Simulation of Large‐Scale Field Infiltration Experiments Using a Hierarchy of Models Based on Public, Generic, and Site Data

Abstract

Comparative simulations of a large‐scale field infiltration experiment at the Maricopa Agricultural Center (MAC) near Phoenix, AZ, were conducted using a hierarchy of models based on public, generic, and site data joined with pedotransfer functions and an inverse procedure. Our purpose was to investigate the ability of simple models and relatively inexpensive data to reproduce and predict reliably the time evolution of water content profiles in nine 10‐m‐deep neutron monitoring boreholes at the site. By relying solely on public sources of information one might conclude that soil at the MAC site is uniform to a depth of 16 m, with a water table at about 22 m. Upon collecting soil samples at the site, we learned the soils are layered and laterally discontinuous, with a perched water table at about 13 m. To identify the least level of complexity required to simulate infiltration at the MAC, we compared models that consider one‐ and two‐dimensional flow in a uniform soil, a soil consisting of uniform layers, and a stratified soil with laterally distinct zones. There is a paucity of hydraulic characterization data for the site. To investigate the feasibility of obtaining hydraulic parameter estimates for the models on the basis of soil type, we ascribed uniform properties to each layer or zone using mean values of three generic databases. To improve these estimates, we ascribed variable soil hydraulic properties to individual soil samples using regression and neural network pedotransfer functions based on soil type and bulk density; we then used them to obtain Bayesian updates of mean hydraulic properties in each layer or zone. We used the various models and mean parameter estimates to simulate water contents during one of several infiltration experiments at the MAC. None of the results compared well with measured water contents, although one set of generic mean parameter values yielded much better results than the other two. Estimating parameters using pedotransfer functions and Bayesian updating does not lead to improved simulations. Only when the parameters are estimated by means of an inverse procedure does one notice a significant improvement in model fit. We compared and ranked the various models and parameter estimates using likelihood‐based model discrimination criteria and confirmed our choice of best model by successfully simulating flow during an earlier infiltration experiment.