Bayesian inversion of Mualem-van Genuchten parameters in a multilayer soil profile: A data-driven, assumption-free likelihood function

This paper introduces a hierarchical simulation and modeling framework that allows for inference and validation of the likelihood function in Bayesian inversion of vadose zone hydraulic properties. The likelihood function or its analogs (objective functions and likelihood measures) are commonly assumed to be multivariate Gaussian in form; however, this assumption is not possible to verify without a hierarchical simulation and modeling framework. In this paper, we present the necessary statistical mechanisms for utilizing the hierarchical framework. We apply the hierarchical framework to the inversion of the vadose zone hydraulic properties within a multilayer soil profile conditioned on moisture content observations collected in the uppermost four layers. The key result of our work is that the goodness-of-fit validated likelihood function form provides empirical justification for the assumption of multivariate Gaussian likelihood functions in past and future inversions at similar sites. As an alternative, the likelihood function need not be assumed to follow a parametric statistical distribution and can be computed directly using nonparametric methods. The nonparametric methods are considerably more computationally demanding, and to demonstrate this approach, we present a smaller dimension synthetic case study of evaporation from a soil column. The main drawback of our work is the increased computational expense of the inversion.