Artificial neural networks to estimate soil water retention from easily measurable data

Indirect estimation of soil water retention from easily measurable data of soil surveys is needed to extend the applicability of hydrological models. Artificial neural networks (ANN) are becoming a common tool for modeling complex ''input-output'' dependencies. The objective of this work was to compare the accuracy of ANN and statistical regressions for water retention estimation from texture and bulk density. We used data on water contents at eight matric potentials for 130 Haplustoll and 100 Aquic Ustoll soil samples. Although the differences were not always statistically significant, ANN predicted water contents at selected matric potentials better than regression. The performances of ANN and regressions were comparable when van Genuchten's equation was fitted to data for each sample, and parameters of this equation were estimated from texture and bulk density. The precision of parameter estimations was lower than the precision of estimating water contents at a given soil water potential with both ANN and regressions. Grouping samples by horizons improved the precision of the estimates, especially in subsoil. Because they can mimic natural ''many inputs-many outputs'' relationships, ANN may be useful in the estimation of soil hydraulic properties from easily measurable soil data.