Accurate and Full-Coverage Retrieval of Total Column Water Vapor From Chinese UV-VIS Satellites Using an Interpretable Machine Learning Approach

Monitoring total column water vapor (TCWV) with high accuracy and full coverage is critical for understanding Earth’s hydrological cycle. Satellite TCWV observations provide broad spatial coverage but often lack accuracy. Ground-based TCWV measurements, such as global Navigation Satellite System (GNSS), have high accuracy and temporal resolution but low spatial coverage. To address these limitations, we developed an interpretable machine learning (ML) approach to correct TCWV biases in satellite-based blue-band spectral retrievals. Original TCWV retrievals from two Chinese environmental trace gases monitoring instruments (EMIs), with morning and afternoon overpasses, respectively, along with other influencing factors, are used as input features for the model, with GNSS measurements serving as target variables. The corrected dataset provides high-accuracy and full-coverage global TCWV observations twice daily. It overcomes the limitations of individual monitoring techniques and eliminates systematic biases both between different satellites and between satellite and ground-based measurements. The cross-validation against 488 GNSS sites shows an excellent agreement, with a determination coefficient ( $R^{2}$ ) of 0.89 and a mean bias of −0.17 to 0.05 kg/m2, representing an improvement of ~48.8% compared to uncorrected satellite retrievals. The ML interpretability also reveals the impact of physical spectral fitting results, air mass factors (AMFs), geographic information, cloud information, and surface albedo on TCWV retrieval biases. These results provide valuable guidance for enhancing and utilizing satellite TCWV retrievals. Overall, this study enables seamless and accurate global TCWV monitoring, offering crucial insights into the dynamics of the global hydrological cycle and its climatic impacts.