DEEP LEARNING MODELS FOR DISEASE AND PEST DETECTION IN CROPS USING MULTISPECTRAL AND HYPERSPECTRAL IMAGING

Abstract

Deep learning models combined with multispectral imaging (MSI) and hyperspectral imaging (HSI) offer a transformative, non-destructive approach for early detection of crop diseases and pests, addressing the 40% global yield losses caused by biotic stressors. Traditional manual scouting and laboratory methods are slow, subjective, and often too late for effective intervention, whereas spectral imaging captures subtle biochemical and structural changes (e.g., chlorophyll degradation, water content shifts, red-edge shifts) before visible symptoms appear. CNNs (including 1D, 2D, and 3D variants), Vision Transformers, and hybrid architectures (MSCVT) achieve high accuracies (often >95–99%) by extracting spatial-spectral features from hypercubes or selected bands. Key techniques include spectral preprocessing (Savitzky-Golay smoothing, MSC/SNV), dimensionality reduction (PCA, EGPO), and advanced training strategies (transfer learning, few-shot learning, GAN-based augmentation). Applications span fungal/bacterial/viral pathogen detection and pest infestation monitoring across cereals, vegetables, and other crops. Explainable AI (XAI) methods like Grad-CAM and SHAP enhance model transparency for agronomic decision-making, while model compression enables edge deployment on UAVs and IoT devices. Despite challenges such as the Hughes phenomenon, data scarcity, and model generalization, GeoAI-driven spectral analysis supports precision agriculture, reduces pesticide overuse, and advances sustainable crop protection.