A New LSTM-based Model to Determine the Atmospheric Weighted Mean Temperature in GNSS PWV Retrieval

The atmospheric weighted mean temperature ( T m ) is a key parameter in determining the precipitable water vapor (PWV). Conventional meteorological parameter empirical models have a lower spatial resolution and poor regional applicability, resulting in lower accuracy in obtaining the T m values in global navigation satellite system (GNSS) PWV retrieval. We discuss a long short-term memory-based ERA5 temperature (LSTM-ERATM) model and evaluate the accuracy of calculating the T m . Considering T m ’s annual, semi-annual, and daily cycle characteristics, an ERATM model was developed based on the ERA5 data from 2017 to 2020 provided by the European Center for Mesoscale Weather Forecasts (ECMWF). Then, the LSTM model was used to train the differences between the T m values obtained by discrete integration of the ERA5 data and T m values calculated by the ERATM model to enhance the accuracy of the ERATM model. We use the ERA5 and sounding data from 2021 to 2022 to analyze the calculation effect of the LSTM-ERATM, ERATM, GPT3, UNB3, and Bevis models. The results show that the ERATM model has broad regional applicability and can provide high-accuracy T m . Compared with the UNB3, GPT3, and Bevis models, the mean root-mean-square (RMS) values of the ERATM model is reduced by 43.4%, 3.4%, and 11.7% respectively when using the ERA5 data as the reference values, and reduced by 22.9%, 13.9%, and 0.2% respectively when using the sounding data as the reference values. Moreover, the accuracy of the LSTM-ERATM is generally better than ERATM at different time points and regions, which shows that the LSTM model effectively improves the accuracy of the ERATM model in calculating T m . For example, the mean RMS values of LSTM-ERATM were reduced by 50.8%, 37.4%, 26.2%, and 18.9% in the next time points of 6:00, 12:00, 18:00, and 24:00 respectively when using the ERA5 data as the reference values, and reduced by 31.3%, 27.2%, 35.9%, and 8.6% respectively when using the sounding data as the reference values. The LSTM-ERATM model in this study provides a powerful tool to improve the accuracy of calculating T m , which can provide more reliable data for meteorology and climate research.