Defective wheat kernel classification using dual-range hyperspectral imaging and an interpretable spectral-spatial...

The rapid and non-destructive screening of defective wheat kernels is essential for quality assurance and process control, yet reliable identification remains challenging due to subtle spectral and spatial differences between defective and sound wheat kernels. In this study, a spectral-spatial fusion convolutional neural network (SSFCNN) was developed to integrate complementary spectral and spatial information from hyperspectral images for the classification of five wheat kernel categories. An end-to-end fusion framework was constructed, in which squeeze-and-excitation (SE), shuffle attention (SA), and efficient channel attention (ECA) were integrated for spectral channel recalibration, spatial feature refinement, and fusion feature enhancement, respectively. The results demonstrated that the SSFCNN with deep feature fusion outperformed a support vector machine (SVM) and a convolutional neural network (CNN) constructed using conventional feature fusion. The highest overall accuracies of 96.48% in the visible and near-infrared (Vis-NIR) and 95.61% in the short-wave infrared (SWIR) were achieved, together with consistently improved precision, recall, specificity, and F1-score across all wheat kernel categories. Moreover, visualization of classification outputs on the external validation set indicated improved spatial coherence and decision reliability of the SSFCNN. Overall, this study provided a validated and interpretable spectral-spatial fusion framework for hyperspectral screening of defective wheat kernels, offering a methodological basis for future intelligent grading and online quality control applications after further validation under real sorting-line and cross-domain conditions.