A machine-learning-based reconstruction of surface mass balance over the Greenland Ice Sheet from 1950 to 2020.

An accurate estimation of surface mass balance (SMB) is imperative for reliable quantification of the Greenland Ice Sheet (GrIS) mass changes and associated global sea level rise. Here, we present two new reconstructions of gridded annual SMB at the resolution of 0.1° × 0.1° across the GrIS from 1950 to 2020 by means of a CNN-Transformer model. This machine learning model is trained on the mostly recent compiled in-situ SMB observations, integrating with ERA5 Land reanalysis product, and outputs of the polar regional climate model MAR, respectively. When trained on the full dataset, the best-performing model predictions highly and significantly correlate with observations, with correlation coefficients of 0.96 and 0.93 for ERA5-Land-based and MAR-based reconstructions, respectively. Spatial cross-validation over unseen regions shows that reconstruction errors reduce by approximately 50% relative to the original datasets (ERA5-Land and MAR), with both reconstructions yielding statistically comparable RMSE (~53 mm w.e. yr-1) against in-situ observations. Independent validation against IceBridge airborne radar products further confirms their robustness, with 35% and 50% reduction in relative absolute error relative to ERA5-Land and MAR, respectively. The spatially and temporally complete annual SMB datasets can be used for the input of ice-sheet models and surface hydrological studies.