01.017
13. ABNT, Associao Brasileira de Normas Tcnicas. NBR
10576: leo mineral isolante de equipamentos eltri-
cosDiretrizes para superviso e manuteno, Brazil
(2006).
14. ABNT, Associao Brasileira de Normas Tcnicas. NBR
6234: Determinao da tenso interfacial de leo gua,
Brazil (1965).
15. Annual Book of ASTM Standards, Standard Practices for
Infrared Multivariate Quantitative Analysis, E1655-05. ASTM
International, West Conshohocken, Pennsylvania, USA
(2005).
M.S. Godinho, A.E. Oliveira and M.M. Sena, J. Near Infrared Spectrosc. 19, 243251 (2011) 251
16. J.A.F. Fernndez Pierna, F. Wahl, O.E. Noord and D.L.
Massart, Methods for outlier detection in prediction,
Chemometr. Intell. Lab. Syst. 63, 27 (2002).
17. P. Valderrama, J.W.B. Braga and R.J. Poppi, Variable
selection, outlier detection, and gures of merit estima-
tion in a partial least-squares regression multivariate
calibration model. A case study for the determination of
quality parameters in the alcohol industry by near-infra-
red spectroscopy, J. Agric. Food Chem. 55, 8331 (2007).
doi: 10.1021/jf071538s
18. R. Bro and C.M. Andersen, Theory of net analyte signal
vectors in inverse regression, J. Chemometr. 17, 646
(2003). doi: 10.1002/cem.832
19. N.M. Faber, Exact presentation of multivariate calibra-
tion model as univariate calibration graph, Chemometr.
Intell. Lab. Syst. 50, 107 (2000). doi: 10.1016/S0169-
7439(99)00054-4
20. W.F.C. Rocha, A.L. Rosa, J.A. Martins and R.J. Poppi,
Determination and validation of nimesulide in pharma-
ceutical formulation by near infrared spectroscopy, J.
Braz. Chem. Soc. 21, 1929 (2010). doi: 10.1590/S0103-
50532010001000019
21. P. Valderrama, J.W.B. Braga and R.J. Poppi, State of
the art of gures of merit in multivariate calibration,
Quim. Nova 32, 1278 (2009). doi: 10.1590/S0100-
40422009000500034
22. L.C. Rodriguez, A.M.G. Campaa, C. Jimenez-Linares
and M. Romn-Ceba, Estimation of performance char-
acteristics of an analytical method using the data set of
the calibration experiment, Anal. Lett. 26, 1243 (1993).
doi: 10.1080/00032719308019900
23. Annual Book of ASTM Standards, Standard Test
Method for Interfacial Tension of Oil Against Water by
Ring Method D971-99a. ASTM International, West
Conshohocken, Pennsylvania, USA (2004).
24. C. Pasquini, Near infrared spectroscopy: Fundamentals,
practical aspects and analytical applications, J.
Braz. Chem. Soc. 14, 198 (2003). doi: 10.1590/S0103-
50532003000200006
25. J.J. Workman, Jr, Interpretative spectroscopy for
near infrared, Appl. Spectrosc. Rev. 31, 251 (1996). doi:
10.1080/05704929608000571
26 R.W. Kennard and L.A. Stone, Computer aided design
of experiments, Technometrics 11, 137 (1969). doi:
10.2307/1266770
JOURNAL
OF
NEAR
INFRARED
SPECTROSCOPY
253
ISSN: 0967-0335 IM Publications LLP 2011
doi: 10.1255/jnirs.936 All rights reserved
Agricultural soils, having been depleted of much of their native
C stocks, have a signicant CO
2
sink capacity.
1
For this reason,
the quantication of C content in various types of soil, according
to climate and soil management, is of great interest. Moreover,
C sequestration in agricultural soils will also benet productivity
and sustainability, and trading of C credits or tax benets could
be a source of income for agricultural producers.
2
Awarding
these benets would require the monitoring of soil organic C
content. It has been estimated that a soil C monitoring system,
including all US farms, using existing analytical techniques,
could cost more than $1 billion for the rst four-year commit-
ment period.
3
Clearly, a reduction in per-sample cost of analysis
Soil carbon determination in a
Mediterranean vertisol by visible and near
infrared reflectance spectroscopy
Jos M. Fontn, Luis Lpez-Bellido,
*
Juan Garca-Olmo and Rafael J. Lpez-Bellido
Departamento de Ciencias y Recursos Agrcolas y Forestales, Campus Rabanales, edicio C-4, Crta. Madrid km 396, CP: 14071,
University of Crdoba, Spain; and NIR/MIR Spectroscopy Unit Central Service for Research Support (SCAI), Campus Rabanales,
Edicio Ramn y Cajal, Crta. Madrid km 396, CP: 14071, University of Crdoba, Spain. E-mail: cr1lobel@uco.es
The monitoring of the C content of agricultural lands requires long-term experiments, the analysis of a large number of samples,
various sampling depths and adequate analytical tools to perform the experiments. Near infrared (NIR) reectance spectroscopy
is a fast, reliable and low-cost analytical tool for the determination of soil C content. The objective of this study was to present NIR
spectroscopy technology as a reliable method to monitor C concentration (total C, inorganic C and organic C) of soil samples from a
long-term experiment established in 1986 in Crdoba (Spain). The soil is a rain-fed vertisol, characteristic of Mediterranean agricul-
tural systems. The soil samples studied from four different years (1997, 2000, 2003 and 2006) included all of the variabilities existing
in the long-term experiment: tillage system (conventional tillage vs no tillage), crop rotations [wheat (Triticum aestivum L)-chickpea
(Cicer arietinum L), wheat-faba bean (Vicia faba L), wheat-sunower (Helianthus annus L), wheatfallow and wheatwheat], N fertiliser
rate (0 kg N ha
1
, 50 kg N ha
1
, 100 kg N ha
1
and 150 kg N ha
1
), as well as different soil depths (015 cm, 1530 cm, 3060 cm and 6090 cm).
Calibration models were obtained with 492 samples from the aforementioned vertisol using visible and NIR spectroscopy (vis-NIR) and,
when tested using a validation sample set, returned the values: total C (r
2
= 0.92, SEP = 1 g kg
1
, RER = 26, RPD = 4.8); inorganic C (r
2
= 0.95,
SEP = 1.5 g kg
1
, RER = 26.6, RPD = 4.6) and organic C (r
2
= 0.76, SEP = 0.8 g kg
1
RER = 12.3, RPD = 2). The vis-NIR models developed reliably
predicted the total, inorganic and organic C content at plot scale (homogeneous sample set). The advantage deriving from very specic
vis-NIR calibrations for prediction of soil C is that one can rapidly and economically predict C in the study area, enabling monitoring to
compare soil C evolution depending on time and variables such as tillage, crop rotation and fertilisation. To develop vis-NIR models at
local and regional scales would require a more heterogeneous sample set and adequate calibration/validation strategies.
Keywords: vis-NIR, soil carbon, Mediterranean, vertisol, carbon sequestration, monitoring, organic carbon, inorganic carbon, total carbon, CO
2
Introduction
J.M. Fontn et al., J. Near Infrared Spectrosc. 19, 253263 (2011)
Received: 15 November 2010 Revised: 22 June 2011 Accepted: 20 September 2011 Publication: 27 September 2011
254 Soil Carbon Determination in Mediterranean Vertisol
for soil C, coupled with a quicker sampling and sample prepara-
tion method, would be welcome in this context.
During more than two decades of the application of near
infrared (NIR) reectance spectroscopy to soil, a relatively
large and exponentially growing literature has demon-
strated that a number of constituents, parameters and
functions in soils can be successfully determined by NIR
spectroscopy, resulting in signicant benets in time and
cost savings and increased exibility.
2
For these reasons,
NIR spectroscopy would be practical and cost-effective to
use for mapping the large numbers of samples necessary
for landscape-scale monitoring over large geographical
areas,
46
allowing a greater resolution for C determination
on the landscape and for simultaneous determination of
multiple parameters.
Long-term eld experiments are very useful for the moni-
toring of soil C content according to different soil manage-
ments. This type of experiment is of special interest in
semi-arid environments, such as the Mediterranean, given
the low rate of increase in organic C in these climates.
7

Increasing demand for environmental assessment and
monitoring, and emerging environmental issues in the face
of xed or declining budgets, may be encouraging the expo-
nential increase in publications on soil NIR spectroscopy
applications.
2
Recently, several large-scale programmes
have been applying the technology to large numbers of soil
samples over large geographical areas.
8,9
An example of
this is the development of the Global Soil Spectral Library,
which is reported by Stenberg et al.
10
There are reports in the literature of studies which have
used NIR spectroscopy to determine the content of total
C,
1115
organic C
9,16
and inorganic C
17 20
in different types of
soil from different environments. There are authors who have
studied analysis using NIR on a local scale.
21 26
Nevertheless,
currently, few studies exist on the application of NIR spec-
troscopy technology in the vertisols of Mediterranean areas
to monitor the C content. These soils are characterised by
their low content of organic matter.
7
Likewise, studies which
include soil sampling performed below the topsoil (30 cm)
are not frequently found, although, as indicated by Baker et
al.,
27
these deeper horizons must be more widely studied for
the sake of clarifying C sequestration for different tillage
systems.
The aim of this study was to present NIR spectros-
copy technology as a reliable method to monitor soil C
concentration (total C, inorganic C and organic C) of soil
samples from a long-term experiment established in 1986
in Crdoba (Spain), on a Mediterranean vertisol soil, Typic
Haploxeret,
28
on a plot scale. For this purpose, soil samples
from a long-term eld experiment, which started 22 years
ago in southern Spain, were used. The advantage deriving
from very specic local vis-NIR calibrations for predictions
of soil C is it can rapidly and economically predict C in the
study area, which would enable monitoring to compare soil
C evolution depending on time and variables as tillage, crop
rotation and fertilisation.
Material and methods
Site characteristics
The soil samples studied belong to a long-term eld experi-
ment called Malagn (total area 2.5 ha), located in Crdoba,
southern Spain (37 46 N, 4 31 O, 340 m above sea level). The
soil is a vertisol (Typic Haploxeret), typical of the Mediterranean
region. This field experiment, which began in 1986, was
designed as a randomised complete block with a splitsplit
plot arrangement and four blocks. The main plots used the
tillage system (no-tillage and conventional tillage); sub-plots
were preceding crops, with four two-year rotations (wheat
sunower, wheatchickpea, wheatfaba bean and wheatbare
fallow) and continuous wheat; sub-sub-plots were N fertiliser
rate (0 kg N ha
1
, 50 kg N ha
1
, 100 kg N ha
1
and 150 kg N ha
1
)
applied to wheat. Duplicate sets of plots were established to
allow all phases of the rotation to be present each year. The
area of each sub-sub-plot was 50 m
2
(10 m by 5 m).
Soil samples
A total set of 2168 soil samples were taken using a manual auger
(Eijkelkamp; Giesbeek,The Netherlands) during the month of
November in four different years (1997, 2000, 2003 and 2006)
and four depths (015 cm, 1530 cm, 3060 cm and 6090 cm).
For each year, the samples used were: 480 in both 1997 and
2000 from wheat plots (two tillage systems ve crop rota-
tions four N fertiliser rates three replications four depths);
648 in 2003, 520 from wheat plots [two tillage systems ve crop
rotations four N fertiliser rates four replications one depth
(015 cm) and two tillage systems ve crop rotations four N
fertiliser rates three replications three depths (1530 cm,
3060 cm and 6090 cm)] and 134 from faba bean, chickpea,
sunower and fallow plots [two tillage systems four N ferti-
liser rates four replications one depth (015 cm) in each
crop]; 560 in 2006 from wheat plots [two tillage systems ve
crop rotations four N fertiliser rates four replications two
depths (015 cm and 1530 cm) and two tillage systems ve
crop rotations four N fertiliser rates three replications two
depths (3060 cm and 6090 cm). All the soil samples were
air-dried, ground, sieved (2 mm) and frozen at 30C until
analysis.
The total set (n = 2168 samples) was divided into three groups:
a calibration set, a validation set and the remaining samples.
The calibration set consisted of 492 samples and the validation
set consisted of 161 samples. The remaining samples were:
2168 (492 + 161) = 1515)samples.
Reference analytical method
Total C of each soil sample in the calibration set and in the vali-
dation set was determined by the dry combustion method. An
elemental analyser (EA 3000 Eurovector SpA, Milan, Italy) was
used for this purpose. All of the samples were air-dried and
stored in plastic bags until analysis. The previous processing
of the samples consisted of a very ne grinding with an agate
mortar. Part of the ground sample was incinerated in a mufe
furnace at 450C for 16 h, to eliminate all of the organic C
J.M. Fontn et al., J. Near Infrared Spectrosc. 19, 253263 (2011) 255
present.
29
The non-incinerated and incinerated samples were
analysed by dry combustion, determining total C and inorganic
C, respectively. The organic C of the sample was calculated as
the difference between the two values.
The standard error of laboratory (SEL) was calculated as the
uncertainty estimate of the reference method. For this, two
subsamples of each sample were analysed on different days.
The formula used to calculate this error was:
( )
1 2
2 / SEL y y N

=


y
1
and y
2
being the values obtained for a sample and its repeti-
tion and N, the number of data pairs used to calculate the
SEL.
NIR Spectroscopy analysis
All samples (n = 2168) were scanned with a Foss NIRSystems
6500 monochromator [Foss NIRSystems, Silver Spring (now
Laurel), MD, USA] using the WinISI II 1.5 software [Infrasoft
International, Port Matilda (now State College), PA, USA].
This software was also used for the treatment of the spectral
data obtained. The visible and NIR (vis-NIR) spectra (400 nm
to 2500 nm) were collected at 2 nm intervals. Soil samples
were placed in a rectangular cell with a 4.5 5.5 cm
2
quartz
window. Two readings were taken in two different positions,
recording the average of the two readings. A quarter cup
check cell (Infrasoft International LLC, USA) was used to check
FossNIRSystems 6500 monochromator performance over the
course of the scanning.
Spectral data analysis
Spectral data analysis and vis-NIR equation calculations were
performed with WinISI II 1.5 software (Infrasoft International,
Port Matilda, PA, USA).
Spectral outliers of the total sample set (2168 samples) were
detected and eliminated. This was done by applying a standard
normal variate and detrending scatter correction (SNV and D)
to all of the spectral data and using a mathematical pretreat-
ment of 1,4,4,1. The four numbers respectively designate the
derivative degree (first derivative in this case), wavelength
data point gap over which the derivative is calculated, number
of data points in a smoothing correction and number of data
points in a second smoothing. Afterwards, a principal compo-
nent analysis (PCA) was applied and the Mahalanobis distance,
H, was calculated (also referred to as H statistic).
30
Global H
is the multivariate distance of a sample spectrum from the
centroid of a total sample population. Samples with global
H > 3 were considered spectral outliers and were eliminated
from further investigations.
31
After eliminating the spectral outliers, a calibration set and
a validation set were chosen from the total set. The calibra-
tion set included the most representative samples from the
total set and was selected using the WinISI software SELECT
procedure.
31
This procedure consists of applying a principal
component analysis (PCA)
32
to calculate the neighbourhood
H distance of each sample (the Euclidian distance from the
closest sample). The most representative samples of the set
can be chosen according to the neighbourhood H distances
calculated. The calibration set (n = 492 samples) included all
agronomic treatments (tillage systems, crop rotations and
N fertiliser rates), years (1997, 2000, 2003 and 2006) and soil
depths (015 cm, 1530 cm, 3060cm and 6090 cm). The vali-
dation set was formed by removing 161 soil samples. They
were chosen at random from the collection.
Spectral data of the calibrations set were subjected to
different spectral pre-treatments combined with different
mathematical pre-treatments. Standard normal variate and
detrending scatter correction (SNV and D)
33
were used as
spectral pre-treatments. As mathematical pre-treatments,
first and second derivatives were used to reduce baseline
variation and enhance spectral features.
12
The option of not
applying SNV and D (None) or any derivative was also tested.
A total of eight pre-processing methods were evaluated: SNV
and D (1,4,4,1), None (1,4,4,1), SNV and D (2,10,5,1), None
(2,10,5,1), SNV and D (2,4,4,1), None (2,4,4,1), SNV and D
(0,0,1,1) and None (0,0,1,1). The four numbers in parenthesis
respectively designate derivative degree (0 for no derivative,
1 for rst derivative and 2 for second derivative), wavelength
data point gap over which the derivative is calculated, number
of data points in a smoothing correction and number of data
points in a second smoothing.
Calibration, in all cases, was based on cross-validation
(ve groups of cross-validation). Modied partial least square
regression (MPLSR)
31
was used to relate spectra with the
measured values for total C, inorganic C and organic C.
Multivariate regression was carried out using the PLS algo-
rithm. All spectral data were summarised in a few variables
(PLS terms) with greater correlation with reference values.
The residues obtained after the calculation of each regres-
sion term in the MPLS regression were standardised dividing
by the standard deviation before the calculation of the next
regression term.
33
For each C type, regression coefcients
of the MPLS model were represented based on the spectral
information of the soil samples used in the calibration. In each
C type, this representation showed the wavelengths of the soil
spectra that had the strongest correlation with the laboratory
C measurement.
For each type of C, the best pre-treatment method was
chosen based on the method which produced models with
higher R
2
values and smaller standard errors of cross-vali-
dation (SECV). The number of terms was xed when the SECV
during the cross-validation procedure presented a minimum
value before over-tting occurred according to Shenk and
Westerhaus
31, 34
.Moreover, the following indices were calcu-
lated: coefficient of variation (CV) calculated as the ratio
of SECV to the mean reference value over the calibration
sample set; ratio error range (RER) is the ratio of the range
of measured values in the validation set divided by the SECV
35

