A Hierarchical Multitask U-Net for Automated Sea Ice Mapping from AI4Arctic Sea Ice Challenge Dataset

Sea ice monitoring is a vital task for various applications, such as climate research, ship navigation, and Arctic sea route planning. However, manual generation of sea ice condition maps using synthetic aperture radar (SAR) imagery is time-consuming and labor intensive. To address this issue, the European Space Agency (ESA)-lab has created the AutoICE challenge with the AI4Arctic Sea Ice Challenge Dataset provided, which aims to develop deep learning (DL) models capable of producing segmentation maps for ice concentration (SIC), stage of development (SOD), and floe size (FLOE). Among different DL-based models, U-Net is a popular choice used for different sea ice segmentation tasks. However, the ice-water boundaries predicted by U-Net are not always the same between the three different tasks. In this paper, we investigate two methods to address this inconsistency and assess their impact on the model's performance. The first method is the introduction of a new loss function, the water consistency loss, which penalizes inconsistencies in the open-water class between the three outputs. The second method is to use a hierarchical U-Net, which first segments water and ice and then classifies the ice regions into finer categories. The results show that both approaches eliminate the inconsistencies between tasks effectively.