GPR Clutter Removal Based on Weighted Nuclear Norm Minimization for Nonparallel Cases

Li Liu, Chenyan Song, Zezhou Wu, Hang Xu, Jingxia Li, Bingjie Wang, Jiasu Li2023 · DOI

TL;DR

This paper introduces a novel method for removing clutter from Ground-Penetrating Radar (GPR) data, specifically addressing the challenging scenario where the GPR antenna is not perfectly parallel to the ground surface. By employing a weighted nuclear norm minimization technique, the authors achieve superior clutter suppression and target detection compared to existing methods, especially in non-ideal GPR acquisition conditions.

What Problem Does This Paper Address?

Ground-penetrating radar (GPR) is a powerful tool for non-destructively imaging subsurface structures. However, its effectiveness is often hampered by "clutter" – unwanted signals arising from ground reflections, direct antenna coupling, and subsurface discontinuities. This clutter can overwhelm the faint signals from buried targets, making detection difficult. While many clutter removal techniques exist, they typically assume a flat ground surface and parallel antenna alignment. In real-world applications, such as pavement inspection or surveys on uneven terrain, this ideal scenario is rarely met. When the antenna is tilted relative to the ground, the resulting clutter becomes more complex and difficult to remove with conventional methods, leading to reduced detection performance. This paper tackles this specific challenge: improving GPR clutter removal when the antenna-ground angle is non-zero.

Methods

The core of the proposed method is Weighted Nuclear Norm Minimization (WNNM). This approach builds upon the principles of Robust Principal Component Analysis (RPCA), which decomposes a GPR B-scan image (represented as a matrix) into a low-rank component (clutter) and a sparse component (target).

The key innovation lies in how WNNM treats the singular values of the matrix. Traditional RPCA uses a standard nuclear norm, which effectively shrinks all singular values by the same amount. The WNNM, however, introduces weights to individual singular values. The rationale is that larger singular values in GPR data often correspond to stronger clutter (like ground reflections), while smaller ones are more likely to represent targets. By assigning smaller weights to larger singular values and larger weights to smaller ones, WNNM can more effectively suppress dominant clutter components while preserving weaker target signals. This weighted approach is particularly beneficial when the ground clutter is no longer a simple horizontal line due to the nonparallel antenna.

The WNNM optimization problem is solved using an inexact alternating direction method of multipliers (ADMM), which involves iterative updates of the low-rank (clutter) and sparse (target) matrices.

Datasets

The paper evaluates its proposed method using:

  • Simulated Data: Generated using the gprMax3D software, simulating various scenarios with different soil types, target materials (aluminum and plastic pipes), burial depths, and importantly, varying antenna-ground angles (0°, 1°, 3°, 5°).
  • Real Data:
    • Real Data-I: Collected using a stepped-frequency continuous wave (SFCW) GPR system with horn antennas. This dataset features a rough, tilted sand surface with three buried objects (metal pipeline, plastic bottle, PVC pipeline) at different depths.
  • Real Data-II: Obtained from the TU1208 open database, collected with a 400 MHz shielded antenna in a limestone region with a complex underground environment containing various objects (steel pipe, electrical cable, PVC pipes, cavity).

Key Insights and Ideas

  • The "Nonparallel Case" is Crucial: The paper highlights that most existing GPR clutter removal methods are designed for ideal, parallel antenna-ground conditions. Real-world scenarios often deviate from this, and the proposed WNNM directly addresses this gap.
  • Singular Value Weighting Matters: The core insight is that not all singular values are equal in GPR data. Assigning differential weights based on their magnitude allows for more targeted clutter suppression, especially for complex ground reflections.
  • WNNM's Adaptability: The weighted approach makes the method more flexible and adaptable to varying clutter characteristics, which is essential for nonparallel scenarios where clutter is less predictable.
  • Performance Trade-offs: The paper demonstrates that while some methods perform well in ideal conditions, their performance degrades significantly as the antenna-ground angle increases. WNNM shows more robust performance across these varying angles.

Key Results

  • Improved Clutter Suppression: In both simulated and real-world experiments, the WNNM method consistently demonstrated superior clutter removal, particularly in nonparallel antenna-ground scenarios.
  • Higher PSNR and IF: Quantitative evaluations using Peak Signal-to-Noise Ratio (PSNR) and Improvement Factor (IF) showed that WNNM achieved the highest values, especially as the antenna-ground angle increased. For instance, in simulations with a 5° angle, WNNM achieved significantly higher PSNR than other methods.
  • Faster Convergence: The WNNM method converged much faster than RPCA, being approximately five times faster in simulations and over 4.5 times faster in real data experiments. This speed advantage is crucial for practical, real-time GPR applications.
  • Comparable to State-of-the-Art (in ideal cases): In parallel scenarios (0° angle), WNNM performed comparably to or better than other advanced methods like RNMF and RPCA.

Limitations and Future Work

  • Parameter Tuning: The performance of WNNM is influenced by regularization parameters ($\lambda$) and weight parameters ($\rho$). While the paper discusses how to tune these for optimal PSNR, finding these parameters can still require some effort.
  • Computational Complexity: While faster than RPCA, the computational complexity of WNNM is still $O(MN^2)$, similar to RPCA. Further optimizations for extremely large datasets might be beneficial.
  • Future Work: The authors suggest that further research could explore more advanced deep learning-based clutter suppression methods that can also handle the nonparallel case, potentially offering even greater performance gains.

Summary updated

BibTeX
@misc{liu2023gpr,
  author = {Li Liu and Chenyan Song and Zezhou Wu and Hang Xu and Jingxia Li and Bingjie Wang and Jiasu Li},
  title = {{GPR Clutter Removal Based on Weighted Nuclear Norm Minimization for Nonparallel Cases}},
  year = {2023},
  doi = {10.3390/s23115078},
}
Figure