Effective SAR Image Segmentation and Sea-Ice Floe Distribution
Analysis via Kernel Graph Cuts based Feature Extraction and Fusion
Soumitra Sakhalkar
1
, Jinchang Ren
1
, Phil Hwang
2
and Paul Murray
1
1
Centre for Excellence in Signal & Image Processing (CeSIP), Department of Electronic and Electrical Engineering,
University of Strathclyde, Glasgow, G1 1XQ, U.K.
2
Scottish Association for Marine Science (SAMS), Oban, PA37 1QA, U.K.
1 RESEARCH PROBLEM
The Sea Ice that grows in the open seas like the
Arctic sea, forms varying shapes and size due to the
fracturing as well as thickening caused by the strong
gale force winds and sea waves. Over the winter
season, due to the cooler temperature, these sea-ice
regions combine with each other to make a stronger
and larger sea ice block. In the summer however,
due to the higher temperature, they separate into
smaller and weaker floes as shown in Figure 1.
Figure 1: An area of sea ice region captured during the
beginning of the month of the early summer period on the
left and during the ending in the month on the right.
Sea-Ice monitoring has gained significant interest in
recent years, largely due to the fact of the decreasing
area and thickness of the older arctic sea ice (Kwok,
et al., 2009) (Stroeve, et al., 2008). This decline in
older sea ice has been linked largely to the growth of
younger, thinner sea ice regions (Maslanik, et al.,
2007) and also climate changes (Holloway & Sou,
2002), caused by greenhouse gases (Serreze, et al.,
2007).
The study of Polar Regions using Synthetic
Aperture Radar [SAR] has been widely used for
identification of sea ice floes, their size and their
distribution (Burns, et al., 1987) (Rothrock &
Thorndike, 1984), (Soh, et al., 2004), (Soh &
Tsatsoulis, 1998). This is because SAR is not
majorly affected by the harsh weather conditions or
the illumination variations and it is able to cover
large and primarly inaccesible areas (Xu, et al.,
2014). This is particularly important for ensuring
safe marine navigation as well as supporting studies
of climate changes, like ours, of the Polar Regions.
To date, the process of developing an automatic
algorithm for effective segmentation of SAR Sea-Ice
images has not been achievable. As a result, analysis
of sea ice images relies on a time consuming expert
analysis which is performed manually. For this
reason, it is primarily important to develop
techniques to automatically segment the sea-ice
regions from the background and subsequently
extract these sea-ice regions from the SAR image.
When this is completed it will become possible to
build a Floe Size Distribution (FSD) database, where
FSD is a measure of the distribution of the different
size of the sea ice floes. An FSD database will be
constructed in our project by extracting and storing
the total pixel area of these individual sea-ice
regions in the SAR image and grouping them
according to their size. The result will then be used
to generate a graph of the size distribution of the
floes on different days in a year of the Arctic region.
The outcome of our study will further develop
scientist’s understanding of the different trends as
well as the various conditions affecting the size of
the Arctic sea ice floes for that particular year.
Eventually this will improve our understanding of
the changes in the sea-ice extent over the year by
means of comparison with the past several years’
results.
The key research problem addressed by this
work lies in developing a new novel image
segmentation technique which is simple, fast and
robust when used to segment the SAR sea ice
images.
28
Sakhalkar S., Ren J., Hwang P. and Murray P..
Effective SAR Image Segmentation and Sea-Ice Floe Distribution Anlysis via Kernel Graph Cuts based Feature Extraction and Fusion.
Copyright
c
2015 SCITEPRESS (Science and Technology Publications, Lda.)
2 OUTLINE OF OBJECTIVES
The main aim of this study is to create novel
techniques to automatically segment and extract the
sea ice floes from the SAR images of the area in the
Arctic region being monitored. To achieve this it is
important to fulfil the following smaller objectives;
1. To develop an optimal segmentation technique
for accurately segmenting individual ice floes
from the background as well as from each
other.
2. To refine the methods developed in 1 so they
are efficient and inexpensive to compute.
