IEEE Geoscience and Remote Sensing Magazine - September 2021 - 142

OSC maps were determined [13]. However, considering the
fact that GOCI-derived OSC data are actually a displacement
of one template window with a size of more than
10-15 km2 in 1 h, obtaining identical measurements and
simulations that can be used to achieve a point-to-point
validation is unlikely.
SUBPIXEL COREGISTRATION
One common limitation of the previous works is related
to the discretization error that affects the current speed
and direction. Because the image coregistration in previous
works was implemented at the pixel scale (500 m) and
the time interval between two consecutive images was 1 h,
the north and east components of GOCI-derived OSC data
can only have values that are multiples of 13.89 cm/s [13].
However, the MCC approach can provide subpixel accuracy-in
extreme cases, one-tenth of a pixel or even better
[11]. Therefore, it is worthwhile to introduce the subpixel
coregistration technique to decrease the discretization error.
Instead of searching for an identical measurement or
simulation that can be used to achieve point-to-point validation,
a strictly controlled test would be more practical to
assess accuracy at the subpixel level.
QUALITY CONTROL FILTERING
Various types of quality control filtering were specified and
performed on GOCI-derived velocity vectors in the previous
works, including missing data ratios, minimum cross-correlation
cutoffs, reciprocal filtering, neighborhood similarity,
robust optical flow, and edge MCC rejection, with
the intention of removing erroneous estimates. However,
selecting the threshold value for each quality control filtering
process always involves a tradeoff between unreliable
information and less matching. The authors of [27]
pointed out that the MCC method is implicitly assumed to
involve mostly advective motion, with negligible rotational
and deformational motions. When the MCC method is
used to derive an OSC around an eddy, such rotational and
deformational changes may reduce correlations and cause
the derived vector to deviate from neighborhood ones.
Consequently, even though a lot of prominent features are
presented around an eddy structure, most of the derived
vectors are eventually ruled out after quality control filtering.
It is therefore preferable to have objective criteria by
which to initially filter out suspicious vectors with minimum
preset thresholds.
To summarize, after a review of the previous work on
GOCI-derived OSC data, this article is an attempt to develop
a more reliable approach that can be used to detect
and track features to a subpixel level, even with some
cloud cover. Suspicious vectors can be initially ruled out
through objective criteria with minimum preset thresholds.
Instead of seeking point-to-point validation against
existing measurements and simulations, a series of strictly
controlled tests is conducted to provide a detailed accuracy
assessment.
142
METHODS
The approaches proposed in this work were implemented
using a SIMS. To ensure that the SIMS would be able to
achieve all goals in a step-by-step fashion, we began by selecting
one cloudless and landless GOCI image, taken at
8:16 a.m. on 5 April 2011, as the base image. This was also
the image employed in [33] to derive OSCs. Standard GOCI
level-2a products were downloaded from the Korea Ocean
Satellite Center and preprocessed using GDPS v2.0.
SUBPIXEL COREGISTRATION
For a detailed explanation of estimating the shift with subpixel
accuracy, the reader can refer to [25]. The idea is to
find the point (, )xxx 12
=
12
!! a point (, )xx
12
a =
1
2
wx wx
wwl1
11 +
1 +
a = 22
2
wx wx
ww2
$
$
+
+
where
wP xxhhl
ll
11
wP xx
wP xx h
wP xx
Note that a 06 . 5=
21 22
1
11 21
2
ll ll
^
22 12Px x
=+
=+
=+
=+
^
^
h
h
^
^
^^
^
,, ,
,, ,
,, ,
,, .
aa
aa
aa
aa
Px x2
Px x
Px x
ll
l
12
1
h
h
h
(2)
is an empirically optimal value [8],
which can be regarded as a weight of the cross-correlation
coefficient.
Equations (1) and (2) were implemented by a SIMS to
estimate the displacement vectors to the subpixel level from
two consecutive GOCI images. To assess the accuracy of the
SIMS, we follow the work in [33] to select the same GOCI
image and show the true color composite of the level-1b
product in Figure 1(a). The rectangular region off northern
Ulleungdo (540 × 322.5 km) is specified as the study
area, denoted by a red box, because there is no cloud and
land cover. Although the contrast of the sea water within
the box is quite low and the true color image looks rather
dark, the bio-optical product of the ()K 490d
in the same region
reveals numerous spatial patterns and structures that
are ideal to serve as the base image for feature tracking, as
shown in Figure 1(b).
Nine shifted ()K 490d
images are generated by applying
a uniform yet noninteger amount of shift to this base image
in the x and y directions, respectively, with a total of 18
random values ranging from one to five. The built-in Interactive
Data Language function Warp_2D is employed to
perform linear spatial warping with bilinear interpolation.
Note that the box in Figure 1(b) is blown up in Figure 1(c),
and the same region from nine different shifted images is
shown in Figure 1(d)-(l), with white dotted lines to illustrate
the noninteger amount of shift. Apart from using the same
K ()
IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE SEPTEMBER 2021
22
11
$
ll$
,
ll
l
which the value (, )xxP 12
ll is the largest. The coordinates
(, )aa
12
of the shift can be calculated by
,
(1)
for which the cross-correlation
coefficient ()P x takes the largest possible value. Among
four points (, ),xx11
ll is selected for
d 490 bio-optical product, the best setup (a search range
IEEE Geoscience and Remote Sensing Magazine - September 2021

Table of Contents for the Digital Edition of IEEE Geoscience and Remote Sensing Magazine - September 2021
