A deep learning-based approach for directly retrieving GNSS precipitable water vapor and its application in Typhoon monitoring

• Published in: IEEE transactions on geoscience and remote sensing Vol. 62; p. 1
• Main Authors: Huang, Liangke, Lu, Donghui, Chen, Fade, Zhang, Hongxing, Zhu, Ge, Liu, Lilong
• Format: Journal Article
• Language: English
• Published: New York IEEE 01.01.2024; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Artificial neural networks; Atmospheric modeling; Back propagation networks; Bias; Correlation coefficient; Correlation coefficients; Data models; Deep learning; Deep neural network; Delays; Global navigation satellite system; GNSS PWV; Hurricanes; Machine learning; Mathematical models; Meteorological parameters; Meteorology; Monitoring; Navigation; Navigational satellites; Neural networks; Parameters; PWV monitoring; Radiosondes; Real time; Real-time systems; Temperature requirements; Time measurement; Tropical cyclones; Tropospheric delay; Typhoons; Water vapor; Water vapour; Weather; Weather forecasting
• ISSN: 0196-2892; 1558-0644
• DOI: 10.1109/TGRS.2024.3479693

Global Navigation Satellite System (GNSS) offer all-weather and real-time capabilities, enabling the real-time monitoring of precipitable water vapor (PWV). Traditional method for obtaining high-precision GNSS PWV require meteorological parameters such as pressure and the atmospheric weighted mean temperature, which are inconvenient to obtain at GNSS stations, especially in real-time PWV monitoring. To address these challenges, a deep neural network (DNN)-based model, namely D-PWV, has been established to directly convert zenith total delay (ZTD) derived from GNSS signals into PWV without relying on meteorological parameters. The performance of the D-PWV model is comprehensively assessed using GNSS PWV, the fifth generation European Centre for Medium-Range Weather Forecasts reanalysis (ERA5), radiosonde (RS), and data from the 2018 typhoon event Mangkhut in the Guangxi region of China, respectively. The results reveal that the D-PWV model exhibits a bias of -0.25 mm, 0.09 mm, and 0.28 mm when compared to GNSS PWV, RS PWV, and ERA5 PWV, respectively. The corresponding RMSE values are 2.00 mm, 3.37 mm, and 2.74 mm. Notably, when compared to the backpropagation neural network (BP), the D-PWV model demonstrates a smaller range of variation in both Bias and RMSE across the study area and shows greater stability in the latitudinal direction. Besides, the D-PWV model performs stably during Typhoon Mangkhut, with a correlation coefficient of 0.96 with the RS PWV, and the changes on GNSS stations are also highly consistent with the ERA5 PWV, effectively reflecting the PWV variations in the Guangxi region during the typhoon. It shows that the model constructed in this paper can provide real-time PWV values in the absence of measured meteorological parameters.