A Comprehensive Analysis of Leaf Disease Classification: Revolutionizing Detection with Vision Transformer

Leaf diseases present a significant threat to agricultural productivity and crop quality, necessitating swift and precise identification methods to mitigate their detrimental effects and avert economic loss. Conventional diagnostic approaches reliant on manual visual examination are inefficient and susceptible to inaccuracies. Conventional deep learning architectures often struggle with a limited receptive field, a focus on local textures over global structures, and difficulties in recognizing long-range dependencies, hindering their effectiveness in detecting complex disease patterns on leaves. Considering these challenges, we proposed a plant disease classification system that leverages the pre-trained Vision Transformer (ViT), a cutting-edge deep learning architecture renowned for its proficiency in visual data analysis. By successfully capturing long-range dependencies and improving feature representation, the ViT overcomes the drawbacks of conventional models improving leaf disease classification accuracy and robustness. Our methodology encompasses the preprocessing of leaf imagery through resizing, augmentation, and normalization techniques to enhance the input quality. The ViT model employs a patch-based segmentation approach and utilizes self-attention mechanisms to capture both localized and global image features effectively. Using the Major Plant Disease Detection Dataset from Kaggle, our proposed ViT model achieved a test accuracy of $\mathbf{9 8. 1 4 \%}$.