International Journal of Remote SensingVol. 00, No. 00, DD Month 200x, 1–18

Deriving surface global irradiance over the Alpine region from METEOSAT Second

Generation data by supplementing the HELIOSAT method

B. Durr∗† and A. Zelenka†

†Federal Office of Meteorology and Climatology, MeteoSwiss, Krahbuhlstrasse 58, CH-8044 Zurich, Switzerland

(2007)

In the framework of the Satellite Application Facility for Climate Monitoring (CM-SAF) an upgraded formulation of the HELIOSATsurface global irradiance retrieval scheme is proposed, which is suitable for real-time application to METEOSAT Second Generation(MSG) satellite data. The new scheme includes image georeferencing, pixel-wise snow-cover detection, special treatment of clouds abovesnow and correction of terrain effects over the Alpine region. Results show that the mean bias difference between the revised irradianceestimates from satellite and from ground measurements can be substantially reduced by correctly distinguishing between clouds and snow,and by applying the new cloud-index scheme for clouds blue above snow. The increase of root mean square difference with increasingaltitude can be mainly attributed to the increase of the natural irradiance field variability. We strongly recommend the use of MSG-basedirradiance estimates for locations at more than 4km distance to the next measurement site.

1 Introduction

Among the numerous methods developed for retrieving surface global irradiance, respectively surfaceshortwave downward radiation (SDR hereinafter) from satellite observations, several have found a waytowards systematic application, either in climatology (Pinker et al. 1995) or in resource assessment forsolar energy applications (Renne et al. 1999). The ”HELIOSAT” scheme (Cano et al. 1986), and itsupgrades (Beyer et al. 1996, Hammer et al. 2003, Rigollier et al. 2004), is one of the latter. Togetherwith its descendants (e.g. Perez et al. 2002) it has proven robust and stable in operations under variousspacecrafts, mainly because of its original simplicity, which relies on the almost linear coupling that existsbetween irradiance at the top of the atmosphere and at the surface. Straightforward exploitation of thiscoupling leads to (Schmetz 1989)

τatm(n) = τatm(0)(1 − n) (1)

where τatm(n) and τatm(0) are the atmospheric transmittances for the cloudy and cloud-free sky, respec-tively, while cloudiness is expressed with the cloud index n as

n =αtoa − αtoa,min

αtoa,max − αtoa,min(2)

with αtoa being the instantaneous planetary albedo, while αtoa,min corresponds to a cloud-free, clean anddry sky, and αtoa,max corresponds to a heavily overcast sky. These planetary albedos are radiant fluxdensities (irradiances), whereas the satellite observes spectrally filtered radiances.

Originally the scheme has been developed and tuned for the sole VIS channel of the first generationof the METEOSAT spacecrafts, that is up to METEOSAT-7. Its robustness was mainly due to its self-calibrating character. Here we present one possible, among the ongoing (e.g. Muller et al. 2004, Schulzet al. 2005, Betcke et al. 2006), upgrades of it, which takes advantage of the enhanced spectral, temporaland spatial resolution offered by the METEOSAT Second Generation satellites. The goal is to enhance

∗Corresponding author. Email: bruno.duerr@gmail.com

International Journal of Remote SensingISSN 0143-1161 print / ISSN 1366-5901 online c© 200x Taylor & Francis

http://www.tandf.co.uk/journalsDOI: 10.1080/0143116YYxxxxxxxx

2 Duerr and Zelenka

existing HELIOSAT schemes for applications in mountainous regions, specifically in the Alpine region.This requires identification of snow-covered pixels, and cloud detection above such regions, accounting fordeeply cleft orography with shadow casting and for the altitude-dependence of SDR. Results from thisupdated HELIOSAT scheme will be used in the framework of the CM-SAF (Schulz et al. 2005).

Section 2 gives a review of all needed data and their respective sources. Section 3 addresses the problemof converting the satellite’s radiometer counts into planetary albedo counts, and includes the steps requiredto achieve an accurate, pixel-wise scene identification, which is a pre-requisite for a reliable cloud indexdetermination. This section also describes the cloud-free irradiance model used for τatm(0) and how localhorizon obstructions are accounted for, generally leading to a realistic modelling of SDR in mountainousregions. Compact validation and interpretation of results for hourly mean SDR for the year 2006 are givenin sections 4 and 5, respectively, with deliberate focus on overall performance gains rather than detaileddiscussion of individual benefits of the new parameters and thresholds. Finally, we conclude with commentson further possible improvements of our upgraded HELIOSAT scheme in section 6.

2 Observational and auxiliary data

2.1 Geostationary satellite data

Image data from the geostationary METEOSAT Second Generation (MSG) satellites (Schmetz et al. 2002)METEOSAT-8 and METEOSAT-9 in level 1.5 native format from EUMETSAT’s U-MARF archive is thebasis of our study. METEOSAT-8 data is the predominant source during our investigation period from2004 – 2006, except from 25th of September 2006 to 8th of October 2006, where METEOSAT-9 data isused instead. From the 12 available MSG channels, only a subset of 5 channels is used: VIS0.6, NIR1.6,IR3.9, IR10.8 and HRV (see also Appendix A). All channels except HRV have a nominal spatial resolutionof 3x3km at nadir, whereas HRV has a nominal resolution of 1x1km at nadir, i.e. a standard VIS/IR pixelis covered by 3x3 HRV pixels. This corresponds to an approximate standard pixel size of 3.3 km x 5.4 km,resp. 1.1 km x 1.8 km for HRV over the Alpine region. Our investigation area is restricted to the Alpineregion, see figure 1. The MSG time resolution used for this study is 15 minutes, hence a maximum numberof 96 slots is available for each day. The centre of our investigation area is scanned approximately 3.7minutes before the official slot time, e.g. 12:11:18 UTC for the 12:15 UTC slot. Processing of MSG datais limited to a maximum solar zenith angle of 85◦.

