Abstract:
Eggplant is one of the most commercially important crops in tropical and subtropical regions
such as Bangladesh, India, and Pakistan. However, several leaf diseases significantly reduce
yield and quality. Traditional disease detection methods are manual, time-consuming, and
dependent on expert knowledge, which delays timely intervention. Although machine learning
and deep learning approaches like CNN-SVM hybrids, VGG networks, and ResNet-based
techniques have been explored, they often face challenges such as poor real-world accuracy,
limited generalizability, and high computational costs. This study proposes a Vision
Transformer (ViT)-based approach for automated eggplant leaf disease identification to
address these issues. ViT, a state-of-the-art architecture, captures both local and global image
patterns, making it effective for identifying subtle disease symptoms. We used a dataset of
about 9,800 eggplant leaf images classified into seven categories: Wilt, Mosaic Virus, Small
Leaf, Insect Pest, Leaf Spot, White Mold, and Healthy. The ViT model, trained with
preprocessing and augmentation techniques, achieved 98.16% testing accuracy. For
comparison, transfer learning models ResNet50 (91.53%), EfficientNetB5 (91.33%), VGG19
(86.63%), and VGG16 (87.55%) were evaluated, but ViT outperformed them in accuracy and
efficiency. The findings suggest that ViT offers a more reliable and scalable solution for
precision agriculture, enabling early detection, reduced crop loss, sustainable production, and
improved food security.
