Evaluation of model generated cloud cover by means of satellite data

ATMOSPHERIC
RESEARCH
ELSEVIER
Atmospheric Research 39 (1995) 91-111
Evaluation of model generated cloud cover by
means of satellite data
N. M61ders a,., M. Laube
b,
E. Raschke c
"LlM-Institutfiir Meteorologie, Universitiit Leipzig, Stephanstrafle 3, D-04103 LeipzigGermany
Institut fiir Geophysik und Meteorologie, Universitiit zu Kb'ln, Albertus-Magnus-Platz 1,
D-50923 KOlnGermany
c GKSS, Forschungszentrum Geesthacht, Institutfiir Physik, Max-Planck-Strafle, D-21502 GeesthachtGermany
Received 6 April 1994; accepted 25 October 1994
Abstract
An automated cloud retrieval algorithm has been developed and applied to determine cloud cover
from NOAA9 AVHRR (Advanced Very High Resolution Radiometer) satellite data. This satellite
derived cloud cover is used to evaluate the model generated cloud cover provided by two different
cloud cover parameterization schemes established in a 3-D-chemical transport model. In the standard
version of this model cloud cover depends on rain rate for raining clouds and on the relative humidity
at cloud base for fair weather clouds. In the second cloud cover parameterization scheme predictions
of liquid water content and ice content in combination with values of water and ice content derived
from several observations are used to generate the cloud cover by the model. Partial cloudiness is
allowed to form when mesoscale relative humidity is less than 100%. The comparison of the model
generated with the satellite derived cloud cover shows that the second cloud cover parameterization
scheme substantially improves the determination of cloud cover by the model.
1. Introduction
Clouds regularly cover about 50% of the sky (Lee et al., 1992). They modulate radiative
transfer, photolysis (e.g., Madronich, 1987), and, hence, the gas phase chemistry below,
within and above clouds. Furthermore, clouds play an important role in the vertical redistribution, chemical transformation and removal of trace species in the troposphere (e.g.,
W a l c e k and Taylor, 1986). In an attempt to understand transport of pollutants as well as
gas phase and aqueous chemistry processes taking place within the troposphere, several
long-range chemical transport models have been developed in the last decade (e.g., Chang
* Corresponding author.
0169-8095/95/$09.50 © 1995 Elsevier Science B.V. All rights reserved
SSDIOI 69- 8095 ( 9 4 ) 0 0 0 5 9 - X
92
N. M6lders et al. /Atmospheric Research 39 (1995) 91-111
Table 1
Threshold values of relative humidity, -q, used in various models
Model
77
remarks
UK
0.85
ECMWF
0.8
for some cloud types
DNMI
0.85
over sea
UHAM GCM
0.75
0.8
over land
non-convectiveclouds
0.8,0.7,
0.6
low, middle, high
convectiveclouds
height dependent
AFGL GCM
BMRC GCM
0.6
high, low level clouds
0.55
mid-level clouds
source
Dickinson
(1985)
Slingo
(1987)
Sundqvist
etM.
(1989)
Heiseand
Roeckner
(1990)
Lee et al.
(1992)
Colman et
al.
(1994)
et al., 1987). In such models clouds are of subgrid scale and have to be parameterized for
which several assumptions and simplifications are necessary. It is obvious that improved
chemical transport models also demand a cloud cover parameterization scheme to determine
cloud cover.
Unfortunately, one of the most sensitive quantities to parameterize is the cloud amount.
Clouds show a high variability in space and time, as well as in type, in origins, processes
of formation and development.
In numerical models of the meso-a scale, some diagnostic relations are necessary to
derive cloud cover from large scale variables. Following Smagorinsky (1960), most modelers have relied on large scale relative humidity, h = qv/qs, to diagnose nonconvective
cloudiness (Buriez et al., 1988). A tendency to parameterize the cloud cover, b, as a function
of relative humidity can be found in recent models (e.g., Dickinson, 1985; Slingo, 1987;
Saito and Baba, 1988; Sundqvist et al., 1989; Lee et al., 1992). This function is usually
formulated similar to (e.g., Saito and Baba, 1988; Colman et al., 1994)
b(h)=
0
(h_~)n/(l_h)
h<r/
h>~?,
(1)
where n ranges from 0.5 to 3 depending on the model, r/is an arbitrarily chosen threshold
value of relative humidity above which a cloud is assumed to form. This threshold value is,
usually, a characteristic quantity of the model (Table 1 ), i.e., the value is adjusted depending
on the horizontal and vertical grid size. Empirical relationships between the cloud amount
of low-, middle- and high level clouds and relative humidity have been derived by several
authors for different models (e.g., Slingo, 1987; Kvamsto, 1991; Lee et al., 1992). Some
authors have included a dependence on the potential temperature (e.g., Slingo, 1980), on
N. M6lders et al. / Atmospheric Research 39 (1995) 91-111
93
the vertical wind (e.g., Hense and Heise, 1984) or on the precipitation rate (only for
precipitating clouds; Chang et al., 1987) to determine cloud cover. Other parameters that
might be used to diagnose cloud cover are convective activity, atmospheric stability, surface
fluxes and wind shear. Recently, new parameterization schemes of stratiform clouds have
been developed by Sundqvist et al. ( 1989 ) and Smith (1990) which predict the condensated
water content and the cloud cover in a consistent manner.
The meso-a scale model used in our study is the EURAD package (EURopean Acid
Deposition model; Hass et al., 1990, 1993). Herein, cloud cover is parameterized depending
on the rain rate for raining clouds (Chang et al., 1987) and on the relative humidity at cloud
base for fair weather clouds. As shown by M61ders et al. (1994) this cloud cover parameterization scheme leads to poor results, especially under conditions when glaciated clouds
are to be simulated. In our study a new cloud cover parameterization scheme is presented
which was developed for this meso-c~ scale model package by M61ders (1993). This cloud
cover parameterization is based on cloud liquid water, ice content, cloud vertical extension
and relative humidity. In order to evaluate the accuracy of these two cloud cover parameterization schemes the model generated cloud cover and cloud fields are compared with
those analyzed from NOAA9 AVHRR satellite data. In doing so, the satellite data are
processed using a cloud retrieval algorithm which has been developed to classify the pixels
into cloudy or non-cloudy. Then the "observed" cloud cover is determined for each model
grid box.
