Physical and statistical approach for MSG cloud identification
E. Ricciardelli *, F.
F. Romano*, V. Cuomo*
* Institute of Methodologies for Environmental Analysis, National
National Research Council
(IMAA/CNR), Italy
SEVIRI
MODIS
MODIS
INTRODUCTION
INTRODUCTION
Cloud
Cloud detection
detection is
is essential
essential to
to estimate
estimate accurate
accurate
atmospheric
A
atmospheric and
and geophysical
geophysical parameters.
parameters.
A
statistical
statistical and
and physical
physical approach
approach has
has been
been developed
developed
for
MSG
in
order
to
improve
cloud
identification
for
for MSG in order to improve cloud identification for
very
very thin
thin and
and very
very low
low cloud.
cloud. The
The statistical
statistical algorithm
algorithm
is
is aa pattern
pattern recognition
recognition technique
technique that
that uses
uses textural
textural
and
and spectral
spectral features,
features, itit was
was trained
trained on
on the
the basis
basis of
of
MSG
images
collected
for
different
seasons
MSG images collected for different seasons and
and
different
different regions.
regions. The
The physical
physical algorithm
algorithm is
is based
based on
on
dynamic
dynamic threshold
threshold tests
tests and
and itit doesn’t
doesn’t require
require any
any
ancillary
ancillary data.
data. IfIf the
the algorithms
algorithms do
do not
not agree
agree aa
decision
decision has
has to
to be
be taken
taken to
to decide
decide the
the final
final FOV
FOV flag.
flag.
The Spinning Enhanced Visible and Infrared Imager on board the
First Meteosat Second Generation have been available since
February 2003.
Its spatial resolution is 3 km at sub-satellite for all channels except
1km for the high resolution visible channel. The major
improvement is its enhanced spectral characteristic
that,
combined with its higher temporal resolution (15 minutes) allows
an accurate cloud cover analysis.
Spectral
VIS 0.6 VIS 0.8 NIR 1.6
band
1.64
Central
0.635 0.81
Wavelength
(µ m)
IR 3.9
WV 6.2 WV 7.3 IR 8.7
IR 9.7
IR 10.8 IR 12.0 IR 13.4 HRV
3.92
6.25
9.66
10.8
7.35
8.7
12
13.4
0.75
The
The Moderate
Moderate resolution
resolution Imaging
Imaging Spectroradiometer
Spectroradiometer on
on board
board
AQUA
AQUA and
and TERRA
TERRA polar
polar EOS
EOS satellites,
satellites, provides
provides global
global
observations
observations in
in VIS
VIS and
and IR
IR region
region of
of the
the spectrum
spectrum ..
ItIt measures
µm)
m)
measures radiances
radiances at
at 36
36 wavelength(
wavelength( from
from 0.4
0.4 to
to 14.5
14.5 µ
at
at 250m
250m spatial
spatial resolution
resolution in
in two
two visible
visible bands,
bands, 500m
500m resolution
resolution
in
in five
five visible
visible bands
bands and
and 1000m
1000m resolution
resolution in
in the
the infrared
infrared bands.
bands.
Statistical approach
Physical approach
The classification algorithm is based on a K-nn pattern recognition technique and it uses textural and
spectral features estimated in boxes 3x3. Spectral and textural features characterizing each pixel are
extracted for IR SEVIRI radiance data at 3.9µm ( particularly useful in the detection of low level water
clouds) 8.7 µm, 10.8 µm, 12 µm. The features are shown in the table:
The large number of spectral and
Spectral features (describe the grey level range)
Maxima
textural features can be reduced since not
Minima
Ratio between maxima and minima
all of them are useful. The Fisher distance
Mean
is the feature selection criterion:
Textural features (describe aspects of the variability of a scene and are computed from the
grey level differences between each individual pixel and its neighbors)
Dijk = ν ij − vik /(σ ij + σ jk ) . This Description of the
P( d , ϑ ) is the grey level difference histogram,
Contrast
parameter measures the classes
d
1
d is the difference between the grey
C (d ,ϑ ) =
∑ P( d , ϑ ) N
256
levels separated in direction ϑ (0º,45º,90º,
ability of the variable vi Clear over land
Entropy
135º) by the distance ρ = 1 .
Clear over water
to differentiate class
P( d , ϑ ) ⎞
⎛ P( d , ϑ ) ⎞
P( d , ϑ ) denotes the number of times
ENT ( d ,ϑ ) = −∑ ⎛⎜
⎟ log ⎜
⎟
Low clouds
N
N
⎝
⎠
⎝
⎠
difference d occurs between pairs of grey
Ck from class C J .
level at the specified distance and
Cumulus
Angular Second Moment
orientation in the box 3x3.
K nearest neighbor
⎛ P ( d ,ϑ ) ⎞
N
is
the
total
number
of
pairs
of
points
in
the
High stratiform clouds
ASM ( d ,ϑ ) = ∑ ⎜
⎟
classifier
⎝ N ⎠
box separated by distance ρ = 1 in direction
ϑ.
Mean
The
class
likelihoods
have
been estimated
1
d
M (d ,ϑ ) =
∑ P (d ,ϑ ) N
r
K
as follows:
256
P(ν , C) = i
The algorithm is based on multispectral threshold technique applied to each pixel of the image.
The test depends on the solar illumination (day time, night time, sun glint) and on the viewing angles.
The thresholds have been determined on the basis of training database which gather pixels manually
selected and labeled as cloud free.
Description of test seque nces:
over land, daytime
•
•
•
•
•
over se a, daytime
T10.8 < Tresh_ T10.8
R0.6 >Tresh_R 0.6
T10.8 – T12 > Tresh_ T10.8 T12
T8.7 – T10.8 > Tresh_ T8.7 T10.8
T10.8 – T3.9 > Tresh_ T10.8 T3.9 and T8.7 –
T10.8 >-4.5 -1.5(1/cos(asat)-1)
•
•
•
•
•
•
T10.8 < Tresh_ T10.8
R0.8 >Tresh_R 0.8
R1.6 >Tresh_R 1.6
T10.8 – T12> Tresh_ T10.8 T12
T8.7 – T10.8 > Tresh_ T8.7 T10.8
T10.8 – T3.9> Tresh_ T10.8 T3.9
2
255
over land, night time
over se a, night time
d =0
•
•
•
•
•
T10.8 < Tresh_ T10.8
T10.8 – T12 > Tresh_ T10.8 T12
T8.7 – T10.8 > Tresh_ T8.7 T10.8
T10.8 – T3.9 > Tresh_ T10.8 T3.9 and T8.7 –
T10.8 >-4.5 -1.5(1/cos(asat)-1)
T3.9 – T10.8 > Tresh_ T3.9 T10.8
•
•
•
•
•
T10.8 < Tresh_ T10.8
T10.8 – T12> Tresh_ T10.8 T12
T8.7 – T10.8 > Tresh_ T8.7 T10.8
T10.8 – T3.9> Tresh_ T10.8 T3.9
T12.0 – T3.9> Tresh_ T10.8 T3.9
255
e
d =0
2
255
d =0
For the i-th test the probability that the pixel is clear , Pi ,clear , or cloudy, Pi ,cloud , will be determinate.
If the i-th test is successful (the pixel is cloudy) the Pi ,cloud is determined on the basis of the distance
between the real value and the threshold:
Pi ,cloud =
255
d =0
The area-averaged Robert Gradient is a measure of edge strength and it is maximum in
regions of sharp brightness contrast.
M −ρ N −ρ
1
RG =
∑ ∑ I (m, n ) − (I (m + ρ , n + ρ ) + I (m + ρ , n) − I (m, n + ρ )
( M − )( N − )
DiffTreshi
DiffTreshMaxi , cloud
Otherwise, if the i-th test is not successful, it’ll be
Pi,clear =
If
DiffTreshi
DiffTreshMaxi , clear
>0
then Pphysical,cloud = max(Pi,cloud ) ,
∑P
=0
it’ll be Pphysical,clear = min(Pi,clear )
i ,cloud
Finally, we are able to classify the pixel as cloudy if
i
and
The DiffTresh is the difference between the threshold and the brightness temperature or
the brightness temperatures difference. For each test, DiffTresh max values have been calculated
on a sample of MSG imagery collected in different regions of the earth and at different times.
DiffTreshMax has been defined as the mean of max values.
Results
MSG cloud mask has been validated against MODIS cloud mask (Ackerman
et al.,1998) and compared with SAFNWC cloud mask (Schroder et al.,2002).
MODIS cloud mask is collocated within SEVIRI footprint. SEVIRI FOV is
declared clear if a fixed percentage (70%, 90%,100%) of MODIS pixels within
SEVIRI FOV has been determined to be clear. In result interpretation it should
be observed that MODIS cloud mask not always detects cloudy pixels
correctly. Moreover it sometimes classifies as uncertain pixels that CMESP
detects correctly. The blue ring surrounds some pixels that SAF cloud mask
Percentage of MODIS
detects wrongly.
