Enhancing the tree-boosting-based pedotransfer function for saturated hydraulic conductivity using data preprocessing and predictor importance using game theory

Abstract

Machine learning (ML)-based pedotransfer function (PTF) has become a promising alternative to the generic PTFs for predicting the saturated hydraulic conductivity of soils (Ks). This study enhanced the performance of ML-PTF for predicting Ks by utilizing the prominent extreme gradient boosting (XGB) algorithm trained with the cleaned USKSAT containing approximately 18,000 Ks data for U.S. soils. For improving the performance of the developed XGB-PTF, the outliers were detected and eliminated based on the robust Mahalanobis distance (MD). Furthermore, the cooperative game theory was used to quantify the predictor importance on predictions by XGBPTF. High multicollinearity among most predictors in the database indicates the need for including all predictors when using PTFs for Ks and XGB-PTF with the selection of all predictors yielded a comparable performance to ML-PTFs in the literature on the identical database. In addition, the relatively narrow prediction interval reflects the reliability of the presented XGB-PTF, and the substantial improvement on the performance of XGB-PTF was obtained using the robust MD by eliminating 3.7% of the database. Notably, the developed XGB-PTF coupled with the game theory enables identifying the clay content as the most dominant factor affecting the Ks of soils, followed by bulk density (rho b) and 10th percentile particle diameter (d10) for coarse-grained soils and d10 and rho b for fine-grained soils. The three most dominant predictors (clay content, rho b, and d10) found in this study consistent with the observed Ks in the literature indicate the reliable evaluation of predictor importance using the game theory.