An Ensemble Framework Integrating CNNs and Vision Transformers for Cotton Leaf Disease Classification

Abstract

Cotton is a major cash crop in the world, and its productivity is highly influenced by various diseases of the leaf, resulting in significant economic losses. Therefore, early and accurate identification of cotton leaf diseases is essential for improving crop management and increasing crop yield. This paper presents an optimized deep learning (DL) model to automatically classify cotton leaf diseases with the help of image data. The ensemble model classifies five categories of cotton leaves, namely, Alternaria Leaf, Bacterial Blight, Fusarium Wilt, Verticillium Wilt, and Healthy Leaf. Training and evaluation were performed on 4,963 real-field cotton leaf images. The proposed model demonstrates superior performance with an overall classification accuracy of 98.39% and high precision, recall, and F1-score across all classes. Gradient-weighted Class Activation Mapping (Grad-CAM) is used to visualize the discriminative regions that cause the classification decisions thereby improving model interpretability and reliability. Grad-CAM results demonstrate that the model focuses attention to disease-relevant characteristics of lesions, discoloration, and texture changes instead of background data. The model highlights good accuracy and interpretability, making it adequate for precision agriculture and automated systems for diagnosing cotton diseases.