MF-JMoDL-Net: A Deep Network for Azimuth Undersampling Pattern Design and Ambiguity Suppression for Sparse SAR Imaging

Yuwei Wu, Zhe Zhang, Xiaolan Qiu, Yao Zhao, Weidong YuarXiv 2023 · DOI · arXiv

Abstract

Repetition frequency (PRF). Given the system complexity and resource constraints, it is often difficult to achieve high imaging performance and low ambiguity without compromising the swath. In this paper, we propose a joint optimization framework for sparse strip SAR imaging algorithms and azimuth undersampling patterns based on a deep convolutional neural network, combined with matched filter (MF) approximate measurement operators and inverse MF operators, referred to as MF-JMoDL-Net, for sparse SAR imaging methods. Compared with conventional sparse SAR imaging, MF-JMoDL-Net enables us to alleviate the limitations imposed by PRF. In the proposed scheme, joint and continuous optimization of azimuth undersampling patterns and convolutional neural network parameters are implemented to suppress azimuth ambiguity and enhance sparse SAR imaging quality. Experiments and comparisons under various conditions demonstrate the effectiveness and superiority of the proposed framework in imaging results.

Introduction

Synthetic Aperture Radar (SAR) imaging is a cornerstone of modern Earth observation, yet it is perpetually constrained by the Pulse Repetition Frequency (PRF) bottleneck. To achieve high-resolution imagery, a high PRF is traditionally required. However, this generates massive data volumes that often exceed the storage and downlink capacities of satellite platforms. When engineers attempt to lower the PRF to conserve resources, they introduce "azimuth ambiguity"—distorting ghost artifacts that arise because the radar is sampling the scene too slowly, leading to aliasing.

Figure 1: Geometry of the Synthetic Aperture Radar (SAR) imaging system, illustrating the relationship between satellite motion (azimuth) and signal propagation (range).

While Sparse SAR imaging, rooted in compressed sensing, has sought to mitigate this by using fewer pulses, most existing methods rely on fixed, suboptimal sampling patterns such as uniform or random undersampling. MF-JMoDL-Net addresses this by asking a fundamental question: Can deep learning be used to learn the mathematically optimal sampling pattern for a specific scene while simultaneously training a network to reconstruct the image?

Methodology: The MF-JMoDL-Net Framework

The Matched Filter Joint Model-based Deep Learning Network (MF-JMoDL-Net) is a novel framework that breaks the traditional trade-off between radar image swath and resolution. Its architecture is built upon four primary pillars:

  1. Learnable Undersampling: Instead of relying on a fixed grid, the network treats sampling locations as trainable parameters. A U-Net-based model extracts features to determine which pulses provide the most critical information versus background clutter.
  2. Deep Unrolling: The framework "unrolls" a traditional iterative optimization algorithm into a deep neural network. This hybrid approach combines the computational speed of deep learning with the mathematical rigor of classical signal processing.
  3. MF-Based Operators: To ensure the network is efficient enough for large-scale satellite data, the authors incorporate the Chirp Scaling Algorithm (CSA) and Nonuniform Fast Fourier Transforms (NUFFT) directly into the network as approximate measurement operators. This avoids the memory-intensive matrix multiplications typically required by compressed sensing.
  4. Data Consistency (DC) Layer: A Conjugate Gradient (CG) block ensures that the reconstructed image remains consistent with the raw radar data, preventing the network from generating "hallucinated" features.

Experimental Setup and Datasets

The model was rigorously evaluated using the SAR Ship Detection Dataset (SSDD), comprising 1,300 SAR patches standardized to 256×256 pixels. To test the impact of scene sparsity, the data was categorized into three distinct environments:

  • Off-shore Ships: Highly sparse scenes.
  • Coastal Ships: Moderately sparse scenes.
  • Straits: Less sparse, more complex scenes.

To improve model robustness, the authors generated training samples by downsampling original echo data and introducing synthetic noise and phase disturbances.

Key Insights and Results

The core finding of this research is that joint optimization is essential. The sampling pattern and the reconstruction network are interdependent; optimizing them together yields significantly better results than optimizing either in isolation. Furthermore, the "learned mask" (the sampling pattern) adapts to the scene's sparsity, sampling differently for isolated targets versus crowded environments.

Quantitative Performance

When compared against standard patterns like Poisson-disk and staggered sampling, MF-JMoDL-Net showed clear superiority:

  • PSNR Gains: For off-shore ship scenes at 50% undersampling, the proposed method achieved a PSNR increase of 11.99 dB over the aliased input. This significantly outperformed Poisson-disk (5.28 dB) and staggered sampling (3.43 dB).
  • Ambiguity Suppression: Qualitative analysis demonstrated that even at extreme undersampling (down to 12.5%), the network successfully suppressed ghosting artifacts that rendered traditional methods unusable.
  • Stability: Reconstruction quality (PSNR) remained remarkably stable even as the sampling rate decreased, although structural similarity (SSIM) naturally declined at very low rates.

Conclusion and Future Work

MF-JMoDL-Net demonstrates that intelligence can replace brute-force data collection. By learning underlying features, the network preserves target structures that traditional methods might accidentally delete while attempting to remove ambiguities.

However, limitations remain. Performance is most robust in sparse scenes (like maritime environments) and begins to deteriorate as scenes become more complex (e.g., dense urban areas). Future research will explore the framework's performance in different SAR modes, such as ScanSAR or Spotlight, and investigate more complex CNN modules to enhance feature extraction in non-sparse terrestrial environments.

Summary updated

Abstract

repetition frequency (PRF). Given the system complexity and resource constraints, it is often difficult to achieve high imaging performance and low ambiguity without compromising the swath. In this paper, we propose a joint optimization framework for sparse strip SAR imaging algorithms and azimuth undersampling patterns based on a deep convolutional neural network, combined with matched filter (MF) approximate measurement operators and inverse MF operators, referred to as MF-JMoDL-Net, for sparse SAR imaging methods. Compared with conventional sparse SAR imaging, MF-JMoDL-Net enables us to alleviate the limitations imposed by PRF. In the proposed scheme, joint and continuous optimization of azimuth undersampling patterns and convolutional neural network parameters are implemented to suppress azimuth ambiguity and enhance sparse SAR imaging quality. Experiments and comparisons under various conditions demonstrate the effectiveness and superiority of the proposed framework in imaging results.
BibTeX
@misc{wu2023mf,
  author = {Yuwei Wu and Zhe Zhang and Xiaolan Qiu and Yao Zhao and Weidong Yu},
  title = {{MF-JMoDL-Net: A Deep Network for Azimuth Undersampling Pattern Design and Ambiguity Suppression for Sparse SAR Imaging}},
  year = {2023},
  eprint = {2303.10823},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2303.10823},
}
Figure