Environmental factors and critical thresholds affecting colloidal phosphorus concentration in agricultural runoff:...

Colloidal phosphorus (CP) is a highly mobile form of P that significantly contributes to non-point source pollution, yet its loss dynamics in runoff and the thresholds of influencing factors remain poorly understood due to complex environmental interactions. This study monitored surface runoff over 1.5 year across four cropping systems, revealing CP concentrations of 0.02-0.69 mg L^-1. Linear regression failed to capture the complex, non-linear relationships between environmental factors and CP (R² = 0.66), while machine-learning (ML) models significantly improved predictive performance, with the Stacking algorithm achieving R² = 0.85. Feature importance analysis identified runoff electrical conductivity, soil water potential difference, runoff volume, soil organic carbon (SOC), and pH as key drivers of CP loss. Partial-dependence analysis further revealed critical thresholds, such as SOC of ∼12 g kg^-1, above which CP concentrations declined. Targeted incubation experiments validated that increasing SOC beyond this threshold reduced CP via enhanced P activation and decreased soil P saturation. These findings demonstrate that ML approaches not only enhance CP prediction but also provide mechanistic insight and management-relevant thresholds. Integrating real-time soil and water monitoring with ML-guided interventions offers a scalable strategy to mitigate CP losses and protect water quality in agricultural landscapes.