JoWUA

Abstract

This study presents an advanced deep learning approach for automated tomato leaf disease detection, integrating Inception V3 architecture with spatial and channel attention mechanisms. As economically significant crops worldwide, Tomatoes require efficient disease monitoring systems for sustainable agriculture. While convolutional neural networks show promise in plant disease classification, they often lack focus on discriminative features. Our proposed model addresses this limitation by incorporating dual attention mechanisms into Inception V3. The spatial attention module highlights disease-specific regions, while the channel attention module prioritizes informative feature maps. Experiments on 18,160 tomato leaf images across nine disease categories and healthy samples demonstrated that our hybrid attention-enhanced model achieved 96.4% classification accuracy, outperforming baseline Inception V3 (90.2%), V3 with spatial attention achieved 95% accuracy, while Inception V3 with channel attention reached 96% in under 15 epochs. The model showed particular improvement in distinguishing visually similar diseases such as early and late blight. Grad-CAM visualization confirmed the model's capacity to focus on disease-specific lesions. This approach offers a robust solution for real-time tomato disease diagnosis deployable in mobile applications, potentially increasing crop yields while reducing pesticide usage.