Abstract. To resolve the bed elevation of Antarctica, we present DeepBedMap – a novel machine learning method that can produce Antarctic bed topography with adequate surface roughness from multiple remote sensing data inputs. The super-resolution deep convolutional neural network model is trained on scattered regions in Antarctica where high-resolution (250 m) ground-truth bed elevation grids are available. This model is then used to generate high-resolution bed topography in less surveyed areas. DeepBedMap improves on previous interpolation methods by not restricting itself to a low-spatial-resolution (1000 m) BEDMAP2 raster image as its prior image. It takes in additional high-spatial-resolution datasets, such as ice surface elevation, velocity and snow accumulation, to better inform the bed topography even in the absence of ice thickness data from direct ice-penetrating-radar surveys. The DeepBedMap model is based on an adapted architecture of the Enhanced Super-Resolution Generative Adversarial Network, chosen to minimize per-pixel elevation errors while producing realistic topography. The final product is a four-times-upsampled (250 m) bed elevation model of Antarctica that can be used by glaciologists interested in the subglacial terrain and by ice sheet modellers wanting to run catchment- or continent-scale ice sheet model simulations. We show that DeepBedMap offers a rougher topographic profile compared to the standard bicubically interpolated BEDMAP2 and BedMachine Antarctica and envision it being used where a high-resolution bed elevation model is required.

To resolve the bed elevation of Antarctica, we present DeepBedMap -a novel machine learning method that can produce Antarctic bed topography with adequate surface roughness from multiple remote sensing data inputs. The super-resolution deep convolutional neural network model is trained on scattered regions in Antarctica where high-resolution (250 m) ground-truth bed elevation grids are available. This model is then used to generate high-resolution bed topography in less surveyed areas. DeepBedMap improves on previous interpolation methods by not restricting itself to a low-spatial-resolution (1000 m) BEDMAP2 raster image as its prior image. It takes in additional high-spatialresolution datasets, such as ice surface elevation, velocity and snow accumulation, to better inform the bed topography even in the absence of ice thickness data from direct icepenetrating-radar surveys. The DeepBedMap model is based on an adapted architecture of the Enhanced Super-Resolution Generative Adversarial Network, chosen to minimize perpixel elevation errors while producing realistic topography. The final product is a four-times-upsampled (250 m) bed elevation model of Antarctica that can be used by glaciologists interested in the subglacial terrain and by ice sheet modellers wanting to run catchment-or continent-scale ice sheet model simulations. We show that DeepBedMap offers a rougher topographic profile compared to the standard bicubically interpolated BEDMAP2 and BedMachine Antarctica and envision it being used where a high-resolution bed elevation model is required.

Accelerating ice loss from Antarctica's ice sheets is projected to contribute 13-42 cm of sea level rise by the end of the century (Edwards et al., 2021). This contribution is mainly driven by an increase in ice flowing off the continent and into the ocean. Most Antarctic ice discharge becomes part of a floating ice shelf (Rignot et al., 2013), half of which will melt before it reaches the open ocean, while the other half will eventually calve as icebergs (Liu et al., 2015;Rignot et al., 2013). Ice shelves slow the discharge of glacial ice into the ocean. When ice shelves drag past coastlines, islands, and pinning points, they generate back stresses and buttress against ice flow (e.g., Dupont & Alley, 2005;Fürst et al., 2016). This buttressing is an important control on the rate of ice loss from Antarctica. The removal of buttressing when an ice shelf retreats or disintegrates leads to the acceleration of ice loss (e.g., Berthier et al., 2012;Rignot et al., 2004;Scambos et al., 2004).Melt at the base of ice shelves is largely controlled by ocean circulation in the sub-ice-shelf cavity. Theoretical frameworks of ocean circulation under ice shelves were first developed from the interpretation of direct oceanographic observations at ice shelf fronts (e.g., S. S. Jacobs et al., 1979). At a large scale, currents under an ice shelf follow a circulation-cell, described in detail by S. Jacobs et al. (1992) and well approximated by the ice pump mechanism (Lewis & Perkin, 1986). Sea ice formation releases high salinity water which sinks and flows down-

Abstract. To better resolve the bed elevation of Antarctica, we present DeepBedMap – a novel machine learning method that produces realistic Antarctic bed topography from multiple remote sensing data inputs. Our super-resolution deep convolutional neural network model is trained on scattered regions in Antarctica where high resolution (250 m) groundtruth bed elevation grids are available. The model is then used to generate high resolution bed topography in less well surveyed areas. DeepBedMap improves on previous interpolation methods by not restricting itself to a low spatial resolution (1000 m) BEDMAP2 raster image as its prior. It takes in additional high spatial resolution datasets, such as ice surface elevation, velocity and snow accumulation to better inform the bed topography even in the absence of ice-thickness data from direct ice-penetrating radar surveys. Our DeepBedMap model is based on an adapted Enhanced Super Resolution Generative Adversarial Network architecture, chosen to minimize per-pixel elevation errors while producing realistic topography. The final product is a four times upsampled (250 m) bed elevation model of Antarctica that can be used by glaciologists interested in the subglacial terrain, and by ice sheet modellers wanting to run catchment or continent-scale ice sheet model simulations. We show that DeepBedMap offers a more realistic topographic roughness profile compared to a standard bicubic interpolated BEDMAP2 and BedMachine Antarctica, and envision it to be used where a high resolution bed elevation model is required.

This paper introduces a new method, based on Machine Learning, namely a Generative Adversarial Network (GAN), to add short-scale roughness to the bed of Bedmap2. The paper is well written, easy to follow and well illustrated, I really enjoyed reading it. I recommend publication after minor revisions. My main problem while reading the manuscript was that I felt like the authors were overselling their approach and the performance of the GAN.