Evaluation of a probabilistic cloud masking algorithm for ... - CM SAF

EUMETSAT Satellite Application Facility on Climate Monitoring
Visiting Scientist Report
Evaluation of a probabilistic cloud masking
algorithm for climate data record processing:
SPARC: a new scene identification
algorithm for MSG SEVIRI
CDOP VS Study No 14
Date 15. October 2010
Reference CLM_VS10_03

SPARC: a new scene identification
algorithm for MSG SEVIRI
Report on Modifications and Validation
Fabio Fontana
University of Bern
Reto Stockli
MeteoSwiss
Stefan Wunderle
University of Bern
September 2010

Summary
The Separation of Pixels Using Aggregated Rating Over Canada (SPARC) algo-
rithm represents an alternative to conventional scene identification algorithms as
it outputs a single cloud contamination rating (F) instead of a cloud mask with
a limited number of categories. The rating itself is mainly calculated from three
sub-scores generated based on information on the brightness temperature (T-score),
brightness (B-score), and the reflectance in the visible and short-wave infrared por-
tion of the electromagnetic spectrum (R-score). Even though the algorithm was
originally designed for the Advanced Very High Resolution Radiometer (AVHRR),
it may be applied to data from other multispectral sensors.
This report describes a modified version of SPARC and its application to data from
the Meteosat Second Generation (MSG) Spinning Enhanced Visible and Infrared
Imager (SEVIRI) sensor. A validation was performed for the period from July
2004 to June 2005. SPARCMSG output was compared to cloud information from
the Alpine Surface Radiation Budget (ASRB) Network in the Swiss Alps: Cloud
Free Index (CFI) and Partial Cloud Amount (PCA; in eighths). Results show a
good agreement between CFI and F mainly for daytime observations, with linear
correlation coefficients (r) ranging between 0.67 in fall and 0.78 in the winter months.
At the ASRB site in Payerne, observations of complete clear-sky agreed with 85%
(92.3%) during daytime (nighttime) between ASRB PCA and SPARCMSG cloud
mask. However, results also point to an underestimation of nighttime cloud cover
by SPARCMSG. If only two SEVIRI spectral channels are fed into SPARC to simulate
data from the Meteosat First Generation (MFG), a good agreement between CFI and
F is obtained during daytime (r=0.65). Finally, application of SPARC at full disk
level revealed an underestimation of cloud cover compared to the Climate Monitoring
(CM)-Satellite Application Facility (SAF) cloud mask.
I

Summary II
Overall, SPARCMSG has proven to provide interesting new opportunities for cloud
detection using the MSG SEVIRI sensor. Results are particularly promising with
regard to the application of SPARC for data of the MFG satellite series.

Contents
Summary I
Table of Contents III
List of Figures V
List of Tables VII
1 Overview 1
2 SPARC algorithm 3
2.1 Cloud detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2.2 Snow detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.3 Cloud shadow identification . . . . . . . . . . . . . . . . . . . . . . . 8
3 The SPARC algorithm for MSG SEVIRI 9
3.1 SPARCMSG cloud mask . . . . . . . . . . . . . . . . . . . . . . . . . . 9
3.2 Calculation of the T-score . . . . . . . . . . . . . . . . . . . . . . . . 13
3.3 The SPARCMSG snow detection scheme . . . . . . . . . . . . . . . . . 16
3.4 Further modifications . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
4 Validation Study 22
4.1 General Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
4.2 Data and Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
III

Table of Contents IV
4.2.1 ASRB data . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
4.2.2 MSG data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
4.3 Results and Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . 25
4.3.1 Comparison with ASRB Cloud Parameters . . . . . . . . . . . 25
4.3.2 MFG Simulation . . . . . . . . . . . . . . . . . . . . . . . . . 34
4.3.3 Validation at Full Disk Level . . . . . . . . . . . . . . . . . . . 39
4.4 Concluding Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . 42

List of Figures
2.1 SPARC output for a sample slot. . . . . . . . . . . . . . . . . . . . . 5
3.1 HRV channel of a sample slot together with the SPARC output cre-
ated with different thresholds. . . . . . . . . . . . . . . . . . . . . . . 11
3.2 HRV channel of a sample slot together with the SPARCMSG output. . 12
3.3 Examples of a Mannstein function fitted to clear-sky BT108 values. . 15
3.4 Example of a RST map in summer (left) and winter (right) over Swiss
Alps. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
3.5 Output of the novel SPARC snow detection algorithm. . . . . . . . . 19
3.6 Time series of monthly maximum snow cover in the Swiss Alps. . . . 20
4.1 CFI vs. F for all sites and 12 months combined. . . . . . . . . . . . . 26
4.2 CFI vs. F for each season separately. . . . . . . . . . . . . . . . . . . 26
4.3 PCA vs. SPARCMSG cloud mask for the entire 12 month period. . . . 28
4.4 PCA vs. CMMSG for each season separately. . . . . . . . . . . . . . . 29
4.5 Histogram of CMMSG output for each PCA class at the ASRB site in
Payerne. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
4.6 Histogram of CMMSG output for each PCA class at the ASRB site at
Weissflujoch. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
4.7 CFI vs. F for all sites and 12 months combined - SPARCMFG mode. . 34
4.8 Output of SPARCMSG and SPARCMFG on March 20, 2005 (8:45 am). 35
4.9 Histogram of CMMFG output for each PCA class at the ASRB site in
Payerne - SPARCMFG mode. . . . . . . . . . . . . . . . . . . . . . . . 37
V

List of Figures VI
4.10 Histogram of CMMFG output for each PCA class at the ASRB site at
Weissflujoch - SPARCMFG mode. . . . . . . . . . . . . . . . . . . . . 38
4.11 Comparison of the CM-SAF, SPARCMSG, and SPARCMFG cloud masks
for the full disk. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
4.12 Comparison of the CM-SAF, SPARCMSG, and SPARCMFG cloud masks
for five regions of interest within the full disk. . . . . . . . . . . . . . 41
4.13 Modification of the R-score for bright desert surfaces. . . . . . . . . . 42

List of Tables
2.1 Boundary coordinates of the region of interest. . . . . . . . . . . . . . 7
4.1 Sites of the ASRB network used for the validation. . . . . . . . . . . 23
4.2 Channels considered as SPARC input in MSG and MFG mode. . . . 25
VII

