Subsurface Cavity Imaging Based on UNET and Cross–Hole Radar Travel–Time Fingerprint Construction

As a significant geological hazard in large–scale engineering construction, deep subsurface voids demand effective and precise detection methods. Cross–hole radar tomography overcomes depth limitations by transmitting/receiving electromagnetic (EM) waves between boreholes, enabling the accurate determination of the spatial distribution and EM properties of subsurface cavities. However, conventional inversion approaches, such as travel–time/attenuation tomography and full–waveform inversion, still face challenges in terms of their stability, accuracy, and computational efficiency. To address these limitations, this study proposes a deep learning–based imaging method that introduces the concept of travel–time fingerprints, which compress raw radar data into structured, low–dimensional inputs that retain key spatial features. A large synthetic dataset of irregular subsurface cavity models is used to pre–train a UNET model, enabling it to learn nonlinear mapping, from fingerprints to velocity structures. To enhance real–world applicability, transfer learning (TL) is employed to fine–tune the model using a small amount of field data. The refined model is then tested on cross–hole radar datasets collected from a highway construction site in Guizhou Province, China. The results demonstrate that the method can accurately recover the shape, location, and extent of underground cavities, outperforming traditional tomography in terms of clarity and interpretability. This approach offers a high–precision, computationally efficient solution for subsurface void detection, with strong engineering applicability in complex geological environments.