A machine learning network for FY-3A MERSI NIR PWV retrieval under all-weather conditions

Abstract

Accurate precipitable water vapor (PWV) monitoring is critical to understanding weather systems and predicting extreme meteorological events. The Moderate Resolution Spectral Imager (MERSI) onboard the Fengyun-3A (FY-3A) satellite provides PWV products with a spatial resolution of 1 km × 1 km. However, the MERSI PWV products contain a large amount of gross errors due to the stripe noise in the raw MERSI observations. Furthermore, the accuracy of MERSI PWV is also affected by weather conditions and land cover types. To address these limitations, we first applied Gaussian spatial filtering to the MERSI observations to attenuate the stripe noise and generate de-striped observations. Then, the MNPNet (MERSI NIR PWV retrieval Net) was developed based on random forest for the first time to retrieve PWV from the combination of GNSS PWV and MERSI near-infrared (NIR) atmospheric transmittance. The new nets well considered land cover types and weather conditions, and obtained amazing accuracy. The PWV derived from 256 GNSS stations in 2017 is used as a reference to validate the MNPNet PWV. The results show that root mean square errors (RMS) of the new net are 1.95 mm for clear weather, 2.68 mm for uncertain weather, 3.07 mm for cloudy weather, and 2.85 mm for all weather, which indicates improvements of 74 %, 78 %, 87 %, and 85 % over the original FY-3A/MERSI PWV. The validation results verify that the proposed algorithm has excellent performance and can markedly improve the all-weather PWV retrieval from FY-3A/MERSI. This study is essential for improving the utilization of MERSI data, weather forecasting, and climate research.