Krinitskiy_etal_2019.pdf

A profound understanding of the stratospheric wintertime dynamics and its climate changes are important for improving seasonal forecast skill. The primary goal of the research of the wintertime Arctic stratospheric polar vortex (PV hereafter) is defining its states and their clustering. Manual classification is a highly time-consuming task suffering of researcher subjectivity. We apply deep learning methods that let us cluster the PV states based on their spatial structure. We designed the particular kind of neural networks called variational convolutional autoencoder with the sparsity constraint (SpCVAE). We applied the hierarchical agglomerative clustering algorithm to the states pf PV described by their embedded representation generated by SpCVAE. 96-dimensional embedded representation was found to be optimal with high samples reconstruction quality. The best number of clusters was chosen based on "elbow rule" and topic-specific reasoning. The approach applied let us automatically distinguish weak PVs of "displacement" and "split" types, as well as to isolate several strong vortex states of different shift directions. These results are only obtainable when one considers the spatial structure of the PV. We have constructed the calendar of the PV states based on the clustering result. Clustered events of weak PVs were examined and demonstrated good correspondence with the calendar of sudden stratospheric warmings that have been built manually. This result is now the basis for the research of the stratosphere-troposphere interaction for existing and future climate scenarios.