b i o s y s t e m s e n g i n e e r i n g 1 4 7 ( 2 0 1 6 ) 2 4 8 e2 5 8

Available online at www.sciencedirect.com

ScienceDirect

journal homepage: www.elsevier.com/locate/issn/15375110

Research Paper

Performance evaluation of a model for the

classification of contaminants from wheat using
near-infrared hyperspectral imaging

Lankapalli Ravikanth a, Chandra B. Singh a, Digvir S. Jayas a,*,

Noel D.G. White b
a
Department of Biosystems Engineering, Room E2-376, Engineering, Information and Technology Complex, 75A
Chancellor's Circle, University of Manitoba, Winnipeg, Manitoba R3T 2N2, Canada
b
Cereal Research Centre, AAFC, Department of Biosystems Engineering, Engineering, Information and Technology
Complex, 75A Chancellor's Circle, Winnipeg, Manitoba R3T 2N2, Canada

article info
The presence of contaminants in wheat reduces its quality and thereby its grade. The
Article history: identification of these contaminants in wheat is difficult when they are physically and
Received 16 November 2015 sometimes visually similar. Moreover, manual contaminant identification methods are
Received in revised form time-consuming and labour intensive. Near-infrared (NIR) hyperspectral imaging is an
28 February 2016 advanced image processing technique used effectively for quality evaluation of various
Accepted 4 April 2016 food and agricultural products. This technique can be an effective alternative to the
traditionally used manual contaminant identification methods. This study reports the
performance evaluation of the previously developed classification model to differentiate
Keywords: seven foreign material types (barley, canola, maize, flaxseed, oats, rye, and soybean); six
Wheat dockage types (broken wheat kernels, buckwheat, chaff, wheat spikelets, stones, and wild
Foreign material oats); and two animal excreta types (deer and rabbit droppings) from Canada Western Red
Dockage Spring (CWRS) wheat using NIR hyperspectral imaging. The classification model tested in
Animal excreta this study was developed using standard normal variate (SNV) spectral pre-processing
Near-infrared hyperspectral technique and k-nearest neighbours (k-NN) classifier. Two separate experiments were
imaging conducted to identify and quantify (by number) the amount of contaminant type present
along with wheat. The performance of the classification model was compared with the
model validation results. The results of the developed classification model were very close
to the model validation results and thus this model can be used for the classification of
various contaminants in wheat.
Crown Copyright © 2016 Published by Elsevier Ltd on behalf of IAgrE. All rights reserved.

* Corresponding author. Tel.: þ1 204 474 9404.

E-mail address: Digvir.Jayas@umanitoba.ca (D.S. Jayas).
http://dx.doi.org/10.1016/j.biosystemseng.2016.04.001
1537-5110/Crown Copyright © 2016 Published by Elsevier Ltd on behalf of IAgrE. All rights reserved.
 b i o s y s t e m s e n g i n e e r i n g 1 4 7 ( 2 0 1 6 ) 2 4 8 e2 5 8
1. Introduction
A major portion of the world food supply is obtained directly
or indirectly from plants with significant contribution of ce-
reals and pulses. Cereals are considered the staple food of the
world, supplying 30% of calories in developed countries
(Anonymous, 2003), 60% in developing countries, and more
than 80% in poor countries (Awika, 2011). Among the cereals,
wheat is the most widely cultivated crop in the world with a
production of almost 700 Mt (million tonnes) in 2012. Out of
total wheat produced, one-fifth (almost 150 Mt in 2012) is
exported and the remainder is consumed domestically
(FAOSTAT, 2014). Canada is one of the world's leading pro-
ducers and exporters of high quality wheat. The Canada Grain
Act administered by the Canadian Grain Commission (CGC),
sets standards for the grains consumed within Canada and
exported abroad. The Act specifies permissible limits of con-
taminants in wheat (CGC, 2013).
Visual methods are traditionally used to identify the
presence of contaminants in wheat. These techniques are
time-consuming and labour intensive. Moreover, identifica-
tion of contaminants in wheat is difficult as they are physi-
cally and sometimes visually similar. High production and the
demand for high quality wheat leads research teams around
the globe to develop techniques that can provide rapid and
accurate contaminant identification. In recent years, appli-
cation of image processing for the grain quality monitoring
has increased (Majumdar & Jayas, 2000a, 2000b, 2000c, 2000d;
Ramalingam, Neethirajan, Jayas, & White, 2011; Zhang,
Paliwal, Jayas, & White, 2007). Although colour imaging clas-
sified contaminants from cereal grains with good accuracy
using their image features, there is a need to further improve
on classification accuracy because specified tolerances of
contaminants like foreign material and dockage in wheat are
very low (e.g., less than 0.4% foreign grains in No. 1 grade
Canadian Western Red Spring, CWRS wheat).
Hyperspectral imaging is a specialised technique for
quality evaluation as it utilises the spectral features of sam-
ples in a spatially resolved manner. These spectral and spatial
features represent the sample's chemical and physical prop-
erties, respectively. For grain quality monitoring, hyper-
spectral imaging was used to identify damaged kernels,
predict chemical composition, classify grain types and West-
ern Canada wheat classes (Choudhary, Mahesh, Paliwal, &
Jayas, 2009; Kaliramesh et al., 2013; Mahesh,
Manickavasagan, Jayas, Paliwal, & White, 2008; Singh, Jayas,
Paliwal, & White, 2007, 2009, 2010; Vigneau, Ecarnot, Rabatel,
& Roumet, 2011; Wang, Paliwal, & Jayas, 2004; Xing, Symons,
Hatcher, & Shahin, 2011). Pierna et al. (2012) developed a
technique to detect impurities including other grain types
(spelt, barley, and rapeseed) from cereal bulks using NIR
hyperspectral imaging and a support vector machine (SVM)
classifier. The developed model was able to detect cereals
from the rest of the impurities, but their study was not
directed to study the classification of different foreign mate-
rials and animal excreta types from wheat.
Hyperspectral imaging technique is complex and is
affected by several variations caused by imaging components
and external factors. To reduce these variations spectral pre-
249
processing techniques are generally used. If good results are
expected from the hyperspectral data, a pre-processing step is
of vital importance (Elmasry & Sun, 2010). Miralbes (2004)
determined the quality parameter of wheat flour using NIR
hyperspectral imaging. The NIR transmittance intensities
were corrected using SNV and detrending techniques. Wang
et al. (2004) determined the moisture content of ground
wheat samples using NIR spectroscopy. The NIR absorbance
spectra were pre-processed using multiplicative scatter
correction (MSC) and SNV to correct the spectral scatter. Pre-
processing techniques are used extensively in hyperspectral
imaging to reduce the scatter variations and eliminate the
signal noise (Wesley et al., 2001).
Identification of various contaminants in cereal grains is
an important step to identify the grade of the grain. In authors'
previous work (Ravikanth, Singh, Jayas, & White, 2015), three
experiments were conducted to develop a spectral library for
various contaminant types (foreign material, dockage, and
animal excreta) and wheat. Different combinations of spectral
pre-processing techniques and statistical classifiers were tried
to identify the best combination to classify these contaminant
types from wheat. It was found that when the NIR reflectance
spectra in the 1000e1600 nm wavelength range was pre-
processed with the SNV spectral pre-processing technique
and classified using k-NN classifier gave the maximum accu-
racy in classifying various contaminant types from wheat.
After identifying the best combination of spectral pre-
processing technique and statistical classifier, it is very
important to assess the ability of the developed model to
predict the undefined or unknown data set which was not
used in the model development either as a training set or a
testing set. The present study was conducted to predict an
unknown data set using the previously developed spectral li-
brary for each contaminant type and wheat and the model
(SNV pre-processing technique and k-NN classifier). The per-
formance evaluation of this previously developed model was
done with the objectives:
i. to identify the presence of contaminant types along with
wheat, and
ii. to quantify (by number) the amount of each contaminant
type present along with wheat.
2. Experimental design
This study was divided into two experiments based on the
objectives to accurately identify and count the number of the
non-touching contaminant particles present along with
wheat.
To maintain consistency with the previous study
(Ravikanth et al., 2015), experiments for both objectives were
separately conducted for three sample types: foreign mate-
rials (represented by non-wheat grain types) in wheat,
dockage in wheat, and animal excreta in wheat.
For the first objective, the model was tested initially at the
lowest level of contamination present along with wheat on
the same spatial plane and later the degree of contamination
was increased by adding more contaminant types to the
250 b i o s y s t e m s e n g i n e e r i n g 1 4 7 ( 2 0 1 6 ) 2 4 8 e2 5 8

