γ-Net: Superresolving SAR Tomographic Inversion via Deep Learning

Kun Qian, Yuanyuan Wang, Yilei Shi, Xiao Xiang ZhuIEEE TGRS 2022 · DOI

TL;DR

This paper introduces γ-Net, a novel deep learning approach for super-resolving Synthetic Aperture Radar (SAR) Tomography (TomoSAR) inversion. By unrolling an iterative optimization algorithm into a neural network, γ-Net significantly accelerates TomoSAR processing while maintaining state-of-the-art accuracy, making large-scale 3D urban mapping more feasible.

What Problem Does This Paper Address?

Synthetic Aperture Radar (SAR) tomography (TomoSAR) is a powerful technique for creating 3D reconstructions of urban environments. However, achieving high-resolution 3D maps, especially separating closely spaced scatterers (super-resolution), is computationally demanding. Traditional methods like compressive sensing (CS) offer good performance but require extensive computational resources, limiting their scalability. Existing deep learning approaches have shown promise but often lack systematic analysis of their super-resolution capabilities or are difficult to apply to real-world scenarios. This paper aims to bridge this gap by developing an efficient and effective deep learning-based TomoSAR inversion method that can achieve high super-resolution performance.

Methods

The core of γ-Net is a deep neural network designed to mimic the iterative optimization steps of sparse reconstruction algorithms. Specifically, it unrolls a complex-valued learned iterative shrinkage thresholding algorithm (CV-LISTA). Key innovations in γ-Net include:

  • Weight Coupling: Exploiting the correlation between weight matrices in the LISTA architecture to reduce redundant parameters and accelerate training.
  • Support Selection: A technique inspired by LASSO, where a certain percentage of entries with the largest magnitudes are selected before the shrinkage step, helping to preserve important information.
  • Piecewise Linear Thresholding Function: Replacing the conventional soft-thresholding function with a piecewise linear one. This aims to mitigate information loss that occurs when small magnitude elements are simply pruned to zero, thereby improving convergence and accuracy.

Datasets

The paper utilizes simulated data for training and evaluation. The simulation setup mimics real-world SAR acquisitions with varying numbers of spatial baselines, Signal-to-Noise Ratios (SNRs), and different configurations of single and double scatterers. For real-world validation, a stack of six high-resolution TanDEM-X SAR pairs was used.

Key Insights and Ideas

  • Deep Learning as an Optimization Unfolder: The paper demonstrates that deep learning can effectively learn to perform complex iterative optimization tasks, such as sparse reconstruction, by "unrolling" these algorithms into neural network layers.
  • Efficiency Through Architecture Design: By incorporating techniques like weight coupling and support selection, γ-Net achieves significant computational speedups compared to traditional CS methods without sacrificing accuracy.
  • Piecewise Linear Function's Impact: The choice of activation function is crucial. The piecewise linear function proves more effective than the standard soft-thresholding for TomoSAR inversion, leading to better accuracy and faster convergence by preserving more information.
  • Generalization and Transferability: The trained γ-Net shows good generalization ability across different SNRs and amplitude ratios, and even exhibits moderate tolerance to baseline discrepancies between training and testing data.

Key Results

  • Computational Efficiency: γ-Net achieves a computational efficiency improvement of 1 to 2 orders of magnitude compared to first-order CS-based methods. It is also significantly faster than second-order CS methods.
  • Super-Resolution Performance: γ-Net demonstrates comparable super-resolution power to state-of-the-art methods like SLIMMER, particularly in separating closely spaced scatterers.
  • Accuracy: The method approaches the Cramér-Rao lower bound for height estimation and shows high detection rates, even in moderate super-resolving cases (e.g., >90% detection rate at 25 measurements at 6dB SNR).
  • Real Data Performance: Tests on real TerraSAR-X data with only 6 interferograms show high-quality 3D reconstruction with high-density detected double scatterers, outperforming SLIMMER in some scenarios.

Limitations and Future Work

  • Training Time: The training of γ-Net can be time-consuming (around 9 hours for a single model with a powerful GPU), making it less suitable for processing very small datasets where conventional methods might be faster.
  • Baseline Discrepancies: While γ-Net shows moderate tolerance, significant baseline differences between training and testing data can still affect performance. Further research is needed to guarantee performance across large-scale processing with varying baselines.
  • Complex Scenarios (More than 2 Scatterers): The current γ-Net is primarily designed for scenarios with up to two overlaid scatterers. For cases with three or more scatterers, performance degrades. Future work includes enriching training data with more scatterers or exploring out-of-distribution detection methods.
  • Future Directions: The authors suggest extending the deep learning approach to higher-dimensional spectral estimation and exploring recurrent neural network architectures like LSTM to incorporate historical information for improved performance.

Summary updated

BibTeX
@article{qian2022net,
  author = {Kun Qian and Yuanyuan Wang and Yilei Shi and Xiao Xiang Zhu},
  title = {{γ-Net: Superresolving SAR Tomographic Inversion via Deep Learning}},
  journal = {IEEE TGRS},
  year = {2022},
  doi = {10.1109/tgrs.2022.3164193},
}
Figure