<p>The evolution of transformer-based deep learning architectures has significantly transformed the landscape, which was designed to address limitations in Recurrent Neural Networks (RNNs). Unlike RNNs, transformers utilize self-attention mechanisms, allowing them to process input sequences in parallel, which significantly reduces training time and improves efficiency. The original transformer architecture work with two fundamental stages called encoder (to process input data) and decoder (to generate output). Each component is made up of multiple layers, each containing mechanisms for self-attention and feed-forward neural networks. This design enables the transformer to map complex dependencies within the processed data. Although transformers have been developed under research for Natural Language Processing (NLP) that later adapted to perform the tasks associated with several other fields of engineering. One of the popular adaptations is Vision Transformer (ViT) that has been adapted to perform image classification task. ViT has shown that transformer architectures perform better than traditional Convolutional Neural Network (CNNs). Keeping this view of importance of transformer, we have applied ViT for plant-leave disease detection on mango leaf dataset. Hence, used ViT for plant disease detection with multiscale patches reaches to 97.38% of correct detection and model throughout the paper referred as ViT-PDD. Apart from overall accuracy, we have also investigated top-5-accuracy metrics to assess well classified and worst classified class of disease in mango leaves along with other important metrics such as precision, recall, and f1-score. A performance comparison with published works illustrated on the same dataset has been included in the results section.</p> Graphical abstract <p></p>