List of Tables VIII

Chapter 1
Overview
It is of major importance for any satellite data processing system to accurately de-
termine the state of a pixel. Depending on the application it may be necessary to
assign a pixel to the clear-sky (e.g., for land surface applications) or cloudy cate-
gory (e.g., for the retrieval of cloud properties). Further, assigning for instance a
mixed pixel with sub-pixel cloud cover to the clear-sky or cloudy category largely
depends on the final application. For land surface applications a clear sky con-
servative approach is preferred, where for the retrieval of cloud properties a cloud
conservative approach has to be used. To achieve discrimination of various scene
types (e.g., snow, clouds, water, or vegetation), pixel scene identification algorithms
make use of the basic underlying principle that scene types may be distinguished
based on their inherent reflective and emissive properties in certain portions of the
electromagnetic spectrum (Lillesand et al., 2004).
Scene identification algorithms traditionally follow a branching/thresholding ap-
proach: a sequence of spectral tests is applied to each pixel, assigning the pixel
to a certain class (e.g., cloudy, clear-sky, or partly cloudy) based on predefined
thresholds. However, such branching/thresholding approaches have two main dis-
advantages (Khlopenkov and Trishchenko, 2007): first, since a yes/no decision is
made at each node of the classification tree, a pixel that happens to represent an
intermediate state may be assigned to the wrong category, or its classification may
not be possible. Second, the final result of such classification schemes only contains
information on a limited number of classes, such as ’cloudy’, ’clear-sky’, ’partly
cloudy’, or ’uncertain’, but does not provide information on the degree of cloud or
1

SPARC algorithm 2
haze contamination within the field of view. Such information may, however, be
valuable for the generation of clear-sky composites for many applications, e.g., land
surface albedo derivation or vegetation dynamics analysis.
In order to address these limitations, the Separation of Pixels Using Aggregated
Rating over Canada (SPARC) algorithm (Khlopenkov and Trishchenko, 2007) was
developed at the Canada Centre for Remote Sensing (CCRS). The SPARC algorithm
was originally designed for the use with Advanced Very High Resolution Radiometer
(AVHRR) data over Canada, but its design should theoretically enable the algorithm
to be transferred to other multi-spectral sensors such as the Meteosat Second Gen-
eration (MSG) Spinning Enhanced Visible and Infrared Imager (SEVIRI) and other
geographical regions of interest, even though some modification might be necessary.
The aim of this report is to provide a short overview of the SPARC algorithm (Sec-
tion 2) and the modifications specific for its application with the MSG SEVIRI sensor
(Section 3). The report also presents the results of a validation study performed
over the Swiss Alps (Section 4.3.1). The applicability of the SPARC algorithm to
data from the Meteosat First Generation (MFG) satellite series is evaluated (Section
4.3.2). Finally, preliminary results of the applicaiton of SPARC at SEVIRI full disk
dimension are presented (Section 4.3.3) and concluding remarks are given in Section
4.4.

Chapter 2
SPARC algorithm
2.1 Cloud detection
Information provided in this chapter closely follows the work of Khlopenkov and
Trishchenko (2007). The SPARC algorithm outputs a score, which can be regarded
as a measure of the cloud contamination probability for each pixel. The SPARC
score (F) is obtained through the summation of the output of a number of individual
tests, which make use of all five spectral channels of the AVHRR sensor (Eq. 2.1):
where
F = T + B + R + �
Ai, (2.1)
• T is the temperature score, which is based on the comparison of the brightness
temperature at 10.8µm and a dynamic threshold derived from an external sur-
face skin temperature field. In the original SPARC version, surface skin tem-
perature is obtained from the North American Regional Reanalysis (NARR)
at a spatial resolution of 32 km×32 km (Mesinger and Coauthors, 2006).
• B is the brightness score, calculated from the reflectance in the visible (VIS)
and near infrared (NIR) portion of the electromagnetic spectrum over land
and water, respectively,
• R is the reflectance score calculated either from the reflectance values in the
VIS and short wave infrared (SWIR) portions of the spectrum (if a SWIR
3
i

SPARC algorithm 4
channel is available from AVHRR), or from the solar reflective component at
3.7µm otherwise, and
• Ai represents the added scores of four additional tests that are only performed
under certain circumstances due to computational considerations.
These supplementary tests include:
• the simple ratio test (N), exploiting the relatively homogeneous reflective prop-
erties of cloudy scenes in the VIS and NIR spectrum as opposed to clear-sky
observation of, e.g., vegetation cover,
• the texture uniformity test (Utext) based on the NIR spectrum and the ther-
mal uniformity test (Utemp) based on 10.8µm brightness temperature, which
both analyze the variability of the pixel values for a pixel and its four nearest
neighbors, and finally
• the thin cirrus test based on the brightness temperature difference between
12µm and 10.8µm.
The scores are designed such that strong (weak) cloud evidence leads to positive
(negative) values of F, whereas values of F ≈0 represent intermediate, i.e. uncertain
states. In combination with the linear aggregation principle, this design has the
advantage that a cloud mask can be generated even if some of the scores are not
calculated. More detailed information on the calculation of the scores is provided in
Khlopenkov and Trishchenko (2007).
However, some adjustments of Equation 2.1 are required in order for the single
scores to receive appropriate weights under special observation conditions, namely
snow cover, sun glint, and in the proximity to the terminator line. For this reason,
three correction factors are implemented. These are:
• the correction factor (s) for snow conditions, which is necessary since snow-
covered areas observed under clear-sky conditions and clouds can be equally
bright. In the original SPARC scheme, this factor is estimated from the surface
skin temperature fields;

SPARC algorithm 5
• the sun glint correction factor (g; calculated from the observation geometry)
to account for increased spectral reflectance over sun glint areas;
• the nighttime correction factor (n), which accounts for the special circum-
stances at very low Sun zenith angles in the proximity of the terminator line.
Equation 2.1 is thus modified as follows:
F = B(1 − g)sn + T (1 + g)(2 − s)(2 − n) + R(1 − 0.6g)(2 − s)n + N(1 − g)sn
+ Utext(1 − g)sn + Utemp(1 + g)(2 − s)(2 − n) + C (2.2)
Note that all correction factors reduce the weight of the reflective scores (where
applicable for a certain correction factor) and assign more weight to the thermal
tests. By design, reflectance scores are turned off completely at nighttime and the
SPARC output is based only on the test for the thermal channels. Figure 2.1 displays
an example of all SPARC scores as well as the snow correction factor for a MSG
SEVIRI time slot in March 2004. The boundary coordinates of the region of interest
covering the Swiss Alps are provided in Table 2.1. The HRV channel (1A) shows
significant snow cover over the elevated topography of the Swiss Alps under mostly
clear-sky conditions as well as prevalent low stratus clouds over the Swiss Main
Plateau. Note that the absolute values of the single contributing scores vary greatly
in magnitude. In particular, scores of the supplementary tests display significantly
smaller absolute values compared to the three main scores (T, B, and R). Figure
2.1 also highlights a problematic issue around the calculation of the Utext-scores
over the Swiss Alps during wintertime: even under cloud-free conditions the high
reflectance contrast between snow-covered mountain tops and low-lying and snow
free valley bottoms results in increased values of the Utext-score. The same problem,
Figure 2.1 (following page): High resolution visible (HRV) channel of the MSG SE-
VIRI time slot on March 20, 2005 (8:45 am; 1A), together with the corresponding SPARC
scores (see text for more details). The correction factor for snow conditions (s) is also
provided (1E). Note that s was calculated as described in Khlopenkov and Trishchenko
(2007), however, using a clear-sky BT108 mosaic as input (cp. Section 3.2) instead of
an external surface skin temperature map. Details on the calculation of the scores can be
found in Khlopenkov and Trishchenko (2007).

