Advanced SAR-To-Optical Image Translation Techniques using Jaxa's High-Resolution Land-Use and Land-Cover Map
Xuanchao Fu, Toru Kouyama, Suomi Seki, Ryosuke Nakamura, Ichiro Yoshikawa · 2024
The translation from Synthetic Aperture Radar (SAR) to optical imagery is vital in remote sensing. Nevertheless, this translation encompasses profound challenges because of the limited information in SAR data compared to optical images. While progress has been achieved through frameworks like pix2pix, the room for enhancement remains plentiful. We have assembled a unique dataset by harmonizing Sentinel-1 and Sentinel-2 satellite images from the SEN12MS dataset, accompanied by JAXA`s High-Resolution Land-Use and Land-Cover (LULC) Map of Japan. Our study proposes an efficient translation framework that modifies the original pix2pix architecture, by modifying the discriminator architecture to a new one that can be trained with LULC, which enables the supervision of the generator using the LULC map. Our method significantly supersedes the conventional pix2pix, indicating LULC discriminator may improve a result with a conventional one.