The accelerated decline of Arctic sea ice is profoundly reshaping regional climate regimes. Sea ice thickness (SIT), particularly under thin-ice conditions, is an important indicator for assessing early-season Arctic sea ice variability, and accurate prediction of its spatiotemporal evolution during the early freezing period is essential for characterizing short-term sea ice changes. In recent years, deep learning has emerged as a complementary approach to traditional sea ice prediction methods. However, existing deep learning-based studies have not fully exploited the large-scale spatial patterns and temporal contextual dependencies inherent in satellite-derived sea ice thickness. To address this limitation, this study proposes a spatiotemporal prediction framework named EOF-Trans for predicting daily-scale variability of thin sea ice thickness during the early freezing season. The method employs Empirical Orthogonal Functions (EOF) to decompose the sea ice thickness field into temporal mode series, utilizes a Transformer architecture to learn the temporal evolution characteristics, and subsequently reconstructs the predicted outputs back into the spatial thickness field by EOF, thereby enabling spatiotemporal sea ice prediction up to a leadtime of 21 days. Experimental results in the Beaufort Sea indicate that the proposed EOF-Trans framework significantly outperforms numerical models and classical deep learning architectures such as U-Net and ConvLSTM. On the 2022–2023 test set, it achieves a correlation coefficient of 88.04%, representing a 2% improvement over U-Net. Even at a leadtime of 21 days, the correlation remains approximately 84%, with the maximum spatial bias not exceeding 0.5 m. These results indicate that EOF-Trans effectively captures spatiotemporal regularities present in thin sea ice thickness, providing a complementary data-driven perspective for short-term sea ice prediction during the early freezing season.
Liu et al. (Mon,) studied this question.