A Two-Step Machine Learning Framework for High-Resolution Reconstruction of Regional GNSS-Derived PWV Grids Using Meteorological Auxiliary Data

Abstract

High-resolution precipitable water vapor (PWV) gridded products provide essential data for atmospheric state estimation and monitoring of regional heavy precipitation. PWV retrieved from the Global Navigation Satellite System (GNSS) offers high temporal resolution and all-weather observation capability; however, the sparse and uneven distribution of GNSS stations limits its direct application in constructing continuous high-resolution moisture fields. To address this limitation, this study integrates discrete GNSS-PWV observations with high-resolution rapid refresh meteorological information and develops a spatial reconstruction framework to generate PWV grids with a resolution of 0.01$^{\circ }\times$ 0.01$^{\circ }$, enabling high-resolution reconstruction from sparse station observations to a continuous high-resolution water vapor field. First, a light gradient boosting machine (LightGBM)-based PWV spatial interpolation model is developed to generate initial PWV grids. Then, a LightGBM-based bias correction model is constructed using meteorological information to further improve grid accuracy. The results show that PWV spatial interpolation models based on LightGBM, extreme gradient boosting, inverse distance weighting, and ordinary kriging achieve root-mean-square errors (RMSEs) of 0.80, 0.84, 1.19, and 1.26 mm, respectively, with LightGBM showing the best performance. The LightGBM-based bias correction model achieves an RMSE of 0.40 mm. For the PWV grids generated during 2022–2024, the RMSEs relative to GNSS-PWV are 0.38, 0.36, and 0.37 mm, respectively. After bias correction, the RMSEs decrease to 0.30, 0.30, and 0.30 mm, corresponding to accuracy improvements of 21.05%, 16.67%, and 18.92%, respectively. In addition, the reconstructed PWV grids exhibit strong spatiotemporal consistency with precipitation processes, effectively capturing water vapor accumulation prior to heavy rainfall events as well as the spatiotemporal evolution of water vapor during precipitation.