Interpretable multitask deep learning model for detecting and analyzing severity of rice bacterial leaf blight

Abstract

Abstract Rice Bacterial Leaf Blight (BLB), caused by Xanthomonas oryzae pv. oryzae (Xoo), is a major threat to rice production due to its rapid spread and widespread impact. Early detection and stage-specific classification of BLB are essential for timely intervention, particularly in complex environments with cluttered backgrounds and overlapping symptoms. This study introduces RCAMNet, a novel multi-task framework designed for accurate classification and severity analysis of BLB. The proposed approach begins by generating multiclass segmentation masks using three candidate methods: MultiClass U-Net, DeepLabv3, and Detectron2 (used here for its instance segmentation capability). In the second phase, a dual-path attention mechanism is employed. The Convolutional Block Attention Module (CBAM) is independently applied to both the RGB image and its corresponding segmentation mask to emphasize important visual and spatial features. Enhanced features are fused and fed into a lightweight MobileNetV2 classifier for disease severity prediction. RCAMNet achieved a test accuracy of 96.23%, outperforming conventional raw image-based models (89.58%). Interpretability is enhanced through Grad-CAM visualizations. RCAMNet demonstrates robust performance in classifying BLB severity across diverse environmental conditions, confirming its real-world deployment potential. Additionally, the proposed framework supports the development of edge device-compatible solutions, enabling real time monitoring and improved disease management in precision agriculture.