Deep learning-based imaging of underground tunnels’ walls via GPR

Abstract

A B-scan is a two-dimensional radar scan generated by moving a ground-penetrating radar (GPR) device along a survey line and recording the reflected signals at each point, providing structural information about underground objects. This dissertation proposes a two-stage inversion scheme to image the walls of underground tunnels via GPR B-scans. The scheme addresses key challenges such as signal distortion caused by heterogeneous soil conditions and deep burial effects, as well as the lack of specialized tunnel datasets. The first stage introduces a label-guided signature enhancement process, where paired homogeneous-heterogeneous B-scan samples help the model to identify and suppress the noise and clutter while preserving critical tunnel features. The second stage leverages both enhanced and raw B-scans to reconstruct dielectric permittivity and tunnel geometries. A multi-receptive-field (MRF) module is incorporated throughout to extract multi-scale tunnel features and improve inversion accuracy. Experiments on a synthetic dataset generated using gprMax validate the scheme’s effectiveness. Results show that the two-stage design reduces MSE by 0.0745 and improves SSIM by 0.0026 compared to single-stage designs. These findings demonstrate that the proposed approach effectively mitigates soil interference and enhances the imaging of underground tunnels’ walls in complex environments.