Estimation of soil permittivity through autoregressive modeling of time-domain ground-penetrating radar data

Recent advances in context-dependent processing for vehicle-based route clearance suggest that information regarding the environmental context associated with ground-penetrating radar (GPR) data can potentially be exploited to improve target detection performance. In this work, a statistical approach to estimating soil permittivity (dielectric constant) from raw time-domain data is presented as an alternative to electromagnetic model inversion. First, a large set of GPR data was simulated using finite-difference time-domain (FDTD) modeling over a heterogeneous subsurface with a rough air/ground interface. Physics-based features were then extracted from the simulated data through autoregressive (AR) modeling of B-scan time slices. A linear least-squares regression model was applied to the features, and experimental results indicate that the dielectric constant of the base soil can be accurately predicted by the regression model. This approach has several advantages over model inversion techniques for estimating soil permittivity, since it is causal, computationally efficient, and does not require an analytical electromagnetic model a priori.