This paper presents a model based on Vision Transformer for black Sigatoka detection in early stages, which is called ViTSigat. A comparative of the performance with other proposed model based on Convolutional Neural Network is performed. As first step, a pre-processing is applied to a dataset of 98 videos, which were recorded from a banana plantation using a mobile phone with high resolution camera. The obtained images from each video generated a total of 1500 images, which were normalize and divided in four stages/categories. Different setting parameters were used to train ViTSigat and CNN models. Likewise, both models used the same optimizer, loss function and learning rate. The obtained results were evaluated considering the accuracy, precision, recall and f1-score metrics, including the confusion matrix and ROC curves. A saliency map is used to show the relevant areas where the leaves could be affected due to the infection by black Sigatoka. The experimental results of the proposed models show that the ViTSigat model based on transformer encoder obtains better performance since its attention modules focus on important features of input data instead of take attention to non-useful information.