Vadose Zone Journal Почвоведение
Pedotransfer functions in the era of machine learning: Data limitations, spatial scale, transferability, and future directions
Аннотация
Abstract Pedotransfer functions (PTFs) are predictive relationships that estimate difficult‐to‐measure soil hydraulic properties from readily available data such as texture, bulk density, and organic matter content. This review synthesizes PTF development from early regression‐based approaches to modern machine‐learning techniques, including artificial neural networks, Random Forest, and physics‐informed approaches. While methodological advances have improved predictive performance, critical challenges persist: (i) transferability limitations when models are applied to different regions and climates; (ii) data representativeness biases favoring temperate European and North American soils; (iii) failure to capture temporal variability induced by climate change and land‐use transitions; and (iv) measurement protocol inconsistencies affecting data quality. We critically evaluate different PTF types (class, continuous, point, parameter, and pseudo‐continuous) and emphasize that statistical metrics alone ( R 2 , root mean square error, etc.) are insufficient for assessing performance. Functional validation within hydrological models and rigorous uncertainty analysis are essential. We present the Transferability, Representativeness, Adaptability, Validation, Interpretability, and Scalability (TRAVIS) framework for evaluating PTFs, alongside a practical guide to PTF selection and validation that considers data availability, climatic context, and transferability risk. Future research should integrate physics‐informed machine learning, climate‐adaptive dynamic PTFs, expanded datasets for tropical and arid regions, standardized measurement and metadata protocols, and transfer learning approaches to improve model generalizability, reproducibility, and predictive performance across diverse climatic, environmental, and soil conditions worldwide. We argue that no single PTF approach is universally optimal and that selection must balance predictive accuracy, interpretability, and transferability with the application context. More fundamentally, we contend that the primary bottleneck constraining PTF performance is not algorithmic sophistication but the representativeness, consistency, and geographic breadth of available training data.