An adaptive dual-branch deep learning framework achieved 97.0 percent classification accuracy and a 96.4 percent F1-score in identifying cotton leaf diseases across seven categories on the SAR-CLD-2024 benchmark dataset. The evaluated architecture outperformed the baseline model EfficientNet-B0 by 1.5 percent in classification accuracy.
Early detection, identification, and interpretation of cotton leaf disease are critical for reducing crop loss and assisting decision-making within agricultural farming systems. While several deep learning techniques have previously succeeded in classifying plant diseases, most available methods rely primarily on global image features, which struggle to capture subtle diagnostic characteristics.
To capture localized nuances alongside broader leaf structures, the researchers developed an adaptive architecture operating two parallel processing paths. The framework incorporates a global contextual feature extractor to capture context from the entire leaf image, alongside a dedicated local extractor focused specifically on lesion details.
Integration between these paths is managed by a lesion-confidence adaptive fusion mechanism that dynamically adjusts feature contributions based on lesion prominence and the spatial distribution of the disease. Unlike conventional static fusion approaches, this adaptive mechanism performs sample-dependent feature weighting, improving discrimination among morphologically similar disease categories.
The training protocol complements the dual-branch network with a lesion-based augmentation strategy designed to concentrate model learning on disease-relevant regions. In addition, the computational framework employs genetic algorithms to optimize key operational parameters across the system.
To provide clear visual explanations of model decisions, the architecture pairs Grad-CAM with an explicit consistency metric, enhancing the overall interpretability of automated classifications.
Structured ablation experiments confirmed that the local branch, the adaptive fusion mechanism, and the optimization strategy each contributed measurably to overall performance. Additional robustness analyses, ablation trials, and explainability evaluations confirmed the effectiveness of lesion-guided adaptive feature integration for fine-grained agricultural disease recognition in practical field settings.
Subject of Research: Deep learning classification of cotton leaf diseases
Article Title: A Lesion-confidence Driven Adaptive Global-local Deep Framework for Explainable Cotton Leaf Disease Classification
Article References: Ali, F., Khaskheli, Z. H., Muhammad, D., Elkiran, H., Rasheed, J., Alqhtani, S. M., Shaikh, A., & Waheed, Z. (2026). A Lesion-confidence Driven Adaptive Global-local Deep Framework for Explainable Cotton Leaf Disease Classification. International Journal of Computational Intelligence Systems. https://doi.org/10.1007/s44196-026-01686-1
Image Credits: AI Generated
DOI: 10.1007/s44196-026-01686-1
Keywords: cotton leaf disease, plant disease classification, deep learning, global contextual features, lesion extraction, adaptive fusion mechanism, lesion-confidence weighting, SAR-CLD-2024 dataset, EfficientNet-B0, genetic algorithms, Grad-CAM, consistency metric, model interpretability, ablation experiments, agricultural disease recognition, crop loss reduction, farming decision support, lesion-based augmentation
News Source: Denise Maddox. (November 17, 2026). Adaptive Deep Framework Identifies Cotton Leaf Diseases With 97 Percent Accuracy. Scienmag.



