Computers and Electronics in Agriculture 75 (2011) 107–112

Contents lists available at ScienceDirect

Computers and Electronics in Agriculture

journal homepage: www.elsevier.com/locate/compag

Original paper

Detection of Fusarium damaged kernels in Canada Western Red Spring wheat

using visible/near-infrared hyperspectral imaging and principal
component analysis
Muhammad A. Shahin ∗ , Stephen J. Symons
Grain Research Laboratory, Canadian Grain Commission, 1404-303 Main Street, Winnipeg, MB, Canada R3C 3G8

a r t i c l e i n f o a b s t r a c t

Article history: Fusarium damage in wheat reduces the quality and safety of food and feed products. In this study, the
Received 26 April 2010 use of hyperspectral imaging was investigated to detect fusarium damaged kernels (FDK) in Canadian
Received in revised form wheat samples. Eight hundred kernels of Canada Western Red Spring wheat were segregated into three
23 September 2010
classes of kernels: sound, mildly damaged and severely damaged. Singulated kernels were scanned with
Accepted 4 October 2010
a hyperspectral imaging system in the visible-NIR (400–1000 nm) wavelength range. Principal compo-
nent analysis (PCA) was performed on the images and the distribution of PCA scores within individual
Keywords:
kernels measured to develop linear discriminant analysis (LDA) models for predicting the extent of fusar-
Wheat
Fusarium damage
ium damage. An LDA model classified the wheat kernels into sound and FDK categories with an overall
Spectral imaging accuracy of 92% or better. Classification based on six selected wavelengths was comparable to that based
on the full-spectrum data.
Crown Copyright © 2010 Published by Elsevier B.V. All rights reserved.

1. Introduction damaged kernels have a white or pinkish mould or fibrous

Fusarium head blight (FHB), also known as scab or tomb-
stone, is a disease of small grain cereals such as wheat, barley,
and oat (Gallenberg, 2002). The principal causal agent of FHB
is Fusarium graminearum Schwabe (Goswami and Kistler, 2004).
FHB is favoured by wet weather at flowering, but can infect the
grain until harvest, given suitable conditions for infection. How-
ever, infection at the early stages of seed development causes
the greatest physical damage to the seed and the highest lev-
els of mycotoxin production (McMullen et al., 1997; Windels,
2000). In Canada, kernels infected with Fusarium spp. are called
fusarium damaged kernels (FDK). As the result of Fusarium spp.
infection, the kernels may also contain mycotoxins such as deoxyni-
valenol (DON), also known as vomitoxin (http://www.gipsa.
usda.gov/GIPSA/documents/GIPSA Documents/b-vomitox.pdf). A
positive correlation between scab damage and DON levels
has been found (Teich et al., 1987). Fusarium infection may
have detrimental effect on flour color, ash content, and bak-
ing performance (Dexter et al., 1996) as well as other quality
and safety issues (Dexter and Nowicki, 2003). In the Cana-
dian grading system, fusarium-damaged wheat is typically
characterized by thin or shrunken chalk-like kernels. Fusarium-
∗ Corresponding author. Tel.: +1 204 984 7776; fax: +1 204 983 0724.
E-mail addresses: muhammad.shahin@grainscanada.gc.ca,
mshahin@grainscanada.gc.ca (M.A. Shahin), stephen.symons@grainscanada.gc.ca
(S.J. Symons).
growth (http://www.grainscanada.gc.ca/oggg-gocg/04/oggg-gocg-
4e-eng.htm#r), contrasting with the USDA definition which
considers only those kernels that are chalk-like as scabby or
tombstone kernels. In the current Canadian grading system, a
representative sample is visually inspected for kernels showing evi-
dence of Fusarium spp. infection. This process is slow when only
slight damage is apparent as inspectors have to use a 10× magni-
fying lens to examine each suspect seed for the degree of mould
growth. In contrast, severely fusarium damaged wheat is typically
characterized by thin or shrunken chalk-like kernels easily detected
by human inspectors. Visual detection of fusarium damage at an
early stage is challanging. Fast and accurate instrumental methods
are required to meet the needs of grain industry.
Several laboratory methods are available for measuring DON
concentrations in wheat meal and flour including thin-layer chro-
matography (Fernadez et al., 1994), liquid chromatography (Chang
et al., 1984), gas chromatography (Tacke and Casper, 1996), mass
spectrometry (Scott et al., 1981) and enzyme-linked immunosor-
bent assay (Casale et al., 1988). However, these methods are not
suitable for rapid online inspection and quality control proto-
cols. Rapid inspection or sorting methods for grain are typically
based on kernel density using a gravity table (Tkachuk et al.,
1991) or optical properties (Ruan et al., 1998). In a recent prelim-
inary research, Takenaka et al. (2009) have reported encouraging
results on DON decontamination using a combination of gravity
separation and optical sorting. Williams (1997) used near-infrared
(NIR) spectroscopy to model DON levels in whole grain bulk sam-
ples with only moderate success. Dowell et al. (1999) investigated

