Geology Reference
In-Depth Information
have the same limitation. In both cases large‐scale ice
drifts can be estimated [
Kwok et al
., 1998;
Liu and
Cavalieri
, 1998;
Martin and Augstein
, 2000; and
Zhao
et al
., 2002]. SAR data have the advantage of its remarka-
bly finer resolution but at the expense of the poor tempo-
ral frequency of the satellite overpasses. The precision of
the derived motion vectors from passive microwave or
scatterometer data is less than that achievable from SAR.
In all cases, however, ice motion tracking cannot be esti-
mated accurately during the ice melt season or at locations
where variable atmospheric conditions cause decorrela-
tion of the radiometric signature from one day to the next.
Requirements of spatial and temporal resolution for ice
motion tracking vary according to the regional ice char-
acteristics. For example, in the Arctic basin where ice
floes are typically large and move at relatively small veloc-
ities, a spatial resolution of a few kilometers and tempo-
ral resolution around a 10 day revisit interval would be
adequate. In dynamic ice regimes such as marginal ice
zones and low latitude ice‐covered areas (e.g., the Gulf of
St. Lawrence at latitude between 45° and 50°) where the
majority of ice floes are small, a few hundred meter reso-
lution and daily revisit interval is necessary. In highly
dynamic ice regimes where surface signature may change
within hours, swath data should be used instead of daily
averaged gridded data. Validation of motion retrieval is
usually performed using data from drifting buoys, anima-
tion of fine resolution data such as SAR, output from ice
motion models, or comparison to sea level pressure fields.
One of the most common techniques to detect pack ice
motion is called maximum cross correlation (MCC). This
is an application of the well‐known method of matched
filter. It operates on a pair of remote sensing images.
A brief description of the technique is presented here
using the illustration in Figure 10.45. Details are described
in
Ninnis et al
., [1986] and
Fily and Rocthrock
[1987]. The
first step is to co‐register the two images. Then, reference
subareas (also called features in some studies) are identi-
fied in the first image and the match of each subarea
should be found in the second image by convolving the
subarea against the entire second image. Each subarea in
the first image (as shown by the bright-boundary box in
the first image in the figure) should include a unique
structure or texture in order to find a match in the second
image. Otherwise the subarea will not have a unique
match. An example of the structure is shown inside the
subarea in the first image in Figure 10.45. The structure
represents a combination of small ice floes. As the win-
dow of the subarea of (
n
×
n
) pixels is displaced at fixed
steps to perform the convolution against the second
image, the correlation coefficient
R
is calculated:
a
a
b
b
t
t
ipjq
,
kplq
,
Rijkl
,; ,
(10.108)
n
2
pt
qt
ab
where the subscripts (
i
,
j
) and (
k
,
l
) are the coordinates in the
first and second images, respectively,
a
i
,
j
and
b
k
,
l
are the radi-
ometric
v
alues
f
from the windows in the first and second
images,
a
and
b
are the corresponding mean values, and
σ
a
and
σ
b
are the corresponding standard deviations, respec-
tively. The parameter
t
is equal to (
n
− 1)/2. The match
between the subarea from the first image and its location in
the second image (also called tie points) is defined by the
(a)
(b)
(c)
First image
Eulerian displacement vectors
Second image
Figure 10.45
Illustration of the generation of displacement vectors between two images using the MCC tech-
nique. The first image is shown in (a) and the second image shown in (b). A subarea that includes a unique struc-
ture from the first image (defined by the box) is convolved with the second image. The match is identified by the
box with the dark boundary in (b). The displacement vector vectors of grid points that originate in the first image
are shown in (c).