2.2 Digital elevation models

Terrain effects modify SDR in various ways, but mainly through shadow casting by surrounding mountains.Shuttle Radar Topographic Mission (SRTM) 90 m digital elevation model (DEM) data version 3 providedby the CGIAR-CSI GeoPortal at http://srtm.csi.cgiar.org/index.asp is applied in our study to account forthese terrain effects on global irradiance.

Swiss topographical information from the RIMINI 250 m DEM (SFO 1991) is used to replace the heightinformation of the largest Swiss lakes within SRTM. RIMINI-corrected high resolution SRTM data is thenadapted to the satellite spatial scale of the HRV and VIS/IR channels by using the median height z of allheights contained in the pertaining pixels.

2.3 Climatology of Linke turbidity factor

To model the cloud-free shortwave downward radiation (SDRcfr) as described in subsection 3.6, infor-mation about the cloud-free atmospheric transmittance τatm(0) is necessary. Due to the sparse spatialcoverage of transmittance measurements, a climatological mean yearly cycle of the Linke turbidity factorprocessed by Remund et al. (2003) is used instead (from Solar Data resource database at http://www.soda-is.com/eng/services/climat eng.html). Monthly mean Linke turbidity factors for 46.2N, 8.8E (westernSwiss plateau near Payerne) are averaged and applied to the entire investigation area. Monthly meanLinke turbidity values are linearly interpolated to obtain daily values.

MSG based surface global irradiance retrieval 3

2.4 Surface irradiance measurements

Surface irradiance data from the Alpine Surface Radiation Budget (ASRB) network (Marty et al. 2002)serves for the derivation of a MSG pixel classification scheme and for cloud and irradiance validationpurposes. The ASRB network consists of 10 stations between 370 and 3580 meters above sea level (asl)in the Swiss Alps since 1994/95 (see table 1). It measures broadband and hemispherically integratedshortwave downward radiation (SDR) and longwave downward radiation (LDR) as 10 minutes averages.Total cloud-cover except thin cirrus clouds is yielded from the real-time APCADA procedure based onLDR every 10 minutes, 24 hours a day (Durr and Philipona 2004). Missing information about cirrus cloudsduring daytime is gained from a simple SDR based cloud flag (SDRflag). The procedure to derive SDRflag

is explained in detail in Appendix B. Various input parameters and validation results for SDRflag are listedin table 1.

3 Surface global irradiance retrieval from MSG satellite data

The steps for the real-time calculation of SDR from MSG count values are summarised in table 2. Start-ing with georeferencing and orthorectification, we proceed with the correction and normalisation of thesatellite-observed radiance counts and their use in the calculation of the cloud index n. This calculationis only reliable if the minimum counts, which are associated to cloud-free scenes, are smaller than theinstantaneous counts. Snow presence invalidates this assumption and calls for alternative strategies whichare described in subsection 3.5. The deduction chain closes with the description of the cloud-free SDRmodel and how local horizon obstructions are accounted for.

3.1 Georeferencing of MSG-HRV channel

Georeferencing of MSG level 1.5 data is performed by the EUMETSAT ground segment. Bugliaro andMayer (2004) showed that a systematic georeferencing bias with some variation can be found for the HRVchannel. We therefore follow their recommendations to set up a landmark-based georeferencing procedureover the Alpine region for each individual HRV time slot. As our investigation area does not includemaritime coastlines, we use lake-shore information as landmarks instead to georeference HRV images asdescribed in Appendix C. As an example, figure 2 shows an extreme outlier of 8 HRV pixels to the north.The georeferencing offsets found for HRV channel are also used to assign the correspondent offsets for theVIS/IR channels.

3.2 Orthorectification of MSG data

The MSG satellites are seen under an elevation angle of approximately 34◦ – 38◦ from our investigation area.This relatively low viewing angle leads to an important distortion of the HRV image by projecting higherAlpine areas to the North-Northeast. This orography effect can be corrected by using SRTM and satellitegeometry information. The SPT-toolbox version 2.4 (Govaerts et al. 2006) provided by EUMETSAT isapplied to determine the necessary satellite geometry parameters.

3.3 Sun zenith angle normalisation and application of MSG calibration information

As in the original HELIOSAT scheme, a standard Lambertian correction for changing illumination ge-ometry is applied to the HRV georeferenced and orthorectified counts to generate a normalised planetaryalbedo αtoa (Moussu et al. 1989). The nominal sun zenith angle is 50◦. Sun zenith angle and relativeEarth-Sun distance are determined using equations for solar parameters recommended by WMO (2006).

Calibration information included in the header of MSG level 1.5 native files is applied to retrieve the bidi-rectional reflectance factor for VIS0.6 (VIS0 .6BRF) and brightness temperatures for NIR1.6 (NIR1 .6BT),IR3.9 (IR3 .9BT) and IR10.8 (IR10 .8BT) (Govaerts and Clerici 2004). Note that calibrated values are used

4 Duerr and Zelenka

for identification and classification of the MSG pixels for which the straightforward cloud index determi-nation (equation (2)) must be modified. The determination of n itself relies upon the sole HRV counts asbefore, thus further ensuring the robustness of the self-calibrating scheme.

3.4 Classification of MSG pixels

Accurate distinction between cloud-free, cloudy, and snow-covered pixels is the main prerequisite for thecalculation of SDR with the HELIOSAT method, because it shapes the minimum planetary albedo mapαtoa,min and controls the way in which the cloud index n is determined (see subsection 3.5).

Spectral information from different MSG channels can be combined, see section 3.4.1 below, to detectspecific features such as snow-covered pixels. However, some ice clouds and snow have very similar spectralcharacteristics. Therefore, the temporal coherence of the αtoa spatial pattern also needs to be investigatedin order to distinguish between them (see subsection 3.4.2). Fog presence, with or without snow, alsodeserves special attention.