0168-1699/$ – see front matter. Crown Copyright © 2010 Published by Elsevier B.V. All rights reserved.
doi:10.1016/j.compag.2010.10.004
108 M.A. Shahin, S.J. Symons / Computers and Electronics in Agriculture 75 (2011) 107–112
NIR reflectance to measure DON concentration in single kernels
(r2 = 0.64, SE = 44 ppm). Using NIR spectroscopy of individual ker-
nels, sound and FDK were segregated with high accuracy (95–97%)
under controlled laboratory conditions (Delwiche and Hareland,
2004). In contrast, test results using this technique on commer-
cial samples under commercial sorting operational conditions had
a much lower accuracy (50%; Delwiche, 2008). Peiris et al. (2009)
examined NIR absorbance characteristics of various concentra-
tions of DON as well as of sound and fusarium damaged single
wheat kernels. This study indicated that NIR spectrometry in the
1000–2100 nm range could estimate DON levels in kernels hav-
ing more than 60 ppm DON. However, further investigations were
required to detect DON in asymptomatic kernels.
Hyperspectral imaging in the shortwave infrared range
(1000–2500 nm) has shown potential for detecting fungal con-
tamination in wheat (Zhang et al., 2007; Singh et al., 2007) and
fusarium damage in maize corn (Williams et al., 2010) and wheat
(Polder et al., 2005). The extremely high cost of cameras sensitive
in the shortwave infrared range has been a limiting factor in the
development of commercially viable application systems in this
waveband range. Berman et al. (2007) have shown that accuracy of
fungal detection in wheat grain with hyperspectral imaging over
the 420–1000 nm wavelength range could be just as good as over
the full spectral range of 420–2500 nm. Recently, the use of high-
power bichromatic light emitting diodes (LEDs) emitting at red and
green wavelengths have been reported to achieve moderate levels
of overall accuracies (50–85%) for detection of fusarium damaged
wheat kernels (Delwiche, 2008; Yang et al., 2009).
The objectives of this study were: (a) to investigate the
use of hyperspectral imaging in the visible-NIR spectral range
(400–1000 nm) for the detection of varying degrees of fusarium
damage in wheat, and (b) to identify a reduced set of wave-
lengths/wavebands to be used in a future development of a low
cost imaging system.
2. Materials and methods
2.1. Samples
A set of four hundred individual kernels of Canada Western Red
Spring (CWRS) wheat were hand picked from commercial samples
containing a range of FDK types. These kernels were individually
inspected and scored by a trained grain inspector (Industry Ser-
vices Division, Canadian Grain Commission) as sound (SND) or
FDK. The FDK kernels were further classified into severely damaged
(SVR) and mildly damaged (MLD) categories based on the extent of
fusarium damage. Mild damage was characterized as chalky-white
kernels with fungal or mycelial growth around the germ and in
the broadened crease, while severe damage was characterized as
shrivelled chalky-white kernels with abundant mycelial growth on
both seed surfaces with some pink discoloration at the germ. Ker-
nels with no visible symptoms of damage were characterized as
sound. This set of kernels was termed as the calibration set consist-
ing of 200 SND and 200 FDK (100 MLD, 100 SVR) kernels. Another
independent set of four hundred kernels, the validation set, was
collected from the 2009 harvest survey samples of CWRS wheat
received at the Grain Research Laboratory (GRL, CGC). The valida-
tion set consisted of 200 SND and 200 FDK (100 MLD, 100 SVR)
kernels.
2.2. Hyperspectral imaging system
A push-broom type hyperspectral imaging system (VNIR 100E;
Lextel Intelligence Systems, Jackson, MS, USA) in the visible-NIR
wavelength range (400–1000 nm) was used for spectral mea-
surements of wheat kernels. The imaging system consisted of
a prism-grating-prism spectrograph (ImSpector V10E; Specim,
Oulu, Finland), a high-resolution 14-bit CCD camera (PCO Imaging,
Germany), a motorized C-mount focusing lens, and a personal com-
puter. The motorized lens assembly moved in front of the camera
allowing for imaging stationary samples. Two 250 W quartz-
tungsten-halogen lamps were used for sample illumination. Power
to each lamp was regulated through a radiometric power supply
(M-69931; Newport Oriel, Stratford, CT, USA).
2.3. Image acquisition and calibration
For imaging, singulated wheat kernels, in batches of 24–36 per
image, were placed crease-down on a neutral-grey plastic board
and hyperspectral images (also known as hypercubes) were col-
lected in the diffuse reflectance mode. Each hypercube contained
three rows of kernels: sound (bottom row), mildly damaged (mid-
dle row), and severely damaged (top row). Image size was 800 by
400 pixels by 218 wavebands within 400–1000 nm range at a spec-
tral resolution of approximately 2.75 nm and a spatial resolution of
0.028 mm per pixel in both x and y directions. The exposure time
was set at 60 ms. Each kernel was approximately 1800 pixels in
area.
Dark current and white light reference images were collected
before imaging each sample to calibrate spectra at each pixel as
percent reflectance value. A polytetrefluoroethylene panel with
99% reflectance (Spectralon, Labsphere, USA) was used to collect
white light reference images. Dark current response images were
collected with the lamp off and a cap covering the focusing lens.
Calibrated reflectance images (R) were calculated using Eq. (1).
Iraw − Idark
R= (1)
Iwhite − Idark
where, Iraw is the non-calibrated original image of a sample, Iwhite is
the image of the white reference, and Idark is the dark current image.
Calibrated hypercubes were subset to keep 181 bands between
450 and 950 nm for further analyses. Data below 450 nm or above
950 nm were excluded due to the presence of excessive noise in
this range.
2.4. Spectral characteristics and kernel-background separation
In order to separate kernel objects from the image background,
a threshold value was determined based on spectral differences
between kernels and image background. Representative spectra
of wheat kernels and image background were extracted and visu-
ally observed using the region of interest (ROI) tool in the ENVI
software (Version 4.5; ITT Visual Information Solutions, Denver,
CO, USA). Three sound kernels were manually selected at ran-
dom from an image to represent ‘SND’ category (Fig. 1a). Similarly,
three mildly damaged kernels and three severely damaged kernels
were selected to represent ‘MLD’ and ‘SVR’ categories, respec-
tively. For the image background (BKG), roughly 2000 pixels in
groups of 400–500 pixels at different locations in an image were
selected to represent ‘BKG’ category. For each category, a represen-
tative spectrum was computed as the average of all pixel spectra
in the respective category (Fig. 1b). Based on the spectral differ-
ences between wheat kernels and the image background as shown
by the representative spectra, a threshold value was determined
to separate kernel objects from the image background to exclude
image background from subsequent analyses. A binary mask image
was created for each hypercube by thresholding the image band
at 600 nm where pixel intensities greater than 10 were labelled
as kernels (white) and values less than 10 were labelled as back-
ground (dark) (Fig. 1c). The kernel spectra were mean-normalized
by dividing each spectrum with its mean value computed along
 M.A. Shahin, S.J. Symons / Computers and Electronics in Agriculture 75 (2011) 107–112 109

