atmosphere
Article
An Integrated Convective Cloud Detection Method
Using FY-2 VISSR Data
Kuai Liang, Hanqing Shi *, Pinglv Yang and Xiaoran Zhao
Institute of Meteorology and Oceanography, PLA University of Science and Technology, Nanjing 211101, China;
[email protected] (K.L.); [email protected] (P.Y.); [email protected] (X.Z.)
* Correspondence: [email protected]; Tel.: +86-25-8083-0624
Academic Editor: Robert W. Talbot
Received: 26 November 2016; Accepted: 15 February 2017; Published: 20 February 2017
Abstract: The characteristics of convective clouds on infrared brightness temperature (BTIR ) and
brightness temperature difference (BTD) image were analyzed using successive Infrared and Visible
Spin-Scan Radiometer (VISSR) data of FY-2, and an integrated detection method of convective clouds
using infrared multi-thresholds in combination with tracking techniques was implemented. In this
method, BT and BTD thresholds are used to detect severe convection and uncertain clouds, then the
tracking technique including overlap ratio, minimum BT change and cross-correlation coefficient is
used to detect convection activities in uncertain clouds. The Application test results show that our
integrated detection method can effectively detect convective clouds in different life periods, which
show a better performance than any single step in it. The statistical results show that the α-type
clouds are mostly large-scale systems, and the β- and γ-type clouds have the highest proportion of
general type. However, the proportion of weak convective cloud is higher than that of severe ones in
γ-type cloud, and an opposite result is found in the β-type.
Keywords: convective cloud; VISSR; integrated method; detection
1. Introduction
Convective clouds are produced by atmosphere during uplift movement under thermal or
dynamic effect, causing strong winds, hail, lightning, heavy rainfall and other severe weather.
Improving the accuracy of convective cloud detection and early warning is an important component
of modern weather forecast business construction [1].
Since the 1960s, meteorological satellites have rapidly become a powerful tool for studying
convective cloud [2,3]. The use of satellite data not only broadens convective cloud detection methods,
but also monitors areas where conventional data cannot be obtained [4]. The geostationary satellite,
which is able to monitor the same area in successive time, has become one of the main ways to detect
convective cloud [5,6].
There are mainly three convective cloud detection methods based on the geostationary satellite
data. The first one is the single-threshold method, which directly detects clouds by employing
the infrared brightness temperature (BTIR ) and infrared brightness temperature difference (BTDIR )
threshold values [7]. It is the main method for detecting convection activities in the current operational
business [8]. For example, in the established auto nowcasting system of NCAR (National Center for
Atmospheric Research), BT thresholds of multiple channels are used for detecting and nowcasting
of convective cloud [9]. However, it is easy to misinterpret some cirrus clouds as convective cloud,
since their cloud top heights are sometimes similar to each other [10,11]. The second method is the
artificial intelligence algorithm [12], which achieves the purpose of automatic detection through a large
number of samples for training model parameters, such as fuzzy logic theory and neural network [13].
The third method is to track the clouds to confirm if they have the convection characteristics [14–18].
Atmosphere 2017, 8, 42; doi:10.3390/atmos8020042
www.mdpi.com/journal/atmosphere
Atmosphere 2017, 8, 42
2 of 11
For example, Arnaud et al. [19] realized automatic tracking of low-temperature cloud top by applying
the overlap area method; Carvalho et al. [20] used the maximum spatial correlation to track and match
the mesoscale convection systems.
In the past, the way to detect convective cloud was mainly based on some single methods.
However, the research of Huang et al. [21] showed that the combined use of the BT threshold method
and the image processing method can effectively detect convective cloud in a wide area. In this paper,
we integrated the threshold method with the tracking techniques, and designed a step-by-step detection
process for the convection systems of different life periods. Section 2 analyzes the characteristics of
BTIR1 , BTDIR1-IR2 , BTDIR1-IR3 and BTDIR1-IR4 of FY-2F images, and tries to determine some reasonable
thresholds by making a comparison of different thresholds for detecting convective cloud. In addition
to the infrared threshold method, parameters such as the overlap ratio, minimum BT change, and the
cross-correlation coefficient [22,23] at two successive times are also introduced to our method as the
tracking process. Section 3 shows the application test results of every detection step by counting the
number of cloud clusters and other statistics.
The Infrared and Visible Spin-Scan Radiometer (VISSR) on FY-2 provides data of four infrared
channels (IR1 : 10.3–11.3 µm, IR2 : 11.5–12.5 µm, IR3: 6.3–7.6 µm, and IR4: 3.5–4.0 µm) and a visible
channel (0.55–0.9 µm). In this paper, we adopted geographic Lat/Lon projection of 5 km × 5 km
infrared channels data made from the full disc nominal image file in level-1 data and the research
scope was determined to be 70◦ E–140◦ E, 0◦ N–50◦ N.
2. Threshold Determination and Detection Process
2.1. Threshold Determination
In order to evaluate the performance of BTIR and BTD images for convection detection, we made
a comparison case of BTIR1 , BTDIR1–IR2 , BTDIR1–IR3 , and BTDIR1–IR4 images on 16 June 2015, shown in
Figure 1, and tried to determine a feasible threshold range. In Figure 1a, the deep blue area indicates
the convection area. Also, it could be found that the convection area has a clear boundary with a center,
where BT is less than 200 K, which indicates high and strong convection activity. On the BTDIR1–IR2
image, the convective cloud has an irregular texture, and it is easily confused with some cirrus clouds.
