A robust multitask deep learning algorithm for Antarctic ice shelf fracture detection from multisource satellite imagery

Quantitative information on fracture distributions across Antarctic ice shelves is crucial for improving the predictability of ice shelf calving and retreat. Satellite imagery acquired by various optical and radar sensors provides a valuable resource for studying fracture processes at different spatial and temporal scales. Retrieving high-resolution fracture information at extended spatial and frequent temporal scales is particularly important for understanding the complex dynamics of ice shelves and improving the predictive modeling of their evolution. To efficiently retrieve spatiotemporal fracture information, we developed a novel multitask learning U-shaped network architecture (MLU-Net) for automatic and accurate fracture detection (including fracture areas and lines) from optical and synthetic aperture radar (SAR) images. We evaluated the performance and generalizability of MLU-Net through comprehensive experiments on the Amery, Thwaites Eastern, and Brunt ice shelves. Our results show that MLU-Net successfully maps the multitemporal fracture distributions over the three ice shelves in Landsat-8 Operational Land Imager (OLI) and Sentinel-1 SAR images, and that fusing the optical- and SAR-derived results enhances fracture detection. Quantitatively, MLU-Net achieves the highest F1 scores of 0.955 and 0.928 on the optical and SAR images, respectively, with a mean overall accuracy of 0.905 and a mean intersection over union of 0.792 across both image types. Beyond extracting fracture areas and lines, we formulated two fracture indices (fracture area-to-area and length-to-area ratios) that account for fracture distributions to characterize structural damage. Our findings indicate that the fracture area-to-area ratio is more sensitive to rift widening, while the fracture length-to-area ratio more effectively captures newly formed fractures with minimal sensitivity to image types and imaging conditions. Therefore, the proposed algorithm and fracture indices provide a promising framework for measuring ice shelf fractures at a continental scale, advancing the monitoring and modeling of ice shelf dynamics.