Multiscale sugarcane disease classification using EfficientNet transformer fusion

Early monitoring of sugarcane leaf diseases (SLD) is important to reduce yield losses and promote agricultural sustainability. Nevertheless, images captured in the real world tend to contain noise, such as illumination variation, blur, occlusion, or compression artifacts, which leads to less accurate classification. In this work, the VTransNet-E model is proposed. The hybrid CNN–Transformer architecture extracts and fuses feature from different scales using EfficientNet-V2 and cross-scale attention. Shallow, mid-level, and deep features are tokenized and fused with a lightweight Transformer to learn fine-grained local features and global contextual. A domain-randomized augmentation pipeline and robust hybrid loss functions enhance the proposed model’s ability to withstand real-world distortions. VTransNet-E is validated on a public dataset for sugarcane leaf disease and achieves 98.34% accuracy and an F1-score of 98.12%, surpassing CNN-only, ViT, and recently proposed ensembles. The ablation studies further validate the effectiveness of multi-scale (MS) fusion, cross-scale encoding, and noise-aware training.