Dual-head CNN–transformer network for interpretable potato tuber disease phenotyping and severity quantification

Accurate and interpretable assessment of potato tuber diseases is important for postharvest grading, but image-based models may rely on background cues rather than diseased tissue. We evaluated a dual-head CNN–Transformer for disease classification and lesion segmentation, incorporating Segmentation-Aware Pooling (SAP) with stop-gradient to guide classification towards lesions. A dataset of 421 annotated images covering eight disease categories was evaluated using stratified group five-fold cross-validation, training-only augmentation, and repeated experiments. The proposed model achieved 92.4 ± 1.5% classification accuracy, a mean Dice coefficient of 0.706, and a lesion-area mean absolute error of 3.08 %age points. SAP increased the proportion of Grad-CAM attribution within annotated lesions compared with global average pooling (0.234 versus 0.147), but LIME showed no significant difference (p = 0.19). Background replacement substantially reduced classification accuracy, indicating that lesion-focused attribution did not eliminate background dependence. Input-level lesion masking improved accuracy on background-swapped images from 38.4% to 59.5%, although accuracy on unmodified images decreased by approximately 3 %age points. Predictive entropy from test-time augmentation identified misclassifications (AUROC = 0.894), while cross-source evaluation revealed substantial sensitivity to imaging conditions. These findings highlight the distinction between visual attribution, predictive performance, and background robustness, emphasizing the need for complementary evaluation methods in automated potato disease assessment.