Stochastic methods for model assessment of airborne frequency-domain electromagnetic data

Abstract

Bayesian Markov chain Monte Carlo (MCMC) algorithms are introduced for the analysis of one- and two-dimensional airborne frequency-domain electromagnetic datasets. Substantial information about parameter uncertainty, non-uniqueness, correlation, and depth of investigation are revealed from the MCMC analysis that cannot be obtained using traditional least-squares methods. In the one-dimensional analysis, a trans-dimensional algorithm allows the number of layers to be unknown, implicitly favouring models with fewer layers. Assessment of data errors and systematic instrumentation errors can also be incorporated. An example from western Nebraska shows that the MCMC analysis reveals important details about the subsurface that are not identified using a single ?best-fit? model. A geostatistical facies-based parameterization is introduced in order to reduce the number of underlying parameters for the two-dimensional MCMC analysis. This parameterization naturally incorporates lateral constraints in the proposed models, which is important for efficiently sampling the model space. A transdimensional component can be optionally incorporated in the two-dimensional algorithm by allowing the number of facies to vary, but with models that contain fewer facies implicitly favoured.