A Novel Deep Learning Approach for High-Resolution Satellite-Based DEM Filtering

Building detection and shape extraction from satellite imagery are essential requirement for virtual 3D city modelling. For accurate 3D city modeling, the exact heights of buildings are required where digital elevation model (DEM) data is of prime consideration. DEM is a digital representation of the earth’s surface including the heights of objects and it can be generated using in-track or across-track stereo data. In order to determine accurate heights of buildings, filtering of DEM to produce a digital terrain model is very essential. Traditional DEM filtering algorithms are dependent on terrain complexity, size and shapes of building structures. The deep learning based methods are already proven their importance on image classification and segmentation tasks. Due to non linear characteristics and excellent feature extraction capabilities, deep learning methods are able to understand complex features and produce state of the art results. In this study, a novel approach for DEM filtering based on deep learning is sought. Here, an U-Net model with VGG19 at the backbone is used to extract the building shapes automatically from satellite based high resolution DEM data. Our model is trained over dense urban areas and achieved more than 99% training, precision and recall accuracy of 96% and a maximum 92% as IoU score.