The COSI project

In the COSI project, we are developing calibration techniques using supervised machine learning algorithms in order to improve short-term (10 days) sea-ice concentration forecasts produced by the Copernicus Marine Service (CMEMS) TOPAZ prediction system. The probability that the sea-ice concentration exceeds 10 % and 20 % will also be assessed in the calibrated forecasts in order to provide information to seafarers in agreement with the requirements from the International Code for Ships Operating in Polar Waters (International Maritime Organization, 2016). The calibrated forecasts will be available for demonstration from November 2023 and operationally in the CMEMS catalogue in 2025. 

In supervised machine learning, the quality of the observations used for training the models (target variable) can have a major impact on the performances. In COSI, an enhanced Pan-Arctic sea-ice concentration dataset is developed using observations from the Advanced Microwave Scanning Radiometer 2 (AMSR2) in order to train the machine learning models. It is expected that its higher spatial resolution (5 km) will be advantageous for developing accurate calibration methods.

COSI is organized through four work packages (WP):

• WP1 Management and communication
• WP2 Development of a Pan-Arctic sea-ice concentration satellite product for calibration
• WP3 Development of calibration methods
• WP4 Evaluation of the products (satellite observations from WP2 and calibrated forecasts from WP3).