Pedotransfer functions for the soil water characteristics of New Zealand soils using S-map information

Publication date: 15 September 2018
Source:Geoderma, Volume 326
Author(s): S.J. McNeill, L.R. Lilburne, S. Carrick, T.H. Webb, T. Cuthill
Empirical pedotransfer functions (PTFs) have been developed for estimates and the uncertainty of soil water content at different tensions, using explanatory variables from a soil information system for New Zealand (S-map). The explanatory variables include the soil order classification, texture, and parsed information from the S-map functional horizon description. Three models have been considered. The first uses a linear model based on the logit transformation to convert the bounded soil water response range to unbounded form. The second uses beta regression, which models the location and scale of the response separately. Finally, we consider a common response model that includes the tension as an explanatory variable to fit the soil water response at all tensions. A feature of the PTFs is the consistent development of the uncertainty of estimates. All regressions are constrained within the range bounded by 0 and 100%, while the logit transformation and beta regression models are constrained so that response differences are bounded; this ensures that the response is monotonic with respect to tension. For the logit and beta regression methods, the constraint of range (0–100%), monotonicity, and order of uncertainty calculation are simultaneously maintained for all tensions, ensuring that derived estimates such as total available water (TAW), and macroporosity, and their uncertainty, are physically consistent. The available data was split between a fitting and an independent validation dataset used for verification of the uncertainty model. Using the independent validation dataset, none of the models showed any evidence of over-fitting. The logit-transformation model was selected because it provided the lowest reliable estimate of mean absolute error and root mean square error in soil water response, TAW, and macroporosity, with uncertainty estimates based on posterior simulation. In the case of certain soil orders, a bias in the estimated TAW is evident at high values, although its origin is not clear. The selected model is used as the soil hydraulic response pedotransfer function used in the S-map inference engine to provide estimates of water content and available water for a wide range of New Zealand soils.