Predicting the Water Retention Behavior of Biochar-Amended Soils: A Comparative Study of Six Machine Learning Approaches

Abstract

The soil-water retention curve (SWRC) is a key indicator characterizing soil water retention capacity. Biochar incorporation can affect the SWRC by altering soil pore structure and physicochemical properties. The SWRC of biochar-amended soils is influenced by both soil- and biochar-related factors. To address the time-consuming and labor-intensive nature of conventional SWRC measurement methods, this study employed six machine learning models, Decision Tree (DT), Random Forest (RF), Extreme Gradient Boosting (XGBoost), Categorical Gradient Boosting (CatBoost), Light Gradient Boosting (LightGBM), and Artificial Neural Network (ANN), to predict the SWRC of biochar-amended soils based on a constructed dataset. Feature importance analysis and partial dependence analysis were also conducted to identify key variables and reveal their influence patterns. The results indicated that gradient boosting tree models, XGBoost, CatBoost, and LightGBM, achieved the highest accuracy, with R2 = 0.85–0.89 on the inner-test set and R2 = 0.61–0.64 on the held-out source, demonstrating their advantage in modeling complex nonlinear relationships. Suction was the variable most strongly associated with volumetric water content, while soil particle-size distribution and particle packing density also played important roles. Among biochar-related parameters, ash content showed higher importance than carbon content and pH. The model-derived response of volumetric water content to ash content was non-uniform: limited at low levels, increasing after reaching a certain range, and stabilizing at high levels. In contrast, carbon content and pH had relatively weak marginal effects. A staged hyperparameter optimization strategy combining random search and Bayesian optimization was then applied to optimize the XGBoost model, yielding accuracy comparable to grid search while substantially reducing the training time. Based on the optimized XGBoost model, a prediction platform for the SWRC of biochar-amended soils was developed, enabling volumetric water content prediction, SWRC generation, and van Genuchten (VG) model fitting. This study provides a data-driven approach and a prototype tool for rapid, preliminary SWRC estimation for biochar-amended soils, thereby facilitating the application of biochar in engineering practice.