Automatic grounding line delineation of DInSAR interferograms using deep learning

Abstract

The regular and robust mapping of grounding lines is essential for various applications related to the mass balance of marine ice sheets and glaciers in Antarctica and Greenland. Differential Interferometric Synthetic Aperture Radar (DInSAR) enables precise detection of tide-induced ice shelf flexure at a continent-wide scale with temporal resolutions of just a few days. While automated pipelines for generating differential interferograms are well established, grounding line delineation remains largely a manual process, which is labor-intensive and increasingly impractical given the growing data streams from current and upcoming synthetic aperture radar (SAR) missions. To address this limitation, we developed an automated pipeline employing the holistically nested edge detection (HED) neural network to delineate grounding lines from DInSAR interferograms. The network was trained in a supervised manner using 421 manually annotated grounding lines of outlet glaciers and ice shelves of the Antarctic Ice Sheet. We also evaluated the utility of non-interferometric features such as surface elevation, ice velocity, and differential tide levels for enhancing delineation performance. Our recommended neural network, trained on the real and imaginary interferometric features, achieved a median offset of 265 m and a mean offset of 421 m from manual grounding line delineations, as well as a predictive uncertainty of 401 m. Furthermore, we demonstrated this network's capacity to generalize by generating grounding lines for previously undelineated interferograms, highlighting its potential for large-scale, high-resolution spatiotemporal mappings.