and ratio RPD, which is dened as the standard deviation
of the measured values of the validation set divided by the
SECV.
36,37
256 Soil Carbon Determination in Mediterranean Vertisol
During each of the calibration processes, the Student's
t-statistic was used to detect chemical outliers (samples
with large differences between the reference and predicted C
values). Those samples which exhibited signicant differences
between the reference and predicted C values (Student's
t-statistic value greater than 2.5) were considered chem-
ical outliers and, therefore, eliminated from the calibration
process. After removal of outliers, there were still a large
number of samples in the calibration set.
The prediction accuracy of the model was evaluated on an
independent validation set which had not been used for cali-
bration model development. From the validation set (n = 161
samples), the standard error of prediction between predicted
and measured values (SEP), bias (difference between the mean
measured data and predicted data values), r
2
, RER (range/SEP),
RPD (Std dev./SEP) and CV (mean/SEP) were calculated.
Results and discussion
Soil sample characteristics
The set of soil samples used for calibration presented a
narrow range of variation for organic C, as well as excep-
tionally low content average (Table 1), as is typical of rainfed
Mediterranean vertisol soils.
7
It should be pointed out that
the average organic C content of the soil samples used in
other similar studies were approximately four times higher
20,38
(Table 2). For total C and for inorganic C (Table 1), the soil
samples presented wide ranges and averages more consistent
with those obtained for other types of soil (Table 2).
The average, range and standard deviation of the set of
soil samples used to validate the vis-NIR calibration models
presented, for each type of C, values consistent with those for
the set of samples used to make the calibrations (Table 1).
Figure 1 shows mean spectra of soil samples analysed for
the different soil depths. A principal component analysis on
the total set (n = 2168 samples) is represented in Figure 2. No
clear differences between years were observed. The rst and
second principal components (PC1 and PC2) explained about
80% of the variation; PC1 49% and PC2 32%.
Vis-NIR models
Equations were obtained for each type of C studied from 492
soil samples belonging to the calibration set, which included
all agronomic treatments (tillage systems, crop rotations and
N fertiliser rates), years (1997, 2000, 2003 and 2006) and soil
depths (015 cm, 1530 cm, 3060 cm and 6090 cm). The
pre-processing methods which obtained the best results in
Carbon type Calibration set Validation set
n
1
Mean
(g kg
1
)
Range
(g kg
1
)
Std dev.
(g kg
1
)
SEL
(g kg
1
)
n
2
Mean
(g kg
1
)
Range
(g kg
1
)
Std dev.
(g kg
1
)
Total C 492 17.4 2.4 45.3 6.7 1.0 161 18.9 13.6 42.2 5.3
Inorganic C 492 11.6 1.2 41.0 6.9 0.8 161 12.3 4.9 39.0 6.0
Organic C 492 5.8 0 16.6 2.3 0.5 161 6.1 1.9 10.5 1.4
n
1
:

number of samples used in calibration;
n
2
: number of samples used in validation;
SEL: standard error of laboratory reference method.
Table 1. Summary statistics for C types of vertisol analysed.
Figure 1. Visible and NIR spectra of vertisol samples sourced from different soil depths. Malagn long-term experiment, Crdoba
(Spain).
J.M. Fontn et al., J. Near Infrared Spectrosc. 19, 253263 (2011) 257
Author/s Carbon
type
Soil
characteristics
Sample
number
Error model
(g kg
1
)
Range
(g kg
1
)
r
2
/ R
2
RPD RER
Madari et al.
15
Total C Ferralsol soil (Brazil) 120 RMSD = 0.72 7.948.0 R
2
= 0.99
Chang et al.
8
Total C Soils from 448
sites in USA (Depth:
03 cm, 310 cm,
1030 cm)
743 SECV = 7.86 1.3 345.8 r
2
= 0.87 2.79 36.2
Barths et al.
14
Total C Soil samples from
Africa and America
Clayley and sandy
soils. (Depths:
05 cm, 510 cm,
1020 cm, 2030 cm,
6070 cm)
85 SECV = 5.66 5.545.2 r
2
= 0.91 3.34
Reeves et al.
48
Total C Silt loam soils
(020 cm). Tillage and
no tillage.
179 RMSD = 0.94 6.133.9 r
2
= 0.95 4.95 29.6
Genot et al.
49
Total C Four soils
characteristics in
Wallonia (Belgium)
1300 SEP = 11.4 r
2
= 070 1.8
Cozzolino and
Morn
47
Organic C
(in the clay
and silt
fraction)
Silty clay loam soil
(Argiudoll) (Depth:
07.5 cm, 7.515 cm).

87 SECV = 3 50480 R
2
= 0.96 3.2
Chang and Laird
21
Organic C Mixture of ve
surface soils
108 RMSEP = 6.2 15.4144.9 r
2
= 0.89 4.2 20.9
Fystro
50
Organic C 40 Grasslands elds.
Norway. (Depth:
025 cm; 2540 cm.)
80 SEP = 6.8 6.079.0 r
2
= 0.81 2.2 10.7
Confalonieri et al.
13
Organic C Sandy loam soil
(030 cm). Years:
1985 and 1997.
142 SECV = 0.6 5.114.5 r
2
= 0.87 2.5 15.7
Shepherd and
Walsh
9
Organic C African soils. (Depth:
015 cm, 020 cm)
674 RMSE = 2.2 2.355.8 r
2
= 0.91 24.3
Martin et al.
51
Organic C A horizon soil
(Canada)
130
149
274
SEP = 1.9
SEP = 3.3
SEP = 3.5
020
2040
3.837.1
r
2
= 0.80
r
2
=0.51
r
2
= 0.75
2.11
1.32
2
8.3
6.26
9.7
Ben-Dor and
Banin
16
Inorganic C 12 groups of Israeli
soils.
91 SEP = 79 0743 r
2
= 0.73 7.7
McCarty et al.
6
Inorganic C Soils from 14
locations in central
USA. (Depth: 05 cm,
510 cm, 1025 cm).
273 RMSD = 3.1 065.4 r
2
= 0.87 4.3 21.1
Chang and Laird
21
Carbonate
calcic
Mixture of ve
surface soils.
108 RMSEP = 1.54 035.7 r
2
= 0.96 5.5 23.2
SEP: standard error of prediction;
RMSD: root mean square deviation;
SECV: standard error of cross validation;
RMSEP: root mean square error of prediction;
RMSE: root mean square error
Table 2. NIR Spectroscopy models obtained for different authors in several soils and environments.
258 Soil Carbon Determination in Mediterranean Vertisol
the calibrations were different depending on the type of C
considered (Table 3). For total C and inorganic C, the pre-
processing method (1,4,4,1) with no scatter correction (None)
obtained the best results, whereas for organic C the best
pre-processing method was (1,4,4,1) with scatter correction
(SNV and D). All of these regression models showed a high
significance (P < 0.001). Spectral outliers eliminated in the
PCA applied were 53 soil samples. The number of chemical
outliers was: 22 samples for total C, 18 samples for inorganic
C and 51 samples for organic C (Table 3). Most of chemical
outliers were soil samples with low C values which would be
expected to have greater analytical error. This was especially
Figure 2. First and second principal components (PC1 and PC2) of principal component analysis (PCA) from global sample set (n = 2168)
of a Mediterranean rainfed vertisol.
Carbon
type
n Range
(g kg
1
)
Mean
(g kg
1
)
Std dev.
(g kg
1
)
SECV
(g kg
1
)
R
2
Slope
c
RPD RER
Total C
a
417 6.5 43.1 17.9 6.3 1.1 0.97
***
0.99 5.7 34.8
Inorganic C
a
421 3.2 41.0 11.9 6.8 1.3 0.96
***
1.03 5.2 29.4
Organic C
b
388 1.5 12.4 6 1.9 0.7 0.87
***
0.96 2.7 15.6
a
Optimal pre-processing: None (1,4,4,1);
b
Optimal pre-processing: standard normal variate and detrending (1,4,4,1);
***
means signicance at 0.001 probability level; n: number of samples used in calibration after the elimination of spectral and chemical outliers;
SECV: standard error of cross validation;
R
2
: Coefcient of determination;
c
Slope of the linear regression line
Number of factors in all cases was 4.
Table 3. NIR calibration equations developed by cross-validation and obtained from the long-term experiment Malagn (Crdoba,
southern Spain).
J.M. Fontn et al., J. Near Infrared Spectrosc. 19, 253263 (2011) 259
relevant in the organic C case, since its content was lower than
total or inorganic C (Table 1) and it had a higher relative error
(CV) associated with the conventional determination: organic
C is not measured directly, but calculated as the difference
between total and inorganic C. The number of eliminated
outliers (spectral and chemical) was not greater than the limit
recommended by Shenk and Westerhaus
31
(about 20% of cali-
bration sample set).
For total, inorganic and organic C, the errors for the vis-NIR
models obtained (SECV, Table 3) were 10%, 60% and 40%
higher than their respective standard errors of laboratory [SEL,
(Table 1)].
The vis-NIR models for total and inorganic C were more
precise than those obtained for organic C (Table 3), which
could be caused by different error sources. A wider range is
able to obtain higher coefcients of determination. However,
probably the principal cause was the indirect determination of
organic C as the difference between the total C and inorganic
C, which constitutes a source of error apparent in the lesser
predictive capacity of the models obtained for this parameter.
Also, the presence of CaCO
3
could complicate the organic
C NIR spectroscopy calibration, although, this effect is not
well known.
39
The coefcients of determination of the vis-NIR
models obtained for each type of C were very close to the
possible maximum, an especially noteworthy situation in the
cases of total and inorganic C (Table 3).
The wavelengths with the highest inuence in the MPLS
models for predicting total, inorganic and organic C are shown
in Figure 3. For organic C, wavelengths around 16001800 nm
were of great importance, probably due to the rst overtone
of the CH stretching band (CH group is associated with
organic matter). For total and inorganic C, the range with
the highest inuence was 10001800 nm. In the three types
of C, the visible zone of the spectrum (400700 nm) had a low
inuence on calibrations (Figure 3), caused possibly by the
absence of plant residues with pigments which are associated
with absorptions in the visible range.
11,18
According to Morra et
al.
40
and Stevenson,
41
wavelengths around 1700 nm would be
associated with the presence of some mineral clay. However,
Clarck
42
and Stenberg
43
reported that 1710 nm and 1725 nm
regions most probably correspond to organic compounds.
We observed important spikes around this wavelength
(Figure 3), which is logical taking into account the high clay
content in this vertisol (around 700 g kg
1
, see Lpez-Bellido
et al.
44
).
The predictive quality of the models obtained was conrmed
by the results from the validation performed (Figure 4). In
general, for this validation, the SEP and the r
2
(Figure 4), were
of the same order as the SECV and R
2
obtained in the calibra-
tion models developed by cross validation (Table 3).
The values of the dimensionless statistics RER and RPD of
the vis-NIR models were higher for total (26.0 and 4.8, respec-
tively) and inorganic C (26.2 and 4.6, respectively) than for
organic C (12.3 and 2.0, respectively), which is more bene-
cial as it means less model error (RER = range/SECV) and
(RPD = Std Dev/SEC). In all cases, the RPD values were greater
than 2 and were signicantly above this value in the case of
total C and inorganic C (Table 3).
Malley et al.
2
propose guidelines for interpreting RPD and
RER for environmental samples (such as soil). Excellent
calibrations are those with r
2
> 0.95, RPD > 4 and RER > 20.
Successful calibrations are those with r
2
= 0.90 to 0.95, RPD = 3
to 4 and RER = 15 to 20. Moderately successful calibrations have
r
2
= 0.80 to 0.90, RPD = 2.25 to 3 and RER = 10 to 15. According
to these guidelines, our NIR equations for total and inorganic
C were excellent calibrations. The NIR equation for organic C
could be considered as successful or moderately successful,
depending of the statistic evaluated.
Dunn et al.
45
studied the ability of NIR spectroscopy to
predict soil properties using a sample set from three types
of soil. They reported calibrations with RPD values (1.6) which
Figure 3. Modied partial least squares regression coefcients
for the vis-NIR spectra from calibration models for total, inor-
ganic and organic soil C in a Mediterranean rainfed vertisol.
260 Soil Carbon Determination in Mediterranean Vertisol
indicated poor predictions. Chang et al.
8
also evaluated the
ability of NIR spectroscopy to predict diverse soil properties
from a set of 802 samples from different soils. They suggested
grouping the results obtained into three categories. Category A
(RPD > 2.0) includes properties with measured vs predicted R
2

values between 0.80 and 1.00. Category B (RPD 1.42) includes
properties with measured vs predicted R
2
values between
0.50 and 0.80. Category C (RPD < 1.4) includes properties
with measured vs predicted R
2
values < 0.50. These authors
believe that the prediction of soil properties in category B (as
is the case for total and inorganic C in soil clods; Table 2) can
possibly be improved by using different calibration strategies,
but that properties in category C may not be reliably predicted
using NIR spectroscopy.
From vis-NIR models obtained, total, inorganic and
organic C of 2168 soil samples from the long-term experi-
ment Malagn were predicted. The prediction obtained an
average Mahalanobis global H of 1.607 (less than 3) and an
average neighborhood H of 0.586 (less than 0.6). These indices
would suggest a good t of the models and the reliability of
the NIR spectroscopy predictions obtained, given the similarity
between the predicted samples and those which composed
the set used for the calibration.
For the purpose of simplifying the comparison of our results
with those obtained by other authors, Table 2 shows the
most relevant NIR spectroscopy models obtained by different
authors for different locations and types of soil. As can be
observed in the table, rarely are soil sampling depths deeper
than 30 cm included. In order for the NIR spectroscopy models
to be of greater use in the analysis and monitoring of soil
C, soil samples from deeper horizons must be included and,
according to Brown et al.,
46
adequate calibration/validation
strategies must be designed to ensure that really independent
soil samples are included in the validation set. Baker et al.
27

stated that the quantification of C in deeper soil layers is
essential for the comparison of the C sequestration capacity
between different tillage systems.
Variability in the soil sampling also provides the NIR spec-
troscopy models with a better capacity to function as useful
tools for the quantification of soil C and for following the
evolution of its content over time with respect to different soil
management systems. This is the case of our study where,
in addition to the soil layer of the different depths studied
(015 cm, 1530 cm, 3060 cm and 6090 cm), we included soil
samples from two tillage systems (conventional tillage and
no-tillage), different crop rotations and N fertiliser rates and
different years.
The RPD and RER statistics for the calibrations and valida-
tions made for each of the types of C studied are of the same
order and, in many cases, better than those obtained by other
authors (Table 2). Only the NIR spectroscopy models for organic
C obtained by Chang and Laird
21
(for validation statistics:
RER = 20.9 vs 15.6; and RPD = 4.2 vs 2.7), Cozzolino and Morn
47