Each MSG pixel runs through the tests described below and listed in table 4 in descending order untilthe first one it passes. If there are less than 3 of the 4 preceding slots available due to the sun zenithangle limitation in the early morning or due to missing data, spectral tests without any temporal testsare applied as listed in table 5. The result is an updated αtoa,min map, a classified pixel map (CPM ) andsnow-covered pixel map (SPM ). While this approach must be applied to each slot individually, mainlybecause of the terrain induced daily cycles, it avoids the otherwise requested trailing window (typically 15to 30 days) techniques for constructing a slowly evolving αtoa,min map. Note that αtoa,min is updated ona day-by-day basis as indicated by index s, i.e. only information from the same slot time is applied dueto the almost constant sun azimuth angle, which leads to a relatively stable illumination geometry overcomplex terrain. We distinguish 11 CPM classes as a result from the MSG pixel tests, including 2 differentclasses for undefined pixels. The different classes and the corresponding tests are listed in table 3. Theclassified pixel map is required for choosing the adequate formulation of the cloud-index n (see section 1and subsection 3.5) and for getting a realistic estimate of SDR over the Alpine region (see subsection 3.7).

3.4.1 MSG spectral information. All channels with standard spatial resolution (VIS0 .6BRF,NIR1 .6BT, IR3 .9BT and IR10 .8BT) are bilinearly interpolated to HRV spatial resolution. Dependingon the current georeferencing offsets, a sub-area is extracted from these interpolated datasets to matchthe dimensions of αtoa.

The normalised difference snow index (NDSI ) and the simple IR-channel cloud index (CLI ) (Derrienand Gleau 2005) combine spectral information of VIS/IR channels and provide an adequate basis formodifying and constructing tests and thresholds for the classification of MSG pixels. They are defined as:

NDSI =VIS0 .6BRF − NIR1 .6BT

VIS0 .6BRF + NIR1 .6BT(3)

CLI =IR3 .9BT − IR10 .8BT

µ(4)

CLI is dependent on the cosine of the solar zenith angle (µ), because channel IR3 .9 also measures some partof the solar spectrum. Spectral and temporal tests and their thresholds (M. De Ruyter de Wildt, personalcommunication, May 2005, De Ruyter de Wildt et al. 2007) were modified or iteratively supplemented byminimising the difference between SDR estimates from satellite and ground measurements as follows:

(i) Calculation of MSG-based SDR from October 2004 – September 2005(ii) Validation of SDR with ASRB 10 minutes averages(iii) Investigation of SDR outliers by visual inspection of single/combined MSG images or temporal evolu-

tion of HRV channel compared to ASRB SDR measurements and cloud information from APCADA(see subsection 2.4) and SDRflag (see Appendix B)

(iv) Addition of new tests or modification of existing tests/thresholds

MSG based surface global irradiance retrieval 5

(v) Re-iterate starting at (i)

Based on this iteration and the subsequent validation (see section 4), a series of tests has been designedand the corresponding thresholds are summarised in tables 4 and 5.

Single spectral channel information can also provide us with pertinent criteria. Indeed, ice clouds andsnow-covered pixels often have different apparent brightness temperatures in the IR atmospheric windowchannel (IR10 .8BT), where ice clouds are colder than the snow-covered surface. Thus, we can empiricallydetermine a minimum surface brightness temperature for snow (Tsnow−min) by fitting a cosine function toIR10 .8BT measurements above snow-free and cloud-free pixels determined by visual inspection at differentaltitudes from October 2004 – September 2005 as follows:

Tsnow−min = 272 − 12 cos(

2π

366d − π

4

)− 4z

1000, (5)

where Tsnow−min is given in Kelvin, d is the current day of the year (d = 1 on 1st of January) and z themedian altitude of the underlying pixel. We further assume that pixels flagged by the lakemask LM arenot allowed to be snow-covered, i.e. major lakes are not allowed to be frozen. A pixel earlier detected assnow-covered keeps its classification as long as test 4 in table 4 or test 11 in table 5 are not fulfilled.

Some of the updates depend on the successful georeferencing of the current slot, otherwise default valuesare used regardless of the test result: αtoa,min,s for αtoa,min,t0, SPM t0−1 for SPM t0 and 0 for CPM t0.αtoa,min,t0 for undefined pixels is processed with the basic trailing window scheme proposed by Zelenka(2001). The idea is to compare αtoa,t0 with αtoa,min,s ± ε, where αtoa must be less than αtoa,min,s + ε andgreater than αtoa,min,s − ε for being accepted to update αtoa,min,s. We slightly modified this scheme byintroducing a variable upper (εup) and lower (εlow) limit for ε based on the current value of αtoa,min,s:

εlow = 35 + 3(

2(αtoa,min,s − 60)100

)(6)

εup = 50 + 4(

2(αtoa,min,s − 60)100

). (7)

3.4.2 Temporal variability of αtoa. Due to the relatively coarse resolution of the MSG visible (exceptHRV) and IR channels, application of spectral information alone is often insufficient to detect cirrus withsnow-like spectral properties or sub-scale convective clouds and fog. Thus a test must be introduced basedon the analysis of the short-term variability of αtoa. The idea is to separate the variability of αtoa causedby the interdependency of the changing sun position and underlying terrain for cloud-free situations fromthe variability of αtoa caused by moving or changing clouds.

