Antarctic ice shelf surface melt rates derived from observational data and machine learning

The Antarctic ice sheets contribution to sea level rise is dominated by the stability of ice shelves. Intense surface melt and the presence of surface water has been identified as a common precursor to ice shelf disintegration. Therefore, understanding the variability and causes of surface melt across ice shelves is critical for assessing future global sea level rise. Existing techniques used to estimate surface meltwater production include regional climate modeling, satellite remote sensing, and in situ meteorological observations, though each has inherent limitations including computational expense, difficulty in gauging melt intensity, and data scarcity. Here we develop a machine learning model to predict the intensity of Antarctic ice shelf surface melting using satellite data including albedo measurements from the Moderate Resolution Imaging Spectroradiometer (MODIS) and radar backscatter measurements from the Advanced Scatterometer (ASCAT) satellite combined with ERA-5 reanalysis products. Our random forest machine learning model is trained using observations from four automatic weather stations on Larsen C ice shelf on the Antarctic Peninsula from 2009 to 2018 that include the measurements of the full surface energy balance. We demonstrate that the ERA-5 reanalysis well represents most components of the ice shelf surface energy balance including the impacts of fohn winds known to drive surface melt across Larsen C. Likewise, we find that surface melt predicted using the random forest model agrees reasonably well with the seasonal trends in observed melt indicating the combined satellite and reanalysis data can be reliably used to predict ice shelf surface meltwater production. As a sensitivity test, we also examine other reanalysis products for their ability to reproduce observed melt rates. Finally, we examine the feasibility of these datasets and methods for generating melt estimates at the continental scale. Our research implies that machine learning applied to observational datasets can help resolve melt driven by small spatial scale processes that are not typically well simulated by regional climate models, and enable prediction of future surface melt at a low computational cost.