Arctic sea ice drift monitoring is essential for understanding polar climate dynamics. Passive microwave radiometers provide valuable daily observations of the Arctic region. However, their use for sea ice drift extraction involves a trade-off among frequency channels: high-frequency channels offer finer spatial details but are strongly affected by atmospheric factors and melt-related radiometric noise, whereas low-frequency channels are more stable but too coarse to resolve small-scale drift. To address these challenges, this study proposes MF-IceDrift, a novel explainable deep learning framework for year-round, spatially complete Arctic sea ice drift extraction using multi-frequency AMSR2 data. Specifically, a Multi-frequency feature extraction and fusion module is designed to extract and fuse complementary features from different frequency channels. Moreover, a Transformer-based TransUpdate block and an augmentation consistency-based self-supervised strategy are incorporated to refine the flow estimation and ensure the spatial coherence of the drift fields. Validation against buoy observations demonstrates that MF-IceDrift achieves the best accuracy while maintaining 100% spatial coverage, consistently outperforming existing multi-frequency decision-level fusion methods as well as single-frequency baselines. The extracted drift fields exhibit a high degree of correlation with operational sea ice drift products, confirming the reliability of the large-scale drift patterns. Furthermore, the MF-IceDrift demonstrates generalization capabilities on the Microwave Radiation Imager onboard the Feng-Yun 3D (FY-3D MWRI) and Defense Meteorological Satellite Program Special Sensor Microwave Imager/Sounder (DMSP SSMIS) data. The deep learning network developed in this study provided a new approach for large-scale and long-term sea ice drift extraction using multi-frequency passive microwave data.
Jiang et al. (Mon,) studied this question.