Characterizing the Kuroshio Current Using Machine Learning
David Plotkin, University of Chicago
The Kuroshio western boundary current runs along the southeastern coast of Japan. It transports heat and salinity poleward, influencing large-scale weather, industry, and biology. The Kuroshio is widely claimed to exist in one of two persistent states: a small-meander state in which it does not separate from the coast and a large-meander state in which the current axis separates from the coast by about 200 km. The residence time in each meander is reported as five to 10 years, with rare but extreme transitions between meanders occurring on a timescale of several months. This behavior implies a bimodal distribution of the distance of the current axis from the coast. Previous work has made progress on characterizing the two meanders but has not found conclusive evidence of two well-separated, persistent states. Using a diffusion maps and spectral clustering algorithm, we reduce the dimension of a high-dimensional time series of Kuroshio velocity fields. The algorithm treats the data as a random walk that is likely to persist on groupings of points with small pairwise separations, giving rise to clusters that would be missed by standard algorithms. By clustering the low-dimensional data, we distinguish the two meanders: Instead of states characterized respectively by persistently large and small path separation from the coast, we find that the meanders are best characterized by differences in variability. Small-meander months contain only paths that run along the coast, while large-meander months contain paths with a wide range of separations from the coast. Importantly, the mean paths in the small and large meanders are similar. This has made the two states difficult to characterize with other techniques.
Abstract Author(s): David Plotkin