Unraveling multi-scale drivers of Chl-a dynamics via interpretable machine learning: a case study of Shahu Lake

Abstract

Aquatic eutrophication exhibits pronounced temporal variability, yet conventional global interpretation frameworks often fail to capture short-term ecological responses and obscure environmental drivers that dominate only during specific ecological periods. To address this limitation, this study developed a multi-scale interpretable machine-learning framework and applied it to Shahu Lake as a case study. Using high-frequency water-quality monitoring data, an XGBoost model was developed to simulate chlorophyll-a (Chl-a) concentrations, achieving an \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\varvec{R^2}$$\end{document} of 0.889 for the training dataset and 0.810 for the independent testing dataset, indicating both strong predictive accuracy and good generalization performance. Global-scale interpretation showed that water temperature and organic matter mineralization constituted the dominant long-term environmental drivers, while phosphorus remained the fundamental limiting nutrient. Regional-scale interpretation further revealed a distinct seasonal mechanistic shift during the summer bloom period (July–September), when the system transitioned from long-term phosphorus limitation to transient nitrogen limitation under elevated thermal conditions. At the local scale, heterogeneous decision pathways underlying extreme high-Chl-a events were identified, demonstrating that bloom development followed different nutrient-regulation trajectories even under similar environmental backgrounds. These findings suggest that long-term eutrophication management should prioritize phosphorus-load reduction, while adaptive nitrogen management may be required during the summer bloom period. By unraveling ecological mechanisms across global, regional, and local interpretation scales, the proposed framework provides a state-dependent decision-support tool for lake eutrophication management and demonstrates the potential of interpretable machine learning for understanding complex aquatic ecosystems.