Cloud removal in high resolution multispectral satellite imagery: comparing three approaches

Abstract

Clouds and cloud-shadow are a persistent problem in all optical satellite imagery. Plenty of methods have been suggested in the literature to address this problem, and reconstruct the missing part of the optical signal. In this work, three methods representative of different approaches to the cloud removal problem are compared. The first method is temporal fitting using Fourier series, which benefits from the temporal continuity of the signal. The second method uses sparse spectral unmixing to fill in the missing areas. The third method employs radiometric consistency as a tool to determine the missing part of the optical signal. These three methods are first presented and their theoretical background described, followed by a discussion of their implied assumptions, general performance, and failure modes. A set of experiments using Landsat 8 time series with diverse land cover types were conducted. The quantitative results of the three methods using simulated clouds as well as real ones are presented. Finally, some concluding remarks about the relative advantages of the three approaches are listed, in addition to some recommendations about their use.