A quantile regression forest estimate of Greenland’s subglacial topography

Abstract Accurate knowledge of basal topography is required for numerical modelling efforts to predict how Earth’s ice sheets will respond to continued warming. The widely used BedMachine v3 dataset has limitations with respect to its use in modelling studies, particularly in estimating uncertainties. Machine learning approaches offer promise in addressing this gap, with quantile regression forests (QRFs) especially suited to geospatial data. Here, we apply a novel QRF approach to map the basal topography of Greenland’s ice sheet using airborne radio echo sounding (RES) data. Compared to BedMachine, our model reduces the root-mean-squared-error of ice depth predictions by 18%, from 232 to 190 m. It also significantly improves uncertainty calibration: 89.8% of new observations fall within our 90% prediction interval, versus 68% for BedMachine. The QRF model achieves a lower continuous ranked probability score (92 m vs. 130 m), indicating improved balance between accuracy and uncertainty. Our volume estimate for the Greenland ice sheet is 0.7% higher than BedMachine’s, though we emphasise differences in the predicted shape of subglacial features like outlet glacier troughs. This approach offers a computationally efficient, accessible method for deriving subglacial topography from RES data, while providing better-calibrated uncertainty estimates than existing models.