Abstract:Aiming to address the challenges of high spectral data redundancy and the limited accuracy of traditional machine learning models in identifying cotton Verticillium wilt severity levels, Nano Hyperspec hyperspectral cameras were mounted on drones to collect hyperspectral images of cotton fields. The spectral response characteristics of cotton canopies to different severity levels of Verticillium wilt were analyzed. An optimal vegetation index combination was identified and used to establish a monitoring model suitable for severity classification. This approach enabled precise classification of Verticillium wilt severity levels. The minimum redundancy maximum relevance algorithm was applied to rank the importance of features among 17 vegetation indices and 270 spectral bands. Features selected by this algorithm were incrementally grouped and input into an eXtreme gradient boosting model. This process determined the vegetation indices and spectral bands most strongly correlated with Verticillium wilt severity levels. A Transformer FNN ( feedforward neural network) classification model was then developed. Vegetation indices and spectral features were used as inputs to this model for classification. The classification accuracy of vegetation indices and spectral features in identifying Verticillium wilt severity levels was compared. Additionally, classification models based on back propagation neural network (BPNN), Transformer, and support vector machines ( SVM) were constructed. The accuracy of these models was validated and analyzed. The results showed that the optimal vegetation index combination for Verticillium wilt severity classification was MSR and TVI. The optimal spectral band combination included 430 nm, 439 nm, 488 nm, 566 nm, 697 nm, 722 nm, 742 nm, 764 nm, 769 nm, 782 nm, 822 nm, 831 nm, 858 nm, 873 nm, 878 nm, 893 nm, 909 nm, and 985 nm. Using the Transformer FNN model, the overall classification accuracy based on vegetation indices reached 95.6% . This represented a 6.2 percentage points improvement compared with the accuracy achieved by using spectral features, which was 89.4% . For vegetation indices, the Transformer FNN model achieved a classification accuracy of 95.6% . This was 11.2 percentage points higher than the accuracy of the BPNN model, 17.2 percentage points higher than that of the Transformer model, and 30.8 percentage points higher than that of the SVM model. The research proposed a high-accuracy monitoring method for cotton Verticillium wilt based on vegetation indices. It provided an effective approach for large-scale and precise monitoring of cotton Verticillium wilt.