wheat on the same spatial plane. The acquired hyperspectral

images thus ranged from the least possible degree of
contamination with wheat of one contaminant type to the
largest possible degree of contamination with wheat of all
contamination types. The ‘combination without repetition’
mathematical technique (Alferes, 2012; Cavazzuti, 2012;
Gossett, 2009) was used to design the experiments for each
sample type and is shown in Table 1.
With this experimental design, seven foreign material
types in wheat generated 127 combinations, with the two-way
and eight-way classifications as the lowest and highest degree
of contaminations, respectively. Six dockage types in wheat
generated 63 combinations, with the two-way and seven-way
classifications as the lowest and highest degree of contami-
nations, respectively. Two excreta types in wheat generated 3
combinations, with the two-way and three-way classifica-
tions as the lowest and highest degree of contaminations, Fig. 1 e Wheat kernels arrangement model in the field of
respectively. All the combinations were replicated 10 times to view of the NIR hyperspectral imaging system.
test the robustness of the classification model.
For the second objective of the study, 25 non-touching
provided by Conservation and Research Division of the
contaminant particles or wheat kernels were placed in the
Assiniboine Park Zoo, Winnipeg. The moisture content of all
field of view (FOV) of the hyperspectral imaging system (Fig. 1).
the grain types was measured using the oven method (ASABE,
Quantity analysis experiments were conducted by replacing
2012) and for the entire study their moisture content was
the individual wheat kernels with contaminant particles. For
maintained at 13.0 ± 0.8%, wet basis.
each contaminant type the number wheat kernels replaced
were 0, 1, 2, 3, 4, and 5 which was equal to 0, 4, 8, 12, 16, and 20
percentage contamination (by number), respectively. All the 3.2. Hyperspectral imaging system, calibration and
quantity experiments were replicated 10 times to test the image acquisition
robustness of the classification model.
The imaging system, calibration and image acquisition
method used in this study are described in Ravikanth et al.
3. Materials and methods (2015), Mahesh, Jayas, Paliwal, and White (2011), and Singh
et al. (2010). For better the ready reference and enhanced
3.1. Grain, dockage, and animal excreta samples readability, the main details of the imaging system are pro-
vided in this paper. The imaging system consisted of a long
Three categories of contaminants used in this study: foreign wave NIR InGaAs camera (Model: SU640-1.7RT-D, Sensors
materials, dockage, and animal excreta were the same as in Unlimited Inc., Princeton, NJ) with 640  480 pixels resolution
the previous study (Ravikanth et al., 2015). The clean grain and 960e1700 nm operational range. This camera was com-
types: barley (Canadian Western Malting, six rows), canola plemented with a C-mount lens (Model: L25F1.3, Sofradir EC,
(NEX4-105RR), maize (Pioneer 3995), flaxseed (Canadian Inc., Fairfield, NJ). Double VariSpec liquid crystal tunable fil-
Western, CDC Bethune), oats (AC Pinnacle), rye (mixed vari- ters (LCTFs) (Model: MIR06, Cambridge Research and Instru-
eties), soybean (OAC Ayton) and wheat (CWRS) were supplied mentation Inc., Woburn, MA) with transmission bandwidth
by the Industry Services Division of the CGC, Winnipeg. The of 10 nm and aperture size of 20 mm were used to rapidly
six dockage fractions of wheat (broken wheat kernels, buck- select the imaging wavelengths. A frame grabbing board
wheat, chaff, wheat spikelets, stones, and wild oats) were also (Model: NI PCI-1422, National Instruments Corp., Austin, TX)
supplied by the CGC, Winnipeg. The two animal excreta was used to grab the signal produced by the camera. Two
samples (mule deer droppings and field rabbit droppings) were 300 W tungsten-halogen lamps (Model: Ushio Lighting Inc.,