First the mean spatial difference of αtoa to the 8 surrounding pixels is calculated for the current and the3 preceding slots:

∆αtoa,t =i=1,j=1∑

i=−1,j=−1

(αtoa,t,i=0,j=0 − αtoa,t,i,j)8

, t = t0, t0 − 1, t0 − 2, t0 − 3 (8)

where i and j indicate indices in column and row direction, respectively. Afterwards the temporal variabilityof ∆αtoa,t is calculated by summarising the absolute differences of ∆αtoa,t between the current and theprevious slots and normalised with respect to αtoa,min,s for snow-free, or (partly) snow-covered pixels:

γ =

∑ m=2

m=0|∆αtoa,t0−m−∆αtoa,t0−m−1|αtoa,min,s

if αtoa,min,s ≥ 0.39αtoa,max,s∑ m=2m=0|∆αtoa,t0−m−∆αtoa,t0−m−1|

0.39·αtoa,max,sif αtoa,min,s < 0.39αtoa,max,s

. (9)

The maximum αtoa,min,s value for snow-free and cloud-free pixels, respectively the minimum αtoa,min,s

6 Duerr and Zelenka

value for only partly snow-covered pixels, has been determined empirically by visual inspection of singleαtoa,min images to a value of 0.39αtoa,max,s. The threshold of γ for testing snow-cover is expressed as:

γlim ={

0.45 − 0.15∣∣ z−z0

1000

∣∣ if µ > 0.500.45 + 3(0.50 − µ) − (0.65 − µ)(

∣∣ z−z01000

∣∣) if µ ≤ 0.50 , (10)

where µ is the cosine of the sun zenith angle, and z0=2000 m, as this altitude shows the largest cloud-freevariability due to exposure and ground albedo in the Alpine area. This observation agrees with the factsthat in most regions 2000 m asl is just above the tree limit and that the intricate system of valleys liesbelow it, while the sole structure above it are the prominent, regularly disposed main ridges.

3.4.3 Determination of αtoa,max. According to equation (2), determination of the cloud-index n re-quires the current maximum normalised planetary albedo αtoa,max. Following the original ideas of Moserand Raschke (1983) and Cano et al. (1986), once a day at 1127 UTC, a trailing 90 days window of spatiallyaveraged αtoa values at a given pixel over the Lakes of Constance, Geneva and Garda is applied to calculatethe 90th percentile of the αtoa distribution (αtoa,p90) for cloudy pixels. Lake areas are favoured for the de-termination of αtoa,max due to their damping effect on convective processes, which increases the chance toobserve stratiform clouds over lakes. Vertically expanded clouds have the disadvantage of overestimatingαtoa due to strong reflection of sunlight at their outsides. For the first 90 days a default value of 388 countsis taken for αtoa,p90, which is the mean αtoa,p90 value from 2004 – 2006. Cloud-free pixels over the lakes areexcluded by removing all values below αtoa,p90,s/3. The spatial averaging is done by applying the followingkernel X to convolve αtoa:

X =

0.0 0.5 1.0 0.5 0.00.5 1.0 2.0 1.0 0.51.0 2.0 3.0 2.0 1.00.5 1.0 2.0 1.0 0.50.0 0.5 1.0 0.5 0.0

/

23

Finally the definitive value for αtoa,max is calculated by multiplying αtoa,p90 by a factor of 1.03 except forpixels with fog (CPM == 1 |CPM == 8), where a factor of 0.96 is used instead to account for the specialreflectivity and absorption properties of fog. Both factors were determined by the iterative proceduredescribed above in subsection 3.4.1.

3.5 Modification of HELIOSAT cloud-index n for clouds above snow

Figure 3 shows time series of αtoa, αtoa,min, αtoa,max and αtoa,p90 at ASRB site SLF-Versuchsfeld for 2006.The current αtoa measurements are often below αtoa,min and αtoa,min can be substantially higher thanαtoa,max during snow-covered periods in winter- and spring time. This seasonal cycle clearly demonstratesthe difficulty introduced by snow-cover in winter and spring for the determination of the cloud-index nfrom equation (2). Indeed, αtoa and even αtoa,max fall substantially below αtoa,min because of the veryhigh albedo especially of fresh snow. Thus the determination of n using equation (2) no longer holds forsnow-covered conditions, where αtoa < αtoa,min as shown in figure 3.

Therefore, for clouds above snow-covered pixels, n is estimated depending on the classified pixel map(CPM ) as follows:

n =

{αtoa−0.48·αtoa,max

αtoa,max+0.41·αtoa,max−0.48·αtoa,maxif (CPM == 1) | (CPM == 8) | (CPM == 10)

1 − αtoaαtoa,min

if CPM == 9(11)

For opaque clouds above snow, the normalised minimum planetary albedo αtoa,min is substituted by a fixedfraction (0.48) of the normalised maximum planetary albedo αtoa,max, and αtoa,max is increased by a fixed

MSG based surface global irradiance retrieval 7

fraction (0.41) to artificially increase the dynamic range. This increase is mandatory in periods of snowmelting (e.g., April and May in figure 3), where snowy and snowless patches equally fill the pixel areaand bring minimum and maximum αtoa too close together. Setting αtoa,min to a comparatively high valueaccounts for part of the multiple reflection effect between snow-covered pixels and clouds above snow,which considerably amplifies SDR and also αtoa.

In contrast to the normal formulation of n in equation (2) we observed a quasi-linear increase of n withdecreasing αtoa for transparent clouds above snow, which αtoa straightforwardly led to the new formulationof n in equation (11) for CPM == 9.

