Dem Generator from Single Swath Radargrams

The ice sheet dynamics in Antarctica that directly impact the polar ice mass balance and the glacier erosion caused to the bedform are predicted by models that rely on several hard-to-estimate variables, including the bed topography itself. Antarctica’s bed topography is hard to estimate because it is covered by several layers of ice that could be up to several kilometers thick. Sparse, higher-resolution along-track measurements of its bed topography collected using Ice-Penetrating Radar (IPR) data are interpolated to create coarser-resolution gridded bed topography models. However, the significant gaps between IPR profiles mean there is significant scope for improving the measurements and interpolation approaches to fill those gaps. Here, we propose a deep learning (DL) generative adversarial network (GAN) approach to generate a realistic model of the bed topography from single-channel IPR acquisitions. The model takes advantage of the clutter caused by the IPR antenna to predict a Digital Elevation Model (DEM) accordingly to the information in the IPR acquisitions. The method is tested with synthetic data from regions with a high-resolution DEM available.