To obtain high-precision three-dimensional (3D) nutrient distributions in the Zhejiang coastal waters, this study develops a high-resolution reconstruction framework based on machine learning. Focusing on the period from April to October between 2011 and 2020, the study first incorporates river runoff data to effectively correct biases in the GLORYS reanalysis salinity at multiple river estuaries, thereby enhancing the accuracy of input data. Subsequently, the corrected salinity field is integrated with machine learning algorithms to achieve the 3D inversion of nitrate, phosphate, and silicate in shallow layers, followed by an in-depth analysis of their spatiotemporal evolution characteristics. The results demonstrate that salinity constraints significantly improve the accuracy of 3D nutrient reconstruction, which holds potential for the development of integrated marine monitoring and early warning systems for ecological disasters.