The convection area of the BTDIR1–IR3 image on Figure 1c is below about 10 K. The margin outline
is the same prominence as the BTIR1 image; and on the BTDIR1–IR4 image, the convection area is
generally below −20 K. However, its low value zone does not correspond exactly to the position of the
convection center, for extremely low value occurs in the non-convection section and the convective
boundary. According to the comparison of every image in Figure 1, we found that the convection
activities on the BTIR1 image have more distinct characteristics than other images in the large-scale
convective systems. So it could play a major role in our detection method. As regards the other images,
they have the ability to detect some convection area and eliminate some non-convection area.
Based on the comparison results, we eventually regarded the BTIR1 image as the major method
for “Detection”, and the BTD images as the supplement for ”Elimination”, which means that the
BTIR1 image is used to detect the main convection area, and other BTDs are used to eliminate some
other clouds, respectively. The next problem is to determinate feasible thresholds for each image.
Here, we analyzed the detection result with a series of continuous thresholds. Figure 2 shows the
detection result when the BTIR1 threshold changes from 210 K to 250 K. The result shows that the
central area of convective cloud is widely detected out when the threshold is under 220 K. When the
threshold exceeds 220 K, the detection area spreads from the convection center to the surrounding
convection area; and when the threshold changes from 240 K to 250 K, the only enlarged area is mostly
cirrus clouds and some possible convective cloud. In order to detect the convective cloud system and
not involve much cirrus clouds at the same time, we considered using 240 K to detect the possible
convective cloud.
Atmosphere 2017, 8, 42
Atmosphere 2017, 8, 42
Atmosphere 2017, 8, 42
Latitude
Latitude
32°N
32°N
3 of 11
3 of 3
3 of 3
(a)
(a)
280
280
260
260
30°N
30°N
240
240
28°N
28°N
220
220
26°N
26°N
72°E 76°E 78°E 80°E 82°E 84°E
72°E 76°E 78°E 80°E 82°E 84°E
Latitude
Latitude
32°N
32°N
(c)
(c)
200
200
40
40
30
30
20
20
10
10
0
0
30°N
30°N
28°N
28°N
26°N
26°N
72°E 76°E 78°E 80°E 82°E 84°E
72°E 76°E 78°E
80°E 82°E 84°E
Longitude
Longitude
32°N
32°N
(b)
(b)
30°N
30°N
28°N
28°N
26°N
26°N
72°E 76°E 78°E 80°E 82°E 84°E
72°E 76°E 78°E 80°E 82°E 84°E
32°N
32°N
(d)
(d)
30°N
30°N
28°N
28°N
26°N
26°N
72°E 76°E 78°E 80°E 82°E 84°E
72°E 76°E 78°E
80°E 82°E 84°E
Longitude
Longitude
10
10
8
8
6
6
4
4
2
2
0
0
10
10
0
0
-10
-10
-20
-20
-30
-30
-40
-40
Figure
caseofofconvective
convective cloud on
on brightness temperature
(BT)
Figure
1. 1.
AA
case
(BT)and
andbrightness
brightnesstemperature
temperature
Figure
1.
A
case of convectivecloud
cloud on brightness
brightness temperature
temperature (BT)
and
brightness
temperature
difference
(BTD)
images
(/K):
(a)
infrared
brightness
temperature
(BT
IR1) image; (b) BTDIR1–IR2 image;
difference
(BTD)
images
(/K):
) image; (b) BTDIR1–IR2
image;
difference
(BTD)
images
(/K):(a)
(a)infrared
infraredbrightness
brightnesstemperature
temperature(BT
(BTIR1
IR1) image; (b) BTDIR1–IR2
image;
(c) BTDIR1–IR3 image; (d) BTDIR1–IR4 image. The deep blue areas indicate the convection activeties.
(c) (c)
BTD
image;
(d)
BTD
image.
The
deep
blue
areas
indicate
the
convection
activeties.
IR1–IR3
IR1–IR4
BTDIR1–IR3 image; (d) BTDIR1–IR4 image. The deep blue areas indicate the convection activeties.
Latitude
Latitude
32°N
32°N
30°N
30°N
28°N
28°N
26°N
26°N
(a) 210 K
(a) 210 K
(b) 215 K
(b) 215 K
(c) 220 K
(c) 220 K
(d) 225 K
(d) 225 K
(e) 230 K
(e) 230 K
(f) 235 K
(f) 235 K
Latitude
Latitude
32°N
32°N
30°N
30°N
28°N
28°N
26°N
26°N
Latitude
Latitude
32°N
32°N
30°N
30°N
28°N
28°N
26°N
26°N
(g) 240 K
(g) 240 K
74°E 76°E 78°E 80°E 82°E 84°E
74°E 76°E 78°E
80°E 82°E 84°E
Longitude
Longitude
(h) 245 K
(h) 245 K
74°E 76°E 78°E 80°E 82°E 84°E
74°E 76°E 78°E
80°E 82°E 84°E
Longitude
Longitude
(i) 250 K
(i) 250 K
74°E 76°E 78°E 80°E 82°E 84°E
74°E 76°E 78°E
80°E 82°E 84°E
Longitude
Longitude
Figure 2. (a–i) Results of convective cloud detection by Infrared BTIR1 thresholds.
Figure 2. (a–i) Results of convective cloud detection by Infrared BTIR1 thresholds.
Figure 2. (a–i) Results of convective cloud detection by Infrared BTIR1 thresholds.
The same method for determining the detection threshold is also used for the BTD images in
The same method for determining the detection threshold is also used for the BTD images in
Figure
where
three BTD’s
detection results
vary fromthreshold
four thresholds.
can for
be seen
in Figure
3, it in
The 3,
same
method
for determining
the detection
is alsoAs
used
the BTD
images
Figure
3, where
three BTD’s
detection results
vary from four thresholds.
