A unified method of cloud detection and removal robust to spectral variability

Abstract

This paper proposes a unified method to detect a cloud cover and to remove thin clouds from multispectral satellite images. Unlike conventional methods, the variability of a cloud spectrum is taken into account. To estimate multiple spectra of a cloud, the method first identifies probable cloud pixels and then forms their clusters each of which has a representative spectrum, both based on spatial-spectral properties of a cloud. A spectral unmixing technique is employed to determine the extent of spectral contamination by clouds. A cloud cover consisting of thick and thin clouds is thereby detected and then thin clouds are removed based on a physical model of radiative transfer. Evaluation results demonstrate that the proposed method detects a cloud cover most appropriately among well-known conventional methods, and that radiometric accuracy of thin cloud removal is improved by on average 22% compared to one of the state-of-the-art methods.