3. To remove/reduce the speckle noise present in
almost every SAR image using appropriate
filters. That is, using filters which retain the
original image characteristics as well as
removing/reducing the presence of speckle
noise.
4. To make the techniques developed completely
automatic and dynamic so that they can
process any SAR image to segment the ice
floes.
The result of completing these objectives will be
a robust and efficient global machine learning
algorithm for segmentation. Furthermore the
algorithm parameters will be computed
automatically depending on the image statistics
derived from the current image under study. This
will introduce a step change in the way the sea ice
floes are analysed.
3 STATE OF THE ART
Due to the decrease in sea ice floe extent and
thickness, it has become increasingly important to
generate a better understanding of the environmental
as well as the social impacts on the transformation of
Arctic sea ice over the duration of each day in a
year.
The need for automated segmentation techniques
for ice floe analysis has led to some popular floe
related studies: examples include techniques for
classification using dynamic thresholding for
separation and heuristic geophysical knowledge
(Haverkamp, et al., 1995), deformation of sea ice by
means of measuring the opening and closing of leads
(Fily & Rothrock, 1990) and floe size identifications
and measurement (Korsnes, 1993). More recent
studies have tried to achieve this by using the two
texture analysis methods, Markov Random Fields
(MRF) and Gray-Level Co-occurrence Probabilities
(GLCP) (Clausi & Yue, 2004), or by using
stochastic ensemble consensus approach (Wong, et
al., 2009), utilisation of Bayesian segmentation
approach with MRF model (Deng & Clausi, 2005)
and using pulse-coupled neural networks (PCNN)
(Karvonen, 2004).
Although many of these approaches have been
implemented, tested and provided in the literature,
they do not meet the necessary criteria for
application to our data. This is largely due to the fact
that the existing techniques are either data specific
and can only be applied to segment sea ice floes in
certain parts of the world or they involve
manipulation of the data or in other cases take too
long to compute. As a result, the currently used
approach relies on the lengthy technique of manually
segmenting the regions based on expertise and
knowledge of a sea-ice expert.
There have also been some automated
approaches for Sea-Ice segmentation; using an
intelligent system named Advanced Reasoning using
Knowledge for Typing Of Sea ice (ARKTOS).
ARKTOS automatically segments the regions and
generate the descriptors for the segmentation of sea
ice floes. Others use expert rules for classification
(Soh, et al., 2004) using the analytical tool:
Automated sea ice segmentation (ASIS). ASIS uses
local thresholding for obtaining and retaining
information in the image. Image quantization has
also been used to obtain the different classes within
an image and computing spatial attributes of each
class using Aggregated Population Equalization
(APE) concept (Soh & Tsatsoulis, 1998).
Nonetheless all these have been deemed
unsatisfactory to be used for our study due to the
same reasons indicated before.
Recent developments in image segmentation
techniques have led to a new approach called Graph
Cuts [GC] with kernel mapping which is based on
the work of (Salah, et al., 2011). GC is based on
energy minimization for effectively finding a cut
{segmentation} between regions (Boykov, et al.,
2001). GC have been used in many studies for image
segmentation, for instance in (Rother, et al., 2004),
an iterative and interactive GC based approach has
been used for effective foreground extraction,
whereas in (Boykov & Funka-Lea, 2006) &
(Boykov, et al., 2001) GC is used for segmenting the
regions using image histogram analysis.
In fact many other approaches based on energy
minimization, similar to GC have also been used for
image segmentation; Region Competition (Zhu &
Yuille, 1996), Active Contours (Caselles, et al.,
1997) (Chan & Vese, 2001) and Level Set using
EffectiveSARImageSegmentationandSea-IceFloeDistributionAnlysisviaKernelGraphCutsbasedFeatureExtraction
andFusion
29
Figure 2: Methodology of our study.
Mumford and Shah model (Vese & Chan, 2002).