2. Cloud cover parameterization schemes
In the EURAD model package the meteorological field quantities are predicted by the
meteorological preprocessor MM4 (Mesoscale Meteorological Model Version 4; Anthes
et al., 1987) which can be run with different options for the hydrological cycle (cf. Anthes
et al., 1987; M61ders et al., 1994). Usually a Kuo-type cumulus parameterization scheme
(Anthes et al., 1987) is used where cumulifrom clouds are assumed to form if sufficient
instability and horizontal moisture convergence are present above a grid area. Applying this
cloud parameterization scheme MM4 provides 3-D-fields of temperature, wind and water
vapor mixing ratio as well as the surface temperature, the sea level pressure and the rain
rates. No liquid water or ice is predicted. Another option to treat the hydrological cycle in
MM4 is the ice parameterization scheme developed by M/31ders (1993), which predicts
additionally to the meteorological quantities noted above, the mixing ratios of cloud water,
rainwater and ice as well as precipitation rates of snow (see M61ders et al., 1994). The
meteorological field quantities serve as input for the CTM (Chemistry Transport Model;
Hass et al., 1990; Hass et al., 1993) which is the chemical part of the model package
mentioned above and which predicts the transport and transformation of trace constituents
as well as wet and dry deposition. This model is based on RADM (Regional Acid Deposition
Model; Chang et al., 1987). In CTM cloud cover is required for the calculation of the
photolysis rates (also called photolysis frequencies) and wet deposition. These facts emphasize the importance of having a realistic representation of clouds, cloud distribution and
cloud cover.
94
N. M61derset al. /Atmospheric Research 39 (1995) 91-111
2.1. Standard cloud cover parameterization scheme
In the standard version of CTM a cloud module (Walcek and Taylor, 1986) is applied
wherein cloudiness is parameterized similar to the cumulus parameterization scheme of
MM4 using the water vapor, temperature and rain data delivered by the meteorological
preprocessor. Note that the mixing ratios of cloud water, rainwater and ice predicted by
MM4 with the ice parameterization scheme are ignored by this CTM version. In accord
with the eqs. ( 2 3 ) - ( 3 0 ) of Chang et al. (1987) the cloud cover, B, of a raining cloud is
determined by
B = ('/',.Pr) / ( EOxs),
(2)
where Qxs is the amount of excess water within a cloud generated during its lifetime %
(usually 1 h), • = 0.3 is the storm efficiency and Pr is the precipitation rate predicted by the
meteorological preprocessor. The cloud cover of fair weather clouds only depends on the
relative humidity at cloud base, hb (Hass et al., 1993)
B=4hb-3.
(3)
Fair weather clouds higher than the 650 hPa level and also those with a'relative humidity
at cloud base of less than 0.75 are rejected.
2.2. N e w cloud cover parameterization scheme
This cloud cover parameterization needs the mixing ratios of rainwater, cloud water and
ice predicted by MM4 with the ice parameterization scheme as input data. This means that
in contrast to the standard version of the model these quantities are not neglected.
The mixing ratios predicted by the meteorological preprocessor are representative values
for a grid volume of 80 km × 80 km in the horizontal and of some dekameters to some
kilometers in the vertical directions. Therefore, the actual in-cloud mixing ratios are higher
than these representative values. Assuming that in a model layer k the cloud always ranges
from the bottom to the top of this model layer and that the horizontal cloud extension does
not vary with height within that model layer, the predicted water content can be interpreted
as the product of the actual in-cloud water content, L', and the cloud cover, bk, of the grid
box at the level k (M~51ders, 1993)
L'bk =Lk + Ik.
(4)
Herein, Lk and Ik are the liquid water and ice content in the kth model layer calculated from
the mixing ratios of cloud water, rainwater and ice.
Rearranging Eq. (4) the cloud cover of layer k can be calculated if the water content, L',
is known. Since clouds vary according to their form, vertical extension, altitude, particulate
composition, ice and water content, L' is chosen individually for the different cloud types
considering statistics of observations by Matveev (1984). According to its main characteristics the cloud class or type of a simulated cloud is determined and used to derive the value
L' from lookup tables (Tables 2-6) which are based on the observations noted above. For
cumuliform clouds that do not match the conditions given in Table 2 cloud cover is calculated
by Eq. (3). Since cirriform clouds are often of subgrid scale with respect to the vertical
N. MOlders et al. / Atmospheric Research 39 (1995) 91-11l
95
Table 2
Lookup table of cloud characteristics as used in the new cloud cover parameterization scheme. The functions
f(Az) are explained in Tables 3-6
Cloud class water content
( g / m 3)
cloud extension
(kin)
cloud base
(km)
Cb
Cu
As/Ac
As
Ns/St
Ns/As
Ns
St/Sc
Ci, Cs, Cc
Ci
3.8
2.5
3.0
1.1
2.3
2.3
2.7
0.65
variabl&
variable"
>_ 1.3
0.9-2.5
_>2.0
_>2.4
_>2.4
_>2.4
_> 1.0
> 0.6
zh
_>5.7
1.0
f( Az )
0.1
f(Az)
f(Az)
f(Az)
0.1
J(Az, T)
j[zt,)
0.06
cloud top (km)
_>5.3
< 4.0
<4.7
<4.7
<4.7
--<4.7
_<5.0
_< 1.8
-< 10.5
variable"
further characteristics
RR> 0.1 mm/h
RR-<O.I mm/h
RR<_O.1m m / h
1> 0.001 g / m 3
RR-<O.I m m / h
RR> 0.1 m m / h
RR-<O.Im m / h
see Table 6
aHere, see text.
resolution ( I-3 km) of the upper model layers, the fact that a modeled cloud consists
prevailingly or totally of ice is taken as the last criterion for its classification as cirrus.
Since stratiform clouds are strongly related to relative humidity (e.g., Kvamsto, 1991 ) a
dependency on relative humidity, h, is also considered in the parameterization (MOlders,
1993)
bk = 0.75 (Lk + 1,0/L' + 0.25 ( 2 ( h - 0.85) )3.
