Stacking machine learning model for vertical interpolation of precipitable water vapor using GNSS networks and radio occultation data

This study proposes a stacking machine learning (SML) model for the vertical interpolation of precipitable water vapor (PWV), addressing the issue of missing near-surface water vapor information in radio occultation (RO) profiles, where measurements do not reach the surface. The model integrates data from over 1500 ground-based Global Navigation Satellite System (GNSS) stations and more than 300,000 profiles from the Constellation Observing System for Meteorology, Ionosphere, and Climate-2 (COSMIC-2), and is trained and validated in two regions of the Northern Hemisphere. Results show that in the North American region, the SML model reduces the root-mean-square error (RMSE) of PWV estimates by over 33% compared to conventional models. In China and Southeast Asia, the RMSE reduction is about 42%, demonstrating notable improvements over conventional model approaches. External validation with radiosondes shows close agreement between the SML-interpolated RO-PWV and radiosonde-PWV. Additionally, the SML-interpolated RO-PWV exhibits high consistency with PWV estimates from RO profiles of the Meteorological Operational satellites (Metop by ROMSAF), which contain complete (background) near-surface information. The SML model performs reliably across various terrains and climatic conditions. This study also preliminarily explores the model performance for extreme weather conditions, providing insights for future enhancements. The new SML model provides a high-accuracy and effective solution for PWV estimation and contributes to advancements in water vapor monitoring, weather forecasting, and atmospheric science research.