Proxy-supervised Sim2Real single-trace GPR clutter reconstruction with adversarial adaptation

Single-trace ground-penetrating radar (GPR) analysis is relevant to data-scarce settings where batch-wise imaging is unavailable or delayed. However, supervised learning for trace-wise clutter reconstruction is constrained by the lack of paired field data, and direct transfer from simulation to measurement is further limited by physical domain mismatch. This study presents a bounded proxy-supervised Sim2Real reconstruction chain for single-trace GPR clutter estimation. The method uses a physics-informed semi-synthetic training strategy, where simulated target responses are injected into measured background traces through spectral adaptive resampling (SpAR), and combines it with adversarial latent-space adaptation for source–target alignment. The network is explicitly trained to predict the clutter component, while the target-related response is treated only as a residual. Experiments on controlled semi-synthetic data show that the proposed framework improves clutter reconstruction over non-adaptive learning baselines and conventional signal-processing methods when the source and target domains retain sufficient distributional overlap. A real-data study on the GPR-SD dataset is further formulated as a proxy-based stress test to examine the behavior of the reconstruction chain under uncontrolled physical mismatch, which bounds the achievable transfer performance. These findings support the feasibility of single-trace clutter reconstruction under controlled semi-synthetic supervision, while also clarifying the physical validity boundary of the current proxy construction.