Geology Reference
In-Depth Information
Clausi and Deng
[2003] concluded that the combination
of image attributes (tone + texture) coupled with the
K
‐means algorithm and MRF labeling algorithm was the
most appropriate combination to produce accurate seg-
mentation. The
K
‐means method is used as an initial
guess and the MRF labeling algorithm is used for refine-
ment purposes. Although the conclusion is certainly jus-
tified based on the shown image, it should be noted
that the original SAR image (Figure 10.3a) shows only
leveled ice. The recommended approach may not be
suitable in the case of deformed or rough ice surfaces or
other surface morphologies covered. It cannot accurately
classify surfaces covered with frost flowers or metamor-
phosed snow.
An algorithm that combines image segmentation and
classification was developed by
Yu and Clausi
[2007] to
classify major operational ice types in SAR imagery
provided by the CIS. The algorithm is called iterative
region growing using semantics (IRGS). A low‐level
unsupervised region growing segmentation technique is
initialized. This is followed by a high‐level supervised
classification applied to the segmented regions to refine
the segmentation and produce semantic class labels. The
algorithm proceeds in an iterative manner. For the seg-
mentation process, gray tone and texture are usually
used. To conduct the higher level classification, attrib-
utes other than tone and texture need to be incorpo-
rated. The purpose is to mimic the high‐level knowledge
used by ice analysts at CIS to identify ice types during
their visual analysis of SAR images. Two shape attrib-
utes were used. The first is a geometric measure to
quantify elongated shapes. This is used to identify leads
within the surrounding ice cover. The second is a meas-
ure of how close a given segmented region is to the
elliptical shape. This is used to identify MY ice floes.
Successful use of the shape‐based analysis requires a
fairly accurate output from the low‐level segmentation;
i.e. the image segmentation should neither be over‐ nor
undersegmented.
Yu and Clausi
[2007] integrated the
two processes of the segmentation and clustering under
the Bayesian framework. Both aim at reducing a defined
energy. The interactions allow the classification result to
have some degree of control on the region growing seg-
mentation process. Figure 10.4 is an example of classify-
ing two stable winter ice types (thick FY ice and MY ice)
in a Radarsat image of an area in the Baffin Bay. A few
MY ice floes are identified by the algorithm.
The IRGS method has been used in a recent study to
classify (or rather reclassify) ice types in the analyzed
Radarsat images produced in the CIS [
Ochilov and Clausi
,
2012]. The CIS analysis delineates polygons of uniform
ice attributes in terms of ice types and concentrations
(section 11.2). The IRGS is used to segment each polygon
in the CIS's image analysis product into regions of uni-
form ice type and then label the regions. The labeling tech-
nique uniquely models the spatial relationship of regions
between the polygons in the form of a neighborhood sys-
tem embedded in a Markov random field framework. The
criterion of the maximum a posteriori probability (MAP),
derived from Bayesian theory, is used. Adjacent regions
that have the same assigned class are merged until the sys-
tem energy cannot be decreased further. These two steps
(class assignment and merging) are iterated until no more
merging can be performed. The technique offers ice type
(a)
(b)
Thick FY ice >120 cm
Second Year ice
Figure 10.4
Radarsat‐1 SAR (a) image over the Baffin Bay acquired on 7 February 1998 and (b) the results from
the IRGS method of segmentation and classification. Two ice classes are shown, and a few second‐year ice floes
are identified by their marked boundaries [
Yu and Clausi
, 2007, Figure 7, with permission from IEEE].
