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

Zixin Wang \affilnum 1; Ishfaq Aziz \affilnum 2; Mohamad Alipour \affilnum 3 · 2025

Read Bridging simulation and reality in subsurface radar-based sensing: physics-guided hierarchical domain adaptation with deep adversarial learning by Zixin Wang \affilnum 1; Ishfaq Aziz \affilnum 2; Mohamad Alipour \affilnum 3 (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2512-17831.txt.

Introduction

Zixin Wang, Ishfaq Aziz, and Mohamad Alipour want ground radar to read what lies below the soil, such as how wet it is. That reading matters for farm planning and fire risk. Radar models trained on computer made data often fail on real scans, because the two kinds of data differ in subtle ways.

Their fix guides a learning system with physics and aligns the fake and real data step by step. Lab and field tests on one and two layer materials show closer matches to measured values than two standard methods give. The work helps radar tools trained in simulation stay reliable outdoors.

Catalog record

CREATORZixin Wang \affilnum 1; Ishfaq Aziz \affilnum 2; Mohamad Alipour \affilnum 3
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE80 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2512-17831

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Bridging simulation and reality in subsurface radar-based sensing: physics-guided hierarchical domain adaptation with deep adversarial learning Zixin Wang \affilnum 1; Ishfaq Aziz \affilnum 2; Mohamad Alipour \affilnum 3 Abstract † † runninghead: Wang et al. † † affiliation: \affilnum 1Postdoctoral Research Associate, Ph.D., Department of Civil and Environmental Engineering, University of Illinois Urbana-Champaign, Urbana, IL, 61801, USA (corresponding author). Email: [email protected] \affilnum 2Ph.D. Candidate, Department of Civil and Environmental Engineering, University of Illinois Urbana-Champaign, Urbana, IL, 61801, USA. Email: [email protected] \affilnum 3Research Assistant Professor, Ph.D., Department of Civil and Environmental Engineering, University of Illinois Urbana-Champaign, Urbana, IL, 61801, USA. Email: [email protected] Abstract 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. Keywords Domain adaptation; Sim-to-real; Physics-guided machine learning; Ground-penetrating radar; Material property estimation; Soil moisture 1 Introduction In nature, soil is often overlaid by a top layer of vegetation or organic matter, known as biomass. Determining soil moisture is critical in precision agriculture [ 48 ] , water resource management [ 67 ] , identification of wildfire hotspots [ 68 ] , and assessing the likelihood and size of wildfires [ 35 , 44 ] . The top biomass layer directly acts as wildfire fuel, exerting direct control over ignitability and spread [ 58 ] . Moreover, the biomass amount, characterized by depth or volume, is necessary for simulating wildland fires [ 4 ] . Hence, estimating properties of multiple layers formed by soil and the overlaying biomass layer carries significant importance. Traditional methods of estimating soil and biomass moisture include labor-intensive gravimetric tests or in-situ sensors [ 26 , 60 , 13 ] , which are time-consuming and provide only point measurements. Moreover, in the presence of a top vegetation layer or biomass, the deployment of in-situ sensors poses additional challenges. Satellite or passive remote sensing overcomes some of these challenges by providing remote measurement, but suffers from low resolution [ 47 , 40 ] . In contrast, radar-based sensing provides high-resolution measurements, and is used across a wide range of applications, including soil and biomass moisture estimation for agriculture and wildfire assessment [ 56 , 10 , 78 , 79 , 62 ] , permafrost characterization [ 38 ] , nondestructive testing of structures [ 31 ] , building material evaluation [ 2 , 39 ] , and pavement monitoring [ 1 ] . In these applications, ground-penetrating radar (GPR) scans are often used to image the subsurface or estimate the dielectric properties of subsurface materials, such as dielectric permittivity and electrical conductivity [ 79 , 9 , 39 ] . Automated estimation of these dielectric properties from radar scans is often conducted using waveform inversion or data-driven methods [ 62 , 25 , 29 , 24 , 57 , 49 , 80 , 84 , 9 ] . Waveform inversion is the process of estimating the subsurface material properties from recorded radar signals. It involves global optimization or model updating to iteratively update and optimize material properties so that the simulated response matches the recorded radar response in the real world. This often requires hundreds or thousands of forward simulations, which is computationally expensive, even for estimating multiple material properties for a single scenario. On the other hand, data-driven methods, such as deep learning, overcome this limitation by yielding fast inference and retrieval of material properties. However, deep learning methods require a large amount of labeled real-world data for adequate training and accurate prediction, which is often costly and impractical to obtain. In radar-based subsurface sensing, obtaining ground truth data typically requires collecting soil samples for gravimetric analysis and installing through-depth sensors within the soil, which is labor-intensive and time-consuming. Synthetic data generated by physics-based simulations offers the advantages of being easily obtainable and capable of representing the physics of the phenomenon. When the training data is derived from a physics-based model, the resulting framework can be regarded as a physics-guided machine learning paradigm [ 23 ] . However, simulated data cannot fully capture the complex factors that influence the actual system response, resulting in a domain discrepancy between the simulated data (source domain) and real-world data (target domain). Consequently, if the data-driven model is trained on the simulated data, it may not perform well on the real-world data due to the domain shift problem. To mitigate the domain shift inherent in sim-to-real transfer problems, numerous techniques have been proposed to minimize the discrepancy between simulated and real-world domains. Model updating is a widely adopted approach that calibrates the parameters of a simulation model so that its dynamics closely match those observed in real systems [ 46 ] . However, model updating alone cannot fully eliminate the domain gap and often serves as a crucial pre-processing step before applying domain adaptation or domain randomization techniques for sim-to-real transfer [ 77 ] . Domain randomization, as another sim-to-real transfer technique, enhances the diversity of simulated data by randomizing key simulation parameters, thereby enabling the trained model to generalize more effectively to real-world variations [ 69 ] . Nevertheless, this approach comes at the cost of an extensive number of training samples to learn a task, and excessive randomization across multiple parameters can significantly increase task complexity, making it challenging for the model to converge to optimal solutions [ 37 ] . In contrast, rather than relying solely on model calibration or extensive parameter randomization, domain adaptation, a transfer learning technique, leverages shared representations between domains to transfer knowledge learned from a source domain for performing inference tasks in a different but related target domain, thereby providing a promising solution to the sim-to-real problem. It has been widely applied across various sim-to-real tasks, including structural health monitoring [ 28 , 75 ] , machinery fault diagnosis [ 50 ] , image-based condition assessment [ 3 ] , semantic segmentation [ 11

Read the complete text