Coherent heat patterns revealed by unsupervised classification of Argo temperature profiles in the North Atlantic Ocean

A quantitative understanding of the integrated ocean heat content depends on our ability to determine how heat is distributed in the ocean and identify the associated coherent patterns. This study demonstrates how this can be achieved using unsupervised classification of Argo temperature profiles. The classification method used is a Gaussian Mixture Model (GMM) that decomposes the Probability Density Function of a dataset into a weighted sum of Gaussian modes. It is determined that the North Atlantic Argo dataset of temperature profiles contains 8 groups of vertically coherent heat patterns, or classes. Each of the temperature profile classes reveals unique and physically coherent heat distributions along the vertical axis. A key result of this study is that, when mapped in space, each of the 8 classes is found to define an oceanic region, even if no spatial information was used in the model determination. The classification result is independent of the location and time of the ARGO profiles. Two classes show cold anomalies throughout the water column with amplitude decreasing with depth. They are found to be localized in the subpolar gyre and along the poleward flank of the Gulf Stream and North Atlantic Current (NAC). One class has nearly zero anomalies and a large spread throughout the water column. It is found mostly along the NAC. One class has warm anomalies near the surface (50 m) and cold ones below 200 m. It is found in the tropical/equatorial region. The remaining four classes have warm anomalies throughout the water column, one without depth dependance (in the southeastern part of the subtropical gyre), the other three with clear maximums at different depths (100 m, 400 m and 1000 m). These are found along the southern flank of the North Equatorial Current, the western part of the subtropical gyre and over the West European Basin. These results are robust to both the seasonal variability and to method parameters such as the size of the analyzed domain. (C) 2017 Elsevier Ltd. All rights reserved.