Temporal unmixing-based cloud removal algorithm for optically complex water images

Abstract

ABSTRACT Remote sensing has become an essential technique in water environment monitoring. However, the cloud contamination of remote sensing images breaks its spatial and temporal continuity in practical application. The current cloud removal algorithms always based on the hypothesis that the ground surface has no significant changes temporally, which is not proper for optically complex water body. Therefore, to overcome the weakness of current cloud removal algorithms on temporally fast-changing landscapes, a temporal unmixing-based cloud removal (TUBCR) algorithm is proposed to estimate the under-cloud information of optically complex water images. In this algorithm, the spatial information and temporal information were separated into temporal abundance images and temporal endmembers. By extracting the high temporal resolution endmembers, the fast-changing water information were estimated. The performance of the proposed algorithm was compared with those of two widely used cloud removal algorithms: the modified neighbourhood similar pixel interpolator (mNSPI) algorithm and the weighted linear regression (WLR) algorithm. The results of four simulated datasets (500 images in total) and two real Geostationary Ocean Colour Imager (GOCI) datasets (8 images in total) show that the proposed algorithm performed better in both quantitative indexes and visual effect compared with the mNSPI algorithm and the WLR algorithm. In satellite image time series, the proposed algorithm could accurately estimate under-cloud water information and correct the offsets of the statistical temporal trends that caused by cloud contamination. Hence, theoretically, the proposed TUBCR algorithm altered the similar pixel-based strategy of classic methods and got a satisfied performance in fast-changing water images. In applications, the TUBCR algorithm has great potential to improve the spatiotemporal coverage of optically complex water remote sensing.