The European geostationary Meteosat Second Generation (MSG) satellite offers a variety of channels to use for various purposes, including nowcasting of convection. A number of applications have also been developed to make use of these new capabilities for nowcasting, especially for the detection and prediction of severe weather. The MSG infrared channel selection makes it possible to assess the air stability in preconvective, that is, still cloud-free, conditions. Instability indices are traditionally derived from radiosonde profiles. Such indices typically combine measures of the thermal and moisture properties and often only use a small quantity of vertical profile parameters. MSG-based temperature and moisture retrievals are used for the derivation of stability indices, which are a part of the MSG meteorological products derived centrally at the European Organisation for the Exploitation of Meteorological Satellites (EUMETSAT). Such indices are of an empirical nature, are often only applicable to certain geographic regions, and their thresholds are dependent on seasonal variation, but they can assess the likelihood of convection within the next few hours, thus providing a warning lead of about 6–9 h. Numerous test cases and the more quantitative verification process that was initiated by the South African Weather Service show the generally good warning potential of the derived instability fields. The added capability of a nearly continuous monitoring of the instability fields that is guaranteed by MSG’s 15-min repeat cycle is most valuable, since it provides nowcasters with new information much more regularly than the twice-a-day soundings at only a limited number of radiosonde stations. The current EUMETSAT instability product is aimed at helping forecasters to focus their attention on a certain region, which can then be monitored more closely with other means, like satellite imagery and radar data, over the next few hours.
Since 2002, the European Organisation for the Exploitation of Meteorological Satellites (EUMETSAT) has operated an advanced generation of a geostationary imager, the Spinning Enhanced Visible and Infrared Imager (SEVIRI) instrument on board the Meteosat Second Generation (MSG) satellites. Two of the MSG satellites have been launched and are identified under their operational names: Meteosat-8 and Meteosat-9. The SEVIRI instrument was specially designed to provide enhanced nowcasting information by scanning the earth’s disk every 15 min in 11 spectral channels; an additional high spatial resolution broadband visible channel is also available (Schmetz et al. 2002). The channel selection is based on the well-known Advanced Very High Resolution Radiometer (AVHRR) imagers of polar orbiters, together with additional channels to allow for the detection of specific surface, cloud, and atmospheric features, like identification of fog, dust storms, fires, air masses, and cloud microphysical parameters. Table 1 gives a summary of the MSG channels.
Since the launch of MSG, a number of applications have been developed to make use of these new capabilities for nowcasting, especially for the detection and prediction of severe weather.
As a correct and timely prediction of convective processes—especially their severe stages—is an important area of nowcasting, many MSG-based observations of clouds and their temporal evolution have been used in both qualitative and quantitative aspects to identify the most severe parts of a convective cloud system (Setvák and Rabin 2005; Rosenfeld and Lensky 2006). The MSG infrared channel selection, however, makes it also possible to assess the air stability in preconvective, that is, still cloud-free, conditions. Air instability indices as single-valued numbers have a long history of evaluating the convective potential of the atmosphere. A comprehensive summary can be found in Peppler (1988). They are usually used such that some convective potential can be inferred if the respective index exceeds a certain threshold. Traditionally derived from radiosonde profiles, such indices typically combine measures of the thermal and moisture properties and often only use a small quantity of vertical profile parameters. Due to the limited spectral resolution, MSG-based temperature and moisture retrievals will only have coarse vertical resolution, which is, however, fully sufficient for the derivation of instability indices, as they typically utilize a lower quantity of observations within a vertical profile (Peppler 1988; Fuhrhop et al. 2000).
It should be noted that such indices are of a highly empirical nature and are often only applicable to certain geographic regions. Also, the above-mentioned threshold criteria may change between regions and seasons. Such indices can only assess the likelihood of convection within the next few hours, and should still be used in combination with other triggering/lifting mechanisms.
A satellite-derived map of instability parameters has the advantage of very good spatial and temporal resolutions—for example, 3-km pixel size and 15-min repeat cycle for MSG—which is a significant improvement over the sparse locations of radiosonde stations with at best two daily soundings.
In the following, we will first outline how the MSG-based instability indices are derived, and the general performance of the results will be shown by a number of case studies. The last section will focus on a more dedicated validation study in terms of real warning potential, where the satellite indices are compared to a later occurrence of lightning as a signature of severe convection.
