Cloud removal in remote sensing images using generative adversarial networks and SAR-to-optical image translation

Satellite images are often contaminated by clouds. Cloud removal has received\nmuch attention due to the wide range of satellite image applications. As the\nclouds thicken, the process of removing the clouds becomes more challenging. In\nsuch cases, using auxiliary images such as near-infrared or synthetic aperture\nradar (SAR) for reconstructing is common. In this study, we attempt to solve\nthe problem using two generative adversarial networks (GANs). The first\ntranslates SAR images into optical images, and the second removes clouds using\nthe translated images of prior GAN. Also, we propose dilated residual inception\nblocks (DRIBs) instead of vanilla U-net in the generator networks and use\nstructural similarity index measure (SSIM) in addition to the L1 Loss function.\nReducing the number of downsamplings and expanding receptive fields by dilated\nconvolutions increase the quality of output images. We used the SEN1-2 dataset\nto train and test both GANs, and we made cloudy images by adding synthetic\nclouds to optical images. The restored images are evaluated with PSNR and SSIM.\nWe compare the proposed method with state-of-the-art deep learning models and\nachieve more accurate results in both SAR-to-optical translation and cloud\nremoval parts.\n