Explainable artificial-intelligence-based hyperspectral image analysis for leaf disease detection in intercropping system

Introduction: Intercropping regimes enhance the efficiency of land use and ecological sustainability but present serious problems to automated disease analysis since the overlapping canopy and the similarity of symptoms in crop species are visually indistinguishable. Methods: This work presents an explainable artificial intelligence (XAI)-based hyperspectral analysis on leaf disease in intercropping systems. The framework combines the spectral-spatial feature generators that utilize transformers including vision transformer (ViT), Swin transformer, pyramid vision transformer (PVT), and detection transformer (DETR) to identify nuanced biochemical and structural changes in crop combinations for maize-soybean and pea-cucumber. In order to reduce spectral redundancy and high dimensionality, an enhanced greedy political optimization (EGPO) algorithm is used as a wrapper-based feature selection strategy. A capsule spatial shift neural network (CSSNet) is used to predict the classification of diseases. Explainable AI methods, such as Local Interpretable Model-agnostic Explanations (LIME) and SHapley Additive exPlanations (SHAP) feature attribution analysis and gradient-weighted class activation mapping (Grad-CAM) visualization of disease-relevant regions, provide model transparency. The DETR + EGPO + CSSNet framework is tested on the conventional feature selection methods. Results and discussion: The results or findings on publicly available hyperspectral datasets on intercropping show an average recall of 99.998% with high region consistency (Dice score: 99.997%) of activation maps and expert-marked disease regions. These findings affirm that the proposed framework is highly accurate, stable, and interpretable to identify subtle and overlapping disease in leaves in a complex system of intercropping.