Skip to content
Open access

Interpretable Machine Learning for Predicting Soil Hydraulic Properties from Texture and Organic Carbon Using Global Soil Data

Oct 2026 · Civil Engineering Journal · 0 citations · 46 references

Abstract

Accurate characterization of soil hydraulic properties is essential for hydrological modeling and agricultural water management, yet direct measurements remain costly and spatially sparse, creating a need for reliable approaches that can estimate these properties at large spatial scales from readily available soil information. The objective of this study was therefore to develop and evaluate an interpretable machine learning framework for predicting volumetric water content at field capacity (θ₃₃), permanent wilting point (θ₁₅₀₀), and available water capacity (AWC) using particle-size distribution and soil organic carbon (SOC). The analysis used 14,126 soil-layer observations from 4,390 profiles obtained from the World Soil Information Service (WoSIS). Random Forest (RF) and Extreme Gradient Boosting (XGBoost) models were evaluated using 10 repeated profile-level splits and leave-region-out validation (RF only). The findings indicated that incorporating SOC improved predictive performance for all three hydraulic properties. RF with texture and SOC achieved mean R² values of 0.716, 0.703, and 0.477 and mean RMSE values of 0.079, 0.057, and 0.052 cm³ cm⁻³ for θ₃₃, θ₁₅₀₀, and AWC, respectively. Geographical transferability varied among regions and target properties. The novelty of this study lies in integrating globally distributed soil data, readily available predictors, geographical validation, and model interpretation within a unified predictive framework.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.