Machine Learning for Predicting Spatially Variable Lateral Hydraulic Conductivity: A Step Toward Efficient Hydrological Model Calibration and Global Applicability

Recent advances in machine learning (ML) techniques show promise for estimating soil hydraulic properties from soil data sets. Pedo‐transfer functions (PTFs) can facilitate the mapping of the complex relationship between soil properties and soil hydraulic properties, for example, lateral hydraulic conductivity—a necessity for estimating lateral subsurface flow in distributed hydrological models. In wflow_sbm model, the horizontal‐to‐vertical saturated hydraulic conductivity ratio is a sensitive parameter, but no established PTF exists. Our objective is to investigate the potential of ML algorithms in estimating PTFs for prediction. In this study, publicly available calibrated (i.e., optimized) across Great Britain were utilized to train two ML algorithms: Random Forest (RF) and Boosted Regression Trees (BRT), employing SoilGrids data set. Both algorithms effectively predicted in 92 of the 115 tested sub‐basins (i.e., 80%), demonstrating a high correlation with the optimized values, with RF slightly outperforming BRT. As a next step, we compared wflow_sbm simulated discharge results using uncalibrated (default value) and our predicted values. The predictions notably improved discharge simulations, with a median Kling‐Gupta Efficiency (KGE) increasing from 0.55 to 0.75. Subsequently, we generated two globally distributed maps to investigate the transferability of the ML‐based PTFs in the Loire basin, France. ML‐based PTFs improved performance in 75% of sub‐basins, with an average KGE increase of 0.06. Finally, we assessed the uncertainty in predictions, confirming the robustness of the ML‐based PTFs. Our study highlights the potential of ML methods for estimating soil hydraulic properties, aiding parameter estimation for distributed hydrological models.