Terrestrial water storage is a comprehensive manifestation of land water. Analyzing the spatio-temporal changes of terrestrial water storage is vital for improving the understanding of hydrological processes and water resource management. However, the low spatial resolution of existing terrestrial water storage anomalies derived from GRACE limits their applications in small and medium basins. To improve their spatial resolution, the random forest models were utilized to downscale terrestrial water storage anomalies data derived from GRACE satellites and its follow-up mission GRACE-follow on into 0.25°×0.25° spatial resolution at three spatial scales, including grid cell, regional (basin) and regional (China). Groundwater storage anomalies were calculated by combining the vertical water budget and GLDAS model output. The performance of the downscaling data was evaluated based on models' indicators and in-situ groundwater levels across China. Results show that random forest model can accurately establish the statistical relationship between input variables (precipitation, temperature, vegetation condition index, and soil water storage) and GRACE data. The average correlation coefficient of the grid cell downscaling method during the validation period is generally around 0.6. The average Nash efficiency coefficient, correlation coefficient and root mean square error of the regional downscaling method are greater than 0.5, 0.75 and less than 6.6 cm, respectively. Overall, the accuracy of different downscaled data was promising. From 2003 to 2021, the deficit of China's terrestrial water storage from original, grid cell downscaling-based, regional downscaling-based (basin) and regional downscaling-based (China) GRACE data were about 119.5×10⁸, 62.4×10⁸, 121.1×10⁸ and 121.8×10⁸ m³/a, respectively. The storage of groundwater storage was approximately 230.0×10⁸, 171.8×10⁸, 235.6×10⁸ and 236.4×10⁸ m³/a, respectively. The simulation accuracy of the grid cell downscaling results is relatively poor due to the small sample size. Change rates of water storage obtained by the regional downscaling methods were generally consistent with the original GRACE data, indicating that the regional downscaling methods were better than the grid-cell downscaling method. Compared with the grid-cell downscaling method, results obtained by the regional downscaling method were smoother in space, and the regional downscaling (basin) refined to the basin could improve the accuracy of groundwater storage anomalies. Regional downscaling has the advantages of strong applicability, high simulation accuracy and high computational efficiency than grid cell downscaling. Findings of this study can provide refined water storage anomalies data for sustainable utilization and water resources planning at basin-scale.