Machine learning techniques for estimating saturated soil hydraulic conductivity at the watershed scale: advances in pedotransfer functions

Abstract Saturated soil hydraulic conductivity (Ksat) is a key property for soil and water management in agricultural and environmental contexts. Due to the high cost and complexity of direct Ksat measurement, pedotransfer functions (PTFs) based on readily available soil variables are widely used. However, studies focusing on subtropical soils remain limited. This study aimed to develop PTFs for a monitored river basin with subtropical Brazilian soils and evaluate machine learning (ML) techniques for Ksat estimation. A dataset with 105 samples from the Ellert Creek watershed (Rio Grande do Sul, Brazil) was used. Twelve model sets were trained using different combinations of predictors, and six ML algorithms were tested: multiple linear regression, decision tree, random forest, support vector regression, artificial neural networks (ANN), and ANN combined with principal component analysis. Random forest and ANN showed the highest predictive performance, followed by linear regression and support vector regression. Decision trees performed least effectively. The best PTFs, using variables such as sand, clay, bulk density, and macroporosity, achieved R² = 0.75. These results represent a significant advancement for estimating Ksat in subtropical soils, supporting the use of ML-based PTFs where direct measurements are scarce. The developed models can improve hydrological simulations and contribute to sustainable water and soil resource management in subtropical regions. Graphical abstract