Two factors ultimately led to the final classification
procedure used in this study. First, there was the
identification of the sPCA components into spatial
(component-1) and temporal (components 2 and 3). These
representations led to a natural division for interpretable
layers. Secondly, the efficiency of utilising classified layers
for assessment or management purposes had to be taken
into account. As a large number of separate layers can be
unwieldy for a conclusive assessment, all possible effort
was made to reduce the number of final layers. In this
respect the combination of relevant SST and CSAT layers
for a single classification was explored. This method would
have resulted in a single layer for both the spatial and
temporal dynamic representations. Although in the spatial
realm this is a feasible option, the temporal dynamic realm
proved more problematic. Important areas of high dynamic
in one dataset were often cancelled by low dynamic values
in the same area in the second dataset. For example, the
Leeuwin Current, which is an important feature in the
distribution of many commercial marine species, has a
strong dynamic signature in SST components. However,
due to is oligotrophic nature it is almost absent in CSAT
dynamics. The combination of these two datasets tends to
reduce the signature of the.Leeuwin, which could lead to an
assumption of reduced importance. On this factor it was
decided to maintain SST and CSAT datasets separately.
The Iterative Self Organising Data Analysis (ISO)
clustering technique was used in this study. This method
iteratively classifies an image, recalculates statistics and
reclassifies again. Cells are assigned to a specific class
Each existing regional polygon was given a unique
identifier and the vector layer was converted back to a
raster image with the same extent and resolution as the
original. The final classified rasters were able to have
statistics extracted per the individual regions, allowing
distinct spatial regions to be assessed independently or in
conjunction with surrounding regions.
In the final result of the classification process, relevant
oceanographic regions were isolated. The Leeuwin Current
in the west was separated into regions, reflecting it
weakening signal as the current moves eastward. The
northern Bass Straight showed the opposing influences
from both western and eastern sources. Upwelling regions
in the Great Australian Bight, and the Bonney Coast leading
into Encounter Bay, were maintained as well as major eddy
regions such as those South of the Leeuwin and in the
central GAB. The classification results for southern
Australia are shown in Figure 3 and the results for the west
Australian coast are shown in Figure 4.
define it is a more important factor. In other words, a
uniform degree of high variability within a region is not a
problem, as long as the region's bounds delimit the region
of variability relatively accurately.
using minimum spectral distance to class means. The initial
first pass is based on arbitrary spectral means. The process
is repeated until little change in classes occurs or, either a
minimum spectral distance or a maximum number of
iterations is reached. The raw classified images were
imported into the ArcGIS environment and dendrograms
were created. Using the grouping hierarchies in the relevant
dendrograms and knowledge of key oceanographic systems,
similar class groupings were assessed on repetitiveness and
merged. One of the important knowledge-based
assumptions in the classification process is the
identification of spatially significant regions. Maintaining
these often large areas while removing redundant regions
can be as simple as eliminating regions under a certain area
threshold. In this case these processes took several steps
and were conducted in the vector environment after
converting the classified raster images. Vector features
were eliminated by being merged into a larger bordering
feature which shared the longest border. Initially a hard
threshold required all features with areas smaller than 0.5
decimal degree squared (dD2) to be eliminated. Features up
to 0.1dD2 were assessed and despite a few exceptions all
were eliminated. Notable were relatively small regions
located in the open ocean regions. As the open ocean is
subject to a high degree of spatio-temporal influences, some
of the regions fail to spatially define a significantly
common or periodic occurrence. Open ocean regions with
depths greater than 500m were subjected to a larger area
threshold of 1 dD2, although care was taken not to
eliminate features attributed to islands and well known
anomalies.
CSAT Spatial SST Spatial
CSAT Seasonal SST Seasonal
Figure 3: Southern Australian Boundaries
The verification of these results proved a problem since
conventional accuracy assessment techniques did not apply.
A pixel's relative fit to a region's character is not in itself
an accurate measure of the suitability of regional
boundaries. While interpreting error magnitude in the case
of this project, it should be remembered that a region that
evenly contains a high degree of internal heterogeneity is
just as viable as a region that is heterogeneous. The degree
to which the region bounds
pix
de
wit
reg
ran