In view of the high cost of traditional river water quality monitoring and the sparse ground monitoring stations， based on Sentinel-2 satellite multispectral remote sensing data， combined with MODIS surface temperature， vegetation index， aerosol optical thickness data products， and the surface wind speed data in ERA5 meteorological data products， the monitoring data of the surface water quality monitoring stations with non-optical active parameters Dissolved Oxygen （DO）， Chemical Oxygen Demand （COD） and ammonia nitrogen （NH₃-N） are taken as reference， three machine learning methods， Support Vector Regression （SVR）， Random Forest （RF） and Multilayer Perceptron （MLP）， were used to select the optimal inversion model of each water quality parameter and its corresponding input feature combination through comparative experiments. The experimental results of the model performance test show that the determination coefficients （R²） of DO， COD and NH₃-N are 0.896，0.781 and 0.529， respectively，and the Root Mean Square Error（RMSE） are 0.263 mg/L，0.383 mg/L and 0.061 mg/L， respectively. Compared with the retrieval results using only Sentinel-2 multi-spectral remote sensing data， R² increased by 7.04%， 19.05% and 18.34% respectively， and RMSE decreased by 34.58%， 37.42% and 14.08% respectively. It shows that multi-source remote sensing and meteorological data are of great significance to improve the retrieval accuracy of DO， COD and NH₃-N water quality parameters. Finally， the model robustness evaluation experiment shows that the trained model has better space-time robustness when the representativeness of the model training data is close to the global data.