Interpretable Deep Learning for Rice Leaf Disease Detection: A Self-Attention and XAI Framework

Rice leaf diseases pose a significant challenge to food security and agricultural productivity. To address this challenge, innovative disease detection methods are required. A deep learning-based method for detecting rice leaf diseases is proposed in this study, utilizing a Convolutional Neural Network (CNN) enhanced with a self-attention mechanism. The model was evaluated on an augmented dataset consisting of 22,974 images, which were expanded from 3,829 using data augmentation techniques. Disease identification and classification were performed using CNNs, including Self-Attention CNN, customized CNN, InceptionV3, MobileNetV2, and ResNet-50. The effectiveness of these models was demonstrated through experimental results, with the Self-Attention CNN achieving an accuracy of 0.9974, the customized CNN achieving 0.9857, InceptionV3 at 0.9665, MobileNetV2 at 0.9458, and ResNet-50 at 0.8847. SHapley Additive explanations (SHAP) were employed to provide pixel-level insights into the model’s predictions by integrating deep learning techniques with model interpretability methods. SHAP heatmaps confirmed that the model focused on disease-specific regions, aligning with domain knowledge and enhancing trust in the system. This transparency improves the reliability of the AI-driven disease detection system and supports its applicability in agriculture. The research demonstrates the effectiveness of integrating self-attention mechanisms with explainable AI, offering a scalable, interpretable, and highly accurate solution for rice leaf disease detection. By enabling precise disease identification and timely intervention, the study advances precision agriculture and promotes sustainable crop management.