Deep Learning Inversion of Ground-Penetrating Radar With Prior Physical Velocity Field for Complex Subsurface Root System

The inversion of ground-penetrating radar (GPR) data is beneficial for studying the subsurface media distribution, especially for the root system and their environment. Emerging GPR inversion models focus on constructing a direct mapping from radar wave B-Scan image to permittivity distribution. They not only rely on the diversity of the dataset, but also are hard to specify the effects of electromagnetic wave propagation properties. These make it difficult to accurately recover complex underground scenes. Addressing the above issues, this article proposes a radar wave inversion model based on the embedded velocity field (velocity field inversion UNet, VFIUNet), which utilizes the estimated global velocity information within the detection domain to reconstruct the subsurface permittivity distribution from B-Scan data. First, the propagation velocity of radar wave in the detection domain is estimated through self-focusing FK migration. Then, an embedded physical velocity field inversion model based on the multiscale transformation module (MSTM) was designed, which integrates the prior velocity field and the collected B-Scan data to estimate the difference in permittivity distribution. Finally, the parameters of the VFIUNet are optimized with the hybrid loss function [including L2 and structural similarity index measure (SSIM)] and the root scene radar wave dataset. Comparative studies with existing inversion models exhibit that inversion models based on a priori velocity information provide better performance in the permittivity distribution reconstruction. The predictions of numerical experiments and in situ root detection indicate that the permittivity, shapes, and positions of the subsurface media could be restored by the proposed inversion model with high accuracy. Accurate GPR inversion would facilitate the 3-D reconstruction of subsurface root scenes.