Edge‐Optimized Sugarcane Disease Detection Using CLIP ‐Distilled Swin Transformer and Grad‐ CAM Visualization

ABSTRACT

Early and accurate diagnosis of plant diseases is crucial to ensure the security of crop production, food security, and economic viability. One of the most critical cash crops in the world is sugarcane, that is very susceptible to fungal and bacterial diseases like “red rot” and “rust” which can cause significant losses in yield if not detected. While traditional machine learning and deep learning approaches have success in automated disease diagnosis, they typically involve complex models that are impractical to efficiently run on farmer's mobile or edge devices. Furthermore, in practical applications, the accuracy of the classification is often affected by noise in the field scene, changing lighting conditions and complicated background scenes, which further restricts the field applicability. To tackle these challenges, we introduce a new framework in which we use a fuzzy c‐means (FCM) pre‐segmentation approach and transfer knowledge from a CLIP ViT‐B/32 teacher model to a lightweight Swin Transformer student model. The FCM step extracts lesion sensitive regions, which result in the decrease of noise and enhancement of feature extraction relevant to the lesions. The main idea of Knowledge Distillation is to pass semantic rich and robust representation of the large‐scale CLIP teacher model to the lightweight MobileViT student model for improving performance without losing efficiency. The proposed model is tested on a publicly available sugarcane dataset and custom field collected dataset and performances are presented under various classification metrics, localization accuracy of lesions and deployment performance in edge devices. Experimental results show that our approach achieves 5%–8% higher robustness in terms of F1‐score compared to baseline MobileViT, a 12% higher lesion localization IoU and real‐time inference (< 100 ms) on mid‐range Android devices with a smaller model size (< 40 MB). The results demonstrate the feasibility of this approach as a viable and scalable solution for smart agriculture, where farmers can make informed decisions about preventing disease on the farm. This research is a step toward enhancing smart, farmer‐oriented systems for crop protection by combining research‐level precision with field deployment efficiency.