Abstract
Cloud removal from satellite imagery is a well-known problem in both remote sensing and deep learning. Many methods have been developed to address the cloud removal problem in a supervised setting. These methods require gathering of huge datasets to learn the mapping from cloudy images to cloud-free images. In this paper, we address cloud removal as an inverse problem. We convert this problem into an inpainting task in which cloudy regions are treated as missing pixels and completely solve the cloud removal problem in an unsupervised setting. We show that the structure of a network is a good prior by itself and is sufficient to remove clouds from satellite imagery using Deep Image Prior algorithm. Experimental results on Sentinel-2 Imagery have quantitatively and qualitatively demonstrated the effectiveness of our technique on a diverse range of clouds. KeywordsDeep image priorCloud image priorConvolutional neural networksInpaintingU-NETMeshgridPeak signal-to-noise ratioStructural similarity index measure
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have