Deep learning based automatic grounding line delineation in DInSAR interferograms

• Main Authors: Ramanath Tarekere, Sindhu, Krieger, Lukas, Floricioiu, Dana, Diaconu, Codrut-Andrei, Heidler, Konrad
• Format: Article in Journal/Newspaper
• Language: English
• Published: Copernicus Publications 2025
• Subjects: SAR-Signalverarbeitung; EO Data Science; Antarctic; Greenland; The Antarctic; Antarc*; Antarctica; Ice Sheet; Ice Shelf; Ice Shelves; The Cryosphere
• Online Access: https://elib.dlr.de/208792/; https://egusphere.copernicus.org/preprints/2024/egusphere-2024-223/egusphere-2024-223.pdf

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. With Differential Interferometric Synthetic Aperture Radar (DInSAR) interferograms, it is possible to accurately capture the tide-induced bending of the ice shelf at a continent-wide scale and a temporal resolution of a few days. While current processing chains typically automatically generate differential interferograms, grounding lines are still primarily identified and delineated on the interferograms by a human operator. This method is time-consuming and inefficient, considering the volume of data from current and future SAR missions. We developed a pipeline that utilizes the Holistically-Nested Edge Detection (HED) neural network to delineate DInSAR interferograms automatically. We trained HED in a supervised manner using 421 manually annotated grounding lines for outlet glaciers and ice shelves on the Antarctic Ice Sheet. We also assessed the contribution of non-interferometric features like elevation, ice velocity and differential tide levels towards the delineation task. Our best-performing network generated grounding lines with a median distance of 222.2 m and mean distance of 340.5 m $\pm$ 373.88 m from the manual delineations. Additionally, we applied the network to generate grounding lines for undelineated interferograms, demonstrating the network's generalization capabilities and potential to generate high-resolution temporal and spatial mappings.