Geology Reference
In-Depth Information
original version of each algorithm and the other was a
universal (consistent) set obtained from the Round Robin
Data Package (RRDP). This package encompasses a set
of ice draft and thickness measurements. It is the back-
bone of the ESA Climate Change Initiative (CCI).
The adaptation of a universal set of tie points
reduces the differences between the algorithm results
remarkably. This allows fair comparison between algo-
rithms as it eliminates any possible error caused by
applying the algorithm in a region different than the
region from which the tie points were established. In
this manner differences between algorithms' result can
be attributed to the methodology or the framework of
the algorithm. Fairer comparison between algorithms
can then be achieved. Another interesting remark from
Figure 10.25 is the higher ice area and extent from
using the original tie points compared to using the uni-
versal tie points (the difference is around one million
square kilometer). No explanation for this observation
is offered, but the lower values obtained from using the
consistent set of tie points must be closer to the truth
because the tie points are more realistic. The spread in
the estimation of the sea ice area is larger than the esti-
mation of its extent, which is expected because of the
coarser calculations of the latter (see section 10.3).
Differences between algorithms are manifested more
in the calculations of the area.
likely to happen in the case of co‐polarization SAR
observations. According to
Eriksson et al
. [2010] this
problem exists in all operational SAR frequencies:
X‐, C‐, or L‐band data. According to
Geldsetzer and
Yackle
[2009] cross‐polarization backscatter from OW
is almost independent of the wind‐generated water
surface roughness (yet
Voronovich and Zavorotny
[2011]
argued otherwise). The authors suggested that the
combination of co‐ and cross‐polarized data could
potentially lead to improved algorithms for the
retrieval of ice concentration.
Estimating ice concentration using remote sensing
imagery data is equivalent to discriminating sea ice
from OW. Traditionally, a threshold approach is used
to discriminate ice from water pixels in SAR backscat-
ter data. Other advanced approaches that depend on
SAR image analysis include the use of texture features
[
Clausi
, 2001], neural networks [
Kern et al
., 2003],
Markov random field model, [
Deng and Clausi
, 2005],
learning vector quantization [
Bovith and Anderson
,
2005], and iterative region growing using semantics
[
Yu and Clausi
, 2007].
Karvonen
[2012] developed an algorithm that uses the
autocorrelation as a statistical texture measure to avoid
the dependence of OW backscatter signature on wind
speed. The algorithm was applied to Radarsat‐2 HH
polarization from the ScanSAR mode. Although it has
not been widely applied or verified, it carries potential
because the autocorrelation is also insensitive to the inci-
dence angle (the Radarsat ScanSAR mode incidence
angle has a wide range from 20° to 50°). The algorithm is
described briefly in the following. It starts with a segmen-
tation technique to produce a predetermined number of
clusters (typically 10). Small segments are then merged
with larger segments. Directional autocorrelation COR
is calculated using a sliding 11 × 11 pixel block
B
within
each segment:
10.2.3. Ice Concentration Using SAR
Although SAR imagery is the prime data source for
generating operational ice charts in national ice cent-
ers, a quantitative algorithms for calculating ice con-
centration from these data have not matured to satisfy
the operational requirements. Therefore, after more
than three decades of research on ice type and concen-
tration retrieval from SAR, this information is still
extracted based on visual analysis of the images by
trained ice analysts. They use ancillary data that
include weather, climatic as well as recent history of
the ice regime to support the image interpretation
(section 11.2). SAR data are preferred over the coarser‐
resolution passive microwave data, although the latter
have been used for synoptic‐scale information [
Bertoia
et al
., 2004]. The main difficulty that has hindered the
development of an automated or semiautomated capa-
bility for ice classification (hence concentration) is the
overlap of backscatter signatures from different ice
types and with OW (section 8.1.1). Low backscatter of
smooth ice surface overlaps with the low backscatter
of calm open water. Likewise high backscatter gener-
ated by wind‐driven ocean surface roughness overlaps
with the high backscatter signature from rough young
and FY ice. These overlapping signatures are more
Ii kj l
,
I ij
( .
)
B
B
ijB
,
COR
kl
2
B
B
(10.56)
where
I
(
x
,
y
) is the image pixel,
μ
B
and
σ
B
are the sample
mean and standard deviation in the given set
B
. The
directional estimates are achieved by using four different
lag values (
k
,
l
) = (0, 1), (1.0), (1, 1), (1, − 1). The autocor-
relation distribution
p
(
x
) for each segment is assumed
to be a mixture of
M
Gaussian distributions, each one
representing either an ice type or OW:
M
px
gx
(10.57)
kk
k
1
