A Fast Denoising Dataset-Generation Method for Ground Penetrating Radar Based on a Transformer Network Constrained by an Underground Clutter Physical Model

Ground-penetrating radar (GPR) denoising requires large paired datasets that preserve weak target echoes while representing realistic underground clutter, but field measurements are difficult to label, and high-fidelity electromagnetic simulations are costly to scale. This paper proposes a fast denoising dataset-generation method based on a transformer network constrained by an underground clutter physical model. The method decomposes each generated B-scan into a noisy image, a clean target response, a target mask, and three physically interpretable clutter components: direct antenna coupling, random background clutter, and interface-induced multipath. A reference B-scan is used only to estimate gray-level, vertical-energy, and spectral statistics, while transformer modulation branches adjust the physically generated clutter components without copying reference image details. In a 100-sample experiment, the proposed generator produced 512×512 B-scans on a CUDA-enabled GPU in 38.36 s, corresponding to 2.61 B-scans/s and 15.64 exported maps/s. The dataset contained 600 maps, including 500 supervision or component maps. The mean reference correlation was −0.0210, indicating low pixel-level dependence on the reference image. Compared with representative GAN, diffusion, and forward-modeling generators, the proposed method provides explicit target and clutter labels for denoising-oriented dataset construction. Five matched gprMax/FDTD conditions yielded a direct-referenced propagation-delay MAPE of 4.70% and a hyperbola-curvature MAPE of 14.41%, while larger amplitude and multipath errors revealed the expected fidelity–efficiency trade-off of the simplified generator.