2. Theoretical background
Atmospheric instability parameters are routinely extracted from the MSG imagery within the Meteorological Products Extraction Facility (MPEF) at EUMETSAT. Since these parameters are provided on a global scale (i.e., the entire MSG field of view), the product has been termed the Global Instability Index (GII). From the algorithm side, the GII parameters can be produced on any spatial scale raging from a single MSG pixel and as averages over n × n pixels. The current operational setup is such that the product is derived as 15 × 15 pixel averages, that is, over an area of approximately 50 × 50 km2. Some of the case studies presented here, however, come from an offline test environment, where 3 × 3 or 5 × 5 MSG pixel averages are computed.
Instability indices are usually calculated from atmospheric profiles of temperature and humidity to provide some information concerning the vertical stability of the atmosphere. Various indices are used by forecasters for different applications and regions, and these indices are defined as the differences in profile parameters at different pressure levels (Galway 1956; Kurz 1993). Such data are usually derived from radio soundings, but a few satellite-derived indices also exist [e.g., derived from Geostationary Operational Environmental Satellites (GOES); Hayden (1988); Rao and Fuelberg (1997)]. The airmass parameters can be used to issue severe weather warnings if the corresponding index exceeds a certain threshold. These thresholds are usually determined empirically and should not be regarded as fixed values; they may vary from season to season and from region to region. A skilled local forecaster is absolutely necessary for a correct interpretation of the provided indices.
The MPEF GII product includes two instability indices, the lifted index and the K index, which will be used in this study, as well as the total precipitable water (TPW) content as a further airmass analysis parameter. The instability indices are defined as
where T is the air temperature at the indicated levels and TD is the observed dewpoint temperature at the indicated levels. The atmospheric layer is potentially unstable for a negative lifted index, and a K index of more than 20°C is a good indicator of unstable conditions.
An operational satellite-based retrieval of the lifted index parameter and the total precipitable water has been performed since 1988 using first the GOES Visible and Infrared Spin Scan Radiometer (VISSR) Atmospheric Sounder (VAS) instrument and, later, the GOES Sounder (Hayden 1988; Huang et al. 1992; Menzel et al. 1998; Dostalek and Schmit 2001; Schmit et al. 2002). Studies have shown that the good spatial and temporal resolutions of the VAS instrument and of the derived parameters provide good potential for the identification of preconvective conditions (Kitzmiller and McGovern 1989). For the GOES data, these results are presented in a pictorial form with color-coded lifted index–precipitable water areas together with an overlay of the apparent cloud-top temperature as provided by the IR 11-μm channel. Animations of these images can be used to visually enhance the areas of high likelihood of strong convection and of cloud growth (information online at http://cimss.ssec.wisc.edu/goes/rt/). The MPEF GII retrieval algorithm is in its theory very similar to the physical retrievals developed for the GOES Sounder instrument, but was completely redeveloped, due to a number of differences in the channels and the details of the applied radiative transfer model. The algorithm only works for clear-sky conditions; that is, no instability information is inferred for cloudy pixels.
The GII retrieval is a “physical” method; that is, it tries to infer an actual temperature and humidity profile from the satellite-observed radiances in a given set of channels. The airmass parameters are then derived from this profile. Purely statistical methods exist, which use regression schemes between the satellite observations and the actual index, where the entire regression is based on a dataset such as radio soundings. The results of such statistical methods are often of poorer quality and have a number of other disadvantages, like the total dependence on a given instrument and training data.
The physical retrieval is often referred to as an optimal estimation or one-dimensional variational data assimilation (1DVAR) type of retrieval. An inversion algorithm is applied to find the atmospheric profile that best reproduces the observations (Rodgers 1976; Hayden 1988; Ma et al. 1999). In general, this is a multisolution problem: a wide range of temperature and moisture profiles as well as of surface conditions may exist that produce identical satellite measurements in such a limited number of spectral channels. A suitable “background” or “first guess” profile is used as a constraint to the solution. As this retrieval problem has an iterative solution, the first guess is fed into the iteration scheme as an initial proposal for a solution. The original first guess is then modified in a controlled manner until its radiative properties (the simulated radiances at the top of the atmosphere for the MSG channels) fit the satellite observations. A typical first-guess field is a short-term forecast. The major limitations of this method are the high computational effort and the fact that the retrieved profiles tend to retain features of the first guess, especially when coupled with the low-spectral resolution.
