Mapping Ice Shelf Calving Fronts at Thwaites Glacier using Deep Learning and Satellite Imagery in a Cloud-Based Workflow

Abstract

Thwaites Glacier (TG) is one of the primary sources of ice mass loss from the West Antarctic Ice Sheet, making it a critical site for monitoring changes in the calving front location. The long duration of the Landsat mission provides a valuable opportunity to analyze over 50 years of historical imagery and produce near-real-time calving front monitoring solutions for the future. Here, we have developed a tool that allows users to produce calving front maps from cloud-hosted Landsat imagery using a U-Net, a deep learning architecture commonly used for semantic segmentation. The tool utilizes open-source Python packages for rapid querying of the Landsat catalog stored in a the Spatio-Temporal Asset Catalog (STAC) standardized metadata format, and for scalable and distributed cloud processing. This cloud-based workflow will provide researchers with access to pre-trained calving front segmentation models and decades of Landsat imagery from Thwaites Glacier. This workflow may be expanded in the future to provide historical analysis and near-real-time monitoring of other important ice shelves and glaciers in Antarctica.