From Pedo to Pedon: Towards the next generation of transfer functions to estimate saturated hydraulic conductivity

Pedotransfer functions (PTFs) are widely used tools to predict soil properties across different spatial scales and are commonly built using regression-based techniques (e.g., multiple linear regression or regression trees) and, more recently, machine learning methods (e.g., artificial neural networks). In these techniques, soil material arising from different soil horizons are treated as independent samples despite the depth dependency that exists for horizons within individual pedons. Here we propose a new approach to build PTFs that takes into account the depth dependency of saturated hydraulic conductivity (K_sat) and refer to this type of depth-dependent PTFs as a “pedontranfer” function (PnTF). Slope (β₁) and intercept (β₀) parameters describing the relationship of log-scale K_sat with soil horizon depth were fit to pedons selected from the Pedogenic and Environmental DataSet (PEDS). The intercept parameter can be interpreted as the K_sat at a 0 cm depth (i.e., K_sat at the soil surface) and β₁ as the rate of change of K_sat with respect to depth. In order to build the PnTF, we used field-based pedon information from PEDS, encompassing approximately 2,000 pedons and >13,000 soil horizons across the United States and estimated K_sat using a generalized Kozeny-Carman equation. Our results show a strong negative linear relationship between β₁ and β₀ (r² = 0.80; P < 0.01). When we predicted the fitted line of the linear relationship between β₁ and β₀ using a multiple linear regression with different soil and climatological variables we found a significant (P < 0.01) and direct relationship, with relatively good agreement (R² = 0.38). Our results suggest that the PnTF approach represents a step forward in the development of the next generation of PTFs, although further research is needed to improve its precision and accuracy. We believe that PnTFs, in principle, have significant advantages over PTFs that should be of interest to the community of developers and users of Earth system and community land models. For example, soil K_sat at depth may be predicted from knowledge only of the surface K_sat since β₁ can be predicted from β₀. Future work should incorporate other soil databases in order to account for systematic biases of the different methods to measure or estimate K_sat.