Groundwater is a critical resource for sustainable development, particularly in arid regions facing water scarcity. The Gravity Recovery and Climate Experiment (GRACE) and its Follow-On, GRACE-FO, offer valuable data on groundwater storage anomalies (GWSA). However, while their coarse resolution has been improved using machine learning approaches such as the global random forest (RFG) model, the aspatial nature of the RFG model limits its ability to capture spatial heterogeneity when downscaling GRACE (-FO) data. Downscaling GWSA data to higher resolutions is crucial for assessing small-scale groundwater variations. To address this, a novel spatially weighted random forest (RFSW) model has been proposed to downscale GWSA to a high resolution (0.1 degrees) across the North China Plain (NCP) from 2003 to 2023. We found that the RFSW model outperforms the RFG model, reducing RMSE by 44.44% and residuals by 43.57%. The downscaled GWSA data strongly correlate with in-situ measurements from 559 monitoring wells (correlation coefficients: 0.52-0.85), revealing significant groundwater depletion in the Piedmont Plain (PP) and East-Central Plain (ECP) sub-regions, with the most severe losses in Shijiazhuang (-17.08), Xingtai (-16.67), and Handan ( -16.02 mm/yr), respectively. The winter wheat area doubling from 2.5 million to 5.8 million hectares, reducing GWSA from-180 mm to-480 mm. This improved downscaling technique enhances our understanding of local groundwater dynamics and their relationship to agricultural practices. This method's high-resolution GWSA data can inform more targeted and effective water management strategies in water-stressed regions worldwide.