Table 1 e Experimental design to identify the presence of contaminant type present with wheat.
First degree Second degree nth Degree
a b
F1 W F1 F2 W F2 F3 W F3 F4 W F4 F5 W F5 F6 W … F1 F2 F3 F4 F5..Fn W
F2 W F1 F3 W F2 F4 W F3 F5 W F4 F6 W
F3 W F1 F4 W F2 F5 W F3 F6 W
F4 W F1 F5 W F2 F6 W
F5 W F1 F6 W
Fn W
a
Different contaminant types.
b
Wheat.
 b i o s y s t e m s e n g i n e e r i n g 1 4 7 ( 2 0 1 6 ) 2 4 8 e2 5 8
Cypress, CA) which can emit light in the 400e2500 nm
wavelength range were used to illuminate the samples dur-
ing image acquisition. The imaging system was controlled
through a personal computer (Model: Dell Optiplex GX280,
Dell Inc., Round Rock, TX) with Windows 7 operating system
and Intel duel core processor.
The imaging system was calibrated before every imaging
session to ensure the focal length of the camera was con-
stant, which in turn confirmed that the spatial size of each
pixel remained consistent throughout the study period. The
spatial calibration of the imaging system was performed
before every imaging session using a Canadian five cent coin
(diameter 21.2 mm). As the NIR hyperspectral camera used
in this study had a wavelength-dependent dark current, a
dark image was also acquired before every imaging session
by closing the camera lens with a dark cardboard. The in-
tensity data was converted into relative reflectance using a
99% reflectance white standard plate (Model: Labsphere,
North Sutton, NH). Equations given in Mahesh, Jayas,
Paliwal, and White (2010) were used to calculate the rela-
tive reflectance.
Image acquisition and hyperspectral data storage were
performed using the LabView (Version: 1, National In-
struments, Austin, TX, USA) data acquisition platform installed
on the personal computer. Hypercubes in this entire study were
constructed by acquiring images at every 10 nm interval in
1000e1600 nm range. The camera was set to the centre wave-
length of 1300 nm to minimise geometric distortion.
The spectral library of foreign material, dockage, animal
excreta types and wheat for this paper were taken from the
previous experiments (Ravikanth et al., 2015). For the devel-
opment of spectral signature of each sample type, 300
randomly selected particles of each this sample type were
used. A total of 4800 individual sample particles (300
kernels  16 contaminant types) were imaged to build a
spectral library. The spectral signature of each contaminant
type and wheat were pre-processed using SNV technique and
used as a spectral library for the identification and quantifi-
cation experiments.
For both the identification and quantification experiments,
contaminant particles and wheat kernels were placed in a
non-touching manner on a black cardboard and hypercubes
were built by acquiring 61 images in the 1000e1600 nm
wavelength range at 10 nm interval. For the identification
experiment, a total of 1930 hypercubes (193 combinations  10
replicates) and for the quantification experiment, a total of 960
hypercubes (16 contaminant types  6 contamination
levels  10 replicates) were built.
These three dimensional hypercubes (width, height, and
wavelength) were saved in .spc file format. An algorithm
written in MATLAB (Version: 8.1, Mathworks Inc., Natick,
MA, USA) software was used to read these hypercubes,
remove dead pixels in each image, perform operations like
labelling, thresholding, and background separation on each
image of hypercube. The algorithm converted the intensity
data to reflectance data and arranged them in a two
dimensional matrix (number of particles on the 2D
plane  61 wavelengths) which contains reflectance data
(mean of all pixels covered by each particle) at all the 61
wavelengths.
251
3.3. Spectral pre-processing technique and statistical
classifier
From the previous study the best combination of spectral pre-
processing technique and the statistical classifier for
discriminating the contaminants from wheat were the SNV
and k-NN classifier, respectively. A single MATLAB program
was developed to convert the intensity data of each particle of
the hypercube into reflectance spectra, pre-process the
spectra using the SNV technique, and classify the particles of
the hypercube by using k-NN classifier. The suitable k value
was determined by the technique given by Duda, Peter, and
Stork (2001). During classification, the SNV pre-processed
spectral library of the previous study was used as a training
set and reflectance spectra of each particle of the hypercube
from this study were used as testing set for the k-NN classifier.
The SNV technique removed the offset and multiplicative ef-
fects from each individual spectrum by mean centring and
scaling the spectrum using its own standard deviation
(Barnes, Dhanoa, & Lister, 1989; Fearn, Riccioli, Garrido-Varo,
& Guerrero-Ginel, 2009; Rinnan, Berg, & Engelsen, 2009). This
technique is not only known to perform multiplicative
correction, but also removes the baseline effect and surpasses
the effect of interfering compounds. This technique also has
the capacity to induce the curved structures in the score plot,
which are generally derived from the treated spectra.
Rodriguez-Cobo, Garcia-Allende, Cobo, Lopez-Higuera, and
Conde (2012) suggested that the k-NN classifier would be an
effective classifier for the classification of sample with high
variations like foreign materials from grains.
For the identification experiment, the classification using
k-NN classifier was performed as multi-way classification, but
for the quantification experiment, the classification was per-
formed as two-way classification based on the type of
contaminant in wheat. Each classification was performed in
ten replicates (using 10 hypercubes) and the mean classifica-
tion accuracy and confidence intervals (a ¼ 0.05) were
reported.
4. Results and discussion
4.1. Identification of the presence of contaminant types
along with wheat
4.1.1. Foreign material
The mean classification accuracies and confidence intervals
for the identification of the presence of each foreign material
type individually and in combinations along with wheat are
given in Table 2.
In the first degree of contamination, where wheat was
present along with one foreign material type, except wheat-
maize (WF3, 95.0 ± 9.8%) and wheat-rye (WF6, 90.0 ± 13.1%),
all the foreign material types were perfectly classified
(100.0 ± 0.0%). These results were similar to the model vali-
dation results, where among all the foreign material types,
maize and rye were least accurately classified from wheat.
In the second degree of contamination, where wheat was
present with two foreign material types, misclassifications
252 b i o s y s t e m s e n g i n e e r i n g 1 4 7 ( 2 0 1 6 ) 2 4 8 e2 5 8

