Calculating Shadows with U-Nets for Urban Environments

Shadow calculation is an important prerequisite for many urban and environmental analyses such as the assessment of solar energy potential. We propose a neural net approach that can be trained with 3D geographical information and predict the presence and depth of shadows. We adapt a U-Net algorithm traditionally used in biomedical image segmentation and train it on sections of Styria, Austria. Our two-step approach first predicts binary existence of shadows and then estimates the depth of shadows as well. Our results on the case study of Styria, Austria show that the proposed approach can predict in both models shadows with over 80% accuracy which is satisfactory for real-world applications, but still leaves room for improvement.