Optimization of extracted building footprints from UAV images

Building footprints is important information in many types of applications, including optimization of rescuer response in case of catastrophic events, urban planning, urban dynamic monitoring, 3D building modeling etc. Traditionally, in remote sensing, building footprints are detected from very high-resolution images or point clouds. Convolution Neural Network (CNN) based semantic image segmentation model has become a common way to extract buildings footprints from remote sensing data with high accuracy regardless of differences in landscapes, shapes, texture, and used materials. However, the results of extraction usually represent rooftop outlines with overhangs rather than true building footprints. This paper presents the methodology for the optimization of building footprints by using contour information, which is derived from the UAV point cloud. First, the CNN model was used to extract rooftops from high-resolution UAV-based orthophoto. After that, the cross-section of the mesh model was performed in order to detect the outline of the building. The optimum height of the mesh cross section was derived based on an analysis of the Digital Elevation Model and Digital Surface Model. The generated results were compared with Open Street Map (OSM) and reference cadastral datasets. Quantitative and qualitative evaluations show that the proposed methodology can significantly improve the accuracy of CNN-extracted building footprints (and OSM data) compared to cadastral data. In addition, the high of buildings is simultaneously derived. Therefore, our approach opens up the possibility to use the full potential of UAV products for generating accurate building footprints and 3D building models of LoD1 with compatible accuracy as cadastral.