Pedotransfer functions (PTFs) are widely used as an efficient alternative for estimating soil hydraulic properties. However, insufficient understanding of how specific input data characteristics drive PTF prediction performance, coupled with challenges in assessing PTF transferability beyond their development datasets, limit their robustness and broad application. Here, we employed the hierarchical Rosetta3 as the development dataset and evaluated its PTF performance using two independent application datasets (National Cooperative Soil Survey (NCSS) and HYBRAS-V2), comprising over 51,900 samples. To further investigate the effect of the input similarity on PTF performance, the Chamfer Distance (CD) was used to quantify the similarity between the development and application datasets. The extensive NCSS database allowed us to stratify the application dataset by soil temperature regimes, texture classes, and depths for a detailed performance evaluation. Results showed that incorporating additional inputs (e.g., bulk density, field capacity, and wilting point) moderately reduces the correlations between these newly added inputs and the estimation residuals, and that higher residual-input correlations are associated with inferior PTF performance. Furthermore, a lower CD (better resemblance of development and application dataset) leads to better PTF performance. However, increasing input complexity using the hierarchical Rosetta3 models mitigates this effect of resemblance, enhancing robustness across diverse soil and environmental conditions. These findings highlight the importance of analyzing residual-input correlations and suggest that quantifying input-data similarity between PTF development and application datasets can serve as a practical approach to assess the transferability of PTFs.
Shi et al. (Thu,) studied this question.