Machine learning for predicting spatially variable lateral hydraulic conductivity: a step towards efficient hydrological model calibration and global applicability

Recent advances in the application of machine learning techniques to estimate soil hydraulic properties using soil datasets have shown promising results. PedoTransfer Functions (PTFs) can facilitate the mapping of the complex relationship between soil properties and soil hydraulic properties, e.g., lateral hydraulic conductivity—a necessity for estimating lateral subsurface flow in spatially distributed hydrological models like wflow_sbm. The vertical-to-horizontal saturated hydraulic conductivity ratio (f_Kh0) is crucial for model calibration, but an established PTF is currently lacking. Our objective is to investigate the potential of ML algorithms in estimating PTFs for f_Kh0 prediction. First, optimized f_Kh0 across Great Britain (GB) resulting from a sensitivity analysis of the wflow_sbm model (Weerts et al., 2024) were used to train two ML algorithms; Random Forest (RF) and Boosted Regression Trees (BRT), using seven soil parameters from SoilGrids v1.0. Both algorithms effectively predicted f_Kh0 of 92 subbasins (i.e., test set of 25%) with high performance as compared against the optimized values, and RF slightly outperformed BRT. As a next step, we compared wflow_sbm simulated discharge results using uncalibrated f_Kh0 (default value of 100) and predicted values. The predictions notably improved wflow_sbm predictive accuracy by rising the median KGE from 0.55 (using uncalibrated f_Kh0) to 0.75 (using predicted f_Kh0). Following, we generated two globally distributed f_Kh0 maps, allowing us to further investigate the transferability of the ML-based PTFs. Therefore, we tested the predicted f_Kh0 across 559 gauge stations within the Loire basin in France. The utilization of either RF or BRT improved performance in around 75% of these subbasins with a KGE that was, on average, 0.06 higher. Furthermore, f_Kh0 prediction uncertainty and the impact of model spatial resolution were further analyzed. In conclusion, our study demonstrates the potential of ML methods to find relationships between soil properties and (model) soil hydraulic properties, which assists in parameter estimates for distributed hydrological models in gauged and ungauged basins.