SPARC algorithm 6

SPARC algorithm 7
Table 2.1: Boundary coordinates of the region of interest (ROI) covering the Swiss Alps,
where ’lat’ is the latitude [ ◦ ] and ’lon’ the longitude [ ◦ ].
ROI Swiss Alps
latmin
latmax
lonmin
lonmax
yet less pronounced, is observed for the Utemp-score. Similarly, the N-score adopts
high values both for cloud and snow cover. Even though the contributions of these
additional scores to the final score F are small, it is suspected that erroneously large
values of these scores may lead to some misclassifications in uncertain conditions.
2.2 Snow detection
Snow detection is originally performed in two steps based on the T-, B-, and R-
scores:
45.5
48.0
6.5
10.0
• In step one, a number of threshold tests is performed, testing whether the pixel
is sufficiently bright (B-score threshold), limiting the R-score to small values to
exclude cloud contamination, and ensuring that the brightness temperature at
10.8µm is close to the surface skin temperature map as outlined above. Once
a pixel is classified as snow-covered in this first step, it undergoes another test
in
• step two, which is designed to detect clouds over snow based on the R- and
T-score. A pixel is classified as clear-sky snow, if both the R- and T-score are
smaller than zero. The pixel is classified as snow-covered and cloud contami-
nated, if these criteria are not met.
Additional details as well as a flowchart of the SPARC snow algorithm can be found
in Khlopenkov and Trishchenko (2007, Figures 4 and 5).

SPARC algorithm 8
2.3 Cloud shadow identification
Cloud shadow detection is performed based on geometrical considerations, i.e. by
computing the extent of the cloud shadow on the Earth’s surface (shadows cast
by high clouds on the top of low clouds are not considered). The output mask
includes information on three different categories: shadow-free and, according to
two thresholds to F, thin cloud shadow and thick cloud shadow. For more details
on the cloud shadow retrieval it is again referred to Khlopenkov and Trishchenko
(2007).

Chapter 3
The SPARC algorithm for MSG
SEVIRI
Prior to the application of the SPARC algorithm to data from the MSG SEVIRI
sensor (hereafter referred to as SPARCMSG), a number of modifications relative to
the original version (SPARCORIG) were conducted. These will be explained in the
following sections.
3.1 SPARCMSG cloud mask
Other than in SPARCORIG, the primary output of the SPARCMSG cloud detection
scheme is not the score F itself, but rather a cloud mask with three different classes:
clear (0), thin cloud (1), and thick cloud (2). Similar to the criteria in Khlopenkov
and Trishchenko (2007) for the distinction between shadows caused by optically
thin or thick clouds, respectively, this simplified cloud mask is obtained by applying
thresholds to F:
0 : F < ct : clear-sky; (3.1)
1 : 4 > F > ct : thin cloud;
2 : F > 4 : thick cloud,
where ct is the cloud threshold, which for each pixel is calculated incrementally as
ct = −4.0 : as a basic cloud threshold (3.2)
9

SPARC algorithm 10
ct = ct − 5.0s : stricter threshold if T-score was not applied
ct = ct − 5.0s : even stricter if C-score was not applied.
Even though SPARC was originally designed to overcome limitations of traditional
cloud detection schemes as outlined above, the cloud mask generation based on the
application of thresholds to F represents an interesting feature of the new SPARCMSG
algorithm: by changing the thresholds in either direction, the resulting cloud mask
may be shifted dynamically from a clear-sky conservative to a cloud conservative
mode without modifying the retrieval algorithm itself. This is illustrated in Figure
3.1 which displays the high resolution visible (HRV) channel of a sample slot (A)
covering the Swiss Alps together with F (B) and a number of cloud masks (D-
I). Cloud masks were generated based on basic cloud thresholds varying between
−8≤ct≤4 (thresholds for thick cloud were changed accordingly and were set to
ct+8). The cloud mask displayed in Figure 3.1, panel F, was generated following
Equations 3.1 and 3.2. Note that significant differences are observed depending on
the selected threshold, especially in areas where the cloud mask error (see below)is
large.
In a subsequent step, the error (ε) of the cloud mask is calculated based on F as
follows:
(F −ε)2
−
ε = e (2σ2 ε ) , (3.3)
where ε = 0.0
σε = 10.0
(F − ε) 2 � 2000.0
The cloud mask error may adopt values between 0 and 1 is designed to be largest at
values of F=0, i.e., in situations where the summation of the SPARC scores did not
give clear evidence for either cloudy or clear-sky conditions. Hence, ε is a potentially
useful parameter, e.g., to set up decision criteria in the clear-sky compositing process.
Figure 3.2 displays the high resolution visible (HRV) channel of a MSG SEVIRI time
slot in February, together with the corresponding SPARC score F, cloud mask, and
cloud mask error ε. Cloud mask error is also displayed in Figure 3.1 (C).

SPARC algorithm 11
Figure 3.1: The HRV channel of a sample slot on 11 September 2004, 11 am (A), together
with the SPARC score (B), the cloud mask error (C; see text for explanations) and a
number of cloud masks based on varying thresholds (C-H; see text for further explanations).
Black areas represent clear-sky conditions, optically thick (thin) clouds are delineated in
white (grey).

SPARC algorithm 12
Figure 3.2: The HRV channel of a sample slot channel on 25 February 2005, 10:15 am
(top, left) together with the SPARC score (top, right), cloud mask (bottom, left), and cloud
mask error (bottom, right). Please note the increased values of the error ε along the cloud
boundaries.

SPARC algorithm 13
3.2 Calculation of the T-score
One of the major modifications with regard to SPARCORIG concerns the calculation
of the T-score. The high temporal resolution of the SEVIRI sensor (15 minutes)
opens up new opportunities for surface temperature retrievals, because the chance
of a certain location being observed under clear-sky conditions during the course of a
single day is significantly increased compared to the AVHRR sensor (1-2 observations
daily). For this reason, an external surface temperature map is not implemented in
SPARCMSG. Instead, a clear sky radiative surface temperature (RST) map directly
derived from SEVIRI 10.8µm channel radiances (BT108) is used as a reference in
the calculation of the T-score. As a preliminary solution, brightness temperature is
used instead of the radiative surface temperature, because atmospheric correction
is not yet implemented.
The RST map is generated daily as follows: after the processing of all 96 daily slots
is completed, the BT108 clear-sky values for each pixel and slot are aggregated in
a running array of clear sky BT108 values. In this running array, which stores a
maximum of two clear-sky values per pixel and slot, the oldest clear-sky BT108
values are replaced by the current observations. However, values observed more
than ten days ago are removed even if there are no current clear-sky observations
available for replacement. The resulting array is subsequently used to model the
diurnal evolution of surface brightness temperature for each pixel on the following
day. This is performed by fitting a modified version of the diurnal surface radiative
temperature model (f) by Mannstein et al. (1999). The model is defined as follows
(Dürr, 2009):
2
−2(ωt − a3)
f = a0 + a1(exp( ) + 0.1sin(ωt − a3)) (3.4)
a 2 2
where ω = 2π/96, with 96 the total number of slots per day, a0 is the minimium
brightness temperature (BT) of the diurnal cycle in Kelvin, a1 is the diurnal ampli-
tude of BT in Kelvin, a2 is the half width of the day lenght in radians, and a3 is the
true Sun time in radians. Please consult Dürr (2009) for more detailed information
on the model. Points are weighted as follows, considering the cloud mask error for
the corresponding BT108 clear-sky value as well as the date of observation:
w =
1
ε + t , (3.5)
tmax

