A 3D building reconstruction method for SAR images based on deep neural network

Abstract

Owing to its momentous applications in urban planning, disaster monitoring, smart cities and other fields, 3D building reconstruction is an important topic in many research areas such as computer vision, photogrammetry, and remote sensing. The particularity and complexity of the microwave scattering mechanism bring great challenges to the 3D building reconstruction of SAR images, and the applicability and automation of existing methods need to be improved. This study constructs the overall framework of building detection in SAR images and 3D reconstruction based on deep learning and radar imaging mechanism. It puts forward a method of using a coupled equivalent complex valued convolutional neural network for building facade detection in SAR images, a method for RaySAR-based modeling simulation and point cloud generation for 3D model training, and a method for a 3D generation network for 3D building reconstruction from SAR images. Experiments using TerraSAR-X and GF-3 high resolution SAR images are carried out, producing good 3D reconstruction results. The proposed method is a useful attempt for 3D target reconstruction of SAR images and represents a new technical approach.