Reflectance images of effective wavelengths from hyperspectral imaging for identification of Fusarium head blight-infected wheat kernels combined with a residual attention convolution neural network

Abstract

Identification of the Fusarium head blight (FHB) infection degree of wheat kernels is important to customise the reasonable use of wheat kernels and ensure food safety. In this study, an FHB infection degree identification method using hyperspectral imaging (HSI) and deep learning networks was proposed. Firstly, the reflectance spectra of healthy and mildly, moderately and severely FHB-infected wheat kernels were extracted from HSI images, and five effective wavelengths (EWs) of the spectra were selected by random frog. Secondly, the reflectance images (RIs) of different combinations of the EWs were screened using LeNet-5, and a residual attention convolution neural network (RACNN), which was constructed by increasing width and depth and adding channel attention and residual modules, was adopted to recognise various FHB infection degrees in wheat kernels. Optimal recognition performance was achieved by RACNN and RIs of 940 nm and 678 nm with a classification accuracy of 100%, 98.60% and 98.13% for the calibration, validation and prediction sets, respectively. Meanwhile, class activation maps revealed that the RACNN could effectively extract the distinctive features of the different classes of kernels. The images of only two wavelengths can be quickly acquired and processed, and the simultaneous recognition of multiple targets is easily realised. Overall, the proposed method enables the rapid, accurate and massive analysis of the FHB infection degree of wheat kernels.