with x being the observation vector (temperature and humidity profile); n the iteration step, where n = 0 denotes a first-guess or background profile; TB the observed brightness temperature; TB,n the simulated brightness temperature for profile of iteration step n; 𝗦x the covariance matrix of first-guess errors; 𝗞n the weighting function matrix (Jacobians); and 𝗦ɛ the error covariance matrix of observed brightness temperatures and of the radiation model. The scheme does in principle allow for a bias correction of the observed brightness temperatures, TB. This could, for example, be a correction against the radiative transfer model or against the forecast. The results of the presented cases were derived without any bias correction.
In its application to the SEVIRI instrument on board the MSG, the physical retrieval uses six channels: the three longwave radiation window channels (IR8.7, IR10.8, and IR12.0), the two water vapor channels (WV6.2 and WV7.3), and the CO2 channel (IR13.4). Accordingly, the matrix 𝗦ɛ contains the instrument’s temperature noise in these six channels with the uncertainty of the radiation model added.
The background profile is taken from the European Centre for Medium-Range Weather Forecasts (ECMWF) global model forecasts, which are provided on a 1° latitude–longitude grid, from 6-hourly forecasts, which are interpolated in time to the actual image time. The covariance matrix 𝗦x was produced from a global (up to 60°N/S) set of atmospheric profiles (Chevallier 2002) to cover a wide range of natural variability. As the full range of natural variability would leave the retrieval problem too unconstrained, that is, it would find highly unrealistic solutions, the original matrix 𝗦x was downscaled to find a compromise between a realistic constraint and a possibility for the retrieval to find a different solution than the first guess. A further spectrally invariant uncertainty of 0.3 K is added to account for uncertainties within the radiative transfer model. The observation vector x contains the full vertical temperature and humidity profiles and the surface skin temperature as the important lower boundary condition for the IR window channels.
The Radiative Transfer for Television and Infrared Observation Satellite (TIROS) Operational Vertical Sounder (TOVS) model (RTTOVS; Eyre 1991; Saunders et al. 1999) is used as it provides a computationally fast method for deriving the Jacobians 𝗞. In addition, the forward model of RTTOV (version 8.7) is used to derive the simulated MSG temperatures for a given observation vector, or profile, x. The iteration is stopped if the root-mean-square (RMS) difference between the observed and simulated brightness temperatures in the six channels is less than a given threshold, usually set to 1.5 K. A spectral variation between the six individual channels is currently not considered.
Figure 1 shows a visualization of the operational product with its 15 × 15 MSG pixel resolution. The averaging process is done here on the brightness temperature level; that is, every box enters the iteration scheme with the average brightness temperatures. The product is derived for every 15-min MSG repeat cycle, for the full disk up to a satellite viewing angle of 70°. It should be noted that a GII value is assigned to a processing box if 50% or more of the box is cloud free, where the cloud information is taken from the MSG Cloud Mask product (Lutz 2007). The brightness temperatures are only averaged over the cloud-free fields of view.
There are always a number of pixels or processing boxes within each image where the physical method never converges to a solution, that is, where the final RMS difference against the observations is not reached within five iteration steps. This usually occurs over clouds, where the entire scheme is not applicable and thus naturally does not find a good solution. In the MPEF installation, as mentioned above, the GII retrieval actually only runs on the cloud-free pixels, and this information is taken from the MSG pixel-based cloud mask product. The necessary brightness temperatures for the retrieval are then only averaged over the cloud-free parts of the processing element. In principle, the GII can also run in a stand-alone mode, without the underlying cloud mask, as the presence of clouds will immediately lead to a “failed” GII retrieval and is thus automatically masked out from the result. Figure 2 shows an example of this stand-alone GII mode: one of the retrieved parameters (K index) is shown together with the infrared window channel (IR10.8) brightness temperatures for those areas where the retrieval never came to a good solution. The IR10.8 image is shown for reference as a visual conformation that indeed the retrieval failed over clouds; that is, clouds are also mostly directly detected within the retrieval. Only partially cloud-filled pixels will—depending on the degree of partial coverage—either not converge or show a too moist retrieval, a feature that is often observed near cloud edges in a pixel-based product.