SPARC algorithm 14
where w =weight for each point, t =the number of days the clear-sky BT108 value
had been stored in the running mosaic, and tmax=10, i.e. the maximum number of
days considered in the clear-sky compositing (may be adjusted in the configurations
file). Hence, smallest weights are assigned to BT108 values observed early in the
preceeding compositing interval of tmax days and under uncertain conditions (F≈0).
Figure 3.3 shows examples of a RST fit for a location on the Swiss Main Plateau
in summer (left) and winter (right). Please note the difference in the evolution and
the magnitude of diurnal brightness temperature.
In addition, a spatial interpolation procedure is implemented in order to fill gaps
in the RST field. The interpolation is based on an elevation-dependent RST filling
using a lapse-rate that is retrieved from neighboring valid pixels with a radius of 50
SEVIRI HRV pixels. Figure 3.4 provides an example of a RST map of the Swiss
Alps in summer 2005 (left) and winter 2004/2005 (right).
The use of a RST map derived directly from MSG data has two major advantages:
first, the suggested SPARCMSG cloud detection scheme represents a stand-alone ap-
proach that does not rely on any auxillary input data, such as the coarse spatial
resolution NARR surface skin temperature fields as described in Section 2. Sec-
ond, in contrast to these modelled surface temperature fields, the RST map ap-
parently provides substantial spatial detail, which is assumed to be beneficial for
cloud detection over the complex and heterogeneous terrain of the Swiss Alps. The
disadvantage of modelling the following day’s diurnal BT108 amplitude from the
previous day’s clear-sky BT108 observations is that day-to-day changes in BT108
at a certain time of the day cannot be captured. Resulting biases in the T-score
could translate into erroneous cloud masks, especially if such day-to-day changes
happen to be significant. However, this problem is likely to occur only in the case of
warm and low clouds, when the temperature difference between ground and cloud
top are small. Dectection of high and cold clouds should not be affected. A solution
might, however, be to apply a two-stage processing, where first the RST of the full
day is calculated, then the cloud mask for the full day is repeated with the cur-
rent day’s RST. Furthermore, this stand-alone approach implies that, under some
circumstances, the T-score cannot be calculated for certain pixels. This is the case
either at the beginning of the processing sequence, when the RST map is not defined
due to the lack of clear-sky BT values in the mosaic, or if no clear-sky BT values

SPARC algorithm 15
Figure 3.3: Examples of a Mannstein function fitted to clear-sky BT108 values for a
location in the Swiss main land in summer (left) and winter (right). The error bars are
calculated as defined in Equation 3.5.
Figure 3.4: Example of a RST map in summer (left) and winter (right) over the Swiss
Alps Region of Interest (ROI; see Table 2.1 for details on the ROI). Please note the bright-
ness temperature contrast between the low-lying valleys and the more elevated topography,
especially in winter.

SPARC algorithm 16
had been observed for more than tmax days (certain gaps cannot be filled using the
spatial interpolation scheme if large areas are affected). Limitations at the beginning
of a processing sequence can be avoided by adding a spin-up time prior to the actual
period of interest. Theoretically, one single clear-sky day would be sufficient to gen-
erate a complete RST field. However, escpecially during the winter months when
low stratus cloud cover over the Swiss Main Plateau is observed frequently, gaps in
the RST field may still be observed after several weeks of observation and longer
spin-up times are required. However, due to the design of the scores and the linear
aggregation principle as mentioned above, SPARCMSG can also be employed if not
all scores are calculated. See Section 4.3.2 for a discussion of a situation where not
all scores are calculated. Another option could be to consider high resolution surface
temperature data from, e.g., the COSMO-2 model over the Swiss Alps (grid spacing
of about 2.2 km; Doms and Schaettler, 2002) as a reference in T-score calculation
where the RST map is not defined.
3.3 The SPARCMSG snow detection scheme
The third major modification of the SPARC algorithm concerns the calculation of
the snow mask. A daily snow mask is generated through the aggregation of all
snow-covered pixels observed under clear-sky conditions during the course of the
day. Snow detection in SPARCMSG includes a number of different tests, whereof
one includes the calculation of the Normalized Difference Snow Index (NDSI). The
NDSI ([−1.0,1.0]) is calculated by dividing the difference of reflectances observed
in MSG SEVIRI channels 1 and 4 by their sum and may be regarded as a measure
of the abundance of snow and ice within the area covered by a pixel (Salomonson
and Appel, 2004). Even though the spectral channels needed for the calculation of
the NDSI are available on the AVHRR sensors of the third generation (AVHRR/3),
historical AVHRR sensors lack the channel at 1.6µm. As a consequence, the NDSI
cannot be implemented systematically for AVHRR, which is the reason why an
alternative method was originally developed as described in Section 2.2.
The idea behind the new snow detection scheme is the calculation of a S-score that
can be regarded as a measure of the snow evidence within a pixel. The S-score is
calculated by summing two newly defined scores, the NDSI-score and the FREEZE-

SPARC algorithm 17
score, as well as the T- and R-scores as described above. The NDSI-score is defined
as follows:
NDSI =
REF 006 − REF 016
REF 006 + REF 016
NDSI-score = (NDSI − 0.4)40.0, (3.6)
where REF 006 is the reflectance in SEVIRI channel 1 at 0.6µm and REF 016 the
reflectance in channel 3 at 1.6µm. Positive (negative) values of the NDSI-score are
obtained for snow-covered (snow-free) pixels. The FREEZE-score is calculate as
follows:
FREEZE-score = −|BT 108 − (Tfreeze) + 5|, (3.7)
where Tfreeze is calculated as a sine function oscillating between a maximmum of
+2 ◦ C in spring and a minimum of −2 ◦ C in fall (Khlopenkov and Trishchenko, 2007).
The FREEZE-score is designed to adopt negative values if the difference between
BT108 and Tfreeze is large (positive or negative), and positive values if BT108 is
close to Tfreeze. The S-score is finally aggregated as
S-score = (R − 3.0)(−1.0) + NDSI-score + (T − 3.0)(−1.0) + FREEZE-score (3.8)
The snow mask is subsequently generated from the S-score by classifying pixels
as snow-covered if the S-score is greater than zero. Again, the sensitivity of the
algorithm may be adjusted by tuning the selected threshold. Snow masks are only
generated for observations made at Sun zenith angles (SZAs) smaller than 75 ◦ , since
the suggested approach in some cases failed to accurately detect snow cover in the
proximity of the terminator line, mainly due to erroneously high avalues of the
NDSI-score. In addition, two criteria are applied to the snow mask:
• For pixels where BT>tfreeze+10.0, the snow mask is set to zero. This threshold
is needed to account for subpixel snow in mountains in summer.
• For pixels with reflectance values in the visible channel (HRV or R0.6) of

