Automatic Reconstruction of 3-D Building Structures for TomoSAR Using Neural Networks

Tomographic Synthetic Aperture Radar (TomoSAR) has become a competitive remote sensing method of three-dimensional (3-D) reconstruction over urban areas. Raw TomoSAR point clouds are inevitably noise-corrupted, which would severely obstruct the reconstruction of building structures. Whereas, data segmentation and parameter tuning are required in current methods of 3-D building reconstruction, which influences the precision and efficiency of the reconstruction process. In this paper we propose a novel method using neural networks to reconstruct 3-D building structures from TomoSAR data. By using the proposed method, the precise 3-D surface of the building structure can be retrieved quickly. The proposed method also performs commendably in point cloud denoising. More importantly, our method achieves full automation of the reconstruction process, which does not require data segmentation or complex parameter adjusting. Experiment results demonstrate the effectiveness of the proposed method.