Comparison of SAR-Optical translation frameworks based on Generative Adversarial Networks (GAN)

Abstract

The potentials of multitemporal SAR data acquired from Sentinel-1 data have been reported for forest mapping and managing forest. However, SAR data have difficulty in visual interpretation due to speckle noise and geometric distortion caused by the distance-dependence along the range axis and the characteristics of radar signal wavelengths. Recently, the advantages of Generative Adversarial Networks (GANs) have been reported in image-to-image translation and there could be also effective methods for SAR-Optical image translation. In the current study, comparisons among CycleGAN, pix2pix, pix2pixHD and feature-guided SAR-to-optical image translation (FGGAN) were performed using Sentinel-1 and 2 data acquired in the forests. As a result, FGGAN was the best technique, achieving the structural similarity values of 0.664, 0.708 and 0.725 for red, green and blue bands, respectively.