Explainable Attention-Based Plant Disease Classification Using Deep Feature Fusion and Bidirectional Neural Network for Sustainable Crop Management

Abstract

Agriculture plays a significant role in economic development, and plant diseases can substantially reduce crop quality and yield. Early and accurate detection of plant diseases is essential for improving crop management and reducing losses. This paper presents an Explainable Attention-based Plant Disease Classification framework based on a bidirectional neural network (XAPDC-BINN) for reliable plant disease detection from leaf images. Initially, bilateral filtering and CLAHE are applied for noise removal and contrast enhancement. Then, GhostNetV2, RegNet, and SECapsNet are utilized for extracting discriminative features from leaf images, followed by concatenation-based feature fusion for comprehensive feature representation. Adam optimizer is incorporated to improve feature learning performance. Further, an attention-based bidirectional temporal recurrent network is employed for disease classification. To improve model interpretability, Grad-CAM++ is utilized to highlight disease-affected regions. Experiments on the selected plant disease datasets show that the proposed XAPDC-BINN model achieves competitive classification performance compared with the evaluated approaches.