In this paper, we introduce a novel application domain which is first of its kind to the vision transformer based deep learning model. We propose a model for ripeness grading of mangoes using vision transformers. Our approach divides the mango image into patches, which are then linearly projected and transformed into a sequence of embeddings. To retain positional information, positional encodings are added to these patches. Additionally, for image classification, learnable class tokens are included at the start of this sequence of embeddings. The resulting sequence is passed through multiple multi-head self-attention (MSA) layers to capture both local and global dependencies and to interpret the spatial relationships among patches. Further to improve the classification performance, we explored five data augmentation strategies to synthetically induce additional data for training. Moreover, different vision transformer models are investigated with and without pre-trained weights while training of neural network. This study is demonstrated through an experimentation on a dataset of 979 images of Alphonso mango variety belonging to four classes particularly, unripen, ripened, over-ripened without internal defects and over-ripened with internal defects. The vision transformers viz., ViT_Base_16, ViT_Large_16 and ViT_Huge_14 is considered for experimentation. The results of the experimentation demonstrated that, the ViT_Huge_14 with pre-trained weight and with data augmentation gives average accuracy of 92.78%, which is better than 85.40% quoted in the existing work of mango grading using conventional machine learning on the same dataset (Raghavendra et al., 2020).