Fig. 1. (a) True color representation of a hyperspectral image of CWRS wheat kernels with varying degree of fusarium damage—sound (bottom row), mildly damaged (middle
row) and severely damaged (top row). (b) Spectral response of sound (SND), mildly damaged (MLD) and severely damaged (SVR) kernels and image background (BKG). (c)
Mask image showing kernel blobs separated from the image background. (d) Normalized spectral response of sound (NSND), mildly damaged (NMLD) and severely damaged
(NSVR) wheat kernels.

the wavelength direction to visualize spectral similarities or differ- 2.6. Classification

ences due to the extent of damage thus minimizing the effect of
lighting inconsistencies within the image plane (Fig. 1d).
2.5. Principal component analysis (PCA)
Principal component analysis (PCA) is a widely used statisti-
cal technique for data reduction. Using inbuilt ENVI subroutine
pc rotate, PCA was performed on the full-spectrum (450–950 nm)
hyperspectral images. Image background was excluded from these
calculations using a binary mask image created for each hyper-
cube to separate kernels from the background as explained
earlier. For each kernel, mean values of the first 10 PCA scores
(mpc1–mpc10) and the corresponding standard deviation values
(sdpc1–sdpc10) were extracted for further analysis. The ENVI sub-
routine envi stats doit was used to compute these image statistics.
A macro was written in IDL software (Version 7.0.2; ITT Visual Infor-
mation Solutions, Denver, CO, USA) to automate the process of PCA
calculations and extraction of mean and standard deviation values
of PCA scores for each kernel in all the images in a batch mode.
Based on the loadings (eigenvectors) of the first five principal
components (PC), a set of six significant wavelengths were selected
maintaining the overall behaviour of all five loading functions. PCA
scores of the images were re-calculated for the selected wave-
lengths and kernel by kernel mean and standard deviation values of
PCA scores were extracted using an IDL macro. The purpose of the
second pca-rotation was not data reduction: it was done to generate
orthogonal components and to compute within kernel variations
in the rotated planes so that the same set of features as for the
full-spectrum could be used for kernel classification. The PCA score
measurements from the full-spectrum and selected wavelengths
were used to develop and compare classification models.
The PCA scores data along with the target values (inspector
scores) were imported into the SAS software (version 9.1; SAS Insti-
tute Inc., Cary, NC, USA) to develop classification models. First,
variables with significant contribution towards desired classifi-
cation were selected from the PCA score measurements of the
calibration dataset using the SAS procedure stepdisc. Second, linear
discriminant analysis (LDA) classifier models were developed based
on the selected variables using the SAS procedure discrim. Data
from the calibration sample set was used to develop the LDA clas-
sifier models using the full cross-validation (leave one out) option
while data from the validation sample set was used for performance
evaluation of the models.
Two LDA models were developed based on PCA scores data
extracted from the full-spectrum images: LDA-1a was developed to
classify wheat kernels into two categories: SND and FDK (mild and
severe damage combined); LDA-1b was developed to further clas-
sify wheat kernels assigned to FDK category into two sub-classes:
MLD and SVR. Two more classification models were built using the
wavelengths selected based on the PCA-scores data: LDA-2a for
SND–FDK classification and LDA-2b for MLD-SVR classification. Per-
formance of the LDA classifiers was determined in terms of overall
or total accuracy of classification as well as class-by-class accuracy
of classification on the validation dataset.
3. Results and discussion
3.1. Visual and spectral characteristics
True color representation of hyperspectral images of CWRS
wheat kernels with different levels of fusarium damage exhibited
110 M.A. Shahin, S.J. Symons / Computers and Electronics in Agriculture 75 (2011) 107–112

Table 1a Table 1b
Feature selection for classification into sound (SND) and FDK categories. Feature selection for classification into mild (MLD) and severe (SVR) fusarium dam-
age categories.
Variable in the modela Partial R2 p-Value Wilk’s Lambda
Variable in the modela Partial R2 p-Value Wilk’s Lambda
None – – 1.000
sdpc1 0.3722 <0.0001 0.6278 None – – 1.000
mpc2 0.3366 <0.0001 0.4165 mpc3 0.3621 <0.0001 0.6379
sdpc6 0.0686 <0.0001 0.3879 mpc2 0.0893 <0.0001 0.5809
sdpc3 0.0286 <0.0001 0.3767 mpc6 0.0463 <0.0001 0.5541
mpc5 0.0248 <0.0001 0.3675 mpc4 0.0753 <0.0001 0.5123
mpc4 0.0113 0.0026 0.3633 mpc1 0.0773 <0.0001 0.4727
mpc1 0.0126 0.0015 0.3587 sdpc4 0.0125 0.0264 0.4668
mpc3 0.0117 0.0023 0.3545 sdpc1 0.0259 0.0013 0.4547
mpc6 0.0071 0.0179 0.3520 mpc5 0.0132 0.0228 0.4487
a
sdpc2 0.0069 0.1018 0.4456
mpc1 to mpc6: mean values of the 1st 6 principal component scores computed
a
over a single wheat kernel. sdpc1 to sdpc6: standard deviation values of the 1st 6 mpc1 to mpc6: mean values of the 1st 6 principal component scores computed
principal component scores computed over a single wheat kernel. over a single wheat kernel. sdpc1 to sdpc6: standard deviation values of the 1st 6
principal component scores computed over a single wheat kernel.

observable visual differences (Fig. 1a). Sound kernels (bottom row)

Table 2a
appeared much darker in color than severely damaged kernels (top
Sound (SND) and FDK classification results for an LDA model based on selected PCA
row). Mildly damaged kernels (middle row) appeared somewhat score features extracted from full spectrum (450–950 nm).
brighter than the sound kernels but darker than the severely dam-
Actual class Number of kernels classified into class
aged ones. Representative spectra of CWRS wheat kernels with
different degree of fusarium damage as well as the image back- Calibration set Validation set
ground extracted from one of the hyperspectral images are shown SND FDK Total Accuracy % SND FDK Total Accuracy %
in Fig. 1b. Each spectrum shown is the average of all pixel spectra
SND 188 12 200 94 185 15 200 92.5
in the respective category. In contrast to the spectral response of
FDK 16 184 200 92 17 183 200 91.5
the kernels, the image background exhibited near-zero response Total 204 196 400 93 202 198 400 92
in the visible range of the spectrum. Based on these spectral dif-
Bold values indicate correct classification.
ferences, the image background could be separated from kernels
using a simple threshold value in a single image band in the vis-
ible range as shown in Fig. 1c produced from the image band at kernel mean of PC2 (mpc2) and spatial variability of PC6 (sdpc6).
600 nm using a threshold value of 10. A mask so produced could be A similar set of significant features was selected for classification
used to extract and analyse kernel spectra excluding image back- of damaged kernels into MLD or SVR categories (Table 1b) consist-
ground from analyses. Fig. 1d shows the normalized spectra of ing of six mean score features (mpc1–mpc6) and three standard
the kernels eliminating intensity differences caused by illumina- deviation features (sdpc1, sdpc2, sdpc4). Kernel means for PC3, PC2
tion inconsistencies within the image plane. Sound and severely and PC6 (mpc3, mpc2 and mpc6) were the top three contributors
damaged kernels exhibited differently shaped spectral characteris- towards the MLD–SVR damage classification, in this order.
tics, however, none of the spectra showed any distinct absorption
bands within the entire wavelength range (450–950 nm). Sound 3.3. PCA based classification
kernels showed a curvilinear spectral response approximating a
3rd-order polynomial increase with wavelength. Spectral response Classification results for SND and FDK categories based on PCA
of severely damaged kernels exhibited approximately monotonic scores computed from the entire spectrum had an overall accuracy
gradual increase from visible to NIR spectral region while the of 93% and 92% for the calibration and validation datasets, respec-
response of mildly damaged kernels was in the middle, resembling tively (Table 2a). Class-by-class classification accuracies for SND
more to that of sound kernels in the case presented. The spectral and FDK were 92.5% and 91.5%, respectively on the validation set.
response of mildly damaged kernels would be expected to appear False positives and false negatives for the FDK category were 7.5%
as an intermediate transition spectra between those for the sound and 8.5%, respectively on the validation set. As would be expected,
and severely damaged ones. kernel classification accuracy was higher for the calibration data set
than for the independent validation set. Classification results for the
3.2. Principal component analysis (PCA) separation of MLD and SVR FDK show an overall accuracy of 87.5%
and 86% for the calibration and validation datasets, respectively
The first 10 principal components (PC1–PC10) explained over (Table 2b). Class-by-class accuracies for MLD and SVR categories
99.9% of the variance in hyperspectral image data with the first were 88% and 84%, respectively on the validation set.
three components explaining more than 99.5% of the variance in the The behaviour of the 1st five loading functions could be approx-
data. Contribution of PC4–PC10 combined was approximately 0.4%. imated by a set of six wavelengths or wavebands (Fig. 3). Using the
The first four principal component scores of the kernels (Fig. 1a) are
shown in Fig. 2. These score images contain information related to Table 2b
kernel classification based on fusarium damage, however, none of Mild (MLD) and severe (SVR) classification of fusarium damaged kernels based on
the principal components alone contained sufficient information to selected PCA score features extracted from full spectrum (450–950 nm).

