Explainable AI for Precise Leaf Disease Diagnosis: A Comparative Study

Sustainable agriculture depends on the timely identification of plant diseases. While deep learning, particularly Convolutional Neural Networks (CNNs), has shown promise in leaf disease classification, its black-box nature undermines transparency and farmer trust. This study presents a framework integrating EfficientNet-B4/B7 with multiple Explainable AI (XAI) techniques—Grad-CAM++, Score-CAM, Integrated Gradients, and LIME—to enhance interpretability. The proposed framework was developed and tested on the PlantVillage dataset. Although the classification model itself demonstrated modest performance (AUC 0.53), highlighting challenges in generalization, the primary contribution of this study lies in the comparative analysis of XAI methods. A qualitative expert evaluation revealed Grad-CAM++ as the most consistent and visually coherent method for highlighting disease-related features. A prototype farmer-friendly GUI was developed to visualize input images, predicted classes, and XAI heatmaps. This work underscores the importance of interpretability in building trustworthy AI systems for agriculture and provides a comparative baseline for XAI techniques in plant pathology, establishing a foundation for future work on more robust classifiers.