Multi-task predictions of the Arctic sea ice by a transformer-based deep learning model

The Arctic sea ice has been retreating dramatically in recent years in summer and fall. The navigation season for open water vessels along the Northeast Passage has lengthened to sub-seasonal scales. Accurate perditions of Arctic sea ice in sub-seasonal scales are essential for planning shipping activities. The numerical model cannot achieve a high accuracy of daily sea ice predictions on a sub-seasonal scale. The advanced deep learning brings new solutions for the data-driven-based sea ice prediction.

This study proposed a transformer-based deep learning model to predict multiple sea ice parameters, including sea ice concentration (SIC), sea ice thickness (SIT), and sea ice drift (SID), in the Pan-Arctic in a sub-seasonal scale, 90 days’ lead. An encoder and decoder are constructed based on transformer modules to extract spatio-temporal dependencies from daily SIC, SIT, and SID sequences. The spatio-temporal dependencies at different scales are fused to form the final feature maps. Three SIC, SIT, and SID output modules are designed based on the final feature maps to output different parameters for the next 90 days. The satellite-observed sea ice data from the National Sea Ice Data Center (NSIDC) are employed to train the proposed model. We compared our model with anomaly persistence and the European Centre for Medium-Range Weather Forecasts (ECMWF) ensemble predictions to demonstrate the model’s prediction skill.

Further, based on the proposed model, we discuss the effects of typical thermal and dynamic factors on sub-seasonal scale daily sea ice prediction. The selected factors include surface air temperature (SAT), sea surface temperature (SST), surface solar radiation downwards (SSRD), and geopotential height. Finally, we conclude with some scientific guidelines for the sub-seasonal sea ice predictability of the Arctic.