Cloud shadow detection and removal for high spatial resolution optical satellite data

Kurzfassung

Optical satellite imagery contain in many cases clouds and cloud shadows, so that an automated classification of objects on the Earth’s surface is difficult. Therefore, on the one hand, it is important to create a good cloud and cloud shadow map and, on the other hand, to correct the shadow areas in such a way that a classification via the spectral properties of the objects is uniformly possible. Of course, this also includes the process of atmospheric correction of the optical satellite image data. Especially for land applications the amount of scenes with usable data is of high importance due to the exact timing being significant (crop yield estimation) or because the scene is not free-of-charge and one has to pay for the next acquisition if the current one contains cloud shadow over the location of interest. The masking of clouds, cloud shadows, water and snow/ice in optical satellite imagery is therefore an essential step in automated processing chains. Furthermore the exact masking of cloud shadows is a very important task prior to the removal of cloud shadows. Due to the Earth having an annual cloud coverage of approximately 70%, the contamination of multi-spectral satellite imagery is inevitable and scientist will have to work with and around this shortcoming. For this study, the satellite data from the Sentinel-2 mission is used which provides a five day revisit time at the equator. The swath width of a Sentinel-2 scene is 290 km and the data is acquired in 13 bands with a spatial resolution of 10 m, 20 m and 60 m. For a first comparison of available masking codes, Function of mask (Fmask), ATCOR and the scene classification of Sen2Cor are evaluated. All three masking codes use rules that are based on the physical properties such as the Top of Atmosphere Reflectance (TOA) in order to differentiate clear pixels from cloud pixels. For the prediction of cloud shadows, the sensor view angle and solar illumination geometry are used. Furthermore, a special focus is set on the correct and automatic detection of cloud shadows. A new method for cloud shadow detection in multi-spectral satellite images is proposed and compared to current methods. This method is based on the evaluation of Thresholds, Indices and Projections. Following the detection of cloud shadows, an improved cloud shadow removal algorithm is presented for high spatial resolution optical satellite data over land. It is based on the Matched Filter method which calculates the covariance matrix and a corresponding zero-reflectance matched filter vector. The new cloud shadow map is added to the removal of cloud shadows as well as further evaluations performed on the shadow function to improve the removal algorithm.