fully segregate wheat kernels into target fusarium damage classes. Actual class Number of kernels classified into class
Accurate kernel classification required a combination of PC score
Calibration set Validation set
features. The stepdisc procedure selected a set of nine significant
features for the sound-damaged (SND–FDK) kernel classification MLD SVR Total Accuracy % MLD SVR Total Accuracy %
(Table 1a). Six mean score features (mpc1–mpc6) and three stan- MLD 90 10 100 90 88 12 100 88
dard deviation features (sdpc1, sdpc3, sdpc6) were selected. Spatial SVR 15 85 100 85 16 84 100 84
variability or standard deviation of PC1 (sdpc1) contributed the Total 105 95 200 87.5 104 96 200 86

most towards the target SND–FDK classification followed by the Bold values indicate correct classification.
 M.A. Shahin, S.J. Symons / Computers and Electronics in Agriculture 75 (2011) 107–112 111

Fig. 2. Score images of the first four principal components (PC1–PC4) of the image of CWRS kernels shown in Fig. 1a.

0.5 Table 3b
PC1 PC2 PC3 PC4 PC5 Mild (MLD) and severe (SVR) classification of fusarium damaged kernels based on
selected PCA score features extracted from 6 selected wavebands (484, 567 684,
817, 900, 950 nm).
0.25
Actual class Number of kernels classified into class

Calibration set Validation set

Value

0 MLD SVR Total Accuracy % MLD SVR Total Accuracy %

MLD 86 14 100 86 86 14 100 86

SVR 14 86 100 86 18 82 100 82
-0.25 Total 100 100 200 86 104 96 200 84

Bold values indicate correct classification.