Also in many studies like (Muller, et al., 2001),
(Scholkopf, et al., 1999), (Dhillon, et al., 2007),
(Schölkopf, et al., 1998), (Girolami, 2002) and
(Zhang & Chen, 2002), the so called kernel trick has
been used for effective and efficient clustering of
complex data.
Similarly many studies have used Kernel mapped
Graph cuts (KGC) for image segmentation,
examples include spine image fusion to replace CT
and MR (Miles, et al., 2013), for classification of
brain images (Harini & Chandrasekar, 2012), for
segmentation of abdomen MR images (Luo, et al.,
2013), for segmentation for MR images with
intensity inhomogeneity correction (Luo, et al.,
2013).
Thus it can be seen how Kernel mapped Graph
Cuts, have until now, been mostly used for medical
image processing as opposed to SAR sea ice image
processing like our study. In fact, to the author’s
knowledge KGC based techniques have not yet been
used to segment images containing ice floes. It is
therefore proposed that a technique using GC will be
applied to automatically segment ice floes in SAR
imagery. To deal with the speckle noise, we also
propose the addition of an effective pre-processing
and image filtering stage which will lead to an
optimal segmentation result.
4 METHODOLOGY
The entire processing procedure for our study is
shown by the flow chart given in the Figure 2. As
seen in Figure 2, the methodology for our study is
split in two major stages;
1. This will involve segmentation of sea ice floe
using the existing KGC algorithm. It will also
incorporate our proposed contribution of the
addition of a pre-processing stage to improve
the results and a technique to allow the
optimization of the parameters which will
make the algorithm adapt automatically for
processing the image under study.
2. On completion of Stage 1, the next step will
involve the extraction of individual sea ice
floes to build the FSD analysis for each image
for each day in that particular year of study.
We will now briefly explain the existing
algorithm; KGC and then explain the importance of
our contribution to the algorithm for the
improvement of results.
4.1 The Kernel Graph Cuts
Our implementation of the existing KGC algorithm
is based on the implementation by Salah, et. al.
(Salah, et al., 2011), who initially proposed this
technique. This algorithm is based on a three stage
processing procedure;
1. K-means clustering to find the initial clusters
and their centroids.
2. Kernel mapping of the image into higher
dimensional feature space.
3. Image Segmentation achieved using Graph
Cuts.
SENSORNETS2015-DoctoralConsortium
30
4.1.1 K-means based Clustering
K-means is a popular un-supervised and easy to
implement clustering method, first introduced by
Lloyd (Lloyd, 1982). In fact a survey of clustering
algorithm some years ago showed how K-means,
even after 25 years was still the most widely used
clustering algorithm (Berkhin, 2002) at that time.
The algorithm partitions/clusters a given set of
data into k clusters depending on the least squared
distance of each point from that cluster’s centroid.
K-means is an iterative process, which continually
estimates the least squared distance from the cluster
centroid and re-assigns the data into these ‘k’
clusters until the process is stabilized. An example
of a set of data points clustered into 3 clusters
represented by the three different colours for each
one, is illustrated in Figure 3.
Figure 3: K-means based clustering example.
4.1.2 Kernel Mapping
Kernel mapping or the kernel trick is a popular
technique used in many recent image segmentation
algorithms (Muller, et al., 2001), (Scholkopf, et al.,
1999), (Dhillon, et al., 2007), (Schölkopf, et al.,
1998), (Girolami, 2002) , where a kernel function is
used to map a data set into a higher dimensional
feature space, so that partition of regions is possible.
Figure 4, first seen in (Salah, et al., 2011),
illustrates how the kernel mapping aids a better and
faster separation /segmentation of result, due to the
implicit mapping of data set into higher dimensional
space, so that the GC algorithm can be applied.
There are many popular kernel functions
commonly used in the field of digital image
processing; examples include the Gaussian/ radial
basis function kernel, polynomial kernel, sigmoid
kernel and many more as mentioned in (Genton,
2002).
Figure 4: Illustration of a non-linear data separation. Data
separation is non-linear in data space. The data is mapped
into a higher dimensional feature space using a kernel. The
separation is now linear in feature space, separated by a
hyper plane.
