The Orientation Estimation of Elongated Underground Objects via Multipolarization Aggregation and Selection Neural Network

Hai-Han Sun, Yee Hui Lee, Chongyi Li, Lai Fern Ow, Mohamed Lokman Mohd Yusof, Abdulkadir C. YucelIEEE GRSL 2022 · DOI · OpenReview

Abstract

Estimating the 3D orientation of buried elongated objects is a significant challenge in subsurface imaging. Traditional Ground Penetrating Radar (GPR) methods are often limited to horizontal angle estimation or restricted vertical ranges. This paper presents the Multi-Polarization Aggregation and Selection Neural Network (MASNet), a deep learning framework that utilizes multi-polarimetric GPR data to provide high-precision orientation estimates across the full spatial range. By aggregating co-polarized and cross-polarized features, MASNet resolves orientation ambiguities and achieves mean absolute errors of 1.2° and 0.9° for horizontal and vertical angles, respectively, with near-instantaneous processing speeds.

Introduction

Accurately identifying the 3D orientation of buried elongated objects—such as utility pipes, rebar, or tree roots—is a critical task for infrastructure mapping and environmental monitoring. Ground Penetrating Radar (GPR) is the industry standard for non-destructive subsurface imaging. However, traditional GPR analysis often struggles to provide a complete picture of an object's orientation, specifically its "tilt" (vertical inclination, $\theta$) and "heading" (horizontal orientation, $\phi$).

GPR scanning scenario for an elongated object with horizontal orientation $\phi$ and vertical inclination $\theta$.

Existing algorithms like Alford rotation are typically limited to horizontal angle estimation. Furthermore, mathematical models capable of estimating both angles often operate within restricted ranges (e.g., 0–45° for vertical inclination) and suffer from ambiguity when using single-polarization data. To address these limitations, this research introduces MASNet, a deep learning framework designed to leverage multi-polarimetric GPR data for high-precision, full-range 3D orientation estimation.

The MASNet Framework

The Multi-Polarization Aggregation and Selection Neural Network (MASNet) is engineered to mimic and enhance domain-specific radar knowledge through a multi-stage architecture.

Architecture of the Multi-Polarization Aggregation and Selection Neural Network (MASNet).

The architecture consists of three primary modules:

  1. Multi-Polarimetric Feature Aggregation: The network processes three scattering components ($S_{xx}$, $S_{xy}$, and $S_{yy}$) alongside various concatenated combinations. This approach forces the model to learn the complementary relationships between different polarization channels, which is essential for resolving orientation ambiguities.
  2. Distinguished Feature Selection: Utilizing residual blocks and convolutional layers, the network suppresses redundant noise and extracts the most salient patterns from the radargrams. Residual learning is particularly important here, as GPR signal values can be extremely small, and deep networks might otherwise lose this information.
  3. Multi-Task Angle Estimation: The final stage of the network branches into two parallel paths, each dedicated to estimating one of the two orientation angles (horizontal $\phi$ or vertical $\theta$).

Experimental Setup and Datasets

The model was trained and validated using extensive simulated data generated via gprMax, an electromagnetic wave simulator. The datasets included:

  • Training Set: 1,260 B-scan image sets of a metal bar in sand, with angles varied from 0° to 175° in 5° increments.
  • Testing Set: 60 sets with random orientation angles to evaluate real-world accuracy.
  • Generalization Set: Over 9,000 sets with varying object lengths (0.15m to 0.9m) and diameters (0.01m to 0.04m) to ensure the model learned general physical principles rather than specific object dimensions.

Key Insights: Polarization and Asymmetry

The success of MASNet lies in its ability to interpret two key physical phenomena:

  • The Power of Cross-Polarization ($S_{xy}$): While co-polarized components ($S_{xx}$, $S_{yy}$) might appear identical for different angles, the cross-polarized component ($S_{xy}$) exhibits a phase difference. This "smoking gun" allows the network to distinguish between orientations that would otherwise be ambiguous.
  • Hyperbolic Asymmetry: Vertical inclination ($\theta$) manifests as asymmetry in the hyperbolic reflection pattern of a GPR scan. When one end of an object is closer to the surface, the radar signal returns earlier, tilting the hyperbola. MASNet effectively captures these subtle geometric distortions.

Performance Evaluation

MASNet demonstrates significant improvements over both traditional algorithms and existing mathematical models:

  • Precision: The model achieved a Mean Absolute Error (MAE) of 1.2° for horizontal angles and 0.9° for vertical angles.
  • Range and Speed: Unlike previous models, MASNet maintains high accuracy across the full 0–180° spatial range. Furthermore, inference is exceptionally fast, taking only 0.0007 seconds—approximately nine times faster than the Alford rotation method.
  • Robustness: Testing on the generalization set confirmed that the model remains accurate despite variations in the buried object's length and diameter, proving it learned the underlying physics of reflection.

Conclusion

MASNet represents a significant advancement in subsurface sensing, providing a fast, accurate, and automated solution for 3D orientation estimation. By aggregating multi-polarimetric features, the network overcomes the inherent ambiguities of single-polarization GPR. Future work will focus on validating the model in more complex, heterogeneous soil environments and extending its application to non-metallic objects.

Summary updated

Abstract

The horizontal orientation angle and the vertical inclination angle of an elongated subsurface object are key parameters for object identification and imaging in ground-penetrating radar (GPR) applications. Conventional methods can only extract the horizontal orientation angle or estimate both angles in narrow ranges due to limited polarimetric information and detection capability. To address these issues, this letter, for the first time, explores the possibility of leveraging neural networks with multipolarimetric GPR data to estimate both angles of an elongated subsurface object in the entire spatial range. Based on the polarization-sensitive characteristic of an elongated object, we propose a multipolarization aggregation and selection network (MASNet), which takes the multipolarimetric radargrams as inputs, integrates their characteristics in the feature space, and selects discriminative features of reflected signal patterns for accurate orientation estimation. Numerical results show that our proposed MASNet achieves high estimation accuracy with an angle estimation error of less than 5°. The promising results obtained by the proposed method encourage one to think of new solutions for GPR-related tasks by integrating multipolarization information with deep learning techniques. The data and code implemented in the letter can be found at <uri xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">https://haihan-sun.github.io/GPR.html</uri> .
BibTeX
@article{sun2022orientation,
  author = {Hai-Han Sun and Yee Hui Lee and Chongyi Li and Lai Fern Ow and Mohamed Lokman Mohd Yusof and Abdulkadir C. Yucel},
  title = {{The Orientation Estimation of Elongated Underground Objects via Multipolarization Aggregation and Selection Neural Network}},
  journal = {IEEE Geosci. Remote. Sens. Lett.},
  year = {2022},
  doi = {10.1109/lgrs.2021.3072923},
}
Figure