Hybrid CNN-Attention Framework With Texture Feature Fusion For Multi-Label Detection Of Co-Infections In Rice Leaves

Abstract

Early and accurate detection of co-infections caused by multiple diseases, pests and weeds  in rice plants is essential for minimizing yield losses and enabling timely intervention. Traditional image-based classification models often fail to capture the subtle inter-class and intra-class similarities that arise from overlapping symptom patterns. In this study, we propose a hybrid deep learning framework that integrates handcrafted feature extraction with an advanced convolutional neural network architecture for robust multi-label classification and similarity detection. The framework leverages Gray-Level Co-occurrence Matrix (GLCM) and Local Binary Pattern (LBP) to extract texture-based features from rice leaf images, which are then fused with raw image data and fed into a ResNet50 backbone enhanced with a Convolutional Block Attention Module (CBAM). The model is trained on a custom-structured dataset of rice leaf images categorized into diseases and pests, with multi-label annotations representing potential co-infections. Experimental results demonstrate the model’s capability to accurately predict co-infections with high confidence and quantify inter-class (e.g., disease–pest-weeds) and intra-class (e.g., disease–disease) similarities using learned feature embeddings. The proposed hybrid approach achieves notable improvements in classification performance, interpretability, and generalization across visually similar classes. This system offers significant potential for real-time deployment in precision agriculture, particularly in the early diagnosis and management of biotic stressors in rice cultivation.