Deep-learning-based mapping reveals multi-decadal crevasse evolution on Antarctic ice shelves from Landsat imagery

Antarctic ice shelves buttress grounded ice and play a critical role in regulating glacier stability and the mass balance of the Antarctic Ice Sheet. Over recent decades, many Antarctic ice shelves have exhibited significant mass loss. Crevasses, which formed in response to internal stresses, are key indicators of ice shelf stability and their development is intricately coupled to the evolving ice dynamics. However, consistent long-term and high-resolution records of crevasse evolution over Antarctic ice shelves remain limited.

Here, we present a continuous and automated mapping of crevasses on representative Antarctic ice shelves from 1999 to 2024 using Landsat 7 and Landsat 8 imagery and a deep-learning-based segmentation framework at 30 m resolution. A manually delineated dataset based on Landsat 8 RGB imagery from multiple ice shelves, encompassing a wide range of crevasse morphologies, was constructed for model training and validation. To address the scan-line corrector (SLC) failure of Landsat 7 ETM+ since 2003, we developed a diffusion-based gap-filling approach trained on a dataset specifically constructed for this study, enabling consistent crevasse mapping across the full Landsat 7/8 archive.

Our results reveal pronounced crevasse development on Pine Island, Thwaites, and Larsen B Ice Shelves over the past two decades, while other mapped ice shelves exhibit more moderate or minimal changes. This long-term, high-resolution crevasse mapping provides new insights into ice shelf damage evolution and offers valuable constraints for damage parameterization and assessments of ice shelf stability. The developed pipeline is readily extendable to additional ice shelves and remains computationally efficient.