SPARC algorithm 18
with Smin=−15.0 and Smax=10.0. Given the limitations of the new snow detection at
high SZAs, the snow correction factor s must be adjusted for SZAs greater than 75 ◦ .
The preliminary implementation of this correction at high SZAs reads as follows:
tf = e SZA−75.0
10.0
s75 = (s + 0.3)tf, (3.10)
where tf is the correction factor in the proximity of the terminator line and s75 is
the adjusted snow correction factor. In order to overcome limitations due to low
SZAs, a possible future solution could be not to generate snow masks at all for
observations made at SZA>75 ◦ , or, adopting the idea of the nighttime correction
factor, to reduce (increase) the weights of the reflective (thermal) scores in Equation
3.8 in the proximity to the terminator line.
Figure 3.5 displays the single scores of the new snow detection scheme (1A, 2A,1B),
the resulting snow mask (1C) as well as the corresponding snow correction factor
s (2B) for a day in March 2005. See Figure 4.8 (left) for additional SPARCMSG
output for the same time slot. The corresponding snow mask generated based on
the original snow detection scheme is provided as a reference (2C).
Major differences are observed both for s and the daily snow mask. Given the
importance of the snow correction factor s to reduce the weight of the reflective
scores in Equation 2.2, such differences in s inevitably translate into discrepancies
in the magnitude of F in SPARCMSG compared to SPARCORIG. Comparing the
values of F from SPARCORIG in Figure 2.1 (2E) with SPARCMSG F in Figure 4.8
(1E), significant differences become apparent in the Swiss Alps. Panel 2E in Figure
2.1 shows increased values of F over snow-covered areas, except where reflective
scores were turned of by s (seen as dark areas; see also Panel 1E in Figure 2.1).
Regarding the comparison of the daily snow masks, from a purely visual inspection
the novel snow mask seems to be more accurate, with snow cover being understi-
mated by SPARCORIG. Figure 3.6 (left) displays a preliminary time series of monthly
snow cover from SPARCMSG, SPARCORIG, and Moderate Resolution Imaging Spec-
troradiometer (MODIS) data over the Swiss Alps. Monthly maximum snow cover
composites for the period between July 2004 and June 2005 were generated for
both SPARC versions by considering all pixels that were classified as ’snow’ at any
time during the month. Monthly MODIS snow composites provide average monthly

SPARC algorithm 19
Figure 3.5: Output of the novel SPARC snow detection algorithm on March 20, 2005
(8:45am). The corresponding snow mask generated based on the original snow detection
scheme is provided as a reference (2C). See Figure 4.8 for the corresponding HRV image
and SPARCMSG.

SPARC algorithm 20
Figure 3.6: Time series of monthly maximum snow-covered area (SCA; [%]) in the Swiss
Alps region of interest for SPARCORIG, SPARCMSG, and MODIS (see the text for more
information on the generation of the composites).
snow cover for a certain pixel. For comparison with SPARC SCA, MODIS data
were transformed into a binary snow mask by classifying a pixel as snow-covered
if the average monthly snow cover was ≥50%. Differences [%] relative to MODIS
are provided in the right panel. An examination of SCA time series reveals that
the underestimation of SCA by SPARCORIG relative to SPARCMSG is observed in-
dependently of the season. Relative to MODIS, SPARCMSG underestimates (overes-
timates) SCA in summer (fall, winter, and spring), whereas SPARCORIG SCA does
not follow the same pattern. During fall, winter, and spring, an overestimation of
SCA by SPARCMSG is in accordance with theory since ephemeral snow cover may
not be captured by MODIS due to the lower temporal resolution of the MODIS
sensor system. However, during summer we would expect the discrepancies to be
small. Further analysis is required to explain the underestimation of SCA by both
SPARC algorithms.
3.4 Further modifications
In the original version of SPARC the additional scores are only calculated in uncer-
tain situations to obtain more information on the state of the pixel. As of now, this

SPARC algorithm 21
distinction is not made in SPARCMSG and all scores are calculated for each pixel
and fed into Equation 2.2.
Furthermore, the coefficients for the calculation of the R-score needed to be modified,
since the contour lines in Khlopenkov and Trishchenko (2007, Figure 3) could not be
reproduced with MSG SEVIRI reflectances using the original coefficients. The new
coefficients for the calculation of the R-score are: offset=0.02 and scale factor=320
(original coefficients: offset=0.1, scale factor=160).

Chapter 4
Validation Study
4.1 General Remarks
The focus of the validation study was set on a region of interest (ROI) covering the
Swiss Alps. The boundary coordinates of the subset are listed in Table 2.1. A time
period of 12 months (from July 1, 2004 to June 30, 2005) was considered. Further-
more, seasons were analyzed independently, where winter included the months from
December to February, spring from March to May, summer from June to August,
and fall from September to November. In addition, differences between the Climate
Monitoring (CM) Satellite Application Facility (SAF) cloud mask (EUMETSAT CM
SAF , 2009; Schulz et al., 2009) and SPARCMSG were visually analyzed at full disk
level for the 12 UTC slot on June 13, 2006, which is the day selected for the cloud
mask intercomparison study of the EUMETSAT workshop (Walther et al., 2009).
4.2 Data and Methods
4.2.1 ASRB data
Information on cloud coverage derived from data of the Alpine Surface Radia-
tion Budget (ASRB) network (Philipona et al., 1996; Marty and Philipona, 2000),
was used for the validation of the SPARCMSG cloud mask. Two parameters were
compared to SPARCMSG output at each ASRB site in the Swiss Alps (Dürr and
22

SPARC algorithm 23
Philipona, 2004): the Cloud-Free Index (CFI) derived from downward longwave ra-
diation, temperature, and relative humidity in the Swiss Alps, and Partial Cloud
Amount (PCA), which is derived from the CFI. The CFI adopts values >1.0 for
cloud cover and values

