Machine-learning downscaling of GPM satellite precipitation products in mountainous regions: A case study in Chongqing

Abstract

High-quality precipitation data are essential for research in hydrology, meteorology and ecology. Nevertheless, in mountainous regions with intricate terrain, the reliability of gridded precipitation data derived from station data interpolation is low due to the limited number of stations caused by the difficulty of station setup. Current satellite precipitation products suffer from low spatial resolution, making them unsuitable for hydrological and meteorological research at small and medium scales. Their application in mountainous regions with significant spatiotemporal heterogeneity is even more challenging. To this end, downscaling satellite precipitation products has become an effective method for obtaining accurate spatial distribution information of precipitation in these regions. This study employs a method of first calibration followed by downscaling analysis of GPM daily precipitation product in the Chongqing area using random forest (RF), extreme gradient boosting (XGBoost), and long short-term memory (LSTM) algorithms. Ultimately, the spatial resolution of GPM product is improved from 0.1° to 0.01° (∼1 km). The findings demonstrate that: (1) the station-calibrated GPM precipitation product performed better than the original GPM product, and it is closer to the station measurements; (2) in practical applications, the LSTM downscaling algorithm can effectively enhance spatial resolution without compromising accuracy, whereas RF and XGBoost incur considerable accuracy loss when enhancing spatial resolution; (3) the downscaled results from all three algorithms were consistent with the calibrated GPM precipitation maps and significantly improved the spatial details of precipitation. Among them, the results of the LSTM method exhibited greater continuity in the spatial distribution of precipitation, aligning more closely with the characteristics of precipitation distribution. In summary, the LSTM algorithm demonstrates greater potential for the downscaling of GPM precipitation product in the study area. This research provides a promising high-quality precipitation data generation scheme for mountainous regions with sparse station coverage and complex terrain and landforms.