Pan-Arctic Weekly Sea-Ice Forecasts: A Large-Scale Baseline Study and Meta-Learning Ensemble Model

Abstract

This thesis introduces a standardized benchmark for a diverse set of architectures for weekly Pan-Arctic sea-ice forecasts with a 13-week (91-day) horizon. This benchmark is introduced in combination with a meta-learning ensemble model which consistently produces improved ice forecasts. The diverse suite of deep learning baselines is trained and evaluated under a common protocol to predict sea-ice concentration (SIC), sea-ice thickness (SIT), and sea-ice presence (SIP) on a Pan-Arctic grid. Building on these baselines, the proposed Meta-Learner employs stacked generalization within a sequence-to-sequence forecasting setup, learning to fuse model outputs with spatio-temporal and lead-time context to improve robustness and reliability. Utilizing atmospheric reanalysis from ERA5 and oceanographic reanalysis from GLORYS12, the study assesses skill across short-, medium-, and long-lead regimes up to the 13-week horizon. Results demonstrate that the Meta-Learner outperforms the best individual baseline, with reductions in ice concentration MAE of 11% and a reduction in ice thickness MAE of 21% while reducing the cross-entropy in ice presence classification by 5%. Improvements are most pronounced in lower variance across random initializations for all tracked metrics, indicating enhanced stability. The bench- mark and ensemble framework provide a reproducible foundation for Pan-Arctic weekly forecasting and highlight learned ensembling as a practical pathway to more accurate and dependable SIC/SIT/SIP predictions at operationally relevant horizons.