Calibration of sea ice drift forecasts using random forest algorithms

Abstract

Developing accurate sea ice drift forecasts is essential to support the decision-making of maritime end-users operating in the Arctic. In this study, two calibration methods have been developed for improving 10 d sea ice drift forecasts from an operational sea ice prediction system (TOPAZ4). The methods are based on random forest models (supervised machine learning) which were trained using target variables either from drifting buoy or synthetic-aperture radar (SAR) observations. Depending on the calibration method, the mean absolute error is reduced, on average, between 3.3 % and 8.0 % for the direction and between 2.5 % and 7.1 % for the speed of sea ice drift. Overall, the algorithms trained with buoy observations have the best performances when the forecasts are evaluated using drifting buoys as reference. However, there is a large spatial variability in these results, and the models trained with buoy observations have particularly poor performances for predicting the speed of sea ice drift near the Greenland and Russian coastlines compared to the models trained with SAR observations.