Does model complexity of a pedotransfer function for soil hydraulic properties hamper its transferability?

Land surface models need information on soil hydraulic properties (SHP) that are often estimated using pedotransfer functions (PTFs). Due to a lack of calibration data, PTFs must be applied that were trained for regions and land use types outside the region of interest. In this study, we quantify the transferability of PTFs to new regions as function of mathematical complexity and number of covariates. For that purpose, we trained new PTFs for forest soils based on a dataset of 25 soil profiles from climatically moderate regions of Switzerland. In a second step, we tested the new and some existing PTFs in a drier and hotter Swiss region (Valais). Tests of transferability showed that increasing the mathematical complexity (from a linear model to a highly non-linear random forest model) was always beneficial for the predictive power in new regions. Increasing the number of covariates revealed a trade-off between improving the accuracy of the predicted soil water retention curve and reducing accuracy of the soil hydraulic conductivity. Interestingly, the use of environmental predictors (climate data) hampers transferability the most due to large climatic contrasts between the calibration and validation regions. These results suggest that transferability works better for PTFs using fewer predictors. We recommend the use of non-linear PTFs based on soil texture, soil density, and organic carbon content for optimal prediction accuracy in regions without training data. This work highlights that the models with the highest number of predictors are not optimal for achieving transferability and that reducing the number of predictors can be beneficial.