Precipitable water vapor (PWV) is a critical factor in precipitation formation, the hydrological cycle, and climate change. In this study, a neural network–based PWV estimation model was established using near‐surface meteorological observation and numerical weather prediction (NWP) products. The inputs of the model include in situ air temperature, relative humidity, and surface pressure; vertical atmospheric profiles of the Global Forecast System (GFS); and other auxiliary data. The SuomiNet Global Positioning System (GPS) data over United States at 2019 and 2020 were used to train and evaluate the model, respectively. The validation results indicate that the estimated PWV achieves a correlation coefficient ( R ) of 0.986, a root mean square error (RMSE) of 0.212 cm, and an average bias of 0.006 cm. In comparison, the R , RMSE, and bias of the model without GFS data are 0.874, 0.63 cm, and 0.08 cm, respectively. These results demonstrate that integrating GFS data significantly enhances the accuracy of PWV estimation. The RMSE of the model shows spatiotemporal dependence. Generally, the RMSE in summer (0.27 cm) is higher than that in winter (0.14 cm), and it is usually lower in the western United States and higher in the eastern. The model’s performance also exhibited a dependence on the water vapor content. Generally, the RMSE increased with water vapor content. In addition, the model proposed in this paper mainly requires traditional meteorological station observations and GFS data, achieving a high correlation coefficient ( R ) of 0.986 and a low RMSE of 0.212 cm, which provides a highly accurate and cost‐effective option for PWV estimation.
Chenghua Xie (Thu,) studied this question.