Transferability of pedotransfer functions for estimating soil hydraulic properties: An analysis of controlling factors for forest soils in Switzerland

• Published in: Geoderma Vol. 460; p. 117397
• Main Authors: Schoch, Julian, Nussbaum, Madlene, Walthert, Lorenz, Carminati, Andrea, Lehmann, Peter
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 01.08.2025; Elsevier
• DOI: 10.1016/j.geoderma.2025.117397
• ISSN: 0016-7061

•Parsimonious Lasso models and Random Forest models were trained for Swiss data.•The transferability of pedotransfer functions was evaluated for forest soils.•Differences in laboratory methods are suggested to influence transferability.•Small covariate set is essential for better generalizability.•A novel approach for testing the controlling factors of PTF transferability. Soil hydraulic properties (SHP) are essential for estimating fluxes in terrestrial ecosystems, plant available water, and root water uptake. To provide SHP for large scale applications, pedotransfer functions (PTFs) are used. Many PTFs are trained for a specific region and its applicability outside this region is controversial. In this study, we analyse the controlling factors affecting PTF transferability across forest soils in Switzerland, focusing on confounders, and the entire modelling framework that we denote as model-building-and-form-of-statistical-function (i.e., the statistical method used to link covariates and responses, as well as model training and selection). We trained parsimonious Lasso models and Random Forest models with data from 24 forest sites located in the Swiss Central Plateau to create new PTFs (SwiPT). These were then transferred, alongside existing European PTFs, to forest soils of another Swiss region (Valais), which is topographically, climatically, and geologically considerably different. Our key finding is that PTFs using fewer covariates (specifically, only sand and clay content) demonstrated in average higher predictive performance when transferred, compared to PTFs using up to 11 covariates. We identify the presence of covariates acting either as confounders or whose measurement uncertainty undermines any predictive gains they might offer, as the main contributors to the limited transferability of PTFs with many covariates. In the context of measurement uncertainty, we discuss how bias introduced by different methods and laboratories could potentially contribute to this limited transferability. In addition, based on our analyses related to model-building-and-form-of-statistical-function, we conclude that effectively limiting or reducing the number of covariates is essential for developing transferable PTFs. This work advances our understanding of the mechanisms limiting PTF transferability and highlights key aspects for improving their generalisation.