Research Progress on Ground-Penetrating Radar Inversion Algorithms: From Physics-Driven, Data-Driven to Fusion Methods

Transparent detection of underground space is crucial to urban safety and resource development, and ground-penetrating radar (GPR) inversion is the core technology to achieve this goal. In recent years, GPR inversion has undergone major changes, shifting from traditional physics-driven methods to deep-learning-driven methods and their fusion strategies. This article systematically reviews the technological evolution of GPR inversion technology, taking 257 documents screened by bibliometric analysis of the Web of Science Core Collection from 2018 to 2025 as the main sample, focusing on the ground-penetrating radar electromagnetic parameter inversion algorithm and on the reconstruction of parameters such as dielectric constant and conductivity. First, traditional physics-driven methods, including ray tomography, hyperbola fitting inversion, and full waveform inversion (FWI), are reviewed; their advantages in physical interpretability are discussed, while their limitations, such as computational inefficiency and strong dependence on initial models, are also pointed out. Subsequently, pure data-driven architectures were classified and sorted according to the three dimensions of A-scan, B-scan and C-scan, covering convolutional neural networks (CNNs), recurrent neural networks (RNNs) and generative adversarial networks (GANs). Additionally discussed is the fact that, although these have the advantage of millisecond-level real-time reasoning, they still lack physical consistency, rely on annotation or simulation data and have limited cross-scenario generalization capabilities. In addition, the article focuses on fusion methods, including model parameter prior regularization, physical information neural networks (PINNs), closed-loop architecture and multi-task collaborative guidance. The article also clarifies the mechanism by which to improve physical consistency through model parameter prior constraints, Maxwell equation residual embedding, differentiable forward closed-loop or network structure enhancement. Finally, three development directions are proposed: enhancing physical interpretability, optimizing real-time inversion calculations, and integrating multi-source heterogeneous data. This review aims to provide a comprehensive reference for the theoretical research and engineering application of GPR inversion methods.