Enhanced Rice Leaf Disease Detection Through InceptionV3 Integration With Dual Attention and GRAD-CAM Explainability

Abstract

In this study, we introduce an innovative deep learning strategy to automate the detection of rice leaf disease using the InceptionV3 model, together with dual attention mechanisms and GRAD-CAM. Rice production is threatened by numerous diseases, to which there are no effective solutions other than early identification and due treatment of where possible. Although classical CNNs are promising in plant disease recognition, they still face difficulty to capture the most discriminative disease-related features. To handle this we integrated both the spatial and channel attention modules into the InceptionV3 backbone. The spatial attention component learns to highlight disease-affected regions on leaves, while the channel attention mechanism prioritizes the most informative feature maps for accurate diagnosis. Through extensive experiments on a dataset of 18,445 rice leaf images spanning 10 disease categories, our attention-enhanced model achieved remarkable results: 96.45% classification accuracy in just 17 epochs. This substantially outperformed the baseline InceptionV3 model (80.21% accuracy), InceptionV3 with spatial attention alone (88.94%), and InceptionV3 with channel attention alone (92.06%). The dual attention approach proved particularly effective at distinguishing visually similar diseases such as bacterial leaf blight and leaf blast. Grad-CAM visualization techniques confirmed that our model successfully focuses on actual disease lesions rather than irrelevant background features. This research offers a practical, deployable solution for real-time rice disease diagnosis that could be integrated into mobile applications, potentially helping farmers reduce crop losses and optimize pesticide usage for more sustainable agriculture.