Dynamic Similar-Year Feature Construction for Seasonal Arctic Sea Ice Prediction
Year: 2026
Type: dataset
Abstract: This model is a SwinUNet for seasonal Arctic sea ice concentration (SIC) forecasting. It replaces the usual multi-year climatology with a dynamic similar-year feature, so the historical background reflects interannual differences in sea ice evolution rather than only the mean seasonal cycle. At forecast initialization, the method compares the current SIC spatial pattern with historical SIC fields since 1979 and selects the year with the most similar sea ice extent using the Intersection over Union (IoU). The SIC evolution of that year over the forecast window is used as the historical-evolution feature. Continuous SIC and sea ice thickness (SIT) from the preceding four months are included as well, combining recent sea ice conditions with same-period evolution from the selected year. Forecasts are made at pentad intervals for four months, using NSIDC SIC and PIOMAS SIT from 1979 to 2024. Within the same SwinUNet framework, the similar-year scheme outperforms both the continuous-history and climatology feature schemes. Over the first 12 forecast periods, it reduces the mean absolute error (MAE) by 5.41% and 6.97% relative to those two schemes, and improves the anomaly correlation coefficient (ACC) by 1.63% and 4.03% and the binary accuracy (BACC) by 0.09 and 0.42. Against persistence and ECMWF SEAS5, the annual-mean MAE decreases by 22.19% and 42.24%, and BACC improves by 4.67% and 8.43%. The gain comes from the input features, without increasing model complexity.
Author: Wang, Mingfeng
Institution: Christian-Albrechts-Universität zu Kiel
Language: English
Subfield: Atmospheric Science
Field: Earth and Planetary Sciences
Domain: Physical Sciences
SDG: Climate action
Open Access status: green
OpenAlex ID: W7221479146