A machine learning-assisted integrated framework for linking water quality and phytoplankton carbon fixation...

Water quality and carbon-related processes in drinking water reservoirs are increasingly recognized as interrelated challenges in sustainable water management. This study focuses on ecologically managed drinking water reservoirs (hereafter referred to as "ecological reservoirs") in Shanghai and investigates the relationship between water quality and phytoplankton carbon fixation potential. An integrated framework combining water quality assessment, VGPM-based primary productivity estimation, and interpretable machine learning was developed and applied to three reservoirs. The results indicate that improved water quality is associated with higher phytoplankton carbon fixation potential under the studied conditions. The estimated annual phytoplankton carbon fixation potential reached 18.12 kt C yr^-1 based on a VGPM-derived primary productivity proxy. The XGBoost model was used as a data-driven surrogate of the VGPM formulation for predicting primary productivity, and as a data-driven model for predicting the water quality index (WQI), effectively reproducing nonlinear and context-dependent response patterns under the observed environmental conditions. SHAP analysis identified chlorophyll-a (Chla) as a dominant predictor within the surrogate modeling framework. In addition, 2-MIB and fecal coliforms were identified as important indicators of water quality across the three reservoirs. Scenario analysis further indicated that management measures may be associated with concurrent variations in water quality and phytoplankton carbon fixation potential under the modeled conditions. These findings provide management-relevant insights into the association between water quality and phytoplankton carbon fixation processes in managed reservoir systems, while focusing on carbon fixation potential derived from primary productivity rather than net carbon sequestration or full carbon balance.