Deep Neural Network Based Automatic Grounding Line Delineation In Dinsar Interferograms

Accurate identification of grounding lines is of immense importance for estimating the mass budgets of ocean-terminating ice sheets and glaciers of Antarctica and Greenland. In Differential Interferometric SAR (DInSAR) interferograms, human experts still largely manually digitize grounding lines. The time-consuming nature of this task makes it infeasible to produce timely, continent-wide grounding line mappings. This study employed a Deep Neural Network (DNN) to automate delineation. The Holistically-Nested Edge Detection (HED) network was trained in a supervised manner on features derived from interferometric phase, elevation data, ice velocity, tidal amplitude, atmospheric pressure and corresponding manual delineations. HED-generated lines achieved a median deviation of 209 m with a median absolute deviation of 153 m from manual delineations. The developed automatic pipeline demonstrates the potential for generating spatially and temporally dense mappings of the grounding line.