Attention-Driven Deep Learning Framework for Multi-Class Rice Leaf Disease Classification

Abstract

Crop productivity can be significantly impacted by rice leaf diseases, and prompt and precise disease detection is essential for efficient crop management. Using a combination of main and secondary image data sets, this work attempts to offer an attention-based deep learning model for multi-class rice leaf disease classification. 942 photos from five classes— bacterial leaf blight, fake smut, healthy, leaf blast, and sheath blight—were included in the final dataset after cleaning. A stratified train-validation-test split was created after duplicate, cross-labeled, and inconsistent unlabelled images were eliminated. The suggested AttentionMobileNetV2-SE model combines a lightweight backbone, MobileNetV2, with a Squeeze-and-Excitation attention module to improve the representation of disease-relevant data. Class-weighted cross entropy loss was employed to lessen the effects of class imbalance in training. The model yielded a weighted F1-score of 0.8656, validation accuracy of 90.07%, and test accuracy of 86.62%. By emphasising disease-relevant picture regions, Grad-CAM visualisation was also utilised to interpret model predictions. The outcomes show that the suggested framework can successfully differentiate between several rice leaf disease categories and that the model's judgements are visually interpretable.