Evaluation of Parametric and Nonparametric Machine-Learning Techniques for Prediction of Saturated and Near-Saturated Hydraulic Conductivity

Abstract

Parametric and nonparametric supervised machine learning techniques were used to estimate saturated and near-saturated hydraulic conductivities ( and , respectively) from easily measurable soil properties including the name of the pedological horizon (HOR), soil texture (sand, silt, and clay), organic matter (OM), bulk density (BD), and water contents (θ, θ, θ, and θ) measured at four different matric heads (−10, −100, −1000, and −15,848 cm, respectively). Using a stepwise linear model (SWLM) and the Lasso regression as parametric methods with 316 data in training and 135 data in the testing phase, four pedotransfer functions (PTFs) were obtained in which water contents for both methods play an important role compared with other variables. The SWLM showed better performance than Lasso in the testing phase for log() and log() prediction, with RMSE values of 0.666 and 0.551 cm d and of 0.26 and 0.65. Nonparametric supervised machine learning methods trained and tested with a similar data set significantly improved the accuracy of prediction, with of 0.52, 0.36, and 0.53 for Gaussian process regression (GPR), support vector machine (SVM), and ensemble (ENS) methods in the testing stage. These methods also described 74.9, 66.7, and 72.5% of the variation of log(). Bootstrapping validated the strong performance of nonparametric techniques. The feature selection capability of GPR determined that instead of using a model with all predictors, HOR, silt, θ, and θ are sufficient for the prediction of log(), while HOR, silt, and OM can predict log() as accurate as the comprehensive model with all variables.