An Improved Tiny Vision Transformer-based Approach for Explainable Apple Leaf Disease Classification

Apples are one of the most important and widely cultivated fruits around the world, and they hold significant economic value in many regions. However, apple production and fruit quality are greatly affected by leaf diseases such as Rust and Scab. Therefore, early and accurate detection of apple leaf diseases is essential for sustainable orchard management and effective disease prevention. Traditional inspection methods are slow, labor-intensive, and prone to human error, which highlights the need for automated and reliable solutions. This paper proposes an optimized Tiny-ViT-5M-224 model for three-class apple leaf disease classification (Healthy, Rust, Scab) using a Kaggle dataset of 1,730 images. The proposed model is based on the combination of transfer learning and domain specific preprocessing, structured data augmentation, and regularization techniques, which allows it to achieve strong generalization at a low computational cost. Tiny-ViT achieves a test accuracy of 99.23 % and means 5-fold cross-validation accuracy of 98.38 % which is way better than other evaluated lightweight transformer baselines. The model also has strong statistical reliability as its Matthews correlation coefficient is 0.9884, Cohen kappa is 0.9901, and the 95 % interval is small$([0.9811,0.9981])$. Notably, Grad-CAM visualizations indicate that biologically relevant lesion areas of the prediction are directed by artifacts of the background instead of biologically important areas, thus increasing transparency and credibility. Collectively, these results show that Tiny-ViT is a reliable, explainable and practical object detector of precision agriculture, demonstrating how tiny transformer designs can also be used to achieve high performance in detecting plant diseases in resource-limited conditions.