(for calibration statistics: RPD = 3.2 vs 2) and Shepherd and
Walsh
9
(for validation statistics: RER = 24.3 vs 15.6), exceeded
the models obtained in our study for this same type of C
(Table 2). Often, NIR spectroscopy models of heterogeneous
and multi-location sample sets have less predictive capacity.
Although these authors employed soil sample sets which were
more heterogeneous than our sample sets, they obtained
better results. A possible explanation for these better results is
that it could be attributed to the use of a much wider range of
values for organic C. The range for the organic C content in our
vertisol was 016.6 g kg
1
(Table 1), much lower than the ranges
presented in the three studies of the previously mentioned
authors (Table 2). Confalonieri et al.,
16
with a quite similar range
of organic C, showed similar results to ours (Tables 1 and 2).
Moreover, organic C results could be explained by the error
Figure 4. Comparison of values analysed by dry combustion
(actual value) and predicted by vis-NIR (predicted value) for
total C (g kg
-1
), inorganic C (g kg
-1
) and organic C (g kg
-1
) in a
Mediterranean rainfed vertisol.
J.M. Fontn et al., J. Near Infrared Spectrosc. 19, 253263 (2011) 261
associated with the conventional determination of this compo-
nent (which is not measured directly, but calculated as the
difference between total and inorganic C).
Table 4 shows a correlation matrix between total carbon,
organic carbon and inorganic carbon. Correlation between
total and inorganic carbon is close to unity. However, no corre-
lations were observed between total carbon vs organic carbon
and inorganic carbon vs organic carbon. It could be due to the
important contribution of inorganic carbon to the total carbon
pool in this kind of soil.
On the other hand, the good results obtained in this study
could be due to the characteristics of the soil sample set.
Although soil samples with different agronomic treatments,
soil depths and years were employed, the set was relatively
homogenous. The soil sample set was obtained in only one
location and type of vertisol from a long-term eld experiment.
Therefore, the application of the results is limited to this experi-
ment. According to Brown et al.,
46
an NIR spectroscopy model
obtained from small-scale sampling would over estimate C
values if it was used on a regional scale, for example, to predict
soil C in other vertisols with similar characteristics. In that
case, soil C prediction would produce high errors caused by the
dependence between calibration and validation sets.
Conclusions
NIR spectroscopy technology, due to its speed, precision and
low cost, can be considered an appropriate analytical method
for the determination and monitoring of the levels of C of rain-
fed Mediterranean vertisols on a local scale. This technology
could be a very useful tool for the estimation of the capacity of
such soils to act as sinks for excess atmospheric CO
2
, which
requires the analysis of a large number of samples and at
different levels of soil depth.
The vis-NIR models developed reliably predicted the total,
inorganic and organic C content in the Mediterranean rain-fed
vertisol studied. These models include several years of study,
different tillage systems and, above all, consider supercial
and deep soil horizons, a key aspect for the estimation of the
true capacity of a soil to sequester atmospheric CO
2
according
to different soil managements. Although vis-NIR models were
obtained with a homogeneous soil sample set, the precision
level indicates soil C measurement by vis-NIR could be an
optimal, rapid and cheap method to be used for C credits
purposes at plot scale (monitoring to compare soil C evolution
depending on time and variables as tillage, crop rotation and
fertilisation.). The potential for NIR Spectroscopy technology
being applied to other vertisols and on a wider scale must
be studied taking into account adequate strategies for the
calibration/validation procedure.
Acknowledgements
We thank the following for their excellent assistance in the
laboratory and field work: Joaqun Muoz and Jos Muoz.
Our thanks are also expressed to ABECERA for providing the
land and allowing us to use their eld facilities. Special thanks
are extended to INIA for assisting with nancial resources to
conduct this long-term eld experiment. This study was funded
by Spains Plan Nacional I+D Projects (AGF95-0553, AGF97-0498,
AGL2000-0460, AGL2003-03581 and AGL2006-02127/AGR).
References
1. K. Paustian, O. Andrn, H.H. Janzen, R. Lal, P. Smith,
G. Tian, H. Tiessen, M. Van Noordwijk and P.L. Woomer,
Agricultural soils as a sink to mitigate CO
2
emissions,
Soil Use Manage. 13, 230 (1997). doi: 10.1111/j.1475-
2743.1997.tb00594.x
2. D.F. Malley, P.D. Martin and E. Ben-Dor, Application
in analysis of soils, in Near-Infrared Spectroscopy
in Agriculture, Agronomy 44, Ed by C.A. Roberts, J.
Workman, Jr and J.B. Reeves, III. American Society
of Agronomy, Inc., Crop Science Society of America,
Inc., Soil Science Society of America, Inc. Publishers,
Madison WI, USA, p. 729, (2004).
3. S. Subak, Agricultural soil carbon accumulation in
North America: considerations for climate policy, Global
Environ. Change 10, 185 (2000). doi: 10.1016/S0959-
3780(00)00026-1
4. R.C. Izaurralde, W.B. McGill, A. Bryden, S. Gram, M.
Ward and P. Dickey, Scientic challenges in developing
a plan to predict and verify carbon storage in Canadian
prairie soils, in Management in Carbon Sequestration in
Soil, Advances in Soil Science, Ed by R. Lal, J.M. Kimble,
R.F. Follett and B.A. Stewart. CRC Press, Boca Raton,
Florida, USA, p. 433, (1997).
5. F.B. Metting, J.L. Smith and J.S. Amthor, Science needs
and new technology for soil carbon sequestration, in
Carbon Sequestration in Soils: Science, Monitoring and
Beyond: Proceedings of the St. Michaels Workshop, Ed by
N.J. Rosenberg, R. Izaurralde and E. Malone. Battelle
Press, Columbus, Ohio, USA, p. 1 (1998).
6. G.W. McCarty and J.B. Reeves, III, Development of
rapid instrumental methods for measuring soil organic
Total C
Inorganic
C
Organic C
Total C 1
Inorganic C 0.99 1
Organic C 0,035 0,078 1
Table 4. A correlation matrix (correlation coefcient values)
between total carbon, organic carbon and inorganic carbon
obtained from soil samples of the long-term experiment Malagn
(Cordoba, southern Spain).
262 Soil Carbon Determination in Mediterranean Vertisol
carbon, in Assessment Methods for Soil Carbon, Advances
in Soil Science, Ed by R. Lal, J.M. Kimble, R.F. Follett and
B.A. Stewart. CRC Press, Boca Raton, Florida, USA, p.
371 (2001).
7. L. Lpez-Bellido, Mediterranean cropping sytems, in
Ecosystems of the World. Field Crop Ecosystems, Ed by C.J.
Pearson. Elsevier, Amsterdam, The Netherlands, p. 311
(1992).
8. C.W. Chang, D.A. Laird, M.J. Mausbach and C.R.
Hurburgh, Jr, Near infrared reectance spectroscopy.
Principal components regression analyses of soil prop-
erties, Soil Sci. Soc. Am. J. 65, 480 (2001). doi: 10.2136/
sssaj2001.652480x
9. K.D. Shepherd and M.G. Walsh, Development of reect-
ance spectral libraries for characterization of soil prop-
erties, Soil Sci. Soc. Am. J. 66, 988 (2002). doi: 10.2136/
sssaj2002.9880
10. B. Stenberg, R.A.V. Rossel, A.M. Mouazen and J.
Wetterlind, Visible and near infrared spectroscopy
in soil science, in Advances in Agronomy, Volume 107,
Ed by Donald L. Sparks. Elsevier, Amsterdam, The
Netherlands, pp. 163215 (2010).
11. R.C. Dalal and R.J. Henry, Simultaneous determi-
nation of moisture, organic carbon and total nitro-
gen by near infrared reectance spectrophotom-
etry, Soil Sci. Soc. Am. J. 50, 120 (1986). doi: 10.2136/
sssaj1986.03615995005000010023x
12. J.B. III Reeves, G.W. McCarty and T. Mimmo, The poten-
tial of diffuse reectance spectroscopy for the determi-
nation of carbon inventories in soil, Environ. Pollut. 116,
277 (2002). doi: 10.1016/S0269-7491(01)00259-7
13. C.A. Russell, Sample preparation and prediction of soil
organic matter properties by near infra-red reectance
spectroscopy, Soil Sci. Plant Anal. 34, 1557 (2003). doi:
10.1081/CSS-120021297
14. B.G. Barths, D. Brunet, H. Ferrer, J.L. Chotte and
C. Feller, Determination of total carbon and nitrogen
content in a range of tropical soils using near infrared
spectroscopy: inuence of replication and sample grind-
ing and drying, J. Near Infrared Spectrosc. 14, 341 (2006).
doi: 10.1255/jnirs.686
15. B.E. Madari, J.B. Reeves, III, P.L.O.A. Machado, C.M.
Guimaraes, E. Torres and G.W. MacCarthy, Mid- and
near-infrared spectroscopic assessment of soil com-
positional parameters and structural indices in two
Ferralsols, Geoderma 136, 245 (2006). doi: 10.1016/j.
geoderma.2006.03.026
16. M. Confalonieri, F. Fornasier, A. Ursino, F. Boccardi, B.
Pintus and M. Odoardi, The potential of near infrared
reectance spectroscopy as a tool for the chemical
characterisation of agricultural soils, J. Near Infrared
Spectrosc. 9, 123 (2001). doi: 10.1255/jnirs.299
17. E. Ben-Dor and A. Banin, Near infrared reect-
ance analysis of carbonate concentration in
soils, Appl. Spectrosc. 44, 1064 (1990). doi:
10.1366/0003702904086821
18. A. Salg, J. Nagy, J. Tarnoy, P. Marth, O. Plmai and G.
Szab-Kele, Characterisation of soils by the near infra-
red technique, J. Near Infrared Spectrosc. 6, 199 (1998).
doi: 10.1255/jnirs.137
19. G.W. McCarty, J.B. Reeves, V.B. Reeves, R.F. Follett and
J.M. Kimble, Mid-infrared and near-infrared diffuse
reectance spectroscopy for soil carbon measure-
ment, Soil Sci. Soc. Am. J. 66, 640 (2002). doi: 10.2136/
sssaj2002.6400
20. D.J. Brown, K.D. Shepherd, M.G. Walsh, M.D. Mays
and T.G. Reinsch, Validation requirements for diffuse
reectance soil characterization models with a case
study of VNIR soil prediction in Montana, Geoderma 132,
273 (2006). doi: 10.1016/j.geoderma.2005.01.001
21. C.W. Chang and D.A. Laird, Near infrared reectance
spectroscopy analysis of soil C and N, Soil Sci. 167, 110
(2002).
22. J.B. Sankey, D.J. Brown, M.L. Bernard and R.L.
Lawrence, Comparing local vs global visible and near-
infrared (VisNIR) diffuse reectance spectroscopy (DRS)
calibrations for the prediction of soil clay, organic C and
inorganic C, Geoderma 148, 149 (2008). doi: 10.1016/j.
geoderma.2008.09.019
23. J. Wetterlind, B. Stenberg and M. Sderstrm, The use
of near infrared (NIR) spectroscopy to improve soil map-
ping at the farm scale, Precision Agriculture 9(1/2), 57
(2008). doi: 10.1007/s11119-007-9051-z
24. J. Wetterlind, B. Stenberg and M. Sderstrm,
Increased sample point density in farm soil mapping by
local calibration of visible and near infrared prediction
models, Geoderma 156(3/4), 152 (2010). doi: 10.1016/j.
geoderma.2010.02.012
25. J. Wetterlind and B. Stenberg, Near-infrared spectros-
copy for within-eld soil characterization: small local
calibrations compared with national libraries spiked
with local samples, Eur. J. Soil Sci. 61(6), 823 (2010). doi:
10.1111/j.1365-2389.2010.01283.x
26. C. Guerrero, R.A. Viscarra Rossel and A.M. Mouazen,
Diffuse reectance spectroscopy in soil science and
land resource assessment, Geoderma 158(1/2), 1 (2010).
doi: 10.1016/j.geoderma.2010.05.008
27. J.M. Baker, T.E. Ochsner, R.T. Venterea and T.J. Grifs,
Tillage and soil carbon sequestrationWhat do we
really know?, Agric. Ecosyst. Environ. 118, 1 (2006). doi:
10.1016/j.agee.2006.05.014
28. World Soil Resources, Report 84. Food and Agriculture
Organization, Rome, Italy (1998).
29. A.L. Page, D.E. Baker, D.R. Keeney, R.H. Miller,
J.D. Rhoades, and E. Jr. Roscoe, Methods of soils
analysis, in Chemical and Microbiological Properties,
Number 9, Part 2. American Society of Agronomy, Soil
Science Society of America, Madison, Wisconsin, USA
(1982).
30. H.L. Mark, and D. Tunnell, Qualitative near-infrared
reectance analysis using Mahalanobis distances, Anal.
Chem. 58, 379 (1985). doi: 10.1021/ac00284a061
J.M. Fontn et al., J. Near Infrared Spectrosc. 19, 253263 (2011) 263
31. J.S. Shenk and M.O. Westerhaus, Population denition,
sample selection and calibration procedures for near
infrared reectance spectroscopy, Crop Sci. 31, 469
(1991). doi: 10.2135/cropsci1991.0011183X003100020049x
32. T. Naes, T. Isaksson, T. Fearn and T. Davies, A User-
Friendly Guide to Multivariate Calibration and Classication.
NIR Publications, Chichester, UK. (2002).
33. R.J. Barnes, M.S. Dhanoa and S.J. Lister, Standard
normal variate transformation and de-trending of near-
infrared diffuse reectance spectra, Appl. Spectrosc. 43,
772 (1989). doi: 10.1366/0003702894202201
34. J.S. Shenk and M.O. Westerhaus, Population struc-
turing of near infrared spectra and modied partial
least square regression, Crop Sci. 31, 1548 (1991). doi:
10.2135/cropsci1991.0011183X003100060034x
35. C. Starr, A.G. Morgan and D.B. Smith, An evaluation
of near-infrared reectance analysis in some plant-
breeding programmes, J. Agric. Sci. 97, 107 (1981). doi:
10.1017/S0021859600035929
36. P.C. Williams and D.C. Soebering, Comparison of com-
mercial near infrared transmittance and reectance
instruments for analysis of whole grains and seeds, J.
Near Infrared Spectrosc. 1, 25 (1993). doi: 10.1255/jnirs.3
37. P.C. Williams, Implementation of near-infrared tech-
nology, in Near-Infrared Technology in the Agricultural
and Food Industries, Ed by P. Williams and K. Norris.
American Association of Cereal Chemists, St Paul,
Minnesota, USA, p. 145 (2001).
38. A.M. Mouazen, M.R. Maleki, J. De Baerdemaeker and
H. Ramon, On-line measurement of some selected soil
properties using a VIS-NIR sensor, Soil Tillage Res. 93,
13 (2007). doi: 10.1016/j.still.2006.03.009
39. J.B. Reeves, III, Does the spectral format matter in dif-
fuse reection spectroscopy?, Appl. Spec. 63(6), 669
(2009). doi: 10.1366/000370209788559692
40. M.J. Morra, M.H. Hall and L.L. Freeborn, Carbon and
nitrogen analysis of soils fractions using near-infrared
reectance spectroscopy, Soil Sci. Soc. Am. J. 55, 288
(1991). doi: 10.2136/sssaj1991.03615995005500010051x
41. F.J. Stevenson, Humus Chemistry. Genesis, Composition,
Reactions. John Wiley & Sons Inc., New York, USA (1994).
42. R.N. Clark, Remote Sensing for the Earth Sciences. Manual
of Remote Sensing. John Wiley & Sons Ltd, Chichester,
UK (1999).
43. B. Stenberg, Effects of soil sample pretreatments and
standardised rewetting as interacted with sand classes
on vis-NIR predictions of clay and soil organic carbon,
Geoderma 158(1/2), 15-22 (2010). doi: 10.1016/j.geo-
derma.2010.04.008
44. L. Lpez-Bellido, F.J. Lpez-Bellido, M. Fuentes, J.E.
Castillo and E.J. Fernndez, Inuence of tillage, crop
rotation and nitrogen fertilization on soil organic matter
and nitrogen under rainfed Mediterranean conditions,
Soil Tillage Res. 43, 277 (1997). doi: 10.1016/S0167-
1987(97)00018-4
45. B.W. Dunn, H.G. Beecher, G.D. Batten and S. Ciavarella,
The potential of near-infrared reectance spectroscopy
for soil analysisa case study from the Riverine Plain
of south-eastern Australia, Aust. J. Exp. Agric. 42, 607
(2002).
46. D.J. Brown, R.S. Bricklemyer and P.R. Miller, Validation
requirements for diffuse reectance soil characteriza-
tion models with a case study of VNIR soil prediction in
Montana, Geoderma 129, 251 (2005). doi: 10.1016/j.geo-
derma.2005.01.001
47. D. Cozzolino and A. Morn, Potential of near infrared
spectroscopy and chemometrics to predict soil organic
carbon fractions, Soil Tillage Res. 85, 78 (2002). doi:
10.1016/j.still.2004.12.006
48. J.B. Reeves, III, G.W. McCarty and J.J. Meisinger, Near
infrared reectance spectroscopy for the analysis of
agricultural soils, J. Near Infrared Spectrosc. 7, 179
(1999). doi: 10.1255/jnirs.248
49. P. Dardenne, Y. Reusen, D. Vanvyve, L. Bock, G. Colinet
and V. Genot, Near-infrared reectance spectroscopy
for estimating soil characteristics valuable in the diag-
nosis of soil fertility, J. Near Infrared Spectrosc. 19, 117
(2011). doi: 10.1255/jnirs.923
50. G. Fystro, The prediction of C and N content and their
potential mineralization in heterogeneous soil samples
using VISNIR spectroscopy and comparative methods,
Plant Soil 246, 139 (2002). doi: 10.1007/s11104-010-
0501-4
51. P.D. Martin, D.F. Malley, G. Manning and L. Fuller,
Determination of soil organic carbon and nitrogen at the
eld level using near-infrared spectroscopy, Can. J. Soil
Sci. 82, 413 (2002). doi: 10.4141/S01-054
JOURNAL
OF
NEAR
INFRARED
SPECTROSCOPY
265
ISSN: 0967-0335 IM Publications LLP 2011
doi: 10.1255/jnirs.931 All rights reserved
Near infrared (NIR) analysis is a rapid, non-destructive
and environmentally friendly technique widely used in the
pharmaceutical industry
13
for the determination of active
pharmaceutical ingredient(s),
46
the determination of water,
7,8