(5)
Moreover, it is assumed that the cloud cover of a predicted cloud may not be lower than
one octa. Assuming maximum overlapping the total cloud amount, B, within a grid column
is given by (Sundqvist et al., 1989)
Table 3
Lookup table for mean water content ( g / m 3) in stratus and stratocumulus clouds according to Matveev (1984)
Cloud thickness (m)
100
150
200
250
300
350
400
450
500
550
600
temperature (°C)
-15to-10
-10to-5
-5 to0
0to5
5 t o 10
0.02
0.04
0.05
0.07
0.08
0.09
0.10
0.11
0.12
0.14
0.15
0.03
0.05
0.07
0.09
0.1 I
0.13
0.15
0.16
0.17
0.19
0.20
0.04
0.08
0.11
0.14
0.16
0.18
0.20
0.22
0.24
0.26
0.28
0.06
0.10
0.15
0.19
0.22
0.24
0.26
0.29
0.31
0.33
0.35
0.08
0.15
0.19
0.23
0.26
0.29
0.32
0.34
0.36
0.38
0.40
96
N. M61ders et al. / Atmospheric Research 39 (1995) 91-111
Table 4
Lookup table for mean water content (g/m 3) in stratus and stratocumulus clouds according to Matveev (1984)
for temperatures that are not within the range demanded in Table 3
Cloud thickness (m)
100-200
200-300
300-400
400-500
500~500
> 600
Height above cloud base (m)
< 100
100-200
200-300
300--400
> 400
0.11
0.11
0.10
0.11
0.13
0.12
0.16
0.18
0.19
0.18
0.19
0.19
0.24
0.25
0.25
0.26
0.24
0.30
0.31
0.31
0.31
0.32
0.33
0.32
Table 5
Lookup table for mean water content (g/m 3) of nimbostratus and altostratus clouds according to Matveev (1984)
Cloud thickness (m)
< 500
500-1000
1000-2000
2000-3000
3000--4000
Water content (g/m 3)
0.14
0.19
0.25
0.39
0.42
Table 6
Lookup table for mean ice content (g/m 3) in ice clouds according to Matveev (1984)
Height above ground (km)
Ice content (g/m 3)
<1
1-2
2-3
3-4
4-5
>5
0.170
0.088
0.032
0.026
0.016
0.006
B = 1 - I~1.1 - m a x ( b k _ l,bk)
k=~
i ---~k- i
(6)
3. Cloud c o v e r d e r i v e d f r o m satellite data
T o d a y there are satellite data that afford the possibility to obtain cloud cover data both
over land and sea ( K v a m s t o , 1991 ). In this study high resolution data of the A V H R R on
board of N O A A 9 serve to evaluate the cloud cover parameterization schemes described
above. The A V H R R is a five c h a n n e l instrument, providing in a cross track s c a n n i n g m o d e
data of u p w a r d radiances with a spatial resolution of about 1.1 km × 1.1 k m at the subsatellite
point. Each scan line consists o f 2048 pixels for each sensor. Table 7 lists the spectral
intervals of the radiometer. The spectral bands of channel 1 measure in the visible spectrum.
Herein, all clouds except haze increase the reflectance of the system E a r t h - a t m o s p h e r e when
they are located over sea or land (see Fig. l a ) . This m e a n s that the spectral values of the
ground (dark colors in Fig. l a ) differ from that of the clouds (light colors in Fig. l a ) .
N. Mi~lders et al. / Atmospheric Research 39 (1995) 91-111
97
Table 7
Spectral sensitivities of the NOAA9 AVHRR and their exploitation for cloud detection
Channel
Spectral range (/zm)
Measuring
Detection of
1
0.580-0.68
0.725-1.10
3.550-3.93
10.500-11.30
11.500-12.50
solar, visible
solar, near infrared
solar and terrestrial
terrestrial radiation
terrestrial radiation
clouds, snow, land, water
clouds, snow, land, vegetation
water clouds
ice-, water clouds
ice-, water clouds
2
3
4
5
Channel 2 measures in the near infrared region. Its sensor is sensitive for vegetation, for
which it can be used for the distinction between cloudfree vegetated land and clouds.
Channel 3 receives a mixed solar and terrestrial signal, while channels 4 and 5 measure in
the infrared region (Table 7). Fig, lb shows the spectral values obtained by channel 4 after
being converted to temperature values. Since temperature usually decreases with height
cloud tops are cooler than the cloudfree Earth's surface. In Fig. lb the colder clouds appear
in brighter colors than the warmer Earth's surface. Due to the spectral differences of land,
water, snow and clouds in the different channels a cloud detection can be carried out as a
Fig. 1. NOAA9 pass over Europe on 27th April 1986, 1325 GMT encompassing in the lower left corner North
Africa and Spain, at the upper left comer Island, in the upper right comer Scandinavia and in the lower right comer
Italy and Greece. (a) Image of channel 1 data of the AVHRR. Light colors represent high radiances (clouds and
snow). Dark colors represent low radiances (land and water). (b) Image of channel 4. Dark tones represent high
radiances which are representative for land and water surface temperatures. High and low clouds are presented in
white and grey, respectively.
98
N. MiJlders et al. /Atmospheric Research 39 (1995) 91-111
Fig. 1. (continued).
multichannel technique (e.g., Raschke et al., 1986). A simple cloud retrieval algorithm
using a threshold technique has been developed to analyze the satellite data. Such a threshold
technique offers a reasonable compromise between efficiency and accuracy (Inoue, 1987).
In a threshold method each imagery pixel is taken to be overcast if the properties observed
for the pixel differ from that observed under cloudfree conditions by certain threshold values
(Chang and Coakley, 1993).
First, the data measured by the radiometer were converted into reflectances for channel
1 and 2, Pl, P2 and into equivalent blackbody temperatures for channel 4, T4, respectively.
Since the spectral responses of channel 4 and 5 are very similar only channel 4 data are
used in the cloud retrieval algorithm. The reflectance and emittance properties of ice clouds
and water clouds differ the strongest at 3.7/xm, for which, the data of channel 3 can be used
for the distinction of water clouds and ice clouds (e.g., Raschke et al., 1986). The consideration of these data often presents difficulties for an automated procedure, because clouds
may appear warmer than the ground or alternatively colder, depending on their size and
depth as well as on sun elevation (Phulpin et al., 1983). Moreover, some data of channel 3
were missing or erroneous. In view of these aspects the data of channel 3 were not considered.
