Early detection of multi-pest infestations with a hierarchical 3D-SSA-Net framework using synergistic RGB and hyperspectral imaging data

Early pest detection is often hindered by the phenotypic latency of RGB imaging and the computational costs of hyperspectral imaging (HSI). This study proposes a coarse-to-fine hierarchical monitoring framework that integrates RGB and HSI modalities. First, ResNet-50 achieved 98.88% recognition accuracy among four key greenhouse pests, including Tetranychus urticae Koch, Liriomyza sativae Blanchard, Bemisia tabaci Gennadius, and Aphis gossypii Glover. To quantify faint early-infestation signals, we developed 3D-SSA-Net, integrating 3D convolutions with a Spatial-Spectral Attention (SSA) mechanism. This model targets critical red-edge and near-infrared bands, enhancing feature extraction robustness. Results show that 3D-SSA-Net improves grading accuracy from a 48.33%–79.84% baseline to 76.67%–99.17%. To reduce spectral redundancy, XGBoost and SHAP were used to identify a biologically interpretable 10-wavelength subset from the 248-dimensional Point Percentage Change (PPC) representation. When evaluated using conventional machine-learning classifiers, Random Forest (RF) and Support Vector Machine (SVM) maintained high Overall Accuracy (OA) on both datasets, with each achieving an OA of 99.83% on the Mid-to-Late-stage dataset, while the number of spectral features was reduced by 95.97%. These findings demonstrate the compactness and predictive sufficiency of the selected wavelength subset and provide a preliminary spectral basis for future multispectral sensor development.