Clustering global ocean profiles according to temperature-salinity structure

TL;DR

This paper introduces a novel clustering method for global ocean profiles using Gaussian mixture models and path signatures, improving the identification of oceanic conditions from temperature-salinity data.

Contribution

It is the first to cluster nearly all Argo profiles by fully utilizing temperature, salinity, and pressure sequences with path signatures in an unsupervised framework.

Findings

Enhanced clustering accuracy over traditional methods

Effective identification of distinct oceanic conditions

Demonstrated applicability to large-scale Argo data

Abstract

An unsupervised clustering using a Gaussian mixture model is applied to Argo profiles distributed over the entire ocean. We employ as the coordinate components in feature space the path signature, which is a central notion in rough path theory. This allows us to better identify the oceanic condition at each horizontal point with distinct clusters, than by using conventional temperature and salinity coordinate. To the best of my knowledge, it is the first attempt at clustering almost all of the existing Argo profiles with the full use of measured sequences of temperature, salinity, and pressure. We will also discuss why the path signature is relevant to representing the property of a profile.