For our study, we will use the Gaussian/ radial basis
function (RBF) (Buhmann, 2003) kernel, due to its
simplicity and ease of implementation, for mapping
into higher dimensional feature space. The equation
for the RBF kernel is given by,
,

–

(1)
4.1.3 Graph Cut based Image Segmentation
The implementation of the algorithm of KGC as first
proposed by Salah (Salah, et al., 2011), is based on
the GC technique first introduced in (Boykov, et al.,
2001), which implements the energy minimization
based on the minimization of the two energy terms
given by,
E
f
E

E

(2)
Here, E
Smooth
is the Smoothness cost which measures
the extent to which a label f is no longer piecewise
constant and E
Data
is the Data cost which measures
the disagreement of the current labelling f with the
observed data. The labelling f mentioned here is
assigned by the K-means clustering algorithm in the
initial stage of the KGC algorithm. A labelling f is
said to be piecewise constant if it varies smoothly on
the surface of the object but changes dramatically at
the object boundaries.
Figure 7 shows the step-by-step procedure of
how a GC algorithm finds a cut {segmentation}.
Figure 7(a) shows the initial labelling for the pixels
before loading the graph for the GC algorithm; blue
EffectiveSARImageSegmentationandSea-IceFloeDistributionAnlysisviaKernelGraphCutsbasedFeatureExtraction
andFusion
31
pixels (p - s) belong to label A and red pixels (t - x)
belong to label B. In Figure 7(b) the GC algorithm
then assigns the labelling and weights of each pixel
with each of the labels A and B using the Data cost
term. The darker arrows denote the more likelihood
of a label being assigned to a pixel, while the dotted
arrow denotes low likelihood of a label being
assigned to a pixel. It can be seen how the pixel u is
now indicated to be more likely to present label A. In
Figure 7(c) the weights of each pixel with its
neighbouring pixel are then assigned using the
Smoothness cost term. Similar to the previous
section, the darker lines denote the more likelihood
and the weaker lines show less likelihood of a pixel
being associated to be similar to each other. In
Figure 7(d) the GC algorithm finds a cut between the
labelling one neighbouring pixel pair at a time. It
can be seen how the GC now assigns pixel u to label
A based on the most likelihood (Data Cost) and
similarity measure (Smoothness Cost).
4.1.4 Drawbacks
Although the KGC algorithm has various advantages
over the conventional image segmentation
techniques, it still has limited number of drawbacks
which need to be addressed to make the image
segmentation more robust and efficient.
Figure 5 shows an area where the KGC works
really well and produces really good results and in
Figure 6, it can be seen how the KGC produces very
poor results due to the heavy presence of speckle
noise in that region of the original SAR image.
Figure 5: One example result of the KGC segmentation.
4.2 Pre Processing
For our study, we propose the addition of a two
stage pre-processing routine to help improve the
Figure 6: Other example result of the KGC segmentation.
performance of the existing KGC based
segmentation algorithm. This pre-processing step
involves filtering the image using adaptive filters to
remove the speckle noise before applying
morphological processing. The aim is to improve the
classification result from the K-means and
subsequently the segmentation result from the KGC
algorithm by first applying these pre-processing
techniques.
4.2.1 Adaptive Filtering
Most SAR images contain the multiplicative noise
known as speckle. This presence of speckle noise
reduces the detection of targets or patterns present in
the SAR sea ice images (Sheng & Xia, 1996).
Hence for this purpose we have used the
Adaptive Median (AM) filter (Qiu, et al., 2004),
which uses the local statistics within a filter window
to mark a pixel as speckle noise and remove/reduce
this speckle noise present in the image. We have
compared our results with other popular speckle
filtering techniques like the Lee filter (Lee, 1980)
(Lee, 1981), Frost filter (Frost, et al., 1982) (Frost, et
al., 1981), Bilateral filter (Tomasi, 1998), Median
filter & Wiener filter (Lim, 1990), Local Sigma filter
(Eliason & McEwen, 1990) and found that the AM
filter to be the most suitable for our study. For the
scope of this paper, we have not added the
comparison results.
