Study on Vehicle-Mounted GPR Clutter Processing Algorithm

Preprocessing of ground-penetrating radar (GPR) B-scan data is fundamental for effective clutter suppression and weak target enhancement, especially for deeply buried objects with low signal-to-noise ratios. Conventional preprocessing approaches typically depend on manual parameter tuning and multi-stage signal processing pipelines, which limit automation, adaptability, and robustness across varying operational environments. Modern GPR systems commonly operate in equivalent time sampling mode. In vehicle-mounted deployments, mechanical vibrations and environmental electromagnetic interference introduce substantial low-frequency clutter in the deeper sections of B-scan data. Critically, the dominant frequency components of this clutter overlap with the effective signal bandwidths, making frequency-domain separation ineffective. Consequently, traditional filtering-based methods fail to suppress clutter without degrading useful target responses, while alternative enhancement strategies often introduce signal distortion and amplify noise. To overcome these challenges, we propose PreP-GAN, a generative adversarial network (GAN)-based preprocessing framework incorporating multi-scale attention mechanisms for GPR B-scan enhancement. Through adversarial learning, the generator learns to produce high-quality preprocessed B-scan data that preserve intrinsic radar signal characteristics while selectively enhancing weak target reflections. A multi-scale attention module integrating Attention Gates and the Convolutional Block Attention Module (CBAM) is embedded within the network to improve feature representation, enabling the extraction of weak and deeply buried target signatures from clutter-dominated backgrounds. Quantitative evaluation demonstrates that, compared with conventional preprocessing techniques and state-of-the-art deep learning approaches, PreP-GAN achieves a reduction of 106.56 in mean squared error (MSE), along with improvements of 3.81% in structural similarity index (SSIM) and 2.81% in multi-scale SSIM (MS_SSIM) relative to their respective optimal values. Experiments conducted on multi-frequency GPR datasets collected across diverse urban environments further validate the robustness and generalization capability of the proposed framework.