Constructing High-Precision and Spatial Resolution Precipitable Water Vapor Product Using Multiple Fusion Models

Abstract

Water vapor is a critical parameter in the earth's climate system, affecting precipitation and global warming. Precipitable water vapor (PWV) is a measure of atmospheric water vapor content that can be obtained by multiple methods. In this study, we construct two deep learning models and four other models to obtain a high-precision and spatial resolution PWV fusion product, namely, convolutional neural networks (CNNs), multilayer perceptron, random forest, gradient boosting regression, elastic network regression, and multiple linear regression. The fusion data sources are mainly from Fengyun (FY)3D/medium resolution spectral imager (MERSI) PWV and FY4A/advanced geostationary radiation imager (AGRI) PWV. Moreover, we use PWV derived from 207 global navigation satellite system (GNSS) stations in mainland China to help with model training and testing. The experimental duration lasts two years, from May 2019 to April 2021. The results show that the CNN PWV has the best consistency with the GNSS PWV, with a correlation coefficient of 0.98, a root-mean-square error (RMSE) of 3.01 mm, and a mean bias of 1.00 mm. Regarding the RMSE, CNN PWV shows an improvement of 81.5% and 27.3% when compared to FY3D/MERSI PWV and FY4A/AGRI PWV, respectively. Furthermore, the water vapor maps produced by the CNN model exhibit more precise details than the original PWV products. We also find that the performance of six models has spatiotemporal characteristics, such as winter being better modeled than in summer, and Eastern China being better modeled than in Western China.