From cascading drivers to precision management: Integrating causal inference and machine learning for lake trophic state analysis

Understanding the cascading drivers of lake trophic states is critical for advancing sustainable water resource management. This study addresses a key research gap by analyzing large-scale patterns across 863 lakes in the United States, incorporating anthropogenic and natural drivers that are often neglected in single-lake assessments. Using generalized additive models and Spearman correlation analysis, we identified significant associations between trophic states ranging from oligotrophic to hypereutrophic and multiple quantitative variables, including population density, geographic coordinates, physicochemical properties, biological indices, and habitat complexity. Collectively, these drivers accounted for a high proportion of deviance (86.3 %) in trophic states. Strong positive correlations were observed between trophic state and turbidity, total nitrogen (TN), and total phosphorus (TP), while a significant negative correlation was found with Secchi depth (SD, |r| > 0.7, p < 0.05). Furthermore, categorical analyses indicated significant variations in the lake origin, surface area, and the integrity of riparian vegetation across different trophic states. Structural equation modeling (SEM) revealed a pivotal causal cascade wherein urbanization and artificial lake typologies intensified anthropogenic disturbances, leading to riparian habitat degradation, increased nutrient loading, elevated turbidity, reduced light penetration, and diminished zooplankton function, ultimately accelerating eutrophication. Complementary analysis using an eXtreme Gradient Boosting model confirmed turbidity, TN, SD, and TP as dominant nonlinear drivers, reflecting ecological threshold responses and multi-stable states. Zooplankton, as key biological mediators, exhibited complex indirect regulatory roles under multistressor conditions. This integrated framework, linking SEM-derived causal inference with machine learning interpretability, advances a scalable, mechanistically grounded approach to lake eutrophication management. The findings support precision restoration strategies that emphasize habitat complexity and zooplankton community dynamics, offering a compelling alternative to conventional mitigation models for conserving and restoring lakes under increasing anthropogenic pressures.