PaperlayerView exact revision on arXiv

Source-linked AI summary

FFDT: Sim-to-Real Domain Transfer Method for Far-Field Ground Penetrating Radar

Chi Wang, Hai-Han Sun

arXiv:2610.19265v1eess.SP

TL;DR

Scarce RealGPR data and system differences leave a sim-to-real gap that limits DL training. FFDT transfers far-field SimGPR data using angle-dependent transfer functions and ray tracing; its transferred data closely match RealGPR ground truth in most cases. The approach also supports DL dataset generation without task-specific field measurements, subject to the dominant-reflection approximation’s limits.

  • Problem

    RealGPR data are scarce, and system differences create a sim-to-real gap between simulated training data and real measurements.

    Source: §I · p. 1

  • Method

    FFDT combines angle-dependent transfer functions representing three-dimensional antenna radiation with ray tracing to transfer far-field SimGPR signals into the corresponding RealGPR domain.

    Source: §I · p. 2, §VII · p. 17

  • Results

    FFDT-transferred data differ from corresponding RealGPR ground truth by less than −30 dB in most cases, and trained models achieve robust performance on RealGPR data.

    Source: §VII · p. 17

  • Takeaways & Limitations

    FFDT enables RealGPR-domain dataset generation for DL training without task-specific RealGPR field measurements.

    Source: §VII · p. 17

    The dominant-reflection approximation is less accurate when multiple strong scattering components with substantially different propagation angles contribute comparably.

    Source: A. Applicability and Limitations of… · p. 14

Abstract

from arXiv · show

The scarcity of real-world ground penetrating radar (RealGPR) data is a major barrier to the development of deep learning (DL) methods for GPR data processing. Although simulated GPR (SimGPR) datasets can be used to train DL models, differences in hardware configuration and system parameters between SimGPR and RealGPR create a sim-to-real domain gap that leads to poor generalization of DL models to RealGPR measurements. To address this gap, we present a physics-based far-field GPR domain transfer (FFDT) method to transfer far-field SimGPR data to the corresponding RealGPR-domain data in the same scenario. In FFDT, SimGPR and RealGPR are modeled using angle-dependent transfer functions that capture three-dimensional antenna radiation characteristics. These transfer functions are then combined with ray tracing to enable the sim-to-real data transfer. Extensive studies across diverse objects and soil conditions demonstrate that FFDT-transferred data closely match RealGPR ground truth. To further demonstrate its value for RealGPR-domain dataset generation, FFDT is implemented to build a RealGPR-domain dataset for training a DL model for clutter removal. The model achieves robust performance on RealGPR measurements. With one-time SimGPR and RealGPR system calibration, FFDT enables RealGPR-domain dataset generation without requiring task-specific RealGPR field measurements for DL model training or fine-tuning, thereby addressing the RealGPR data scarcity bottleneck and supporting reliable development and deployment of DL methods for real-world GPR applications.

I. INTRODUCTION

RealGPR data scarcity and SimGPR–RealGPR mismatch constrain DL training and generalization. FFDT uses angle-dependent antenna transfer functions and ray tracing to transfer simulated signals into the corresponding RealGPR domain, with close agreement to measured data in most cases.

I. INTRODUCTION

  • Differences between simulated and real systems create a sim-to-real gap, while collecting and labeling RealGPR data is costly and time-consuming.Existing learning-based strategies still require task-specific RealGPR measurements and depend on data quality, diversity, and sometimes label accuracy.Source: §I · p. 1
  • Angle-dependent transfer functions capture three-dimensional antenna radiation characteristics, addressing the directional behavior omitted by angle-independent system representations.The method combines SimGPR and RealGPR transfer functions with ray tracing to transfer signals between domains.Source: §I · p. 2
  • FFDT targets (quasi-)monostatic, co-polarized GPR systems operating in the far field and transfers SimGPR signals to those of a specified RealGPR system in the same scenario.The stated objective is to replace task-specific RealGPR training-data collection with physics-based domain transfer.Source: §I · p. 2
  • FFDT-transferred data differ from RealGPR data by less than −30 dB in most evaluated cases.The evaluation spans object geometries, material properties, soil heterogeneity, surface roughness, and GPR systems.Source: §I · p. 2
  • FFDT-transferred datasets improved DL-model performance on RealGPR radargrams over training on SimGPR data in experimental clutter removal and numerical detection and classification.The studies demonstrate RealGPR-domain dataset generation without task-specific field measurements.Source: §I · p. 2

