Deep graph neural networks for spatiotemporal forecasting of sub-seasonal sea ice: A case study in Hudson Bay

This study introduces GraphSIFNet ( Graph S ea I ce F orecast neural Net work), a novel graph-based deep learning framework for spatiotemporal sea ice forecasting. GraphSIFNet employs a Graph Long-Short Term Memory (GCLSTM) module within a sequence-to-sequence architecture to predict daily sea ice concentration (SIC) and sea ice presence (SIP) in Hudson Bay over a 90-day time horizon. The use of graph neural networks (GNNs) allows the domain to be discretized into arbitrarily specified meshes, allowing more explicit spatial modeling than approaches based on the convolutional neural network (CNN). This study demonstrates the model’s ability to forecast over an irregular mesh with higher spatial resolution near shorelines. The model is trained using atmospheric data from ERA5 and oceanographic data from GLORYS12. Results demonstrate the model’s superior skill over a linear combination of persistence and climatology as a statistical baseline. The model showed skill particularly in short- to medium-term (up to 35 days) SIC forecasts, with a noted reduction in root mean squared error (RMSE) by up to 10% over the statistical baseline during the break-up season, and up to 5% in the freeze-up season. Long-term (up to 90 days) SIP forecasts also showed significant improvements over the baseline, with increases in accuracy of around 10% even at a lead time of 90 days. The use of an attention-based convolution offered the additional benefit of interpretability by highlighting the primary direction and magnitude of information flow that aligned with the direction of freezing and melting. The study lays the groundwork for future exploration into dynamic graph-based forecasting, and future work forecasting ice-ocean phenomena.