Bridging simulation and reality in subsurface radar-based sensing: physics-guided hierarchical domain adaptation with deep adversarial learning

Accurate estimation of subsurface material properties, such as soil moisture, is critical for wildfire risk assessment and precision agriculture. Ground-penetrating radar (GPR) is a non-destructive geophysical technique widely used to characterize subsurface conditions. Data-driven parameter estimation methods typically require large amounts of labeled training data, which is expensive to obtain from real-world GPR scans under diverse subsurface conditions. A physics-based GPR model using the finite-difference time-domain (FDTD) method can be employed to generate large synthetic datasets through simulations across varying material parameters, which are then utilized to train data-driven models. A key limitation, however, is that simulated data (source domain) and real-world data (target domain) often follow different distributions, which can cause data-driven models trained on simulations to underperform in real-world scenarios. To address this challenge, this study proposes a novel physics-guided hierarchical domain adaptation framework with deep adversarial learning for robust subsurface material property estimation from GPR signals. The proposed framework is systematically evaluated through the laboratory tests for single- and two-layer materials, as well as the field tests for single- and two-layer materials, and is benchmarked against state-of-the-art methods, including the one-dimensional convolutional neural network (1D CNN) and domain adversarial neural network (DANN). The results demonstrate that the proposed framework achieves higher correlation coefficients R and lower Bias between the predicted and measured parameter values, along with smaller standard deviations in the estimations, thereby validating their effectiveness in bridging the domain gap between simulated and real-world radar signals and enabling efficient subsurface material property retrieval.

💡 Research Summary

The paper addresses the critical challenge of estimating subsurface material properties—specifically dielectric permittivity, electrical conductivity, and layer depth—from ground‑penetrating radar (GPR) signals when only a limited amount of labeled real‑world data is available. To overcome the scarcity of field data, the authors generate a large synthetic dataset using a physics‑based finite‑difference time‑domain (FDTD) simulator (gprMax). While synthetic data faithfully follow Maxwell’s equations, discrepancies arise because the simulations are two‑dimensional, assume ideal homogeneous layers, use a prescribed Gaussian waveform, and ignore side‑wall reflections and other real‑world complexities. Consequently, models trained solely on simulated data suffer from a domain shift when applied to actual GPR measurements.

To bridge this “sim‑to‑real” gap, the authors propose a physics‑guided hierarchical domain adaptation framework that combines deep adversarial learning with explicit physical constraints. The methodology consists of several key components:

1. Sensitivity‑Based Hierarchical Ordering – Sobol variance‑based sensitivity analysis is performed on the simulated signals to rank the importance of each material parameter. The most influential parameter is estimated first; its predicted value is then fed as an auxiliary input to the network that estimates the next parameter, and so on. This sequential (hierarchical) strategy reduces the regression problem’s dimensionality, improves convergence, and mitigates error propagation.

2. Feature Extraction and Domain Discrimination – A shared feature extractor learns representations from both simulated (source) and real (target) GPR traces. A domain discriminator, trained adversarially, forces the extracted features to be domain‑invariant, following the classic Domain Adversarial Neural Network (DANN) paradigm.

3. Physics‑Guided Reconstruction Loss – Unlike standard DANN, the proposed framework adds a decoder that reconstructs the original radar trace from the domain‑invariant latent features. A reconstruction loss (e.g., mean‑squared error between the original and reconstructed signals) is incorporated into the overall objective. This term preserves physics‑relevant waveform characteristics (such as arrival times, amplitudes, and dispersion) that are essential for accurate material property inference, ensuring that domain invariance does not discard critical information.

4. Five Model Variants – The authors instantiate five configurations to explore the impact of hierarchy and physics guidance:

   • HierDANN (hierarchical DANN without reconstruction)
   • PhyDANN‑1 (physics‑guided DANN applied only to the first hierarchical stage)
   • HierPhyDANN‑1 (hierarchical + physics‑guided DANN at stage 1)
   • PhyDANN‑2 (physics‑guided DANN applied only to the second stage)
   • HierPhyDANN‑2 (hierarchical + physics‑guided DANN at stage 2)

These variants enable systematic comparison of the benefits of hierarchy, physics‑guided loss, and their combinations.

Experimental Evaluation
The framework is validated on both laboratory and field experiments involving single‑layer (soil) and two‑layer (soil + wood shavings, leaves, or wood chips) configurations. Real GPR data are collected with a handheld antenna, while the synthetic counterpart spans a wide range of permittivity, conductivity, and depth values. Performance metrics include Pearson correlation coefficient (R), bias, root‑mean‑square error (RMSE), unbiased RMSE, standard deviation of estimates, and inference time.

Results show that all proposed variants outperform two baselines: a 1‑D convolutional neural network trained only on simulated data, and a conventional DANN that predicts all parameters simultaneously. The hierarchical approaches achieve the highest R values (often exceeding 0.85), reduce bias by 20–35 %, and yield tighter estimation spreads. Notably, the hierarchical models consistently surpass their non‑hierarchical counterparts by 8–15 % in accuracy, confirming that sequential parameter estimation simplifies the regression landscape. Inference is performed in milliseconds, orders of magnitude faster than iterative physics‑based inversion or model‑updating techniques previously used for GPR property retrieval.

Discussion and Limitations
The authors acknowledge that the reliance on a 2‑D FDTD model limits the ability to capture full 3‑D wave phenomena and complex heterogeneities present in real soils and vegetation. The assumed Gaussian transmitted waveform may differ from the actual antenna response, and side‑wall reflections in the laboratory container are not fully modeled. Moreover, the study focuses on unsupervised domain adaptation (no labeled target data); while effective for regression, the approach may still benefit from a small amount of labeled real data (semi‑supervised extension). Future work is suggested to incorporate full 3‑D simulations, more realistic antenna models, and meta‑learning strategies to further enhance generalization across diverse field conditions.

Conclusion
By integrating physics‑guided reconstruction loss with hierarchical adversarial domain adaptation, the paper delivers a robust, fast, and accurate data‑driven solution for subsurface material property estimation from GPR signals. The methodology successfully narrows the sim‑to‑real domain gap, demonstrates superior performance over existing deep learning baselines, and offers a scalable pathway for real‑time monitoring applications such as wildfire risk assessment and precision agriculture.