Deep Learning for Advancing Ground-Penetrating Radar under Transport Infrastructure: A Comprehensive Review

Abstract

Ground penetrating radar (GPR) is a non-destructive monitoring and inspection method that has been applied widely across various transport infrastructures such as railways, roads/highways, tunnels and bridges. However, its scalability remains constrained by the reliance on expert-led interpretation, which is increasingly challenged by the large volumes of data generated by modern vehicle-mounted and multi-channel GPR surveys. In this context, deep learning has emerged as a powerful approach, demonstrating considerable efficiency and robustness across multiple stages of the GPR workflow. This paper aims to provide a systematic review of this rapidly emerging research direction, with a particular focus on its applications to transport infrastructure. It begins with an overview of the fundamental concepts of GPR and conventional application methods, highlighting their inherent challenges and establishing the emerging role of deep learning in addressing these limitations. Based on this, the review develops around three major stages of the deep learning-integrated GPR: (i) data acquisition, (ii) data processing including data augmentation, signal attenuation compensation and denoising; and (iii) data interpretation covering object detection, segmentation, and waveform inversion. It synthesises the methodologies and findings of state-of-the-art studies on deep learning for GPR, thereby identifying current limitations and highlighting future research directions. The review highlights the role of physics-based numerical simulation and generative augmentation in producing annotated datasets, while also noting the performance gap between simulated training environments and complex field conditions. It also identifies the practical difficulty of validating continuous inversion models against real-world data.