A New Machine-Learning-Based Calibration Scheme for MODIS Thermal Infrared Water Vapor Product Using BPNN, GBDT, GRNN, KNN, MLPNN, RF, and XGBoost

The knowledge of atmospheric water vapor distribution is vital to our understanding of weather and climate. In this article, we propose a new calibration scheme based on machine learning to enhance the observational performance of official all-weather precipitable water vapor (PWV) data records from thermal infrared (IR) measurements of the Moderate Resolution Imaging Spectroradiometer (MODIS) sensor. The calibration scheme takes several influence factors into consideration, which are linked with the performance of satellite-retrieved IR PWV measurements. The ground-based water vapor data, acquired from 214 global positioning system (GPS) sites across China in 2016, are regarded as reference PWV to train the machine-learning-based calibration approaches. The evaluation result during 2017-2019 across China shows that the calibrated MODIS IR all-weather PWV product agrees better with GPS-retrieved reference PWV observations, with $R^{2}$ of 0.88-0.94, root-mean-square error (RMSE) of 2.79-4.08 mm, and mean bias (MB) of 0.16-0.52 mm. The RMSE between water vapor measurements from MODIS and GPS can be reduced by 41.74%, 45.76%, 44.29%, and 49.04% in confident-clear, probably-clear, probably-cloudy, and confident-cloudy conditions, respectively. Our methods, developed based on the new calibration scheme, could be a promising tool to the calibration of other satellite-derived IR all-weather water vapor products, which could be also extended to other regions or time periods.