Large-Scale 3-D Urban Mapping by Fusing InSAR and Optical Data

Kurzfassung

The rapid urbanization process has strongly affected the life of humanity. 3-D urban models are fundamental information for sustainable urban development and making policies to manage urban growth. Modern advanced remote sensing techniques enable the task of large-scale urban mapping. The aim of the thesis is to establish a generic framework for large-scale 3-D urban modeling. In order to achieve this, the work mainly addresses the following aspects: building height estimation, building footprint extraction, 3-D urban modeling by fusing the building height and footprint. SAR tomography is an advanced InSAR technique for building height estimation. Although TomoSAR can provide unprecedented results for large extent of areas, if our goal is the global mapping, following isusses need to be considered: (1) the resolution of images; (2) the limited number of images; (3) the computational cost. The state-of-the-art TomoSAR algorithms usually require fairly large interferometric stacks (> 20 images) for a reliable reconstruction. Hence, they are usually not directly applicable for large-scale 3-D urban mapping using TanDEM-X data where only a few acquisitions (i.e., 3-5 interferograms) are available in average for each city. In order to solve the abovementioned issues, a fast and accurate basis pursuit denoising algorithm is proposed to solve the L1 regularized least squares minimization problem for TomoSAR inversion. Moreover, a new SAR tomographic processing framework is proposed in the thesis to those extremely small stacks, which intergrates the nonlocal filtering into SAR tomography inversion. The applicability of the algorithm is demonstrated using a TanDEM-X multi-baseline stack with 5 bistatic interferograms over the whole city of Munich, Germany. Systematic comparison of our result with TanDEM-X raw digital elevation models (DEM) and airborne LiDAR data shows that the relative height accuracy of two thirds of the buildings is within two meters. Building footprint information is an essential ingredient for 3-D reconstruction of urban areas. The automatic generation of building footprints from satellite images presents a considerable challenge due to the complexity of building shapes. Semantic segmentation is comparatively inexpensive and time saving for automatic building footprint extraction. In recent years, a great success in semantic segmentation has been obtained through the use of deep learning. In particular, the deep convolutional neural networks (DCNNs) have shown promising results thanks to their high capacity for data learning. DCNNs have brought about compelling advancement over traditional semantic segmentation methods. However, exploiting DCNN for semantic segmentation tasks still faces significant challenges. In this thesis, a conditional Wasserstein generative adversarial network (WGAN) with graidient penalty is proposed to improve the generalization property of DCNN in order to achieve better accuracy. Deep architectures generally fail to produce fine-grained segmentation with accurate boundaries due to their progressive down-sampling. Moreover, DCNN fails to fine local details without the consideration about the interactions between pixels. To overcome these issues, a novel deep learning framework “gated graph convolutional neural network with deep structured feature embedding” is proposed, which enables a local and global spatial information propagation to improve the performance of the segmentation. Finally, the LOD1 3-D urban models are generated by fusing the building height obtained from TomoSAR point cloud and building footprint extracted from the optical satellite images.