Preserving heterogeneity and consistency in hydrological model inversions by adjusting pedotransfer functions

TitlePreserving heterogeneity and consistency in hydrological model inversions by adjusting pedotransfer functions
Publication TypeConference Paper
Year of Publication2015
AuthorsSchaap M, Zhang Y, Xu C, Levi M, Rasmussen C, Larsen J
Conference Name3rd Brazilian Soil Physics Meeting
Date Published05/2015
Conference LocationCuritiban, Parana, Brazil
ARIS Log Number321884

Numerical modeling is the dominant method for quantifying water flow and the transport of dissolved constituents in surface soils as well as the deeper vadose zone.  While the fundamental laws that govern the mechanics of the flow processes in terms of Richards' and convection-dispersion equations are relatively simple in principle, the practical implementation and parametrization of realistic “problems” remains difficult and fraught with many uncertainties.  Besides defining appropriate boundary conditions (e.g., atmospheric forcing by rain and evapotranspiration as well as groundwater fluctuations), the practitioner must decide upon the dimensionality, space and time discretization and the internal structure of the problem (e.g., pedological structure and stratigraphy) and assign realistic hydraulic and chemical properties (water retention and unsaturated hydraulic conductivities and absorption coefficients) to all elements in the numerical grid.   It is well-known that hydraulic properties are difficult to measure, making it virtually impossible to completely cover the variability within the simulated domain.  Instead, inversion methods are often deployed which allows the determination of effective parameters by optimizing hydraulic parameters on observed time-series of moisture contents or matric potentials.  Similar to laboratory measurements, field monitoring is expensive and problems with non-uniqueness of the inversion results often limits the level of detail that can be resolved in the subsurface.  Without additional data or methods, lab measurements nor model inversions can completely characterize all the heterogeneity present at a site and the resulting model may therefore not be able to provide reliable flow estimates.

Pedotransfer functions are an alternative method for estimating hydraulic properties from “cheap” and ubiquitous co-variates such as soil particle size and bulk density.    The method is imperfect, however,  because usually only approximate relations can be established that are database or site-specific and which do often not “transfer” accurately to other locations.   In this presentation we will discuss a method that will allow the preservation of heterogeneity by combining a 3D geostatistical model of texture and bulk density at a deep vadose zone site, with a pedotransfer function and model inversion.  Instead of optimizing hydraulic parameters directly ('direct optimization'), the new method optimizes parameters in simple functions that scale pedotransfer estimates for the 3D geostatistical model.  The new method has several advantages over typical inversion methods: 1) It does not require that the pedotransfer (PTF) model is perfect.  Instead, the method allows the PTF to be adjusted to better represent the site. 2) It does not require that the site is sub-divided in homogeneous subdomains. as is typical for direct inversion methods.  We will show that sub-division is still advisable, but that such sub-domains can retain their internal variability. 3) The new method leads to better results than direct optimization of hydraulic parameters and is amenable to cases where direct optimization is impossible (e.g. when data is sparse).   The new method has a limited number of free parameters, ensuring a quick convergence upon a final solution.

Also presented some further refinements regarding a previously published PTF (Rosetta).  Rosetta was previously optimized to predict van Genuchten-type  (VG) hydraulic parameters without regarding the uncertainty of the VG parameters when they were optimized on measured data.  We show that weighting by uncertainty improves the models performance (in terms of root mean square error) and virtually eliminates model bias.  The new model also allows the computation of a full covariance matrix which is helpful for doing uncertainty analyses in numerical model simulations.