A. Far-Field Monostatic GPR System Model

The model represents far-field monostatic GPR signals with antenna and environment terms, then uses calibrated system transfer functions and ray tracing to map SimGPR reflections into the RealGPR domain.

A. Far-Field Monostatic GPR System Model

  • The frequency-domain model separates antenna and transceiver behavior from the environment’s Green’s function, which describes the scattered field and is independent of the GPR system.The antenna’s directional transmitting and receiving functions account for radiation intensity and phase delay in three-dimensional space.Source: A. Far-Field Monostatic GPR System Model · p. 2
  • For quasi-monostatic systems, the model also applies when the TX and RX antennas are close together, with a small bistatic angle and nearly identical propagation paths.In that configuration, Ha(f) represents direct TX-RX coupling.Source: A. Far-Field Monostatic GPR System Model · p. 2
  • Transfer functions are calibrated using free-space measurements and measurements over a PEC plate at multiple distances; the resulting calibration is reusable across sensing environments.Using more than three distances forms an overdetermined system to solve for the combined antenna transfer functions and feedback function.Source: A. Far-Field Monostatic GPR System Model · p. 3

B. Far-Field GPR Domain Transfer (FFDT) Method

  • FFDT separates reflections, estimates their propagation angles with ray tracing and Snell’s law, transfers each through the SimGPR and RealGPR functions, then recombines them with the RealGPR antenna reflection term.The frequency-domain workflow can operate over sub-bands within the systems’ overlapping frequency range, with inverse Fourier transformation available for time-domain signals.Source: B. Far-Field GPR Domain Transfer… · p. 3

C. Case Study

The case study applies FFDT to B-scans of a buried cylindrical PEC pipe measured by different SimGPR and RealGPR systems. It separates surface and object reflections, transfers them using angle-dependent system responses, and combines them into a B-scan that closely matches RealGPR data.

C. Case Study

  • The case study compares a gprMax-modeled GSSI 1.5-GHz RealGPR system with a simplified dipole-and-field-probe SimGPR system, which use different excitation waveforms and configurations.Both systems use a 0.06 m TX–RX separation; RealGPR uses a Gaussian waveform and SimGPR a Ricker waveform.Source: C. Case Study · p. 3
  • The comparison excludes Ha, which the implementation treats as clutter that can be readily subtracted from far-field GPR data.The simulated time-domain signal is transformed to the frequency domain and restricted to 0.5–2.0 GHz for transfer.Source: C. Case Study · p. 4
  • At each survey position, ray tracing and Snell’s law estimate the reflection angles; the corresponding SimGPR and RealGPR transfer functions map each separated reflection before the components are recombined into A-scans and a B-scan.The surface reflection is obtained from a soil-only simulation and subtracted from the full B-scan to isolate the object reflection.Source: C. Case Study · p. 4
  • FFDT reduces the maximum Diff from 4.67 dB to −45.88 dB, bringing the transferred B-scan almost identical to the RealGPR B-scan from 0.5 GHz to 2.0 GHz.The scenario uses a 0.1 m diameter pipe buried 0.2 m deep; the antennas are 0.6 m above the surface.Source: C. Case Study · pp. 3–4

III. NUMERICAL STUDY

The numerical study tests FFDT across object geometries and dielectric properties, alongside varied soil conditions and far-field applicability. In the PEC geometry cases, angle selection based on dominant reflections yields substantial reductions in B-scan difference, with residual mismatch when reflections have different angles.

