Distinguishing meanders of the Kuroshio using machine learning

Abstract

AbstractThe Kuroshio south of Japan is often described as being bimodal, with abrupt transitions between a straight path state that stays near the coast (small meander) and a meandering state that deviates from the coast (large meander). Despite evidence of the existence of two or more states of the Kuroshio, previous data‐driven studies have shown only high variability of the current; they have not, however, demonstrated bimodality in the sense of two states of relatively high probability separated by a region of relatively low probability. We use singular value decomposition (SVD), a standard time series analysis method for characterizing variability, and diffusion maps and spectral clustering (DMSC), a machine learning algorithm that seeks multimodality, to investigate Kuroshio reanalysis output. By applying these methods to a time series of velocity fields, we find that (1) the Kuroshio is bimodal, with high inflow and low path variability in the small meander and low inflow and high path variability in the large meander, (2) the state of the system correlates highly with the location of the recirculation gyre south of Japan, and (3) the meanders are better characterized by path variability than by mean path. Because these results are consistent with satellite sea surface height data, they are not an artifact of the model used for reanalysis. Further, our results provide evidence for a previously proposed transition mechanism based on the strengthening, migration, and weakening of the recirculation gyre south of Japan and can therefore help direct future modeling studies.