As
can be seen
in Figure
3, it
is
difficult
to
distinguish
convection
from
cirrus
clouds
at
any
threshold
in
the
BTD
IR1–IR2 image. So
Figure
3, where
three BTD’s
detectionfrom
results
vary
from at
four
Asthe
can
beIR1–IR2
seen image.
in Figure
is difficult
to distinguish
convection
cirrus
clouds
anythresholds.
threshold in
BTD
So 3,
it is difficult to distinguish convection from cirrus clouds at any threshold in the BTDIR1–IR2 image.
Atmosphere 2017, 8, 42
4 of 11
Atmosphere 2017, 8, 42
Atmosphere 2017, 8, 42
4 of 4
4 of 4
BTD
BTD
IR1-IR4
IR1-IR4
BTD
BTD
IR1-IR3
IR1-IR3
BTD
BTD
IR1-IR2
IR1-IR2
So the
BTD
canused
be used
to eliminate
the and
low middle
and middle
at a threshold
of about
4 K.
IR1–IR2
the
BTD
IR1–IR2
can be
to eliminate
the low
cloudcloud
at a threshold
of about
4 K. The
the
BTD
IR1–IR2 can be used to eliminate the low and middle cloud at a threshold of about 4 K. The
The
BTD
image
has
a
similar
effect
to
the
BT
,
and
it
can
eliminate
most
non-convective
areas
IR1it can eliminate most non-convective areas when
BTDIR1–IR3IR1–IR3
image has a similar effect to the BTIR1, and
BTD
IR1–IR3 image has a similar effect to the BTIR1, and it can eliminate most non-convective areas when
when
the
threshold
value
is
set
at
about
10
K.
For
the BTD
image,
a better
threshold
of K
−16
K
the threshold value is set at about 10 K. For the BTD
IR1–IR4IR1–IR4
image,
a better
threshold
of −16
was
the
threshold
value iswhen
set atthe
about
10 K.varies
For the
BTD
IR1–IR4 image, a better threshold of −16 K was
was
selected
because
threshold
from
−
16
to
−
11
K,
the
main
convection
area
remains
selected because when the threshold varies from −16 to −11 K, the main convection area remains
selected
because when
the only
threshold varies
from
−16 to −11clouds.
K, the main convection area remains
mostly unchanged
unchanged
and the
the
mostly
and
only enlarged
enlarged area
area is
is some
some cirrus
cirrus clouds.
mostly unchanged and the only enlarged area is some cirrus clouds.
32°N
32°N
30°N
30°N
28°N
28°N
26°N 3.2 K
26°N 3.2 K
3.6 K
3.6 K
4K
4K
4.4 K
4.4 K
10 K
10 K
15 K
15 K
20 K
20 K
32°N
32°N
30°N
30°N
28°N
28°N
26°N 5 K
26°N 5 K
32°N
32°N
30°N
30°N
28°N
28°N
26°N -16 K
26°N -16 K
74°E 76°E 78°E 80°E 82°E 84°E
74°E 76°E 78°E 80°E 82°E 84°E
-14.5 K
-14.5 K
74°E 76°E 78°E 80°E 82°E 84°E
74°E 76°E 78°E 80°E 82°E 84°E
-13 K
-13 K
74°E 76°E 78°E 80°E 82°E 84°E
74°E 76°E 78°E 80°E 82°E 84°E
-11.5 K
-11.5 K
74°E 76°E 78°E 80°E 82°E 84°E
74°E 76°E 78°E 80°E 82°E 84°E
Figure
3.
Results
of
convective
cloud
detection
by
infrared
BTD
thresholds.
Figure 3.
3. Results
Resultsof
ofconvective
convectivecloud
clouddetection
detection by
by infrared
infrared BTD
BTD thresholds.
thresholds.
Figure
After determining the four thresholds, we did a step-by-step threshold test to ensure the
After
After determining
determining the
the four
four thresholds,
thresholds, we
we did
did aa step-by-step
step-by-step threshold
threshold test
test to
to ensure
ensure the
the
performance of the BT threshold method seen in Figure 4. First, the BTIR1 is used to detect the main
performance
seen
in in
Figure
4. First,
the BT
isIR1
used
to detect
the main
performance of
ofthe
theBT
BTthreshold
thresholdmethod
method
seen
Figure
4. First,
theIR1BT
is used
to detect
the
convective cluster, and then the three BTD thresholds are used to eliminate the non-convection area.
convective
cluster,
and then
three
are used
eliminate
the non-convection
area.
main convective
cluster,
andthe
then
the BTD
threethresholds
BTD thresholds
are to
used
to eliminate
the non-convection
Figure 4a shows that the integrated BT threshold method is able to detect out the most convective
Figure
4a shows
that the
BT threshold
method
is able
to to
detect
area. Figure
4a shows
thatintegrated
the integrated
BT threshold
method
is able
detectout
outthe
themost
mostconvective
convective
cloud. The detection results are not affected by the un-convection area. Compared with other single
cloud.
cloud. The
The detection
detection results
results are
are not
not affected
affected by
by the
the un-convection
un-convection area.
area. Compared
Compared with
with other
other single
single
thresholds above, we found that the integrated BT threshold method has a better performance in
thresholds
above,
we
found
that
the
integrated
BT
threshold
method
has
a
better
performance
thresholds above, we found that the integrated BT threshold method has a better performance in
in
cirrus cloud elimination. However, there are still some dotted “noise” clouds found in Figure 4a. At
cirrus
are
still
some
dotted
“noise”
clouds
found
in Figure
4a. At
cirrus cloud
cloudelimination.
elimination.However,
However,there
there
are
still
some
dotted
“noise”
clouds
found
in Figure
4a.
this moment, four pixels are used as a threshold to remove small area clouds and the result of Figure
this
moment,
fourfour
pixels
are used
as a threshold
to remove
small area
and theand
result
Figure
At this
moment,
pixels
are used
as a threshold
to remove
smallclouds
area clouds
theofresult
of
4b is obtained. By comparing the two results before and after the removal of broken clouds, it was
4b
is obtained.
By comparing
the twothe
results
after
theafter
removal
of broken
clouds, clouds,
it was
Figure
4b is obtained.
By comparing
two before
resultsand
before
and
the removal
of broken
discovered that broken clouds are effectively eliminated after the removal process.
discovered
that broken
cloudsclouds
are effectively
eliminated
after the
removal
process.
it was discovered
that broken
are effectively
eliminated
after
the removal
process.
Latitude
Latitude
32°N
32°N
(a)
(a)
32°N
32°N
30°N
30°N
30°N
30°N
28°N
28°N
28°N
28°N
26°N
26°N
26°N
26°N
72°E
72°E
76°E
76°E
78°E
80°E
78°E
80°E
Longitude
Longitude
82°E
82°E
84°E
84°E
(b)
(b)
72°E
72°E
76°E
76°E
78°E
80°E
78°E
80°E
Longitude
Longitude
82°E
82°E
84°E
84°E
Figure 4. Integrated detection result of convective cloud: (a) Before the removal of “broken cloud”;
Figure
4.
Integrated
detection result
of
Figure
4. the
Integrated
result
of convective
convective cloud:
cloud: (a)
(a)Before
Beforethe
theremoval
removalof
of“broken
“brokencloud”;
cloud”;
(b)
After
removaldetection
of “broken
cloud”.
(b)
After
the
removal
of
“broken
cloud”.
(b) After the removal of “broken cloud”.
The geostationary satellite provides users with data of successive time which inspires researchers
The geostationary satellite provides users with data of successive time which inspires researchers
to find the way to match and track the clouds. If a convective cloud cluster is growing, its minimum
to find the way to match and track the clouds. If a convective cloud cluster is growing, its minimum
Atmosphere 2017, 8, 42
5 of 11
The 2017,
geostationary
Atmosphere
8, 42
satellite provides users with data of successive time which inspires researchers
5 of 5
to find the way to match and track the clouds. If a convective cloud cluster is growing, its minimum
(BTmin
becomeslower
lowerfast
fastand
andthe
theratio
ratioof
ofthe
the overlap
overlap area
area between
between the
the two successive
successive images
BTIRIR (BT
) )becomes
min
decreasing
rate
threshold
and set
set 50%
50% as
becomes high. In this paper, we employed
employed 88 K/h
K/h as the BTmin
decreasing
rate
threshold
and
min
the overlap ratio
threshold
to
confirm
whether
the
detection
area
has
convection
characteristics
after
BT
ratio threshold to confirm whether the detection area has convection characteristics after
threshold
detection
process.
Furthermore,
there
are are
sometimes
several
cloud
clusters
meeting
the two
BT
threshold
detection
process.
Furthermore,
there
sometimes
several
cloud
clusters
meeting
the
criterions.
To determine
whether
therethere
is a possible
cloudcloud
cluster
which
is growing,
we calculated
their
two
criterions.
To determine
whether
is a possible
cluster
which
is growing,
we calculated
cross-correlation
coefficient.
If thereIfis athere
coefficient
value beyond
0.35,beyond
we thought
their
cross-correlation
coefficient.
is a coefficient
value
0.35, the
wecorresponding
thought the
cloud cluster belonged
to convective
or we would
skip
current
cluster
to restart
corresponding
cloud cluster
belongedcloud,
to convective
cloud,
or to
wenext
would
skipcloud
to next
current
cloud
our detection
process.
cluster
to restart
our detection process.
2.2. Detection
Detection Process
Process
2.2.
On the
BTBT
threshold
method
and and
the tracking
method,
a step-by-step
detection
process
On
the basis
basisofofthe
the
threshold
method
the tracking
method,
a step-by-step
detection
for
convective
cloud
is
designed
as
Figure
5:
process for convective cloud is designed as Figure 5:
Tracking and matching method
Severe convection
center
BT threshold method
BTIR1
Uncertain
cloud
BTDIR1-IR2
Small area
eliminated
Minimum BTIR1，
overlap ratio，
correlation matching
Preliminary
detection result
Severe
convection
BTDIR-IR3
Un-severe
convection
Integrated detection
result
BTDIR1-IR4
Cloud classification
Figure 5. Convective cloud detection process. The detection process goes from the BT threshold
Figure 5. Convective cloud detection process. The detection process goes from the BT threshold
method and ends with cloud classification.
method and ends with cloud classification.
•
••
••
•
•
•
•
••
••
•
Step 1: detect severe convection center by employing 220 K as the BTIR1 threshold.
Step 2:
1: detect
detect preliminary
severe convection
centercloud
by employing
220 K240
as K
theasBT
threshold.
IR1BT
Step
convective
by employing
the
IR1 threshold.
Step 2:
preliminary
convective
by employing
the−16
BTIR1
threshold.
Step
3: detect
eliminate
non-convection
areacloud
by employing
4 K,240
10 K
K as
and
K as
the threshold of
StepIR1–IR2
3: eliminate
non-convection
by employing 4 K, 10 K and −16 K as the threshold of
BTD
, BTDIR1–IR3
and BTDIR1–IR4,area
respectively.
BTDIR1–IR2
, BTDIR1–IR3
and BTDIR1–IR4
respectively.
Step
4: matching
the convection
center, with
the preliminary convective cloud. The cloud cluster
Stepbe
4: retained
matchingifthe
convection
center with
the preliminary
Thecloud.
cloud cluster
will
it contains
a convection
center,
or it will beconvective
labeled as cloud.
uncertain
will be
retained ifthe
it contains
a convection
center,by
oremploying
it will be labeled
as uncertain
cloud.
Step
5: eliminate
small area
of broken cloud
four pixels
as the threshold.
5: use
eliminate
the steps
smallabove
area oftobroken
by employing
Step 6:
the same
detect cloud
convective
cloud an four
hourpixels
before.as the threshold.
Step 6:
7: use
tracking
andsteps
matching
uncertain
cloud cloud
clusters
thebefore.
previous time t−1 and the
the same
abovethe
to detect
convective
an at
hour
current
time
t
0
.
Firstly,
a
grid
spacing
of
and
the
central
position
is1used
to express
m cloud
× n clusters at the previous P(i,
Step 7: tracking and matching the uncertain
timej)t−
and the
current
the
range of
the uncertain
cloud
where
m isposition
the longitude
andthe
n
timespace
t0 . Firstly,
a grid
spacing of
m ×clusters,
n and the
central
P(i, j) direction
is used tolength
express
is
the
latitude
direction
length.
We
then
search
the
current
time
cloud
cluster
in
the
space
range
space range of the uncertain cloud clusters, where m is the longitude direction length and n is
M =time
4× m
= 4 ×inn the
of
the central
P(i, j)search
at t−1,the
where
and N
. In space
the searching
× N with
the M
latitude
direction
length.point
We then
current
cloud
cluster
range of
area
t−1, we
the point
cloudP(i,
clusters
BT
min
increases
beyond
8
K
and
the
overlap
ratio
M × at
N with
thesave
central
j) at t−whose
,
where
M
=
4
×
m
and
N
=
4
×
n.
In
the
searching
area
1
beyond
50%.
Finally,
we
calculate
the
cross-correlation
coefficient
between
the
satisfying
cloud
at t−1 , we save the cloud clusters whose BTmin increases beyond 8 K and the overlap ratio beyond
clusters
at t−1 we
andcalculate
the current
uncertain cloud coefficient
cluster at tbetween
0. The cross-correlation
coefficient
is
50%. Finally,
the cross-correlation
the satisfying cloud
clustersr at
defined
t−1 and as:
the current uncertain cloud cluster at t0 . The cross-correlation coefficient r is defined as:
 (T (i, j ) - T )(T (i, j ) - T )
m
n
0
r=
0
-1
-1
i =1 j =1
 (T (i, j ) - T )  (T
m
n
2
0
i =1 j =1
m
n
0
i =1 j =1
−1 (i, j ) - T-1 )
2
,
(1)
Atmosphere 2017, 8, 42
6 of 11
m n
∑ ∑ T0 (i, j) − T0 T−1 (i, j) − T−1
i =1 j =1
s
,
r= s
m n
2
2 m n
∑ ∑ T0 (i, j) − T0
∑ ∑ T−1 (i, j) − T−1
i =1 j =1
(1)
i =1 j =1
where T− 1 and T0 are the BT matrices of two cloud clusters to be matched at t−1 and t0 , respectively.
If the highest cross-correlation coefficient we have calculated is beyond 0.35, we save the current cloud
cluster as convective cloud, or we will eliminate it.
When the detection process is accomplished, the convective cloud is divided into weak convection
(BTmin > 230 K), general convection (210 K < BTmin ≤ 230 K), and severe convection (BTmin ≤ 210 K) by
BTmin , and is divided into α(200 km ≤ L < 2000 km), β(20 km ≤ L < 200 km), and γ(0 ≤ L < 20 km)
type by L, where L is defined as:
p
L = m2 + n2
(2)
We can then obtain nine categories that represent the characteristics of different kinds of cloud
clusters as shown in Table 1.
Table 1. Classification standard of convective cloud clusters (L: km; BTmin : K).
L > 240
220 < L ≤ 240
L ≤ 220
200 < BTmin ≤ 2000
20 < BTmin ≤ 200
0 < BTmin ≤ 20
α-weak convection (α-SC)
α-general convection (α-GC)
α-severe convection (α-WC)
β-weak convection (β-SC)
β-weak convection (β-GC)
β-severe convection (β-WC)
γ-weak convection (γ-SC)
γ-general convection (β-GC)
γ-severe convection (γ-WC)
3. Application
In June 2016, southern China suffered from heavy rainfall affected by El Nino, causing some
casualties and serious economic losses. Figure 6 shows the FY-2F BTIR1 image at every half hour from
8:30 to 12:30 on 14 June, where two main convection systems are observed in the image. One is the
long Meiyu Front cloud belt in the Yangtze River Basin extending from 108◦ E to 140◦ E, another is
the larger-scale convection system in the Bay of Bengal, which is affected by the South Asian summer
monsoon region. At the same time, some isolated convection systems in small scale are also found
in the west plateau, Pacific Ocean and other regions. For these large-scale cloud systems, their BTmin
always never change much in short time. Different from the large-scale convection systems, most of
the isolated convective cloud only last for several hours. Their BTmin changes fast within an hour, and
will remain almost unchanged or increased when convective activities begin to disappear. To evaluate
the application performance of the integrated detection method, we use the 10:30 image as the case to
execute the step-by-step procedures, and track the detected clouds after each step.
Figure 7a shows that the severe convection centers, which distributed almost in the Meiyu Front
Rain Belt and the Bay of Bengal, and some severe convection centers are also located in the southern
part of the plateau and the Philippines. Figure 7b shows the preliminary convective cloud detected
with the BTIR1 and BTD process. Although mainly convective clouds are included in the result, many
cirrus clouds are also included; for example, some ”noise” areas where no convection activities emerge
in the summer monsoon cloud system of the Bay of Bengal. Figure 7c shows the severe convection
after matching the cloud in Figure 7a,b,d, showing the uncertain cloud, which is widely distributed in
the image. The uncertain cloud contains not only the convective cloud but some cirrus clouds such
as the small-scale convection in the northwestern Pacific Ocean. The result of the tracking process
is displayed in Figure 7e based on the former steps, where some dotted cloud “noises” have been
eliminated. Figure 7f is the integrated detection result combining Figure 7c,e, where some isolated
cloud clusters are also well detected. On the whole, the integrated detection process has a better
performance than any single-threshold method, especially for the isolated convective cloud.
Atmosphere 2017, 8, 42
Atmosphere
Atmosphere2017,
2017,8,8,42
42
7 of 11
77ofof77
Figure
image
at
every
Figure
6.6.(a–i)
FY-2F
BT
IR1
08:30–12:30.
10:30
isis
IR1
Figure6.
(a–i)FY-2F
FY-2FBT
BT
IR1image
imageat
atevery
everyhalf
halfhour
houron
on14
14June
June2016,
2016,08:30–12:30.
08:30–12:30.The
Theimage
imageatat10:30
10:30is
the
object
to
test
our
detection
process
in
this
paper.
the
theobject
objecttototest
testour
ourdetection
detectionprocess
processininthis
thispaper.
paper.
Figure 7. Cont.
Atmosphere 2017, 8, 42
Atmosphere 2017, 8, 42
8 of 11
8 of 8
Figure
Figure 7.
7. (a–f)
(a–f) Results
Results after
after each
each detection
detection step.
step. The
The pink
pink area
area on
on the
the image
imageisisthe
thedetection
detectionresult.
result.
In order
order to
to evaluate
evaluate the
the detection
detection result
result after
after each
each detection
detection step,
step, we
we counted
counted the
the number,
number, mean
mean
In
scale and
and mean
mean BT
BT of
of the
the clouds
clouds on
on Figure
Figure 7.
7. Table
Table 22 shows
shows that
that there
there are
are 19
19 severe
severeconvective
convective cloud
cloud
scale
clusterswith
withan
an average
averagescale
scale of
of 69.7
69.7 km,
km, and
and 61
61 un-severe
un-severe convective
convective cloud
cloud clusters
clusters with
with an
an average
average
clusters
scaleofof27.4
27.4km.
km.The
The
contrast
of their
average
suggests
that more
small-scale
cloud clusters
scale
contrast
of their
average
scalescale
suggests
that more
small-scale
cloud clusters
might
might
be
detected
out
by
using
the
integrated
detection
method
in
this
paper.
Due
to
the
fact
that
several
be detected out by using the integrated detection method in this paper. Due to the fact that several
severe convection
convection centers
centers may
may belong
belong to
to the
the same
same convective
convective cloud,
cloud, the
the number
number of
of cloud
cloud clusters
clusters is
is
severe
reducedfrom
from38
38to
to19
19after
after matching
matchingthe
the severe
severeconvection
convectioncenters
centersand
andpreliminary
preliminarydetection
detectionresults.
results.
reduced
In
the
154
uncertain
cloud
clusters,
there
are
only
61
un-convective
cloud
clusters.
This
means
that
In the 154 uncertain cloud clusters, there are only 61 un-convective cloud clusters. This means that
the tracking
tracking process
process using
using successive
successive time
time data
data has
has eliminated
eliminated 96
96 cloud
cloud clusters,
clusters, indicating
indicating that
that our
our
the
tracking
process
is
valid
to
distinguish
the
cirrus
clouds
from
convective
cloud.
The
severe
tracking process is valid to distinguish the cirrus clouds from convective cloud. The severe convection
convection
is lower
about than
9.8 Kun-severe
lower than
un-severeinconvection
in mean
termsBT,
of the
mean
BT,be
which
could
be
is
about 9.8 K
convection
terms of the
which
could
inferred
that
inferred
that there
more deepinconvection
in thecloud
large-scale
cloud
system.with
Compared
6, we
there
is more
deep is
convection
the large-scale
system.
Compared
Figurewith
6, weFigure
see that
in
see
that
in
the
integrated
detection
process,
the
BT
threshold
method
mainly
detected
the
severe
the integrated detection process, the BT threshold method mainly detected the severe convection in
convection
large
andprocess
the tracking
process
tended
to detect
more which
convection
which is growing.
large
scale, in
and
the scale,
tracking
tended
to detect
more
convection
is growing.
Table2.2.Cloud
Cloudstatistical
statisticalresult
resultofofeach
eachstep.
step
Table
Severe convection centers
Severe convection centers
Preliminary
detection
Preliminary
detection
result result
Severe
convection
Severe
convection
Uncertain cloud
Uncertain cloud
Un-severe convection
Un-severe
Integrated
detectionconvection
result
Integrated detection result
Number
38
38
173173
19 19
154
154
61
80 61
80
Number
Mean Scale (km)
Mean Scale (km)
20.6
20.6
23.9
23.9
69.7
69.7
18.3
18.3
27.4
27.4
37.4
37.4
Mean BTIR1 (K)
Mean BTIR1 (K)
204.8
204.8
225.2
225.2
223.5
223.5
232.5
232.5
232.3
232.3
225.1
225.1
According to the scale and the BTmin of the convective cloud, the detection result is classified and
According to the scale and the BTmin of the convective cloud, the detection result is classified and
statistically analyzed, as shown in Figure 8. In the classification result, the convection activities are
statistically analyzed, as shown in Figure 8. In the classification result, the convection activities are
mainly distributed in the mid–low latitude. There is more severe convection in the large cloud belt,
mainly distributed in the mid–low latitude. There is more severe convection in the large cloud belt,
while the general and weak convection do not have a distinct tendency in the distribution, for we could
while the general and weak convection do not have a distinct tendency in the distribution, for we
see them in any area. In the statistical result, it could be discovered that the total number of β-type
could see them in any area. In the statistical result, it could be discovered that the total number of
cloud clusters is the highest, followed by γ-type cloud clusters. In addition, there are only two α-SC
β-type cloud clusters is the highest, followed by γ-type cloud clusters. In addition, there are only two
clouds which exist in the summer monsoon cloud system of the Bay of Bengal. Among β-type cloud
α-SC clouds which exist in the summer monsoon cloud system of the Bay of Bengal. Among β-type
clusters, general convective cloud clusters have the highest number and the weak convective cloud
cloud clusters, general convective cloud clusters have the highest number and the weak convective
has the lowest number. In the γ-type cloud, a majority of general convective cloud is also found, but
cloud has the lowest number. In the γ-type cloud, a majority of general convective cloud is also
the number of severe convection is lower than that of weak convection, probably due to the fact that
found, but the number of severe convection is lower than that of weak convection, probably due to
the γ-type cloud tends to last for a short time and has difficulty attaining a very high cloud top height.
the fact that the γ-type cloud tends to last for a short time and has difficulty attaining a very high
cloud top height.
Atmosphere 2017, 8, 42
Atmosphere 2017, 8, 42
9 of 11
9 of 9
30
27
20
15
15
10
10
2
9
2
0
0
Figure 8.
8. Classification
Classification of
of convective
convective cloud
cloud clusters
clusters (a)
(a) and
and its
its statistical
statistical result
Figure
result (b).
(b).
4. Conclusions
Conclusions
4.
An integrated
cloud
detection
method
basedbased
on FY-2
of successive
An
integratedconvective
convective
cloud
detection
method
onVIRRS
FY-2 infrared
VIRRS data
infrared
data of
time
was
put
forward
in
this
paper.
We
employed
some
reasonable
thresholds
according
to the
successive time was put forward in this paper. We employed some reasonable thresholds according
threshold
test
and
designed
a
step-by-step
detection
process.
In
this
method,
both
the
BT
threshold
to the threshold test and designed a step-by-step detection process. In this method, both the BT
detection process
and
the tracking
areprocess
used toare
detect
of convective
cloud, and
threshold
detection
process
and theprocess
tracking
useddifferent
to detectkinds
different
kinds of convective
the
application
test
shows
that:
cloud, and the application test shows that:
(1) The
The integrated
integrated method
method is
is effective
effective for
for convection
convection in
in different
different scales and life periods, especially
the
isolatedconvective
convective
cloud.
However,
the large-scale
cloud itsystem,
it misinterpret
is easy to
the isolated
cloud.
However,
in theinlarge-scale
cloud system,
is easy to
misinterpret
some
cirrus
clouds
as
convective
cloud
as
well.
some cirrus clouds as convective cloud as well.
(2) The
detecting convective
convective cloud
cloud with
with aa low
low BT
BTmin
min more
more
(2)
The BT
BT threshold
threshold method
method was
was capable
capable of
of detecting
efficiently,
and
the
tracking
methods
are
more
capable
of
detecting
convection
which
efficiently, and the tracking methods are more capable of detecting convection which is
is
growing.
temperature change
change
growing. Also,
Also,the
the combined
combined use
use of
of overlap
overlap ratio,
ratio, minimum
minimum brightness
brightness temperature
and
and cross-correlation
cross-correlation coefficient
coefficient shows
shows aa remarkable
remarkable effect
effect on
on the
the elimination
elimination process
process for
for
non-convection
area.
non-convection area.
(3) The
cloud
clusters
detected
by by
thethe
integrated
method
are
(3)
Thestatistical
statisticalresult
resultshows
showsthat
thatthe
theα-type
α-type
cloud
clusters
detected
integrated
method
mostly
large-scale
cloud
systems,
andand
thethe
β- βand
highest
are mostly
large-scale
cloud
systems,
andγ-type
γ-typecloud
cloudclusters
clusters have
have the
the highest
proportion
of
general
convective
cloud.
However,
the
proportion
of
weak
convection
is
proportion of general convective cloud. However, the proportion of weak convection is higher
higher
than
that
of
severe
ones
in
γ-type
cloud,
but
it
is
the
opposite
in
β-type
cloud.
than that of severe ones in γ-type cloud, but it is the opposite in β-type cloud.
Atmosphere 2017, 8, 42
10 of 11
Acknowledgments: This research was supported by the National Nature Science Foundation of China (41576171).
We are grateful to the editors and anonymous reviewers for their constructive comments on the manuscript.
Author Contributions: Hanqing Shi designed the research and the general outline of the detection method.
Kuai Liang performed the experiments and wrote the manuscript. Pinglv Yang and Xiaoran Zhao revised
the manuscript.
Conflicts of Interest: The authors declare no conflicts of interest.
References
1.
2.
3.
4.
5.
6.
7.
8.
9.
10.
11.
12.
13.
14.
15.
16.
17.
18.
19.
Fujita, T.; Bradbnry, D.L.; Murino, C. A Study of Mesoscale Cloud Motions Computed from ATS-1 and Terrestrial
Photographs; R. Satellite Mesometeorological Research Project Research Paper No. 71; Department of
Geophysical Sciences University of Chicago: Chicago, IL, USA, 1968.
Baldauf, M.; Seifert, A.; Forstner, J.; Majewski, D.; Raschendorfer, M.; Reinhardt, T. Operational
convective-scale numerical weather prediction with the cosmo model: description and sensitivities.
Mon. Weather Rev. 2011, 139, 3887–3905. [CrossRef]
Long, Q.; Chen, Q.; Gui, K.; Zhang, Y. A case study of a heavy rain over the southeastern Tibetan Plateau.
Atmosphere 2016, 7, 118. [CrossRef]
Ma, Y.; Wang, X.; Tao, Z.Y. Geographic distribution and life cycle of mesoscale convection system in China
and its vicinity. J. Prog. Nat. Sci. 1997, 7, 583–589.
Tao, Z.Y.; Wang, H.Q.; Wang, X. A survey of meso-α-scale convection systems over China during 1995.
Acta Meteorol. Sin. 1998, 56, 166–177.
Zheng, Y.G.; Zhu, P.J.; Chen, M. General investigation and analysis of Meso-α-scale convection systems
in the Yellow Sea and surrounding areas from 1993 to 1996. Acta Sci. Nat. Univ. Pekin. 2004, 40, 66–72.
(In Chinese)
Bai, J.; Wang, H.Q.; Tao, Z.Y. Recognition and tracing of severe convective cloud from IR images of GMS.
J. Trop. Meteorol. 1997, 13, 158–167.
Csekits, C.; Zwatz, V.M.; Jann, A. Automatic detection of convective cells-a nowcast module at the
Austrian meteorological service. In Proceedings of the 2000 EUMETSAT Meteorological Satellite Data
Users’ Conference, Bologna, Italy, 29 May–2 June 2000; pp. 715–721.
Mueller, C.; Saxen, T.; Roberts, R.; Wilson, J.; Betancourt, T.; Dettling, S.; Oien, N.; Yee, J. NCAR auto-nowcast
system. Weather Forecast. 2003, 18, 545–561. [CrossRef]
Doelling, D.R.; Gambheer, A.V.; Stassi, J.C. Deep convective cloud calibration. In Proceedings of the
29th CERES Science Team Meeting, Hampton, VA, USA, 17–18 November 2003. Available online: https:
//ceres.larc.nasa.gov/documents/STM/2003-10/pdf/Doelling_DeepC.pdf (accessed on 11 July 2016).
Liu, J. Improvement of dynamic threshold value extraction technic in FY-2 cloud detection. J. Infrared
Millim. Waves 2010, 29, 288–292. (In Chinese)
Li, H.J.; Kong, Y.S. To fetch convective cloud cell with continuous wavelet transaction method. J. PLA Univ.
Sci. Technol. 2005, 6, 181–186. (In Chinese)
Chen, G.Y.; Ding, X.X.; Zhao, L.Y. An automatical pattern recognition technique of cloud based on fuzzy
neural network. Chin. J. Atmos. Sci. 2005, 29, 537–844. (In Chinese)
Li, Z.J. Estimation of cloud motion using cross-correlation. Adv. Atmos. Sci. 1998, 15, 277–282.
Liu, K.F.; Zhang, R.; Sun, Z.B. A cloud movement short-time forecast based on cross-correlation.
J. Image Graph. 2006, 11, 586–591. (In Chinese)
Liu, Y.A.; Wei, M. Application of FY-2C data in rainstorm nowcasting. In Proceedings of the First International
Conference on Information Science and Engineering, Nanjing, China, 26–28 December 2009; pp. 4766–4769.
Ohsawa, T.; Ueda, H.; Hayashi, T.; Watanabe, A.; Matsumoto, J. Diurnal variations of convective activity and
rainfall in tropical Asia. J. Meteorol. Soc. Jpn. Ser. II 2001, 79, 333–352. [CrossRef]
Vila, D.A.; Machado, L.A.T.; Laurent, H.; Velasco, I. Forecast and tracking the evolution of cloud clusters
(For TraCC) using satellite infrared imagery: Methodology and validation. Weather Forecast. 2008, 23, 233–245.
[CrossRef]
Arnaud, Y.; Desbois, M.; Maizi, J. Automatic tracking and characterization of African convection systems on
Meteosat pictures. J. Appl. Meteorol. 1992, 31, 443–453. [CrossRef]
Atmosphere 2017, 8, 42
20.
21.
22.
23.
11 of 11
Carvalho, L.M.V.; Jones, C. A satellite method to identify structural properties of mesoscale convection
systems based on maximum spatial correlation tracking technique (MASCOTTE). J. Appl. Meteorol. 2001, 40,
1683–1701. [CrossRef]
Huang, Y.; Liu, H.J.; Zhai, J.; Feng, Y. FY2E convective cloud detection method based on multi-thresholds.
Remote Sens. Technol. Appl. 2014, 29, 915–922. (In Chinese)
Shah, S.; Notarpietro, R.; Branca, M. Storm identification, tracking and forecasting using high-resolution
images of short-range x-band radar. Atmosphere 2015, 6, 579–606. [CrossRef]
Li, P.W.; Wong, W.K.; Chan, K.Y.; Lai, E.S. SWIRLS—An Evolving Nowcasting System; Technical Note, No. 100;
Hong Kong Observatory: Hong Kong, China, 2000.
© 2017 by the authors; licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons Attribution
(CC BY) license (http://creativecommons.org/licenses/by/4.0/).