Comparing stochastic FDEM inversion methods for near-surface modelling

Abstract:

The spatial distribution of physical properties in the critical zone is often complex due to existing small-scale heterogeneities resulting from both natural and anthropogenic sources. Thus, obtaining numerical three-dimensional models that accurately describe the spatial behavior of these properties is often challenging, but essential for sustainable land management. Geophysical survey methods have proven their potential to image these properties in detail. In particular, frequency-domain electromagnetic induction (FDEM) methods are of interest due to their versatility during field operation, and their sensitivity to two key subsurface properties, namely electrical conductivity (EC) and magnetic susceptibility (MS). Converting recorded FDEM signals into models of subsurface EC and MS describing the spatial distribution of the physical properties requires solving an ill-posed, nonlinear inverse problem with multiple solutions. We address this FDEM inversion problem using to two different stochastic approaches, which allow for uncertainty quantification of the inverted models. The two methods are the Kalman ensemble generator (KEG) and iterative geostatistical FDEM inversion, which are compared for benchmark synthetic and field data sets. The KEG is a Monte Carlo implementation of a Gaussian Bayesian update problem. The iterative geostatistical FDEM inversion uses stochastic sequential simulation and co-simulation as model perturbation technique and converges iteratively based on the misfit between synthetic and observed data. Both inversion methods allow for simultaneous prediction of EC and MS. We discuss the main assumptions and limitations of both methods and compare their inversion results in terms of the quality of spatial prediction and corresponding uncertainty assessment.