FOVs detected exactly
pixels determined to be
Total number of clear
Percentage of MODIS
pixels determined to be MODIS pixels used
clear within SEVIRI FOV for validation
Total number of cloudy
clear within SEVIRI FOV
MODIS pixels used for
validation
70 %
90%
100%
2225386
2296947
2386948
1546074
1487811
1416786
MODIS pixels have been collocated at SEVIRI spatial resolution
70%
90%
100%
NiV
}
m =1 n =1
When the Euclidean distance measure is used, some features (the ones with the largest variance
across the design set) tend do dominate this measure so it has been necessary to normalize
the features.
i
if
ρ
i =1
∑P
i ,cloud
{
where Ni is the number of samples in
class Ci and K i are the samples closest to the
B( m, n) is the measured brightness at coordinate (m,n) in the MxN (3x3) box and ρ is the
r
separation distance (in this case ρ=1)
features vector ν , V is the volume in which
K
i
those reside. In order to define the distance between two points in the feature space and calculate
the volume V, we have assumed that the feature space is a metric space and the dfunction which
r
r
r r
expresses the distance between ν 1 and ν 2 is the Euclidean distance : d (ν 1 ,ν 2 ) = (∑ (ν 1,i −ν 2,i ) 2 )1/ 2
ρ
Cloud Mask CMESP
Cloud Mask SAF
cloudy
90.4%
90.7%
89.7 %
cloudy
88.3%
88.5%
87.5%
clear
93.6%
95.2%
94.6%
clear
92.8%
94.1 %
93.5 %
r
r
Pknn,cloud = max(P(ν , Clow _ cloud ), P(ν , Ccumulus ), P(ν , Chigh _ stratiform
. ))
Otherwise, the pixel is clear if
and
r
r
r
r
s
r
( P(ν ,C clear ) < P(ν , Clow _ cloud )) ∨ ( P(ν , Cclear ) < P(ν , Ccumulus )) ∨ ((P(ν , Cclear ) < P(ν ,C High _ stratiform))
r
r
Pknn,clear = P(ν , Cclear )
.
Summary
Ensemble of Statistical and Physical
methods (CMESP)
The
physical
and
statistical
approach
independently, then the individual decisions
matched; if the two methods do not agree, the
FOV flag will be clear if Pphysical ,clear > Pknn ,cloud or
Pknn ,clear > Pphysical ,cloud
If the difference between the probabilities is lower
5%, a further test on the Robert Gradient ( R G ) ,
has to be done:
RG > R Gtreshold the pixel is cloudy.
If
R Gtreshold is a dynamic threshold.
r
r
r
r
r
r
( P(ν ,C clear ) > P(ν , Clow _ cloud )) ∧ ( P(ν , Cclear ) > P(ν , Ccumulus )) ∧ ( P(ν , Cclear ) > P(ν ,C High _ stratiform))
In this work, a physical and statistical approach has been developed for
run
are
final
MSG and has been combined in order to improve cloud identification of
very low and thin clouds. This approach does not need ancillary data which
are not always available. The two approaches compensate each other. The
statistical algorithm is very useful in low clouds and high stratiform clouds
than
detection, while the physical one is essential in discriminating clouds from
snow. Moreover, physical approach can be decisive in detecting clear pixels
that sometimes are classified cloudy by the statistical method, because of
SEVIRI ch 4, 3.9 µm
their very high entropy or contrast, especially over not homogeneous land.
Comparison between SEVIRI and MODIS at SEVIRI resolution cloud mask
References
2005-10-07 13:00
M. Derrien,H. L. Gleau, 2005, MSG/SEVIRI cloud mask and type from SAFNWC,International Journal of
Remote Sensing, Vol..26, No. 21,pp. 4707-4732;
J.Parikh, 1977, A comparative Study of Cloud classification Techniques,Remote Sensing of environment,
Vol. 6,pp.67-81;
E. Ebert,1988, Analysis of Polar Clouds from Satellite Imagery Using Pattern Recognition and a Statistical
Cloud Analysis Scheme, Journal of Applied Meteorology, Vol.28, pp. 382-399
M. Schroder, R. Bennartz, 2002, Generating cloudmasks in spatial high-resolution observation of cloud
using texture and radiance information, Int. Journal of Remote sensing, vol.23,no 20, pp.4247-4261
S. Ackerman, K. Strabala, W. P. Menzel,1998, Discrimating Clear –Sky from clouds with
Modis, Journal of Geophysical Research;
R. Haralick, K. Shanmugam,1973,Textural Features for Image Classification, IEEE Transactions on
Systems, Man and Cybernetics, vol.3 , no.6,
a) SEVIRI Channel 9, 10.8 µm
b) SEVIRI Cloud Mask CMESP
cloudy
clear
c) MOD IS Cloud Mask
cloudy
clear
uncertain
d) SEVIRI Cloud Mask SAF
cloudy
clear