Satellite Image Classification Through Stable Diffusion and Vision Transformers

Various satellite images are used for scientific purposes; however, their availability is limited. To solve this problem, data augmentation is used. It is a widely used method to decrease model overfitting by increasing the number of training examples. Our goal is to increase the existing dataset while maintaining its credibility. For this, we have used stable diffusion to increase the dataset, followed by a vision transformer for image categorization. Vision transformers, in contrast to traditional techniques, divide images into patches and allow spatial connections, allowing data augmentation to expand the dataset via stable diffusion. This work provides insights into the technology architecture and training process and establishes a strong foundation for upcoming developments in the field by demonstrating the impact of ViT on satellite imagery interpretation coupled with stable diffusion. The efficiency of this model has been established as it achieved an accuracy of 99.16% in the SAT 6 dataset.