Similar-Year Feature Construction Improves Deep-Learning Arctic Sea Ice Forecasts

Abstract. Seasonal Arctic sea ice prediction is important for understanding Arctic climate change and its impacts. Existing deep-learning models commonly use the multi-year mean climatology of the forecast period as a historical-evolution background. This provides stable seasonal information, but has limited ability to represent interannual differences in sea ice evolution. We propose a dynamic similar-year feature construction method based on sea ice spatial similarity and incorporate it into the SwinUNet forecasting framework. Using sea ice concentration (SIC) at forecast initialization and the Intersection over Union (IoU) of sea ice extent (SIE), the method identifies the most similar historical year since 1979 and extracts its SIC evolution over the forecast period as the historical-evolution feature. Continuous SIC and sea ice thickness (SIT) from the preceding four months are also included, combining recent conditions with historical same-period evolution. Using NSIDC SIC and PIOMAS SIT data from 1979 to 2024, four-month Arctic sea ice forecasts are conducted at pentad intervals. The Similarity scheme is compared with Continuous and Climatology features in SwinUNet, as well as Persistence and ECMWF SEAS5. Similarity generally outperforms Continuous and Climatology. Over the first 12 forecast periods, it reduces MAE by 5.41% and 6.97%, improves ACC by 1.63% and 4.03%, and increases BACC by 0.09% and 0.42%, respectively. Relative to Persistence and SEAS5, the annual-mean MAE decreases by 22.19% and 42.24%, while BACC improves by 4.67% and 8.43%. These results show that dynamic similar-year features can enhance SwinUNet skill without increasing model complexity.