So-called pseudo images [difference between the albedo derived from the measurements
of channel 2 and 1, P2 - Pl; ' 'vegetation index" = (p2 - P~) / (P2 + Pl ) ] are calculated as
additional tools for the scene identification (Fig. 2). Often the difference between the
radiance temperatures of channel 4 and 5 is utilized in the classification of thin cirrus clouds
(e.g., Raschke et al., 1986; Yamanouchi et al., 1987). Here, this pseudo image is renounced
since the differences and, hence, the threshold value strongly depends on sun zenith angle.
Threshold values were derived from I-D- and 2-D-histograms over regions of known
class (Table 8). Such histograms were calculated for different regions of the satellite passage
N. MOlders et al. / Atmospheric Research 39 (1995) 91-111
99
Fig. 2. As Fig. 1, but for the (a) difference of albedo derived from channel 2 and 1. The difference is positive for
land (light color ) and negative for water (light grey) and snow (black). Clouds appear dark grey. (b) Vegetation
index. Cloudfree vegetated areas appear bright. Water is presented in black, clouds are given in grey.
and for all data sets available to take into account temporal and regional differences. To
receive statistically significant threshold values all considered areas were chosen to include
more than 30 pixels. The threshold value for a certain scene is determined from the left side
(lowest value) of the m a x i m u m in the histogram calculated for this scene. The minimum
value between the m a x i m a in bimodal histograms determined over areas that contain both
ground and clouds are fixed as the threshold value for clouds.
The analyses of the satellite data works pixelwise, i.e., no neighbored pixels are taken
into account. Assuming that a pixel belongs homogeneously to a scene, every pixel is either
classified according to its characteristics or rejected as unclassifiably. Herewith, a systematic
error may arise due to mixed scenes because a pixel can hold information from the reflected
radiation of different scenes. This is likely to occur for certain cloud types, e.g., cumulus
clouds, optically thinner clouds or the edges o f stratus clouds. In such cases the infrared
radiation from the cloud top combines with the radiation from the ground and produces a
grey level (indistinguishable) signal. Moreover, the reflectance is low and does not allow
to distinguish the cloud from the ground. Furthermore, the radiometric value of a mixed
scene pixel can yield the value of an other scene. Here, the combination of the different
characteristics (Table 8) will avoid misclassification in most of the cases so that errors from
this may be rare. Mixed scene pixels will occur more frequently at the edges of the passage,
where the area represented by the pixel is the largest due to the earth curvature and the
geometric conditions of the system Earth-satellite. Further error sources are the direction
dependency of the solar reflection and the terrestrial emittance, calibration errors (Raschke
N. MiJlders et al. / Atmospheric Research 39 (1995) 91-111
100
Fig. 2. ( c o n t i n u e d ) .
et al., 1988) as well as shadows of clouds and mountains. The effect of mixed pixels on the
cloud classification and on the satellite derived cloud cover will be discussed later.
Defined characteristics or unequivocal combinations of the different spectral compounds
of the scenes have been searched which allow to classify the pixels into different classes
(Table 8). First it is checked if the pixel belongs to the class land. The vegetation index is
positive for cloudfree land covered with vegetation while it is negative for all other scenes.
To ensure the classification as land also over regions with poor or none vegetation (e.g.,
Africa, Iberian peninsula, etc.), in doubtful cases, the last criterion to classify a pixel as
land is that the pixel has an albedo, Pl, of less than 15%, because the reflectance of land is
the smallest for channel 1 and, hence, the contrast to the clouds is higher than in channel 2
data. Pixels not classified as land are examined if they belong to the class water. The
difference P 2 - Pl is slightly negative for sea (e.g., Sakellariou and Leighton, 1988), but
Table 8
Threshold values for multispectral cloud analyses. Pt,2 is the albedo derived f r o m channel 1 and 2, "1"4is the
radiation temperature derived f r o m channel 4 data, respectively
Ocean
Land
Snow/ice
W a t e r clouds
Clouds
Pl ( % )
P2 ( % )
P2-P~ ( % )
Vegetation index
T4 (K)
< 15
< 15
>_ 25
>__15
> 15
< 15
< 15
>_25
> 15
> 15
-7-0
>0
< - 7
> 0
> 0
<0
> 0
< 0
< 0
< 0
>270
variable
< 273
> 273
< 273
N. MOlders et al. / Atmospheric Research 39 (1995) 91-111
101
Fig. 3. As Fig. 1, but for the results of the cloud retrieval algorithm. Clouds with cloud top radiation temperatures
lower than - 18°C, 0°C to - 18°C, and above 0°C are given in white, light grey and grey, respectively. Open
water and open land appear in black and dark grey, respectively.
greater than - 7%, and the radiation temperature, T4, has to be higher than 270 K. Additionally, the albedo p~ and P2 may not exceed 15%. Pixels not classified so far are now
investigated if they can be identified as snow. The difference P2 - Pl is negative for snow
and ice (e.g., Sakellariou and Leighton, 1988) and has to be less than - 7 % . Furthermore,
the albedo Pl and P2 have to be higher than 25% and T4 has to be lower than 273 K. Pixels
being not classified by now are examined if they are cloudy. Otherwise, the pixels are
rejected as unclassified. Fig. 3 shows the classification obtained by this algorithm for the
satellite pass on 27th April 1986 1325 GMT. Obviously, the distinction between land and
water is not satisfying (see for comparison, e.g., Fig. 2 where the continent occurs in white
if no clouds are present). This lack could be removed by help of a land-sea-map (which
was not available), but it does not affect the satellite derived cloud cover. Sometimes when
P2 - P~ is slightly positive water is misclassified as land (see North Atlantic, Ireland, Fig.
3). Allowing slightly positive values of P 2 - Pl yields in misclassifications the other way
round. A distinction between pure water clouds, mixed clouds and ice clouds, as often made
by cloud retrieval algorithms, was disclaimed for the benefit of an evaluation of the prediction of glaciated clouds by the two different cloud parameterization schemes. Since the
lower limits of coexistence of ice and supercooled water in the CTM cloud module and in
the ice parameterization scheme are - 1 8 ° C and - 3 5 ° C , respectively (cf. Walcek and
Taylor, 1986; M~lders et al., 1994), the class cloud has been divided into clouds with cloud
top radiation temperatures above 0°C, between - 1 8 ° C and 0°C, between - 3 5 ° C and
- 18°C and lower than -35°C.
