Hybrid Swin-CNN Autoencoder for Arctic Sea Ice Anomaly Detection Using Monthly Passive Microwave Satellite Observations

Abstract

Arctic sea ice is important for the global climate system, polar ecosystems, and polar regions’ navigation. Precise recognition of abnormal patterns of Arctic sea ice concentration is significant to observe climate variability and detect the unusual change in polar regions. In this study, we propose a Hybrid Swin-CNN Autoencoder for Arctic sea ice anomaly detection based on monthly satellite observations of the NOAA/NSIDC Climate Data Record of passive microwave sea ice concentration. Monthly sea ice concentration in the northern hemisphere from 2010 to 2024 are used, data in 2010–2020 are used as the training set and data in 2021–2024 are used as the test set. Our model has the combination of convolutional encoder-decoder layer to do the local spatial feature reconstruction and Swin-style window attention layer for discovering the global spatial feature in the sea ice concentration distribution. Comparison of the proposed method To verify our method, four comparison models were evaluated: statistical climatology baseline model, CNN Autoencoder, Light Swin Autoencoder, and Swin-CNN refinement variant. Experimental results show that the proposed Hybrid Swin-CNN Autoencoder outperforms the other models in every evaluation measure, for example, the error measure (MSE = 0.000143, MAE = 0.003391, RMSE = 0.011673, SSIM = 0.994328), indicating that our proposed model is able to better reconstruct the pixel-level value and preserve the spatial structure of sea ice than the baseline models. Error maps also verify that high-error zones mainly happen at or near sea ice boundaries and marginal ice areas, showing that the reconstruction-error-related anomaly maps are meaningful for localizing abnormal sea ice patterns.

Introduction

The text presents a research study on Arctic sea ice anomaly detection using deep learning. Arctic sea ice is a key indicator of climate change, and satellite passive microwave data from NOAA/NSIDC (1978–present) is widely used to monitor sea ice concentration. Traditional methods like climatology comparison and statistical anomaly detection struggle to capture complex spatial patterns and nonlinear changes in sea ice.

To address this, the study proposes a Hybrid Swin-CNN Autoencoder for unsupervised anomaly detection. It combines CNNs to capture local spatial features (such as ice edges and textures) with Swin Transformer-based attention to model broader spatial dependencies. The model uses reconstruction error to detect anomalies, trained on monthly sea ice data from 2010–2024 (resized to 64×64 grids).

Key ideas include:

• Deep learning methods (CNNs, U-Net, Transformers) outperform traditional statistical approaches in modeling sea ice patterns.
• Autoencoders detect anomalies by reconstructing normal patterns; high reconstruction error indicates unusual sea ice conditions.
• The hybrid model improves over pure CNN or pure Transformer approaches by balancing local detail preservation and global spatial understanding.
• Performance is evaluated using MSE, MAE, RMSE, and SSIM, along with visual analysis of reconstruction errors.

The dataset used is the NOAA/NSIDC passive microwave sea ice concentration dataset, split into training (2010–2020) and testing (2021–2024). The study concludes that the proposed hybrid model better captures sea ice spatial structure and improves anomaly detection compared to baseline methods.

Conclusion

This study introduced a Hybrid Swin-CNN Autoencoder to detect anomalies in Arctic sea ice by reconstructing monthly sea ice concentration data from the NSIDC, covering 2010 to 2024. The model combines CNNs for capturing local features with Swin-style window attention in the bottleneck, which helps keep local sea ice patterns intact while also considering broader spatial relationships. Gated skip connections were used to bring back important spatial details without simply copying the input to the output. The new model was tested against four others: a statistical monthly climatology baseline, a CNN Autoencoder, a Lightweight Swin Autoencoder, and a Swin-CNN Refine Autoencoder. Results from 2021 to 2024 showed that the Hybrid Swin-CNN Autoencoder outperformed them all based on measures like MSE, MAE, RMSE, and SSIM. This suggests the method does better at both reconstructing sea ice concentration at the pixel level and keeping the overall structure. When looking at where reconstruction errors occurred, they mainly appeared around sea ice edges and marginal zones. This means the model is picking up meaningful anomalies rather than random noise. So, the Hybrid Swin-CNN Autoencoder offers a relatively simple but effective way to spot anomalies in Arctic sea ice using satellite data. Looking ahead, there are several ways to improve this approach. One option is to use sea ice data with higher resolution to better identify small-scale anomalies. Another is to include other environmental factors like sea surface temperature, air temperature, wind, or ice thickness to create a system that detects anomalies using multiple data types. It might also help to analyze longer time periods to track how anomalies change over time. Finally, testing the method on Antarctic sea ice or other climate anomaly detection problems could show how well it works in different settings.

