Abstract The variational autoencoder (VAE), a deep generative model, can extract a good feature representation for clustering from complex data; however, the use of this algorithm in the geophysical fluid circulation has been limited. The sample size for a geophysical phenomenon is generally small because of a large dimensional size, especially for extreme (i.e., rare) events. This study proposes a method for clustering geophysical circulations with a small sample size based on the VAE algorithm, leveraging an augmentation technique based on noise injection scaled by the principal components for the target data. For idealized circulation fields, we confirm that this principal‐component‐scaled augmentation greatly improves the accuracy of the identification of cluster means and decision boundaries compared with the case without augmentation. We apply the proposed clustering method to circulation fields (velocity vector fields) in a coastal ocean region for a Kyucho, which is an extreme oceanic phenomenon that causes a sporadic strong current along the coast. The proposed method successfully identified four modes for a data set consisting of about 150 samples; the method without augmentation cannot identify them. In a group‐wise ensemble time series analysis, these four modes are shown to be caused by different processes, suggesting that the VAE clustering is capable of identifying dynamically significant modes and can facilitate finding their precursors.
Aoki et al. (Tue,) studied this question.