Real-Time GNSS Integrated Water Vapor Sensing Based on Time Series Correction Deep Learning Models

Abstract

AbstractIn the past three decades, GNSS-based Integrated Water Vapor (IWV) retrieval has been intensively investigated, and its products have been widely used in meteorology like severe weather event monitoring. The physical model for the inversion of IWV from the tropospheric Zenith Total Delay (ZTD) requires meteorological data at the location of the GNSS station, such as the surface pressure and the atmospheric weighted mean temperature. However, real-time acquisition of the meteorological data is a very challenging task for most GNSS stations. While proposed empirical models such as Global Pressure and Temperature 3 (GPT3) can provide the meteorological data based on their historical information, larger estimation distortions are found in specific mid- and high-latitude regions. Moreover, we analyzed the seasonal variations in GPT3 prediction errors. In view of the above-mentioned problems, this study implements an IWV conversion model based on a feedforward Deep artificial Neural Network (DNN) and Long Short-Term Memory Network (LSTM) network, which learns historical data from GNSS stations and allows real-time ZTD to IWV conversion without the need of actual meteorological observation but of values only GPT3. Results at four selected mid- and high-latitude GNSS stations show that the Root Mean Square Error (RMSE) of the proposed deep learning method decreases from an average of 3.97 mm to 2.84 mm compared to GNSS IWV retrieved from GPT3. The proposed model provides a broad applicability in real-time GNSS IWV prediction without the availability of real-time measured meteorological data.