Stochastic geometrical model and Monte Carlo optimization methods for building reconstruction from InSAR data
Stochastic geometrical model and Monte Carlo optimization methods for building reconstruction from InSAR data
复制标题
DOI：
10.1016/j.isprsjprs.2015.06.004
复制
发表时间：
2015-10
影响因子：
12.7
通讯作者：
Yue Zhang;Xian Sun;A. Thiele;S. Hinz
中科院分区：
文献类型：
--
作者：
Yue Zhang;Xian Sun;A. Thiele;S. Hinz
合成孔径雷达（SAR）系统，如TanDEM-X、TerraSAR-X和cosmos - skymed，可以获得高空间分辨率（HR）的图像，从而可以高细节地观察城市地区的物体。在本文中，我们提出了一个新的自顶向下的框架，用于从HR干涉SAR （InSAR）数据中重建三维（3D）建筑物。与之前提出的大多数方法不同，我们采用生成模型，并通过蒙特卡罗方法通过最大化后验估计（MAP）来利用重建过程。之所以采用这种策略，是因为SAR图像的噪声需要一个彻底的先验模型，以更好地应对其固有的幅度和相位波动。在重建过程中，根据雷达配置和建筑物几何形状，将三维建筑物假设映射到SAR图像平面上，并分解为中途停留、角线、阴影等特征区域。然后，分别探讨了各区域的强度、干涉相位和相干性的统计特性，并将其作为区域项包含。屋顶不直接考虑，因为在大多数情况下，它们与墙壁混合在一起成为中途停留区。在估计建筑物假设与实际数据的相似度时，考虑了先验、区域项以及与过境点和边线轮廓相关的边缘项。在优化步骤中，为了获得收敛的重构输出并消除局部极值，设计了特殊的过渡核。该框架在TanDEM-X数据集上进行了评估，并在建筑物重建中表现良好。
Synthetic aperture radar (SAR) systems, such as TanDEM-X, TerraSAR-X and Cosmo-SkyMed, acquire imagery with high spatial resolution (HR), making it possible to observe objects in urban areas with high detail. In this paper, we propose a new top-down framework for three-dimensional (3D) building reconstruction from HR interferometric SAR (InSAR) data. Unlike most methods proposed before, we adopt a generative model and utilize the reconstruction process by maximizing a posteriori estimation (MAP) through Monte Carlo methods. The reason for this strategy refers to the fact that the noisiness of SAR images calls for a thorough prior model to better cope with the inherent amplitude and phase fluctuations.In the reconstruction process, according to the radar configuration and the building geometry, a 3D building hypothesis is mapped to the SAR image plane and decomposed to feature regions such as layover, corner line, and shadow. Then, the statistical properties of intensity, interferometric phase and coherence of each region are explored respectively, and are included as region terms. Roofs are not directly considered as they are mixed with wall into layover area in most cases. When estimating the similarity between the building hypothesis and the real data, the prior, the region term, together with the edge term related to the contours of layover and corner line, are taken into consideration. In the optimization step, in order to achieve convergent reconstruction outputs and get rid of local extrema, special transition kernels are designed. The proposed framework is evaluated on the TanDEM-X dataset and performs well for buildings reconstruction.