Answers are grounded in this paper and cite their sources.
A central challenge in ground-penetrating radar permittivity inversion is that strong near-surface heterogeneity causes complex superposition of multiple scattering and diffractions, making the inverse problem highly nonlinear and ill-posed. Under limited bandwidth, noise contamination, and imperfect initial models, solutions are prone to nonuniqueness and local minima. Conventional full-waveform inversion (FWI) can exploit waveform information, but it is highly sensitive to the starting model and data quality, often suffering from cycle skipping, excessive interface smoothing, and spurious anomalies. Purely data-driven deep-learning inversion is computationally efficient, yet it lacks explicit physical constraints; when training and field conditions mismatch, it may produce physically implausible structures and exhibits limited stability and generalization. To address these issues, we propose a physics-guided fusion inversion framework. We first construct physically consistent training pairs via finite-difference time-domain modeling and obtain a coarse permittivity estimate using FWI as a physics-derived prior. The radargram (B-scan) and FWI prior are then fused as dual inputs to an enhanced UNet++ network, where a Vision Transformer captures global structural correlations, SimAM emphasizes fine details, and depthwise separable convolutions reduce computational cost. A composite MSE–TV loss is used to suppress noise and artifacts. Synthetic and highway field experiments demonstrate clear improvements over FWI and single-input deep inversion in boundary delineation, layer continuity, and artifact suppression, with enhanced robustness to noise and scenario variability.
No takes yet. Share an insight, caveat, or question.
Li et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: