Improving short-term sea ice concentration forecasts using deep learning

Reliable short-term sea ice forecasts are needed to support maritime operations in polar regions. While sea ice forecasts produced by physical-based models still have limited accuracy, statistical post-processing techniques (often called calibration) can be applied to reduce forecast errors. In this study, post-processing methods based on supervised machine learning have been developed for improving the skill of sea ice concentration forecasts from the TOPAZ4 prediction system for lead times from 1 to 10 days. The deep learning models use predictors from TOPAZ4 sea ice forecasts, weather forecasts, and sea ice concentration observations. On average, the forecasts from the deep learning models have a root mean square error 41 % lower than TOPAZ4 forecasts, and 29 % lower than forecasts based on persistence of the sea ice concentration observations. They also significantly improve the forecasts for the location of the ice edges, with similar improvements as for the root mean square error. Furthermore, the impact of different type of predictors (observations, sea ice and weather forecasts) on the predictions has been evaluated. Sea ice observations are the most important type of predictors, and the weather forecasts have a much stronger impact on the predictions than sea ice forecasts.