Using neural networks to predict soil water retention and soil hydraulic conductivity

Direct measurement of hydraulic properties is time consuming, costly, and sometimes unreliable because of soil heterogeneity and experimental errors. Instead, hydraulic properties can be estimated from surrogate data such as soil texture and bulk density with pedotransfer functions (PTFs). This paper describes neural network PTFs to predict soil water retention, saturated and unsaturated hydraulic properties from limited or more extended sets of soil properties. Accuracy of prediction generally increased if more input data are used but there was always a considerable difference between predictions and measurements. The neural networks were combined with the bootstrap method to generate uncertainty estimates of the predicted hydraulic properties.