Identifying Ocean Submesoscale Activity From Vertical Density Profiles Using Machine Learning

Yao, Leyu; Taylor, John R. ORCID: https://orcid.org/0000-0002-1292-3756; Jones, Dani C. ORCID: https://orcid.org/0000-0002-8701-4506; Bachman, Scott D. ORCID: https://orcid.org/0000-0002-6479-4300. 2025 Identifying Ocean Submesoscale Activity From Vertical Density Profiles Using Machine Learning. Earth and Space Science, 12 (1), e2022EA002618. 18, pp. 10.1029/2022EA002618

Before downloading, please read NORA policies.

Preview

Text (Open access) © 2025. The Author(s). Earth and Space Science - 2025 - Yao - Identifying Ocean Submesoscale Activity From Vertical Density Profiles Using Machine.pdf - Published Version Available under License Creative Commons Attribution 4.0. Download (6MB) | Preview

Abstract/Summary

Submesoscale eddies are important features in the upper ocean where they mediate air-sea exchanges, convey heat and tracer fluxes into ocean interior, and enhance biological production. However, due to their small size (0.1–10 km) and short lifetime (hours to days), directly observing submesoscales in the field generally requires targeted high resolution surveys. Submesoscales increase the vertical density stratification of the upper ocean and qualitatively modify the vertical density profile. In this paper, we propose an unsupervised machine learning algorithm to identify submesoscale activity using vertical density profiles. The algorithm, based on the profile classification model (PCM) approach, is trained and tested on two model-based data sets with vastly different resolutions. One data set is extracted from a large-eddy simulation (LES) in a 4 km by 4 km domain and the other from a regional model for a sector in the Southern Ocean. We show that the adapted PCM can identify regions with high submesoscale activity, as characterized by the vorticity field (i.e., where surface vertical vorticity Submesoscale eddies are important features in the upper ocean where they mediate air-sea exchanges, convey heat and tracer fluxes into ocean interior, and enhance biological production. However, due to their small size (0.1–10 km) and short lifetime (hours to days), directly observing submesoscales in the field generally requires targeted high resolution surveys. Submesoscales increase the vertical density stratification of the upper ocean and qualitatively modify the vertical density profile. In this paper, we propose an unsupervised machine learning algorithm to identify submesoscale activity using vertical density profiles. The algorithm, based on the profile classification model (PCM) approach, is trained and tested on two model-based data sets with vastly different resolutions. One data set is extracted from a large-eddy simulation (LES) in a 4 km by 4 km domain and the other from a regional model for a sector in the Southern Ocean. We show that the adapted PCM can identify regions with high submesoscale activity, as characterized by the vorticity field (i.e., where surface vertical vorticity ζ is similar to Coriolis frequency f and Rossby number Ro = ζ/f ∼ O(1)), using solely the vertical density profiles, without any additional information on the velocity, the profile location, or horizontal density gradients. The results of this paper show that the adapted PCM can be applied to data sets from different sources and provides a method to study submesoscale eddies using global data sets (e.g., CTD profiles collected from ships, gliders, and Argo floats).

Item Type:	Publication - Article
Digital Object Identifier (DOI):	10.1029/2022EA002618
ISSN:	2333-5084
Additional Keywords:	oceanography, machine learning
Date made live:	22 Jan 2025 15:31 +0 (UTC)
URI:	https://nora.nerc.ac.uk/id/eprint/538787

Actions (login required)

View Item

Document Downloads

More statistics for this item...