4.2.2 Morphological Processing
Other popular techniques, widely used for pre and
post processing are the morphological filters such as
dilation and erosion (Matheron, 1975) (Serra, 1982)
(Dougherty & Lotufo, 2003). In terms of image
processing, Dilation, enlarges the image features
SENSORNETS2015-DoctoralConsortium
32
Figure 7: Step-by-Step procedure of GC algorithm finding a cut using min cut-max flow algorithm.
Figure 8: Morphological processing on the KGC segmentation result.
based upon the size and shape of the structuring
element chosen. Erosion shrinks the image features
and can also remove them based upon the size and
shape of the structuring element chosen. The
morphological processing is extensively explained
extensively in (Gonzalez, et al., 2004). For the
purpose of our study we have used morphological
closing, which is a dilation followed by an erosion
using the same structuring element.
For our study, we have used morphological
processing to overcome the aforementioned
drawbacks of the KGC algorithm and produce better
K-means clustering results.
For this purpose, we build a mask image, by
thresholding the original grayscale image using
standard deviation. We use standard deviation rather
than grey level threshold for thresholding to make
the process more adaptive to the current image under
study.
This is then followed by our addition of
EffectiveSARImageSegmentationandSea-IceFloeDistributionAnlysisviaKernelGraphCutsbasedFeatureExtraction
andFusion
33
morphological processing. We then multiply this
mask image with the original grayscale image to
produce a morphologically enhanced image as seen
in Figure 8.
We have used the disk shaped structuring
element for our work, since the shape almost
represents the shape of an ideal sea ice floe. Figure
8, shows an example result achieved using the disk
structuring element.
5 EXPECTED OUTCOME
It is anticipated that a major outcome of this study
will be a novel, fast and reliable algorithm for sea
ice floe image segmentation. The algorithm will be
easy to use so that environmental experts are able to
replace the current manual analysis with this
sophisticated and robust technique.
Beyond this, the outcome of this work will help
us develop our understanding of the environmental
as well as social factors affecting the Arctic Sea ice
floe cover. For this, a detailed analysis of sea ice
floe needs to be done, with the implementation of
the FSD analysis. This will be done to monitor the
sea ice floe extent on each day of subsequent year.
This data will then be compared with the results
of the FSD analysis done in the same area in the
Arctic Region from the previous years and will then
help us validate the theory that the older Arctic sea
ice floes are indeed getting reduced and replaced by
younger, weaker sea ice floes.
Beyond this it is also anticipated that this study
can be applied to a wide range of applications. Some
examples would include segmenting medical images
of a biological nature or microscopic images of
metals and other similar materials. Our addition of
the adaptive filtering process can also be used for
other noise removal applications.
6 STAGE OF THE RESEARCH
6.1 Results
We now present our results by means of a
comparison between the results produced with the
original KGC algorithm and the results produced as
a result of our addition of the proposed pre-
processing stage. To validate the efficacy of the
proposed approach, real SAR images with a high
resolution of 16k by 16k have been used for both
visual assessment and quantitative analysis.
The algorithm is coded in Matlab running on a
Dell Inspiron 5537 laptop with 2.3 GHz processor, 4
GB RAM and 64 bit Windows 8.1 operating system.
It requires approximately 41 minutes for obtaining
Figure 9: Comparison Result on Image 1.
SENSORNETS2015-DoctoralConsortium
34
Figure 10: Comparison Result on Image 2.
the segmentation result for the entire 16k by 16k
original SAR image. The processing speed can be
further optimised and reduced to run even faster, but
it is currently not the main area of focus of our
study.
In Figures 9 and 10, the images on the left are the
original real SAR sea ice images, the images in the
middle are the Original KGC segmented results and
the images on the right are the results produced after
our addition of a pre-processing stage. It can be seen
how our addition improves the segmentation result
of the five sample sub sections of the main 16k by
16k real SAR Sea-Ice image, visually.
