PLANT FRUIT DISEASE SEVERITY ALERT NETWORK USING HYBRID DEEP LEARNING AND A RETRIEVAL-AUGMENTED GENERATION MODEL

Abstract

Agricultural fruit production is significantly affected by plant diseases that reduce crop yield, degrade fruit quality, increase pesticide dependency, and threaten global food security. Disease symptoms vary considerably in lesion texture, shape, color, infection spread, and environmental background, making accurate disease identification challenging. Existing machine learning approaches rely on handcrafted features and demonstrate poor generalization under varying field conditions. Conventional deep learning methods also have limitations, including inaccurate lesion boundary extraction, privacy concerns in distributed agricultural data, and insufficient severity-aware decision support. To address these challenges, this research proposes an integrated intelligent approach for fruit disease detection and severity estimation using segmentation, privacy-preserving classification, and explainable severity prediction through the Plant Fruit Disease Severity Alert Network (PDANet). The fruit disease image dataset was preprocessed using an anisotropic diffusion filter to suppress noise while preserving disease boundaries and texture information. The images were then segmented using the proposed Deformable Enhanced Multi-Channel Attention Swin U-Net (DES-UNet), which incorporates Enhanced Multi-Channel Attention and a Deformable Convolution-based Multi-Layer Perceptron to improve lesion boundary extraction, contextual feature learning, and irregular symptom localization. Disease classification was performed using the proposed Quantum ChronoNet-based Federated Learning (QC-FL) framework, in which the local model is based on Quantum ChronoNet. Finally, disease severity was estimated using a Retrieval-Augmented Generation (RAG)-based Large Language Model (LLM), which combines segmented lesion measurements, classified disease information, and retrieved agricultural knowledge to estimate disease severity and generate contextual farmer alerts. The proposed PDANet model achieved 98.8 % accuracy, 98.5 % F1-score, 98.2 % sensitivity, 98.6 % specificity, 0.98 Kappa coefficient, 0.98 Matthews Correlation Coefficient (MCC), and 0.97 General Detection Rate (GDR). The experimental results demonstrate that the proposed model provides accurate and stable fruit disease classification that supports lesion-based severity assessment.