Machine learning reveals dual controls on global lake chlorophyll-a by temperature regime and land use

Understanding the drivers of chlorophyll-a (Chl-a) is crucial for effective water quality management. Accurate Chl-a prediction remains challenging when relying solely on total nitrogen (TN) and total phosphorus (TP); only a few studies have additionally incorporated climate and landscape variables. However, few studies have integrated short-term temperature regimes, multi-scale land use, and hydrological variables, or examined their synergistic effects with TN and TP in predicting Chl-a. We hypothesize that incorporating these variables into machine learning (ML) models could significantly enhance the prediction accuracy of Chl-a and reveal more complex relationships beyond TP and TN stoichiometry. We used 38 variables in 8 ML models across 2 seasons (spring, summer) and 2 depths (shallow, deep) for 564 global lakes. The XGBoost model ( R 2 = 0.80 and RMSE = 13.68) consistently outperformed the other seven models ( R 2 = 0.36–0.64, RMSE = 10.27–16.52), achieving a robust predictive performance. The XGBoost model built with the three variables (TP, TN, and TN/TP ratio) was also outperformed by the XGBoost model with 38 variables. TP and TN showed limited predictive power for Chl-a, especially in shallow lakes during summer and spring, but TP strongly predicts Chl-a in deep lakes during summer. The partial dependence plot analysis revealed that TP and TN alone were insufficient to explain the complex mechanisms underlying Chl-a in both deep and shallow lakes, whereas land use and temperature variation (i.e., at 7- and 30-day intervals) triggered the release of TN and TP, thereby affecting Chl-a. This study highlighted the importance of integrating land use, temperature variation, and nutrient availability levels into machine learning models to capture accurate Chl-a predictions across global lakes.