Tomato TransDeepLab: A Robust Framework for Tomato Leaf Segmentation, Disease Severity Prediction, and Crop Loss Estimation

This research aims to reduce crop loss in tomato crops by developing an effective disease detection and segmentation system. To meet this objective, the researchers prepared three annotated real-world datasets consisting of tomato leaf images, including samples with simple and complex backgrounds, and young plants aged two to four weeks. A transformer-based deep learning model, ‘Tomato TransDeepLab’, is proposed for accurately segmenting multiple and minute disease lesions on a single leaf. The model demonstrates high segmentation precision even with complex backgrounds and indistinct lesion boundaries. It achieves a maximum Intersection over Union (IoU) of 89.09% and accuracy of 98.17%. It also recorded the lowest training time of 989 minutes. Comparative analysis with state-of-the-art architectures including U-Net, ResUNet, DeepLabV3, and Tuned DeepLabV3+ across all three datasets confirms its superior performance. Furthermore, the model is employed to assess disease severity and estimate crop loss using a severity scale designed by plant pathologists. This scale links lesion area to severity level and crop loss. Results indicate a crop loss between 1 to 5% for Datasets 1 and 2. The proposed approach offers an automated solution for disease segmentation, severity assessment, and crop loss estimation in tomato crops.