Abstract:
Soybeans are prone to mold during storage and transportation. In order to quickly and non-destructively detect the mold status of stored soybeans and improve the efficiency of quality inspection, this experiment employed a hyperspectral imaging system to collect samples of soybeans with different levels of mold. The spectral data of the samples were preprocessed using SNV, MSC, 1ST, and 2ND. Spectral features were extracted using the SPA and the CARS algorithm. Color features were obtained by calculating the first three moments of color moments. SVM and PLSR models were separately established by combining the spectral features and color features. The results showed that the SVM model, built using 1ST-processed spectral data, 19 spectral features extracted by CARS, and 9 color features, achieved the best performance. The SVM model achieved accuracy, precision, recall, and
F1 score of 98.13%, 98.25%, 98.15%, and 98.20%, respectively, on the test set. The recognition accuracy for each mold level was 100%, 100%, 92.6%, and 100%, with only 2 misclassified samples. This model demonstrated strong identification capabilities for soybeans with different levels of mold and provided a theoretical basis for rapid and non-destructive detection of mold levels in soybean kernels.