Building Target Extraction Methods in Array SAR Tomography Based on Machine Learning

By settling multiple antennas of different heights in intersection direction, synthesizing aperture in azimuth direction and using wide bandwidth signals in slope range, 3D imaging is now possible in array SAR Tomography system by employing 3D point clouds obtained from censors with a single flight in the observation area. Restricted to the number of arrays and the finite baseline length in reality, the high resolution in the elevation direction becomes a difficult task. In addition, the process of building feature extraction is still of low efficiency due to the layover phenomenon existence in urban areas. In order to overcome this problem, a novel method for building feature detection and extraction based on machine learning is proposed. The facades, roofs and grounds have to be extracted from other points, which is accomplished by a point cloud segmentation based on multiple linear regression, an edge extraction using gradient operator and a partition reconstruction exploiting clustering analysis. The method ensures adequate footprints information crossed by the fa?ade and the ground, and improves the efficiency of building feature extraction. Experimental studies with respect to the actual airborne array TomoSAR data obtained from the first 3D imaging experiment in China were carried out to demonstrate the validity.