Improving Fengyun-3D satellite atmospheric precipitable water vapor products through machine learning-based post-processing correction

Abstract

Satellite remote sensing has become essential for observing precipitable water vapor (PWV). However, limitations in sensor performance, algorithmic assumptions, and estimation methods can result in errors in satellite retrievals of PWV, which limits the accuracy of these products. This study analyzes the bias and post-processing correction of the Medium Resolution Spectral Imager-II (MERSI-II) PWV operational products aboard China's polar-orbiting meteorological satellite, Fengyun-3D. Initially, a global quality assessment is conducted using AERONET observations from May 2019 to May 2020. Afterward, variations in the product's bias are analyzed using various influencing factors. Validation results show that the four PWV products of MERSI-II tend to underestimate values to varying degrees. Bias varies based on factors such as solar zenith angle, view zenith angle, solar azimuth angle, view azimuth angle, digital elevation model, and normalized difference vegetation index. Based on the bias analysis, factors are selected, and a post-processing correction is implemented on the PWV products using the Extreme Gradient Boosting model. Post-processing correction results show that the mean bias of the PWV products is nearly zero, with minimal impact from the selected parameters. The correlation coefficient of the three-channel weighted product is 0.975, and 76.0 % of the matchups fall within the expected error envelope of ±(0.03 + 0.1 × PWV_AERONET). These research findings assist in minimizing bias and enhancing the quality of MERSI-II PWV products.