A further important input parameter to the retrieval scheme is the correct knowledge of the spectral surface emissivity. This is especially important for the SEVIRI IR8.7 channel, which is centered at 8.7-μm wavelength. Nonvegetated, desert-type surfaces have a rather low surface emissivity of 0.75–0.80 in this channel, and this needs to be accounted for within the retrieval’s radiation model. Figure 3 shows an example of the total precipitable water product as one of the retrieval products. The left panel in Fig. 3 shows the retrieval results using a uniform spectral surface emissivity of 0.95 for all channels. Over most parts of the Saharan Desert the retrieval could not find a suitable solution; in the visualization, this is shown by the corresponding IR channel grayscale. However, using the spectral emissivities from the Global Infrared Land Surface Emissivity (IREMIS) database [information online at http://cimss.ssec.wisc.edu/iremis/; see also Seemann et al. (2008)], which are remapped to the MSG pixel locations, this problem was resolved, as the right panel in Fig. 3 shows. Depending on the location, the inclusion of the correct surface emissivity can also lead to slightly modified results.
As mentioned before, a typical feature of such an inversion scheme is its dependence on the first-guess field and the fact that the retrieved profiles tend to retain the general characteristics of the first-guess profiles. In many cases, the first guess already matches the observations so closely that the profile is not changed at all; that is, the RMS to the first-guess field is already below 1.5 K. However, there are many cases where the first-guess field is quite significantly changed within the retrieval; that is, the satellite adds information. This added information often modifies extreme values and local gradients. Figure 4 shows an example: the ECMWF 12-h forecast is used as the first guess for the retrieval. The forecast K index already shows a zone of moderate instability over Germany, extending to the southeast over Austria and neighboring countries. In the retrieval, this zone is not only slightly enhanced in values, but also shifted toward the east. This is actually supported by local radio soundings, as summarized in Table 2.
Figure 5 shows the typical behavior of the retrieval scheme, in terms of how many iterations of Eq. (2) it takes before a good match with the satellite observation is reached. The area of high K index changes with respect to the forecast over eastern Germany and Poland is well dominated by orange; that is, several iterations were needed here to change the not so appropriate first guess to the final result. Generally speaking, this means that more iterations usually imply a higher deviation from the first guess.
3. GII applications
Section 3a will show a few examples of the GII results, which demonstrate the area of application and possible operational usage. We will focus on a case of high instability, which was in this case a good predictor for strong convection, and we will also show an example of the GII results correctly identifying very stable air masses. Section 3b will then highlight a special application and verification environment that was set up during the convective season in the summer of 2006/07 in South Africa.
a. GII examples for selected test cases
Figure 6 shows the K index result from the operational MPEF dissemination at 0800 UTC 26 October 2006. Each box in this visualization represents a color-coded view of the operational 15 × 15 MSG pixel average values together with the precise K index value written in the box [EUMETSAT (2006, p. 44); and information online at http://www.weathersa.co.za/SUMO]. Again, cloudy areas are shaded according to the corresponding IR10.8 image to give some visual impression of the cloud situation.
Usually, K index values exceeding 30°C are indicative of strong convective potential, and K > 40°C indicates a highly likely chance of convective storm occurrence (Peppler 1988). In this case, K index values of around 40°C were reached just east of the Pretoria–Johannesburg region (in red), where a few hours later a convective storm occurred (Fig. 7).
While numerous test cases and examples for both the K index and the lifted index have been compiled to demonstrate this storm predictive potential of the GII results, users raised the question of whether the GII retrievals were not excessively driven by surface heating processes, which would in turn always produce an unstable atmosphere over an underlying warm or hot surface. The second test case shows such a situation on 31 August 2005 over Poland. This day was characterized as being a very hot summer day, but practically no clouds of a convective nature developed during the day, as seen in the top panel in Fig. 8, which shows the MSG IR10.8 image over the area in question for 1800 UTC, that is, at the end of the day. Correspondingly, both the K index and the lifted index for 1200 UTC indicate very stable conditions (Fig. 8, bottom panel).
b. Verification against lightning data over South Africa
Although the above-presented test cases and examples give a good qualitative impression of the GII results, an objective and more long-term verification and assessment of the nowcast potential is difficult, as radiosondes have their obvious limitations in space and time.
