|
Abstract:
|
In recent decades, climate change has caused sharp reductions in the volume of sea ice in the Arctic Ocean. This has created demand for accurate forecasts of Arctic sea ice, for decisions about resource management and shipping. Such forecasts have two main components: where sea ice will be present, and how thick the ice will be if it is present. Existing methods rely on ensembles of deterministic dynamic models, but we show that these can be both biased and poorly calibrated. We propose a probabilistic contour model to predict the area where sea ice will be present, which corrects the bias in existing physical models and assesses their uncertainty. We then develop a Gaussian random field model for ice thickness, conditional on the ice-covered region. We apply our method to forecast Arctic sea ice thickness, and find that point predictions and prediction intervals from our model offer improved accuracy and calibration compared with existing forecasts. We also show that our model can generate well-calibrated short-term forecasts of aggregate quantities such as overall sea ice volume.
|
