Detection of sorghum variety based on CW-MST++ network reconstruction of hyperspectral images

In order to realize the rapid detection of sorghum varieties, a sorghum variety detection method based on spectral reconstruction technology combined with machine learning classification model was established. Firstly, the soft threshold (ST) function and squeeze-and-excitation (SE) module were embedded in the deep structure of the multi-stage spectral-wise transformer (MST++), and the channel-wise (CW)-MST++ was constructed. Using MST++, hierarchical regression network (HRNet) and CNN-based hyperspectral recovery from RGB images (HSCNN+) as controls, the performance of rebuild model was compared using mean relative absolute error (MRAE) and root mean square error (RMSE). Secondly, the Support Vector Machine (SVM) model was established by hyperspectral data of four reconstructed networks, and the performance of the detection model was evaluated by average accuracy (ACC), average recall rate (Recall) and F₁ scores. The results showed that compared with MST++, HRNet and HSCNN+ networks, the hyperspectral image reconstructed by CW-MST++ had the smallest error (MRAE=0.017 5, RMSE=0.007 6 in validation set), low number of network parameters and floating point operations (Params was 1.77M, Flops was 20.80 G), and the SVM model established by the reconstructed hyperspectral data had the optimal prediction effect (ACC=94.52%, Recall=94.24%, F₁ scores=94.14 in the testing set). Compared with the original hyperspectral data, the differences in ACC, Recall and F₁ scores of hyperspectral data reconstructed by CW-MST++ were only 2.06%, 2.54% and 2.52%, respectively, which realized the rapid and accurate detection of sorghum.