Sparse radar imaging based on <italic>L</italic><sub>1/2</sub> regularization theory

Zongben Xu, Yirong Wu, Bingchen Zhang, Yao WangChinese 2018 · DOI

Abstract

Sparse radar imaging aims to provide high-performance radar imaging for observed scenes by leveraging the $L_{1/2}$ regularization method. This article reviews a decade of systematic exploration and innovative practices led by Zongben Xu and Yirong Wu, which earned the 2018 Tan Kah Kee Science Award. Key contributions include a new sparse radar imaging theory based on $L_{1/2}$ regularization, a novel model utilizing an echo simulation operator, and a 3D phase diagram approach for system evaluation. These advancements enable effective Synthetic Aperture Radar (SAR) imaging at sub-Nyquist sampling rates with reduced computational complexity, paving the way for large-scale real-world applications.

The Practicality Walls of Sparse Radar

Traditional radar imaging is governed by the Nyquist-Shannon sampling theorem, which necessitates high sampling rates. This leads to massive data volumes, excessive power consumption, and hardware complexity. While Compressed Sensing (CS) theoretically allows for sub-Nyquist sampling by exploiting the inherent sparsity of a scene (e.g., a few ships in a vast ocean), its practical application in radar has been hindered by two major obstacles: the "Sparsity Gap" and the "Computational Wall."

Comparison of traditional vs. compressed sensing workflows.

Standard $L_1$ regularization, often considered the gold standard for CS, frequently fails to produce sufficiently sparse results in noisy, real-world radar environments. Furthermore, the computational burden of using a massive measurement matrix ($\Phi$) to represent the sensing process is prohibitive for large-scale SAR scenes, where matrices can reach terabytes in size, making iterative reconstruction impossible for practical use.

Core Innovations: $L_{1/2}$ and the Echo Simulation Operator

To overcome these challenges, the researchers introduced three pivotal innovations that bridge the gap between theoretical promise and engineering reality:

1. $L_{1/2}$ Regularization Theory

The team proved that among all non-convex $L_q$ ($0 < q < 1$) penalties, $L_{1/2}$ is the "sparset" one that still possesses an analytical solution, known as the Half-thresholding algorithm. This discovery provides a unique bridge in the $L_q$ space, offering the aggressive sparsity of non-convex optimization with computational efficiency comparable to convex methods.

2. Radar Echo Simulation Operator ($M^\dagger$)

Instead of storing a gargantuan $\Phi$ matrix, the authors developed an operator that mimics the radar's physical process, specifically range-Doppler decoupling. This allows the algorithm to perform matrix-vector multiplications using Fast Fourier Transforms (FFTs), slashing computational complexity from $O(N^2)$ to $O(N \log N)$. This shift from a static matrix representation to a functional operator bypassed the memory wall entirely.

3. 3D Phase Diagrams

To evaluate system performance, the research moved beyond simple 2D plots to 3D phase transition volumes. These diagrams map the probability of successful reconstruction against three axes: Sparsity, Sampling Ratio, and Signal-to-Noise Ratio (SNR). This allows engineers to precisely calculate the minimum sampling rate required for a specific environment's noise level, providing a robust framework for system design.

Experimental Validation and Results

The research utilized a combination of simulated data (using RadarSat-1 images as ground-truth) and real-world airborne experimental data. A custom-built Sparse Radar Principle Prototype was flown over the Tianjin Binhai area in China to verify the theory's effectiveness under real-world atmospheric and motion-error conditions.

Key results include:

  • Efficiency Gain: A 100x speedup in reconstruction efficiency compared to traditional CS frameworks.
  • Memory Reduction: Requirements were reduced from terabytes to the same order as standard radar imaging algorithms.
  • Hardware Validation: The airborne prototype successfully reconstructed high-resolution images at sampling rates significantly below the Nyquist limit, proving the universal value of the $L_{1/2}$ theory.
  • Broad Impact: Beyond radar, the $L_{1/2}$ theory has been successfully applied to CT imaging, wireless communications, and MRI.

Future Directions

While 2D SAR imaging has been significantly advanced by this framework, the authors identify several frontiers for future work. Extending the theory to 3D-SAR (TomoSAR) introduces new sparsity challenges in the elevation direction, requiring super-resolution capability for height reconstruction. Additionally, moving from vectors and matrices to higher-order tensors is proposed for handling complex data like multi-polarization or time-varying scenes. Finally, implementing these iterative solvers on satellite-grade FPGAs for real-time onboard processing remains a significant engineering hurdle.

Summary updated

Abstract

Sparse radar imaging seeks to provide high-performance radar imaging technique for observed scenes by using <italic>L</italic><sub>1/2</sub> regularization method. This article surveys our systematic explorations and innovative practices for tackling such problem. The key contributions include a new sparse radar imaging theory based on <italic>L</italic><sub>1/2</sub> regularization theory, a new sparse radar imaging model based on echo simulation operator, a new sparse radar imaging system evaluate approach based on 3D phase diagram. The synthetic aperture radar (SAR) imaging can be formulated as a <italic>L</italic><sub>1/2</sub> regularization problem, which optimizes a quadratic error term of radar imaging. The method can implement SAR imaging effectively at lower sampling rates than Nyquist rate and produce higher quality images with reduced sidelobes. Echo simulation operator is constructed to generate SAR raw data by means of inverse of matched filter-based imaging procedure, instead of the exact observation matrix. The echo simulation operator based model forms a new radar imaging method, which reduces the computational complexity and memory occupation to the same order as matched filter imaging algorithm, and hence, makes <italic>L</italic><sub>1/2</sub> regularization reconstruction of large-scale considered scene become possible. 3D phase diagram is employed to analyze and evaluate the performance of sparse radar imaging system. In 3D phase diagram, success rates of signal recovery can be derived from the statistics of relative errors between the ground truth and the recovered signals. The success rates for each combination of SNR, sparsity and sampling ratio are provided to form a three-dimensional phase diagram. The trend of success rates is presented accurately by taking advantage of 3D phase transition boundary. We design the sparse radar prototype and conduct a series of airborne experiments, which demonstrate the applicability and potentials of the sparse radar imaging. 3D phase diagram is used to evaluate the relationship between reconstruction performance and signal-to- noise ratio. The experimental results verity the effectiveness of the <italic>L</italic><sub>1/2</sub> regularization method and the feasibility of 3D phase diagram. The sparse radar imaging based on <italic>L</italic><sub>1/2</sub> regularization has promising application prospects. Firstly, it could be generalized to deal with 3D-SAR imaging, which achieves better flexibility for system design and super-resolution capability of height dimension for reconstruction property. Then, we review some of the most popular techniques to exploit sparsity, for recovering high-dimensional matrices and higher order tensors from compressive measurements, and further emphasize the broad application prospects of higher order compressive sensing in 3D-SAR imaging.
BibTeX
@inproceedings{xu2018sparse,
  author = {Zongben Xu and Yirong Wu and Bingchen Zhang and Yao Wang},
  title = {{Sparse radar imaging based on \&amp;lt;italic\&amp;gt;L\&amp;lt;/italic\&amp;gt;\&amp;lt;sub\&amp;gt;1/2\&amp;lt;/sub\&amp;gt; regularization theory}},
  booktitle = {Chinese},
  year = {2018},
  doi = {10.1360/n972018-00372},
}
Figure