ACCM-Net: an adaptive CLAHE-CBAM-MobileNet architecture for wheat leaf disease classification

Wheat disease classification is critical for early detection and crop management, but manual identification is challenging due to visually similar symptoms in early infection stages. Existing deep learning approaches are not adept at addressing class imbalance, complex lesion patterns, and non-uniform field conditions, and thus, practical use is obstructed. To address this, this paper presents ACCM-Net, a fusion architecture founded on a frozen MobileNetV2 backbone and Convolutional Block Attention Module (CBAM) for enhanced feature refinement, combined with adaptive CLAHE and Sobel-based preprocessing to emphasize disease-discriminative regions. The model incorporates dynamic learning rate scheduling for stable training and enhanced generalization. Evaluated on a Kaggle-sourced dataset of 407 wheat leaf images, ACCM-Net attained an overall accuracy of 97.56%, F1-score of 97.60%, sensitivity of 98.41%, and specificity of 98.92%, beating baseline CNN and five other pre-trained models. The results show that ACCM-Net's combination of contrast enhancement, attention mechanisms, and adaptive learning techniques significantly enhances classification accuracy and interpretability and, thus, is a promising practical tool for agricultural diagnostics. Future work will extend the model to other disease classes and explore transformer-based enhancements to further contextual understanding.