Data-Centric Approach to SAR-Optical Image Translation

Paritosh Tiwari, Muneendra Ojha

Abstract: Image-to-image translation, especially from SAR to Optical domains, is of great importance in many Earth observation applications. However, the highly complex and unregularised nature of satellite images makes this problem challenging. Existing methods usually work on improving the underlying model to get better results and end up moving towards highly complex and unstable models tested on limited variations of the Earth’s surface. In this paper, we propose a new data-centric approach toward SAR-Optical image translation. The methodology of segregating and categorising the data according different types of land surfaces leads to improvements in image quality and simplifies the task of scaling the model and improving the results. The model is able to effectively capture and translate features unique to different land surfaces and experiments conducted on randomised satellite image inputs demonstrate that our approach is viable in significantly outperforming other baselines.