A multi-objective intelligent self-distilled lightweight network for explainable grape leaf disease recognition and decision support

Grapevine foliar diseases require timely scouting, yet many image-based diagnostic models remain too demanding for routine edge deployment. This study presents MOHEOA–SDM-LiteNet , a lightweight, explainable classification framework for Black Rot, Esca, Leaf Blight, and Healthy grape leaves, evaluated on a public dataset of 9,027 images. The framework combines a MobileNetV2-derived search space, an auxiliary training exit with logit-level self-distillation, and multi-objective human evolutionary optimization over validation accuracy, inference latency, and parameter count. Methodological safeguards include perceptual-hash grouping, a post-split Hamming-distance audit, and an independent calibration split for temperature scaling. Representative pretrained backbones and focused ablations are evaluated using the same leakage-resistant partitions. The selected self-distilled model contains 0.54 million parameters and achieves 98.78% test accuracy, 0.9882 macro-F1, and an expected calibration error (ECE) of 0.0055, reduced to 0.0048 after temperature scaling. Across three matched seeds, the compact configuration without the auxiliary exit or self-distillation and the full model attain the same mean test accuracy (0.9893). The compact variant uses 0.338 million parameters and is therefore the recommended resource-efficient operating point, whereas the full model provides lower calibrated test ECE (0.00490 versus 0.00792) in every paired seed and is retained as a secondary calibration-oriented variant. Thus, self-distillation is not claimed to improve accuracy. Grad-CAM++ provides qualitative evidence that predictions often focus on symptomatic regions, without establishing formal lesion localization. Overall, the workflow offers a reproducible within-dataset framework for uncertainty-aware grape leaf decision support. Reported latency is a controlled CUDA-based efficiency proxy, not a device-specific deployment guarantee.