102
N. MOlders et al. /Atmospheric Research 39 (1995) 91-111
Fig. 4. As Fig. 1, but for the cloud cover as determined form the satellite data projected onto the grid of the model
(from M61derset al., 1994).
After being classified into cloudy or non-cloudy the analyzed pixels are projected onto
the model grid. Then the cloud cover is calculated for each model grid box by counting the
fraction of its pixels identified as cloudy. Unclassified pixels are treated as being cloudfree.
This may lead to slight too low cloud cover values. Fig. 4 shows the satellite derived cloud
cover for 27th April 1986 1325 GMT. It is important to recognize the limitation of such a
cloud cover analysis. The accuracy of the pixel coordinates is about 5 km in the subsatellite
point (Steffens, 1988). Therefore, pixels partly belonging to more than one model grid box
are associated to that grid box which they cover prevailingly. For this reason partly cloudy
pixels are also not considered separately. It can be assumed that the amount of partly cloudy
pixels classified as cloudfree is of the same order of magnitude like that classified as cloudy.
Therefore, the error in the cloud cover resulting from partly cloudy pixel will remain small.
Note that a model grid box encompasses approximately 5289 pixels.
Sensitivity studies on the influence of the chosen threshold values on the derived cloud
cover were carried out. Therefore, the threshold values of land and w a t e r for the albedo Pl
and P2 were decreased and set equal to 10%. The so obtained total cloud cover within the
passage is higher than that determined using the original algorithm due to partly cloudy
pixels (Table 9). In a second study the threshold values of Pl and P2 for land were raised
103
N. Mrlders et al. / Atmospheric Research 39 (1995) 91-111
Table 9
Total cloud cover integrated over the pass region as derived from the satellite data by the cloud detectionalgorithm
and its sensitivity studies and their correlation with the original algorithm
threshold value changed
cloud cover (%)
correlation
original algorithm
land P~.2< 10%
land Pl,2< 20%
ice P~,2> 30%, P2- P~< - 5%
47.7
54.5
45.7
47.5
0.94
0.90
0.85
to be equal to 20%. Herewith, insignificantly more pixels were classified as l a n d due to
partly cloudy pixels (Table 9). The passage integrated cloud cover is slightly lower than
that of the original algorithm. In a third study the threshold value for P2 - Pl is set equal to
- 5 % and the threshold values p~ and P2 for s n o w are raised from 25% to 30%. Slight
differences in the classification of c l o u d and s n o w occur in the regions of the Alps and
Scandinavia. Some pixels are classified as c l o u d and some pixels as s n o w which have been
determined the other way round by the original algorithm. This may be due to pixels
representing broken or thin clouds over snow. The total cloud cover determined for the
passage decreases hardly (Table 9). In view of these sensitivity studies the error in the
satellite derived cloud cover can be expected to be less than 10%.
4. Comparison of model generated and satellite derived cloud cover
The model domain encompasses the troposphere to the 100 hPa level over Europe with
a horizontal grid resolution of 80 km x 80 km (46 X 61 grid points) and 15 non-equidistant
~r-levels (8 levels below and 7 levels above 2 km) in the vertical direction. The initial data
to start the simulations were derived from analysis of the NMC (National Meteorological
Center) data.
The meteorological situation chosen for our study is the three day episode from 25th to
28th April 1986, 1200 G M T governed by a cyclone near Iceland and high pressure over
northeastern Europe. On April 25th the weather situation in Europe showed little pressure
gradients. In front of a trough laying over western Europe warm air masses were advected
from southwest. A low located over Iceland moved slowly eastwards and decreased. A low
located over the Alps enforced and moved northwards during the 26th April. On the 27th
April the low moved from north Germany towards the North Sea. A new cyclone came
from the Adriatic Sea moving northwards.
Indeed, the validation of cloud cover parameterization schemes is one of the most important issues for a proper treatment of clouds, clouds radiation interactions (Buriez et al.,
1988) and cloud related chemical and physical processes in chemical transport models. As
shown by M61ders et al. (1994) the standard cloud cover parameterization scheme performs
hardly better when using the input data from MM4 simulations with cumulus parameterization than when using those of MM4 simulations with the ice parameterization scheme
although the model package was designed for the first mentioned combination. For fairness,
the standard cloud cover parameterization scheme uses the results from the MM4 simulation
104
N. MOlders et al. /Atmospheric Research 39 (1995) 91-111
with the cumulus parameterization scheme while the new cloud cover parameterization
scheme uses those of the MM4 simulation with the ice parameterization scheme.
Since the model delivers no information at what place within a model grid box a predicted
cloud is located, the distribution of clouds determined by the threshold algorithm (Fig. 3)
cannot be directly verified against the model results. Therefore, the distribution of cloud
cover determined from the classified satellite data (Fig. 4) is used to verify the cloud cover
distributions predicted by the model.
The differences in the predicted cloud cover reported by MOlders et al. (1994) when the
CTM standard cloud cover parameterization is applied with data from MM4 simulations
with different cloud schemes were caused by the differences in the precipitation rates and
the relative humidity of the different MM4 results. For instance, the MM4 run with the
cumulus parameterization predicts a larger horizontal pattern of precipitation than the MM4
run with the ice parameterization (cf. figs. 8 and 10 in M/31ders et al., 1994). This yields to
a larger amount of raining clouds. Further differences in cloud cover result from the fact
that the upper model atmosphere predicted by the simulation with the ice parameterization
scheme is drier than that predicted by the run with the cumulus parameterization scheme.
This results from falling ice crystals and rain drops that shift condensate and, hence, moisture
downwards. Since the ice parameterization allows the evaporation and/or sublimation of
rain and/or ice the predicted relative humidity below cloud is often higher than that predicted
by the run with the cumulus parameterization where no evaporation is included. In the latter
scheme all condensate and also the excess moisture that has not been removed by the
cumulus parameterization is treated to fall out instantaneously (within one time step) if the
air is supersaturated.
For the comparison of the results of the new and the standard cloud cover parameterization
scheme the differences in relative humidity discussed above yield in discrepancies of the
cloud cover both determined according to Eq. (3). Moreover, these differences in humidity
cause differences in the distribution of high and low level clouds.
