Abstract: The precise positioning of underground pipelines is crucial for preventing construction accidents and ensuring urban development. As an efficient and non-destructive detection method, ground penetrating radar (GPR) can rapidly acquire subsurface structural information. However, existing GPR data processing largely relies on manual interpretation of B-Scan images, which suffers from complexity, subjectivity, and limited accuracy. To address these issues, this paper proposes a convolutional neural network (CNN)-based inversion method capable of automatically identifying pipeline features from GPR data to achieve precise positioning. First, underground pipeline models with varying materials, dimensions, and burial depths were constructed to simulate the diversity and complexity of real pipelines. Subsequently, the gprMax module was utilized to perform forward modeling on these models to generate a training dataset. An end-to-end deep learning network was then established, taking normalized GPR forward responses as input and outputting target permittivity models to learn the mapping between them. Finally, the trained network was applied to invert observed GPR data to derive the permittivity models of underground pipelines. Quantitative experimental results demonstrate the robustness of the proposed inversion scheme. Even in challenging scenarios where the signal-to-noise ratio ranges from 0 dB down to -15 dB, the method maintains a structural similarity index above 0.86 and a mean squared error below 0.1. These metrics confirm that the algorithm is capable of high-precision inversion regarding pipeline location and dimensions while exhibiting excellent noise immunity. Moreover, validation using measured field data further substantiates the reliability and practical effectiveness of the method for actual GPR applications.

Key words: pipeline detection, ground penetrating radar, deep learning, convolutional neural networks