Lesion-Aware Multiple-Instance Learning for Grapevine Disease Recognition from Field RGB Images

This retrospective offline study examines how dense manual regional annotations can be organized for image-level grapevine disease recognition using pre-acquired HERMOS RGB images. After audit and filtering, 491 source images and 13,765 Pascal VOC regions were retained. A leakage-safe source-image split was used across full-image classification, regional crop classification, crop-to-image aggregation, annotation-guided multiple-instance learning (MIL), and YOLO11 localization. MIL bags in training, validation, and testing were constructed from crops derived from manual boxes; the results therefore describe annotation-assisted aggregation rather than raw-image deployment. The original ConvNeXt-Tiny MIL model achieved macro-F1 0.922 (95% source-image bootstrap CI 0.882-0.956), compared with 0.836 (0.754-0.895) for full-image ConvNeXt-Tiny and 0.869 (0.804-0.917) for the best fixed crop aggregation. Paired macro-F1 differences favored MIL over these baselines by 0.087 (95% CI 0.029-0.163; Holm-adjusted p = 0.013) and 0.053 (0.005-0.114; adjusted p = 0.036), respectively. Downy mildew precision and recall were 1.000 on nine positive test images, but the corresponding exact 95% binomial intervals were 0.664-1.000. Controlled ablations produced lower point estimates without learned attention or the auxiliary instance loss, whereas removing context expansion did not reduce performance; none of the three paired ablation tests was significant after Holm correction. Increasing detector capacity from YOLO11n to YOLO11s did not improve full-image mAP@0.50 (0.175 versus 0.170). These findings support annotation-guided local-evidence aggregation as an offline exploratory benchmark, while not establishing split-to-split stability, computational efficiency, or external-vineyard generalization.