Refining daily precipitation estimates using machine learning and multi-source data in alpine regions with unevenly distributed gauges

The Qilian Mountains region, located in the northeast edge of the Tibetan plateau. Reliable high-spatiotemporal-resolution and long-term precipitation data are critical for agriculture, hydrology, and climate change impact analysis. However, in the cold and arid Qilian Mountains, the uneven distribution of rain gauges and the high spatial heterogeneity of topography pose great challenges to obtaining such data. To over these limitations, a downscaling-merging framework based on XGBoost (XDMF) is proposed to generate high accuracy precipitation dataset with 1 km by combining gauges, satellite, and reanalysis precipitation products. XDMF includes three critical steps: precipitation downscaling, identification, and estimation, focusing on simultaneously improving the spatial resolution, precipitation detection capability and estimation capability. This framework is applied in the Qilian Mountains and generated two datasets: QL-DMP 2 P (1981–2020) and QL-DMP 4 P (2001–2020). The results demonstrate that QL-DMP significantly outperforms original products at different temporal and spatial scales. Compared to methods that use XGBoost only in two steps (downscaling and estimation, or identification and estimation), XDMF can better reproduce the precipitation variability at the small-scale and reduce precipitation detection errors. This study offers high-quality and long-term alternative data for hydrometeorology research. Meanwhile, XDMF is a promising algorithm for enhancing precipitation in high-altitude mountain areas, which can be flexibly transferred to other regions, different machine learning algorithms, and various hydrometeorological variables. • A precipitation downscaling-merging framework based on XGBoost (XDMF) is proposed and applied in the Qilian Mountains. • XDMF effectively improves the spatial resolution, the identification and estimation abilities of precipitation products. • The performance of XDMF is superior to other traditional methods. • The final 1km dataset outperforms original precipitation products in different temporal and spatial dimensions.