SAR-to-optical image translation with UNSB-FFC considering shadow area pixels

Deep-learning-based image translation based on high-resolution synthetic aperture radar (SAR) images is investigated as a method of restoring cloud-occlusion areas in optical satellite images. To train SAR and optical images, a Fast Fourier Convolution block was added to the unpaired image-to-image translation via the neural Schrödinger bridge (UNSB) algorithm, an unsupervised learning model for translating high-resolution images. The effect of shadow-area pixels on deep-learning-based image simulations was analyzed. The analysis results specific to urban, road, forest, and water areas indicated that the proposed algorithm performed better than the existing UNSB algorithm, and improved performance evaluation results were obtained when shadow correction was conducted. The proposed method utilizes the shadow correction effect to improve the learning model's performance between optical and radar images, and to increase the accuracy of the results of converting high-resolution radar images into optical images.