Radiosonde-informed machine learning correction of ERA5 integrated water vapor: spatiotemporal generalization and external GNSS evaluation

Accurate integrated water vapor (IWV) estimates are essential for atmospheric monitoring, yet reanalyses can retain systematic discrepancies relative to in situ observations. This study develops a radiosonde-informed machine learning framework to correct ERA5-derived IWV over Europe and assess spatial and temporal generalization. Radiosonde profiles from the University of Wyoming and the Integrated Global Radiosonde Archive were collocated with ERA5 and VMF3 products at 116 stations from November 2021 to November 2025. CatBoost used spatiotemporal, geographical, meteorological, atmospheric structural, and climate-zone predictors to estimate ERA5 IWV differences. Spatial generalization was assessed by strict station holdout and temporal generalization by an independent one-year test. The station holdout achieved RMSE values of 1.33 and 2.40 mm and R 2 values of 0.94 and 0.92 in the cold and warm seasons, respectively. In the temporal test, RMSE decreased to 1.11 and 2.06 mm, with R 2 values of 0.96 and 0.93 and near-zero mean biases. Relative to ERA5, RMSE was reduced by 51.5% and 40.1% in the two seasons. The model was then transferred to external GNSS station coordinates to generate radiosonde-informed, spatially collocated IWV estimates. Across 47,418 common observations from 95 valid radiosonde-GNSS pairs, the collocated product reduced RMSE from 3.15 mm for nearby radiosonde observations to 2.25 mm. These results demonstrate robust radiosonde-based correction and spatiotemporal generalization, while highlighting the value of transferring radiosonde observational constraints to locations without collocated sounding measurements.