Imaging techniques can be used to evaluate the quality and safety of
agricultural products. Fusarium head blight (FHB) results in reduced barley yields and
also diminished value of harvested barley. Deoxynivalenol (DON) is a mycotoxin
produced by the causal Fusarium species that pose health risks to humans and
livestock. DON has currently measured via gas chromatography (GC) methods that are
time-consuming and expensive. We seek to apply imaging technology to rapidly and
non-destructively quantify DON in high throughput and less expensive method. The
feasibility of hyperspectral imaging to determine DON contents of barley kernels was
evaluated using machine learning algorithms. Partial least square discriminant analysis
(PLSDA) was able to discriminate kernels into four separate classes corresponding to
their DON levels. Barley kernels could be classified as having low (<5 ppm) or high
DON levels, with Matthews's correlation coefficient in cross-validation (M-RCV) of as
high as 0.823. PLSR showed good performance in linear algorithms for DON
detection, but higher accuracy was obtained by non-linear algorithms, including
weighted partial least squares regression (LWPLSR), support vector machine
regression (SVMR), and artificial neural network (ANN). Among all algorithms, the
non-linear LWPLSR achieved the highest accuracy, with the coefficient of
determination in prediction (R2
P) of 0.728 and root mean square error of prediction
(RMSEP) of 3.802. The results demonstrate that hyperspectral imaging and machine
learning algorithms have the potential to assist the FHB resistance breeding process by
accelerating the quantification of DON in barley samples.
Keywords: Food safety, Hyperspectral imaging, Machine
learning.
