Cloud Removal of Optical Remote Sensing Imageries using SAR Data and Deep Learning

Satellite optical images are useful for creating updated land cover maps but suffer the cloud contamination problem with region below the clouds are not mapped. On the other hand, the Synthetic Aperture Radar (SAR) imageries are all-weather sensor able to see through cloud but with different sensing properties and normally poorer resolution. In this work, a SAR data-based deep learning recovery approach is proposed and explored to remove cloud and reconstruct realistic ground mapping. The model utilized an autoencoder structure which intakes the SAR image and other optical features to provide a cloud free edge map, later use the edge map, together with the SAR image, as a guide to an inpainting model, which provides a reasonable cloud-free image for both human perception and industrial application. Numerical examples show that the proposed method works well with Sentinel dataset. With the resolution of SAR data improving rapidly through recent years, using SAR imageries to enhance data from other sources could be a more promising topic for research as well as practical applications.