Geology Reference
In-Depth Information
available from Radarsat‐2 and ALOS PALSAR only.
Multifrequency data have not been available from any
space‐borne SAR but ESA is planning to launch the
Earth Explorer Cold Regions Hydrology (CoReH
2
O) sat-
ellite to measure snow on glaciers and sea ice. Classification
of thin ice types has been identified as a requirement
from this system. The satellite will measure co‐ and cross
polarization from X‐band and Ku‐band radars. Only a
few studies have addressed the potential of using multi-
frequency data for ice classification while more studies
have addressed the use of multipolarization data.
As mentioned in section 7.6.2.2, the first experimental
fully polarimetric multifrequency airborne SAR system
(AIRSAR) was installed onboard NASA's DC‐8 aircraft
in 1988. It operated at C‐, L‐, and P‐bands (wavelength
5.6, 23.5, and 68 cm, respectively).
Drinkwater et al
. [1991]
compiled statistics of polarimetric parameters from the
three bands to characterize the broad categories of FY
ice and MY ice along with the subcategories of thin and
thick FY ice in the Beaufort Sea and the marginal ice
zone in the Bering Sea during March 1988. Although the
study did not encompass any classification scheme, the
backscatter observations from existing ice types led
the authors to conclude that the various combinations
of wavelength and polarization offer better capability
to distinguish between sea ice of different fabrics or
morphologies.
An early attempt to explore the difference in the
discriminating capability between multifrequency and
multipolarization data from AIRSAR is presented in
Shokr et al
. [1995]. A set of 18 images of a sea ice scene in
the Beaufort Sea obtained on 11 March 1988 was used.
The set consisted of three subsets, one from each fre-
quency of AIRSAR, and each subset has six images gen-
erated for polarization states: HH, VV, HV, RR, LL, and
RL. The last three are circular polarizations. Two sets of
training areas for FY ice and MY ice were selected in the
near range of the image and that was repeated in the far
range. Gray tone and texture (using the GLCM parame-
ters) were sampled from each training area to examine
their distributions in terms of their separability or
overlap. Statistical tests were carried out to determine
the discriminating power of each channel or combina-
tion of channels in separating the two ice types. Linear
and multiple discriminant analyses were used for this
purpose. The Wilkes' lambda criterion was used to deter-
mine the ratio of within‐class variance to the total vari-
ance in the entire set of samples. A simple supervised
classification technique was used to assign pixels to the
appropriate class based on the nearest neighbor criterion.
Results showed that combining texture with gray tone
improved the classification accuracy. Better accuracies
were obtained when combining frequency channels of
the same polarization rather than combining polarization
channels from the same frequency. No substantial dif-
ference in classification accuracy was noticed between
co‐ and cross‐polarization data in all frequencies. The
classification performance from the circular polarization
data was similar to the linear polarization except for the
RL polarization in the P‐band where substantially lower
classification accuracy was obtained. The far range data
could be classified at higher accuracies.
A few investigations were conducted to explore the
potential uses of polarization variety of ENVISAT in ice
classification.
Geldsetzer and Yackle
[2009] explored the
potential of combining the co‐polarization backscatter
coefficients
0
to classify ice types and discrimi-
nate ice from OW. Twenty scenes of sea ice and OW in an
area surrounding Cornwallis Island, Nunavut, Canada,
obtained in April 2004, were used. Figure 10.6 shows two
images of
0
and
hh
vv
0
and
0
in addition to the co‐polarization
hh
vv
0
0
0
is
ratio
/
. The backscatter of OW from
co
hh
vv
vv
0
, but the reverse is true for
the smooth FY ice. The co‐polarization ratio from OW
and thin ice surfaces is remarkably high. This is consist-
ent with the polarization‐dependent Fresnel reflection
that predicts higher reflection in the horizontal polariza-
tion (section 7.3.2 and Figure 7.17). The MY ice floes
cannot be identified in the
γ
co
image. The smooth FY ice
has a relatively wide range of backscatter. The classifica-
tion technique used in the study is rather simple, but it
reveals information on the discriminating capability of
the examined co‐polarization channels. The threshold
for separating ice types (TH
0
) is based on backscatter
magnitude:
considerably higher than
hh
TH
0
0
0
s
s
/
2
(10.3)
l
l
u
u
0
,
s
l
, and
s
u
are the lower and upper limits of
the mean backscatter coefficient and its standard devia-
tion, respectively, from a given ice type. The backscatter
coefficient in this equation applies to any ice type in
either horizontal or vertical polarization. A similar
threshold for separating ice types and OW, based on the
co‐polarization ratio, is defined by an equation similar
to equation (10.3):
0
,
where
l
u
TH
s
s
/2
(10.4)
co
l
l
co
uu
Geldsetzer and Yackle
[2009] found the classification
accuracies of the MY ice, rough FY ice, smooth FY ice,
thin ice, and OW to be 99%, 32%, 89%, 67%, and 50%,
respectively, while the total classification accuracy was
approximately 70%. Except for MY ice and smooth FY
ice, the rest of the surface types could not be classified
accurately. This shows the limitation of both the dual
