Predicting the evolution of Antarctic sea ice with data science

Details

Supervisors

Delvenne, Jean-Charles ; Massonnet, François

Faculty

Ecole polytechnique de Louvain

Degree label

Master [120] en science des données, orientation technologies de l'information, à finalité spécialisée

Abstract

Data science makes it possible to apply a wide range of methods to problems in different fields. These include the prediction of sea ice concentration at the poles, which has a number of implications, including navigation, economic and diplomatic interests, and a better understanding of our planet. At the same time, during the year 2023, Antarctica experienced its lowest peak of sea ice extents ever recorded. This thesis evaluates the ability of neural networks, and more specifically the U-Net architecture, to predict Antarctic sea ice concentration by using and replicating data from a dynamical model. The models are compared with two baselines that represent the minimum performance that a more complex model must achieve in order to be relevant, a climatological forecast and a linear model. In addition, two different-sized versions of the U-Net architecture are used to compare the impact of model size on the results. Each model is evaluated from different angles to understand its strengths and weaknesses. The results show that the two U-Net architectures achieve better results than the baselines in terms of both mean error and error variance. However, no improvement is detected when using a larger model, with both versions delivering similar performances.