Abstract: Abstract：ObjectiveThe rainfall erosivity factor (R) is a crucial input parameter for soil erosion models, and the accuracy of its values and spatial distribution is essential for regional and watershed soil erosion prediction and prevention. In high mountain canyon regions with sparse and unevenly distributed stations and complex terrain, traditional spatial interpolation methods struggle to accurately reflect the spatial variability of precipitation. While statistical downscaling that integrates multi-source environmental variables theoretically offers advantages, systematic evaluations of its accuracy differences and variation patterns compared to ground-based interpolation results are still lacking. MethodsTo address this, this study focuses on the high mountain canyon area of southwestern China. Based on GPM remote sensing precipitation data and the Multiscale Geographically Weighted Regression (MGWR) downscaling method, we obtained spatial distribution data of the multi-year average R factor at 250m resolution for the entire region from 2015 to 2020. Under different validation station distribution scenarios (10%, 20%, 30%), we used two common interpolation methods, Inverse Distance Weighting (IDW) and Ordinary Kriging (OK), to derive R factor data with the same temporal sequence and resolution. We then systematically evaluated the accuracy differences among these methods. The results indicate: Results(1) The multi-year average R factor values for the entire region obtained by IDW interpolation, OK interpolation, and MGWR downscaling were 1781, 1773, and 2484 MJ·mm/(hm²·h·a), respectively. These differences were primarily caused by the variation in values in areas without ground stations, mainly in the Himalayan region. The two ground-based interpolation results (IDW and OK) showed relatively consistent values and distributions across the entire region but differed significantly from the downscaling (MGWR) results in areas with few or no stations. (2) Under different proportions of validation station scenarios, the downscaling (MGWR) results showed a reduction in RMSE by 14%–17% and 13%–20% compared to IDW and OK interpolation, respectively, at the regional scale; MAE decreased by 18%–24% and 16%–25%, respectively; and R² significantly increased. There was no significant difference in accuracy between the two ground-based interpolation methods. (3) MGWR downscaling can significantly improve the accuracy of spatial R factor data compared to the two ground-based interpolation methods (IDW and OK), with particularly noticeable improvement in areas with sparse ground stations.Conclusion MGWR downscaling can significantly improve the spatial accuracy of the R factor, with particularly evident advantages in regions with sparse ground observations. The findings can provide a basis for selecting optimal methods for spatial mapping of rainfall erosivity in regions with complex terrain or insufficient data.