SPARC algorithm 24
computational considerations, a shorter spin-up time of only 5 days was selected for
the full disk processing.
The SPARC algorithm was run both in a MSG as well as a MFG mode. While
the MSG mode includes five spectral channels, the MFG mode only relies on MSG
channels 12 and 9 in order to simulate the spectral properties of the MFG sensor
(see Table 4.2 for more information). The restriction to only two spectral channels
aimed at investigating the algorithm’s capability to detect clouds with only a limited
amount of spectral information, analysis that may be useful with regard to the re-
processing of MFG long-term data records. While differences to the MSG mode
should be small at nighttime (when both algorithms rely on thermal scores only),
we would expect some discrepancies during daytime, when information from all
scores except the B-, UTEMP, and T-scores is not available in the MFG mode.
As listed in Table 4.2, channels 12 (HRV) and 1 can be used interchangeably in MSG
mode. If the HRV channel is selected, all other bands are resampled from native
0.06 ◦ ×0.06 ◦ to the higher spatial resolution of the HRV channel (0.02 ◦ ×0.02 ◦ ). This
implies that scores calculated using the HRV channel (Table 4.2) are influenced by
eventual sub-grid clouds that are not represented in the remaining scores to the same
extent. For the analysis presented here the HRV channel was selected, neglecting
uncertainties due to subgrid-scale features.
For the comparison of ASRB measurements with SPARC output, a 9 pixels ×
9 pixels area around each ASRB site was considered. At a spatial resolution of
0.02 ◦ ×0.02 ◦ this approximately corresponds to an area of 15×15 km 2 . As for the
comparison with CFI, all values of F within the subset were averaged to obtain a
single value of F per time slot and ASRB site. In order to compare the cloud mask
with ASRB PCA, former was transformed into partial cloud amount in ninths of
the 15×15 km 2 area. This was achieved by counting the number of pixels classified
as ’cloud’ (thin and opaque) within the 9 pixels × 9 pixels subset. The resulting
82 classes were then stratified into 10 classes of increasing cloud contamination (the
0/9 class only contained entirely cloud free, i.e. 0/81, scenes).

SPARC algorithm 25
Table 4.2: The spectral channels of SEVIRI used as input for SPARC in MSG and MFG
mode, together with the SPARC scores associated with each spectral channel. Channels
12 (HRV) and 1 can be used interchangeably in MSG mode, however, all channels are
resampled to higher spatial resolution (0.02 ◦ ×0.02 ◦ ) if the HRV channel is selected instead
of Channel 1.
Mode
Channel involved scores MSG MFG
HRV (12) B (over land), N X X
1 B (over land), N (X) -
2 B (over water), N, Utext X -
3 R X -
9 T, Utemp, C X X
10 C X -
4.3 Results and Discussion
4.3.1 Comparison with ASRB Cloud Parameters
CFI vs. F
Figure 4.1 displays a scatter-density plot for the comparison of CFI with F for day
(left) and night observations (right; all ASRB sites and months were combined). A
Sun zenith angle (SZA) threshold of 88 ◦ was used to separate day and night obser-
vations. The dotted lines indicate the thresholds used for CFI (Section 4.2.1) and F
(Section 3.1), respectively, to distinguish between cloudy and clear-sky observations
(values of CFI

SPARC algorithm 26
Figure 4.1: Scatter-density plot of CFI vs. F for all sites and 12 months combined for
day (left) and night observations (right). Dotted lines indicate the thresholds used for
CFI (Section 4.2.1) and F (Section 3.1), respectively, to distinguish between cloudy and
clear-sky observations. Colors indicate the frequency of occurence for a certain CFI/F
combination. A bin size of 0.07 (15.0) was selected for CFI (F); r: linear correlation
coefficient.
and night. Very large positive values of F are also observed, yet at very small
frequencies. The agreement for nighttime observations was clearly weaker (r 2 =0.22).
For values of CFI>1.0, F mostly adopts slightly negative values, which is indicative
of an underestimation of nighttime cloud cover by SPARCMSG, even though Dürr and
Philipona (2004) report that the application of the CFI was found to overestimate
nighttime cloud cover. A comparison with synoptic measurements will be needed to
gain additional and more detailed insight into nighttime perormance of SPARCMSG.
Similar to Figure 4.1, the CFI/F relationship is shown in Figure 4.2, but for each
season separately as defined in Section 4.1. For daytime observations, a strong
Figure 4.2 (following page): Similar to Figure 4.1, but for each season separately.

SPARC algorithm 27

SPARC algorithm 28
Figure 4.3: Observation frequencies of all possible PCA (in eights)/CMMSG (in ninths)
combinations for the 12 month validation period during daytime (left) and nighttime
(right). Colors indicate the frequency of occurence for a certain PCA/CMMSG combi-
nation; corresponding percentage values are also provided.
DFI/F relationship was observed in all seasons, even though better agreements were
found for winter and spring (r 2 =0.59 and r 2 =0.61, respectively). Similar to Figure
4.1, predominantly negative values of F are observed where CFI>1.0, showing that
the underestimation of nighttime cloud cover by SPARCMSG relative to ASRB CFI
is observed independently of the season.
PCA vs. SPARCMSG Cloud Mask
Two different analyses were performed in order to validate the SPARCMSG cloud
mask (hereafter referred to as CMMSG) with ASRB PCA. Firstly, frequencies of
occurence of certain PCA (in eights)/CMMSG (in ninths) cloud cover combinations
were analyzed (Figures 4.3 and 4.4). Secondly, the accuracy of CMMSG was ana-
lyzed separately for specific ASRB sites (Figure 4.5 and 4.6). Interestingly, Figure
4.3 shows that in the majority of the cases during daytime (left) and nighttime
(right), either complete clear-sky (PCA: 0/8, CMMSG: 0/9) or complete cloud cov-
erage (PCA: 8/8, CMMSG: 9/9) was observed at the ASRB sites. Partial cloud

SPARC algorithm 29
overcast is only observed in a limited number of cases, with

SPARC algorithm 30

SPARC algorithm 31
plots suggest a clear-sky conservative behavior of SPARCMSG, meaning that a pixel
is only classified as ’clear-sky’ where cloud evidence is very low. In partial cloud cover
situations, there is likely to be some cloud evidence at a certain pixel if the neigh-
boring pixel is cloud covered, which eventually causes CMMSG to assign the pixel
to the ”thin cloud” or ”opaque cloud” category. This is reflected in the tendency
to classify partial cloud cover as completely cloud covered. At nighttime, results
for partial cloud cover exhibit a more bimodal distribution, with partial cloud cover
being classified as either clear-sky or complete overcast in most cases. Results for
the high elevation site at Weissfluhjoch (Figure 4.6) were similar, however, the per-
centage of accurately classified daytime clear-sky scenes (70%) was lower compared
to the Payerne site. Furthermore, clear-sky scenes were found to be misclassified as
complete overcast in 8.4% of the cases during daytime.
Figure 4.5 (following page): Histogram of CMMSG output for each PCA class for
daytime (top) and nighttime observations (bottom). Numbers represent relative frequencies
[%]. Results are shown for the ASRB site in Payerne (cp. Table 4.1).

SPARC algorithm 32

SPARC algorithm 33
Figure 4.6: Similar to Figure 4.5, but for the ASRB site at Weissfluhjoch (cp. Table
4.1).

