Cloud removal in image time series through unmixing

Satellite Image Time Series (SITS) are sets of images acquired by the same or different spaceborne sensors over the same area at different acquisition times. In recent years, SITS acquired by multispectral sensors have increased, due to the large availability of these kind of data and the decrease in revisit time. The latter aspect is also a consequence of the advent of multispectral satellite constellations such as RapidEye and Pleiades, with the Sentinel 2 mission being able to derive such products at global scale in the near future. The high dimensionality of SITS unlocks the use of methods usually restricted to the processing of hyperspectral images, typically characterized by hundreds of narrow, contiguous spectral bands. This paper proposes a new cloud removal algorithm for scenes within a SITS based on the concept of spectral unmixing, which decomposes hyperspectral image elements into fractional abundances of reference spectra related to the materials present in a given scene. In the case of SITS, we consider a reference spectrum as a pixel which presents a characteristic evolution in time of its spectral features. While traditional methods aim at substituting the affected pixels with other ones taken from other scenes, the proposed algorithm allows exploiting the full spectral and temporal information of the SITS to synthesize the expected values below a cloudy area. This is done independently from 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.