On sea-ice forecasting

Abstract

Accurate sea-ice prediction is essential for safe operations in the Arctic and potentially also for weather forecast at high-latitudes. The increasing number of sea-ice related satellite observations in the Arctic can be used to improve the model predictions through data assimilation. For sea ice, sea-ice concentration (SIC) observations have been available for many years. Observational information of SIC can be used to constrain the sea-ice extent in models. In addition to SIC, other sea-ice related observations such as sea-ice thickness (SIT) and snow depth have recently become available. The assimilation of these observations is expected to have a substantial impact on the sea-ice forecast. In this thesis, the main goal is to enhance the sea-ice model forecast accuracy by improving the initial model state on which the forecast is based. Primarily, the assimilation of sea-ice-related observations that are previously little used in sea-ice data assimilation is investigated. This includes the assimilation of SIT, snow depth and high-resolution SIC observations. A secondary objective of this thesis is to reduce the computational cost of both sea-ice assimilation and modelling. A new direct and computationally cheap method for data assimilation, the Multi-variate nudging (MVN) method, is proposed as an alternative to more complex assimilation methods for sea-ice. In addition, to reduce the computational cost of the sea-ice prediction, two machine-learning methods were applied for sea-ice forecasting, a fully convolutional network and a k nearest neighbours. It is found that the assimilation of observations other than SIC has the potential to enhance the accuracy of sea-ice models and improve predictions. The proposed new assimilation method, the MVN, proves to be a valid assimilation alternative to the Ensemble Kalman Filter when few observation types are available, and the computational resources are limited. The machine-learning forecasts are found to improve upon persistence and show comparable skills to the dynamical model. Hence there is a potential for machine-learning methods for sea-ice predictions which should be developed further.