Enhancing satellite-based turbidity monitoring in drinking water source lakes using optical classification–guided machine learning

Turbidity is a key indicator of inland water quality and a critical parameter for the management of drinking water source lakes. In optically complex lakes, pronounced spatial heterogeneity in water optical properties often limits the reliability of conventional remote sensing–based turbidity retrieval methods, constraining their application in regional water source management. This study proposes an optical classification–guided remote sensing framework integrating machine learning to improve turbidity estimation in heterogeneous drinking water source lakes. Waters are first classified into algae dominated water, transitional water, and suspended particulate matter dominated water using K-means++ clustering based on optical characteristics, enabling water type–specific model calibration. Model evaluation shows that eXtreme gradient boosting (XGBoost), BP neural network (BP), and random forest (RF) achieve optimal performance for algae-dominated (MAPE = 13.8%), transitional water (MAPE = 19.7%), and suspended particulate matter–dominated (MAPE = 13.7%) water, respectively. Application of the framework to three large regional drinking water source lakes reveals strong spatial heterogeneity in turbidity and consistent seasonal patterns, with higher turbidity in autumn and lower levels in spring. Mean turbidity ranges from 36.1 to 62.5 NTU across the lakes, indicating substantial variability in regional source water quality. Further analysis indicates that meteorological forcing is the dominant control on turbidity variability, explaining 88%–91% of the observed changes. Overall, the proposed framework provides a robust and transferable approach for regional-scale turbidity monitoring in optically complex lakes and supports drinking water source protection and adaptive water quality management.