3.6 Modelling of cloud-free global irradiance

The HELIOSAT cloud-index basically provides only information about the presence and the effects ofclouds on SDR. Therefore a model is needed to estimate SDR under a cloud-free sky, but with varyingconcentrations of aerosols and atmospheric water content. For this SDRcfr we use the model of Kastenet al. (1984) based on the Linke turbidity coefficient (see subsection 2.3 above) in its altitude-dependentversion (F. Kasten, personal communication, 1990), which reads

SDRcfr = a1 · I0 · cos(θz) · exp(−a2 · m · (fh1 + fh2 · (TL − 1))), (12)

where I0 is the instantaneous solar constant, m the relative optical airmass, TL the Linke turbidity factor(see also Appendix A), and fh1 and fh2 are altitude-dependent functions describing the decrease of theLinke turbidity factor TL with height. In the original model’s version, a1 and a2 are maintained constantat 0.86 and 0.027, respectively. Ineichen and Perez (2002) noted that a1 and a2 should also exhibit a slightaltitude dependence, because 1 − a1 accounts for the fraction absorbed by the atmosphere, and becausea2 also slightly varies with altitude. Based on the SDRcfr measurements in the ASRB records selected byAPCADA values ≤ 1/8 and SDRflag == 0 for snow-free conditions from 2004–2005, we determined

a1 = 1.74 · 10−5 · z + 0.868 (13)

a2 = 6.81 · 10−6 · z + 0.0387. (14)

Mean monthly values of Linke turbidity TL are not representative for the clearest sky conditions asrepresented by the ASRB SDRcfr values. Therefore we replaced TL in equation (12) with T ′

L = TL + ∆TL,where ∆TL denotes an average Linke turbidity offset for very clear days. We use a value of ∆TL = −0.9(P. Ineichen, personal communication, March 2007).

3.7 Surface global irradiance

The modelled shortwave downward radiation for cloud-free conditions (SDRcfr) is finally combined withthe satellite cloud information represented by the cloud-index n. Considering k ≈ (1− n) as a “clearness”index, SDR is expressed as:

SDR = k · SDRcfr. (15)

For all variants of n described in this paper we us the k(n) formulation by Fontoynont et al. (1997), whichaccounts for deviations from linearity at both ends of the n range.

3.8 Correction of shadow casting and terrain effects

To account for shadow casting and surrounding terrain effects, each constituent of SDR is investigatedseparately to yield the corrected SDRcor:

SDRcor = (σ · SDRdir + fsky · SDRdiffuse) · (1 + α · (1 − fsky)) , (16)

8 Duerr and Zelenka

where σ indicates a binary shadow map, SDRdir the direct horizontal component of SDR, SDRdiffuse thediffuse horizontal component of SDR, α the ground albedo and fsky the sky view factor (Dozier and Marks1987). The latter is estimated under assumption of isotropic distribution of the sky diffuse radiance andis, thus, the result of a semi-analytical integration along the local horizon screen yielded by the SRTMdigital elevation model. To obtain SDRdir and SDRdiffuse from the satellite SDR estimates, an empiricalglobal-to-diffuse model (Skartveit and Olseth 1987) is first applied. The basic idea of this model is toestimate SDRdiffuse as a function of the clearness index kt = τm

atm (see equation (B1)) and the solar zenithangle θz. The model was fitted empirically to measured SDR and SDRdiffuse values.

To avoid the smoothing of SRTM topography to HRV spatial scale, every HRV pixel is sub-divided into4x4 sub-pixels, upon which the following calculations are based. Elevation angles of the local horizon hare determined with SRTM orography on sub-pixel scale. h is calculated for 24 azimuthal sectors with asearch radius of 50km. The binary shadow map σ is based on the classified pixel map CPM and on thecomparison of the sun elevation angle (90◦ − θz) with the local horizon of the closest azimuthal sector:

σ ={

0 if ((90◦ − θz) ≤ h) | (CPM == 3) | (CPM == 4)1 if (90◦ − θz) > h

(17)

For ground albedo α we use a very simple approximation with constant values over land and above snow:

α ={

0.14 if SPM == 00.50 if SPM == 1 (18)

Finally SDRcor is calculated by using equation (16), and SDRcor of the 16 sub-pixel points is averagedarithmetically to obtain SDRcor on HRV pixel scale.

4 Validation of satellite-derived global irradiance over the Alpine region

In this paper we deliberately focus on overall performance gains, because a parameter- and threshold-wisediscussion exceeds the scope of this paper and may, at least partly, be found in the cited, pertaining refer-ences. Thus, we only concentrate on the validation of the impact of snow detection and special treatmentof clouds above snow without any correction of shadow casting and other terrain effects. The validationbases on hourly average SDR estimates for the year 2006, which is completely independent from the set-upperiod (October 2004 – September 2005) of our upgraded HELIOSAT scheme.

Three different experiments are presented to show the effect of missing snow-cover detection and accuratetreatment of clouds above snow-covered pixels introduced in this paper compared to SDR measurements:

(a) No distinction between clouds and snow-covered pixels for αtoa,min retrieval, αtoa,min of snow-coveredpixels is set equal to an average αtoa,min value over land (αtoa,min = 120 counts)

(b) Distinction between clouds and snow-covered pixels for αtoa,min retrieval, but SDR calculation basedon the HELIOSAT standard definition of n (equation (2))

(c) Same as experiment (b), but n determined by equation (11) where adequate

Figure 4 shows the mean bias difference (MBD) and root mean square difference (RMSD) for hourlymean SDR between estimates from MSG and ASRB measurements in increasing order of altitude (see table1). The two low-land sites Locarno (southern side of the Alps) and Payerne (western Swiss plateau) arerarely snow-covered, and MBD and RMSD of the three different experiments are nearly identical. RMSDfor these sites is between 17% – 20%, which is slightly below or similar to other results (e.g. Zelenka et al.1999, Betcke et al. 2006). All elevated sites (except Cimetta on the southern side of the Alps with onlya short period of snow-cover) show an increasing negative SDR bias with increasing altitude and hencesnow-cover duration for experiment (a), where snow-covered pixels are mistaken by interpreted cloudypixels. For the permanently snow-covered site Jungfraujoch the MBD is below -30%, and RMSD morethan 70%. The correct distinction between cloudy pixels and snow-covered pixels substantially reducesMBD and RMSD, especially for high-altitude sites as indicated by experiment (b). A further substantial

MSG based surface global irradiance retrieval 9

reduction of RMSD by about 3% up to 13% is achievable by using the new cloud-index n formulation forclouds above snow-covered pixels as indicated by experiment (c). A small systematic negative MBD biasbetween -2.2% and -7.3% remains for all the sites above 2500 m asl. The absolute MBD values are mostlywithin ±10 Wm−2 except for Eggishorn and Gornergrat.

