CLOUD MASKING
Spacemetric has implemented improved cloud masking and cloud shadow identification methods based on monitoring and analysis of the state of the art in this field via projects funded through national programmes and internal R&D funds.
The primary aim of our cloud masks is to help users identify useful pixels to facilitate robust, automated user workflows. The masking techniques flags those pixels with a significant chance of contamination by cloud or cloud shadow.
The philosophy has been to tend to overestimate the number of contaminated pixels at the potential cost of discarding a small proportion of uncontaminated pixels. The remaining pixels are those judged to have minimal contamination. The method also aims to preserve the pixels, such as bright urban targets, that otherwise are falsely flagged as clouds by other methods.
Cloud and shadow detection consist of labelling image pixels to indicate if they correspond to a cloud, a shadow, or neither of these. For simplicity, we refer to the label map as a “cloud mask” (Figure 1).
Figure 1: Sentinel-2 image (left), and corresponding cloud & shadow mask (right)
Traditional cloud detection relates brightness levels, or indices derived therefrom, to thresholds, sometimes based on physical models specific to clouds, but usually in combination with morphological image processing. The use of parallax effects between spectral bands as well as the geometrical relationship between clouds and their shadows have also been used.
Such methods need to be carefully customised to the characteristics of each specific sensor and this has been achieved for Sentienl-2. These methods individually are reported to perform at accuracies above 90% for Sentinel-2, while they are used in combination within Spacemetric’s hybrid technique.