Abstract
Abstract
Accurately predicting Antarctic sea ice on a subseasonal‐to‐seasonal scale remains a challenge for current numerical models, partly due to imperfect model parameterizations and the extensive computational resources required. Here, we have developed a lightweight machine learning model, Ice‐k‐nearest neighbor (kNN)‐South, to predict Antarctic sea ice concentration anomaly (SICA) up to 90 days in advance. The lightweight sea ice prediction model can be executed efficiently with limited computational resources. Compared with anomaly persistence, climatology, and the European Centre for Medium‐Range Weather Forecasts predictions, the Ice‐kNN‐South exhibits improved performance for almost 90 lead days, as evaluated by the mean absolute error (MAE), anomaly correlation coefficient (ACC), and integrated ice edge error (IIEE). Besides, the improvements of Ice‐kNN‐South are more pronounced in regions with high SICA variability by implementing a multi‐scale structure that combines field‐to‐field and grid‐to‐grid predictions. Furthermore, Ice‐kNN‐South shows the most significant enhancement in predictive skill during the summer and autumn seasons, particularly in summer. In contrast, relatively smaller enhancements are observed in the West Pacific Ocean and the Ross Sea during winter and spring. Even during the years of minimum sea ice areas (SIAs) in 2017, 2022, 2023, and 2024, Ice‐kNN‐South accurately captures sea ice variations, maintaining monthly averaged SIA errors below 8.5% compared to satellite observations within a 30‐day lead time. These findings suggest a potential approach to improve the accuracy of Antarctic sea ice predictions, particularly for the use of data‐driven models.