GNSS Zenith Wet Delay prediction from ERA5 using Machine Learning with cross-station generalization

Accurate estimation of the Zenith Wet Delay (ZWD) is essential for GNSS meteorology and atmospheric water vapor monitoring, with important applications in weather forecast and climate monitoring. With the growing availability of reanalysis data sets such as ERA5 and dense GNSS networks, machine learning (ML) offers a powerful means to integrate these data sources and learn the statistical relationships between atmospheric variables and tropospheric delays.

This study presents a machine-learning framework for predicting ZWD using ERA5 atmospheric profiles and a multi-year data set of GNSS observations across Europe. We applied the GNSS ZTD observations from 2018 to 2023, from which ZWD is obtained using Zenith Hydrostatic Delay (ZHD) computed from ERA5. An XGBoost model is trained using GNSS stations from 2018–2022 and evaluated on independent stations excluded from training to ensure that the results reflect true spatial generalization. Under this station-based cross-validation strategy, the model reaches an RMSE of approximately 9 mm on the validation stations and about 9.5 mm on entirely independent test stations in 2023. These results demonstrate that our method can effectively capture ZWD variability and generalize across heterogeneous environments.

By learning a data-driven mapping between ERA5 atmospheric fields and GNSS-derived delays, the proposed approach enables rapid, spatially continuous estimation of ZWD, supporting applications in GNSS meteorology, numerical weather prediction, and climate monitoring.