the identication of counterfeit drugs
9,10
and process analytical
technology.
1113
Moreover, the recently developed universal
NIR models can effectively eliminate interferences from
excipients or other types of process disturbances and have
been employed to analyse pharmaceutical products with the
same International Non-proprietary Names from different
manufacturers.
1417
Various quantitative methods have also
been developed for the determination of low dosages. For
example, Chalus et al.
18
reported quantitative methods for the
determination of low doses of bromazepam and clonazepam
using near infrared diffuse reectance spectroscopy; Meza et
al.
19
reported the use of a transmission near infrared spectro-
scopic method to determine the drug content in tablet formu-
lations containing less than 1% w/w active ingredient; Alcal et
al.
20
used Fourier transform near infrared (FT-NIR) transmis-
sion spectroscopy to detect an active ingredient (05% w/w) in
pharmaceutical tablets; Zimonsa et al.
21
developed a robust
NIR calibration method to determine the acetaminophen
Quantitative calibration models for the
determination of azithromycin and
decladinosylazithromycin in azithromycin
injection powder by using diffuse
reflectance near infrared spectroscopy
Ji-Xiong Dong,
a,b
Wen-Bo Zhou,
a
Yan-Chun Feng,
a
Dan-Qing Song
b
and Chang-Qin Hu
a,*
a
National Institutes for Food and Drug Control, Beijing 100050, PR China. E-mail: hucq@nicpbp.org.cn
b
Institute of Medicinal Biotechnology, Chinese Academy of Medical Science and Peking Union Medical College, Beijing 100050, PR China
A quantitative analysis method based on near infrared (NIR) spectroscopy was developed for determining the azithromycin content and
its degradation product impurity (decladinosylazithromycin) in azithromycin for injection. Two quantitative models were obtained by
evaluating 18 laboratory samples and 44 commercial samples formulated as hydrochlorate, phosphate, lactate and citrate salts. The
samples had an azithromcyin content ranging from 74.2% to 91.0% (w/w) and decladinosylazithromycin ranging from 0.1% to 7.7% (w/w).
The root mean square error of cross-validation for the azithromycin model and decladinosylazithromycin model were 1.54% and 0.49%,
respectively. The estimated limit of detection and limit of quantication for the decladinosylazithromycin calibration were about 1.2%
and 4.1% (w/w), respectively. In addition, the quantitative models were validated in terms of specicity, linearity, accuracy, repeatability
and intermediate precision, according to the International Conference on Harmonisation guidelines, by evaluating the characteristics of
the NIR spectra. The results showed that the models were satisfactory. Our results demonstrated that a universal quantitative model
based on NIR diffuse reectance spectroscopy could be used as a quick and reliable method for quality control of low-dose impurities
in pharmaceutical drugs.
Keywords: NIR, quantitative model, azithromycin, decladinosylazithromycin, low dosage, impurity
Introduction
J.-X. Dong et al., J. Near Infrared Spectrosc. 19, 265275 (2011)
Received: 23 November 2010 Revised: 5 April 2011 Accepted: 27 July 2011 Publication: 16 August 2011
266 Determination of Azithromycin and Decladinosylazithromycin in Azithromycin Injection Powder
content in a low-dose syrup formulation (2%, w/v) and tested
the method for monitoring real-time changes in the concen-
tration of the active pharmaceutical ingredient; and Blanco et
al.
22
developed a partial least squares (PLS) calibration method
for analysis of 2-ethylhexanol (201600 ppm) in an industrial
ester. However, reports on the use of diffuse reectance NIR
spectroscopy for quantitative analysis of low concentrations of
impurities in pharmaceutical products are scant.
Azithromycin is a macrolide antibiotic usually formulated
as tablets, oral suspensions and powders for intravenous
injection. For intravenous injection, azithromycin is typically
reacted with an acid to give the corresponding hydrochlo-
rate, phosphate, lactate or citrate salt and then freeze dried
with excipients. The main impurity present is decladinosyl-
azithromycin, which is the degradation product formed from
salt-forming reactions and storage. The degradation reaction
of azithromycin in aqueous solution follows rst-order kinetics
with a T
1/10
of 20.1 min (pH 2 at 37C) and the degradation will
increase at higher temperatures.
23
Decladinosylazithromycin
has a similar molecular structure to azithromycin (Figure
1). The content of decladinosylazithromycin in intravenous
azithromycin formulations sold on the Chinese market was
found to range from 0.1% to 8% (w/w) depending on the manu-
facturing process.
24
In this study, quantitative calibration methods were devel-
oped to determine the concentrations of decaldionosylazithro-
mycin and azithromycin in different salts of azithromycin for
injection.
Materials and methods
Fourier transform near infrared equipment
NIR spectra were obtained on a MATRIX-F2 series FT-NIR
spectrometer (Bruker Optik GmbH, Ettlingen, Germany)
equipped with a 1.5 m bre-optic diffuse reectance probe
(1 cm i.d. and 3 mm spot size) and an extended thermo-
electrically-cooled indium gallium arsenide detector. Bruker
OPUS software version 6.5 was used for all data collections
and analysis.
Samples
Forty-four commercial samples of azithromycin for injec-
tion were collected from 15 different manufacturers from
the Chinese market between 2008 and 2009 by the Chinese
National Institute for Food and Drug Control (NIFDC), which
included the hydrochlorate, phosphate, lactate and citrate salts
of azithromycin. As seen in Table 1, the content of decladinosyl-
azithromycin among the samples was not consistent and the
range at 1.54.5% (w/w) was lacking. To provide a better range
of decladinosylazithromycin concentration for analysis, 18
laboratory samples were prepared by mixing either the refer-
ence sample of azithromycin or decladinosylazithromycin
(both were provided by NIFDC) with azithromycin injection
powder in different proportions.
The calibration set consisted of 28 commercial samples from
15 different manufacturers and eight laboratory samples; the
concentrations of azithromycin and decladinosyl azithromycin
ranged from 74.2% to 91.0% and 0.1% to 7.7% (w/w), respec-
tively. The test set consisted of 16 commercial samples from
eight different manufacturers and ten laboratory samples;
the concentrations of azithromycin and decladinosylazithro-
mycin ranged from 74.8% to 90.3% and 0.2% to 7.4% (w/w),
respectively. Ten commercial samples in the test set were
from four new manufacturers and were not included in the
calibration set. Detailed information of the samples is shown
in Table 1.
Acquisition of near infrared spectra
Powder samples more than 5 mm thick were put into trans-
parent glass vials and analysed directly using a bre-optic
probe from the bottom of the glass vial at 8 cm
1
resolution with
32 co-added scans over the spectral range (400012,000 cm
1
).
In order to reduce measurement errors, five spectra from
each batch were recorded and then averaged. The average
spectrum from each batch was used to develop the models.
Figure 1. Chemical structures of azithromycin and decladinosylazithromycin.
J.-X. Dong et al., J. Near Infrared Spectrosc. 19, 265275 (2011) 267
Reference method
The reference method used for determination of the contents
of azithromycin and decladinosylazithromycin was the HPLC
method recommended by the Chinese Pharmacopoeia (2010) for
azithromycin.
25
A Waters HPLC device that consisted of a 2690
separation module and a PDA996 DAD detector was used (Waters,
Mil,ford, MA, USA). Each sample was dissolved to 1 mg mL
1

by the mobile phase. The 50 L solution sample was injected
onto a Shiseido MGIIUC18 column (250 mm 4.6 mm.i.d., 5 m,
particle size; Shiseido Co. Ltd, Tokyo, Japan) with a ow rate of
1.0 mL min
1
. Two determinations were made for each sample
and the average result was used as the reference value.
Modelling
In this study, two calibration models were developed with a
calibration set composed of 36 sample batches using the
PLS-1 algorithm (PLS regression for one y-variable)
26
avail-
able in the Quant 2 package of Bruker OPUS software. In
order to avoid the over-tting, a random leave-four-out cross-
validation method was applied in this work. Rank is the number
of factors determined by performing an F-test ( = 0.05) on the
prediction error sum of squares (PRESS) as shown in Equation
(1). This F-test consists of searching the model with the lowest
number of factors to yield a PRESS which is not signicantly
different from the lowest one within the rst 20 factors. The
coefcient of determination (R
2
) is given by Equation (2), which
reects the percentage of reference variance with respect to
total variance in the reference values:
( )
2
1
ref pred
(1)
, ,
n
i
i i
PRESS C C
=
=
( )
( )
2
1
ref pred
2
2
1
ref
1 100 (2)
, ,
,
n
i
i i
n
i
i
C C
R
C C
=
=



=



where n is the number of samples in the calibration set, C
ref
is the reference content, C
pred
is the content predicted by NIR,
and C

is the mean of reference content.

The root mean square errors of cross validation (RMSECV)
[Equation (3)] was used to describe the results of cross-
validation, where n
c
is the number of samples in the calibra-
tion set:
( )
c
2
1
ref pred
c
(3)
, ,
n
i
i i
C C
RMSECV
n
=

=
Method validation
The validation processes were performed according to the
guidelines of the International Conference on Harmonisation
(ICH).
27
In this work, besides the evaluation of specicity, line-
arity, accuracy, repeatability and intermediate precision, the
limit of determination (LOD) and limit of quantication (LOQ)
for decladinosylazithromycin calibration were also estimated.
Specicity is the ability to assess unequivocally the analyte in
the presence of components which are expected to be present.
The Mahalanobis distance limit or threshold [Equation (4)] in
the Quant 2 package of the OPUS software was used in the
specicity test:
factor rank
threshold (4)

=
M
where rank is the number of partial least squares, M is the
number of modelling samples, and factor is a coefcient
that can range from 2 to 10, but was set to 2 in this study.
The linearity was tested by evaluating the slope, the inter-
cept and the correlation coefcient of the plot of NIR predic-
tion content in comparison with the reference content values.
Salt Manufacturers Samples
Content range of
azithromycin
Content range of
decladinosylazithromycin
(w/w) (w/w)
Hydrochloride 2 11 81.687.1% 4.77.7%
Phosphate 6 11 75.485.2% 0.41.4%
Lactate 1 11 74.277.6% 0.71.2%
Citrate 6 11 76.282.5% 0.11.3%
Laboratory
samples
18 87.591.0% 1.74.5%
Calibration set Test set
Commercial samples 28 10
Laboratory samples 8 16
Manufacturers 15 8
*
Content range of azithromycin (w/w) 74.291.0% 74.890.3%
Content range of decladinosylazithromycin (w/w) 0.17.7% 0.27.4%
*
Ten samples were from four new manufacturers not included in the calibration set.
Table 1. Information regarding the samples.
268 Determination of Azithromycin and Decladinosylazithromycin in Azithromycin Injection Powder
In the accuracy test, a test set comprised of 26 batches was
used to evaluate the performance of the two NIR calibration
models. The root mean standard error of prediction (RMSEP)
[Equation (5)] was used to describe the differences between
the predicted content and the reference content:
( )
2
1
ref pred
v
(5)
, ,
n
i
i i
C C
RMSEP
n
=

=
The number of samples in the calibration set is represented by
n
v
, while C
ref
is the reference content and C
pred
is the content
predicted by NIR.
There are different approaches for estimating the LOD and
LOQ in multivariate calibration. One of them is based on the
idea that the standard deviation of the noise signal is similar to
the standard deviation of the signal of a sample with a very low
content of the parameter of interest.
28
Therefore, in this case,
LOD and LOQ were approximated by LOD = 3 and LOQ = 10,
where is the standard deviation, estimated by 10 times predic-
tion for a low content decladinoaziyhromycin sample (1.0%).
Results and discussion
Selection of calibration samples and
validation samples
With the goal of developing universal calibration models in
mind, different variables generated from different manufac-
turers and formulations were considered. These variables were
reected on the NIR spectra of the samples, demonstrating
that it is important to choose representative spectra as the
calibration set, which can be achieved by using cluster analysis
method. In this study, Wards algorithm in the cluster analysis
package of the Bruker OPUS software was used to obtain the
sample variance of the entire spectra from 62 samples. Based
on the hierarchical cluster analysis results, the 62 samples
were divided into 36 groups from the first done (Figure 2).
One representative sample was chosen from each group to
constitute the calibration set and the remaining samples in
each group were treated as the test set. Principal components
analysis (PCA) was used to verify whether the calibration and
test samples, selections were reasonable. The spectral range
from 10,000 cm
1
to 4250 cm
1
and ve factors were adopted in
PCA without using any spectral pre-processing method. The
PCA distribution results for the calibration and test sets by the
rst two principal components are shown in Figure 3.
Selection of wavelength range
The near infrared spectra of azithromycin and decladinosylazi-
thromycin reference material were processed using vector
normalisation and the rst derivative; the overlapping spec-
trum (75004000 cm
1
) is shown in Figure 4. Signicant differ-
ences between azithromycin and decladinosylazithromycin
from the rst-derivative spectra were found in the A and C
regions at 72006200 cm
1
and 55004500 cm
1
, respectively,
while the peaks in the B region might be the rst overtone
absorption of the CH
3
, CH
2
and CH groups. Our previous study
indicated that NIR models can lead to better predictions when
established by combining the structural relevant spectrum
bands with content relevant spectrum bands.
29,30
In this study,
the A and C regions were considered as the structural relevant
bands belonging to azithromycin and decladinosyl azithromycin,
whereas the B region corresponded to the content relevant
bands of both compounds. After optimisation, absorptions
at 63006050 cm
1
, 59255650 cm
1
, 55505460 cm
1
and
53005050 cm
1
were applied to the modelling of azithromycin,
and the absorptions at 72007050 cm
1
, 64506300 cm
1
and
60005700 cm
1
were applied to the modelling of decladinosyl-
azithromycin.
Value of rank
The value of rank used for the nal models was 7. For azithro-
mycin for injection, there were at least four types of salt
(hydrochlorate, phosphate, lactate and citrate) on the Chinese
market. Due to the complexity of the spectral composition,
which at least has four kinds of spectra corresponding to four
kinds of salt, and laboratory samples in the calibration set, 7
was set as the rank value for these models to be acceptable
although the rank values for NIR models is usually about
35. Our previous study on moisture content determination of
-lactam powder injections by NIR could reveal the relation-
ship between rank and the spectral complexity.
31
The major
results are shown in the Table 2 and it can be seen that the
rank values were gradually increased with the increasing of
composition and type of calibration spectra.
Calibration models
The leave-four-out cross-validation was employed in the
modelling experiment, which yielded two calibration models
for azithromycin and decladinosylazithromycin. With respect
to spectral preprocess, the combination of second derivative,
rst derivative and vector normalisation was optimised and
employed in model construction. Vector normalisation is dened
in the Quant 2 package of OPUS software as follows: vector
normalisation rst calculates the average y-value of spectra
and only uses data points within the selected spectral ranges.
The average value calculated will then be subtracted from the
spectrum, which causes the spectrum to be centred at around
y = 0. This is followed by calculating the sum of squares of all y
values and the respective spectrum is divided by the square root
of this sum. The vector norm of the resulting spectrum is 1:
( )
m
(6)