-0.5
450 550 650 750 850 950 lem using a low cost imaging system built around a monochrome
Wavelength, nm digital camera and a set of optical filters in a motorized filter
wheel. For industrial uptake, the lower cost of such a multi-spectral
Fig. 3. Loadings (eigenvectors) of the first five principal components (PC1–PC5); approach would be appealing. Such a system would initially have
Wavelengths marked with circles approximate the behaviour of the five loading the potential to identify samples requiring further chemical analy-
functions. sis for toxin contamination. Subsequent studies in this laboratory
will continue to explore this possibility.
six selected wavebands (484 nm, 567 nm, 684 nm, 817 nm, 900 nm
and 950 nm), the overall classification accuracy for SND and FDK
4. Summary and conclusions
categories (Table 3a) and MLD and SVR categories (Table 3b) was
92% and 84% respectively on the validation dataset. Kernel classifi-
A visible-NIR hyperspectral imaging system was used to detect
cation based on the six selected wavebands was comparable to that
fusarium damage in CWRS wheat over the 450–950 nm range.
based on the full-spectrum (450–950 nm). These results demon-
Using principal component analysis (PCA), the hyperspectral image
strate that two levels of FDK in CWRS wheat can be detected with
data were reduced to 10 PCA score images and linear discriminant
six selected wavebands.
analysis (LDA) models developed for kernel classification in two
The fact that a few specific wavebands show promise for the
stages, sound and damaged and then subsequently the damaged
detection of FDK suggests that it may be possible to solve this prob-
kernels into mild and severe. Based on the results of this research,
it can be concluded that hyperspectral imaging over 450–950 nm
Table 3a
can be used to detect varying degrees of fusarium damage in CWRS
Sound (SND) and FDK classification results for an LDA model based on selected
PCA score features extracted from 6 selected wavebands (484, 567 684, 817, 900, wheat. Using LDA, sound and damaged kernels can be classified
950 nm). with an overall accuracy of 92% and the extent of fusarium damage
can be predicted with an accuracy approaching 86%. A set of six
Actual class Number of kernels classified into class
wavebands can be used to achieve accuracies similar to those with
Calibration set Validation set the entire range of 450–950 nm.
SND FDK Total Accuracy % SND FDK Total Accuracy %

