Early diagnosis of Sugarcane leaf diseases through CNN and Vision Transformer hybrid model

Abstract Timely and precise identification of foliar diseases in sugarcane is imperative for yield optimization and disease management. This work proposes a hybrid deep learning framework leveraging Convolutional Neural Networks (CNNs) and Vision Transformers (VITs) for automated multi-class classification of sugarcane leaf diseases, including healthy, yellow rust, mosaic, rust, and red rot. Initially, baseline CNN architecture was employed to extract spatially localized features, attaining a classification accuracy of 84.3%. Subsequently, a pre-trained VIT model, capable of modelling long-range dependencies through self-attention mechanisms, was fine-tuned on the same dataset, achieving 93.07% accuracy. To further enhance feature representation, a hybrid CNN + VIT model was constructed by integrating CNN-based local feature encoders with VIT-based global context modelling. The proposed ensemble architecture achieved a superior accuracy of 97.43%, demonstrating robust generalization and discriminative power. The results affirm the efficacy of transformer-based architectures in plant disease detection tasks and validate the synergy between convolutional and attention-based models for high-resolution agricultural image analysis.