The modeled atmosphere and the observed atmosphere are not well phased for both
simulations. A temporal offset of about two hours in the forecast, the positions of the cloud
fields and the distribution of high relative humidity can be found, i.e., these quantities
predicted for the time two hours after the satellite passes (not shown here) match at best
with the positions observed from the satellite during its passage. This offset may be caused
by the neglect of subgrid scale and microphysical processes (MtJlders et al., 1994). Here
some further research is needed to get a better (spatial and temporal) coincidence between
the synoptic charts and the prediction. Obviously, the results of both cloud cover parameterization schemes also depend on how well MM4 predicts the meteorological situation,
and if the individual cloud parameterization scheme predicts a cloud or not. That is the
resulting cloud cover also depends on how well the cloud relevant processes are represented.
Note that temporal and also spatial offsets in the positions of the clouds may lead to large
differences in the distribution of photochemically active gases, because their precursors
may be strongly variable with location and because the photolysis rates depend on sun
zenith angle and, hence, on the time of the day. The time aspect also plays a role, for the
reason that emissions vary with the time of the day and aqueous chemistry processes depend
on temperature which has a diurnal variation. Moreover, if the clouds are at the right place
at the wrong time this may affect aqueous chemistry processes and wet deposition, too. A
105
17. MOlders et al. / Atmospheric Research 39 (1995) 91-111
<_. 100~
< 00X
< 80X
<
60X
l
<
<
<
50X
40X
30X
<
10X
Fig. 5. Cloud cover as predicted by the standard cloud cover parameterization scheme for 27th April 1986, 1325
GMT (from Mtilders et al., 1994).
cloud passing over a conurbation, for instance, will be different polluted than that passing
over agricultural land. Taking these points into account, a better coincidence of the predicted
and observed cloud positions two hours after the satellite pass is at least not satisfying, but
a hint that the principle works in the right direction.
Fig. 5 illustrates the clouds and cloud cover obtained by the standard cloud cover parameterization scheme. Generally, the CTM cloud module diagnoses too few clouds. The
cloud cover of both precipitating and fair weather clouds is underestimated. The total cloud
cover predicted for the region of the satellite pass is less than half the cloud cover observed
from satellite. The reasons are manifold. The CTM cloud module simulates the cloud from
the water vapor data obtained by MM4 which already removed some moisture as precipitation. To compensate this a 0.5 K and 0.5 g / k g warmer and moister air parcel than in the
MM4 cumulus parameterization is used in the CTM cloud module for the determination of
the lifting condensation level. This raises the possibility for the formation of a cloud, but
may cause higher excess water, Qxs, which lets down the cloud cover (cf. Eq. 2). The fact
that the cumulus parameterization of the CTM cloud module is very restrictive to allow
cloud formation, consequently, also reduces the domain integrated cloud cover. Another
reason for the underestimation of cloud cover may be that the storm efficiency that has been
well chosen for North America may not be adequate under European conditions.
106
N. Mrlders et al. / Atmospheric Research 39 (1995) 91-111
_( 100:{
< 90:(
< 80~
<
60:(
~
<
<
<
50:{
40~
30Z
<
10:(
Fig. 6. As Fig. 5, but for the new cloud cover parameterization scheme (from MOlders and Laube, 1994).
Note that the parameterization of a vertically integrated cloud cover of a representative
cloud may lead to discrepancies between observation and prediction. If two atmospheric
layers each contained a cloud of 10% cloud cover and do not overlap, the satellite would
see 20%, while the cloud module would (correctly) predict 10% for both layers. The often
observed overlapping of clouds in different atmospheric layers is also not considered in the
standard cloud cover parameterization. Note that from satellite low level clouds under high
level clouds cannot be detected.
Fig. 6 shows the cloud cover predicted by the new cloud cover parameterization scheme
for the satellite pass on April 27th 1986 1325 GMT. Since the ice parameterization scheme
only demands supersaturation for cloud formation (Mrlders et ai., 1994) this scheme is
less restrictive and, hence, at more gridpoints clouds are predicted than with the cumulus
parameterization scheme. The total cloud cover predicted for the region of the satellite pass
by the new cloud cover parameterization scheme is 43% for 27th April 1986 1325 GMT.
Apart from the temporal offset in the cloud positions noted before, the ice parameterization
performed an adequate forecast of the fields of cloud water and ice. On 27th April 1986 the
cloud band over the Mediterranean Sea (Fig. 4) is not predicted by the ice parameterization
scheme, for which, in this region no cloud cover can be determined (Fig. 6). A possible
explanation may be that the evaporation from the sea may be insufficient. Here, further
N. Mglders et al. /Atmospheric Research 39 (1995) 91-111
28
100.9
21
75.7
14
50.5
107
Z
t.d
la..
25.2
<.1
<.3
<5
CLASS
<.7
,.9
=1 .
MIDVALUES
Fig. 7. Distribution of cloud cover as predicted by the new cloud cover parameterization scheme for 28th April
1986, 1200 GMT.
research on the whole hydrological cycle is required. Nevertheless, the new cloud cover
parameterization scheme results in a more adequate forecast of the horizontal cloud distribution and cloud cover than that of the standard model version.
Since cloud water, rainwater and ice are not included in the analysis and in the data
assimilation scheme, the simulation with the ice parameterization scheme starts without
initial values of these parameters. These values, of course, are also set to zero at the
boundaries. In view of these facts, some spin-up time is required to form water substances,
i.e., at the beginning of the simulation and at the edges of the model domain cloud cover
may be underpredicted or unrealistic. Usually the CTM domain is chosen smaller than the
MM4 domain, for which, the effect of the underestimation of cloud cover at the boundaries
of the model on other processes will remain small. The spin-up time can be avoided by
applying a pre-run.
Investigations on the distribution frequency of different cloud cover in tenth showed that
the cloud cover determined by the new cloud cover parameterization scheme provides the
U-shaped distribution typically obtained by the results derived from satellite data, i.e., there
exist a lot of areas with low and high amounts of clouds (Fig. 7 ). A tendency to overestimate
the amount of grid boxes with fractional cloud cover higher than 90% and the amount of
grid boxes with fractional cloud cover less than 40% can be found. The U-shaped distribution
is not always achieved by the standard cloud cover parameterization scheme. The shapes
are often skewed to lower values or W-shaped (essentially few clouds, about 50% or full
cloud cover) and the distribution is flatter (Fig. 8) than that provided by the new cloud
cover parameterization scheme or than that derived from satellite data. Moreover, the amount
of grid boxes fully covered by clouds is often strongly underestimated.
