Robust Apple Leaf Disease Classification Using Vision Transformer and Data Augmentation Techniques

Vision Transformer has been one of the most exciting developments in precision agriculture in recent years. It is beneficial to improve the performance for more tasks with few datasets through knowledge transferred from the pre-trained models. In this work, we examined the efficiency of ViT in classifying apple leaf diseases. Diseases were labeled into five major classes: Alternaria, Apple Scab, Black Rot, Cedar Apple Rust, and Healthy. To increase the model's robustness and evade possible overfitting, it incorporates augmenting the data using rotation, zoom, flip, and brightness adjustment. We applied the proposed ViT to the dataset at hand from Mendeley, where it achieved an accuracy of 99.878%, quite above the accuracy of no augmented models at 97.11%. To ensure the model was effective in capturing subtle patterns of disease manifestation, we did an extensive evaluation of its performance through precision, recall, and F1-score. The confusion matrices and ROC curves established the disease classification capability of the ViT model; further, they achieved high accuracy for the more challenging diseases like Black Rot and Alternaria. The result of increased generalization from an augmented dataset was such that the model could also perform well in changing conditions, which again established the principle of data augmentation during the training of deep learning models in agricultural applications. In particular, this work presents the ability of the ViT to lead to real-time and accurate disease detection, providing a very valuable development in the process of precision agriculture for apples and in conducting their disease management accordingly.