A downscaling-fusion framework based on machine learning to improve daily precipitation estimates in the dry-hot valley of southwest China

The Jinsha River dry-hot valley Accurately obtaining high-precision spatiotemporal precipitation information is crucial for hydrological modeling and soil erosion research in the dry-hot valley of southwest China. However, existing satellite precipitation products suffer from low spatial resolution and significant errors, which are not suitable for small-scale research. To address these gaps, this study constructed a downscaling-fusion framework based on random forest (RF) to improve the estimation accuracy of daily gridded precipitation in the dry-hot valley region. In the downscaling stage, the framework selected key influencing factors; in the fusion stage, it adopted a two-stage strategy of classification and regression, incorporating auxiliary variables and the spatial autocorrelation characteristics of satellite precipitation. Using the Jinsha River dry-hot valley as a case study, the results showed that: (1) The RF downscaling model more precisely characterized spatial distribution patterns of precipitation. (2) The two-stage fusion strategy outperformed single regression models in both precipitation event identification and quantitative estimation; however, its accuracy decreased with increasing elevation, and the underestimation of heavy precipitation was particularly pronounced in high-altitude areas. (3) Over the past two decades, no significant trend was observed in the total annual precipitation in the Jinsha River dry-hot valley, whereas the dry season became drier and precipitation concentration increased during the rainy season, leading to a rising risk of local heavy precipitation events. Overall, this study provides an effective scheme for constructing high-quality precipitation data in dry-hot valley regions. • An RF downscaling-fusion framework is developed to optimize precipitation estimates for the Jinsha River dry-hot valley. • The proposed framework demonstrates prominent capability in accurately capturing diverse precipitation conditions. • Vertical zonation characteristics of precipitation distribution were explicitly evaluated. • In the dry-hot valley, the dry season is becoming drier, and the risk of local heavy rainfall increases during the rainy season.