Hierarchical Multi-Stage Framework for Robust and Explainable Tomato Leaf Disease Identification

Accurate identification of tomato leaf diseases is essential for improving crop health management and reducing agricultural losses. However, the wide variety of tomato diseases, the limited availability of real-world datasets, and the variability of agricultural environments make developing reliable and explainable tomato disease identification systems challenging. Traditional deep learning models trained on controlled datasets often fail to generalize in real-world farm environments due to complex backgrounds and variations in illumination and image conditions. This paper presents a hierarchical, multi-stage framework that integrates leaf detection, segmentation, and classification with interpretability to achieve robust and explainable disease identification. The proposed pipeline employs YOLO11 for leaf detection, the Segment Anything Model (SAM) for segmentation, a ResNet-50 classifier, and LIME interpretability. To evaluate the impact of the pipeline on robustness and generalization, we conducted cross-domain experiments in two settings, using PlantVillage for laboratory conditions and PlantDoc for real-world conditions. The proposed pipeline improved real-world accuracy from 22.79 % with a flat ResNet-50 to 55.33 %, cutting the accuracy drop by 34.0 %. The pipeline stages focused on leaf isolation to ensure that predictions are driven by symptomatic tissue rather than background. Experiments demonstrate that our approach significantly improves generalization and transparency, addressing major gaps in existing plant disease identification systems.