Mapping the grounding line of Antarctica in SAR interferograms with machine learning techniques

Abstract

The grounding line marks the transition between ice grounded at the bedrock and the floating ice shelf. Its location is required for estimating ice sheet mass balance, modelling of ice sheet dynamics and glaciers and for evaluating ice shelf stability, which merits its long-term monitoring. The line migrates both due to short term influences such as ocean tides and atmospheric pressure, and long-term effects such as changes of ice thickness, slope of bedrock and variations in sea level. Of the numerous in-situ and remote sensing methods currently in use to map the grounding line, Differential Interferometric Synthetic Aperture Radar (DInSAR) is, by far, the most accurate technique which produces spatially dense delineations. Tidal deformation at the ice sheet-ice shelf boundary is visible as a dense fringe belt in DInSAR interferograms and its landward limit is taken as a good approximation of the grounding line location (GLL). The GLL is usually manually digitized on the interferograms by human operators. This is both time consuming and introduces inconsistencies due to subjective interpretation especially in low coherence interferograms. On a large scale and with increasing data availability a key challenge is the automation of the delineation procedure. So far, a limited amount of studies were published regarding the delineation processes of typical features on the ice sheets using deep neural networks (DNNs). The objectives of this thesis were to further explore the feasibility of using machine learning for mapping the interferometric grounding line, as well as exploring the contributions of complementary features such as coherence estimated from phase, Digital Elevation Model, ice velocity, tidal displacement and atmospheric pressure, in addition to DInSAR interferograms. A dataset composed of manually delineated GLLs generated within ESA’s Antarctic Ice Sheet Climate Change Initiative project and corresponding DInSAR interferograms from ERS-1/2, Sentinel-1 and TerraSAR-X missions over Antarctica together with the above mentioned features was compiled and used for training two DNNs: Holistically-Nested Edge Detection (HED) andUNet. The developed processing chain handles creation of the training feature stack, training the DNNs and performing post processing functions on the resulting predictions. HED outperformed UNet and was able to achieve a median deviation (from manual delineations) of 209.23 m with a median absolute deviation of 152.91 m. Analysis of the individual feature contributions revealed that only the phase and derived features (real and imaginary interferogram components and coherence estimates) substantially influence the predicted GLLs. This finding is advantageous in terms of saving time, computational effort and memory in creating and storing the above mentioned feature stack. Although the delineations generated from HED do not perfectly follow the true GLL in all locations, the gains in efficiency and consistency are considerable, compared to the time and effort spent for manual digitizations. This study shows the potential of DNNs for automating the interferometric GLL delineation process.