=
k
a k
a
N
( ) ( )
m
(7) = a k a k a
( )
( )
( )
2
(8)
k
a k
a k
a k

where a
m
is the mean spectral intensity in the frequency range,
while a(k) is the original spectral intensity, k is all data points of
J.-X. Dong et al., J. Near Infrared Spectrosc. 19, 265275 (2011) 269
the frequency range, N is the number of the original spectra and
a(k) is the spectral intensity pretreated by vector normalisation.
The smoothing point for both models was 13. The values of R
2
and RMSECV in the model for azithmycin were 0.91 and 1.54%,
respectively. The values of R
2
and RMSECV in the model for
decladinosylazithmycin were 0.94 and 0.49%, respectively. The
value of rank used for the nal models was 7. In this case, the
reason for the slightly high value of 1.40% for RMSECV might
Figure 2. The hierarchical cluster analysis result for all samples: (a) the overall plot and (b) the magnied plot. The les were labelled
with the letters C, V, Hyd, Lac, Pho, Cit and LaS for the calibration sample, test sample, hydrochlorate, lactate, phosphate, citrate and
laboratory samples, respectively. The letters were preceded by a three-digit alpha numeric code of a specic sample. For example, the
spectrum of hydrochlorate calibration sample No.01 is labelled as C01-Hyd.
270 Determination of Azithromycin and Decladinosylazithromycin in Azithromycin Injection Powder
be the error of the reference method (about 1.0%). Therefore,
a value of 1.54% as the error of NIR prediction is acceptable as
a secondary method. The detailed parameters of the models
are shown in Table 3. The NIR predicted concentrations were
plotted against reference contents for the calibration and test
samples [Figures 5(a) and (b)]. Figure 5(c) shows the plot of
RMSECV against rank for both models, where the rank values
are circled.
Method validation results
Specicity
Five batches of azithromycin powder samples with four formu-
lations and nine other macrolide samples (supplied by NIFDC)
with chemical structures very similar to that of azithromycin
and decladinoazithromycin were selected to test for specicity.
The Mahalanobis distance

is used to detect the outliers and
can be calculated in the relevant sub-space spanned by the
calibration PLS vectors. Based on the calibration models, the
thresholds were set at 0.39. When comparing the Mahalanobis
distance of an unknown sample with the threshold, if the
Mahalanobis distance of the unknown macrolide sample was
greater than the threshold of the corresponding model, or the
Mahalanobis distance of the unknown azithromycin sample
was lower than the threshold, the unknown sample was iden-
tied correctly. As shown in Table 4, all identications were
correct. These results illustrate that both the azithromycin
model and decladinosylazithromycin model have a satisfac-
tory specicity.
Linearity
Unlike a typical univariate calibration method, the linearity in
NIR validations is evaluated by examining the slope and inter-
cept of the plot of NIR predicted values with reference values,
as shown in Figures 5(a) and( b). The correlation coefcient, r
2
,
and the 95% condence interval for the slope and intercept are
summarised in Table 5. The condence interval for slope and
intercept of azithromycin calibration data did not include 1 and
zero, respectively. Therefore, the calibration curve is nonlinear.
Similar nonlinear results were also observed in a study of the
parameters of edible oils by NIR spectrometry.
28
Figure 3. Distribution of the samples in calibration set and test
set by rst two principal components in principal component
analysis. Grey blocks indicate the samples of calibration set
and grey triangles indicate the samples of test set.
Figure 4. The rst-derivative spectra of azithromycin and
decladinosylazithromycin.
Complexity
of spectral
composition
Number of
calibration
spectra
Type of
calibration
spectra
Spectral
range
Preprocessing R
2
RMSECV Rank
A 61 1 7502.34198.9 FD+VN 0.994 0.181 5
B 100 4 8917.96217.8 FD+VN 0.990 0.275 9
C 148 6 100005002.3 FD+VN 0.990 0.285 12
D 179 7 7502.34196.6 FD+VN 0.990 0.247 16
A: Cefazolin sodium.
B: Cefazolin sodium, ceftaroxime sodium, ceftriaxone sodium, cefotaxime sodium.
C: Cefazolin sodium, ceftaroxime sodium, ceftriaxone sodium, cefotaxime sodium, cefoperazone sodium, cefepime hrdrochloride.
D: Cefazolin sodium, ceftaroxime sodium, ceftriaxone sodium, cefotaxime sodium, cefoperazone sodium, cefepime hrdrochloride, ceftizoxime sodium.
FD+VN: rst derivative + vector normalisation.
Table 2. The major results reported by Zhang et al.
31
J.-X. Dong et al., J. Near Infrared Spectrosc. 19, 265275 (2011) 271
Accuracy
A test set with 26 sample batches was used to evaluate the
performance of the NIR calibration models. The accuracy
of the method is described by RMSEP (Table 3). The NIR
predicted values of the two models are shown in Table 6.
A paired t-test was also performed to determine whether
the NIR values and the reference values were signicantly
different. The t-test results showed that the P values for
Parameter Azithromycin Decladinosylazithromycin
Content range (w/w) 74.291.0% 0.17.7%
Wavelength range (cm
1
) 63006050, 59255650 72007050, 64506300
55505460, 53005050 60005700
Spectral preprocess SD FD+VN
Smoothing point 13 13
R
2
0.91 0.94
Rank 7 7
RMSECV (%) 1.54 0.492
RMSEP (%) 1.65 0.606
SD: second derivative; FD+VN: rst derivative + vector normalisation.
Table 3. Parameters of two nal calibration models.
Figure 5. The results for the near infrared models: (a) prediction results of azithromycin versus the reference results; (b) prediction
results of decladinosylazithromycin versus the reference results; and (c) RMSECV versus ranks of both models.
272 Determination of Azithromycin and Decladinosylazithromycin in Azithromycin Injection Powder
both models were greater than 0.05, which indicated that
the NIR predicted values were not signicantly different from
the reference values (Table 5).
Repeatability
As dened in the ICH guidelines, repeatability expresses the
precision under the same operating conditions over a short
interval of time. In this study, a single batch of azithromycin
was measured six times by the same operator on the same
day and quantied for repeatability test. The statistical results
are shown in Table 5.
Intermediate precision
This level of precision is useful for studying the inuence of
random events during the analysis. For the precision study,
two random events were considered: the analysis of a single
sample on three different days and two analysts performing
the analysis on the same day. The variances due to different
Samples Formulation Mahalanobis distance Outlier
Azithromycin
(threshold = 0.39)
Decladinosylazithromycin
(threshold = 0.39)
*
Azithromycin powder Citrate 0.15 0.086 No
*
Azithromycin powder Hydrochloride 0.17 0.33 No
*
Azithromycin powder Lactate 0.072 0.11 No
*
Azithromycin powder Phosphate 0.24 0.35 No
*
Azithromycin powder Laboratory
sample
0.18 0.096 No
Acetylkitasamycin powder NA 3 3.6 Yes
Kitasamycin powder NA 6.6 76 Yes
Midecamycin powder NA 27 69 Yes
Roxithromycin powder NA 11 22 Yes
Meleumycin powder NA 6.2 84 Yes
Erythromycin powder NA 48 24 Yes
Clarithromycin powder NA 23 9.3 Yes
Acetylspiramycin powder NA 13 70 Yes
Erythromycin ethylsuccinate powder NA 57 97 Yes
*
Powder including azithromycin and decladinosylazithromycin
Table 4. Specicity for two nal calibration models.
Test Description Azithromycin Decladinosylazithromycin
Linearity NIR = a REF + b Calibration samples: Calibration samples:
a = 0.905 0.045 a = 0.902 0.034
b = 7.732 3.687 b = 0.190 0.098
r
2
= 0.960 r
2
= 0.977
Accuracy Paired t-test of NIR values and reference
values
Test samples: Test samples:
Avg. diff. = 0.565 Avg. diff. = 0.019
S
d
= 1.425 S
d
= 0.6713
t = 2.022 t = 0.146
P = 0.054 ( = 0.05) P = 0.885 ( = 0.05)
Repeatability One sample analysed six times by the
same operator on the same day
Avg. = 89.100 Avg. = 6.300
SD = 0.783 SD = 0.103
RSD = 0.879 RSD = 1.634
r
2
: correlation coefcient; Avg. diff: mean of difference; S
d
: standard deviation of difference. SD: standard deviation; RSD: relative standard deviation.
Table 5. Validation results of the methods.
J.-X. Dong et al., J. Near Infrared Spectrosc. 19, 265275 (2011) 273
operators and sampling day were determined jointly by anal-
ysis of variance. Based on the results listed in Table 7, these
factors had no signicant inuence on each model.
Limit of detection and limit of quantication
In this case the estimated LOD and LOQ for decladinosyl-
azithromycin calibration were about 1.2% and 4.1% (w/w),
Sample
Azithromycin Decladinosylazithromycin
Reference
(HPLC, mg mg
1
)
Prediction
(NIR, mg mg
1
)
Reference
(HPLC, mg mg
1
)
Prediction
(NIR, mg mg
1
)
T01 84.3 85.5 7.4 6.7
T02 85.9 84.4 5.9 6.8
T03 82.5 85.3 6.3 6.1
T04 87.1 84.4 7.3 6.8
T05 83.6 84.7 5.3 5.9
T06 81.8 82.1 5.3 6.2
T07 74.8 76.4 1.1 0.4
T08 76.8 76.6 0.8 0.3
T09 77.6 77.8 0.5 0.8
T10 83.0 81.0 1.1 0.9
T11 80.1 80.0 0.8 1.2
T12 77.8 77.3 0.7 1.2
T13 82.5 81.6 0.6 0.9
T14 80.9 80.4 0.8 0.4
T15 90.3 88.3 1.7 2.4
T16 90.2 87.4 1.8 2.1
T17 89.8 88.7 2.2 2.4
T18 77.1 77.6 0.8 0.6
T19 89.5 87.9 2.5 3.2
T20 89.3 89.0 2.7 3.2
T21 77.3 76.9 0.2 0.9
T22 88.8 88.2 3.3 2.9
T23 88.6 86.5 3.4 1.9
T24 88.2 85.7 3.8 4.7
T25 87.7 88.5 4.3 3.3
T26 87.5 86.1 4.5 3.6
Table 6. Accuracy of the two calibration models was determined by comparing the NIR predicted values with HPLC reference values of the
test set.
Azithromycin Operator 1 Operator 2 Decladinosylazithromycin Operator 1 Operator 2
(88.76%, w/w) (6.29%, w/w)
Day 1 89.5 89.2 Day 1 5.8 5.7
Day 2 88.8 88.1 Day 2 5.9 5.9
Day 3 89.0 89.3 Day 3 5.9 5.7
Statistics F-value P-value Statistics F-value P-value
Rows 3.7346 0.2112 Rows 2.1168 0.3208
Columns 0.4747 0.5620 Columns 1.5478 0.3395
Table 7. Results from intermediate precision checks.
274 Determination of Azithromycin and Decladinosylazithromycin in Azithromycin Injection Powder
respectively. As most of the content of decladinosylazithromycin
is more than 2% in azithromycin for injection on the Chinese
market,
24
the model can be used as a useful tool for quick
screening, although some of the results are not satisfactory.
Conclusion
Two quantitative calibration models based on diffuse reect-
ance NIR spectroscopy were developed for determining the
concentration of azithromycin and decladinosylazithromycin in
azithromycin for injection. Our results demonstrated the feasi-
bility of generating calibration spectra for simultaneous quan-
tication of the active pharmaceutical ingredient and residual
impurities in a drug. Moreover, this NIR method allows for
reliable and rapid analysis of various intravenous azithro-
mycin formulations from different manufacturers and might
be applicable to other drugs or formulations.
References
1. Y. Roggo, P. Chalus, L. Maurer, C.L. Martinez, A. Edmond
and N. Jent, A review of near infrared spectroscopy
and chemometrics in pharmaceutical technologies,
J. Pharm. Biomed. Anal. 44, 683 (2007). doi: 10.1016/j.
jpba.2007.03.023
2. J. Luypaert, D.L. Massart and Y.V. Heyden, Near-
infrared spectroscopy applications in pharmaceutical
analysis, Talanta 72, 865 (2007). doi: 10.1016/j.
talanta.2006.12.023
3. A.A. Gowen, C.P. ODonnell, P.J. Cullen and S.E. Bell,
Recent applications of chemical imaging to pharma-
ceutical process monitoring and quality control, Eur.
J. Pharm. Biopharm. 69, 10 (2008). doi: org/10.1016/j.
ejpb.2007.10.013
4. J. Mantanus, E. Zimons, P. Lebrun, E. Rozet, R.
Klinkenberg, B. Streel, B. Evrard and P. Hubert, Active
content determination of non-coated pharmaceutical
pellets by near infrared spectroscopy: method develop-
ment, validation and reliability evaluation, Talanta 80,
1750 (2010). doi: org/10.1016/j.talanta.2009.10.019
5. M. Haag, M. Brning and K. Molt, Quantitative analysis
of diphenhydramine hydrochloride in pharmaceutical
wafers using near infrared and Raman spectroscopy,
Anal. Bioanal. Chem. 395, 1777 (2009). doi: org/10.1007/
s00216-009-3130-0
6. D. Xiang, M. Konigsberger, B. Wabuyele, K. Hornung and
J. Cheney, Development of robust quantitative methods
by near-infrared spectroscopy for rapid pharmaceuti-
cal determination of content uniformity in complex tab-
let matrix, Analyst 134, 1405 (2009). doi: org/10.1039/
b821836f
7. H. Grohganz, D. Gildemyn, E. Skibsted, J.M. Flink and
J. Rantanen, Towards a robust water content deter-
mination of freeze-dried samples by near-infrared
spectroscopy, Anal. Chim. Acta 676, 34 (2010). doi:
org/10.1016/j.aca.2010.07.035
8. J. Mantanus, E. Zimons, P. Lebrun, E. Rozet, R.
Klinkenberg, B. Streel, B. Evrard and Ph. Hubert,
Moisture content determination of pharmaceutical pel-
lets by near infrared spectroscopy: method development,
validation and reliability evaluation, Anal. Chim. Acta
642, 186 (2009). doi: org/10.1016/j.aca.2008.12.031
9. T. Puchert, D. Lochmann, J.C. Menezes and G. Reich,
Near-infrared chemical imaging (NIR-CI) for counterfeit
drug identicationa four-stage concept with a novel
approach of data processing (Linear Image Signature),
J. Pharm. Biomed. Anal. 51, 138 (2010). doi: org/10.1016/j.
jpba.2009.08.022
10. M.B. Lopes, J.C. Wolff, J.M. Bioucas-Dias and M.A.
Figueiredo, Determination of the composition of coun-
terfeit Heptodin tablets by near infrared chemical imag-
ing and classical least squares estimation, Anal. Chim.
Acta 641, 46 (2009). doi: org/10.1016/j.aca.2009.03.034
11. Y. Sulub, R. LoBrutto, R. Vivilecchia and B.W. Wabuyele,
Content uniformity determination of pharmaceutical tab-
lets using ve near-infrared reectance spectrometers:
a process analytical technology (PAT) approach using
robust multivariate calibration transfer algorithms,
Anal. Chim. Acta 611, 143 (2008). doi: org/10.1016/j.
aca.2008.02.016
12. T.R.M. De Beer, P. Vercruysse, A. Burggraeve, T. Quinten,
J. Ouyang, X. Zhang, C. Vervaet, J.P. Remon and W.R.G.
Baeyens, In-line and real-time process monitoring of
a freeze drying process using Raman and NIR spec-
troscopy as complementary process analytical tech-
nology (PAT) tools, J. Pharm. Sci. 98, 3430 (2009). doi:
org/10.1002/jps.21633
13. J. Mrk, M. Andre, M. Karner and C.W. Huck, Prospects
for multivariate classication of a pharmaceutical inter-
mediate with near-infrared spectroscopy as a process
analytical technology (PAT) production control supple-
ment, Eur. J. Pharm. Biopharm. (2010). doi: 10.1016/j.
ejpb.2010.06.015
14. H.H. Pang, Y.C. Feng, X.B. Zhang and C.Q. Hu,
Construction of universal quantitative models for the
determination of cefoperazone sodium/sulbactam
sodium for injection from different manufacturers using
near-infrared reectance spectroscopy, Chin. Pharm. J.
17, 22 (2008).
15. Y.C. Feng and C.Q. Hu, Construction of universal quan-
titative models for determination of roxithromycin and
erythromycin ethylsuccinate in tablets from different
manufacturers using near infrared reectance spec-
troscopy, J. Pharm. Biomed. Anal. 41, 373 (2006). doi:
org/10.1016/j.jpba.2005.11.027
16. X.P. Liu, Y.C. Feng, C.Q. Hu and L. Ding, Construction of
universal quantitative models for determination of cefra-
dine capsules, Chin. J. Pharm. Anal. 28, 722 (2008).
17. X.L. Wang, Q. Fua, J.F. Sheng, X. Yang, J.Z. Jia and W.
Du, Construction of a universal quantitative model
J.-X. Dong et al., J. Near Infrared Spectrosc. 19, 265275 (2011) 275
for ibuprofen sustained-release capsules from differ-
ent manufacturers using near-infrared diffuse reec-
tion spectroscopy, Vib. Spectrosc. 53, 214 (2010). doi:
org/10.1016/j.vibspec.2010.03.002
18. P. Chalus, Y. Roggo, S. Walte and M. Ulmschneider,
Near-infrared determination of active substance content
in intact low-dosage tablets, Talanta 66, 1294 (2005). doi:
org/10.1016/j.talanta.2005.01.051
19. C.P. Meza, M.A. Santos and R.J. Romaach,
Quantitation of drug content in a low dosage formula-
tion by transmission near infrared spectroscopy, AAPS
PharmSciTech. 7, E206 (2006). doi: org/10.1208/pt070129
20. M. Alcal, J. Len, J. Ropero, M. Blanco and R.J.
Romaach, Analysis of low content drug tablets by
transmission near infrared spectroscopy: selection of
calibration ranges according to multivariate detection
and quantitation limits of PLS models, J. Pharm. Sci. 97,
5318 (2008). doi: org/10.1002/jps.21373
21. E. Zimons, J. Mantanus, P. Lebrun, E. Rozet, B. Evrard
and Ph. Hubert, Acetaminophen determination in low-
dose pharmaceutical syrup by NIR spectroscopy, J.
Pharm. Biomed. Anal. 53, 510 (2010).
22. M. Blanco, M. Castillo, A. Peinado and R. Beneyto,
Determination of low analyte concentrations by near-
infrared spectroscopy: effect of spectral pre treatments
and estimation of multivariate detection limits,
Anal. Chim. Acta 581, 318 (2007). doi: org/10.1016/j.
aca.2006.08.018
23. E.F. Fiese and S.H. Steffen, Comparison of the acid sta-
bility of azithromycin and erythromycin A, J. Antimicrob.
Chemother. 25, 39 (1990).
24. D.S. Zhang, C. Wang, W. Liu, L.X. Wang, M.J. Wang, L.
Zhang, Y.L. Yang and C.Q. Hu, Sampling results and
assessment the quality of domestic azithromycin for
injection in 2009, Chin. J. Pharm. Anal. 30, 1804 (2010).
25. The State Pharmacopoeia Commission of The Peoples
Republic of China, Pharmacopoeia of the Peoples Republic
of China. Chemical Industry Press, Beijing, PR China
(2005).
26. D.A. Burns and E.W. Ciurczak, Handbook of Near-Infrared
Analysis, 3rd Edn. CRC Press, Boca Raton, Florida, USA,
p. 198 (2008)
27. A.C. Moffat, A.D. Trafford, R.D. Jee and P.
Graham, Meeting the International Conference on
Harmonisations Guidelines on Validation of Analytical
Procedures: Quantication as exemplied by a near-
infrared reectance assay of paracetamol in intact tab-
lets, Analyst 125, 1341 (2000). doi: org/10.1039/b002672g
28. S. Armenta, S. Garrigues and M. de la Guardia,
Determination of edible oil parameters by near infrared
spectrometry, Anal. Chim. Acta 596, 330 (2007). doi:
org/10.1016/j.aca.2007.06.028
29. Y.Y. Liu, C.Q. Hu and T.J. Hang, Discussion on spectral
wavelength selection principles and the strategy for
eliminating the solvent interference in the near infrared
quantitative models for determination of Acanthopanax
senticosus injection, Sci. Sinica Chim. 40, 1164 (2010).
30. Y.Y. Liu and C.Q. Hu, Construction of predictive models
for gentamicin contents in aqueous solution using spe-
cic near infrared spectral regions, Vib. Spectrosc. 55,
241 (2011). doi: 10.1016/j.vibspec.2010.12.002
31. X.B. Zhang, Y.C. Feng and C.Q. Hu, Feasibility and
extension of universal quantitative models for moisture
content determination in beta-lactam powder injections
by near-infrared spectroscopy, Anal. Chim. Acta 630, 131
(2008). doi: org/10.1016/j.aca.2008.09.050
JOURNAL
OF
NEAR
INFRARED
SPECTROSCOPY
277
ISSN: 0967-0335 IM Publications LLP 2011
doi: 10.1255/jnirs.934 All rights reserved
Protecting the environment, reducing emissions of green-
house gases and increasing the use of renewable energy
sources in electricity production represent one of the major
strategic objectives for most countries in the world.
1
As a
traditional agricultural country, China has abundant biomass
resources. In 2010, China produced about 820 million tons

