Insights into the automated delineation of glacier and iceshelf calving fronts and multi-sensor radar datasets

Kurzfassung

Accurate calving-front locations of ice shelves and tidewater glaciers in Antarctica and Greenland are necessary observations for monitoring dynamic changes of ice-sheet margins. Floating ice shelves buttress the flow of the interior ice sheet, and their frontal locations are critical indicators of long-term changes in ice-sheet dynamics. Continuous monitoring of calving fronts has been historically challenging due to the scarcity and discontinuity of satellite data over the polar regions, and the time-consuming nature of manually delineating calving-front positions. Recent advances in edge detection and deep learning techniques have enabled automatic calving-front extraction from synthetic aperture radar (SAR) and optical satellites. We evaluate the performance of multi-class semantic segmentation techniques for 681 SAR scenes of glaciers located in Greenland, Antarctica, and Alaska using the CaFFe dataset of calving fronts. We aim to compare different model architectures such as vision transformers and foundation models to the baseline UNet model, with initial results showing promising performance metrics. The results of this study will be applied to an in-progress, novel benchmark dataset of Antarctic ice shelf satellite imagery that will provide the first, long-term record of calving-front locations for the continent and improve our understanding of historical and present-day frontal changes.