Ensemble machine learning enables continental-scale mapping of saturated hydraulic conductivity across South America

Saturated soil hydraulic conductivity (Ksat) is a critical parameter governing water infiltration, groundwater recharge, and land–atmosphere interactions. However, comprehensive continental-scale Ksat data are limited due to marked spatial variability and insufficient field campaigns at continental scale. This study aims to generate the first continent-wide, machine learning-based Ksat dataset for South America, accompanied by independent hydrological validation. Utilizing an ensemble machine learning framework trained on a newly compiled and harmonized dataset of 778 field observations—the most extensive to date for the continent—and integrating environmental covariates selected for their mechanistic relevance to infiltration processes, we systematically assessed three strategies for outlier management: (A) no noise reduction, (B) basic univariate thresholding (>6 m day -1 ), and (C) a robust multivariate approach using the Minimum Covariance Determinant (MCD). The comparison showed that strategy B adversely affected model performance by excluding hydrologically significant high-conductivity values, whereas strategy C retained the natural range of Ksat while discarding statistically inconsistent data. Among four machine learning algorithms evaluated (Random Forest, SVM, GLMnet, XGBoost), XGBoost applied to strategy C delivered optimal results (RMSE = 0.62; R 2 = 0.57), explaining a substantial proportion of the variance in Ksat. The resulting Ksat map demonstrates pronounced spatial gradients, with higher conductivities in humid, deeply weathered soils of the Amazon and Cerrado, and lower values in the Andes and semi-arid northeastern Brazil. Beyond statistical validation, hydrological relevance was established through positive, albeit region-dependent, correspondence with the Baseflow Index (BFI) across multiple watersheds. The produced dataset effectively captures broad regional trends for hydrological modeling purposes across South America, although its application at finer spatial scales requires caution due to smoothing of small-scale variability. Prediction confidence varies spatially, being higher in well-sampled regions such as southeastern Brazil, and lower in data-scarce areas (Amazon, Andes, Patagonia), where extrapolation beyond the training set introduces greater uncertainty.