Answers are grounded in this paper and cite their sources.
Ground penetrating radar (GPR) full waveform inversion (FWI) is a powerful technique for achieving high-resolution subsurface imaging, but its broader application has been constrained by the computational demands of wavefield simulation and gradient evaluation. To address these challenges, DeepGPR is presented as a dual-parameter GPR-FWI framework for relative permittivity and electrical conductivity, in which CUDA kernel functions are integrated with PyTorch automatic differentiation (AD) through a hybrid computation strategy. The framework incorporates gradient checkpointing and Distributed Data Parallelism (DDP) to reduce GPU memory requirements and to support multi-GPU computation. Through this design, both computational efficiency and algorithmic flexibility are achieved, allowing diverse loss functions and optimizers from the PyTorch ecosystem to be seamlessly employed. Building on this flexibility, a new multi-scale structural similarity (M-SSIM) loss function is introduced for GPR-FWI. Unlike traditional pointwise measures such as mean squared error, the M-SSIM loss emphasizes structural features of the waveform and demonstrates improved robustness to time-shifted signals, thereby mitigating the cycle-skipping problem. Numerical experiments on complex synthetic models show that inversion accuracy is maintained by DeepGPR, while inversion stability is significantly enhanced by the proposed M-SSIM loss, with particularly clear improvements in conductivity inversion. A field example yields an interpreted groundwater-table variation consistent with an independent hydrological observation, supporting the practical feasibility of the workflow. Taken together, these developments establish DeepGPR as a practical, flexible, and extensible platform for advanced GPR imaging.
No takes yet. Share an insight, caveat, or question.
Liu et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: