Deep learning to infer eddy heat fluxes from sea surface height patterns of mesoscale turbulence

Tom M George,Georgy E Manucharyan,Andrew F Thompson

doi:10.1038/s41467-020-20779-9

Abstract

Mesoscale eddies have strong signatures in sea surface height (SSH) anomalies that are measured globally through satellite altimetry. However, monitoring the transport of heat associated with these eddies and its impact on the global ocean circulation remains difficult as it requires simultaneous observations of upper-ocean velocity fields and interior temperature and density properties. Here we demonstrate that for quasigeostrophic baroclinic turbulence the eddy patterns in SSH snapshots alone contain sufficient information to estimate the eddy heat fluxes. We use simulations of baroclinic turbulence for the supervised learning of a deep Convolutional Neural Network (CNN) to predict up to 64% of eddy heat flux variance. CNNs also significantly outperform other conventional data-driven techniques. Our results suggest that deep CNNs could provide an effective pathway towards an operational monitoring of eddy heat fluxes using satellite altimetry and other remote sensing products.

Highlights

Mesoscale eddies have strong signatures in sea surface height (SSH) anomalies that are measured globally through satellite altimetry
The dynamical variables contain signatures of eddies and filaments that appear to be correlated between the two layers (Fig. 1b); yet, it is the decorrelated components of the two streamfunctions that are associated with the eddy heat flux (Methods)
We find that there exists an optimal Convolutional Neural Network (CNN) complexity for this problem: simpler networks cannot achieve the highest possible skill, while complex networks struggle with overfitting and computational cost (Fig. 3f)

Summary

Introduction

Mesoscale eddies have strong signatures in sea surface height (SSH) anomalies that are measured globally through satellite altimetry. The divergent component of the eddy heat flux is crucial when considering the evolution of the local heat content in the ocean[7] Despite their profound role in ocean circulation, oceanic mesoscale eddies are not fully resolved in long-term climate projection models due to current computational limitations[20,21] so their impact on largerscale circulations and tracer fields must be represented in other ways. More fundamentally, the dynamics of baroclinic instability depend on interactions between the upper and lower layers of the ocean[1] and direct calculations of eddy heat fluxes require simultaneous observations of the near-surface horizontal velocity field and the temperature and velocity fields at eddy scales. Global in situ observations of subsurface properties, e.g. by ARGO floats[26], moorings[6], and ship transects, remain too spatially and/or temporally sparse to resolve the three-dimensional structure of mesoscale eddies, and a direct evaluation of eddy heat fluxes globally is not possible

Objectives

Methods

Results

Conclusion