Beyond Sea-Ice Concentration: Multivariate and Regime-Robust Learning for Seasonal Sea-Ice Prediction

摘要

Reliable seasonal sea-ice forecasting is challenging because the underlying climate system is strongly nonstationary, interactions among atmospheric and oceanic variables occur at multiple temporal scales, and forecast skill often deteriorates sharply across extreme years and long prediction horizons. Most learning-based approaches primarily optimize average predictive performance under historical train-test splits, providing limited guarantees that the learned relationships remain useful when climatic regimes shift. We introduce IceShift, a multivariate forecasting framework designed to learn predictive representations that remain stable across heterogeneous climate regimes. Rather than forecasting sea-ice concentration in isolation, IceShift jointly models sea-ice and relevant atmospheric-oceanic variables and decomposes representations into temporally persistent and regime-sensitive components. A regime-aware learning objective encourages the forecasting model to retain relationships that generalize across years while reducing reliance on unstable correlations. We further model delayed responses between environmental drivers and sea-ice evolution through multi-scale temporal interactions. The framework is evaluated using satellite and reanalysis observations under standard forecasting settings as well as targeted evaluations involving extreme years, seasonal transitions, and distribution shifts. Beyond aggregate prediction errors, we analyze degradation across forecast horizons, regional behavior, extreme-event performance, and calibration of predictive uncertainty. This work aims to shift sea-ice forecasting evaluation from average in-distribution accuracy toward robustness under the nonstationary conditions that matter most for real-world climate prediction.