Development of pedotransfer functions for estimation of soil hydraulic parameters using support vector machines

A number of hydrological models used to simulate situations ranging from field-scale water flow to global climate change rely on numerical techniques that simulate heat, water, and solute fluxes in the vadose zone. The use of flow models for variably saturated conditions requires accurate estimates of the hydraulic characteristics that govern water retention and water flow in soils (Wosten et al., 2001). The hydraulic characteristics of soils vary spatially from one location to another, and are also scale-dependent (Hopmans et al., 2002). The temporal variability can also occur as a result of various biological and human activities, such as root-growth, soil management and agricultural practices, or Modeling flow in variably saturated porous media requires reliable estimates of the hydraulic parameters describing the soil water retention and hydraulic conductivity. These soil hydraulic properties can be measured using a wide variety of laboratory and field methods. Frequently, this proves to be an arduous task because of the high spatial and temporal variability of soil properties. In the last decade, researchers have shown a keen interest in developing a class of indirect approaches, called pedotransfer functions (PTFs), to overcome this problem. Pedotransfer functions predict soil hydraulic parameters using easily obtainable soil properties such as textural information, bulk density and/or few retention points. In this paper, we use a new methodology called Support Vector Machines (SVMs) to derive a new set of PTFs. Support vector machines represent a pattern recognition approach where the overall prediction error and complexity of the SVM structure are minimized simultaneously. We used the same database that was utilized to develop ROSETTA to generate the SVM-based PTFs. The performance of the SVM-based PTFs was analyzed using the coefficient of determination, root mean square error (RMSE) and mean error (ME). All soil hydraulic parameters estimated using the SVM-based PTFs showed improved confidence in the estimates when compared with the ROSETTA PTF program. Estimates of water contents and saturated hydraulic conductivities using the hydraulic parameters predicted by the SVM-based PTFs mostly improved compared with those obtained using the artificial neural network (ANN)-based ROSETTA. The RMSE for water contents decreased from 0.062 to 0.034 as more predictors were used, while the RMSE for the saturated hydraulic conductivity decreased from 0.716 to 0.552 (dimensionless log_Ļ0 units). Similarly, the bias in the water contents estimated using the SVM-based PTF was reduced significantly compared with ROSETTA.