Shelf-Bench: A benchmark dataset for Antarctic ice shelf front and coastline delineation from multi-sensor radar satellite data

Abstract. Continuous monitoring of Antarctic ice shelf fronts is essential for understanding ice sheet dynamics, detecting iceberg calving events, supporting operational logistics, and generating up-to-date continental maps. However, the automated and continuous delineation of ice shelf fronts has been held back by a lack of suitable training data for deep learning models. We present Shelf-Bench, a comprehensive benchmark dataset comprising 161 manually annotated SAR scenes from three satellite sensors (ERS, Envisat, and Sentinel-1), providing spatial coverage of the Antarctic coastline with multi-temporal seasonal acquisitions spanning 1992–2021. The dataset features manually delineated masks paired with pre-processed imagery at moderate spatial resolution. Through complexity analysis, we characterize delineation challenges, including fast ice, crevassed surfaces, dense iceberg mélange, and limited spatial context. We evaluate five state-of-the-art semantic segmentation architectures, establishing baseline performance metrics. Baseline models showed strongly contrasting behaviour on Shelf-Bench: architectures that achieved higher pixel-wise accuracy tended to produce larger boundary errors, while models with better geometric precision obtained lower semantic scores. This trade-off indicates that the dataset jointly challenges ice-ocean classification and fine-scale calving front delineation, revealing complementary challenges which make it a profound benchmark for automated ice front mapping. By providing this open-access, standardized benchmark, Shelf-Bench enables accelerated development of deep learning methodologies for automated Antarctic coastline detection and supports continuous monitoring across current and future SAR satellite missions. The Shelf-Bench dataset is available at https://doi.org/10.5281/zenodo.17610870.