Table 2 e Percentage accuracy (average ± confidence interval (a ¼ 0.05), %) for the identification of the presence of foreign
material types with wheat at different degrees of contamination.
First degree contamination
WF1 WF2 WF3 WF4 WF5 WF6 WF7
100.0 ± 0.0 100.0 ± 0.0 95.0 ± 9.8 100.0 ± 0.0 100.0 ± 0.0 90.0 ± 13.1 100.0 ± 0.0
Second degree contamination
WF1F2 WF1F3 WF1F4 WF1F5 WF1F6 WF1F7 WF2F3
100.0 ± 0.0 96.7 ± 6.5 100.0 ± 0.0 100.0 ± 0.0 93.3 ± 8.7 100.0 ± 0.0 100.0 ± 0.0
WF2F4 WF2F5 WF2F6 WF2F7 WF3F4 WF3F5 WF3F6
100.0 ± 0.0 100.0 ± 0.0 100.0 ± 0.0 100.0 ± 0.0 93.3 ± 8.7 93.3 ± 8.7 86.7 ± 14.5
WF3F7 WF4F5 WF4F6 WF4F7 WF5F6 WF5F7 WF6F7
100.0 ± 0.0 93.3 ± 8.7 93.3 ± 8.7 100.0 ± 0.0 100.0 ± 0.0 100.0 ± 0.0 100.0 ± 0.0
Third degree contamination
WF1F2F3 WF1F2F4 WF1F2F5 WF1F2F6 WF1F2F7 WF1F3F4 WF1F3F5
100.0 ± 0.0 100.0 ± 0.0 100.0 ± 0.0 100.0 ± 0.0 100.0 ± 0.0 97.5 ± 4.9 97.5 ± 4.9
WF1F3F6 WF1F3F7 WF1F4F5 WF1F4F6 WF1F4F7 WF1F5F6 WF1F5F7
75.0 ± 0.0 97.5 ± 4.9 100.0 ± 0.0 95.0 ± 6.5 100.0 ± 0.0 95.0 ± 9.8 100.0 ± 0.0
WF1F6F7 WF2F3F4 WF2F3F5 WF2F3F6 WF2F3F7 WF2F4F5 WF2F4F6
95.0 ± 6.5 100.0 ± 0.0 95.0 ± 6.5 75.0 ± 0.0 97.5 ± 4.9 100.0 ± 0.0 95.0 ± 6.5
WF2F4F7 WF2F5F6 WF2F5F7 WF2F6F7 WF3F4F5 WF3F4F6 WF3F4F7
100.0 ± 0.0 97.5 ± 4.7 100.0 ± 0.0 95.0 ± 6.5 100.0 ± 0.0 97.5 ± 4.9 100.0 ± 0.0
WF3F5F6 WF3F5F7 WF3F6F7 WF4F5F6 WF4F5F7 WF4F6F7 WF5F6F7
92.5 ± 7.5 100.0 ± 0.0 82.5 ± 7.5 100.0 ± 0.0 100.0 ± 0.0 100.0 ± 0.0 100.0 ± 0.0
Fourth degree contamination
WF1F2F3F4 WF1F2F3F5 WF1F2F3F6 WF1F2F3F7 WF1F2F4F5 WF1F2F4F6 WF1F2F4F7
100.0 ± 0.0 100.0 ± 0.0 96.0 ± 7.8 98.0 ± 3.9 100.0 ± 0.0 98.0 ± 3.0 100.0 ± 0.0
WF1F2F5F6 WF1F2F5F7 WF1F2F6F7 WF1F3F4F5 WF1F3F4F6 WF1F3F4F7 WF1F3F5F6
88.0 ± 10.5 98.03.9 92.0 ± 6.4 96.0 ± 7.8 96.0 ± 5.2 100.0 ± 0.0 80.0 ± 0.0
WF1F3F5F7 WF1F3F6F7 WF1F4F5F6 WF1F4F5F7 WF1F4F6F7 WF1F5F6F7 WF2F3F4F5
98.0 ± 3.9 96.0 ± 7.8 96.0 ± 7.8 100.0 ± 0.0 100.0 ± 0.0 96.0 ± 5.2 84.0 ± 12.8
WF2F3F4F6 WF2F3F4F7 WF2F3F5F6 WF2F3F5F7 WF2F3F6F7 WF2F4F5F6 WF2F4F5F7
94.0 ± 8.4 96.0 ± 5.2 96.0 ± 5.2 100.0 ± 0.0 88.0 ± 12.0 80.0 ± 13.1 100.0 ± 0.0
WF2F4F6F7 WF2F5F6F7 WF3F4F5F6 WF3F4F5F7 WF3F4F6F7 WF3F5F6F7 WF4F5F6F7
80.0 ± 0.0 96.0 ± 5.2 80.0 ± 0.0 96.0 ± 5.2 86.0 ± 6.0 92.0 ± 6.4 96.0 ± 7.8
Fifth degree contamination
WF1F2F3F4F5 WF1F2F3F4F6 WF1F2F3F4F7 WF1F2F3F5F6 WF1F2F3F5F7 WF1F2F3F6F7 WF1F2F4F5F6
98.3 ± 3.3 98.3 ± 3.3 98.3 ± 3.3 85.0 ± 7.6 95.0 ± 5.0 98.3 ± 3.3 95.0 ± 5.0
WF1F2F4F5F7 WF1F2F4F6F7 WF1F2F5F6F7 WF1F3F4F5F6 WF1F3F4F5F7 WF1F3F4F6F7 WF1F3F5F6F7
96.7 ± 4.4 95.0 ± 5.0 96.7 ± 4.4 93.3 ± 7.2 96.7 ± 6.5 85.0 ± 9.0 85.0 ± 5.9
WF1F4F5F6F7 WF2F3F4F5F6 WF2F3F4F5F7 WF2F3F4F6F7 WF2F3F5F6F7 WF2F4F5F6F7 WF3F4F5F6F7
95.0 ± 7.0 93.3 ± 7.2 95.0 ± 7.0 93.3 ± 7.2 95.0 ± 7.0 96.7 ± 6.5 85.0 ± 3.3
Sixth degree contamination
WF1F2F3F4F5F6 WF1F2F3F4F5F7 WF1F2F3F4F6F7 WF1F2F3F5F6F7 WF1F2F4F5F6F7 WF1F3F4F5F6F7 WF2F3F4F5F6F7
87.1 ± 6.5 97.1 ± 3.7 97.3 ± 3.5 91.8 ± 7.8 95.7 ± 4.3 87.1 ± 6.5 95.7 ± 4.3
Seventh degree contamination
WF1F2F3F4F5F6F7
92.5 ± 6.5

W e Wheat, F1 e Barley, F2 e Canola, F3 e Maize, F4 e Flax, F5 e Oats, F6 e Rye, and F7 e Soybean.

were identified between maize-oats (F3F5), maizeerye (F3F6),
flaxeoats (F4F5), and flaxerye (F4F6). The overall accuracy of
the second degree of contamination was 97.6 ± 1.6%. The
overall classification accuracies further reduced as the com-
bination of wheat with foreign material types (degree of
contamination) increased. The seventh degree of contami-
nation where wheat was present with all the seven foreign
material types the classification accuracy was 92.5 ± 6.5%.
The classification accuracies of the first and seventh degree
contamination of the foreign material identification experi-
ment were compared with the two-way and multi-way model
validation classification results (Ravikanth et al., 2015) and are
shown in Fig. 2. The first degree contamination identification
experiment result (97.9 ± 2.9%) of the foreign material was
close to the two-way model validation result (99.9 ± 0.1%). But
the seventh degree contamination identification experiment
result (92.5 ± 6.5%) of the foreign material was lower than the
multi-way model validation result (98.3 ± 0.2%).
4.1.2. Dockage
The mean classification accuracies and confidence intervals for
the identification of the presence of each dockage type individ-
ually and in combinations with wheat are given in Table 3.
In the first degree of contamination, where wheat was
present with one dockage type, except wheat-chaff (WF3,
95.0 ± 9.8%) and wheatewheat spikelet (WF5, 95.0 ± 9.8%), all
the dockage types were perfectly (100.0 ± 0.0%) classified.
These results were similar to the model validation results,
 b i o s y s t e m s e n g i n e e r i n g 1 4 7 ( 2 0 1 6 ) 2 4 8 e2 5 8 253

accuracy of the second degree of contamination was

Fig. 2 e Comparison of the identification experiment
results from the model validation results for the foreign
material types.
where among all the dockage types chaff and wheat spikelets
were least accurately classified.
In the second degree of contamination, where wheat was
present with two dockage types, misclassifications were
identified between chaff-wheat spikelets (F3F5), chaff-wild
oats (F3F6), and wheat spikelets-wild oats (F5F6). The overall
98.0 ± 1.8%. The overall classification accuracies further
reduced as the combination of wheat and dockage types (de-
gree of contamination) increased. The sixth degree of
contamination where wheat was present with all the dockage
types the classification accuracy was 94.3 ± 6.2%.
The classification accuracies of the first and the sixth de-
gree contamination of the dockage type identification exper-
iment were compared with two-way and multi-way model
validation classification results (Ravikanth et al., 2015) and are
shown in Fig. 3. The first degree contamination identification
experiment result (98.3 ± 2.1%) for dockage was close to the
two-way model validation result (100.0 ± 0.0%). But the sixth
degree contamination identification experiment result
(94.3 ± 6.2%) of dockage was lower than the multi-way model
validation result (98.8 ± 0.2%).
4.1.3. Animal excreta
The mean classification accuracies and confidence intervals
for the identification of the presence of each animal excreta
(fecal) type individually and in combinations with wheat are
given in Table 4. In the both first and second degree of
contamination, both deer droppings and rabbit droppings
were perfectly (100.0 ± 0.0%) classified. These results were
similar to the model validation results, where both the animal
excreta types were perfectly (100.0 ± 0.0%) classified.

