Deep learning for automated surveillance of sea ice dynamics

Effective monitoring of sea ice conditions is required for safe maritime operations and research on sea ice. Common surveillance systems rely on radar imagery from ships and coastal stations, which are often handled through a time-consuming process. Currently existing automated systems offer continuous monitoring but are often limited to coarse spatial resolutions and with computational efficiency restricting their use for detailed analysis and real-time navigation support. We leverage state-of-the-art deep learning-based optical flow architectures for continuous, high-resolution surveillance of sea ice dynamics with shipborne and coastal radar systems. By employing these architectures, we can accurately capture the relative motion and deformation fields of sea ice with considerably lower computational overhead. We demonstrate the applicability of the methods with operational radar systems to demonstrate their efficiency and accuracy.