Data-Efficient Wheat Disease Detection using Shifted Window Transformer: Enhancing Accuracy, Sustainability, and Global Food Security

Wheat is an essential crop that plays a vital role in global food security, but is susceptible to a variety of diseases, which have the potential to drastically decrease crop productivity. Detection of disease at an early stage and in an accurate manner is crucial to minimize crop losses. This research presents a deep learning technique based on the Shifted Window (Swin) Transformer, a powerful attention-based model that effectively captures both local and global information for enhanced classification output. Unlike conventional CNN-based methods, which often face limitations in feature extraction, the Swin Transformer utilizes hierarchical attention mechanisms to improve disease detection accuracy. The proposed modelis trained on a dataset of 9,346 wheat leaf images, categorized into eight disease classes and one healthy class. Using Bayesian hyperparameter optimization, we tuned key parameters such as learning rates, batch sizes, and dropout rates, and achieved an accuracy of 99.3%. We also conducted comparative analyses with baseline self-attention, CNN-based feature extraction, and hybrid attention layers. To enhance interpretability, Grad-CAM visualization techniques were applied, confirming model reliability. This research advances precision agriculture by improving the efficiency of wheat disease identification and supporting sustainable farming through state-of-the-art deep learning methodologies.