Clusters of interannual sea ice variability in the northern hemisphere

We determine robust modes of the northern hemisphere (NH) sea ice variability on interannual timescales disentangled from the long-term climate change. This study focuses on sea ice thickness (SIT), reconstructed with an ocean–sea-ice general circulation model, because SIT has a potential to contain most of the interannual memory and predictability of the NH sea ice system. We use the K-means cluster analysis—one of clustering methods that partition data into groups or clusters based on their distances in the physical space without the typical constraints of other unsupervised learning statistical methods such as the widely-used principal component analysis. To adequately filter out climate change signal in the Arctic from 1958 to 2013 we have to approximate it with a 2nd degree polynomial. Using 2nd degree residuals of SIT leads to robust K-means cluster patterns, i.e. invariant to further increase of the polynomial degree. A set of clustering validity indices yields K = 3 as the optimal number of SIT clusters for all considered months and seasons with strong similarities in their cluster patterns. The associated time series of cluster occurrences exhibit predominant interannual persistence with mean timescale of about 2 years. Compositing analysis of the NH surface climate conditions associated with each cluster indicates that wind forcing seem to be the key factor driving the formation of interannual SIT cluster patterns during the winter. Climate memory in SIT with such interannual persistence could lead to increased predictability of the Artic sea ice cover beyond seasonal timescales.