Deep Learning-based Leaf Segmentation and Disease Classification

Timely and precise diagnosis of tomato foliar diseases is crucial to minimize crop losses and enhance productivity in the farming sector. Yet deep learning models that are trained on controlled data sets are not as effective in field settings because they are unable to tackle background clutter, lighting changes and multiple objects in a field scene. To overcome these problems, the present study suggests a deep learning-based system which segments leaf regions prior to disease diagnosis. The images of tomato leaves from PlantVillage dataset are preprocessed using YOLOv8 to detect the leaf and Segment Anything Model (SAM) to segment the leaves and remove noises. The segmented leaf images are then used to train an EfficientNetV2 classifier for nine tomato disease classes. To evaluate real-world applicability, the trained model is tested on the PlantDoc dataset. Since complex backgrounds negatively affect classification performance, U2-Net is employed to segment leaf regions from PlantDoc images prior to classification. Experimental results show a validation ac curacy of 95% on the PlantVillage dataset and a real-world accuracy of 55.76% on the PlantDoc dataset after segmentation, compared with approximately 30% accuracy without segmentation. The results demonstrate that leaf segmentation significantly improves disease classification performance in challenging field conditions.