Antarctic sea ice predictions are becoming increasingly important scientifically and operationally due to climate change and increased human activities in the region. We evaluated the Antarctic sea ice edge prediction skill of the Copernicus Climate Change Service (C3S) and Subseasonal to Seasonal (S2S) projects with a probabilistic metric, the spatial probability score (SPS). We found that predictions by individual dynamical systems remain skillful for up to 38 days (i.e., the ECMWF system). Our analysis reveals that the model initialization is the crucial prerequisite for skillful subseasonal sea ice prediction. For those systems with the most realistic initialization, the model physics dictates the propagation of initialization errors and, consequently, the temporal length of predictive skill. Additionally, we found that the SPS-characterized prediction skill could be improved by increasing the ensemble size to gain a more realistic ensemble spread. Based on the C3S systems, we constructed a multimodel forecast from the above principles. This forecast consistently demonstrated a superior prediction skill compared to individual dynamical systems or statistical observation-based benchmarks. Further, we trained a convolutional long short-term memory (ConvLSTM) deep neural network using only satellite-derived sea ice concentration (SIC) from 1989 to 2016. The network is skillful for approximately one month in predicting the daily spatial distribution of Antarctic SIC between 2018 and 2022, with the best predictive skill found in austral autumn (MAM) and winter (JJA). The spatiotemporal features of the ConvLSTM prediction skills are comparable to the dynamical systems, implying that it has the potential to capture sea ice evolution features.
 

sea ice, deep learning