Automatic Shadow Detection and Removal in Aerial and Satellite Imagery using CNNs

Kurzfassung

The aim of this thesis was to detect and remove shadows from aerial and satellite images without the presence of shadow-free ground truth, using deep learning techniques. Hence as a first step, pre-existing neural networks to detect shadows were trained on the newly introduced DLR-SkyScapes-Shadow dataset. Then, for the shadow removal, several known techniques and experiments were conducted, leading to the presentation of a novel approach to remove shadows and therefore providing a method to extract information from shadowed areas. Despite not quite achieving the desired output quality, invaluable insights into the architectural properties required to obtain shadow free or shadow lighter images were gained. Experimental results show that the correct direction was pursued, and suggestions are made to which direction further research should proceed.