IEEE Geoscience and Remote Sensing Magazine - September 2021 - 147

travels at any point in time. The streamline products derived
from two consecutive GOCI images would be even more
useful to describe OSCs qualitatively and quantitatively.
To highlight the necessity of removing suspicious vectors,
Figure 5(a) includes the streamlines derived from all
the displacement vectors [labeled in red, green, blue, and
yellow in Figure 4(a)], while Figure 5(b) has those derived
only from reliable displacement vectors [labeled in yellow
in Figure 4(a)]. Five regions affected by suspicious vectors
are denoted by cyan ellipses in Figure 5(a) and (b). Note
that some suspicious vectors have minor effects on the
streamlines. However, the suspicious vectors labeled as
red circles deflect the streamlines and generate unrealistic
flows, as denoted by the cyan ellipses in Figure 5(a).
These unrealistic flows are corrected after removing the
suspicious vectors, resulting in more realistic streamlines,
as in Figure 5(b).
PATCHY CLOUD COVER
Small, ragged cloud fragments are frequently found in optical
imagery of Earth's surface. They are generally sheared by
strong winds and have a shredded appearance. If larger image
regions have to be processed through matching, image
patches with low or no signal content pose a serious problem
[11]. The authors of [30] investigated feature tracking
performance in the presence of variable gaps due to land
and clouds in spatial data within the template and master
images. They concluded that a maximum threshold level of
25% missing data in the combined overlap using the MCC
algorithm would be expected to have a success rate of 80%.
We demonstrated that a robust feature is resilient to a
slight change in the STW. It is thus reasonable to expect that
a robust feature would also be immune to a patchy cloud cover,
as long as it was retrieved from the uncovered valid data
and clear enough to be tracked on two consecutive images.
Since the movement of clouds is different from that of OSCs,
the patterns of patchy cloud cover will not be the same on the
template and master images. To address this issue, a union
mask is set by combining the patchy cloud cover within the
template and master windows. The cross-correlation coefficient
is then calculated on all valid data after excluding pixels
outside the union mask. Note that the union mask and the
cross-correlation coefficient must be recalculated and updated
at each step of the search process, which requires more
computer time and resources than other alternatives.
To provide a detailed accuracy assessment, we conducted
a series of strictly controlled tests by applying nine randomly
generated cloud masks to manually shifted images
that were generated earlier for the purpose of assessing the
subpixel-level accuracy of the SIMS. Figure 6 presents one
pair of the original and the manually-shifted images [the
same ones shown in Figure 1(b) and (d)] with randomly
generated cloud masks. Note that the same approach was
proposed in [30], where the authors used a series of elliptical
patches to mimic real cloud cover. Each elliptical
patch was specified by five random numbers, including the
SEPTEMBER 2021 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE
center coordinates (x, y), lengths of the semimajor axis a
and semiminor axis b, and angle to rotate the ellipse
i . New
elliptical patches were generated and added until the total
amount of cloud cover reached 20%. The SIMS was employed
to calculate the subpixel-level shifts, and the RMSE
was 0.2461 of a pixel (3.42 cm/s), shown as the circle in Figure
7. This result was slightly higher than, yet remarkably
close to, the results when there was no cloud cover [0.2435;
see Figure 3(b)], which confirms that a robust feature is indeed
resilient against patchy cloud cover. Although there is
no limitation on the cloud cover ratio in the template and
the master window, the default size-irrelevant filtering of
the SIMS helps to rule out suspicious vectors and ensures
that all features retrieved from the uncovered and valid data
are clear enough to be tracked on two consecutive images.
To summarize, the SIMS can be used to detect and track
features from two consecutive GOCI images to approximately
a 0.25-pixel level of accuracy, even with some cloud
cover. Suspicious displacement vectors can be ruled out using
the size-irrelevant and majority filtering methods, resulting in
a set of reliable displacement vectors that can produce more realistic
streamlines. For days that have only a few hours of clear
130° E 131° E 132° E 133° E 134° E 135° E
130° E 131° E 132° E 133° E 134° E 135° E
(a)
130° E 131° E 132° E 133° E 134° E 135° E
100
km
130° E 131° E 132° E 133° E 134° E 135° E
(b)
100
km
FIGURE 6. Randomly generated cloud masks overlaid on (a) the
base image [as in Figure 1(b)] and (b) the manually shifted image [as
in Figure 1(d)].
147
38° N 39° N 40° N
38° N 39° N 40° N
38° N
39° N
40° N
38° N
39° N
40° N
IEEE Geoscience and Remote Sensing Magazine - September 2021

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