III. NUMERICAL STUDY

  • The study evaluates FFDT across object geometries and dielectric properties, soil heterogeneity, surface roughness, layered media, and the far-field criterion.The supplied PEC geometry cases use rectangular and triangular prisms to introduce multiple reflection points beyond the cylindrical case.Source: §III · p. 5, 1) PEC Objects of Different Geometries · p. 5

1) PEC Objects of Different Geometries

  • For the rectangular prism, maximum Diff falls from 5.17 dB before FFDT to −37.47 dB after using the dominant-reflection angle at each survey position.Minor reflections such as edge reflections may retain residual differences because they do not use their corresponding transfer functions.Source: 1) PEC Objects of Different Geometries · p. 5, Table I · p. 6
  • The triangular prism at 0° has maximum Diff of 4.67 dB before and −45.89 dB after FFDT, while at 45° it is 4.73 dB before and −31.59 dB after.At 45°, selecting the tilted-face reflection angle can leave edge reflections mismatched.Source: 1) PEC Objects of Different Geometries · p. 5, Table I · p. 6
  • For the triangular prism, the transfer angle follows a specular face reflection when available; otherwise it uses the angle to the closest vertex.The rectangular-prism strategy selects the top-left or top-right corner angle when the antenna lies left or right of the object.Source: 1) PEC Objects of Different Geometries · p. 5

2) Dielectric Objects of Different Geometries

FFDT is evaluated on dielectric objects and targets with complex internal structures, where transmitted waves and internal boundaries add scattering components. The reported results show reduced differences for both full and object-only B-scans, with angle coverage qualifying performance for internally complex targets.

2) Dielectric Objects of Different Geometries

  • Dielectric objects can transmit electromagnetic waves and scatter from internal boundaries, increasing the complexity of the backscattered field relative to PEC objects.The tested dielectric properties model concrete with 12% moisture content using a single-pole Debye dispersion.Source: 2) Dielectric Objects of Different… · p. 6
  • For dielectric objects, FFDT reduces full B-scan Diff to below −45 dB and object-only Diff to below −30 dB across PEC and dielectric cases.The weaker dielectric-object reflections relative to the soil-surface reflection help explain the smaller full-B-scan Diff; object-only measurements isolate the more affected reflections.Source: 2) Dielectric Objects of Different… · p. 6

3) Objects with Complex Internal Structures

  • For two tree-root-like targets with a high-permittivity core and lower-permittivity bark, maximum Diff after FFDT is less than −45 dB for full B-scans and less than −40 dB for object-only B-scans.The targets are buried 0.2 m deep and produce more complex scattering than homogeneous dielectric objects.Source: 3) Objects with Complex Internal… · p. 6
  • The internally complex targets remain transferable because dominant received reflections occupy a limited angular range, allowing estimated-angle transfer functions to transfer those components effectively.The targets have irregular shapes and internal material contrasts.Source: 3) Objects with Complex Internal… · p. 6

B. FFDT Performance with Different Soil Conditions

FFDT is tested under heterogeneous soil and rough surfaces, with angle estimation based on average soil permittivity and a flat-surface approximation. Differences generally increase as heterogeneity or roughness increases, but remain below −30 dB in the reported challenging or most cases.

B. FFDT Performance with Different Soil Conditions

  • The soil-condition experiments separately assess heterogeneous soil with a flat surface and homogeneous soil with a rough surface, using a cylindrical PEC target buried 0.2 m deep.The section also identifies layered media as part of its broader soil-condition evaluation.Source: B. FFDT Performance with Different… · p. 6

1) Heterogeneous Soil

  • For heterogeneous soil, ray tracing estimates reflection angles using average permittivity over the VWC distribution rather than fine-resolution soil properties.The study generates 50 stochastic soil fractions, with ten random distributions for each VWC variation range.Source: 1) Heterogeneous Soil · p. 6
  • Even at 30% VWC variation, FFDT-transferred B-scans differ from RealGPR B-scans by below −30 dB.Greater VWC variation increases Diff, especially for object-only B-scans, because local permittivity departures from the average increase angle-estimation error and transfer-function mismatch.Source: 1) Heterogeneous Soil · p. 7

2) Soil Surface Roughness

  • For rough-surface tests, FFDT assumes a flat surface to estimate reflection angles, while roughness height is varied from 0 to 10 cm and three frequency bands are examined.The flat approximation reflects the stated practical constraint that field surface profiles are typically unknown and difficult to measure accurately.Source: 2) Soil Surface Roughness · p. 7
  • Across roughness frequency bands, Diff increases with roughness height, but FFDT reduces it below −30 dB in most cases.Degradation is less pronounced at lower frequencies such as 0.4–1.4 GHz, where longer wavelengths reduce roughness sensitivity.Source: 2) Soil Surface Roughness · p. 7

3) Layered medium

FFDT closely matches RealGPR data in layered media when ray tracing estimates the transmitting/receiving angle well. Its adopted far-field criterion bounds reliable operation, although near-field transfer can retain some domain alignment.

3) Layered medium

  • After FFDT, maximum Diff in the two-layer case falls to −42.71 dB for the full B-scan and −43.06 dB for the object-only B-scan.The angle is estimated with ray tracing and Snell’s law across the air–medium and interlayer interfaces; performance remains high when the angle is well estimated.Source: 3) Layered medium · pp. 7–8

C. Determination of the Far-Field Criterion

  • At fixed distance, Diff increases with frequency, while increasing antenna-to-plate distance improves FFDT performance.The distance study covers five frequency bands from 0.4–1.4 GHz to 1.2–2.2 GHz and antenna-to-plate distances from 0.05 m to 0.6 m.Source: C. Determination of the Far-Field… · p. 8
  • When antenna-to-PEC distance exceeds the Fraunhofer distance, Diff remains below −30 dB; the adopted criterion is R ≥ 2D²/λ_h.λ_h is the wavelength at the operating band’s highest frequency.Source: C. Determination of the Far-Field… · p. 8
  • Below the far-field threshold, Diff rises above −30 dB, though it remains around −20 dB at a 0.10 m antenna-to-surface distance.The paper suggests such near-field data may still retain sufficient alignment for basic detection and localization; distance-dependent antenna radiation effects are not modeled.Source: C. Determination of the Far-Field… · pp. 9–10

IV. MEASUREMENT RESULTS

The measurement studies compare FFDT-transferred scans with RealGPR measurements, beginning with a controlled metallic-rod experiment in air using a custom VNA+Vivaldi system. Calibration uses PEC-plate measurements across angles and distances to estimate system transfer functions.

A. FFDT Validation Using a Custom GPR System

  • The custom-system validation uses a metallic rod in air, providing a controlled setting to assess transfer between calibrated SimGPR and RealGPR systems.The rod is 0.032 m in diameter and 0.79 m long, at a 0.6 m antenna standoff; soil is omitted because its properties cannot be replicated exactly between simulation and measurement.Source: A. FFDT Validation Using a Custom GPR… · p. 10
  • Calibration suppresses indoor clutter by subtracting background measurements and time-gating the PEC-plate reflection, then combines measurements at 12 distances into an overdetermined system for each angle.This multi-distance formulation reduces the influence of residual clutter in any individual plate measurement.Source: A. FFDT Validation Using a Custom GPR… · p. 10
  • After FFDT, maximum Diff between the transferred and RealGPR B-scans decreases from 0.02 dB to −20.80 dB, a 20.82 dB reduction.The residual difference is attributed primarily to small positioning errors during calibration and B-scan acquisition.Source: A. FFDT Validation Using a Custom GPR… · p. 10

B. Experimental Validation Using a Commercial GPR System

A commercial Proceq GP8800 experiment tests FFDT with practical GPR hardware. After calibration, the transferred scan more closely matches the GP8800 measurement, while measurement repeatability remains a constraint.

B. Experimental Validation Using a Commercial GPR System

  • The GP8800 is treated as quasi-monostatic because its 3.6 cm antenna separation is small relative to the sensing distance.The test uses the same metallic rod as the custom-system experiment, positioned in air at a 0.45 m standoff.Source: B. Experimental Validation Using a… · p. 11
  • Antenna-coupling correction reduces repeated-measurement maximum Diff from −13.69 dB to −33.91 dB.Even after correction, GP8800 repeatability remains lower than that of the custom system, with repeated-measurement Diff of −33.91 dB versus −51.06 dB.Source: B. Experimental Validation Using a… · p. 11
  • After FFDT, maximum Diff between normalized SimGPR and GP8800 B-scans decreases from 0 dB to −16.28 dB.The transferred scan closely reproduces the measured reflection characteristics.Source: B. Experimental Validation Using a… · p. 11

C. FFDT-Enabled GPR Clutter-Removal DL Model

The clutter-removal study compares the same U-Net trained on SimGPR data or FFDT-transferred data and evaluates both on RealGPR scans. The FFDT-Net better removes clutter while preserving object reflections, supporting RealGPR-domain dataset generation without task-specific field measurements.

C. FFDT-Enabled GPR Clutter-Removal DL Model

  • The comparison trains the same U-Net separately on SimGPR and FFDT-transferred datasets, then evaluates both models on RealGPR B-scans.Sim-Net denotes the model trained on SimGPR; FFDT-Net denotes the model trained on FFDT-transferred data.Source: C. FFDT-Enabled GPR Clutter-Removal DL… · p. 12
  • The simulated clutter-removal dataset contains 1,300 paired scans, doubled to 2,600 through horizontal-flip augmentation.Pairs combine raw and clutter-free B-scans; simulated conditions vary rod location and depth, soil-surface roughness, and antenna-to-surface distance.Source: C. FFDT-Enabled GPR Clutter-Removal DL… · p. 12
  • On RealGPR scans, FFDT-Net removes background clutter and preserves object reflections better than Sim-Net for both shallow and deep targets.For shallow targets, Sim-Net removes much of the overlapping object reflection; for deep targets, it leaves residual clutter and does not fully restore the target reflection.Source: C. FFDT-Enabled GPR Clutter-Removal DL… · p. 13, Figure 18 · p. 13

V. DISCUSSION

  • FFDT enables simulated-data construction of RealGPR-domain training datasets without task-specific RealGPR field measurements for training or fine-tuning.The paper also discusses applying FFDT to learning-based GPR tasks and dataset generation for sites with unknown, spatially variable soil permittivity.Source: C. FFDT-Enabled GPR Clutter-Removal DL… · p. 13, §V · p. 13

A. Applicability and Limitations of the Dominant-Reflection Approximation in FFDT Implementation

FFDT’s dominant-reflection approximation works best when important returns occupy a narrow angular range, but comparable reflections from different directions can cause transfer mismatch. Angle-estimation errors matter most when the SimGPR and RealGPR antennas have different radiation patterns.

A. Applicability and Limitations of the Dominant-Reflection Approximation in FFDT Implementation

  • The single-angle approximation assumes one dominant scattering component or several components in a narrow angular range; comparable strong returns at substantially different angles leave residual magnitude and phase errors.The paper associates this limitation with targets reflecting strongly from both faces and edges.Source: A. Applicability and Limitations of… · p. 14
  • Around −40 dB for small triangular prisms, Diff rises to around −30 dB at 0.2 m, then returns to around −40 dB for larger prisms.At 0.2 m, face and edge reflections can be comparable; for larger targets, one component tends to dominate while spherical spreading weakens the others.Source: A. Applicability and Limitations of… · p. 14

