High-resolution glacier elevation observations are essential for resolving rapid changes in glacier geometry, but existing satellite datasets face a trade-off between spatial and temporal resolution. Here, we present U-DEM, a machine-learned glacier elevation time series that combines complementary satellite datasets to provide surface elevations at 100 m spatial and 3-month temporal resolution. We apply the approach to glaciers and ice caps of the Canadian Arctic from 2014 to 2024. High-resolution glacier elevation patterns are learned from ArcticDEM strips and reconstructed through time using CryoSat-2 elevations and Sentinel-1 backscatter as temporally repeated inputs.</p> <p>Validation of U-DEM against independent ICESat-2 laser altimetry observations yields a root-mean-square error (RMSE) of 10.01<span> </span>m, approximately 65<span> </span>% lower than the underlying CryoSat-2 elevations, with improvements across all 15 study regions, including regions not represented during training. Comparing U-DEM with kinematic Global Navigation Satellite System (GNSS) measurements on the narrow White Glacier yields RMSE values of 1.70–3.35<span> </span>m, comparable to those of the ArcticDEM strips. Application examples further demonstrate the potential of U-DEM for resolving glacier elevation changes across different spatial scales. Firstly, U-DEM resolves the migration of the dynamic balance line during the ongoing surge of Chapman Glacier. Secondly, U-DEM captures localized elevation variability associated with several active subglacial lakes that is poorly resolved by CryoSat-2.</p> <p>U-DEM thus demonstrates how machine learning can integrate satellite observations with complementary sampling characteristics to bridge the spatial-temporal resolution gap of individual sensors. The approach provides a framework for generating spatially detailed and temporally continuous records of glacier surface evolution from existing and future Earth observation missions.