Sea Ice Edge Constraint Improves Antarctic Sea Ice Seasonal Prediction in Deep Learning Models

Predicting Antarctic sea ice is of substantial academic and practical significance. However, current prediction models, including deep learning (DL)-based models, show notable bias in the marginal ice zone. In this study, we developed a pure data-driven DL model for predicting the Antarctic austral summer monthly-to-seasonal sea ice concentration (SIC) by incorporating a novel hybrid sea ice edge constraint loss function (HybridLoss). The model is referred to as ASICNet. Independent testing based on the last five years (2019–23) demonstrates that ASICNet with HybridLoss achieves significantly higher skill metrics than without, with a reduced mean absolute error of 0.021 from 0.022, a reduced integrated ice edge error of 1.714 × 10⁶ from 1.794 × 10⁶ km², but an increased pattern correlation coefficient of 0.40 from 0.38, although both ASICNet versions outperform dynamical and statistical models. Furthermore, enhanced heat maps were developed to interpret the predictability sources of sea ice within DL-based models, and the results suggest that the predictability of Antarctic sea ice is attributable to factors like the Antarctic Dipole (ADP), Amundsen Sea Low (ASL), and Southern Ocean sea surface temperature (SST), as revealed in previous studies. Thus, ASICNet is an efficient tool for austral summer Antarctic SIC prediction.