6.2 Progress to Date
We have now met the first 3 objectives of our study.
These involved selecting a fast, accurate and
adaptive algorithm for segmenting the sea ice floes
and refining the algorithm to make it more efficient.
These can be verified from the descriptions of
the KGC algorithm in the previous sections and as
evidently seen in Figures 9 & 10. Figures 9 & 10
portray how our proposed pre-processing stage,
removes/reduces the speckle noise present in the
SAR images and validates how our addition
improves the KGC segmented results for the SAR
sea ice images.
6.3 Future Work
We now need to focus on extracting the floe size
information of these segmented sea ice floes. In
order to achieve this, it is first necessary to further
separate the floe regions which are not currently
separated in some regions but which can be visually
predicted to be separated. For achieving this, we are
currently implementing another popular energy
based image segmentation algorithm; Active
Contours (Caselles, et al., 1997) or also referred to
as “snakes”. Active Contours currently being used
for our study are based on the algorithm developed
by Chan & Vese (Chan & Vese, 2001).
The active contour is although known to be
notoriously slow due to the large number of
iterations required to achieve a good segmentation.
To reduce this processing time, we are currently also
building an adaptive algorithm to only extract the
regions where separations of sea ice floe need to be
implemented as per our visual perception. We have
been able to achieve some minor improvements but
more work needs to be done in order to achieve the
optimal results for the extraction of these sea ice floe
regions.
EffectiveSARImageSegmentationandSea-IceFloeDistributionAnlysisviaKernelGraphCutsbasedFeatureExtraction
andFusion
35
ACKNOWLEDGEMENTS
We would like to thank Scottish Association for
Marine Science (SAMS) and NERC for providing us
with such a challenging and interesting topic for our
study and for their funding support "NE/L012707/1"
and "NE/M00600x/1" to make this study possible.
We would also like to thank the University of
Strathclyde for their motivation and support for
conducting this study.
REFERENCES
Berkhin, P., 2002. A survey of clustering data mining
techniques, San Jose, CA: Accrue Software.
Boykov, Y. & Funka-Lea, G., 2006. Graph cuts and
efficient ND image segmentation. International
journal of computer vision, 70(2), pp. 109-131.
Boykov, Y., Veksler, O. & Zabih, R., 2001. Fast
approximate energy minimization via graph cuts.
IEEE Transactions on Pattern Analysis and Machine
Intelligence, 23(11), pp. 1222-1239.
Buhmann, M. D., 2003. Radial basis functions: theory and
implementations. Cambridge: Cambridge university
press.
Burns, B. A. et al., 1987. Multisensor comparison of ice
concentration estimates in the marginal ice zone.
Journal of Geophysical Research: Oceans (1978–
2012), 92(C7), pp. 6843-6856.
Caselles, V., Kimmel, R. & Sapiro, G., 1997. Geodesic
active contours. International journal of computer
vision, 22(1), pp. 61-79.
Chan, T. F. & Vese, L. A., 2001. Active contours without
edges. IEEE transactions on Image processing, 10(2),
pp. 266-277.
Clausi, D. A. & Yue, B., 2004. Comparing Cooccurrence
Probabilities and Markov Random Fields for Texture
Analysis of SAR Sea Ice Imagery. IEEE Transactions
on Geoscience and Remote Sensing, 42(1), pp. 215-
228.
Deng, H. & Clausi, D. A., 2005. Unsupervised
segmentation of synthetic aperture radar sea ice
imagery using a novel Markov random field model.
IEEE Transactions on Geoscience and Remote
Sensing, 43(3), pp. 528-538.
Dhillon, I. S., Guan, Y. & Kulis, B., 2007. Weighted
Graph Cuts without Eigenvectors:A Multilevel
Approach. IEEE Transactions on Pattern Analysis and
Machine Intelligence, 29(11), pp. 1944-1957.
Dougherty, E. R. & Lotufo, R. A., 2003. Hands-on
morphological image processing. Bellingham: SPIE
press.
