FusionInv-GAN: Advancing GPR Data Inversion With RTM-Guided Deep Learning Techniques
Xiangyu Wang, Guiquan Yuan, Xu Meng, Hai Liu
Abstract
Inversion of ground-penetrating radar (GPR) data is an effective technique for imaging subsurface structures and restoring the physical parameters of mediums. However, the traditional full waveform inversion (FWI) algorithm often produces imaging artifacts and suffers from low computational efficiency. In addition, deep learning-based inversion algorithms frequently overlook the inherent time-depth relationships in the GPR data, leading to contradictions between the mapping of diverse data features to a single model and the uniqueness mapping principle of deep learning algorithms. To address these challenges, a Fusion Inversion Pix2PixGAN (FusionInv-GAN) is proposed for GPR data inversion. This approach utilizes the fused data features of reverse time migration (RTM) imaging results and GPR data to provide correct time-depth relationships for deep learning inversion, with the RTM imaging results serving as a guidance term for precise model predictions. The effectiveness and robustness of the proposed inversion framework are tested on one synthetic and two field GPR data, proving its suitability for geophysical inversion tasks.