A Downscaling–Calibration Framework for GPM Precipitation Data Based on Geographically Weighted Machine Learning: A Case Study of the Yellow River Basin, China

Satellite precipitation products alleviate sparse gauge coverage, but their coarse resolution limits hydrometeorological applications. This study develops and evaluates an integrated downscaling–calibration framework for the Yellow River Basin, China, using Geographical-XGBoost (G-XGBoost) for downscaling and Geographical Difference Analysis (GDA) for calibration. Geographically weighted regression, Random Forest, and XGBoost were used for comparison. Fraction disaggregation was compared with an optimal-lag Normalized Difference Vegetation Index (NDVI) method for monthly disaggregation. Using Global Precipitation Measurement (GPM) data and four explanatory variables, including NDVI, Land Surface Temperature (LST), Digital Elevation Model (DEM), and slope, the spatial resolution was improved from 0.1° to 1 km. The downscaled products covered both annual and monthly temporal resolutions. Results show that: (1) G-XGBoost achieved relatively strong overall performance among the four downscaling models before GDA correction across the dry year 2015, wet year 2018, and near-average year 2019, while after GDA correction, all products improved and G-XGBoost remained among the best-performing methods across the three years; and (2) at the monthly scale, fraction disaggregation showed modest numerical improvements in pooled accuracy across the three-year comparison, although its relative performance varied among years and metrics. The framework provides a practical approach for generating high-spatial-resolution precipitation estimates in the Yellow River Basin.