Self-SIDNet: A Self-Supervised Network for Winter Arctic Sea Ice Drift Extraction From AMSR2 Passive Microwave Data

Accurate monitoring of sea ice drift is essential for Arctic climate studies, navigation safety, and modeling polar ocean-atmosphere interactions. Passive microwave radiometers, such as Advanced Microwave Scanning Radiometer 2 (AMSR2), provide wide-swath, all-weather observations, making them valuable for drift extraction. However, existing template matching methods often produce sparse drift fields and rely on manually tuned thresholds, while conventional optical flow algorithms exhibit limited performance in terms of extraction accuracy. In this study, we propose a novel deep learning framework, Self-SIDNet, for extracting dense and accurate Arctic sea ice drift fields from AMSR2 89 GHz brightness temperature data. A cycle consistency loss is introduced as a self-supervision signal that enforces forward-backward drift coherence, which improves the performance of the method without relying on ground-truth labels. Self-SIDNet integrates photometric consistency, smoothness regularization, and cycle consistency into a unified loss function, and applies a sea ice concentration mask to restrict learning to ice-covered regions. Experiments show that Self-SIDNet achieves RMSEs of 6. 17 and 6. 75nbsp;cm/s for horizontal (u) and vertical (v) components, consistently outperforming traditional methods, including the traditional template-matching-based continuous maximum cross-correlation (CMCC) method and the traditional optical flow-based Farneback method. Notably, it provides full spatial coverage of the drift field, resolving the precision-coverage trade-off in CMCC. Comparisons with the National Snow and Ice Data Center (NSIDC) Polar Pathfinder product show high correlation and minimal bias, confirming the reliability of the extracted drift fields, while providing higher spatial resolution for fine-scale drift monitoring. These results demonstrate the potential of self-supervised deep learning for sea ice drift extraction from passive microwave data and suggest its applicability for large-scale, long-term monitoring in polar regions.