108
N. MOlders et al. / Atmospheric Research 39 (1995) 91-111
28
106.4
21
79.8
14
53.2
26.6
<.1
<.3
<.5
CLASS
<,7
<.9
=1 .
HIDVALUES
Fig. 8. As Fig. 7, but for the standard cloud cover parameterization scheme,
For both simulations categorical and bias scores according to Anthes et al. ( 1 9 8 9 ) were
determined for different thresholds of cloud cover (Fig. 9). The bias score measures the
tendency of a model to predict too small or too large an area for a given a m o u n t of cloud
cover. It is defined as ( A n t h e s et al., 1989)
bias = F A / O A
(7)
where F A represents the forecast area and O A is the observed area. The bias scores calculated
for our study indicate a slight underprediction of the cloud cover for thresholds lower than
90% cloudiness and an overprediction of highly cloud covered grid boxes by the n e w cloud
1.0 I
.
°osF
.
.
.
-
.
.
.
.
.
]
06~ \ ,
¢0.2 t" . . . . . . . .
0
~-~
20
40
60
80
°2°I
<
lol-
Jl
•
o.o/
,
0
100
_
,
20
+
_
, --7+-T-,--.--,---,--
40
60
80
Threshold volue
100
Fig. 9. Bias- and categorical scores for different threshold values as determined for 27th April 1986, 1325 GMT,
......
as for the standard and
as for the new cloud cover parameterization scheme.
N. MOlders et al. / Atmospheric Research 39 (1995) 91-111
109
cover parameterization scheme. A categorical forecast is a yes or no forecast of an event,
such as occurence of an amount of cloud cover at a given location. A measure of success
of the categorical forecasts is the percentage of correct forecast. The categorical score is
higher for the simulation with the new parameterization scheme. Both the bias scores and
the categorical scores indicate a better performance of the new scheme than that of the
standard scheme (Fig. 9).
The phase of water is important for the scavenging efficiency (e.g., Scott, 1981; Chang,
1984), radiative transfer (e.g., Hunt, 1973) and, hence, photolysis rates as well as for
aqueous chemistry processes. (Note that in most chemical models cloud ice is assumed to
be chemically conservative.) Therefore, investigations on the amounts of totally glaciated
clouds, mixed phase clouds and water clouds were carried out. The evaluation of the model
predicted water phases with the radiation temperatures derived from satellite data proved
to be complicated for several reasons. The simulated cloud top temperature is an average
value representing the temperature of a volume of 80 km × 80 km and several dekameters.
The simulated cloud is a representative cloud for the model box considered. On the contrary,
the mean cloud top temperature derived from satellite data by averaging of the radiance
temperatures of the pixels classified as cloudy is the average over the different cloud surface
temperatures of higher and lower clouds. Furthermore, both cloud cover parameterization
schemes do not distinguish between parts of the boxes covered by ice clouds, mixed clouds
or water clouds. Nevertheless, taking into account the above noted temporal offset it was
found that ice clouds were simulated in almost those regions where ice topped clouds were
observed (if a cloud was predicted at that place at all). Further research, for validation of
the prediction of ice clouds, mixed phase clouds and water clouds is still required.
5. Final remarks and conclusion
The results generated by two different cloud cover parameterization schemes were evaluated by means of satellite data. In doing so, an automated cloud retrieval algorithm has
been developed and applied to derive cloud cover from AVHRR data. The comparison of
model generated data with those derived from satellite data illustrates that the results
provided by the cloud cover parameterization scheme usually applied in the chemical
transport model is too low and that the positions of the cloud systems are poorly represented.
This underestimation is both large and ambiguous. On the contrary, the cloud distribution
and the cloud cover predicted by the new developed cloud cover parameterization scheme
agree much better with the satellite data. Moreover, here the predicted total cloud amount
is in broad agreement with that derived from satellite data.
Furthermore, it was pointed out that the results of any cloud cover parameterization
scheme depend also on the accuracy with which the clouds are parameterized. One reason
for the underprediction of cloud occurrence when using the standard cloud cover parameterization scheme is that the cumulus parameterization scheme restricts the formation of a
cloud. To improve the performance, sensitivity studies on such restrictions are required and
evaluation by satellite data, of course, different from those used for the adjustment are
unavoidable.
110
N. MOlders et al. /Atmospheric Research 39 (1995) 91-111
Acknowledgements
We gratefully acknowledge the support from U. Steffens, R. Huber, S. Zinn~ker from
the Institut ftir Geophysik und Meteorologie at the Universit~it zu KOln, in the data processing
of the NOAA9 satellite data, especially in the calculation of the pixel coordinates. Thanks
also to H.J. Jakobs from the EURAD group at the Rheinisches Institut ftir Umweltforschung
at Cologne, who helped in questions with the processing of MM4 and the MM4 data. We
should like to express our thanks also to A. Ebel and H. Hass from the EURAD group, C.
Simmer from the Institut ftir Meereskunde at the Universit~it Kiel and to G. Kramm from
the Fraunhofer-Institut fiir Atmospharische Umweltforschung at Garmisch-Partenkirchen
for fruitful discussions and helpful comments. Computational support was performed by
the Forschungszentrum Ji.ilich (KFA), in particular by HLRZ, ZAM, ICH2 and ICH3. This
work was funded by the Ministry of Research and Technology (BMFT) of Germany and
from the Ministry of Science and Research of NRW (Ministerium ffir Wissenschaft und
Forschung des Landes Nordrhein-Westfalen).
References
Anthes, R.A., Hsie, E.-Y. and Kuo, Y.-H., 1987. Description of the Penn State/NCAR Mesoscale Model Version
4 (MM4), NCAR/TN 282 + STR, available from the Publication Office of NCAR.
Anthes, R.A., Kuo, Y.-H., Hsie, E.-Y., Low-Nam, S. and Bettge, T.W., 1989. Estimation of skill and uncertainty
in regional numerical models. Q. J. R. Meteorol. Soc., 115: 768-806.