The South African Weather Service (SAWS) has received and displayed the operational GII product since 2005 and is now in the process of incorporating it into their operational nowcast environment. Because of the high convective activity during the summer months and the high level of other technical equipment in support of severe storm nowcasting, SAWS provides an excellent test environment for a GII verification. SAWS operates 12 volume scan mode radar systems across the country (Terblanche et al. 2000). For the past few years, data have also been received from MSG, and SAWS has developed its own software to display the different channels, and combinations thereof, as well as the red–green–blue combinations. In 2005, SAWS commissioned a National Lightning Detection Network using 19 Vaisala lightning detection sensors that provide real-time information on lightning activity associated with convective activity (Gill 2008).
To see how these different remote sensing tools can be used in combination to improve operational nowcasting methods in South Africa, lightning data, which are seen as an observation of real-time convection, are utilized to provide a quantitative evaluation of the GII indicators. Statistics from the summer months have shown that the most frequent time for lightning occurrence is from 1100 to 1800 UTC. Although instability indices like the K index or the lifted index are not specifically designed to provide a measure of the likelihood of lightning, this comparison is nevertheless useful from a user’s perspective: Forecasters are interested to know whether areas of strong instability, as derived from the satellite, indeed experience strong convective processes, within the next few hours, that are at least strong enough to produce cloud-to-ground lightning. The lightning data between these times were compared to the lifted index and K index values between 0400 and 0800 UTC of the same days in order to verify the occurrence of convection together with an early warning from the GII.
Collocation of the lightning and the GII data was done within a 0.5° × 0.5° latitude–longitude box. The GII results were seen as a correct forecast of severe convection if
more than five lightning strikes occurred in the box between 1100 and 1800 UTC,
the K index exceeded 35°C between 0400 and 0800 UTC (for the K index verification), and
the lifted index was less than −5 K between 0400 and 0800 UTC (for the lifted index verification).
Software developed for this purpose has the ability to visualize and compare the GII values with the occurrence of lighting strokes later in the day, as shown in Fig. 9. A calculation of statistical values such as the probability of detection (POD), false alarm ratio (FAR), and accuracy are done in a quantitative manner. It should be noted that this comparison can only be done over the regions covered by the lightning detection network (Fig. 10); that is, GII values outside this area are not considered.
This verification environment was applied to a number of severe storm cases in 2006 and 2007. To further demonstrate the system, we will look at the case of 17 January 2007 in more detail.
Early in the evening of this day a large and severe storm line developed and touched parts of the South African provinces of Gauteng, Mpumalanga, and Northwest, before moving into the greater Johannesburg area. Wind damage occurred in various places and large hail was reported. Very large advertising signboards were dislodged by the wind at Ellis Park in Johannesburg during a soccer match and players were injured. The storm dissipated rapidly north of Johannesburg as it lost its surface support, sparing Pretoria from damage.
The day started with little cloud cover over the country, which means that a good GII coverage could be achieved in the morning. Figure 11 shows the GII results for the K index and the lifted index for 0600 UTC: the maximum instability values occurred where the storm developed—over the Northwest Province and Gauteng—and the storm then moved eastward toward Mpumalanga. The prevailing wind conditions were westerly flow, so the eastward movement could be well anticipated.
Radar reflectivities obtained at 1700 UTC show the situation for the fully developed storm (Fig. 12). Radar reflectivities of more than 50 dBZ are visible in the line of thunderstorms southwest of Pretoria and Johannesburg. Values of 50 dBZ or more normally are associated with heavy thunderstorms, or perhaps with hail, but as with most other quantities, there are no reliable threshold values to confirm the presence of hail or severe weather in a given situation (Branick 2006).
Although the radar reflectivities alone qualitatively verify the early morning GII results as being a good storm indicator, a quantitative comparison was done against the lightning data (Fig. 13).
The results of the evaluations of the early morning K index and the lifted index against the midday and afternoon lightning occurrences are shown in Table 3. “Accuracy” is defined here as
where the CorrectHits is the number of correctly forecasted storm occurrences according to the above index and lightning criteria, and the CorrectNonHits is the number of correctly forecasted no-storm conditions, that is, where none of the above criteria were met.
Using this verification system for five cases in the storm season of 2006/07, we get POD, FAR, and accuracy values as shown in Table 4.
In summary, we see fairly high values for POD, together with low false alarm ratios over South Africa. These findings encourage forecasters to focus their attention on a specific area early in the morning and then watch the convective development more closely with radar as time progresses.
