Reliability-Aware Explainability for Rice Leaf Disease Classification: Weak Localization, Cross-XAI Comparison, and Cross-Source Generalization

This study presents a reliability-oriented framework for evaluating whether high classification performance in rice leaf disease recognition is accompanied by spatially faithful and generalizable explanations. The primary dataset comprised 4,684 images from three disease classes: bacterial blight, brown spot, and leaf smut. Perceptual-hash grouping and group-preserving partitioning were applied to reduce duplicate-related data leakage, and three ImageNet-pretrained residual networks were evaluated under a common training protocol. ResNet50 achieved the highest observed held-out performance, with an accuracy of 0.9968 and macro-F1 of 0.9967. Explanation reliability was evaluated using a 286-image Mini-GT benchmark with lesion-level bounding boxes, together with localization metrics, calibration analysis, perturbation robustness, and threshold sensitivity. Despite perfect classification on this evaluation subset, Grad-CAM++ produced limited localization, with a mean IoU of 0.1831 and pointing-game accuracy of 0.4720. A comparative evaluation of six XAI methods found no significant overall difference in paired per-image IoU (Friedman χ²(5)=4.9796, p=0.4184), indicating that the observed localization limitation was not specific to Grad-CAM++. Independent cross-source evaluation on a non-overlapping 120-image dataset yielded an accuracy of 0.5667 (95% CI: 0.4773–0.6519) and macro-F1 of 0.5453, revealing substantial degradation in external generalization. Overall, the findings demonstrate that high internal classification performance does not necessarily imply reliable lesion localization, stable explanations across XAI methods, or strong cross-source generalization.