Retrieving Precipitable Water Vapor Over Land From Satellite Passive Microwave Radiometer Measurements Using Automated Machine Learning

• Published in: Geophysical research letters 2023-11, Vol.50 (22), p.n/a
• Main Authors: Xia, Xinran, Fu, Disong, Shao, Wei, Jiang, Rubin, Wu, Shengli, Zhang, Peng, Yang, Dazhi, Xia, Xiangao
• Format: Article
• Language: English
• Subjects: Algorithms; AMSR‐2; automated machine learning; Automation; Cognitive tasks; Global positioning systems; GPS; Ground-based observation; Hydrologic cycle; Hydrological cycle; Infrared sensing; Learning algorithms; Machine learning; Microwave imagery; Microwave radiometers; passive microwave; Positioning systems; Precipitable water; precipitable water vapor; Radiometers; Railway stations; Remote sensing; Retrieval; Satellites; Snow cover; Training; Water vapor; Water vapour; Wavelengths
• ISSN: 0094-8276; 1944-8007
• DOI: 10.1029/2023GL105197

Accurately retrieving precipitable water vapor (PWV) over wide‐area land surface remains challenging. Unlike passive infrared remote sensing, passive microwave (PMW) remote sensing provides almost all‐weather PWV retrievals. This study develops a PMW‐based land PWV retrieval algorithm using automated Machine learning (ML) (AutoML). Data from the Advanced Microwave Scanning Radiometer 2 serve as the main predictor variables and high‐quality Global Positioning System (GPS) PWV data as the target variable. Unprecedentedly large GPS training samples (over 50 million) from more than 12,000 stations worldwide are used to train the AutoML model. New predictors with clear physical mechanisms enable PWV retrieval over almost any land surface type, including snow cover and near open water. Validation shows good agreement between PWV retrievals and ground observations, with a root mean square error of 3.1 mm. This encouraging outcome highlights the potential of the algorithm for application with other PMW radiometers with similar wavelengths. Plain Language Summary Precipitable water vapor plays a critical role in the global hydrological cycle, but retrieving its value from remote‐sensed data is challenging, especially for scientific purposes that require high resolution and accuracy. This work proposes a new retrieval algorithm, which is attractive on three accounts. First is the use of information from the microwave radiometer onboard a solar‐synchronous‐orbit satellite, which has a high spatiotemporal resolution. The second attraction is the use of automated machine learning (AutoML), which could circumvent the complex model selection and tuning processes that are typically involved in machine‐learning tasks. Third, an unprecedented large ground‐based data set is gathered from Global Positioning System stations worldwide, which is to be used as target variables for AutoML training. The validation results reveal that the precipitable water vapor retrieval is remarkably successful over all land surface types, which is rarely seen before. The proposed algorithm can also be transferred and used with radiometers onboard other satellites. Key Points A machine‐learning‐based passive microwave land precipitable water vapor (PWV) retrieval method is developed using the latest enhanced Global Positioning System PWV data set Adding new features with clear physical meanings improves the PWV retrieval accuracy by about 30% The model performs well in areas that have been exclud