of

straw,
2
which is a major biomass resource. Co-ring biomass
with coal can reduce CO
2
, SO
2
and NOx emissions. Biomass
coal co-ring in existing boilers to generate electricity has
been proposed as a near-term, low-cost way to increase
the biomass use in electricity production while reducing CO
2

emissions.
36
Government policies which promote the generation of elec-
tricity from renewable energy sources would result in the
increased use of biomass in existing power plants. However,
there is no effective method for detecting the amount of
biomass in combustion materials and this is one of the critical
reasons for the lack of any government promotion policies or
appropriate subsidies.
Current research for a method to classify biomasscoal
blends mainly focuses on image recognition, gamma-ray
analysis, contents analysis of combustion gas and spectral
analysis methods. However, image recognition methods
cannot produce accurate detection results, gamma rays are
harmful to the human body and contents analysis of combus-
tion gas does not have the capacity for independent clas-
sication. Spectral analysis methods are considered to have
the greatest potential for the identication of biomass, coal
and biomasscoal blends due to its low cost, shorter analysis
time, non-destructive nature, accuracy and reliability when
combined with chemometric methods.
79
Near infrared reectance (NIR) spectroscopy is based on
the absorption of electromagnetic radiation in the region from
780 nm to 2500 nm (12,8204000 cm
1
).
10
It can also provide
A feasibility study on using near infrared
spectroscopy to classify strawcoal blends
Cheng He, Zengling Yang, Guangqun Huang, Longjian Chen and Lujia Han
*
College of Engineering, China Agricultural University, Beijing 100083, China. E-mail: hanlj@cau.edu.cn
Rapid classication of biomass, coal and biomasscoal blends is very important for developing appropriate subsidy policies and meth-
ods for biomass co-ring power generation in China. The use of near infrared reectance (NIR) spectroscopy to classify straw, coal and
strawcoal blends was explored in this study. Eighty-one straw samples, 9 coal samples, 81 strawcoal blends samples with straw
content from 91% to 99% (blends1) and 90 strawcoal blends samples with straw content from 1% to 30% (blends2) were prepared and
separated into a calibration set and an external prediction set. Spectra were scanned by a Fourier transform-NIR spectrometer. Raw
spectra of the samples were mean centred and compared using discriminant analysis. Overall, the correct classication percentages for
straw, blends1, blends2 and coal samples were 98.8%, 90.1%, 83.3% and 66.7%, respectively. The results show that the NIR technique
could provide excellent classication between straw and coal. The spectra of strawcoal blends with straw content less than 98% can
be classied exactly from calibrations based on pure straw and the spectra of strawcoal blends with straw content greater than 20%
can be well classied from calibrations based on pure coal. It is concluded that NIR is a feasible method for rapid classication of straw,
coal and strawcoal blends.
Keywords: near infrared spectroscopy, NIR, co-ring, biomass, straw, coal, renewable energy
Introduction
C. He et al., J. Near Infrared Spectrosc. 19, 277284 (2011)
Received: 14 June 2011 Revised: 18 August 2011 Accepted: 31 August 2011 Publication: 26 September 2011
278 A Feasibility Study Classifying StrawCoal Blends
comprehensive information at molecular level, concerning the
components and properties of samples.
11
It is fast, accurate,
simple to operate and non-destructive in nature.
12
Earlier research on the application of NIR to straw and coal
was mainly focused on analysis of fuel properties. Huang et
al.
1315
used NIR with chemometric methods to predict the
contents of moisture, ash, volatile matter, K, Na, Ca, Mg, Fe,
C, H, N and heating value in straw. Kim et al.
16
used NIR for
on-line measurement of coal properties such as contents of
moisture, ash, volatile mater, xed carbon, C, H, N, O, S and
heating value. Bona and Andres
1719
used NIR with different
chemometric methods to predict contents of moisture, ash,
volatile mater, xed carbon, C, H, N, O, S and heating value in
coal. However, there is no research about using NIR to classify
the straw, coal and strawcoal blends.
The aim of this study was to explore the feasibility of using
NIR for the classication of straw, coal and strawcoal blends.
The power generation industry is able to utilise strawcoal
blends with a straw content of between 1% and 99%. If the
blends contain less than 30% straw, the plants do not need
to adjust the boiler systems during power generation. But
government subsidies encourage the industry to use straw
despite it having only about half the energy content of coal.
Hence there is interest in blends with a high straw content.
For these reasons the extreme situations for straw content in
strawcoal blends were considered in this study.
Materials and methods
Sample collection and preparation
This study was carried out using samples in four classes:
straw, coal, strawcoal blends samples with straw content
from 91% to 99% (blends1) and strawcoal blends samples
with straw content from 1% to 30% (blends2).
Eighty-one straw samples, which were different in terms of
location, variety, soil characteristics, growing climate and farm
management, were all wheat straw, harvested in the year of
2010 and collected from seven provinces of China. After being
air dried and cleaned by removing impurities, straw samples
were put into a mill (Knifetec 1095; FOSS, Hillerd, Denmark)
and ground for 20 s with a rotor speed of 20,000 rpm. Nine coal
samples, which were collected from nine coal mines in China,
were put into a small herbal mill (Taisite FW135; Tianjin, China)
and ground for 20 s with 24,000 rpm rotor speed. The particle
sizes of straw and coal samples were less then 5 mm.
Eighty-one blends1 samples were produced using straw
and coal samples with straw content from 91% to 99% and 1%
content increments. Ninety blends2 samples were prepared
using straw and coal samples with straw content from 1% to
30% and 1% content increments. In order to mix the straw
and coal sufciently, blends1 and blends2 samples were put
into a mixer (REAX 20/8; Heidolph, Schwabach, Germany) and
mixed for 2 h with a rotor speed of 10 rpm. Before spectral
analysis, all samples were stored in plastic cups with a volume
of 1177 cm
3
in a dry and shady area.
The samples were divided into a calibration set and a predic-
tion set. First, straw and coal samples were divided into a cali-
bration set and a prediction set according to a 2/3 and 1/3 ratio
of samples based on number, respectively. Then straw and
coal samples in the calibration set were used to prepare cali-
bration set blends1 and blends2 samples; meanwhile, straw
and coal samples in the prediction set were used to prepare
prediction sets for blends1 and blends2 samples. The calibra-
tion set samples were used to develop calibration models
and then used to predict unknown samples. The number of
samples and their composition in the calibration and predic-
tion sets are presented in Table 1.
Near infrared spectroscopy scanning
NIR spectra were recorded in the reectance mode using an
FT-NIR spectrometer (Spectrum 400; PerkinElmer, Shelton,
CT, USA) with an integrating sphere. The spectrometer was
equipped with a halogen lamp as a light source and an InGaAs
detector. Reectance energy was referenced to a pure white
standard. Each spectrum used for the data analysis was
recorded for the range 10,000 cm
1
to 4000 cm
1
with 8 cm
1

resolution and was the average of 32 scans. Each spec-
trum had 3001 data points. Samples were placed in a quartz
cylinder-shaped sample cup with a diameter of 90 mm and a
height of 16 mm. NIR radiation was injected into samples from
the bottom of the sample cup and samples were scanned in
three replicates. The three spectra from each sample were
averaged before calibration and prediction.
Spectral classication method
We used a principal components analysis (PCA) model to
evaluate the distribution of samples based on a scores plot.
Before developing models to predict class membership for
new samples, the specicity of models was evaluated to check
if the classes overlapped or were sufciently distant from each
other. A PCA model of spectral data of all samples, which
included 81 straw samples, 81 blends1 samples, 90 blends2
samples and 9 coal samples, was calculated, considering a
maximum number of ten principal components.
Prior to spectral classification, NIR spectra were mean
centred. Then discriminant analysis was used for clas-
sifying the four categories of samples. A measurement of
the Mahalanobis distance
20
between the unknown samples
and each known class centre was calculated. The result of
a discriminant analysis is the name of the class that is most
similar to the spectrum of the unknown sample. The closer
each Mahalanobis distance value is to zero, the better is the
match.
20,21
Mahalanobis distance can be defined as a dissimilarity
measure between two random vectors x

and y

of the same
distribution with the covariance matrix S as shown below:
1
(1) ( , ) ( ) ( )
T
d x y x y x y

= S
G G G G G G
The analysis was done by using TQ Analyst software (Thermo
Fisher Scientic Inc., USA).
C. He et al., J. Near Infrared Spectrosc. 19, 277284 (2011) 279
The correct classication rate (CC) is the criterion for the
classication result. The equation of CC is:
CC = N
c
/(N
c
+ N
ic
) 100% (2)
where N
c
is the number of correct classications and N
ic
is the
number of incorrect classications.
22
Results and discussion
Spectral characteristic
NIR spectra of straw, blends1, blends2 and coal samples are
shown in Figures 1(a)(d), respectively. The spectra of straw
and blends1 samples showed absorption peaks around
6897 cm
1
, 5618 cm
1
, 5495 cm
1
, 5155 cm
1
, 4440 cm
1
,
4394 cm
1
, 4280 cm
1
and 4252 cm
1
, while the spectra of
coal and blends2 samples showed absorption peaks around
5155 cm
1
, 4587 cm
1
and 4329 cm
1
. The assignments for
these absorption peaks, according to Li,
23
are presented in
Table 2. Water absorption peaks are around 6897 cm
1
and
5155 cm
1
. Peaks around 5618 cm
1
, 5495 cm
1
, 4280 cm
1

and 4252 cm
1
are attributed to cellulose; peaks around
4440 cm
1
and 4394 cm
1
are attributed to lignin. These
peaks appear in spectra of both straw and blends1 samples
because cellulose and lignin are an important part of straw,
and blends1 has a high content of straw. Bands around
4587 cm
1
and 4329 cm
1
arise from coal; these bands
appear in spectra of samples of both blends2 and coal. A
direct comparison between the spectra of straw for blends1
samples shows some similarities due to the high content
of straw in blends1, while the spectra of coal and blends2
are similar because of the high content of coal in blends2.
However, the absorbance of coal and blends2 are much
higher than straw and blends1 due to the strong absorp-
tion of near infrared radiation by coal. It is evident that the
signicant differences in the NIR spectra of the four sample
types can provide a sound base for the classification of
straw, blends1, blends2 and coal.
Principal component analysis
Figures 2(a) and (b) show the scores plot obtained by PCA for
samples of straw and coal, respectively. In Figure 2(a), prin-
cipal components 1 (PC1) accounts for 82.1% and principal
components 2 (PC2) accounts for 16.1% of total variance. In
Figure 2(b), PC1 accounts for 96.0% and PC2 accounts for
3.7% of total variance. The wide distribution of the scores indi-
cated that straw and coal samples were quite diverse.
Figure 3 shows scores plots for PC1 and PC2 plane of
all spectra. PC1 accounted for 99.8% and PC2 accounted
for 0.2% of the total variance. Samples from straw and
blends1 both appear on the left of the plot, while samples
from blends2 and coal both appear on the right of the plot,
but samples of straw and samples of blends1 are very close,
and samples of blends2 and samples of coal are overlapped,
which indicates that it is very difcult to distinguish blends2
samples from coal.
C
l
a
s
s
N
u
m
b
e
r

o
f

s
a
m
p
l
e
s
S
t
r
a
w

c
o
n
t
e
n
t

i
n
c
a
l
i
b
r
a
t
i
o
n

s
e
t

(
%
)
S
t
r
a
w

c
o
n
t
e
n
t

i
n
p
r
e
d
i
c
t
i
o
n

s
e
t

(
%
)
C
a
l
i
b
r
a
t
i
o
n
s
e
t
V
a
l
i
d
a
t
i
o
n
s
e
t
R
a
n
g
e
I
n
c
r
e
m
e
n
t
M
e
a
n
S
t
a
n
d
a
r
d
d
e
v
i
a
t
i
o
n
R
a
n
g
e
I
n
c
r
e
m
e
n
t
M
e
a
n
S
t
a
n
d
a
r
d
d
e
v
i
a
t
i
o
n
S
t
r
a
w
5
4
2
7
1
0
0
0
1
0
0
0
1
0
0
0
1
0
0
0
B
l
e
n
d
s
1
5
4
2
7
9
1

9
9
1
9
5
.
0
2
.
6
9
1

9
9
1
9
5
.
0
2
.
6
B
l
e
n
d
s
2
7
5
1
5
1

3
0
1
1
5
.
5
8
.
7
1

2
9
2
1
5
.
0
8
.
9
C
o
a
l

6

3
0
0
0
0
0
0
0
0
T
a
b
l
e

1
.

