## Abstract

Estimates of soil water retention characteristics using pedotransfer functions (PTFs) are useful in many studies, such as hydrological modelling and soil mapping. The objective of this study was to calibrate and validate point and parametric PTF models based on neural networks and the Bootstrap method using different sets of predictors. The point PTF models estimated retention points at -1, -10, -100, and -1500 kPa pressure and the parametric PTF models estimated the van Genuchten retention parameters. A Danish soil data set (3226 horizons) dominated by sandy and sandy loamy soils was used in the analysis. The data were split up into a calibration data set (N=1618 horizons) and a testing data set (N=1608). The data were evaluated with the root mean square residuals (RMSR) and the Akaike Information Criterion (AIC), both obtained from measured and predicted water contents at the four retention points. The results for the point PTF models show similar predictions using detailed soil textural classification (seven textural classes) compared to a simplified textural classification (sand, silt, and clay). In general, we found that adding bulk density (BD) and soil organic matter (SOM) as predictors increased the prediction capability. The RMSR values varied between 0.037 and 0.051 cm ^{3} cm^{-3} and the lowest RMSR values were found for the models that used the most detailed data. The AIC followed the change in RMSR closely. The RMSR values for the parametric PTF models were generally 0.011 cm^{3} cm^{-3} higher than the point PTF models using the same predictors. This was mainly due to an imperfect fit of the van Genuchten retention model to the retention data at 1500 kPa. Adding measured soil water content at -10 kPa in the parametric PTF models reduced the RMSR by 0.006 cm^{3} cm^{-3}. Adding additional soil water contents measured at -1 kPa, -100 kPa, and -1500 kPa improved the predictions only to a minor degree. The uncertainty in the prediction of water content using both the point and parametric PTF increased with increasing clay content. Another finding was that the absolute uncertainty in point PTF predictions of water content at -10 kPa was generally higher than the uncertainty found at -1, -100, and -1500 kPa.

Original language | English (US) |
---|---|

Pages (from-to) | 154-167 |

Number of pages | 14 |

Journal | Geoderma |

Volume | 127 |

Issue number | 1-2 |

DOIs | |

State | Published - Jul 2005 |

## Keywords

- Hydraulic properties
- Neural network
- Pedotransfer functions

## ASJC Scopus subject areas

- Soil Science