AN EXPLAINABLE ATTENTION–TRANSFORMER HYBRID FRAMEWORK FOR INTELLIGENT COTTON LEAF DISEASE

Abstract

Cotton is one of the most important commercial crops worldwide, and its productivity is significantly affected by leaf diseases and pest infestations. Early and accurate disease diagnosis is essential for reducing crop losses and improving agricultural sustainability. This study proposes an explainable Attention–Transformer hybrid framework for intelligent cotton leaf disease and pest classification. The framework integrates VGG16 and VGG19 feature extraction networks with Convolutional Block Attention Module (CBAM), Squeeze-and-Excitation (SE) attention, Transformer encoders, and BiLSTM-attention mechanisms to enhance discriminative feature learning and contextual representation. Experiments were conducted on a publicly available cotton leaf disease dataset containing seven disease and healthy leaf categories. Comprehensive evaluation using Accuracy, Precision, Recall, F1-Score, AUC-ROC, and Cohen’s Kappa demonstrates that the proposed architectures consistently outperform conventional VGG baselines. Among all evaluated models, the VGG16-SE-Transformer achieved the best test accuracy of 95.8% with an AUC-ROC of 0.998. Furthermore, Grad-CAM visualization provides interpretable disease localization, improving model transparency and practical applicability for precision agriculture