Cloud screening

Owned by Marco Peters
Last updated: 2020-11-18 by Sabine Embacher
2 min read

The Cloud Screening Tools is used to mask cloudy pixels in CHRIS images. The cloud masking algorithm described below helps the user to find cloudy regions in the image and provides cloud probability and abundances for each pixel instead of a single flag.

Cloud Screening Algorithm

The cloud screening algorithm consists of the following steps:

  1. TOA Reflectance Computation: Top-of-Atmosphere reflectance is estimated from the CHRIS products to remove in practice the dependence on particular illumination conditions (day of the year and angular configuration).

  2. Feature extraction: physically-inspired features are extracted to increase separability of clouds and surface taking into account that the measured spectral signature depends on the illumination (TOA reflectance is already estimated), the atmosphere (oxygen and water vapor atmospheric absorptions are used to estimate the optical path related to the cloud height), and the surface (spectral whiteness and brightness helps to characterize the cloud's spectral behaviour).

  3. Image clustering: an unsupervised Expectation-Maximization (EM) clustering algorithm is performed on the extracted features in order to separate clouds from the ground-cover while obtaining posterior probabilities of each pixel to each cluster.

  4. Cluster labelling: resulting clusters are subsequently labelled by the user. Once all clusters corresponding to clouds have been identified, it is straightforward to merge all the clusters belonging to a cloud type (cloud-clusters). In the clustering of the extracted features, the EM algorithm provides a probabilistic membership for each cluster, thus the probability of being cloud is computed as the sum of the posteriors of the cloud-clusters.

  5. Spectral unmixing: a Fully Constrained Linear Spectral Unmixing (FCLSU) is applied to the image in order to obtain the fraction of cloud content for each pixel (rather than flags or a binary classification).

The final cloud product is obtained combining the cloud probability and the cloud fraction by means of a pixel-by-pixel multiplication. That is, combining two complementary sources of information processed by independent methods: the cloud probability (obtained from the extracted features), which is close to one in cloud-like pixels and close to zero in remaining areas; and the cloud abundance or mixing (obtained from the spectra).

 

ATBD: