SAT2BUILDING: LoD-2 Building Reconstruction from Satellite Imagery Using Spatial Embeddings

The reconstruction of buildings in level of detail–2, according to the CityGML standard, is an essential feature in applications such as urban planning, environmental simulations, and virtual reality. Existing methods work primarily only on aerial data, depend on an external digital terrain model, or do not accurately separate individual buildings. In this work, we present SAT2BUILDING, a method that predicts roof planes, building sections, and building heights in a single, fully convolutional neural network. The network relies on only orthorectified panchromatic imagery and a photogrammetric digital surface model. The three outputs are jointly processed in a level of detail–2 reconstruction pipeline that generates building models that are seamlessly connected, geometrically accurate and complete, and topologically correct. We use spatial embeddings that enable accurate segmentation of building sections and roof planes from satellite imagery. The model generalizes to data from Bonn, Germany, and Lyon, France, after being trained on data from Berlin, Germany. The training and test data differ in lighting conditions, architectural styles, and ground sampling distances. Thorough comparative evaluation shows the superiority of SAT2BUILDING over three baseline methods.