Figure 5 shows the performance evolution between the three experiments for the mean SDR daily cycleat the ASRB site SLF-Versuchsfeld, which undergoes longer-lasting snow-covered and snow-free episodes.RMSD is improved by up to 50% between experiment (a) and (c) ± 4 hours around solar noon, whereasMBD still shows a systematic underestimation of SDR. The absolute MBD values are mostly within ±30Wm−2 except for the late afternoon in summer, where convective cloudiness upcoming above the closeridges very locally shadows the measuring site.

5 Interpretation of observed RMSD increase with altitude

The overall RMSD between satellite estimates and ground measurements of SDR originates basically fromthree different groups: group (1) includes both the remaining simplifications of the HELIOSAT method andsatellite measurement errors, group (2) ground measurement errors and group (3) the principal differencebetween instantaneous pixels extended in space and single point measurements integrated in time (Zelenkaet al. 1999). For group (2), Marty (2000) found a RMSD of about 1%-2% for ASRB measurements.For group (1) a RMSD in the order of 10%–12% has been estimated for sites without increased spatialheterogeneity due to terrain effects (Zelenka et al. 1999). In the present study, the RMSD is roughlydepending on altitude with values ranging from below 20% at Locarno-Monti (370 m asl) up to 40% atJungfraujoch (3580 m asl) with snow-cover all year long. Now, the question arises whether this enhancedRMSD is dominated by the increase of SDR modelling inaccuracies (group (1)) or by the increase of thenatural SDR field variability in high-mountain areas (group (3)).

To empirically quantify the SDR field variability, hourly mean SDR values from three different ASRBsites are compared to each other for the year 2006. The RMSD between Weissfluhjoch and SLF-Versuchsfeldwith a horizontal distance of about 0.63 kilometres is ≈ 14%, between SLF-Versuchsfeld and Davos (dis-tance: 3.13 km) ≈ 34% and between Weissfluhjoch and Davos (distance: 3.6 km) ≈ 35%. Thus for theaverage HRV pixel extension over the Alpine region of about 1.5 km, a RMSD in the order of 20% is arealistic estimate for the measured SDR sub-pixel variability over often snow-covered high-mountain areas.Adding this estimate to the modelling RMSD of 10%–12% (group (1)) and to the ground measurementRMSD of 1%-2% (group (2)), RMSD values of about 32%–40% for ASRB sites above 2000 m asl canplausibly be explained by the increase in natural SDR sub-pixel variability due to snow-cover. In addition,the results indicate that SDR estimated from MSG is more accurate for a given location in the Alpineregion than ground measurements at more than about 4 km distance.

6 Final comments

The overall agreement between MSG based SDR estimates and ground measurements is already close tothe best possible RMSD values expected for mountainous regions. Nevertheless, our upgraded HELIOSATscheme can still be improved if one or several of the following potential improvements were applied:application of SDR cloud-free model using up-to-date atmospheric input data from NWP or GCM models(e.g. Muller et al. 2004), improved treatment of multiple reflection between snow-covered areas and clouds(requires satellite derived high-resolution ground albedo α over the Alpine region), correction of cloudgeoreferencing errors (correction of parallax error) and application of a spatial averaging scheme to accountfor the ergodicity of spatial (satellite) and temporal (point measurement) averages (Moser and Raschke1983).

10 Duerr and Zelenka

7 Conclusions

The wealth of new information yielded by the MSG spacecrafts has triggered several programmes for re-adapting space-based estimation procedures of surface global irradiance (e.g. Muller et al. 2004, Betckeet al. 2006). The upgrade of the HELIOSAT scheme presented here is especially devised for real-timeretrieval of surface global irradiance in mountainous regions. The enhanced spectral, temporal and spatialresolution offered by the METEOSAT Second Generation satellites allows detection of snow-covered areasand of clouds above them, and notably increases the quality of comparison results with ground measure-ments of the ASRB network in the Swiss Alps. The mean bias difference (MBD) of hourly mean irradianceestimates is within ±3% for all sites except Gornergrat, which is severely influenced by the surroundingpermanently snow-covered mountains. The RMSD increases with altitude from about 17%–20% at twolow-land sites up to 40% at the highest-elevated site Jungfraujoch. The largest part of RMSD increasecan be attributed to the growth of the local irradiance field variability over the Alpine region. Hence theproposed MSG-based irradiance scheme yields realistic estimates of hourly mean irradiance at the ground,even above snow-covered and mountainous areas. We strongly recommend the use of MSG-based irradianceestimates for locations at more than 4km distance to the next measurement site.

AcknowledgementsThis work was part of the EUMETSAT co-sponsored Climate Monitoring Satellite Application Facility (CM-SAF)project, where the Federal Office of Meteorology and Climatology (MeteoSwiss) contributes to as a cooperatinginstitution. The authors express their appreciation to R. Philipona and Ch. Ruckstuhl at MeteoSwiss for providingASRB data, and to R. Muller at DWD, P. Ineichen (CUEPE, Geneva) and Martijn de Ruyter de Wildt (ETH,Zurich) for helpful discussions and suggestions. The CM-SAF project team at MeteoSwiss is also grateful for thesupport by Ch. Appenzeller, Th. Konzelmann and M. Liniger. And we would like to thank S. Fukutome for all hereditorial suggestions.

