Abstract
Prydz Bay represents a critical region for studying East Antarctic coastal ice-ocean-atmosphere interactions, with its sea ice–open water–iceberg system playing a vital role in regional climate feedback and marine ecosystems. Based on austral summer Sentinel-1 SAR imagery from 2016 to 2024, this study employed a DenseNet-enhanced Dual-Branch U-Net (DBU-Net) model to achieve high-precision extraction of sea ice, open water, and icebergs (mean Intersection over Union: 0.964, overall accuracy: 0.9810). The segmentation results reveal significant spatiotemporal changes in the cryospheric system, characterized by an overall decrease in sea ice area coupled with an increase in open water extent, while iceberg coverage shows substantial year-to-year fluctuations. Analysis using the Geodetector method demonstrates that these changes are driven by a wind-dominated, thermally synergistic mechanism. Zonal wind speed (u10) serves as the primary controlling factor for both sea ice (q = 0.532) and open water (q = 0.735) distributions, with its interaction with meridional wind (v10) significantly enhancing the explanatory power (q = 0.773 and 0.931, respectively). Notably, the number of positive temperature days (D0 ≥ 3) acts as a crucial thermal threshold that preconditions sea ice for enhanced wind-driven retreat through albedo feedback and lateral melting. The identified framework of ‘deep learning extraction–spatiotemporal change–driver attribution’ is built upon a task-adapted deep learning model specifically designed for polar SAR imagery interpretation. This framework not only advances our understanding of Antarctic coastal dynamics but also provides valuable insights for scientific operations, including navigation safety for Antarctic expeditions.
Showing the abstract — retrieve the full paper via the Exa API.