Decoupled Foundation Models for Instance Segmentation and Automated Detection of Humidity-Induced Tomato Leaf Necrosis

Tomato cultivation is highly vulnerable to both biotic and abiotic stressors, which together constitute major constraints on global crop productivity and food security. Among these, abiotic stressors such as excessive humidity are particularly challenging because they often induce physiological disorders and necrotic leaf symptoms that resemble biological infections, complicating early diagnosis and timely intervention. Manual scouting is labor-intensive and prone to missed early-stage symptoms, motivating automated deep-learning-based detection systems. In this study, a multi-step AI pipeline for the automated segmentation and classification of humidity-induced necrotic leaf spots in tomatoes is proposed. A dataset of 218 RGB images was collected, yielding 3218 annotations across three classes (brown necrotic spots, yellow necrotic spots, and no necrotic leaves). Six end-to-end instance segmentation pipelines combining YOLO26m (for detection and segmentation), SAM2 (for zero-shot prompted segmentation), and either fine-tuned DINOv2 or EfficientNet-B3 (for downstream classification) were systematically designed and evaluated. Fine-tuned DINOv2 reached a macro F1-Score of 0.926 for per-instance crop classification, above EfficientNet-B3, ResNet-50 and Swin-Small baselines (0.886–0.901). The best-performing configuration (YOLO26m-det + SAM2 + DINOv2) achieved mAP@50 of 0.828, outperforming the single-model YOLO26m-seg baseline by approximately 8%. These results demonstrate that decoupling localization from classification through task-specific foundation models consistently outperforms single-model training on small, class-imbalanced agricultural datasets. By leveraging zero-shot segmentation foundation models like SAM2, this approach effectively bridges the gap in diagnostic performance for data-limited agricultural settings.