Table 3 e Percentage accuracy (average ± confidence interval (a ¼ 0.05), %) for the identification of the presence of dockage
types with wheat at different degrees of contamination.
First degree contamination
WF1 WF2 WF3 WF4 WF5 WF6
100.0 ± 0.0 100.0 ± 0.0 95.0 ± 9.8 100.0 ± 0.0 95.0 ± 9.8 100.0 ± 0.0
Second degree contamination
WF1F2 WF1F3 WF1F4 WF1F5 WF1F6
100.0 ± 0.0 100.0 ± 0.0 100.0 ± 0.0 96.7 ± 6.5 100.0 ± 0.0
WF2F3 WF2F4 WF2F5 WF2F6 WF3F4
100.0 ± 0.0 100.0 ± 0.0 100.0 ± 0.0 100.0 ± 0.0 100.0 ± 0.0
WF3F5 WF3F6 WF4F5 WF4F6 WF5F6
90.0 ± 10.0 90.0 ± 10.0 100.0 ± 0.0 100.0 ± 0.0 93.3 ± 8.7
Third degree contamination
WF1F2F3 WF1F2F4 WF1F2F5 WF1F2F6 WF1F3F4
100.0 ± 0.0 100.0 ± 0.0 100.0 ± 0.0 100.0 ± 0.0 100.0 ± 0.0
WF1F3F5 WF1F3F6 WF1F4F5 WF1F4F6 WF1F5F6
95.0 ± 6.5 97.5 ± 4.9 100.0 ± 0.0 100.0 ± 0.0 100.0 ± 0.0
WF2F3F4 WF2F3F5 WF2F3F6 WF2F4F5 WF2F4F6
100.0 ± 0.0 90.0 ± 8.0 90.0 ± 8.0 100.0 ± 0.0 100.0 ± 0.0
WF2F5F6 WF3F4F5 WF3F4F6 WF3F5F6 WF4F5F6
90.0 ± 13.1 95.0 ± 9.8 95.0 ± 6.5 77.5 ± 8.8 95.0 ± 6.5
Fourth degree contamination
WF1F2F3F4 WF1F2F3F5 WF1F2F3F6 WF1F2F4F5 WF1F2F4F6
100.0 ± 0.0 98.0 ± 3.9 100.0 ± 0.0 100.0 ± 0.0 100.0 ± 0.0
WF1F2F5F6 WF1F3F4F5 WF1F3F4F6 WF1F3F5F6 WF1F4F5F6
90.0 ± 6.5 94.0 ± 8.4 90.0 ± 6.5 96.0 ± 7.8 100.0 ± 0.0
WF2F3F4F5 WF2F3F4F6 WF2F3F5F6 WF2F4F5F6 WF3F4F5F6
96.0 ± 5.2 100.0 ± 0.0 78.0 ± 12.3 90.0 ± 6.5 100.0 ± 0.0
Fifth degree contamination
WF1F2F3F4F5 WF1F2F3F4F6 WF1F2F3F5F6 WF1F2F4F5F6 WF1F3F4F5F6 WF2F3F4F5F6
96.7 ± 4.4 98.3 ± 3.3 85.0 ± 7.6 98.3 ± 3.3 95.0 ± 5.0 96.7 ± 4.4
Sixth degree contamination
WF1F2F3F4F5F6
94.3 ± 6.2

W e Wheat, F1 e Broken kernels, F2 e Buckwheat, F3 e Chaff, F4 e Stones, F5 e Wheat spikelets, and F6 e Wild oats.
254 b i o s y s t e m s e n g i n e e r i n g 1 4 7 ( 2 0 1 6 ) 2 4 8 e2 5 8
Fig. 3 e Comparison of the identification experiment
results from the model validation results for the dockage
types (first degree contamination e only one foreign
material type present along with wheat on the same plane;
seventh degree contamination e all the foreign material
types present along with wheat on the same spatial plane).
Table 4 e Percentage accuracy (average ± confidence
interval (a ¼ 0.05), %) for the identification of the presence
of animal excreta types with wheat at different degrees of
contamination.
First degree contamination
WF1 WF2
100.0 ± 0.0 100.0 ± 0.0
Second degree contamination
WF1F2
100.0 ± 0.0
W e Wheat, F1 e Deer droppings, and F2 e Rabbit droppings.
The classification accuracies of the first and the second
degree contamination of the animal excreta type identifica-
tion experiment were compared with two-way and multi-way
model validation classification results (Ravikanth et al., 2015)
and are shown in Fig. 4. The first and second degree
contamination identification experiment result (100.0 ± 0.0%)
Fig. 4 e Comparison of the identification experiment
results from the model validation results for the animal
excreta types.
for animal excreta was similar to the two-way and multi-way
model validation result (100.0 ± 0.0%).
Spectral properties of the hypercubes were identified to be
working well to accurately classify the foreign materials and
dockage at lower degrees of contamination and animal
excreta at all the degree of contamination. The identification
of foreign grains and dockage was possible if the degree of
contamination was lower. If the contamination degree in-
creases the identification becomes difficult. Use of image
properties of the hyperspectral images can be tried in the
future research to understand if the identification can be
improved at the lower degree of contamination of foreign
materials and dockage.
4.2. Quantification (by number) of each contaminant
type present with wheat
4.2.1. Foreign material
The mean classification accuracies and confidence intervals for
the quantification (by number) of the amount of each foreign
material type present with wheat in a non-touching manner are
given in Table 5. Among all the foreign material types, canola was
perfectly classified (100.0 ± 0.0%) at all the levels of contamina-
tion. This was because the composition of canola kernels being
an oilseed, were completely different from wheat which is a
cereal. The flax was almost perfectly classified (>99.2%) at all the
levels of contamination. The least classification accuracy was
observed for quantifying rye at 16% contamination (93.2 ± 2.9%)
and 12% contamination (93.6 ± 2.1%). Certain chemical proper-
ties of rye were much closer to the wheat than that of barley, like
available carbohydrates and crude fat content (FAO, 2015). This
might be the reason for better results were obtained for the
classification of barley than rye. The quantification results of rye
and barley were in similarity to the classification results of the
model development (Ravikanth et al., 2015).
The overall quantification accuracy of each foreign mate-
rial type was calculated as the average accuracy at all the
levels of contamination. This overall quantification accuracy
was compared with the two-way model validation results
(Ravikanth et al., 2015) and is shown in Fig. 5. The quantifi-
cation results of canola, flax, oats, soybean, and barley were
very close (less than 2.8% variation) to the model validation
results. The classification model gave variation of less than
3.6% in the quantification experiment from the model vali-
dation results for maize and rye.
4.2.2. Dockage
The mean classification accuracies and confidence intervals
for the quantification (by number) of the amount of each
dockage type present with wheat in a non-touching manner
are given in Table 6. Among all the dockage types, stones were
perfectly classified (100.0 ± 0.0%) at all levels of contamination
of wheat. Story and Raghavan (1973) developed online sorting
equipment for the separation of stones and soil clods from
potatoes. Gruna, Vieth, Michelsburg, and Leo  n (2010) used Vis-
NIR hyperspectral imaging system for classification of foreign
materials (wood splitter, paper scraps, cork, cord, stones, and
plastic pieces) from dried onions. The Linear discriminant
analysis (LDA) and Quadratic discriminant analysis (QDA)
statistical classifiers were used for the classification of
 b i o s y s t e m s e n g i n e e r i n g 1 4 7 ( 2 0 1 6 ) 2 4 8 e2 5 8 255

