Comparison of Vis-NIR and SWIR hyperspectral imaging for the non-destructive detection of DON levels in Fusarium head blight wheat kernels and wheat flour
Kun Liang,Jiani Huang,Ruiyin He,Qiujin Wang,Yinyin Chai,Mingxia Shen
DOI: https://doi.org/10.1016/j.infrared.2020.103281
2020-05-01
Abstract:Fusarium head blight is a severe wheat disease that affects yield and food safety. Chemical and biological methods are the main techniques for detection of the deoxynivalenol (DON) content, but despite being highly accurate, these techniques are destructive and time-consuming. A nondestructive, efficient and accurate method for estimating the DON content levels of scabs is required in post-harvest wheat processing. Hyperspectral imaging (HSI) systems have the capability of detecting subsurface defects and the contents or properties of organic substances. Because this method is nondestructive and fast and can be fully automated, the HSI technique was used in this study for detection of the wheat grain quality. Mycotoxins can destroy tissues, cells, and even molecules in cereals. Furthermore, the contents or properties of organic substances such as proteins, carbohydrates, and lipids in wheat infected by DON toxin undergo changes that can be reflected in their spectral characteristics. In this study, data for both wheat kernels and wheat flour obtained using two types of HSI systems, i.e., visible near-infrared (Vis-NIR) (ranging from 400 to 1000 nm) and short-wave infrared (SWIR) (ranging from 1000 to 2500 nm), were investigated. Ninety-six samples of wheat kernels and wheat flour samples were divided into two groups, and the multiplicative scatter correction (MSC) and standard normal variate (SNV) were used for spectral data pre-processing. A support vector machine (SVM) and a sparse autoencoder (SAE) network were used to the build classifiers with the extracted important wavelengths by genetic algorithm (GA). The results suggest that it is possible to detect wheat kernel samples in the Vis-NIR bands and to detect wheat flour samples from SWIR data with high accuracy. Moreover, for wheat kernel samples, the results of the SAE and SVM models were roughly the same. For all types of classification models for wheat kernel samples using the Vis-NIR, the MSC-GA-SAE model achieved the highest prediction accuracy (100% for both the training and test sets). In contrast to the wheat kernels samples, for the wheat flour samples, the SNV-GA-SAE models achieved better results with the SWIR data in terms of correctly classifying the samples (100% for the training set and 96% for the test set), which suggested that it is possible to detect wheat flour samples from SWIR data with high accuracy.
optics,physics, applied,instruments & instrumentation