All-weather precipitable water vapor map reconstruction using data fusion and machine learning-based spatial downscaling

Abstract

Precipitable water vapor (PWV) detection with high spatial resolution and high accuracy is of significant importance for contributing to extreme weather events monitoring and forecasting. Current PWV products, however, suffer from limitations of spatial and temporal discontinuities, low accuracy, and coarse spatial resolution. To overcome this problem, a data fusion and machine learning-based spatial downscaling solution is proposed. At first, spatially complete PWV maps are generated by integrating calibrated PWV of Moderate Resolution Imaging Spectroradiometer (MODIS) and the ERA5 PWV data from 2018 to 2022. Subsequently, three spatial downscaling models based on Gradient Boosting Decision Tree (GBDT), Multi-layer Perceptron Neural Network (MLPNN), and Random Forest (RF), respectively, are developed to produce high-quality, all-weather PWV considering the land-cover type. It has been verified that the high-quality all-weather PWV maps generated by the GBDT, MLPNN, and RF models exhibit strong agreement with Global Navigation Satellite System (GNSS) PWV estimates. The correlation coefficients are 0.95, 0.87, and 0.87, while the overall Bias is 0.29 mm, 0.67 mm, and 0.35 mm, and the root mean square errors (RMSE) are 1.74 mm, 2.98 mm, and 3.06 mm, respectively. These results significantly enhance the accuracy of MODIS PWV products (R2 = 0.73, RMSE = 5.64 mm, Bias = 3.05 mm). Notably, the GBDT model outperforms the other models in terms of performance. Compared to MODIS PWV, the new PWV map with a data fusion and machine learning-based spatial downscaling approach effectively utilizes the advantage of satellite-based and reanalyzed PWV products, providing continuous, detailed, and reasonable variation in time and space. Moreover, it is less influenced by seasonal changes. The new PWV map has a promising application for regional hydrology and meteorology.