References

[1] W. N. Meier, F. Fetterer, A. K. Windnagel, J. S. Stewart, and T. Stafford, “NOAA/NSIDC Climate Data Record of Passive Microwave Sea Ice Concentration, Version 6,” National Snow and Ice Data Center, Boulder, CO, USA, 2026, doi: 10.7265/b18j-z797. [2] T. R. Andersson et al., “Seasonal Arctic sea ice forecasting with probabilistic deep learning,” Nature Communications, vol. 12, no. 1, Art. no. 5124, 2021, doi: 10.1038/s41467-021-25257-4. [3] C. Palerme et al., “Improving short-term sea ice concentration forecasts using deep learning,” The Cryosphere, vol. 18, pp. 2161–2182, 2024, doi: 10.5194/tc-18-2161-2024. [4] J. Park, Y. Cho, J.-J. Jeon, J. Park, H.-C. Kim, and S. Hong, “Unicorn: U-Net for sea ice forecasting with convolutional neural ordinary differential equations,” Scientific Reports, vol. 15, Art. no. 36330, 2025, doi: 10.1038/s41598-025-20097-4. [5] Y. Ren, X. Li, and Y. Wang, “SICNetseason V1.0: A transformer-based deep learning model for seasonal Arctic sea ice prediction by incorporating sea ice thickness data,” Geoscientific Model Development, vol. 18, pp. 2665–2678, 2025, doi: 10.5194/gmd-18-2665-2025. [6] G. E. Hinton and R. R. Salakhutdinov, “Reducing the dimensionality of data with neural networks,” Science, vol. 313, no. 5786, pp. 504–507, 2006, doi: 10.1126/science.1127647. [7] Z. Liu, Y. Lin, Y. Cao, H. Hu, Y. Wei, Z. Zhang, S. Lin, and B. Guo, “Swin Transformer: Hierarchical vision transformer using shifted windows,” in Proc. IEEE/CVF International Conference on Computer Vision (ICCV), 2021, pp. 9992–10002, doi: 10.1109/ICCV48922.2021.00986. [8] D. J. Cavalieri, P. Gloersen, and W. J. Campbell, “Determination of sea ice parameters with the Nimbus 7 SMMR,” Journal of Geophysical Research, vol. 89, no. D4, pp. 5355–5369, 1984, doi: 10.1029/JD089iD04p05355. [9] J. C. Comiso, “Characteristics of Arctic winter sea ice from satellite multispectral microwave observations,” Journal of Geophysical Research, vol. 91, no. C1, pp. 975–994, 1986, doi: 10.1029/JC091iC01p00975. [10] O. Ronneberger, P. Fischer, and T. Brox, “U-Net: Convolutional networks for biomedical image segmentation,” in Proc. Medical Image Computing and Computer-Assisted Intervention (MICCAI), 2015, pp. 234–241, doi: 10.1007/978-3-319-24574-4_28. [11] Y. J. Kim, H. Kim, D. Han, J. Stroeve, and J. Im, “Long-term prediction of Arctic sea ice concentrations using deep learning: Effects of surface temperature, radiation, and wind conditions,” Remote Sensing of Environment, vol. 318, Art. no. 114568, 2025, doi: 10.1016/j.rse.2024.114568. [12] Z. Wang, A. C. Bovik, H. R. Sheikh, and E. P. Simoncelli, “Image quality assessment: From error visibility to structural similarity,” IEEE Transactions on Image Processing, vol. 13, no. 4, pp. 600–612, 2004, doi: 10.1109/TIP.2003.819861.

Copyright

Copyright © 2026 Ouadghiri Ouafae, Jingming Xia, Ibrahim Khaled Q.F. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.