DBPINet: A Physics-Informed Inversion Network for Martian Subsurface Radar Signal

Subsurface exploration of Mars is essential for understanding its geological evolution and potential water ice distribution. Subsurface radar sounding is an effective technique for detecting layered structure and physical parameters beneath the Martian surface. However, existing methods often neglect the influence of loss tangent and rely on data-driven approaches without physical constraints, limiting their accuracy in high-lossy environments and reducing their physical interpretability. To overcome these limitations, this paper proposes a dual-branch physics-informed network (DBPINet) for the joint inversion of layer thickness, permittivity, and loss tangent of Martian layered media. This method introduces a dual signal loss tangent branch (DSLT-Branch) to extract frequency-dependent attenuation features from dual-frequency radar signals and incorporates a physics-informed loss function based on the electromagnetic transmission-line model to embed physical laws into the learning process. Multiple numerical and measured experiments demonstrate the effectiveness of DBPINet. Compared with the MLP-based baseline and the more advanced LMPINet, DBPINet achieves significant improvements in different layered subsurface models. Specifically, on the three-layer models, the mean absolute percentage error (MAPE) for layer thickness, permittivity, and loss tangent is reduced by 4.793%, 3.600% and 4.559%, respectively. Meanwhile, DBPINet exhibits enhanced robustness under noisy conditions. When applied to real Mars Advanced Radar for Subsurface and Ionosphere Sounding (MARSIS) data acquired over the Medusae Fossae Formation (MFF) region, the inversion results reveal a three-layer subsurface structure (a volcanic ash surface layer, an ice-mixed basaltic middle layer, and a basaltic basement) that is consistent with existing geological interpretations.