Buriez, J.-C., Bonnel, B. and Fouquart, Y., 1988. Comparison of model generated and satellite-derived cloud
cover and radiation budget. J. Geophys. Res., 93 (D4): 3705-3719.
Chang, J.S., Brost, R.A., Isaksen, I.S.A., Madronich, S., Middleton, P., Stockwell, W.R. and Walcek, C.J., 1987.
A three-dimensional Eulerian acid deposition model: physical concepts and formulation. J. Geophys. Res., 92:
14,681-14,700.
Chang, F.-L. and Coakley, J.A., Jr., 1993. Estimating errors in fractional cloud cover obtained with infrared
threshold methods. J. Geophys. Res., 98(D5): 8825-8839.
Chang, T.Y., 1984. Rain scavenging of HNO3-vapor in the atmosphere. Atmos. Environ., 18: 191-197.
Colman, R.A., McAvaney, B.J., Fraser, J.R., Rikus, L.J. and Dahni, R.R., 1994. Snow and cloud cover feedbacks
modelled by an atmospheric general circulation model. Clim. Dynam., 9: 253-265.
Dickinson, A., 1985. The weather prediction model. Operational numerical weather prediction documentation
paper. Meteorol. Office, Bracknell, U.K., 4:9.13-9.15.
H~ss, H., Ebel, A., Feldmann, H., Jakobs, H.J. and Memmesheimer, M., 1993. Evaluation studies with a regional
chemical transport model (EURAD) using air quality data from EMEP monitoring network. Atmos. Environ.,
27A: 867-889.
Hass, H., Memmesheimer, M., Geil], H., Jakobs, H.J., Lanbe, M. and Ebel, A., 1990. Simulation of the Chemobyl
radioactive cloud over Europe using the EURAD model. Atmos. Environ., 24: 673-692.
Hense, A. and Heise, E., 1984. A sensitivity study of cloud parameterizations in General Circulation Models.
Contr. Atmos. Phys., 57: 240-258.
Heise, E. and Roeckner, E., 1990. The performance of physically based cloud schemes in general circulation
models. Contrib. Atmos. Phys., 63: 1-14.
Hunt, G.E., 1973. Radiative properties of terrestrial clouds at visible and infrared thermal window wavelengths.
Q. J. R. Meteorol. Soc., 99: 346-369.
Inoue, T., 1987. A cloud type classification with NOAA7 split-window measurements. J. Geophys. Res., 92 (D4):
3991-4O0O.
Kvamsto, N.G., 1991. An investigation of diagnostic relations between stratiform fractional cloud cover and other
meteorological parameters in numerical weather prediction models. J. Appl. Meteorol., 30: 200-216.
N. M6lders et al. /Atmospheric Research 39 (1995) 91-111
1l 1
Lee, J.L., Liou, K.N. and Ou, S.C., 1992. A three-dimensional large-scale cloud model: Testing the role of radiative
heating and ice phase processes. Tellus, 44A: 197-216.
Madronich, S., 1987. Photodissociation in the atmosphere. 1: Actinic flux and the effect on ground reflections and
clouds. J. Geophys. Res., 92: 9740-9752.
Matveev, L.T., 1984. Cloud Dynamics. Reidel, Dordrecht, Netherlands.
M61ders, N., 1993. Wolkenparametrisierung fiir ein Chemie-Transport-Modell. Ph. D. Thesis, 88. Mitt. Inst.
Geophys. Meteorol. Univ. K61n. (in German).
M61ders, N., Hass, H., Jakobs, H.J., Laube, M. and Ebel, A., 1994: Some effects of different cloud parameterizations
in a mesoscale model and a chemistry transport model. J. Appl. Meteorol., 33: 527-545~
M61ders, N. and Laube, M., 1994. A numerical study on the influence of different cloud treatment in a chemical
transport model on gasphase distribution. Atmos. Res., 32: 249-272.
Phulpin, T., Derrien, M. and Brard, A., 1983. A two dimensional histogram procedure to analyze cloud cover
from NOAA satellite high-resolution imagery. J. Clim. Appl. Meteorol., 22: 1332-1345.
Raschke, E., Jacobs, H., Lutz, H.-J. and Steffens, U., 1986. Wolkenerkennung fiber der Anarktis in Satellitenbildern. Polarforschung, 56:69-78 (in German).
Raschke, E., Bauer, P., Mi31ders, N., 1988. Clouds over both polar regions from ISCCP pilot data sets. Proc. 2nd
Conf. Polar Meteorol. Oceanogr., March 29-31, Madison, Wisconsin. Am. Meteorol. Soc., Boston, pp. 137
140.
Saito, K. and Baba, A., 1988. A statistical relation between relative humidity and the GMS observed cloud amount~
J. Meteorol. Soc. Jpn., 66: 187-192.
Sakellariou, N.K. and Leighton, H.G., 1988. Identification of cloudfree pixels in inhomogeneous surfaces from
AVHRR radiances. J. Geophys. Res., 93 (D5): 5287-5293.
Scott, B.C., 1981. Sulfate washout in winter storms. J. Appl. Meteorol., 20:619-625.
Slingo, J.M., 1980, A cloud parameterization scheme derived from GATE data for use with a numerical model.
Q. J. R. Meteorol. Soc., 106: 747-770.
Slingo, J.M., 1987. The development and verification of a cloud prediction scheme for the ECMWF model. Q. J.
R. Meteorol. Soc., 113: 899-927.
Smith, R.N.B., 1990. A scheme for predicting layer clouds and their water content in a general circulation model.
Q. J. R. Soc., 116: 435460.
Smagorinsky, J., 1960. On the dynamical prediction of large-scale condensation by numerical methods. Geophys.
Mon., 5: 71-78.
Steffens, U., 1988. Private communications.
Sundqvist, H., Berge, E, and Kristjansson, J.E., 1989. Condensation and cloud parameterization studies with a
mesoscale numerical weather prediction model. Mon. Weather Rev., 117: 1641-1657.
Walcek, C.J. and Taylor, G.R., 1986. A theoretical method for computing vertical distributions of acidity and
sulfate production within cumulus clouds. J. Atmos. Sci., 43: 339-355.
Yamanouchi, T., Suzuki, K. and Kawaguchi, S., 1987. Detection of clouds in Antarctica from infrared multispectral
data of AVHRR. J. Meteorol. Soc. Jpn., 65: 949-962.