Precision cotton disease detection via transformer models applied to leaf imagery

There is great potential for improving agricultural research, ecological monitoring, and biodiversity conservation through computerized plant species cataloging utilizing leaf photos. This work introduces a deep learning-based framework that uses transformer-based architectures, such as the Vanilla Vision Transformer (ViT), Swin Transformer, DeiT (Data-Efficient Image Transformer), and T2T-ViT (Tokens-to-Tokens Vision Transformer), to automatically classify cotton leaf diseases. Images of cotton leaves from four different classes-curl virus, bacterial blight, fusarium wilt, and healthy leaves-make up the dataset. A stratified K-fold hold-out testing technique (K = 1 to 5) is used to maintain the class distribution across training and testing folds in order to guarantee robust model evaluation and address class imbalance. To improve generalization and guarantee compatibility with transformer models, standard image augmentation and normalizing approaches are used. All models begin training using vast collections of images, afterward honed specifically on cotton leaf data to sharpen their ability to tell differences apart. Results spread across multiple test rounds stay steady, one standout reaching nearly perfect accuracy-99.99 percent. This pattern highlights how transformer-driven systems thrive alongside stratified K-fold checks, crafting a dependable way to spot crop issues early, shifting farm oversight toward quicker, smarter responses.