Database Reference
In-Depth Information
Second, although eddies generally have an ellipse-like shape, the shape's
manifestation in gridded SSH data differs based on latitude. This is
because of the stretch deformation of projecting spherical coordinates
into a two-dimensional plane. As a result, one cannot restrict eddies by
shape (
e.g.
circle, ellipse,
etc.
) Finally, eddy heights and sizes vary by
latitude, which makes having a global “acceptable” eddy size unfeasible
[26]. Therefore, applying a single global threshold would wipe out many
relevant patterns in the presence of spatial heterogeneity.
h = 0
h = h
0
h = -100 cm
Figure 15.5.
Schematic of an anti-cyclonic eddy that is embedded in a large scale
background with a larger amplitude than the eddy. If we were to apply a threshold
at
h
= 0 the eddy would be missed. This is motivation to use multiple threshold from
h
=
−
100
cm
to 100
cm
as suggested by CH11. Figure adapted from [12]
Two major approaches have been used to monitor eddies globally: the
first is spatial and the other temporal. In the spatial approach, eddies
are identified as close-contoured positive or negative anomalies using
classical connected component algorithms. These connected component
algorithms tend to be highly parameterized to encode expert knowl-
edge such as minimal eddy radius or amplitude to reduce the number of
false positives. The state-of-the-art connected component approach was,
however, computationally prohibitive and unable to separate eddies that
were in close proximity. Using the insights provided by [26, 14], a recent
study was able to address both shortcomings by applying a latitude de-
pendent convexity criterion that is both ecient and accurate compared
to the state-of-the-art methods. In the temporal domain, Faghmous
et al. (2012) leveraged the spatio-temporal signature of ocean eddies
on SSH to develop an unsupervised learning algorithm that identified
groups of pixels that exhibited an eddy-like signature of slow decreases
or increases in SSH over significant time periods. The temporal ap-
proach is significantly more ecient and robust than existing spatial
methods alone - its computational complexity is linear in the resolution
of the data compared to quadratic for the spatial approach, and given
that only groups of pixels exhibiting similar eddy behavior are labeled
as eddies, the temporal approach is more robust to outliers than the spa-
tial one. This work, highlights the need of novel spatio-temporal data