Eliason, E. M. & McEwen, A. S., 1990. Adaptive box
filters for removal of random noise from digital
images. Photogrammetric Engineering and Remote
Sensing, Volume 56, pp. 453-458.
Fily, M. & Rothrock, D. A., 1990. Opening and closing of
sea ice leads: Digital measurements from synthetic
aperture radar. Journal of Geophysical Research:
Oceans (1978–2012), 95(C1), pp. 789-796.
Frost, V. S. et al., 1981. An adaptive filter for smoothing
noisy radar images. s.l., IEEE, pp. 133-135.
Frost, V. S., Stiles, J. A., Shanmugan, K. S. & Holtzman,
J., 1982. A model for radar images and its application
to adaptive digital filtering of multiplicative noise.
IEEE Transactions on Pattern Analysis and Machine
Intelligence, Volume 2, pp. 157-166.
Genton, M. G., 2002. Classes of kernels for machine
learning: a statistics perspective. The Journal of
Machine Learning Research, Volume 2, pp. 299-312.
Girolami, M., 2002. Mercer kernel-based clustering in
feature space. IEEE Transactions on Neural Networks,
13(3), pp. 780-784.
Gonzalez, R. C., Woods, R. E. & Eddins, S. L., 2004.
Digital image processing using MATLAB. Upper
Saddle River, NJ Jensen: Prentice Hall.
Harini, R. & Chandrasekar, C., 2012. Efficient Pattern
Matching Algorithm For Classified Brain Image.
International Journal of Computer Applications
(0975–8887), 57(4), pp. 5-10.
Haverkamp, D., Soh, L. K. & Tsatsoulis, C., 1995. A
comprehensive, automated approach to determining
sea ice thickness from SAR data. IEEE Transactions
on Geoscience and Remote Sensing , 33(1), pp. 46-57.
Holloway, G. & Sou, T., 2002. Has Arctic sea ice rapidly
thinned?. Journal of Climate, 15(13), pp. 1691-1701.
Karvonen, J. A., 2004. Baltic sea ice SAR segmentation
and classification using modified pulse-coupled neural
networks. IEEE Transactions on Geoscience and
Remote Sensing, 42(7), pp. 1566-1574.
Korsnes, R., 1993. Quantitative analysis of sea ice remote
sensing imagery. International Journal of Remote
Sensing, 14(2), pp. 295-311.
Kwok, R. et al., 2009. Thinning and volume loss of the
Arctic Ocean sea ice cover: 2003-2008. Journal of
Geophysical Research: Oceans (1978–2012),
114(C7).
Lee, J. S., 1980. IEEE Transactions on Digital Image
Enhancement and Noise Filtering by Use of Local
Statistics. Pattern Analysis and Machine Intelligence,
Volume 2, pp. 165-168.
Lee, J. S., 1981. Speckle analysis and smoothing of
synthetic aperture radar images. Computer graphics
and image processing, 17(1), pp. 24-32.
Lim, J. S., 1990. Two-dimensional signal and image
processing. 1 ed. Englewood Cliffs, NJ: Prentice Hall.
Lloyd, S., 1982. Least squares quantization in PCM. IEEE
Transactions on Information Theory, 28(2), pp. 129-
137.
Luo, Q., Qin, W. J. & Gu, J., 2013. Kernel Graph Cuts
Segmentation for MR Images with Intensity
Inhomogeneity Correction. Applied Mechanics and
Materials, Volume 333-335, pp. 938-943.
Luo, Q. et al., 2013. Segmentation of abdomen MR
images using kernel graph cuts with shape priors.
Biomedical engineering online, 12(124).
SENSORNETS2015-DoctoralConsortium
36
Maslanik, J. A. et al., 2007. A younger, thinner Arctic ice
cover: Increased potential for rapid, extensive seaice
loss. Geophysical Research Letters , 34(24).
Matheron, G., 1975. Random sets and integral geometry.
New York: Wiley.
Miles, B. et al., 2013. Spine image fusion via graph cuts.
IEEE Transactions on Biomedical Engineering, 60(7),
pp. 1841-1850.
Muller, K. et al., 2001. An introduction to kernel-based
learning algorithms. IEEE Transactions on Neural
Networks, 12(2), pp. 181-201.
Qiu, F. et al., 2004. Speckle noise reduction in SAR
imagery using a local adaptive median filter.
GIScience & Remote Sensing, 41(3), pp. 244-266.
Rother, C., Kolmogorov, V. & Blake, A., 2004. Grabcut:
Interactive foreground extraction using iterated graph
cuts. ACM Transactions on Graphics (TOG), August,
23(3), pp. 309-314.
Rothrock, D. A. & Thorndike, A. S., 1984. Measuring the
sea ice floe size distribution. Journal of Geophysical
Research: Oceans (1978–2012), 89(C4), pp. 6477-
6486.
Salah, M. B., Mitiche, A. & Ayed, I. B., 2011. Multiregion
image segmentation by parametric kernel graph cuts.
IEEE Transactions on Image Processing, 20(2), pp.
545-557.
Scholkopf, B. et al., 1999. Input space versus feature space
in kernel-based methods. Neural Networks. IEEE
Transactions on Neural Networks, 10(5), pp. 1000-
1017.
Schölkopf, B., Smola, A. & Müller, K. R., 1998.
Nonlinear component analysis as a kernel eigenvalue
problem. Neural computation, 10(5), pp. 1299-1319.
Serra, J., 1982. Image analysis and mathematical
morphology. s.l.:Academic press.
Serreze, M. C., Holland, M. M. & Stroeve, J., 2007.
Perspectives on the Arctic's Shrinking Sea-Ice Cover.
Science, 315(5818), pp. 1533-1536.
Sheng, Y. & Xia, Z. G., 1996. A comprehensive
evaluation of filters for radar speckle suppression.
Geoscience and Remote Sensing Symposium, 1996.
IGARSS '96. 'Remote Sensing for a Sustainable
Future.', International, May, Volume 3, pp. 1559-
1561.
Soh, L. K. & Tsatsoulis, C., 1998. Automated sea ice
segmentation (ASIS). s.l., IEEE, pp. 586-588.
Soh, L. K., Tsatsoulis, C., Gineris, D. & Bertoia, C., 2004.
ARKTOS: An intelligent system for SAR sea ice
image classification. IEEE Transactions on
Geoscience and Remote Sensing, 42(1), pp. 229-248.
Stroeve, J. et al., 2008. Arctic sea ice extent plummets in
2007. Transactions American Geophysical Union,
89(2), pp. 13-14.
Tomasi, C. &. M. R., 1998. Bilateral filtering for gray and
color images. s.l., IEEE, pp. 839-846.
Vese, L. A. & Chan, T. F., 2002. A multiphase level set
framework for image segmentation using the Mumford
and Shah model. International journal of computer
vision, 50(3), pp. 271-293.
Wong, A., Clausi, D. A. & Fieguth, P., 2009. SEC:
Stochastic ensemble consensus approach to
unsupervised SAR sea-ice segmentation. s.l., IEEE, pp.
299-305.
Xu, L., Li, J., Wong, A. & Wang, C., 2014. A KPCA
texture feature model for efficient segmentation of
RADARSAT-2 SAR sea ice imagery. International
Journal of Remote Sensing , 35(13 ), pp. 5053-5072.
Zhang, D. & Chen, S., 2002. Fuzzy clustering using kernel
method. Xiamen, Fujian Province,, IEEE, pp. 162 - 163
Zhu, S. C. & Yuille, A., 1996. Region competition:
Unifying snakes, region growing, and Bayes/MDL for
multiband image segmentation. IEEE Transactions on
Pattern Analysis and Machine Intelligence, 18(9), pp.
884-900.
EffectiveSARImageSegmentationandSea-IceFloeDistributionAnlysisviaKernelGraphCutsbasedFeatureExtraction
andFusion
37