Statistical and scaling analyses of neural network soil property inputs/outputs at an Arizona field site.

Analyses of flow and transport in the shallow subsurface require information about spatial and statistical distributions of soil hydraulic properties (water content and permeability, their dependence on capillary pressure) as functions of scale and direction. Measuring these properties is relatively difficult, time consuming and costly. It is generally much easier, faster and less expensive to collect and describe the makeup of soil samples in terms of textural composition (e.g. per cent sand, silt, clay and organic matter), bulk density and other such pedological attributes. Over the last two decades soil scientists have developed a set of tools, known collectively as pedotransfer functions (PTFs), to help translate information about the spatial distribution of pedological indicators into corresponding information about soil hydraulic properties. One of the most successful PTFs is the nonlinear Rosetta neural network model developed by one of us. Among remaining open questions are the extents to which spatial and statistical distributions of Rosetta hydraulic property outputs, and their scaling behavior, reflect those of Rosetta pedological inputs. We address the last question by applying Rosetta, coupled with a novel statistical scaling analysis recently proposed by three of us, to soil sample data from an experimental site in southern Arizona, USA.