| Summary: | This thesis introduces GraphSIFNet, 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 networks allows the domain to be discretized into arbitrarily specified meshes. This study demonstrates the model's ability to forecast over an irregular mesh with higher spatial resolution near shorelines, and lower resolution otherwise. Utilizing atmospheric data from ERA5 and oceanographic data from GLORYS12, the model is trained to model complex spatial relationships pertinent to sea ice dynamics. 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 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. Variable importance analysis via feature ablation was conducted which highlighted current sea ice concentration and thickness as critical predictors. Thickness was shown to be important at longer lead times during the melting season suggesting its importance as an indicator of ice longevity, while concentration was shown to be more critical at shorter lead times which suggests it may act as an indicator of immediate ice integrity. The thesis lays the groundwork for future exploration into dynamic mesh-based forecasting, the use of more complex graph structures, and mesh-based forecasting of climate phenomena beyond sea ice.
|
|---|