Comparative analysis of different spatial clustering techniques to model Arctic sea ice dynamics

The dynamics of sea ice in Polar regions has attracted increasing scientific interest over the last few decades, for its role in the global climate processes, as well as economic interest for the potential effect of new arctic routes on international trades, and natural gas and oil exploitation. Sea ice evolution is usually represented by adopting multi-layer thermodynamic-dynamic models coupled with atmosphere-ocean general circulation models. First-order physical representations of sea ice are currently included in large-scale models, but whether they are sufficient or not depends on the application. For instance, their spatial and temporal scales may not be sufficiently resolved for operational forecasting. The most relevant model upgrades for improving sea-ice predictions might be made in the atmosphere-ocean interplay mechanisms, more than details of the ice physics. The combination of unprecedented satellite datasets, increased computational power, and the advances in machine learning offer exciting opportunities for expanding our knowledge of the sea-ice trends and their main drivers. Arctic ice dynamic is usually interpreted as the combination of a long-term trend, most of the times considered as linear, and the deviation from this trend, i.e., the interannual variability. Yet, this approach is highly dependent on the linearity assumption, that appears simplistic and could affect the following analyses. We thus focus on the whole time series of ice data and explore its spatiotemporal evolution via time series clustering. We comparatively analyze the ability of three clustering algorithms to detect patterns in the PIOMAS ice thickness dataset, that reports monthly reanalysis data from 1978 to 2020. K-means, mean-shift, and hierarchical algorithms are adopted to represent centroid-, density- and connectivity-based clustering. Our results show that unsupervised machine learning can advance the interpretability of the complex phenomena occurring in the Arctic region. In addition, the proposed clustering analysis is a promising preprocessing tool for supervised tasks, such as forecasting and input selection. The methodology developed can be applied to other variables and spatial domains, and can also be easily extended to the multivariate case to consider the cross correlations.