SPARC algorithm 34
Figure 4.7: Similar to Figure 4.1 (Section 4.3.1), but for SPARCMFG mode.
4.3.2 MFG Simulation
Similar to Figure 4.1 in Section 4.3.1, Figure 4.7 displays the comparison of CFI
and F for the entire 12 month period, however, for SEVIRI data processed in the
SPARCMFG mode (see Section 4.2.2 for more details). As expected from theory,
differences between both SPARC modes are minor during nighttime. Again, low
values of F at values of CFI>1.0 are indicative of an underestimation of nighttime
cloud cover. For daytime observations the relationship between CFI and F is weaker
for SPARCMFG, with r 2 =0.42 (r 2 =0.52 for SPARCMSG; Figure 4.1).
Note that daytime F adopts smaller positive values for cloudy situations in SPARCMFG
mode due to the smaller number of contributing scores. This becomes clear with a
more detailed look at a SEVIRI sample time slot (8:45 am) on March 20, 2005 (Fig-
ure 4.8). The RST mosaics of both SPARC modes (B1 and B2) generated based on
clear-sky observations of the previous 10 days are quite similar and result in similar
T-scores (C1 and C2; see also Section 2). However, since the R-score is not available
in SPARCMFG, the final SPARC scores (F) clearly differ in magnitide (E1 and E2),
which translates into significantly different cloud masks (F1 and F2). Not only does

SPARC algorithm 35
SPARCMFG classify larger areas as clear-sky compared to SPARCMSG, but it also
assigns more clouds to the ”thin cloud” category.
Following the analysis in Section 4.3.1, the histograms of cloud mask output in
SPARCMFG mode (CMMFG) for each PCA class is displayed in Figures 4.9 and 4.10
for the ASRB sites in Payerne and Weissfluhjoch, respectively. At the Payerne
site, cloud cover appeared to be underestimated by SPARCMFG not only during
nighttime, but also during daytime, with 15% of the PCA 8/8 situations being
classified as clear-sky. This is in contrast to the results in Figure 4.5, where a very
good agreement between PCA (8/8) and CMMSG (9/9) was demonstrated (Section
4.3.1). The same pattern, yet less pronounced, was observed at the Weissfluhjoch
site, with a misclassification (PCA 8/8 vs. CMMFG 9/9) in 3.2% of the cases.
Figure 4.8 (following page): Output of SPARCMSG (B1-F1) and SPARCMFG (B2-F2)
on March 20, 2005 (8:45 am). The HRV channel (A1) and B-score (A2) are displayed
once as they are identical for both SPARC configurations. R-score is not available in
SPARCMFG mode (D2), which results in significantly different cloud masks (0: clear-sky,
1: thin cloud, 2: thick cloud). See the text for further explanations.

SPARC algorithm 36

SPARC algorithm 37
Figure 4.9: Similar to Figure 4.5, but for SPARCMFG mode.

SPARC algorithm 38
Figure 4.10: Similar to Figure 4.6, but for SPARCMFG mode.

SPARC algorithm 39
4.3.3 Validation at Full Disk Level
Similar to Figure 1 in Walther et al. (2009) we compared the output of SPARCMSG
and SPARCMFG at full disk level to the cloud mask produced at the Climate Mon-
itoring Satellite Application Facility (CM-SAF). The 12:00 UTC time slot on June
13 th 2006 was selected. Prior to comparison, SPARCMSG and SPARCMFG output
was transformed into binary cloud masks by combining thin and opaque clouds into
one single class. Similarly, the CMSAF cloud information was converted to a binary
cloud mask as described in Walther et al. (2009). The percentage of total cloud cover
was then determined. The resulting cloud masks are displayed in Figure 4.11. From
a visual comparison cloud cover looks quite similar for SPARCMSG and SPARCMFG,
which is also confirmed by cloud cover percentage of 43% and 42%, respectively.
The CM-SAF cloud mask estimates a cloud coverage of 55% of the full disk.
In order to get more detailed insight into SPARC performance at full disk level,
five subsets were extracted from the full disk, some of them similar to the ROIs
in Walther et al. (2009): Central Europe, North Atlantic, Sahara Desert, Western
Africa, and Southern Atlantic. Results are reported in Figure 4.12. At ROI level,
larger differences between SPARCMSG and SPARGMFG are visible. SPARCMFG dis-
plays a higher cloud cover percentage than SPARCMSG, however, except for the
Sahara ROI, both SPARC algorithms detect significantly less cloud cover compared
to the CM-SAF cloud mask. In the Sahara ROI, the extent of bright desert surfaces
classified as ’cloud’ by both SPARC modes is significant. The reason for the over-
estimation of cloud cover in these areas is twofold: as for SPARCMSG, it is mainly
the very large values of the R-score over bright desert surfaces that translate into
erroneous values of F, since the characteristics of the R-score do not account for the
spectral properties of bright barren land. A possible approach could be to modify
the R-score (Figure 4.13) or to introduce a weighting factor for barren land (similar
to s for snow conditions). The former approach was implemented for the full disk
processing presented here, however, Figure 4.12 suggests that further improvements
are required. Latter approach may be considered for SPARCMFG, as misclassifica-
tion is mainly caused by strongly positive B-scores (R-score cannot be calculated in
MFG mode). The positive B-scores over desert in MFG (and MSG) might also be

SPARC algorithm 40
Figure 4.11: Comparison of the CM-SAF (top, right), SPARCMSG (bottom, left), and
SPARCMFG (bottom, right) cloud masks for the full disk on 13 June, 2008, 12:00 UTC.
The corresponding HRV channel is displayed in the top left panel. The cloud mask data
was transformed into a binary cloud mask (see the text for further explanations). The red
squares (A-E) delineate the selected regions of interest (cp. Figure 4.12). Percent values
indicate the cloud cover percentage on the full disk. Please note that the top and bottom
212 lines were excluded from analysis due to computational constraints, which is also the
reason why cloud cover percentage for the CM-SAF cloud mask differs from the value
provided in Walther et al. (2009). Color legend: green: clear-sky land; blue: clear-sky
water surface; yellow: cloud.

SPARC algorithm 41
Figure 4.12: Comparison of the CM-SAF (second row), SPARCMSG (third row), and
SPARCMFG (fourth row) cloud masks for five selected Regions of Interest (ROIs) within
the full disk. Corresping HRV images are provided in the top row. The ROIs are: A
- Central Europe, B - Northern Atlantic, C - Sahara Desert, D - Western Africa, E
- Southern Atlantic (see also Figure 4.11). Percent values indicate the cloud coverage
within the ROI.
overcome by use of a similar score formula as for the T-score:
B-score = (V IS − ALB − offset)scale (4.1)
instead of: B-score = (V IS − offset)scale,
where ALB is the surface albedo and VIS the reflectance in Channel 1 (REF006)
or 12 (HRV). This would mean that a spin-up for the surface albedo similar to the
spin-up of the surface radiative temperature would need to be implemented. This
would then make the snow-factor obsolete for the B-score. The surface albedo spin-
up might, however, be hampered by rapidly changing surface albedo after snowfall
periods.

