Cloud Removal in Image Time Series Through Sparse Reconstruction From Random Measurements

In this paper, we propose a cloud removal algorithm for scenes within a satellite image time series based on synthetization of the affected areas via sparse reconstruction. The high spectrotemporal dimensionality of time series allows applying pixel-based sparse reconstruction techniques efficiently, estimating the values below a cloudy area by observing the spectral evolution in time of pixels in cloud-free areas. The process implicitly compensates the overall atmospheric interactions affecting a given image, and it is possible even if only one acquisition is available for a given period of time. The dictionary, on the basis of which the data are reconstructed, is selected randomly from the available image elements in the time series. This increases the degree of automation of the process, if the area containing clouds and their shadows is given. Favorable comparisons with similar methods and applications to supervised classification and change detection show that the proposed algorithm restores images locally contaminated by clouds and their shadows in a satisfactory and efficient way.