Table 5 e Percentage accuracy (average ± confidence interval (a ¼ 0.05), %) for the quantification (by number) of the
presence of each foreign material type with wheat at different degrees of contamination.
No. of wheat kernels replaced with contaminant particles 0 (0%) 1 (4%) 2 (8%) 3 (12%) 4 (16%) 5 (20%)
(percentage of contamination)
Barley 100.0 ± 0.0 100.0 ± 0.0 96.8 ± 1.1 95.6 ± 1.9 95.2 ± 2.5 95.6 ± 1.9
Canola 100.0 ± 0.0 100.0 ± 0.0 100.0 ± 0.0 100.0 ± 0.0 100.0 ± 0.0 100.0 ± 0.0
Maize 97.6 ± 1.3 97.6 ± 1.3 96.0 ± 1.7 94.4 ± 1.3 95.2 ± 3.8 95.2 ± 3.3
Flax 100.0 ± 0.0 100.0 ± 0.0 100.0 ± 0.0 100.0 ± 0.0 99.2 ± 0.3 100.0 ± 0.0
Oats 100.0 ± 0.0 100.0 ± 0.0 100.0 ± 0.0 99.6 ± 0.2 99.2 ± 0.3 97.2 ± 0.4
Rye 99.6 ± 0.8 97.6 ± 2.4 94.4 ± 2.1 93.6 ± 2.1 93.2 ± 2.9 94.0 ± 3.5
Soybean 100.0 ± 0.0 100.0 ± 0.0 97.6 ± 0.3 97.2 ± 0.3 97.6 ± 0.4 96.8 ± 0.4

response, where even a very low intensity of light was re-

Fig. 5 e Comparison of the classification accuracy of each
foreign grain type in the quantification (by number)
experiment from the model validation two-way
classification results.
Bhattacharyya distance processed spectra (Bhattacharyya,
1946). The QDA classifier gave better results than the LDA in
the characterisation of foreign materials from dried onions.
The LDA failed to detect the wood splitters, but both the QDA
and LDA accurately detected stones from dried onions. The
accurate classification of stones from died onions can be
compared with the classification results of the stones from
wheat. Gowen and O'Donnell (2013) demonstrated the che-
mometric analysis for identification of foreign bodies (plastic
shards, glass beads, and rubber fragments) from grain kernels
(white rice grains and a mixed variety of grains including
dehydrated legumes) using a pushbroom hyperspectral im-
aging system operated in the NIR range (950e1700 nm). The
results suggested that the glass beads had a very low spectral
flected from their surface. The plastic, rubber, and the grain
samples were identified to have a characteristic strong spec-
tral response. Similarly, in the present study the use of
hyperspectral imaging system for the classification of non-
biological foreign material like stone was proven to be
appropriate. Among various dockage types the stones being
non-biological were accurately classified and identified.
The broken kernels and buckwheat were almost perfectly
classified (>99.5%) at all the levels of contamination. The least
classification accuracy was observed for quantifying chaff at
20% contamination (96.0 ± 2.4%) and wheat spikelets at 16%
contamination (96.4 ± 1.9%) of wheat.
The overall quantification accuracy of each dockage type
was calculated as the average accuracy at all the levels of
contamination. This overall quantification accuracy was
compared with the two-way model validation results
(Ravikanth et al., 2015) and is shown in Fig. 6. The quantifi-
cation results of all the dockage types were very close (less
than 1.6% variation) to the model validation results.
For both the foreign material and dockage the availability
of more contaminant particles in the higher percentage of
contamination gave higher chances for the misclassification
of the contaminant particles with wheat. Hence a general
observation was made for both foreign material and dockage
types that the classification accuracy was reduced with the
increase in percentage of contamination.
4.2.3. Animal excreta
The mean classification accuracies and confidence intervals
for the quantification (by number) of the amount of each an-
imal excreta type present with wheat in a non-touching
manner showed that both deer and rabbit droppings were

