A Window‐Augmented Machine Learning Approach for Direct GNSS Precipitable Water Vapor Retrieval

Abstract Global Navigation Satellite Systems (GNSS) provide an effective means for remote sensing of precipitable water vapor (PWV). However, conventional GNSS‐based PWV retrieval approaches rely heavily on auxiliary meteorological parameters, which are frequently unavailable in real time, while most direct retrieval methods use station data aggregated over large regions, overlooking local characteristics. To address these issues, this study proposes three innovative window‐augmented machine learning models—RF‐G, BP‐S‐G, and BP‐D‐G—that directly estimate PWV from GNSS signals without dependence on meteorological data. They were trained using zenith tropospheric delay (ZTD) and PWV data from the ERA5 reanalysis, requiring only GNSS station coordinates, observation times, and ZTD values as inputs. Validation was performed using PWV derived from GNSS and meteorological observations (GNSS‐M PWV) and radiosonde observations in China. Compared with the previously developed D‐PWV model, the proposed models achieved superior accuracy. Using GNSS‐M PWV as reference, the RMSEs for the D‐PWV, RF‐G, BP‐S‐G, and BP‐D‐G models were 2.00 mm, 1.67 mm, 1.88 mm, and 1.74 mm, representing improvements of 16.5%, 6%, and 13%, respectively. Against radiosonde data, the RMSEs were 3.22 mm, 2.62 mm, 3.07 mm, and 2.92 mm, corresponding to improvements of 18.6%, 4.7%, and 9%. To further examine spatial transferability, the three models were additionally evaluated over Australia using GNSS‐M PWV as reference, yielding RMSEs of 2.60 mm, 3.49 mm, and 2.92 mm. Furthermore, the models maintained stable performance during extreme weather events. These results demonstrate that the proposed models enable accurate GNSS PWV retrieval while eliminating dependence on meteorological parameters.