Extended seasonal prediction of Antarctic sea ice using ANTSIC-UNet

Abstract. Antarctic sea ice has experienced rapid change in recent years, which garners increasing attention for its prediction. In this study, we develop a deep learning model (named ANTSIC-UNet) trained by physically enriched climate variables and evaluate its skill for extended seasonal prediction of Antarctic sea ice concentration (up to 6 months in advance). We compare the predictive skill of ANTSIC-UNet in the Pan- and regional Antarctic with two benchmark models (linear trend and anomaly persistence models). In terms of root-mean-square error (RMSE) for sea ice concentration and integrated ice-edge error (IIEE), ANTSIC-UNet shows much better skills for the extended seasonal prediction, especially for the extreme events in recent years, relative to the two benchmark models. The predictive skill of ANTSIC-Unet is season and region dependent. Low values of RMSE are found from autumn to spring in the Pan-Antarctic and all sub-regions for all lead times, but large values of RMSE are found in summer for most sub-regions which increase as lead times increase. Small values of IIEE are found in summer at 1–3 month lead, large errors occur from November to January as the lead time exceeds 2–4 months. The Pacific and Indian Oceans show better predictive skills at the sea ice edge zone in summer compared to other regions. Moreover, ANTSIC-UNet shows good predictive skill in capturing the interannual variability of Pan-Antarctic and regional sea ice extent anomalies. We also quantify variable importance through a post-hoc interpretation method. It suggests in addition to sea ice conditions, the ANTSIC-UNet prediction at short lead times shows sensitivity to sea surface temperature, radiative flux, and atmospheric circulation. At longer lead times, zonal wind in the stratosphere appears to be an important influencing factor for the prediction.