RSI-Mix: Data Augmentation Method for Remote Sensing Image Classification

Data augmentation is a common strategy to improve the performance of computer vision tasks. Regrettably, current data augmentation methods are often designed for images in RGB format and few are studied for remote sensing images. In this paper, we find that the way in which remote sensing data are obtained provides a realistic possibility for cropping images to create new ones. Based on our analyses, we propose RSI-Mix for Sentinel-2 satellite image classification. RSI-Mix is designed to cut and paste two remote sensing images of the same category according to random masks. The key inspiration of RSI-Mix is that the classification of remote sensing images is not strictly based on image texture but based on band features. The information fusion of the same area from different sources is beneficial to make an area contain more band features. Experiments show that the model with RSI-Mix is more stable and has higher performance.