Enhancing sugarcane leaf disease classification through a novel hybrid shifted-vision transformer approach: technical insights and methodological advancements.

Abstract

In the agricultural sector, sugarcane farming is one of the most organized forms of cultivation. India is the second-largest producer of sugarcane in the world. However, sugarcane crops are highly affected by diseases, which significantly affect crop production. Despite development in deep learning techniques, disease detection remains a challenging and time-consuming task. This paper presents a novel Hybrid Shifted-Vision Transformer approach for the automated classification of sugarcane leaf diseases. The model integrates the Vision Transformer architecture with Hybrid Shifted Windows to effectively capture both local and global features, which is crucial for accurately identifying disease patterns at different spatial scales. To improve feature representation and model performance, self-supervised learning is employed using data augmentation techniques like random rotation, flipping, and occlusion, combined with a jigsaw puzzle task that helps the model learn spatial relationships in images. The method addresses class imbalances in the dataset through stratified sampling, ensuring balanced training and testing sets. The approach is fine-tuned on sugarcane leaf disease datasets using categorical cross-entropy loss, minimizing dissimilarity between predicted probabilities and real labels. Experimental results demonstrate that the Hybrid Shifted-Vision Transformer outperforms traditional models, achieving higher accuracy in disease detection of 98.5%, making it crucial for reliable disease diagnosis and decision-making in agriculture. This architecture enables efficient, large-scale automated sugarcane disease monitoring.

Code and data availability

The paper's sugarcane leaf disease classification experiments use three public Kaggle image datasets (SLI, RRSDL, SLDC). The SLDC dataset is cited with a Kaggle URL matching an allowed URL, making it a paper-specific, publicly actionable phenotyping image asset. No author analysis code, trained model checkpoints, or on