Extended water retention curve models for tropical soils: A critical evaluation using information criteria and Monte Carlo simulation

Abstract The pronounced dry season during the Southern hemisphere’s winter in Brazil and the unusual physical properties of the soils of the Cerrado ecosystem make them candidates for using extended water retention models. In this study, four traditional water retention models: Brooks and Corey (BC), Kosugi (Kos), Rieu and Sposito (RS) and van Genuchten (VG) and five extended water retention models: Freundlund-Xing (FX), Groenevelt and Grant (GG), extended Groenevelt and Grant (GGe), Mehta et al. (MSN) and extended van Genuchten (VGe) were tested in five soils of the Cerrado ecosystem, two Ferralsols, Gleysol, Histosol and Regosol. The dry range was considered as that in which water contents correspond to absolute values of the potential greater than 15000 hPa until air or oven dryness. Extended models extend the water retention curve from saturation into the dry range, while the traditional models describe the water retention phenomena from saturation to a residual water content. The objectives of this research were to critically evaluate the performance of water retention equations using different information criteria (IC) and statistical goodness-of-fit indicators. For the soils evaluated in this study, the BC and VG equations were adequate to describe water retention phenomena corresponding to absolute values of water potential below 15000 hPa . For the extended water retention curve, from saturation to absolute water potentials above 15000 hPa , the VGe followed by FX models were adequate to describe the water retention process. Stochastic simulations show that, overall, the Kashyap information criterion was very sensitive to the number of observations ( N ) and precision ( $$\sigma$$ ) of the data. The performance of the Kashyap information criterion in identifying the best model was affected by decreasing N and increasing $$\sigma$$ , in addition to being model sensitive. In general, the corrected Akaike and Bayesian IC were less sensitive to low N and higher $$\sigma$$ for model selection, which was confirmed by Monte Carlo modeling. Because N in water retention data is usually low, IC should be used with caution for model selection.