Keywords: spatiotemporal, spatial, environmental statistics, computer experiments
Reducing sea ice cover in the Arctic has increased the need for accurate forecasts of sea ice. Predicting the contour of ice-covered regions, or the boundary that surrounds regions with greater than 15% of their area covered in sea ice, is a focus of this effort. Current sea ice forecasts are issued based on output from deterministic prediction systems. While these systems provide useful information, they have biases that vary across space and time. Using historical predictions and satellite observations of sea ice, we develop a spatiotemporal model that anticipates how contours forecast by deterministic prediction systems will differ from what is observed. Forecasted contours can then be moved to correct for these expected errors, creating more accurate predictions. This new statistical post-processing technique, called contour-shifting, results in an average 20.3% reduction in the area misclassified as containing sea ice or not for a test set of monthly predictions and observations from 2001-2013.