Understanding Arctic Sea Ice Thickness Predictability by a Markov Model

The Arctic sea ice decline and associated change in maritime accessibility have created a pressing need for sea ice thickness (SIT) predictions. This study developed a linear Markov model for the seasonal prediction of model-assimilated SIT. It tested the performance of physically relevant predictors by a series of sensitivity tests. As measured by the anomaly correlation coefficient (ACC) and root-mean-square error (RMSE), the SIT prediction skill was evaluated in different Arctic regions and across all seasons. The results show that SIT prediction has better skill in the cold season than in the warm season. The model performs best in the Arctic basin up to 12 months in advance with ACCs of 0.7-0.8. Linear trend contributions to model skill increase with lead months. Although monthly SIT trends contribute largely to the model skill, the model remains skillful up to 2-month leads with ACCs of 0.6 for detrended SIT predictions in many Arctic regions. In addition, the Markov model's skill generally outperforms an anomaly persistence forecast even after all trends were removed. It also shows that, apart from SIT itself, upper-ocean heat content (OHC) generally contributes more to SIT prediction skill than other variables. Sea ice concentration (SIC) is a relatively less sensitive predictor for SIT prediction skill than OHC. Moreover, the Markov model can capture the melt-to-growth season reemergence of SIT predictability and does not show a spring predictability barrier, which has previously been observed in regional dynamical model forecasts of September sea ice area, suggesting that the Markov model is an effective tool for SIT seasonal predictions.