LeafFusionNet: An Interpretable EfficientNetV2-Based Pipeline for Automated Tomato Leaf Disease Recognition with Grad-CAM Visual Explanations
Abstract
The task of reliable and early detection of tomato leaf disease is crucial to both the crop yield and the farm's sustainability in the long run. The diagnosis of the disease at present still relies on the human judgment, which may prove to be errorprone. Convolutional neural networks have been successfully applied to numerous computer vision tasks, including disease detection, but only a limited number of studies are dedicated to tomato leaves, and even fewer explain their outcomes. The paper proposes LeafFusionNet, a transfer-learning-based classifier, which diagnoses the tomato leaf disease and uses Grad-CAM to explain its predictions by highlighting the image region that prompted the detection. The model was trained, validated, and tested on a dataset of approximately 11,000 RGB images. Using the EfficientNetV2B0 architecture, LeafFusionNet's encoder extracts high-level features and classifies the image into a disease category, while Grad-CAM highlights the region of the leaf that contributed most to the diagnosis. The model achieved 95.67% accuracy on the test set with high precision, recall, and F1-score for each category. Overall, the results validate the approach and suggest that LeafFusionNet can be used as part of an automated computer vision system for disease detection.