Abstract:Aiming to achieve rapid and non-destructive detection of Fusarium head blight (FHB) severity levels, the hyperspectral imaging technology combined with machine learning models for analysis was employed. A total of 1660 wheat ear samples with varying degrees of infection were obtained by inoculating Fusarium fungi into the middle grains of wheat ears. Hyperspectral information of the samples was collected by using a hyperspectral imaging system, with the entire wheat ear designated as the region of interest (ROI) to extract average spectral data. By comparing the classification accuracy of four preprocessing methods—normalization, standard normal variate (SNV), multiplicative scatter correction (MSC), and Savitzky-Golay (SG) smoothing derivatives—the SNV algorithm was selected as the optimal preprocessing method. Subsequent analyses were conducted on the SNV-processed spectral data. Feature wavelength selection was performed by using the successive projections algorithm (SPA) and competitive adaptive reweighted sampling (CARS), while dimensionality reduction was implemented via uniform manifold approximation and projection (UMAP) and linear discriminant analysis (LDA). After comparing these algorithms, LDA was ultimately chosen for its ability to reduce data to three dimensions while maintaining classification accuracy (96.05% for the test set and 94.71% for the training set) and low computational complexity (0.09s processing time). It was revealed that the critical spectral range for LDA-based FHB severity discrimination lay between 540nm (chlorophyll reflection peak) and 650nm (red light absorption valley), attributed to the synergistic effects of rapid chlorophyll degradation and structural tissue damage as infection progresses. A lightweight support vector machine (SVM) model integrating SNV and LDA was developed as the optimal framework for FHB severity classification. The results demonstrated that the proposed algorithm achieved high accuracy with excellent generalization capability, enabling efficient FHB severity assessment. The research result can lay a foundation for future large-scale, real-time field detection of FHB.