MSG based surface global irradiance retrieval 11

Appendix A: Nomenclature

MSG EUMETSAT Meteosat Second Generation geostationary satellitesHRV MSG broadband (about 0.4 µm – 1.1 µm) high-resolution visible band (1x1km at nadir)VIS0.6 MSG visible channel centred at 0.635 µm (3x3km at nadir)NIR1.6 MSG near-visible channel centred at 1.64 µm (3x3km at nadir)IR3.9 MSG near-infrared channel centred at 3.9 µm (3x3km at nadir)IR10.8 MSG infrared channel centred at 10.8 µm (3x3km at nadir)DEM Digital Elevation ModelSRTM Shuttle Radar Topography Mission DEMz Median altitude [m] within MSG pixel based on SRTMLM Lake mask based on SRTMI0 1367dsun [Wm−2], instantaneous solar constantdsun Instantaneous relative Earth-Sun distance [1]θz Sun zenith angleµ Cosine of sun zenith anglem Relative optical air mass (Kasten and Young 1989)t0 Index of current time slott0 − 1 Index of previous time slots Index of same time slot as t0, but one or more days old (s = t0 − 96m, m = 1, 2, . . .)CLI IR-channel cloud index based on MSG IR3.9 and IR10.8NDSI Normalised difference snow index based on MSG VIS0.6 and NIR1.6Tsnow−min Minimum snow surface temperature as seen from MSG IR10.8αtoa Normalised HRV planetary albedoαtoa,min Normalised HRV minimum planetary albedoαtoa,max Normalised HRV maximum planetary albedo∆αtoa Mean difference of αtoa to the 8 neighbouring pixelsγ Temporal change of ∆αtoa over the past hourCPM Classified pixel mapSPM Snow-covered pixel map based on CPMτatm Atmospheric transmittancen HELIOSAT cloud indexk(n) Clearness index as a function of n (Fontoynont et al. 1997)TL Linke turbidity∆TL Linke turbidity offset for very clear skiesfh1 exp(−z/8000), with z [m]fh2 exp(−z/1250), with z [m]α Ground albedoSDR Shortwave downward radiationSDRcfr Cloud-free SDRSDRflag Cloud-flag based on SDR measurementsSDRcor Terrain effect corrected SDRSDRdir Direct horizontal SDRSDRdiffuse Diffuse horizontal SDR

Appendix B: Derivation of a SDR-based cloud flag

The basic idea for a cloud flag based on SDR is to detect the signatures of the cloud influence on SDR,i.e. increased temporal SDR variability and reduced SDR values compared to cloud-free skies. To reducethe strong SDR daily cycle, the atmospheric transmittance (τatm) is used here instead of SDR:

τatm =(

SDRI0µ

) 1m

, (B1)

where I0 is the instantaneous solar constant at the top of atmosphere, µ the cosine of the solar zenithangle and m the relative optical air mass (see also Appendix A).

SDRflag is determined based on the current magnitude and temporal variability of τatm over the lasthour, and dependent on the current solar zenith angle:

SDRflag =

0 if (θz < 75) & (τatm ≥ τatm,lim) & (τatm,std ≤ 0.003) & (τatm,range ≤ 0.0075)0 if (θz < 80) & (τatm ≥ τatm,lim) & (τatm,std ≤ 0.005) & (τatm,range ≤ 0.0120)1 otherwise

(B2)

12 REFERENCES

SDRflag = 0 indicates cloud-free conditions.The temporal variability of τatm is characterised by the standard deviation of the residuals (τatm,std) and

the range of the residuals (τatm,range = max(τatm,res) − min(τatm,res)) from a linear fit through the 6 τatm

10 minutes values over the last hour.For testing the magnitude of τatm compared to cloud-free skies, we use a lower τatm limit (τatm,lim)

defined as:

τatm,lim = ∆τatm(d) + (0.98 − ∆τatm(d))(

ti − t

tn − t0

)4

, (B3)

where ∆τatm(d) is a τatm offset dependent on the day of the year, ti roughly indicates the time betweensunrise (t0, θz ≈ 93◦) and sunset (tn, θz ≈ 93◦), and t = t0 + tn−t0

2 the time of local solar noon. ∆τatm(d)is determined for each ASRB site by fitting a cosine function to all τatm measurements with τatm > 0.60and τatm,range < 0.0075 at 1130 UTC from 1999 – 2001. Accordingly, ∆τatm(d) is defined as:

∆τatm(d) = τatm,avg − τatm,ci + τatm,amp cos(ωt − π

12

), (B4)

where τatm,avg is the average measured cloud-free atmospheric transmittance, τatm,ci the mean differenceto the lower 90% confidence interval of the cosine fit, τatm,amp the yearly τatm amplitude and ω = 2π/Pwith period P = 366 days. The corresponding values for all ASRB sites are given in table 1.

Synoptic visual observations of total cloud-cover N provided by MeteoSwiss are used for validation ofsituations detected as cloud-free by SDRflag for 09 UTC, 12 UTC and 15 UTC from 1996 – 1998. Table 1shows that the percentage of synoptic observations with more than two octas total cloud-cover is alwaysbelow 10%. Thus, SDRflag is a useful index to detect cloud-free situations from 10 minutes SDR averagesfor sun zenith angles below 80 degrees.

Appendix C: Georeferencing of MSG HRV data

A lakemask (LM ) including 24 different lakes around the Alpine region is extracted from our SRTMDEM-based z data. This lakemask is then projected according to the altitude of each lake to the geo-graphical position as seen from the geostationary satellite. The dark lake areas are used for a patternmatch procedure, where the HRV raw image is shifted by up to ±20 pixels using 1 pixel steps in eachdirection. Only the darkest HRV pixels (minimum value plus 40 counts) are compared with the lakemask.An example is shown in figure 2. The image with the maximum number of matching HRV and lakemaskpixels is used for further processing as HRV geo. HRV lake pixels from at least two different lakes have to bevisible for matching LM and the HRV image. If there are too few lakes visible due to overcast conditions,time-dependent mean georeferencing offsets ∆i in eastern direction and ∆j in northern direction based onpost-processing of the offsets of the correspondent time-periods are used as default values instead:

(∆i,∆j) =

(−1,−3) if t0 < 01/30/2004(0, 2) if 01/30/2004 ≤ t0 < 11/15/2004(1, 2) if t0 ≥ 11/15/2004

.

