SAR Tomography at the Limit: Building Height Reconstruction Using Only 3-5 TanDEM-X Bistatic Interferograms

Yilei Shi, Richard Bamler, Yuanyuan Wang, Xiao Xiang Zhu

TL;DR

This paper introduces a novel SAR tomography framework that can reconstruct detailed 3D building heights using a very limited number of TanDEM-X bistatic interferograms (3-5). This breakthrough overcomes the previous requirement of large interferogram stacks, making large-scale urban 3D mapping more feasible and accurate, even outperforming existing DEM products.

What Problem Does This Paper Address?

High-resolution synthetic aperture radar (SAR) interferometry, particularly SAR tomography (TomoSAR), is a powerful technique for retrieving 3D information of urban environments. However, traditional TomoSAR methods typically require a large number of interferograms (often over 20) to achieve reliable height reconstruction. This limitation makes them impractical for large-scale urban mapping projects using data like TanDEM-X, where only a few acquisitions are usually available for a given area. The challenge is to develop a TomoSAR approach that can deliver accurate 3D building height information with these severely limited "micro-stacks" of interferograms.

Methods

The core innovation of this work is a new SAR tomographic processing framework that integrates non-local filtering into the TomoSAR inversion process. This addresses the low Signal-to-Noise Ratio (SNR) inherent in micro-stacks. The framework proceeds through several key steps:

  • Non-Local Filtering: This technique significantly enhances the SNR of the interferograms by averaging information from similar patches across the image, crucial for improving estimation accuracy with limited data.
  • Spectral Estimation: The paper explores and compares different spectral estimation techniques, including Singular Value Decomposition (SVD) and Compressive Sensing (CS), to reconstruct the reflectivity profile along the elevation dimension. CS is highlighted for its super-resolution capabilities.
  • Model Selection: A method for estimating the number of scatterers (K) within a resolution cell is employed, assuming a sparse urban environment with a few dominant scatterers.
  • Robust Height Estimation: An M-estimator is used to robustly fuse height estimates from neighboring pixels, mitigating the impact of potential outliers and improving the overall accuracy.

The framework is designed to work with multi-master multi-baseline InSAR configurations, specifically adapted for TanDEM-X bistatic data.

Datasets

The primary dataset used for evaluation is a stack of five co-registered TanDEM-X bistatic interferograms over Munich, Germany. This dataset is characterized by a slant range resolution of 1.8 m and an azimuth resolution of 3.3 m.

For validation, the authors utilized a high-quality LiDAR dataset of Munich, provided by the Bavarian State Office for Survey and Geoinformation, with centimeter-level accuracy. This LiDAR data served as the ground truth for quantitative comparisons.

Key Insights and Ideas

  • SNR is King for Micro-stacks: The research strongly suggests that for very small interferogram stacks (3-5), increasing the Signal-to-Noise Ratio (SNR) through techniques like non-local filtering has a more significant impact on height estimation accuracy than simply increasing the number of interferograms.
  • Compressive Sensing Shines: Compressive Sensing (CS) demonstrates superior performance over traditional SVD for TomoSAR reconstruction, particularly in its ability to achieve better accuracy and super-resolution, which is vital for distinguishing closely spaced scatterers in urban areas (e.g., ground and facade).
  • Non-Local Filtering is Essential: The integration of non-local filtering is a critical enabler for applying TomoSAR to micro-stacks, effectively compensating for the reduced information content and low SNR.
  • Beyond DEMs: The proposed NL-TomoSAR approach can generate more detailed 3D building structures and higher accuracy compared to standard TanDEM-X DEM products, especially for complex urban environments.

Key Results

  • High Accuracy with Limited Data: The proposed framework successfully reconstructs building heights with a relative height accuracy where two-thirds of buildings are within two meters of the LiDAR reference. This is a significant improvement over the standard TanDEM-X DEM.
  • Quantitative Validation: A systematic comparison with LiDAR data for over 34,000 buildings in Munich showed that 38.7% of buildings were within 1m accuracy, and 62.8% were within 2m accuracy. The overall standard deviation of height differences was 1.96 m.
  • Visual Superiority: Visual comparisons clearly show that the NL-TomoSAR point clouds preserve more detailed building structures and road networks compared to the TanDEM-X DEM, particularly in complex areas.
  • CS Outperforms SVD: Monte Carlo simulations indicated that CS-based methods provide better accuracy and super-resolution power for double scatterers compared to SVD.

Limitations and Future Work

  • Co-registration Dependency: The accuracy of the TomoSAR point cloud is dependent on the precise co-registration of the interferograms. While the paper mentions steps for co-registration, achieving perfect alignment can still be challenging.
  • Geocoding and Alignment: The initial output of TomoSAR is in a range-azimuth coordinate system. Transforming this to a georeferenced coordinate system and aligning it with other datasets (like DEM and LiDAR) requires careful processing steps (geocoding and ICP registration), which can introduce their own uncertainties.
  • Future Work: The authors suggest that the proposed framework is a promising solution for high-quality large-scale 3D urban mapping. Future work could explore further optimization of the non-local filtering and spectral estimation techniques, as well as extending the framework to other SAR datasets and urban environments. The potential for using this technique for change detection and monitoring in urban areas also remains an avenue for exploration.

Summary updated

Abstract

Multi-baseline interferometric synthetic aperture radar (InSAR) techniques are effective approaches for retrieving the 3-D information of urban areas. In order to obtain a plausible reconstruction, it is necessary to use more than twenty interferograms. Hence, these methods are commonly not appropriate for large-scale 3-D urban mapping using TanDEM-X data where only a few acquisitions are available in average for each city. This work proposes a new SAR tomographic processing framework to work with those extremely small stacks, which integrates the non-local 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 third buildings is within two meters, which outperforms the TanDEM-X raw DEM. The promising performance of the proposed algorithm paved the first step towards high quality large-scale 3-D urban mapping.
BibTeX
@article{shi2020sar,
  author = {Yilei Shi and Richard Bamler and Yuanyuan Wang and Xiao Xiang Zhu},
  title = {{SAR Tomography at the Limit: Building Height Reconstruction Using Only 3-5 TanDEM-X Bistatic Interferograms}},
  journal = {IEEE TGRS},
  year = {2020},
  doi = {10.1109/tgrs.2020.2986052},
}
Figure