7 June, 2018
What can machine learning tell us about the structure of the Southern Ocean? In this preprint*, we apply unsupervised clustering, a machine learning method, to Southern Ocean temperature data.
* UPDATE: Published paper now available here
The Southern Ocean has complex spatial variability, characterized by sharp fronts, steeply tilted isopycnals, and deep seasonal mixed layers. Methods of defining Southern Ocean spatial structures traditionally rely on somewhat ad hoc combinations of physical, chemical, and dynamic properties. As a step toward an alternative approach for describing spatial variability in temperature, here we apply an unsupervised classification technique (i.e., Gaussian mixture modeling or GMM) to Southern Ocean Argo float temperature profiles. GMM, without using any latitude or longitude information, automatically identifies several spatially coherent circumpolar classes influenced by the Antarctic Circumpolar Current. In addition, GMM identifies classes that bear the imprint of mode/intermediate water formation and export, large‐scale gyre circulation, and the Agulhas Current, among others. Because GMM is robust, standardized, and automated, it can potentially be used to identify structures (such as fronts) in both observational and model data sets, possibly making it a useful complement to existing classification techniques.