SPARC algorithm 42
Figure 4.13: Similar to Figure 3 in Khlopenkov and Trishchenko (2007), but only for the
reflectance in SEVIRI channels 1 (REF006) and 3 (REF016) over bright desert surfaces.
The red line demonstrates how the R-score should be modified in order to account for the
spectral properties of bright desert surfaces.
4.4 Concluding Remarks
The Separation of Pixels Using Aggregated Rating Over Canada (SPARC) algo-
rithm represents an alternative to conventional scene identification algorithms as
it primarily outputs a single cloud contamination rating (F) instead of a cloud
mask with a limited number of categories. The rating itself isessentially calculated
from three sub-scores generated from information on the brightness temperature
(T-score), brightness (B-score), and the reflectance in the visible and short-wave
infrared portion of the electromagnetic spectrum (R-score). Even though the algo-
rithm was originally designed for AVHRR, it may be applied to data from other
multispectral sensors.

SPARC algorithm 43
The application of SPARC to data from the MSG SEVIRI sensor was described in
this report. Three major modifications were presented:
1. In addition to F, SPARCMSG outputs a cloud mask with three classes (clear,
thin cloud, thick cloud) based on thresholds applied to F. The application of
thresholds to F facilitates the transformation of the cloud mask from a clear-
sky conservative to a cloud conservative state. In addition, the cloud mask
error for each pixel is calculated based on F, which may be a valuable source
of information for users of this simplified cloud mask.
2. SPARCMSG is a stand-alone algorithm and does not rely on auxiliary data.
In contrast to the original SPARC version, which uses an external surface
skin temperature dataset as a reference in the calculation of the T-score,
SPARCMSG takes advantage of the high temporal resolution of MSG SEVIRI
to derive a brightness temperature mosaic from SEVIRI clear-sky observations
obtained during a preceeding time interval.
3. A new snow detection module is implemented. Based on two newly defined
sub-scores, it outputs both a daily snow mask as well as a snow score (S-score),
which may be regarded as a measure of snow contamination within a pixel.
For the period from July 2004 to June 2005, SPARCMSG output was compared to
cloud cover information from the Alpine Surface Radiation Budget (ASRB) Net-
work in the Swiss Alps: Cloud Free Index (CFI) and Partial Cloud Amount (PCA).
Results show a good agreement between CFI and F mainly for daytime observa-
tions. However, results also point to an underestimation of nighttime cloud cover by
SPARCMSG. Adjustment of the offset and scale factors used to calculate the single
scores may be required to account for this deficiency.
If only two SEVIRI spectral channels are fed into SPARC to simulate data from
the Meteosat First Generation (MFG), a good agreement between CFI and F is
obtained during daytime. Finally, application of SPARC at full disk level revealed an
underestimation of cloud cover compared to the Climate Monitoring (CM)-Satellite
Application Facility (SAF) cloud mask. Both cloud masks should next be compared
over longer time periods to synoptic cloud observations to evaluate possible biases
and spatiotemporal inconsistencies.

SPARC algorithm 44
Overall, SPARCMSG has proven to provide interesting new opportunities for cloud
detection using the MSG SEVIRI sensor. Results are particularly promising with
regard to the application of SPARC to data of the MFG satellite series, since the
algorithm was found capable of generating cloud inforamtion with only a limited
amount of spectral information from two SEVIRI channels.

SPARC algorithm 45
Bibliography
Doms, G., and U. Schaettler (2002), The nonhydrostatic limited-area model lm
part i: Dynamics and numerics. scientific documentation, deutscher wetterdienst,
offenbach, germany, available online: http://www.cosmo-model.org.
Dürr, B. (2009), Technical report of msg-based ir cloud-index processing chain, Tech.
rep., Sunergy GmbH.
Dürr, B., and R. Philipona (2004), Automatic cloud amount detection by sur-
face longwave downward radiation measurements, J. Geophys. Res., 109 (D5),
D05,201–.
EUMETSAT CM SAF (2009), Algorithm theoretical basis document.
cm-saf product cm-02, cm-08 and cm-14. cloud fraction, cloud type
and cloud top parameter retrieval from seviri, Available online:
http://www.cmsaf.eu/bvbw/generator/CMSAF/Content/Publication/atbd pdf/
SAF CM DWD ATBD CFC CTH CTO SEVIRI 1.0,templateId=raw,
property=publicationFile.pdf/SAF CM DWD ATBD CFC CTH CTO SEVIRI
1.pdf.
Khlopenkov, K. V., and A. P. Trishchenko (2007), SPARC: New cloud, snow, and
cloud shadow detection scheme for historical 1-km AVHRR data over Canada,
Journal Of Atmospheric And Oceanic Technology, 24 (3), 322–343.
Lillesand, T., R. Kiefer, and J. Chipman (2004), Remote Sensing and Image Inter-
pretation, fifth ed., John Wiley & Sons, New York.
Mannstein, H., H. Broesamle, C. Schillings, and F. Trieb (1999), Using a meteosat
cloud index to model the performance of solar thermal power stations, in Eumetsat
Conference, p. 239246, Eumetsat, Copenhagen, Denmark.
Marty, C., and R. Philipona (2000), The clear-sky index to separate clear-sky from
cloudy-sky situations in climate research, Geophys. Res. Lett., 27 (17), 2649–2652.
Mesinger, F., and Coauthors (2006), North american regional reanalysis, Bulletin of
the American Meteorological Society, 87, 343360.

SPARC algorithm 46
Philipona, R., C. Marty, and C. Fróhlich (1996), Measurement of the Longwave
Radiation Budget in the Alps, in IRS96, A. DEEPAK Publishing.
Salomonson, V. V., and I. Appel (2004), Estimating fractional snow cover from
MODIS using the normalized difference snow index, Remote Sensing of Environ-
ment, 89 (3), 351–360.
Schulz, J., P. Albert, H.-D. Behr, D. Caprion, H. Deneke, S. Dewitte, B. Dürr,
P. Fuchs, A. Gratzki, P. Hechler, R. Hollmann, S. Johnston, K.-G. Karlsson,
T. Manninen, R. Müller, M. Reuter, A. Riihelä, R. Roebeling, N. Selbach, A. Tet-
zlaff, W. Thomas, M. Werscheck, E. Wolters, and A. Zelenka (2009), Operational
climate monitoring from space: the eumetsat satellite application facility on cli-
mate monitoring (CM-SAF), Atmospheric Chemistry and Physics, 9 (5), 1687–
1709, doi:10.5194/acp-9-1687-2009.
Walther, A., A. Heidinger, A. Thoss, and R. Roebeling (2009), Report on comparison
study for the EUMETSAT workshop, Ascona, Switzerland.