The global depletion of groundwater poses a significant challenge to water security. However, the coarse spatial resolution of GRACE satellite observations obscures fine-scale hydrological dynamics and limits the separation of localized anthropogenic extraction signals from large-scale climatic forcing. To address this limitation, we present a high-resolution (1 km), continuous monthly dataset of global groundwater storage anomalies (GWSA) covering the period from 2002 to 2020. The dataset is generated using a production framework that integrates Singular Spectrum Analysis (SSA) for temporal gap filling and an aquifer-stratified machine learning approach driven by 19 high-resolution hydroclimatic and geophysical predictors. To ensure robustness and spatial consistency, multiple predictive approaches were evaluated as part of a quality control procedure, and the most stable model was selected for final production based on multi-scale validation. Cross-scale evaluation shows that the downscaled dataset preserves the large-scale spatial patterns of the original GRACE observations with high agreement (<em>R</em>² = 0.972, RMSE = 2.10 cm). Independent validation using 1,518 in situ monitoring wells, combined with a geographically stratified specific yield matrix for dimensional conversion, further demonstrates the ability of the dataset to capture long-term groundwater variability across diverse hydrogeological conditions (<em>R</em>² = 0.44, <em>p</em> &lt; 0.01). The resulting 1 km dataset provides enhanced spatial detail and enables the identification of sharp nonlinear boundaries associated with intensive human pumping, as well as spatial polarization patterns in groundwater storage changes. This dataset offers a reliable, observation-constrained resource for water resource assessment, hydrological modeling, and studies of coupled climate and human influences on groundwater systems.