A Lesion-confidence Driven Adaptive Global-local Deep Framework for Explainable Cotton Leaf Disease Classification

Early detection, identification and interpretation of cotton leaf disease are critical to reduce loss of the cotton crop, and can assist in decision-making for farming systems. In this regard, there exist several deep learning techniques that have been able to classify plant diseases with success, but most of the available methods have been based primarily on global image feature. To more fully capture these subtle disease characteristics, an global-local adaptive approach to cotton leaf disease classification is presented. The proposed framework uses two-branch architecture: One global contextual feature extractor and one local extractor for lesions. The global branch is utilized in capturing the context information from the entire leaf image, whereas the local branch is utilized in capturing the details of the lesions. A lesion-confidence adaptive fusion mechanism is added to adaptively adjust the contribution of features based on the lesion prominence and spatial distribution of the disease. In addition, the framework is complemented with augmentation based on lesions, where training is concentrated on diseases-relevant region. Unlike conventional static fusion approaches, the proposed adaptive mechanism performs sample-dependent feature weighting, enabling improved discrimination of morphologically similar disease categories. The key parameters are optimized with genetic algorithms and Grad-CAM combined with a consistency metric are used to enhance interpretability and obtain more accurate visual explanations of model decisions. The performance was measured with the dataset SAR-CLD-2024 in a classification context of seven classes. Experimental evaluation on the SAR-CLD-2024 dataset demonstrates that the proposed framework achieves 97.0% classification accuracy and 96.4% F1-score, outperforming EfficientNet-B0 by 1.5% in accuracy. The ablation experiments confirmed that all three factors (local branch, adaptive fusion mechanism and optimization strategy) had an impact on overall performance. Overall, the results indicate that the system is reliable and interpretable in achieving fine-grained classification of cotton leaf disease in practical field settings. Additional robustness analysis, ablation experiments, and explainability validation confirm the effectiveness of lesion-guided adaptive feature integration for fine-grained agricultural disease recognition.