PREDICTING SOIL SATURATED HYDRAULIC CONDUCTIVITY USING MACHINE LEARNING-BASED PEDOTRANSFER FUNCTIONS

The saturated soil hydraulic conductivity (K_s) is a crucial parameter in various hydrological and climate models, affecting rainfall partitioning between runoff and infiltration. Thus, K_s is essential in many critical applications, such as optimal irrigation design, groundwater recharge estimation, and predicting natural hazards like floods and landslides. Determination of K_s from direct methods is expensive, time-consuming, and often infeasible for large-scale applications. As an alternative, pedotransfer functions (PTFs) are widely used indirect models to estimate K_s from easily retrievable soil parameters. Despite significant efforts, the performance of conventional PTFs in predicting K_s remains poor. Machine learning-based pedotransfer functions (ML-PTFs) have been found to overcome this drawback and outperform conventional PTFs in predicting K_s. The increasing accessibility of large soil databases, coupled with advances in machine learning algorithms, offers new opportunities to enhance the robustness of ML-PTFs. This study employed machine learning algorithms, such as artificial neural networks, Boosted Regression Tree (BRT), and random forests, based on over 20,000 soil samples to develop new PTFs for estimating K_s. We also evaluated the relative importance of predictor variables (soil parameters) used in developing the ML-PTFs. The PTFs implemented with decision tree-based ensemble models outperformed other ML-based models, with BRT yielding the best performance, achieving the highest R² of 0.873 and the lowest root mean square error of 0.336 (log₁₀(cm/h)). Clay content was identified as the most important predictor, followed by nearly equal contributions from bulk density, sand content, and organic carbon content, with silt content being the least significant contributor. The ML-PTFs developed here can be applied to the recently available high-resolution global-scale digital maps of soil properties to generate spatial distribution maps of K_s from regional to global scales.

[This work was supported by the U.S. Department of Agriculture under Non-Assistance Cooperative Agreement 58-6060-6-009. Any opinions, findings, conclusions, or recommendations expressed in this publication are those of the author(s) and do not necessarily reflect the views of the U. S. Department of Agriculture.]