Cloud shadow area derivation for optical satellite images based on cross-correlation methods

Kurzfassung

In this thesis two cross-correlation based cloud shadow area derivation methods for optical satellite images are developed to detect cloud shadows over different earth surfaces, e.g. flat land area, mountain area and ocean. The two methods are called geometric prediction and region growing detection, respectively. Main inputs into the algorithms are a satellite image, its corresponding cloud mask and the DEM of the area. A channel or a combination of some channels of the satellite image with clear cloud shadow structures together with the given cloud mask are used for cross-correlation. It results normally in two peaks: a maximum for the clouds and a minimum for the shadows in the cross-correlation matrix. The shift vector between these two peaks can be used to estimate the mean shift vector between the clouds and their shadows. Hence a coarse shadow mask can be generated. With it the mean cloud height can be calculated according to the solar and satellite angles. A mean shift value can be calculated from cross-correlation between cloud mask and original image. Terrain variation causes another shift value from each cloud to it shadow. The sum of the two shift values should be the best estimate of the shadow length, which presents the shadow position in non-flat area. Seed points for cloud shadows can be generated from the coarse cloud mask. The coarse cloud mask is shifted by the mean shift calculated from cross-correlation. Consequently t region growing algorithm is implemented for cloud shadow area derivation. An improvement for non-flat area is to use seed points from result shadow mask of the geometric prediction. Because these seed points are results after correcting the shift due to terrain variations. This method can detect the cloud shadow areas very precisely. Three scenes from ETM+ images are tested in this work, one flat area, one mountain area and one ocean area. Two methods are implemented for the flat and mountain area, then a combination method is aimed to improve the result in the mountain area. There are further analyses on the application of these methods. For the ocean area, cross-correlation does not work, both methods fail in this case. There is a discussion about the selection of a cloud shadow detection channel, as well as setting the parameters for cross-correlation. Finally benefits and limitations of these methods are discussed, and then a comparison between these cross-correlation based methods is presented.