Sim-to-Real GPR-Based Subsurface Sensing: Physics-Guided Hierarchical-Domain Adaptation With Deep Adversarial Learning

Abstract

Accurate estimation of subsurface material properties, such as soil moisture, is essential for applications, including wildfire risk assessment and precision agriculture. Ground-penetrating radar (GPR) is a nondestructive geophysical technique widely used for subsurface characterization. Physics-based inversion methods, such as iterative model updating, are computationally intensive. Although data-driven approaches offer significantly improved computational efficiency, their performance is often limited by the scarcity of labeled real-world data. To alleviate this limitation, synthetic datasets generated via finite-difference time-domain simulations can be leveraged to train data-driven models. Nevertheless, discrepancies between simulated (source domain) and real-world (target domain) data introduce a significant domain gap, leading to performance degradation in practical deployment. This study proposes a physics-guided hierarchical-domain adaptation framework with deep adversarial learning to enable robust sim-to-real subsurface material property estimation from GPR signals. The framework introduces a family of models, including the hierarchical-domain adversarial neural network (DANN) and its physics-guided hierarchical variants, which eliminate the need for labeled real-world data during training. Extensive laboratory and field experiments involving single-layer and two-layer material configurations demonstrate that the proposed approach consistently outperforms state-of-the-art methods, including the 1-D convolutional neural network and the DANN. The proposed methods achieve higher correlation coefficients ($R$), lower bias, and reduced estimation uncertainty compared to baseline models in estimating relative permittivity, electrical conductivity, and material depth. These results validate the effectiveness of the framework in bridging the domain gap between simulated and real-world radar signals and enabling accurate and efficient subsurface material property estimation.