SND 186 14 200 93 185 15 200 92.5 Acknowledgements

FDK 14 186 200 93 17 184 200 92
Total 200 200 400 93 201 199 400 92.25 The authors thank the Inspection Division of the Canadian
Bold values indicate correct classification. Grain Commission (CGC) for providing inspected samples for this
112 M.A. Shahin, S.J. Symons / Computers and Electronics in Agriculture 75 (2011) 107–112
research. They would also like to thank Loni Powell of the Image
Analysis & Spectroscopy Program (CGC) for scanning the samples.
References
Berman, M.M., Connor, P.M., WhitbournF L.B., Coward, D.A., Osborne, B.G., SouthanF
M.D., 2007. Classification of sound and stained wheat grains using visible and
near infrared hyperspectral image analysis. J. Near Infrared Spectrosc. 15 (6),
351–358.
Casale, W.L., Pestka, J.J., Hart, L.P., 1988. Enzyme-linked immunosorbent assay
employing monoclonal antibody specific for deoxynivalenol (vomitoxin) and
several analogues. J. Agric. Food Chem. 35, 663–668.
Chang, H.L., De Vries, J.W., Larson, P.A., Patel, H.H., 1984. Rapid determination of
deoxynivalenol (vomitoxin) by liquid chormatography using modified Romer
column cleanup. J. AOAC 67, 52–54.
Delwiche, S.R., 2008. High-speed bichromatic inspection of wheat kernels for mold
and color class using high-power pulsed LEDs. Sens. Instrum. Food Qual. Saf. 1
(2), 103–110.
Delwiche, S.R., Hareland, G.A., 2004. Detection of scab-damaged hard red spring
wheat kernels by near-infrared reflectance. Cereal Chem. 8 (15), 643–649.
Dexter, J.E., Nowicki, T.W., 2003. Safety assurance and quality assurance issues asso-
ciated with Fusarium head blight in wheat. In: Leonard, K.J., Bushnell, W.R. (Eds.),
Fusarium Head Blight of Wheat and Barley. APS, St. Paul, MN, pp. 420–460.
Dexter, J.E., Clear, R.M., Preston, K.R., 1996. Fusarium head blight: effect on the
milling and baking of some Canadian wheats. Cereal Chem. 73, 695–701.
Dowell, F.E., Ram, M.S., Seitz, L.M., 1999. Predicting scab, vomitoxin, and ergos-
terol in single wheat kernels using near-infrared spectroscopy. Cereal Chem.
76, 573–576.
Fernadez, C., Stack, M.E., Musser, S.M., 1994. Determination of deoxynivalenol in
1991 U.S. winter and spring wheat by high-performance thin-layer chromatog-
raphy. J. AOAC Int. 77, 628–630.
Gallenberg, D.J., 2002. Wheat scab Extension Extra, ExEx 8097. Cooperative Exten-
sion Service, SDSU/USDA, http://agbiopubs.sdstate.edu/articles/ExEx8097.pdf.
Goswami, R.S., Kistler, H.C., 2004. Heading for disaster: fusarium graminearum on
cereal crops. Mol. Plant Pathol. 5 (6), 515–525.
McMullen, M., Jones, R., Gallenberg, D., 1997. Scab of wheat and barley: a re-
emerging disease of devastating impact. Plant Dis. 81, 1340–1348.
Peiris, K.H.S., Pumphrey, M.O., Dowell, F.E., 2009. NIR absorbance characteristics
of deoxynivalenol and of sound and fusarium-damaged wheat kernels. J. Near
Infrared Spectrosc. 17, 213–221.
Polder, G., Van Der Heijden, G.W.A.M., Waalwijk, C., Young, I.T., 2005. Detection of
Fusarium in single wheat kernels using spectral imaging. Seed Sci. Technol. 33,
655.
Ruan, R., Ning, S., Song, A., Ning, A., Jones, R., Chen, P., 1998. Estimation of Fusarium
scab in wheat using machine vision and a neural network. Cereal Chem. 75,
455–459.
Scott, P.M., Lau, P.-Y., Kanhere, S.R., 1981. Gas chromatography with electron capture
and mass spectrometric detection of deoxynivalenol in wheat and other grains.
J. AOAC 64, 1364–1371.
Singh, C.B., Jayas, D.S., Paliwal, J., White, N.D.G., 2007. Fungal detection in wheat
using near-infrared reflectance hyperspectral imaging. Trans. ASABE 50, 2171.
Tacke, B.K., Casper, H.H., 1996. Determination of deoxynivalenol in wheat, barley,
and malt by column cleanup and gas chromatography with electron capture
detection. J. AOAC Int. 79, 472–475.
Takenaka, H., Kawamura, S., Sumino, A., Yano, Y. New combination use of gravity
separator and optical sorter for decontamination deoxynivalenol of wheat. Pro-
ceedings of the 5th CIGR International Technical Symposium on Food Processing,
Monitoring Technology in Bioprocesses and Food Quality Management, CD-
ROM, ISBN:978-3-000-028811-1, 1036-1039, 31 August to 2 September, 2009,
Potsdam, Germany.
Teich, A.H., Shugar, L., Smid, A., 1987. Soft white winter wheat cultivar field-
resistance to scab and deoxynivalenol accumulation. Cereal Res. Commun. 15,
109–114.
Tkachuk, R., Dexter, J.E., Tipples, K.H., Nowicki, T.W., 1991. Removal by specific
gravity table of tombstone kernels and associated trichothecenes from wheat
infected with Fusarium head blight. Cereal Chem. 68, 428–431.
Williams, P.C., 1997. Recent advances in near-infrared applications for the agricul-
ture and food industries. In: Steele, J.L., ChungF O.K. (Eds.), Proc. Int. Wheat
Quality Conf. Steele. Grain Industry Alliance, Manhattan, KS, pp. 109–128.
Williams, P., Manley, M., Fox, G., Geladi, P., 2010. Indirect detection of Fusarium
in maize (Zea mays L.) kernels by near infrared hyperspectral imaging. J. Near
Infrared Spectrosc. 18, 49–58.
Windels, C.E., 2000. Economic and social impact of Fusarium head blight: changing
farms and rural communities in the northern Great Plains. Phytopathology 90,
17–21.
Yang, I.-C., Delwiche, S.R., Chen, S., Lo, Y.M., 2009. Enhancement of Fusarium head
blight detection in free-falling wheat kernels using bichromatic pulsed LED
design. Opt. Eng. 48 (2) (February 2009) 023602.
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 sup-
port vector machine. Trans. ASABE 50, 1779.