Vision Transformer-Based Rice Leaf Disease Classification with Imbalanced Data Handling Using Image Augmentation

Abstract

Image-based classification of rice leaf diseases helps identify issues quickly and accurately, but data imbalance poses a challenge in deep learning models. This study analyzes the effect of handling imbalanced data using a Vision Transformer-based image augmentation strategy on the Paddy Doctor dataset which contains 10,407 images across 9 disease classes and 1 healthy class. Experiments tested five scenarios including no augmentation, flipping, rotate, cropping, and noise injection, evaluated using accuracy, precision, recall, and F1-Score metrics. Results show the scenario without augmentation provides the best performance with 96.29% accuracy and a 0.9588 macro F1-score on test data. All augmentation techniques decrease model accuracy by 1.41% to 1.94%. Among the tested techniques, cropping gave the best results while rotate produced the lowest performance. This research concludes that generic image augmentation does not always improve Vision Transformer performance because its effectiveness heavily depends on visual characteristics and data distribution.