Predicting the governing factors for the release of colloidal phosphorus using machine learning

Area of Science:

• Environmental Science
• Soil Science
• Water Quality Management

Background:

• Colloidal phosphorus (CP) release from soils is a key factor affecting water quality.
• Traditional models struggle to capture complex soil-CP release dynamics.
• Understanding CP release drivers across diverse land uses (farmland, desert, forest) is crucial.

Purpose of the Study:

• To identify major determinants of CP release in various soil types.
• To compare the predictive performance of machine learning (ML) and traditional models for CP release.
• To investigate the influence of soil properties (Fe, Al, Ca, TOC) and precipitation on CP release.

Main Methods:

• Employed Structural Equation Modeling (SEM), Multiple Linear Regression (MLR), Random Forest (RF), Support Vector Regression (SVR), and eXtreme Gradient Boosting (XGBoost).
• Utilized soil iron (Fe), aluminum (Al), calcium (Ca), total organic carbon (TOC), and precipitation as independent variables.
• Applied SHapley Additive Explanations (SHAP) and partial dependence plots for model interpretation.

Main Results:

• Colloidal-cations (Fe, Al, Ca) and colloidal-TOC were identified as strong drivers of CP release.
• Precipitation and pH exhibited weaker influences on CP release.
• XGBoost demonstrated superior performance (R² = 0.94, RMSE = 0.09), with colloidal TOC being the most critical predictor.

Conclusions:

• The SHAP-enhanced XGBoost model offers superior accuracy in predicting soil CP release compared to MLR, RF, and SVR.
• Colloidal TOC is the most significant variable influencing soil CP release across different land uses.
• Findings provide valuable insights for managing phosphorus levels and mitigating water quality impacts.