Cloud Masking Scheme Based on Spectral, Morphological and Physical Features

CHRIS/PROBA acquisitions are programmed in order to avoid cloud covers, but sometimes images present clouds to some extent. This paper presents the preliminary results of a novel approach to cloud detection based on the hyperspectral information contained in CHRIS Mode 1 images. After a preprocessing phase intended to reduce drop-out noise and vertical striping, physically relevant features are extracted from the scenes. A first classification stage based on thresholds is applied to obtain a mapping of the areas prone to contain cloudy pixels. A growing algorithm is applied in order to obtain a mask of the cloudy pixels that are clustered in an unsupervised way with the Expectation-Maximization algorithm afterwards. The spectral signatures of the prototypes of every cluster are compared to a library of signatures in order to correctly assign the cluster to well-defined classes. The performance of the method is tested in a series of four acquisition dates. Further work is required to assess its performance in a more generalized validation set and under different conditions, e.g. in presence of ice, sand, or uneven scenes.