Using boosted regression trees to explore key factors controlling saturated and near-saturated hydraulic conductivity

Hydraulic conductivity at and near saturation is difficult to predict. We investigated, for the first time, the potential of boosted regression trees to identify the key factors that determine saturated and near-saturated hydraulic conductivities in undisturbed soils with a global meta-database of tension infiltrometer measurements. Our results demonstrate that pedotransfer functions developed from meta-databases may strongly over-estimate prediction performance unless they are validated against each individual data source separately. For such a source-wise cross-validation, we estimated the hydraulic conductivity at a tension of 10 cm (K10) and the saturated hydraulic conductivity (Ks) with coefficients of determination of 0.36 and 0.15, respectively. The most important predictors for K10 were the average annual precipitation and temperature at the measurement location, which are key variables for pedogenesis and constrain soil management. More research is required for the in-depth interpretation of their influence on hydraulic conductivity. The soil clay and organic carbon contents were also important predictors of K10, with hydraulic conductivity decreasing as organic carbon contents increased up to 1.5% and as clay contents increased between about 10 and 40%. The direction of the tension-sequence with which the infiltrometer data were collected was also a significant predictor. Land use and bulk density were the most important predictors for Ks. The direction of the tension-sequence and the soil texture class were also important, with both coarse and fine-textured soils generally having larger Ks values than medium-textured soils.