Analysis of Forward Model, Data Type, and Prior Information in Probabilistic Inversion of Crosshole GPR Data

The crosshole ground penetrating radar (GPR) is a widely used tool to map subsurface properties, and inversion methods are used to derive electrical parameters from crosshole GPR data. In this paper, a probabilistic inversion algorithm that uses Markov chain Monte Carlo (MCMC) simulations within the Bayesian framework is implemented to infer the posterior distribution of the relative permittivity of the subsurface medium. Close attention is paid to the critical elements of this method, including the forward model, data type and prior information, and their influence on the inversion results are investigated. First, a uniform prior distribution is used to reflect the lack of prior knowledge of model parameters, and inversions are performed using the straight-ray model with first-arrival traveltime data, the finite-difference time-domain (FDTD) model with first-arrival traveltime data, and the FDTD model with waveform data, respectively. The cases using first-arrival traveltime data require an unreasonable number of model evaluations to converge, yet are not able to recover the real relative permittivity field. In contrast, the inversion using the FDTD model with waveform data successfully infers the correct model parameters. Then, the smooth constraint of model parameters is employed as the prior distribution. The inversion results demonstrate that the prior information barely affects the inversion results using the FDTD model with waveform data, but significantly improves the inversion results using first-arrival traveltime data by decreasing the computing time and reducing uncertainties of the posterior distribution of model parameters.

Automated summary

The paper presents a probabilistic inversion algorithm using Markov chain Monte Carlo simulations within the Bayesian framework to infer the posterior distribution of relative permittivity. It is found that using the FDTD model with waveform data successfully infers the correct model parameters, while using first-arrival traveltime data requires an unreasonable number of model evaluations and fails to recover the real relative permittivity field. The smooth constraint of model parameters as the prior distribution significantly improves the inversion results using first-arrival traveltime data by decreasing computing time and reducing uncertainties of the posterior distribution.