Table 6 e Percentage accuracy (average ± confidence interval (a ¼ 0.05), %) for the quantification (by number) of the
presence of each dockage type with wheat at different degrees of contamination.
No. of wheat kernels replaced with contaminant particles 0 (0%) 1 (4%) 2 (8%) 3 (12%) 4 (16%) 5 (20%)
(percentage of contamination)
Broken kernels 100.0 ± 0.0 100.0 ± 0.0 100.0 ± 0.0 98.0 ± 1.3 98.4 ± 1.3 98.4 ± 1.3
Buckwheat 100.0 ± 0.0 100.0 ± 0.0 100.0 ± 0.0 99.6 ± 0.8 98.4 ± 1.7 99.2 ± 1.6
Chaff 100.0 ± 0.0 100.0 ± 0.0 100.0 ± 0.0 97.2 ± 1.2 96.8 ± 1.6 96.0 ± 2.4
Stones 100.0 ± 0.0 100.0 ± 0.0 100.0 ± 0.0 100.0 ± 0.0 100.0 ± 0.0 100.0 ± 0.0
Wheat spikelets 100.0 ± 0.0 99.2 ± 1.1 98.0 ± 1.3 97.2 ± 1.2 96.4 ± 1.9 98.4 ± 1.3
Wild oats 100.0 ± 0.0 100.0 ± 0.0 100.0 ± 0.0 98.8 ± 1.2 97.6 ± 1.3 96.8 ± 1.1
256 b i o s y s t e m s e n g i n e e r i n g 1 4 7 ( 2 0 1 6 ) 2 4 8 e2 5 8
Fig. 6 e Comparison of the classification accuracy of each
dockage type in the quantification (by number) experiment
from the model validation two-way classification results.
perfectly classified (100.0 ± 0.0%) at all the levels of contami-
nation of wheat.
The overall quantification accuracy of each animal excreta
type was calculated as the average accuracy at all the levels of
contamination. This overall quantification accuracy was
compared with the two-way model validation results
(Ravikanth et al., 2015) and is shown in Fig. 7. Both deer and
rabbit droppings quantification results were exactly similar
(100.0 ± 0.0%) to the model validation results. It was clear that
the composition of animal excreta was very different from the
plant products like wheat kernels.
Similar studies were conducted in the past to identify the
ability of hyperspectral imaging to classify various impurities
and classes from cereal grains. Pierna et al. (2012) detected
various impurities in cereal grains and nematode cysts on
sugar beet plant roots using hyperspectral imaging, SNV and
second derivative Savitzk Golay (SG) pre-processing tech-
niques and SVM classifier. The accuracy of the analysis was
about 92% and 84% for the identification of impurities in cereal
grains and nematode cysts, respectively. Singh et al. (2010)
Fig. 7 e Comparison of the classification accuracy of each
animal excreta type in the quantification (by number)
experiment from the model validation two-way
classification results.
used a combination of colour and NIR hyperspectral imaging
to identify damaged wheat kernels and obtained greater than
99% accuracy in classifying the midge damaged wheat ker-
nels. Similarly, Mahesh et al. (2008) and Choudhary et al.
(2009) used different statistical classifiers to classified West-
ern Canadian wheat classes using NIR hyperspectral imaging.
They were able to obtain up to 99% accuracy for classifying
different wheat classes.
It was clearly observed that the classification accuracies
of certain contaminant types like canola stones, and animal
excreta were higher than the classification accuracy of other
contaminant types at certain levels; this variation was
caused due to the chemical composition of the samples. The
canola, stones and animal excreta are of completely
different chemical composition from wheat, but the other
cereal grain types, or the dockage originated from the wheat
were difficult to classify from wheat at various levels of
contamination.
4.3. Perspectives for future developments
This current study relies only on the spectral information of
the hyperspectral images for classifying various contaminant
particles from wheat kernels based on number of particles.
This number basis can be readily used for developing grain
cleaning system, where accurate identification of impurities
before and after cleaning is of prime importance and opti-
misation can be done using this information. Although an
effective grain cleaning unit can be developed from this
study, identifying grain grades is a much difficult objective.
An automatic grain grade identification system can reduce
time and efforts of the grain graders and improve their de-
cision making ability. Expanding the present number basis
quantification to mass basis quantification, further work with
respect to the development of correlation between the
contaminant and grain dimensions and mass is required. One
way of achieving this would be using spatial properties of the
hyperspectral images. Multiple imaging systems can be used
for acquiring images of contaminant particles and wheat
kernels in two dimensions (area on horizontal plane and
height in the vertical plane) to determine the volume of the
particles. Later this volume can be correlated with the weight
of individual particles. To reduce overall variance, it is
important to have a larger data size for developing such
correlations. In this kind of analysis, the accuracy of deter-
mining the quantity on mass basis using hyperspectral im-
aging primarily depends on the accuracy of correlation
between volume and mass of the kernels. Although, this
entire system would be much harder to build, it is still
achievable.
5. Conclusion
The present study was conducted to test the performance of
the previously developed classification model (SNV spectral
pre-processing technique and k-NN classifier) using NIR
hyperspectral imaging to differentiate seven foreign material
types (barley, canola, maize, flaxseed, oats, rye, and soybean);
six dockage types (broken wheat kernels, buckwheat, chaff,
 b i o s y s t e m s e n g i n e e r i n g 1 4 7 ( 2 0 1 6 ) 2 4 8 e2 5 8
wheat spikelets, stones, and wild oats); and two animal
excreta types (deer and rabbit droppings) from CWRS wheat.
The overall classification accuracies to identify foreign ma-
terial types present with wheat during first and seventh de-
gree of contamination were 97.6 ± 1.6% and 92.5 ± 6.5%,
respectively. The overall classification accuracies to identify
dockage types present with wheat during first and sixth de-
gree of contamination were 98.0 ± 1.8% and 94.3 ± 6.2%,
respectively. The overall classification accuracy to identify
animal excreta type present with wheat during both first and
second degree of contamination was 100.0 ± 0.0%. For both
foreign materials and dockage, the identification experiment
results were close to the model validation results at the lower
degree of contamination, but the identification experiment
results were slightly lower than the model validation results
at the higher degree of contaminations. The identification
experiment results were similar to model validation results at
all the levels of contamination for animal excreta types.
Spectral properties of the hyperspectral images were good
enough to identify the foreign materials and dockage at lower
degree contamination levels and animal excreta at all the
degree of contaminations. Use of image properties of hyper-
spectral images should be considered for the identification of
higher degree of contamination of foreign materials and
dockage from wheat.
Among all the foreign material types, canola was perfectly
quantified (by number) (100.0 ± 0.0%) from wheat at all the
percentages of contamination. The least classification accu-
racy was for quantifying rye at 16% contamination
(93.2 ± 2.9%) and 12% contamination (93.6 ± 2.1%). Among all
the dockage types, the stones were perfectly quantified (by
number) (100.0 ± 0.0%) from wheat at all the percentages of
contamination. The least classification accuracy was for
quantifying chaff at 20% contamination (96.0 ± 2.4%) and
wheat spikelets at 16% contamination (96.4 ± 1.9%). Both the
animal excreta types, deer and rabbit droppings were
perfectly quantified (100.0 ± 0.0%) at all the levels of
contamination in wheat. The quantification experiment re-
sults were very close to the two-way model validation results
for dockage and animal excreta types. Variation of chemical
composition of different contaminant types from wheat was
the primary reason for the classification of certain contami-
nants classified better than the others from wheat. Therefore,
the results from this study prove that the classification model
using SNV spectral pre-processing technique and k-NN clas-
sifier can accurately identify and quantify various contami-
nants from wheat.
Acknowledgement
The authors thank the Canadian Grain Commission for
providing the samples and the University of Manitoba Grad-
uate Fellowship, the Graduate Enhancement of Tri-Council
Stipends program and the Natural Sciences and Engineering
Research Council of Canada for funding this study. Authors
also thank Canada Foundation for Innovation, Manitoba
Research Innovation Fund, and several other partners for
creating research infrastructure.
257
references
Alferes, V. R. (2012). Methods of randomization in experimental design.
Thousand Oaks, CA, USA: SAGE Publications, Inc.
Anonymous. (2003). Diet, nutrition and the prevention of chronic
diseases. In WHO technical report series, 916 (pp. 1e150). Geneva.
ASABE. (2012). ASAE S352.2: Moisture measurement-unground grain
and seeds. St. Joseph, Michigan: ASAE.
Awika, J. M. (2011). Major cereal grains production and use around
the world. In J. M. Awika, V. Piironen, & S. Bean (Eds.), Advances
in cereal science: Implications to food processing and health
promotion (pp. 1e13). Washington, DC: American Chemical
Society Symposium Series.
Barnes, R. J., Dhanoa, M. S., & Lister, S. J. (1989). Standard normal
variate transformation and de-trending of near-infrared
diffuse reflectance spectra. Applied Spectroscopy, 43(5),
772e777.
Bhattacharyya, A. (1946). On a measure of divergence between
two multinomial populations. Sankhya : The Indian Journal of
Statistics, 7(4), 401e406.
Cavazzuti, M. (2012). Optimization methods: From theory to design
scientific and technological aspects in mechanics. New York, NY,
USA: Springer Science & Business Media.
CGC. (2013). Official grain grading guide-wheat: August 1, 2013.
Canadian Grain Commission Industry Services. Canadian
Grain Commission. Available from https://www.grainscanada.
gc.ca/oggg-gocg/2013/04-wheat-2013-eng.pdf Accessed
11.11.13.
Choudhary, R., Mahesh, S., Paliwal, J., & Jayas, D. S. (2009).
Identification of wheat classes using wavelet features from
near infrared hyperspectral images of bulk samples.
Biosystems Engineering, 102(2), 115e127.
Duda, R. O., Peter, E. H., & Stork, D. G. (2001). Nonmetric methods.
In Pattern classification (2nd ed.). New York City, NY: Wiley-
Interscience.
Elmasry, G., & Sun, D. W. (2010). Principles of hyperspectral
imaging technology. In D. W. Sun (Ed.), Hyperspectral imaging
for food quality analysis and control (pp. 3e43). London:
Academic Press.
FAO. (2015). Major chemical components of cereal grains.
Available from http://www.fao.org/docrep/x2184e/x2184e04.
htm Accessed 07.10.15.
FAOSTAT. (2014). Food and agricultural commodities production. Food
and Agricultural Organization of the United Nations. Available
from http://faostat.fao.org/site/339/default.aspx Accessed
07.04.14.
Fearn, T., Riccioli, C., Garrido-Varo, A., & Guerrero-Ginel, J.-E.
(2009). On the geometry of SNV and MSC. Chemometrics and
Intelligent Laboratory Systems, 96(1), 22e26.
Gossett, E. (2009). Discrete mathematics with proof. Hoboken, NJ,
USA: John Wiley and Sons.
Gowen, A. A., & O'Donnell, C. P. (2013). Near infrared
hyperspectral imaging for foreign body detection and
identification in food processing. Spectroscopy Europe, 25(6),
6e11.
Gruna, R., Vieth, K.-U., Michelsburg, M., & Leo  n, F. P. (2010).
Hyperspectral imaging e from laboratory to in-line food
sorting. In CIGR workshop on image analysis in agriculture.
Budapest, Hungary: CIGR. August 26e27, 2010.
Kaliramesh, S., Chelladurai, V., Jayas, D. S., Alagusundaram, K.,
White, N. D. G., & Fields, P. G. (2013). Detection of infestation
by Callosobruchus maculatus in mungbean using near-infrared
hyperspectral imaging. Journal of Stored Products Research, 52,
107e111.
Mahesh, S., Jayas, D. S., Paliwal, J., & White, N. D. G. (2010). Near-
infrared (NIR) hyperspectral imaging e an emerging analytical
tool for classification of western Canadian wheat classes from
258 b i o s y s t e m s e n g i n e e r i n g 1 4 7 ( 2 0 1 6 ) 2 4 8 e2 5 8
different locations and crop years. In 2010 and North Central
CSBE/ASABE inter-sectional conference ASABE conference paper. St.
Joseph, Michigan, M.I.: ASABE (Paper No. MB-SK 10-302).
Saskatoon, Saskatchewan. October 7e9, 2010.
Mahesh, S., Jayas, D. S., Paliwal, J., & White, N. D. G. (2011).
Identification of wheat classes at different moisture levels
using near-infrared hyperspectral images of bulk samples.
Sensing and Instrumentation for Food Quality and Safety, 5(1),
1e9.
Mahesh, S., Manickavasagan, A., Jayas, D. S., Paliwal, J., &
White, N. D. G. (2008). Feasibility of near-infrared
hyperspectral imaging to differentiate Canadian wheat
classes. Biosystems Engineering, 101(1), 50e57.
Majumdar, S., & Jayas, D. S. (2000a). Classification of cereal grains
using machine vision: I. Morphology models. Transactions of
the ASAE, 43(6), 1669e1675.
Majumdar, S., & Jayas, D. S. (2000b). Classification of cereal grains
using machine vision: II. Colour models. Transactions of the
ASAE, 43(6), 1677e1680.
Majumdar, S., & Jayas, D. S. (2000c). Classification of cereal grains
using machine vision: III. Texture models. Transactions of the
ASAE, 43(6), 1681e1687.
Majumdar, S., & Jayas, D. S. (2000d). Classification of cereal grains
using machine vision: IV. Combined morphology, colour, and
texture models. Transactions of the ASAE, 43(6), 1689e1694.
Miralbes, C. (2004). Quality control in the milling industry using
near infrared transmittance spectroscopy. Food Chemistry,
88(4), 621e628.
Pierna, J. A. F., Vermeulen, P., Amand, O., Tossens, A.,
Dardenne, P., & Baeten, V. (2012). NIR hyperspectral imaging
spectroscopy and chemometrics for the detection of
undesirable substances in food and feed. Chemometrics and
Intelligent Laboratory Systems, 117, 233e239.
Ramalingam, G., Neethirajan, S., Jayas, D. S., & White, N. D. G.
(2011). Characterization of the influence of moisture content
on single wheat kernels using machine vision. Applied
Engineering in Agriculture, 27(3), 403e409.
Ravikanth, L., Singh, C. B., Jayas, D. S., & White, N. D. G. (2015).
Classification of contaminants from wheat using near-
infrared hyperspectral imaging. Biosystems Engineering, 135,
73e86.
Rinnan, A., Berg, F., & Engelsen, S. B. (2009). Review of the most
common pre-processing techniques for near-infrared spectra.
Trends in Analytical Chemistry, 28(10), 1201e1222.
Rodriguez-Cobo, L., Garcia-Allende, P. B., Cobo, A., Lopez-
Higuera, J. M., & Conde, O. M. (2012). Raw material classification
by means of hyperspectral imaging and hierarchical temporal
memories. IEEE Sensors Journal, 12(9), 2767e2775.
Singh, C. B., Jayas, D. S., Paliwal, J., & White, N. D. G. (2007). Fungal
detection in wheat using near-infrared hyperspectral imaging.
Transactions of the ASABE, 50(6), 2171e2176.
Singh, C. B., Jayas, D. S., Paliwal, J., & White, N. D. G. (2009).
Detection of insect-damaged wheat kernels using near-
infrared hyperspectral imaging. Journal of Stored Products
Research, 45(3), 151e158.
Singh, C. B., Jayas, D. S., Paliwal, J., & White, N. D. G. (2010).
Detection of midge-damaged wheat kernels using short-wave
near-infrared hyperspectral and digital colour imaging.
Biosystems Engineering, 105(3), 380e387.
Story, A. G., & Raghavan, G. S. V. (1973). Sorting potatoes from
stones and soil clods by infrared reflectance. Transactions of the
ASAE, 16(2), 0304e0309.
Vigneau, N., Ecarnot, M., Rabatel, G., & Roumet, P. (2011). Potential
of field hyperspectral imaging as a non destructive method to
assess leaf nitrogen content in wheat. Field Crops Research,
122(1), 25e31.
Wang, W., Paliwal, J., & Jayas, D. S. (2004). Determination of moisture
content of ground wheat using near-infrared spectroscopy (Paper
no. MB04-200). St. Joseph: ASABE.
Wesley, I. J., Larroque, O., Osborne, B. G., Azudin, N., Allen, H., &
Skerritt, J. H. (2001). Measurement of gliadin and glutenin
content of flour by NIR spectroscopy. Journal of Cereal Science,
34(2), 125e133.
Xing, J., Symons, S., Hatcher, D., & Shahin, M. (2011). Comparison of
short-wavelength infrared (SWIR) hyperspectral imaging
system with an FT-NIR spectrophotometer for predicting alpha-
amylase activities in individual Canadian Western Red Spring
(CWRS) wheat kernels. Biosystems Engineering, 108(4), 303e310.
Zhang, H., Paliwal, J., Jayas, D. S., & White, N. D. G. (2007).
Classification of fungal infected wheat kernels using near-
infrared reflectance hyperspectral imaging and support vector
machine. Transactions of the ASABE, 50(5), 1779e1785.