Construction and optimization of machine learning models based on remote sensing inversion of NH3-N and TP pollution in the Qiantang River Basin, China.

Rapid industrialization, urbanization, and intensive agriculture have worsened river basin water pollution globally, including in China. Traditional water pollution monitoring, though widely used, is time-consuming, limited in coverage, and lacking in spatiotemporal continuity, hindering dynamic water quality parameters (WQPs) analysis. As a key water source in the Yangtze River Delta, the Qiantang River Basin lacks systematic research on WQPs-factor correlations, its driving forces, and spatial differentiation mechanisms. This study focused on the basin, constructing and optimizing machine learning models (via remote sensing and in-situ data) to invert ammonia nitrogen (NH3-N) and total phosphorus (TP), enabling large-scale monitoring with high spatiotemporal resolution. Results showed: 1) Remote sensing reflectance from the Sentinel-2 satellite, climatic factors, and land use types were all significantly correlated with NH3-N and TP concentrations. Using these three as inputs greatly improves the model performance. For TP inversion, the coefficient of determination (R2) of the test set increased from 0.1874 to 0.7062, and the root mean square error (RMSE) decreased from 0.0291 mg/L to 0.0255 mg/L; for NH3-N inversion, the R2 of the test set rose from 0.2833 to 0.5513, and the RMSE dropped from 0.2041 mg/L to 0.1372 mg/L 2) Land use exerted the strongest influence on NH3-N and TP. 3) Human activities and policies drove the spatial differentiation of NH3-N and TP: agricultural plains like the Wuyi River basins showed high pollution (NH3-N: 4.5 mg/L; TP: 0.5 mg/L), while upstream controlled areas such as Changshan Port met Class I-II water quality standards. These WQPs inversion models offer a scientific basis for the basin's water environment governance, supporting the sustainability of water resources in the Yangtze River Delta region.