An All-Weather Land Precipitable Water Vapor Retrieval Method Integrating Physical Constraints and Machine Learning

Atmospheric precipitable water vapor (PWV) is an essential geophysical parameter for weather forecasting, climate change research and hydrological studies. Spaceborne passive microwave sensors support all-weather PWV observation, yet accurate retrieval of land PWV still faces great difficulties owing to the spatiotemporal complexity and variability of land surface emissivity (LSE). This study proposes a novel all-weather land PWV retrieval approach using Fengyun-3D microwave radiation imager (MWRI) observations, which combines physical radiative transfer principles with machine learning techniques. By reorganizing the radiative transfer equation, this study constructs two characteristic factors that can effectively characterize LSE and PWV, and further incorporates them into the machine learning retrieval framework. Three typical machine learning algorithms, namely Random Forest (RF), Adaptive Kernel Extreme Learning Machine (AKELM), and Extreme Gradient Boosting (XGBoost), achieve the optimal retrieval accuracy, with an RMSE of 3.80 mm and an R2 of 0.91. Multiple validation results indicate that (1) compared with conventional physically constrained machine learning retrieval methods, the proposed method reduces the root mean square error (RMSE) by 16.5%; (2) against ground-based SuomiNet site observations, the minimum retrieval RMSE is as low as 2.95 mm; (3) the proposed method exhibits stable superiority over the Moderate Resolution Imaging Spectroradiometer (MODIS) MYD05 water vapor products with lower RMSE in all experimental cases. This study offers an effective and reliable technical scheme for all-weather land PWV retrieval, and can facilitate high-precision atmospheric monitoring and related practical operational services.