Applying machine learning to Earth observation of Antarctic ice shelves and icebergs

Abstract

There is a critical need for enhanced monitoring of Antarctic ice shelf dynamics and iceberg tracking to better understand the impacts of climate change. This thesis focuses on methods to map two key elements of the dynamic Antarctic coastal environment: calving fronts and icebergs. Traditional remote sensing approaches often fall short in dealing with the nuances of multispectral and SAR data in addressing these elements, necessitating the development of advanced methodologies. The research tackles two key questions: How can machine learning and advanced image processing improve Antarctic coastal monitoring from Earth observation data? and what are the primary drivers behind Antarctic calving front and iceberg changes? The study applies novel machine learning techniques and cloud-based data processing to analyse these dynamics, with a focus on the Amundsen Sea Sector and Getz Ice Shelf.

The methodology integrates advanced machine learning models, such as optimised UNet architectures and the Vision-Transformer-based Segment Anything Model (SAM), for high-precision satellite image segmentation. It incorporates innovative experiments such as the use of boundary-distance loss functions to refine calving front segmentation accuracy and video object segmentation to track iceberg movements in the Amundsen Sea. These approaches are applied to high-resolution multispectral Landsat data and Sentinel-1 Synthetic Aperture Radar imagery.

Key results indicate that Circumpolar Deep Water influx has had a significant influence over the calving front changes at Getz Ice Shelf, although precise mechanisms remain unclear, while bedrock morphology plays a crucial role in iceberg and coastal icescape stability. These insights highlight the importance of local cryospheric, oceanic and atmospheric processes, and large-scale environmental interactions in shaping Antarctic cryosphere dynamics.

The thesis concludes by emphasising the need for sustainable, integrated Earth observation monitoring systems of the Antarctic Ice Sheet to inform global sea-level rise predictions. It demonstrates the potential of deploying scalable machine learning methodologies for continuous monitoring, advocating for high-quality datasets and interdisciplinary research to advance our understanding of the Antarctic’s responses to climate change.