Double Rank-One Prior: Thin Cloud Removal by Visible Bands

Thin cloud removal for multispectral (MS) or hyperspectral (HS) images is a ubiquitous and fundamental problem in remote sensing. However, it is prohibitively challenging due to the ill-posedness and underdetermination of the image formation. The existing methods for cloud removal are dominated by either cirrus detection bands from satellite sensors or large-scale datasets from training. In this article, by vectorizing the MS/HS image of each band and stacking them columnwisely as a matrix, we develop a double rank-one prior (DROP) method for thin cloud removal. By leveraging some visible bands in satellite sensors, the proposed method is composed of cloud retrieval and removal phases. In brief, we first separate clouds from the observed image by formalizing the top-of-atmosphere (TOA) reflectance and then evaluate the thickness of clouds in each band by the independence of cloud-free and cloud images. Numerical experiments on various Landsat-8 and Sentinel-2 images demonstrate the compelling performances of the proposed method.