B. Sensitivity of FFDT to Angle-Estimation Error

  • Angle errors have limited expected impact for similar antenna patterns but can cause large transfer errors when the patterns differ; ray tracing and dominant-reflection selection can reduce errors in modeled scenarios.The modeled scenario properties, object positions, and characteristics are known.Source: B. Sensitivity of FFDT to… · pp. 14–15
  • −20.80, −13.75, −9.33, and −2.03 dB are the maximum Diff values for estimated θ, 0.9θ, 0.8θ, and angle-independent 0° transfer, respectively.The comparison uses a metallic-rod experiment with different SimGPR and RealGPR antenna patterns; the figure also shows increasing magnitude mismatch and time shift as angle error grows.Source: Figure 20 · p. 14, B. Sensitivity of FFDT to… · p. 15, Figure 21 · p. 15

C. Applicability of FFDT to Different GPR Systems

FFDT calibration with angle-dependent transfer functions more accurately transfers directional GPR signals than the tested downward-looking calibration methods. The discussion also describes system scope, a detection and classification application, and a strategy for generating site-specific datasets across soil variability.

C. Applicability of FFDT to Different GPR Systems

  • FFDT is formulated for time-invariant, far-field, co-polarized monostatic or quasi-monostatic systems; system differences are represented by transfer functions over the systems’ common frequency band.Quasi-monostatic use assumes sufficiently small TX–RX separation relative to sensing distance; calibration requires controlled target distance and orientation, and stored transfer functions can be reused if configuration remains unchanged.Source: C. Applicability of FFDT to Different… · p. 15

D. Applicability of FFDT to Different GPR Tasks

  • Models trained on FFDT-transferred data achieve higher mAP than models trained on SimGPR data at IoU thresholds 0.50, 0.70, and 0.90 on 45 RealGPR B-scans.The improvement is most pronounced for mAP90; the table is titled for target detection and classification performance.Source: Table II · p. 15, D. Applicability of FFDT to Different… · p. 16

E. FFDT-Based Dataset Generation for Real-World Sites

  • For a site with unknown, variable soil permittivity, the paper proposes estimating a plausible range and simulating scenarios across that range and soil heterogeneity before applying FFDT.The proposed dataset is intended to incorporate site-specific permittivity variability for downstream model training.Source: E. FFDT-Based Dataset Generation for… · p. 16

A. Comparison with GPR Calibration Methods

  • −20.80 dB for FFDT versus −1.96, −2.13, and −2.03 dB for wavelet cross-convolution, wavelet deconvolution, and angle-independent transfer, respectively, in the metallic-rod experiment.The comparison attributes residual magnitude and delay differences at nonzero propagation angles to methods using only a downward-looking reference; the figure compares the three calibration concepts.Source: A. Comparison with GPR Calibration… · p. 16

B. Comparison with Unsupervised Learning-Based Domain Adaptation and Data Augmentation Methods

Existing unsupervised adaptation and generative augmentation methods still require substantial task-specific RealGPR data. FFDT transfers simulated data into a RealGPR domain, and the paper reports close data matches and robust model performance while identifying the configurations covered by the current formulation.

B. Comparison with Unsupervised Learning-Based Domain Adaptation and Data Augmentation Methods

  • Unsupervised domain adaptation and GAN-based augmentation methods are described as effective for task-specific performance but still requiring substantial task-specific RealGPR data during training or adaptation.FFDT is contrasted as a physics-based direct transfer approach.Source: B. Comparison with Unsupervised… · p. 17
  • After one-time system calibration, FFDT generates physically paired RealGPR-domain B-scans from new simulations without task-specific RealGPR field measurements.The paired simulations retain subsurface permittivity ground truth, supporting supervised and unsupervised dataset use.Source: B. Comparison with Unsupervised… · p. 17

VII. CONCLUSION

  • FFDT-transferred data differ from RealGPR ground truth by below −30 dB in most studied cases, and models trained on transferred datasets achieve robust performance on RealGPR data.The conclusion cites experimental clutter removal and numerical detection and classification demonstrations.Source: §VII · p. 17
  • The current formulation covers time-invariant, far-field, co-polarized monostatic or quasi-monostatic systems; near-field, strongly bistatic, and fully polarimetric configurations remain future extensions.Source: §VII · p. 17
Loading 2610.19265v1…