A contextual approach to building extraction combining highre-solution InSAR and optical data

Urban areas are highly complex scenes and thus single buildings are often hard to extract. A combination of complementary features of high-resolution interferometric SAR (InSAR) data and optical imagery can add valuable information in case one data source leads to ambiguous results. In addition, contextual information like the sun shadow, front yards, and driveways can act as hints for buildings. We propose a Conditional Random Field (CRF) approach set up on super-pixels of an initial multi-scale segmentation to learn and infer those contextual features in addition to direct building features. We compare our results to CRFs with square image patches and CRFs with single scale super-pixels. Experiments on airborne InSAR data and optical data show that the proposed CRF with multi-scale super-pixels performs best.