Deep learning based automatic grounding line delineation in DInSAR interferograms

Kurzfassung

The grounding line is a subsurface geophysical feature that divides the grounded ice sheet and floating ice shelf. Knowledge of its precise location is required for estimating ice sheet mass balance, as ice discharged from the interior is typically calculated at the grounding line [1], [2]. While grounding lines in Greenland have only a minimal extension, in Antarctica, they span about 75% of its coastline. The bending of ice shelves due to ocean tides causes them to migrate several kilometers over a tidal cycle within a transition region called the grounding zone. This short-term displacement adds to the difficulty in grounding line detection on a featureless ice surface. Nevertheless, various remote sensing methods can currently detect grounding lines on a continental scale. In particular, Differential Interferometric Synthetic Aperture Radar (DInSAR) is used to measure the deformation which occurs at the grounding line due to tidal flexure of ice shelves with sub-centimeter accuracy [3]. If coherence is preserved between the SAR repeat passes, the vertical ice deformation at the grounding zone is visible in the double difference interferogram as a dense fringe belt. The landward-most fringe is considered a good approximation of the actual grounding line. Although the generation of DInSAR interferograms is already automatized, the identification of the landward-most fringe and its digitization is still majorly performed manually by human operators. Besides being labour and time-intensive, manual delineations are inconsistent due to varying interpretations of experts in identifying the landward fringe, especially in areas with poor coherence or intricate fringe patterns. In the present study, we attempt to automate the delineation by employing a Convolutional Neural Network (CNN). We developed an automatic workflow that handles the preparation of the training feature stack, training and inference of the neural network and the post-processing of network-generated delineations. The CNN architecture is based on the Holistically-Nested Edge Detection network [4]. It was trained on 478 georeferenced DInSAR interferograms from ERS-1/2, Sentinel-1 A/B and TerraSAR-X repeat pass acquisitions and their corresponding hand-delineated grounding lines that were generated within the Grounding Line Location (GLL) product of ESA’s Climate Change Initiative (AIS cci) project [5]. The training feature stack consists of four interferogram-based features: real and imaginary components, interferometric phase and pseudo coherence (estimated by applying a boxcar filter to interferometric phase) derived from the corresponding DInSAR interferograms. A median deviation of 209 m between the network-delineated and corresponding manual GLLs was measured for the test set. The trained network delineates an interferogram in milliseconds, considerably shorter than the time required for manual delineation. We propose to automatically and efficiently expand the AIS cci GLL product by applying our trained neural network to interferograms that still need to be manually delineated. In particular, we plan to generate DInSAR interferograms from highly coherent TerraSAR-X data triplets acquired in 2021 using the Integrated Wide Area Processor (IWAP) [6]. These acquisitions were made over Southern Byrd, Amundsen, Lennox-King and Dickey glaciers feeding into the Ross Ice Shelf and Recovery Glacier situated in the Ronne-Filchner Ice Shelf at high latitudes, which Sentinel-1 cannot image. Consequently, no updated grounding lines for these glaciers exist in current DInSAR-based grounding line datasets [7]. In general, the performance of our trained neural network is not dependent on the SAR sensor but on the quality of the interferograms. The automatic delineation can create monthly or half-yearly average GLL time series from all suitable DInSAR interferograms in a certain period. This derived product has a downstream application in analyzing short and long-term migratory patterns of grounding lines. References [1] C. Schoof, “Ice sheet grounding line dynamics: Steady states, stability, and hysteresis,” Journal of Geophysical Research: Earth Surface, vol. 112, no. F3, 2007. [2] E. Rignot and H. Thomas, “Mass balance of polar ice sheets,” Science, vol. 297, no. 5586, pp. 1502–1506, 2002. DOI: 10.1126/science.1073888. eprint: https://www.science.org/doi/pdf/10.1126/science.1073888. [Online]. Available: https://www.science.org/doi/abs/10.1126/science.1073888. [3] E. Rignot, “Tidal motion, ice velocity and melt rate of petermann gletscher, greenland, measured from radar interferometry,” Journal of Glaciology, vol. 42, no. 142, pp. 476–485, 1996. [4] S. Xie and Z. Tu, “Holistically-nested edge detection,” in Proceedings of the IEEE international conference on computer vision, 2015, pp. 1395–1403. [5] A. Groh, Product user guide (pug) for the antarctic ice sheet cci project of esa’s climate change initiative, version 1.0, 2021. [Online]. Available: https://climate.esa.int/media/documents/ST- UL- ESA- AISCCI- PUG- 0001.pdf. [6] F. R. Gonzalez, N. Adam, A. Parizzi, and R. Brcic, “The integrated wide area processor (iwap): A processor for wide area persistent scatterer interferometry,” in ESA Living Planet Symposium, vol. 722, 2013, p. 353. [7] E. Rignot, J. Mouginot, and B. Scheuchl, “Measures antarctic grounding line from differential satellite radar interferometry, version 2,” NASA, 2016. [Online]. Available: https://doi.org/10.5067/IKBWW4RYHF1Q.