As mentioned above, the MPEF operational GII product uses the forecast fields from the ECMWF model with a 1° latitude–longitude horizontal resolution. A local version of the GII code has recently been installed at SAWS that uses a local mesoscale model (the local version of the Met Office’s Unified Model) with a 0.1° latitude–longitude resolution. It is now also possible to calculate the values for a 3 × 3 MSG pixel block, replacing the coarser 15 × 15 pixel processing areas of the MPEF product. (A one-pixel-based product is currently not possible because of CPU time constraints.) Ongoing verification will be done in the coming summer months to show whether the improved resolution will be of further benefit to the already positive results.
4. Summary and outlook
MSG offers the great opportunity to derive airmass parameters as instability indices and total precipitable water from the measurements in the infrared spectrum. As MSG SEVIRI is far from being an instrument with full sounding capabilities like modern hyperspectral sounders, detailed temperature and humidity profiles, taken from a global or local forecast model, are needed as a first guess or background field for the retrieval. It was shown that the MSG measurements can actually add information over the forecast in terms of local extremes and gradients. The MSG instability product is thus in its underlying algorithm and in the applicability of the results very similar to the respective GOES Sounder products (Menzel et al. 1998; Schmit et al. 2002).
Numerous test cases and the more quantitative verification process that was initiated by the South African Weather Service show the generally good warning potential of the derived instability fields.
Of high value is the added capability of a nearly continuous monitoring of the instability fields that is guaranteed by MSG’s 15-min repeat cycle. Over time, better spatial coverage can be achieved; as the chance of finding a cloud-free area for the retrieval process increases, the evolution of instability over time can be studied. A great benefit is also the fact that the fast repeat cycles allow the observations of very quick changes in time, where the temporal rate change itself can be of further nowcasting importance. A full evaluation of this latter aspect still needs to be performed and will be the subject of future work.
The current EUMETSAT GII product is aimed at helping forecasters to turn their attention to a certain region, which they can then monitor more closely with other means like satellite imagery and radar data over the next several hours. The MSG GII data have proven to provide lead times between 6 and 9 h. Within a fully developed convective nowcasting system, the GII can provide a first level of warning, usually covering a larger area. When time progresses, other satellite-derived parameters like cloud coverage, cloud-top cooling, and growth rates, together with the development of microphysical parameters, will then help to further indicate the exact location of the most severe parts of the clouds (Mecikalksi and Bedka 2006; Rosenfeld and Lensky 2006). These satellite-derived convective products are shown to still have some lead time over radar observations, that is, are even useful when good radar coverage is available, but will be especially beneficial for areas without any radar information. This is of special importance for large parts of Africa, where the availability of radar data is extremely limited. Providing useful nowcasting tools for use in African countries thus relies on satellite and numerical weather prediction model output. MSG with its spectral, spatial, and temporal resolution provides an ideal platform for the development and testing of such an automated system of a series of warning levels.
With respect to the GII product itself, the impact of the product resolution and the resolution of the underlying model are current areas of research between EUMETSAT and SAWS, where the above-mentioned local GII installation at SAWS plays a crucial role. South Africa also acts as a vehicle to the rest of Africa in providing forecasting and nowcasting tools. Through this product and the Software for the Utilization of Meteosat in Outlook Activities (SUMO) display system (which can be downloaded online free of charge at http://www.weathersa.co.za/SUMO/) to display it, African countries have access to tools to aid in their nowcasting procedures as well.
In addition to the here-described initial validation studies with lightning data, a long-term comparison with radiosondes is also being done at SAWS. It is expected that this will provide further insights to help in the interpretation of the product. It should also be mentioned that hyperspectral sounders like the Infrared Atmospheric Sounding Interferometer (IASI) and the Atmospheric Infrared Sounder (AIRS) might to some extent replace or amend the first-guess background fields, that is, may help to loosen the dependence of the product on the forecast fields. Studies in this respect are ongoing.
The authors thank Monika Pajek and Piotr Struzik from the Polish Meteorological Service (IMGW) and Jarno Schipper from the Austrian Meteorological Service (ZAMG) for their contributions to the GII-related research.
Corresponding author address: Marianne Koenig, EUMETSAT, Am Kavalleriesand 31, Darmstadt D-64295, Germany. Email: email@example.com