Identifying Southern Ocean fronts using unsupervised classification and
edge detection
Abstract
Fronts are ubiquitous in the climate system. In the Southern Ocean,
fronts delineate water masses, which correspond to upwelling and
downwelling branches of the overturning circulation. A robust
understanding of Southern Ocean fronts is key to projecting future
changes in overturning and the associated air-sea partitioning of heat
and carbon. Classically, oceanographers define Southern Ocean fronts as
a small number of continuous linear features that encircle Antarctica.
However, modern observational and theoretical developments are
challenging this traditional framework to accommodate more localized
views of fronts [Chapman et al. 2020]. In this work, we present two
related methods for calculating fronts from oceanographic data. The
first method uses unsupervised classification (specifically, Gaussian
Mixture Modeling or GMM) and an interclass metric to define fronts. This
approach produces a discontinuous, probabilistic view of front location,
emphasising the fact that the boundaries between water masses are not
uniformly sharp across the entire Southern Ocean. The second method uses
Sobel edge detection to highlight rapid changes [Hjelmervik &
Hjelmervik, 2019]. This approach produces a more local view of fronts,
with the advantage that it can highlight the movement of individual
eddy-like features (such as the Agulhas rings). The fronts detected
using the Sobel method are moderately correlated with the magnitude of
the velocity field, which is consistent with the theoretically expected
spatial coincidence of fronts and jets. We will present our python
GitHub repository, which will allow researchers to easily apply these
methods to their own datasets. Figure caption Two methods for
interpretable front detection. Solid lines represent classical fronts.
(a) The “inter-class” metric, which indicates the probability that a
grid cell is a boundary between two classes. The classes are defined by
GMM of principal component values (PCs) derived from both temperature
and salinity. The different colors indicate different class boundaries.
(b) Sobel edge detection: approximately the magnitude of the spatial
gradient of the PCs divided by each field’s standard deviation, which
highlights locations of rapid change.