Swin Transformer algorithm for Multiple Disease detection on Paddy Crops: A hierarchical Vision Method with shifted window

Abstract

The health monitoring of paddy is a significant research issue since outbreaks of diseases may decrease the yield, quality of grains, and the profitability of the farms. The process of field diagnosis is time-consuming and can be inaccurate at times, which drives automated image-based recognition systems. This paper explores a hierarchical vision architecture that is founded on the Swin Transformer to detect multiple paddy health-classes with shifted-window attention. The experiments were done on a publicly available 10-class paddy leaf image dataset with 10,407 labeled images that were stratified into 7,280 training images, 1,557 validation images and 1,570 test images. To overcome the issue of class imbalance and enhance generalization, a transfer-learning setup of Swin-Tiny was trained using weighted cross-entropy loss, AdamW optimization, and data augmentation. The final model had the best validation accuracy of 96.98% and the final test accuracy of 96.50%. The model also achieved weighted precision, weighted recall, and weighted F1-score values of 96.56, 96.50, and 96.51, respectively, and a macro F1-score of 96.22. Analysis by classes indicated very high performance in dead heart, bacterial panicle blight, hispa and normal classes whereas downy mildew was the most difficult category. The results suggest that a hierarchical vision approach with Swin Transformer can offer a high level of competitiveness in multi-class paddy leaf health recognition.