A comparative evaluation of explainable AI techniques across CNN and vision transformer architectures for mango leaf disease detection

Early and accurate disease diagnosis in mango leaves is essential for minimizing crop loss and supporting sustainable agriculture. Although deep learning models using convolutional neural networks (CNNs) and vision transformers (ViTs) achieve high diagnostic accuracy, their lack of interpretability limits practical adoption. This study presents a systematic comparative evaluation of four XAI methods, SHAP, LIME, Grad-CAM and Integrated Gradients, applied to best-performing ResNet50 and Swin Transformer models for multiclass mango leaf disease classification. Methods were evaluated using faithfulness, sparsity, stability and leaf-region localization accuracy through the pointing game, together with statistical analysis. The Friedman test demonstrated significant differences among the four XAI methods across all metrics in both architectures (p < 0.05). Integrated Gradients achieved the highest mean faithfulness for both ResNet50 (0.8082) and Swin Transformer (0.8323), while Grad-CAM achieved the highest stability (0.8640) and pointing game accuracy (0.6208) for ResNet50. For Swin Transformer, SHAP achieved the highest pointing game accuracy (0.3667), whereas Grad-CAM showed the highest mean stability (0.3512). These results show that no XAI method consistently dominates all criteria, and explanation behavior varies across CNN and Transformer architectures. The findings provide evidence-based guidance for selecting XAI techniques according to the desired interpretability objective in agricultural disease diagnosis.