T
h
e

n
u
m
b
e
r

o
f

s
a
m
p
l
e
s

a
n
d

t
h
e
i
r

c
o
m
p
o
s
i
t
i
o
n

i
n


t
h
e

c
a
l
i
b
r
a
t
i
o
n

a
n
d

p
r
e
d
i
c
t
i
o
n

s
e
t
s
.
280 A Feasibility Study Classifying StrawCoal Blends
Classication results of samples
Table 3 and Table 4 show the results of the prediction of
four categories of samples (straw, blends1, blends2 and
coal) in the calibration set and the validation set, respec-
tively. Samples of straw were classied with CC of 98.2%
in the calibration set, with only one sample of straw being
misclassied as blends1. In the validation set, samples of
straw were correctly classied with CC of 100%. Samples of
blends1 were classied with CC of 90.7% in the calibration
set, with ve samples from blends1 being misclassied as
straw samples. In the validation set, samples of blends1
were classied with CC of 88.9%, with three samples from
blends1 being misclassified as straw samples. Samples
of blends2 were classified with CC of 82.7% and 86.7%
in the calibration set and the validation set, respectively.
All the misclassied blends2 samples were classied as
coal samples. Samples of coal were classied with CC of
83.3% and 33.3% in the calibration set and the validation
set, respectively; the misclassied coal samples were all
classied as blends2 samples. These results indicate that
it was very unlikely to incorrectly classify straw and blends1
as coal or a blends2 sample. Overall, CC of straw, blends1,
blends2 and coal were 98.8%, 90.1%, 83.3% and 66.7%,
respectively.
The distance from each sample of straw, blends1, blends2
and coal to the model of straw and coal is shown in Figure 4.
It can be easily recognised that the two groups of straw and
blends1 are perfectly separated from the other two groups of
4000 5000 6000 7000 8000 9000 10000
0.2
0.3
0.4
0.5
0.6
0.7
0.8
4000 5000 6000 7000 8000 9000 10000
0.3
0.4
0.5
0.6
0.7
0.8
0.9
Wavenumber (cm
-1
)
L
o
g

(
1
/
R
)

(a)
Wavenumber (cm
-1
)
L
o
g

(
1
/
R
)

(b)
4000 5000 6000 7000 8000 9000 10000
0.8
0.9
1.0
1.1
1.2
1.3
1.4
1.5
1.6
4000 5000 6000 7000 8000 9000 10000
0.8
0.9
1.0
1.1
1.2
1.3
1.4
1.5
1.6
Wavenumber (cm
-1
)
L
o
g

(
1
/
R
)

(c)
Wavenumber (cm
-1
)
L
o
g

(
1
/
R
)

(d)
Figure 1. Spectra of straw (a), blends1 (b), blends2 (c) and coal (d).
Band
(cm
1
)
Main attribution
6897 OH stretching, rst overtone
5618 CH stretching, rst overtone
5495
OH stretching + CO stretching,
second overtone
5155 OH stretching + OH deformation
4587 amide I, second overtone, + amide III
4440 OH stretching + OH deformation
4394 OH stretching + CC stretching
4329 CH stretching + CH deformation
4280 CH stretching + CH deformation
4252 CH deformation, second overtone
Table 2. Assignments of the main absorption peaks.
23
C. He et al., J. Near Infrared Spectrosc. 19, 277284 (2011) 281
blends2 and coal. A small overlap between straw and blends1
with different straw content can be observed and a signicant
overlap occurred between coal and blends2 with different
straw content. These results are in line with the spectral char-
acteristic and PCA analysis.
Figure 5 shows the incorrect classication rates of blends1
samples with different straw content. The incorrect classica-
tion rates of blends1 samples with 99% and 98% straw content
were 66.7% and 22.2%, respectively. When the straw content
of blends1 was less than 98%, samples of blends1 and straw
were all correctly classied.
Figure 6 shows the incorrect classication rates of blends2
samples with different straw content. The incorrect classi-
cation rates of blends2 with the straw content at the range
of 16%, 712%, 1318%, 1924% and 2530% were 46.7%,
26.7%, 20.0%, 6.7% and 0.0%, respectively. At the straw content
range of 1924%, only one sample with 20% straw content was
misclassied. Different incorrect classication rates occurred
when using NIR spectra and discriminate analysis to classify
coal and blends2 with high content (larger than 80%) of coal.
As the coal content increased, the percentage of incorrect
-1.5
-1.0
-0.5
0.0
0.5
1.0
1.5
-2.5 -2.0 -1.5 -1.0 -0.5 0.0 0.5 1.0 1.5 2.0 2.5
-3
-2
-1
0
1
2
3
-8 -6 -4 -2 0 2 4 6 8
PC1, 82.1 %
P
C
2
,

1
6
.
1

%

(a)
PC1, 96.0 %
P
C
2
,

3
.
7

%


(b)
Figure 2. Scores plot obtained by PCA for samples of (a) straw
and (b) coal.
Actual
Predicted
Straw Blends1 Blends2 Coal Total
Correct classication
(%)
Straw 53 1 0 0 54 98.2
Blends1 5 49 0 0 54 90.7
Blends2 0 0 62 13 75 82.7
Coal 0 0 1 5 6 83.3
-4
-3
-2
-1
0
1
2
3
4
-40 -30 -20 -10 0 10 20 30 40
PC1, 99.8 %
P
C
2
,

0
.
2

%


Figure 3. Scores plot obtained by PCA for principal components
1 (PC1) and principal components 2 (PC2) for samples of straw
(), blends1 ({), blends2 () and coal (S).
Table 3. Results of prediction of straw, blends1, blends2 and coal in the calibration set.
282 A Feasibility Study Classifying StrawCoal Blends
classications increased. When the straw content of blends2
was higher than 20%, the blends2 and coal samples were
correctly classied.
The classication of straw and blends1 provided an excellent
CC when the coal content was more than 2% in blends1. The
incorrect classication rate was very high when the coal content
was equal or less then 2% in blends1, which was mainly due to
the low content components detection limitation of NIR tech-
nology. When the coal content was larger than 80% of blends2,
it was very difcult to classify blends2 and coal. The results
of classication of blends2 and coal were worse than that of
blends1 and straw. This observation was mainly because of
sample grading when samples were put into the sample cup.
Straw fragments were mainly found in the upper layers in the
sample cup due to their lower density, while coal was generally
in the lower layer in the sample cup because of its higher density.
The NIR radiation was injected from the bottom of the sample
cup, therefore, the coal content dominated in the spectra of
blends2. The strong NIR spectroscopy absorption of coal led
to the situation that most of the spectral information of straw
was covered by coal, so the spectra of blends2 mainly reected
the spectral characteristic of coal. When the coal content of
strawcoal blends was in the range of 280%, the percentage
of straw was signicantly larger in the strawcoal blends due to
the smaller density compared with coal so the straw and coal
information could be presented by NIR. Therefore, it was easy to
classify straw, coal and strawcoal blends with the coal content
in the range of 280%.
0
1
2
3
4
5
6
7
8
0 1 2 3 4 5 6 7 8
Distance to straw
D
i
s
t
a
n
c
e

t
o

c
o
a
l

Figure 4. The Mahalanobis distance from each sample of straw
(), blends1 ({), blends2 () and coal (S). to model of straw
and coal.
Table 4. Results of prediction of straw, blends1, blends2 and coal in the validation set.
Actual
Predicted
Straw Blends1 Blends2 Coal Total Correct classication
(%)
Straw 27 0 0 0 27 100
Blends1 3 24 0 0 27 88.9
Blends2 0 0 13 2 15 86.7
Coal 0 0 2 1 3 33.3
66.7
22.2
0.0 0.0 0.0 0.0 0.0 0.0 0.0
0
10
20
30
40
50
60
70
80
90
100
99 98 97 96 95 94 93 92 91
Straw content (%)
I
n
c
o
r
r
e
c
t

c
l
a
s
s
i
f
i
c
a
t
i
o
n

r
a
t
e

(
%
)

Figure 5. Incorrect classication rates (%) of blends1 samples
with different straw content.
46.7
26.7
20.0
6.7
0.0
0
10
20
30
40
50
60
70
80
90
100
1-6 7-12 13-18 19-24 25-30
Straw content (%)
I
n
c
o
r
r
e
c
t

c
l
a
s
s
i
f
i
c
a
t
i
o
n

r
a
t
e

(
%
)

