Improving the Accuracy of MODIS Near-Infrared Water Vapor Product Under all Weather Conditions Based on Machine Learning Considering Multiple Dependence Parameters

We developed six machine-learning-based calibration models to improve the all-weather accuracy of precipitable water vapor (PWV) product from near-infrared (NIR) observations of the Moderate Resolution Imaging Spectroradiometer (MODIS) instrument, i.e., MOD05 PWV. The six machine learning approaches are: back propagation neural network (BPNN), gradient boosting decision tree (GBDT), generalized regression neural network (GRNN), K-nearest neighbor (KNN), multilayer perceptron neural network (MLPNN), and extreme gradient boosting (XGBoost). The input of the models included MOD05 PWV, latitude, longitude, elevation, cloud, season, and solar zenith angle, in association with the quality of the MOD05 PWV product. PWV data measured from in situ 453 Global Positioning System (GPS) stations in Australia in 2017 were utilized as the target water vapor data for model training. The validation results in Australia in 2018–2019 indicate that the models can significantly improve the all-weather quality [increased
$R^{2}$
, reduced root-mean-square error (RMSE), and reduced mean bias (MB)] of the MOD05 PWV product, exhibiting a reduction in RMSE of 53.33% for BPNN, 55.25% for GBDT, 53.24% for GRNN, 37.81% for KNN, 54.98% for MLPNN, and 55.16% for XGBoost. The
$R^{2}$
was 0.58–0.79 and the MB was 0.12–0.72 mm, much better than the MOD05 all-weather PWV product (
$R^{2} = 0.26$
and MB = −1.75 mm). Different from previous studies that focused on clear-sky conditions only, this work is the first one to enhance the quality of official MODIS NIR PWV products under all weather conditions, reducing the impact of clouds on PWV products.