Figure 6. Incorrect classication rates (%) of blends2 samples
with different straw content.
C. He et al., J. Near Infrared Spectrosc. 19, 277284 (2011) 283
Conclusions
This study explored the feasibility of using NIR for the rapid
classification of straw, coal and strawcoal blends. The
results showed that the NIR technique could provide correct
classifications between straw and coal. The strawcoal
blends with straw content less than 98% can be classied
exactly from straw and the strawcoal blends with straw
content larger than 20% can be well classied from coal.
However, the correct classication rate between coal and
strawcoal blends with high content (larger than 80%) of
coal need to be further improved when using NIR. In conclu-
sion, NIR has proved to be a feasible method for the rapid
classications of straw, coal and strawcoal blends. The
ndings of this study provide a sound base for developing
an effective biomass detection method which can be used
to support the implementation of the government subsidy
policies.
Acknowledgements
This work was supported by Special Fund for Agro-scientic
Research in the Public Interest (Project No. 201003063) and
Natural Science Foundation of China (Project No. 31101072).
References
1. F. Al-Mansour and J. Zuwala, An evaluation of bio-
mass co-ring in Europe. Biomass 34, 620 (2010). doi:
10.1016/j.biombioe.2010.01.004
2. Anonymous, The National Crops Straw Resource Survey
and Evaluation Report. Ministry of Agriculture of the
Peoples Republic of China, China (2011).
3. P. Basu, J. Butler and M.A. Leon, Biomass co-ring
options on the emission reduction and electricity gen-
eration costs in coal-red power plants, Renew. Energ.
36, 282 (2011). doi: 10.1016/j.renene.2010.06.039
4. L. Baxter, Biomasscoal co-combustion: opportunity
for affordable renewable energy. Fuel 84, 1295 (2005).
doi: 10.1016/j.fuel.2004.09.023
5. J. Hansson, G. Berndes, F. Johnsson and J. Kjarstad,
Co-ring biomass with coal for electricity generation
An assessment of the potential in EU27, Energ. Policy
37, 1444 (2009). doi: 10.1016/j.enpol.2008.12.007
6. J. Lintunen and H. Kangas, The case of co-ring:
The market level effects of subsidizing biomass
co-combustion, Energ. Econ. 32, 694 (2010). doi:
10.1016/j.eneco.2009.11.004
7. M.F. Ferrao, M.D.S. Viera, R.E. Panta Pazos, D. Fachini,
A.E. Gerbase and L. Marder, Simultaneous deter-
mination of quality parameters of biodiesel/diesel
blends using HATR-FTIR spectra and PLS, iPLS or
siPLS regressions, Fuel 90, 701 (2011). doi: 10.1016/j.
fuel.2010.09.016
8. C. Pasquini and A.F. Bueno, Characterization of petro-
leum using near-infrared spectroscopy: Quantitative
modeling for the true boiling point curve and
specic gravity, Fuel 86, 1927 (2007). doi: 10.1016/j.
fuel.2006.12.026
9. G. Veras, A.D.A. Gomes, A.C. Da Silva, A.L. Bizerra
De Brito, P.B. Alves De Almeida and E.P. de Medeiros,
Classication of biodiesel using NIR spectrometry and
multivariate techniques, Talanta 83, 565 (2010). doi:
10.1016/j.talanta.2010.09.060
10. R.M. Balabin and R.Z. Saeva, Biodiesel classication
by base stock type (vegetable oil) using near infrared
spectroscopy data, Anal. Chim. Acta 689, 190 (2011). doi:
10.1016/j.aca.2011.01.041
11. Y.H. Wu, Y. Zheng, Q.Q. Li, J. Lqbal, L.Y. Zhang, W.B.
Zhang and Y.P. Du, Study on difference between epider-
mis, phloem and xylem of Radix ginseng with near-infra-
red and infrared spectroscopy coupled with principal
component analysis, Vib. Spectrosc. 55, 201 (2011). doi:
10.1016/j.vibspec.2010.11.004
12. G.Q. Huang, L.J. Han and X. Liu, Rapid estimation of the
composition of animal manure compost by near infrared
reectance spectroscopy, J. Near Infrared Spectrosc. 15,
387 (2007) doi: 10.1255/jnirs.745
13. C.J. Huang, L.J. Han, Z.L. Yang and X. Liu, Ultimate
analysis and heating value prediction of straw by near
infrared spectroscopy, Waste Manag. 29, 1793 (2009).
doi: 10.1016/j.wasman.2008.11.027
14. C.J. Huang, L.J. Han, Z.L. Yang and M. Liu, Prediction
of heating value of straw by proximate data, and near
infrared spectroscopy, Energ. Convers. Manage. 49, 3433
(2008). doi: 10.1016/j.enconman.2008.08.020
15. C.J. Huang, L.J. Han, Z.L. Yang and M. Liu, Exploring
the use of near infrared reectance spectroscopy to pre-
dict minerals in straw, Fuel 88, 163 (2009). doi: 10.1016/j.
fuel.2008.07.031
16. D.W. Kim, J.M. Lee and J.S. Kim, Application of near
infrared diffuse reectance spectroscopy for on-line
measurement of coal properties, Korean J. Chem. Eng.
26, 489 (2009). doi: 10.1007/s11814-009-0083-0
17. J.M. Andres and M.T. Bona, Analysis of coal by diffuse
reectance near-infrared spectroscopy, Anal. Chim.
Acta 535, 123 (2005). doi: 10.1016/j.aca.2004.12.007
18. J.M. Andres and M.T. Bona, ASTM clustering for
improving coal analysis by near-infrared spectroscopy,
Talanta 70, 711 (2006). doi: 10.1016/j.talanta.2006.05.034
19. M.T. Bona and J.M. Andres, Coal analysis by diffuse
reectance near-infrared spectroscopy: Hierarchical
cluster and linear discriminant analysis, Talanta 72,
1423 (2007). doi: 10.1016/j.talanta.2007.01.050
20. P.C. Mahalanobis, On the generalized distance in statis-
tics, Proc. Natl Inst. Sci. India 2, 49 (1936).
21. E.G. Gath and K. Hayes, Bounds for a multivariate
extension of range over standard deviation based on the
Mahalanobis distance, Linear Algebra Appl. 435, 1267
(2011). doi: 10.1016/j.laa.2011.03.039
284 A Feasibility Study Classifying StrawCoal Blends
22. P. Ciosek, Z. Brzozka, W. Wroblewski, E. Martinell, C.
Di Natale and A.D. Amico, Direct and two-stage data
analysis procedures based on PCA, PLS-DA and ANN
for ISE-based electronic tongueEffect of supervised
feature extraction, Talanta 67, 590 (2005). doi: 10.1016/j.
talanta.2005.03.006
23. M.Z. Li, Spectral Analysis Technology and its Application.
Science Press, Beijing, China (2006).
JOURNAL
OF
NEAR
INFRARED
SPECTROSCOPY
285
ISSN: 0967-0335 IM Publications LLP 2011
doi: 10.1255/jnirs.937 All rights reserved
Figures 3(a) and 4(a) of this paper published as J. Near Infrared
Spectrosc.19(3), 171180 (2011) were incorrect. The correct
version of these gures is printed below. We apolgise for any
confusion.
Erratum
Classification of viable and non-viable
spinach (Spinacia oleracea L.) seeds by
single seed near infrared spectroscopy and
extended canonical variates analysis
Merete Halkjaer Olesen,
*
Nisha Shetty, Rene Gislum and Birte Boelt
Department of Genetics and Biotechnology, Aarhus University, Faculty of Agricultural Sciences, Forsgsvej 1, DK-4200 Slagelse, Denmark.
E-mail: merete.olesen@agrsci.dk
M.Hallkjaer Olesen et al., J. Near Infrared Spectrosc. 19, 285286 (2011)
Figure 3. Extended canonical variance component 1 (ECV#1) plot against component 2 of MSC data. (a) Seed with and without pericarp
determined as naked and coated and (a+b) the effect of accelerated aging of the seeds.
(a)
(b)
286 Erratum
Figure 4. Sample number plot against (a) ECV component 1 of EMSC transformed spectra and (b) ECVA loading plot. Results are from
seed lot A with preicarp and germination counted after 21 days.
(a)
(b)
JOURNAL
OF
NEAR
INFRARED
SPECTROSCOPY
The Journal of Near Infrared Spectroscopy (JNIRS) aims to
publish original research papers, short communications,
review articles and letters concerned with near infrared spec-
troscopy and technology, its application, new instrumentation
and the use of chemometric and data handling techniques
within NIR. JNIRS will accept contributions from all areas of
applications where near infrared spectroscopy is in use, such
as:
Agriculture Chemical Industry
Food Life Sciences
Process Control Pharmaceuticals
Textiles Polymers
Wood
Topics for papers may include:
Chemometrics Calibrations
Diffuse Reection NIR Imaging
On-Line Use Fibre Optics
Sampling Spectroscopy
Instrumentation Remote Sensing
Types of article
All articles should be written in English and will be subject to
a refereeing process. Letters may be accepted for publication
following review by the Editor or a Subject/Regional Editor
alone.
Original research papers should report, in detail, the results
of the author(s)s work in a particular area of research. This
work should have been completed or a substantial and identi-
able part of it should have been completed.
Short communications can be used to report significant
results from incomplete work and to report completed work
which does not require the full coverage of a research paper.
Short communications should be restricted to 1500 words in
length.
Review articles will normally be commissioned by the Editor
or one of the Regional or Subject Editors. However, should you
be interested in reviewing a particular area, you are welcome
to discuss this with the Editor or your local Regional or
Subject Editor. Review articles should provide a comprehen-
sive coverage of the area reviewed with extensive references
to published work.
Letters can either express the view of the author on a partic-
ular aspect of NIR spectroscopy and technology that they wish
to comment on, or can very briey report an interesting aspect
of the authors work: perhaps details of a calibration or a
sampling procedure.
The manuscript
For the fastest review and publication, please submit your
paper online at www.impublications.com/authors. If you prefer,
an electronic version may be sent by e-mail or on CD-ROM
Instructions for authors
Aims and scope
Did you know?
JNIRS is the only peer review journal dedicated to NIR spectroscopy.
There are NO page charges.
All papers are reviewed by acknowledged experts in NIR technology.
You are ensured that when your paper has been accepted, it will be published rapidly.
JNIRS is the global audience for your paper on any aspect of NIR spectroscopy, technology or its application.
You can submit your paper online at www.impublications.com/authors.
to the Editor-in-Chief, the publishers or a Subject/Regional
Editor.
It is important to remember that the manuscript is going
to be read by referees. Please assist them by observing the
following specication.
Use Times or Times New Roman with a font size of at
least 12. The manuscript should be double line spaced with
generous margins. Please number the pages and use the line
number option if your word processor has it.
JNIRS is computer typeset and authors should include a
computer-readable version of their manuscriptthere is no
need for any hard copy. This not only speeds up the publication
of the article but also reduces the possibility for errors and
reduces the proof reading burden for author and publisher
alike. We can read most common PC word processor les.
We can also accept illustrations in computer readable form,
such as PDF, TIFF, PCX or EPS format. Please ensure that the
highest quality settings are used for their creation and that
any fonts are embedded or converted to outlines. Bitmaps
should be at least 300 dpi for tone images and at least 1200 dpi
for line work. Please avoid JPGs, because they are a lossy
format: each time you save a le as a JPG it loses quality. If
you do not save with the highest quality settings, your gure
will be very blurry.
Symbols, mathematics and formulae
This is an area where care should be exercised. You should use
SI units whenever possible. Please use the following conven-
tions: all symbols which are variables should be italicised, all
constants should be normal (Roman) text and vectors and
matrices should be bold.
Complex formulae can be prepared in equation editors,
ideally MathType. Please note that we do not have the facility
to process teX.
Units
Units of measurement should be separated from the number
by a space. Please avoid the use of the solidus (/) or per and
use a superscript
1
, i.e. 5 mg L
1
, not 5 mg/L.
References
References should be cited in the text by a superscript Arabic
number, i.e. spectroscopy,
1
and listed at the end of the article.
Please give full bibliographical information, including the title
of all papers and book chapters, as in the examples below:
For edited books:
1. G.D. Batten, A.B. Blakeney and C.R. Blatt, Sample
preparation procedures for use in near infrared
analysis of fruit and vegetable crops, in Near Infrared
Spectroscopy: The Future Waves, Ed by A.M.C. Davies
and P. Williams. NIR Publications, Chichester, p. 107
(1996).
For journals:
2. T. Ns, K. Kvaal, T. Isaksson and C. Miller, Articial
neural networks in multivariate calibration, J. Near
Infrared Spectrosc. 1, 1 (1993).
For authored books:
3. B. Osborne, T. Fearn and P.H. Hindle, Practical NIR
Spectroscopy with Applications in Food and Beverage
Analysis. Longman Scientic and Technical, Harlow, UK
(1993).
It is the responsibility of the author to ensure that permis-
sion is obtained from the copyright holder to reproduce copy-
right material. The copyright holder is almost always the
publisher of the work. Forms for this purpose are available
from the Editor and from IM Publications. Copies of any signed
Permission Forms should accompany the manuscript and
the necessary wording of acknowledgement requested by the
copyright holder should be included in the caption.
Abstract
A short abstract not exceeding 350 words should be included.
This should describe the scope of the article and highlight any
signicant areas.
Keywords
Up to 10 Keywords should be included.
Illustrations
Illustrations must be produced ready for publication and
meeting our guidelines set out above.
Labels on the axes should be clear and of sufcient size
that they will be legible when reduced to t on the page, i.e. a
font size of at least 8 pt when produced at one column width.
(This requires a font size of at least 20 pt if the gure is drawn
full size on A4 paper. Note that ordinate and abscissa scales
should be equal for equal values. Referees expect to see
square plots of results of, for instance, predicted versus
reference method. Please use Helvetica or Arial font. The
amount of text on the illustration should be kept to a minimum.
Keys for different line styles etc. should be given in the caption;
not on the illustration.
Illustrations should be numbered in the text with Arabic
numerals as Figure 1, Figure 4 and 5 etc.; they should be in
a separate sequence from any tables. Captions should be
provided for all gures and these should appear at the end of
the manuscript.
If embedded in the word processor le, each illustration
should be provided on a separate page as large as possible.
Please label each page with the Figure number, author(s) and
title of the paper.
Photographs can be published but they must be of good
photographic quality and supplied as high quality images at a
minimum of 300 dpi. Colour illustrations can only be included
if the author meets the extra cost of their production. However,
any illustrations can be reproduced in colour in the online
edition at no cost and in black-and-white in print. If you include
colour gures, this will happen automatically, so you must
inform the Editor if you wish to pay for colour reproduction.
Tables
Tables should be numbered in a separate sequence from the
illustrations as Table 1 etc. Please consider the size of the
page when compiling your Table. Tables should not be split
width-wise across pages, although if they exceed a page in
length they can be continued on subsequent pages.
Tables should be provided on separate pages and not
included in the text. Each Table should have a caption. As
with illustrations, please number each sheet and include the
author(s) and title of the paper.
Copyright
Authors are required to assign the copyright in their article
to the Publisher. A completed and signed Copyright Form
must be lodged with the Publisher before the article can be
published.
If your institution or funding body requires that you deposit a
copy of the paper in an open repository or you wish your paper
to be freely available, we offer an Open Access publishing
model. Please visit the journals website (www.impublications.
com/jnirs/) for more information. Please note that you should
not request this until after your paper has been accepted for
publication.
Submission of articles
Your article should be submitted via our online system at www.
impublications.com/authors/. If you prefer, it may be sent to
the Editor or one of the Regional Editors whose addresses
appear below.
Please use the Check List at the end of these instructions
before posting your article.
Receipt of the articles will be acknowledged, and the article
will be subjected to a peer review process; should any amend-
ments be necessary you will be asked to make these before
the article is accepted for publication.
After acceptance, your article will be copy-edited, typeset
and proofed to the first named author (unless otherwise
requested). The proof should be checked carefully for typo-
graphical errors and to answer any questions raised by the
Copy Editor. The proof stage is not an opportunity for the
author to make changes to the article. Any signicant changes
from the original accepted manuscript may only be accepted
on payment of the correction costs.
Addresses
EDITOR-IN-CHIEF: Professor Graeme D. Batten, Sea Spec Pty
Ltd, PO Box 487, Woolgoolga, NSW 2456, Australia. Phone / Fax
+61-266-562288. E-mail: thebattens@bigpond.com
Authors distant from Australia are welcome, if they prefer, to
send their manuscript to IM Publications LLP, 6 Charlton Mill,
Charlton, Chichester, West Sussex PO18 0HY, UK, who will be
pleased to forward it to Graeme Batten.
REGIONAL EDITORS: Dr S. Kawano, National Food Research
Institute, 2-1-12 Kannondai, Tsukuba 305-8642, Japan, E-mail:
kawano@affrc.go.jp. Dr G. Downey, Ashtown Food Research
Centre, Ashtown, Dublin 15, Ireland, E-mail: gerard.downey@
teagasc.ie. Dr Franklin E. Barton, II, Light Light Solutions, LLC,
PO Box 81486, Athens, GA 30608-1486, USA, E-mail: barton@
lightlightsolutions.com
SUBJECT EDITORS: Chemistry/Spectroscopy: Professor H.W.
Siesler, Department of Physical Chemistry, University of Essen,
PO Box 103674, D-45117 Essen, Germany. E-mail: hw.siesler@
uni-due.de. Chemometrics: Professor T. Fearn, Department of
Statistical Science, University College London, Gower Street,
London WC1E 6BT, UK. E-mail: tom@stats.ucl.ac.uk. Clinical/
Medical: Mr K.H. Norris, Consultant, 11204 Montgomery Road,
Beltsville, MD 20705, USA, E-mail: knnirs@ gmail.com. Textile
Fibres: Dr J. Montalvo, USDA, ARS, SRRC, PO Box 19687, New
Orleans, LA 70179, USA. E-mail: montalvo@srrc.ars.usda.
gov. Food: Professor T. Isaksson, Department of Chemistry,
Biotechnology and Food Science UMB, Norwegian University
of Life Sciences, PO Box 5036, N-1432 s, Norway. E-mail:
tomas.isaksson@umb.no. Pharmaceutical: Carl Anderson,
Graduate School of Pharmaceutical Sciences, Duquesne
University, Pittsburgh, PA 15282, USA. E-mail: andersonca@
duq.edu. Process Control: Dr L. Weyer, ATK Elkton, 55 Thiokol
Rd, Elkton, MD, USA. E-mail: lweyer2@verizon.net.
Check list
1. Are pages and lines numbered?
2. Is the text double line spaced?
3. Are illustrations prepared according to the instructions?
4. Is there an abstract?
5. Are there keywords?
6. Are all the gures refereed to in the text?
7. Are all the references refereed to in the text, and in
numerical order according to their rst citation?
8. Are the reference citations formed according to the
instructions?
9. Are the tables prepared according to the instructions?
10. Have the equations been carefully checked?
11. Have you prepared text and illustration les in electronic
format?
12. Are you prepared to pay for the cost of printing colour
photographs or illustrations?
13. Is there a signed copyright form?
14. Have you provided full contact information: mailing
address, telephone, fax number and e-mail address.


6 Charlton Mill, Charlton
Chichester
West Sussex PO18 0HY, UK
T: +44(0)1243-811334
F: +44(0)1243-811711
E: info@impublications.co.uk
W: www.impublications.com


Copyright Transfer and Publication Agreement


In order to publish your article we need your agreement in writing. Please take a moment to
read the terms of this agreement, sign the form and return it to us as quickly as possible.

Name of Publication
Journal of Near Infrared Spectroscopy .......................................................................................
Title of Article
...................................................................................................................................................
Name(s) of Author(s)
...................................................................................................................................................

...................................................................................................................................................
...................................................................................................................................................
Name of Copyright Owner, if not Author(s)
...................................................................................................................................................
Address of above (if applicable)
...................................................................................................................................................
...................................................................................................................................................
...................................................................................................................................................
1. By signing this form, you (the author(s) or other copyright owner) hereby transfer
your copyright to us (IM Publications LLP). In particular, this means that you grant us the
exclusive right, for the full term of copyright and any renewals/extensions thereof, both to
reproduce and/or distribute your article (including the abstract) ourselves throughout the
world in printed, electronic or any other medium, and in turn to authorise others
(including Reproduction Rights Organisations such as the Copyright Licensing Agency
and the Copyright Clearance Centre) to do the same. You agree that we may publish
your article in the publication named above, and that we may sell or distribute it within
the publication, on its own, or with other related material. In return, we agree to publish
your article at our expense.
2. You warrant (i.e. promise) that the article is your original work, has not previously been
published, and is not currently under consideration by another publication. If it contains
material which is someone elses copyright, you promise that you have obtained the
unrestricted permission of the copyright owner (please attach evidence of this) and that
the material is clearly identified and acknowledged within the text. You also promise that
the article does not, to the best of your knowledge, contain anything which is libellous,
illegal, or infringes anyones copyright or other rights.

3. We promise that we will respect your rights as the author(s). That is, we will make sure
that your name(s) is/are always clearly associated with the article and, while you do
allow us to make necessary editorial changes, we will not make any substantial alteration
to your article without consulting you. When the article is accepted, we undertake to
publish it as soon as we reasonably can. If it is rejected, this agreement is automatically
cancelled and all the rights come back to you. On publication, we will send you a free
copy. We will do everything we reasonably can to maximise the visibility of the
publication, and of your article within it.
4. Copyright is transferred to us, and we will indicate this in the copyright line which
appears on your article. However, we grant back to you the right to use your own article
(provided you acknowledge the published original in standard bibliographic citation form)
in the following ways, as long as you do not sell it or give it away in ways which would
conflict directly with our commercial business interests. You are free to use your article
for the internal educational or other purposes of your own institution or company;
mounted on your own or your institutions internal website, but not on any publicly-
accessible website or other method of distribution; or in whole or in part, as the basis for
your own further publications or spoken presentations.

Signature(s) of Author(s) ... .......................................................................................................
Date ...................................

........................................... .......................................................................................................
Date ...................................

........................................... .......................................................................................................
Date ...................................

........................................... .......................................................................................................
Date ...................................

........................................... .......................................................................................................
Date ...................................

Note: every named author must sign this form, or a copy of it

Signature of Copyright Owner, if different .................................................................................
Date ...................................
Please return this form by post or fax to IM Publications LLP, 6 Charlton Mill,
Charlton, Chichester, West Sussex PO18 0HY, UK. Fax: +44-1243-811711
V
O
L

1
9
,

N
O
.

4
,

P
A
G
E
S

2
1
1

2
8
6

(
2
0
1
1
)

J
O
U
R
N
A
L

O
F

N
E
A
R

I
N
F
R
A
R
E
D

S
P
E
C
T
R
O
S
C
O
P
Y

I
M

P
U
B
L
I
C
A
T
I
O
N
S
volume 19 / number 4 / 2011
ISSN 0967-0335
IN THIS ISSUE:
review: near infrared microscopy and processed animal by-products / the importance of
balanced data sets / determination of azithromycin and decladinosylazithromycin
MlHSoIutions
for PhermeceuticeIs
BOHl prov|des NlP So|ut|ons for Pharmaoeut|oa| lndustry support|ng PAT and QbD approaoh from
warehouse to na| produot re|ease test. BOHls rugged FT-NlP speotrometer, Ohemometr|o software and
NlP g|oba| network experts ensure the h|ghest qua||ty resu|ts.
Mater|a| |dent|cat|on supported by NlPF|ex N-500 w|th a robust F|ber Opt|os and numerous proven
a|gor|thms and oomprehens|ve speotra| ||brar|es.
Automated tab|et ana|ys|s at-||ne for phys|oa| propert|es and oontent un|form|ty test by the NlP
Oheokmaster.
F|na| tab|ets re|ease poss|b|e by NlPF|ex N-500 w|th the ded|oated measur|ng so||d transm|ttanoe oe||.
Easy and powerfu| ca||brat|on oreated by the Oa||brat|on W|zard funot|on |n the NlPOa| Ohemometr|os
software.
Regu|at|on Comp||ance w|th the GMP requ|rements, 21 OFP Part 11 ru|es and |nternat|ona| pharma-
oopoe|as.
Get up to date w|th the |atest teohno|ogy and nd out more on:
www.buch|.com > NIRSo|ut|ons
OueIity in your hends www.buch|.com
BOHl |aborteohn|k AG
9230 F|aw||/Sw|tzer|and
T +41 71 394 63 63
F +41 71 394 65 65
on w
w
w
.buchi.com
/nirw
in
C
o
d
e
: O
O
B
O
7