References

Height model RIMINI, Swiss Federal Office of Topography, Wabern, Switzerland 1991.Betcke, J., Kuhlemann, R., Hammer, A., Drews, A., Lorenz, E., Girodo, M., Heinemann,

D., Wald, L., Cros, S., Schroedter-Homscheidt, M., Holzer-Popp, Th., Gesell, G., Er-bertseder, T., Kosmale, M., Hildenbrand, B., Dagestad, K.F., Olseth, J., Ineichen, P.,Reise, Ch., Dumortier, D., van Roy, F., Gallego, A.O., Beyer, H.G., Trieb, F., Schillings,

REFERENCES 13

Ch., Hoyer, C., Kronshage, S., Mannstein, H., Bugliaro, L. and Krebs, W., 2006, Energy-Specific Solar Radiation Data from Meteosat Second Generation (MSG): The Heliosat-3 Project. Techni-cal report, Energy and Semiconductor Research Laboratory, Carl von Ossietzky University of Oldenburg,D-26111, Oldenburg, Germany.

Beyer, H.G., Costanzo, C. and Heinemann, D., 1996, Modifications of the Heliosat procedure forirradiance estimates from satellite images. Solar Energy, 56, 207–212.

Bugliaro, L. and Mayer, B., 2004, Study on Quantitative Use of the High Resolution Visible ChannelOnboard the Meteosat Second Generation Satellite, Final Report Phase II. EUMETSAT Study ITT No.03/542, EUMETSAT, Am Kavalleriesand 31, D-64295 Darmstadt, Germany.

Cano, D., Monget, J.M., Albuisson, M., Guillard, H., Regas, N. and Wald, L., 1986, A methodfor the determination of the global solar radiation from meteorological satellite data. Solar Energy, 37,31–39.

Derrien, M. and Gleau, H.L., 2005, MSG/SEVIRI cloud mask and type from SAFNWC. Int. J. RemoteSens., 26, 4707–4732.

De Ruyter de Wildt, M.S., Seiz, G. and Grun, A., 2007, Operational snow mapping using multi-temporal Meteosat SEVIRI imagery. Remote Sensing Environment, 109, 29–41.

Dozier, J. and Marks, D., 1987, Snow mapping and classification from Landsat Thematic Mapper.Annals of Glaciology, 9, 97–103.

Durr, B. and Philipona, R., 2004, Automatic Cloud Amount Detection by Surface Longwave DownwardRadiation Measurements. J. Geophys. Res., 109.

Fontoynont, M., Dumortier, Heinemann, D., Hammer, A., Olseth, J., Skartveit, A., Ine-ichen, P., Reise, C., Page, J., Roche, L., Beyer, H.B. and Wald, L., 1997, SATELLIGHT—processing of METEOSAT data for the production of high quality daylight and solar radiation availableon a world wide web internet server. Technical report SoDa-5-2-4, CNRS-ENTPE Mid-term progressreport, jor3-ct90041.

Govaerts, Y., Wagner, S. and Clerici, M., 2006, SEVIRI native format pre-processing toolbox. Tech-nical report EUM/OPS-MSG/TEN/03/0011, EUMETSAT, Am Kavalleriesand 31, D-64205 Darmstadt,Germany.

Govaerts, Y.M. and Clerici, M., 2004, MSG-1/SEVIRI Solar Channels Calibration CommissioningActivity Report. Technical report EUM/MSG/TEN/04/0024, EUMETSAT, Am Kavalleriesand 31, D-64205 Darmstadt, Germany.

Hammer, A., Heinemann, D., Hoyer, C., Kuhlemann, R., Lorenz, E., Muller, R. and Beyer,H., 2003, Solar energy assessment using remote sensing technologies. Remote Sens. Environ., 86, 423–432.

Ineichen, P. and Perez, R., 2002, A new airmass independent formulation for the Linke turdiditycoefficient. Solar Energy, 73, 151–157.

Kasten, F., Dehne, K., Behr, H.D. and Bergholter, D., 1984, Spatial and temporal distribution ofdiffuse and direct solar radiation in Germany. Technical report Research Rept. No. T 84-125, GermanFederal Ministry of Research and Technology 128 pp.

Kasten, F. and Young, A.T., 1989, Revised optical air mass tables and approximation formula. Appl.Opt., 28, 4735–4738.

Marty, Ch., 2000, Surface radiation, cloud forcing and greenhouse effect in the Alps. PhD thesis, SwissFederal Institute of Technology Zurich.

Marty, Ch., Philipona, R., Frohlich, C. and Ohmura, A., 2002, Altitude dependence of surfaceradiation fluxes and cloud forcing in the Alps: results from the alpine surface budget network. Theor.Appl. Climatol., 72, 137–155.

Moser, W. and Raschke, E., 1983, Mapping of global radiation and of cloudiness from METEOSATimage data. Meteor. Rundsch., 36, 33–41.

Moussu, G., Diabate, L., Obrecht, D. and Wald, L., 1989, A method for the mapping of theapparent ground brightness using visible images from geostationary satellites. Int. J. Remote Sens., 10,1207–1225.

Muller, R., Dagestad, K.F., Ineichen, P., Schroedter-Homscheidt, M., Cros, S., Dumontier,

14 REFERENCES

D., Kuhlemann, R., Olseth, J.A., Piernavieja, G., Reise, C., Wald, L. and Heinemann, D.,2004, Rethinking satellite-based solar irradiance modelling – The SOLIS clear sky module. RemoteSensing of the Environment, 91, 160–174.

Perez, R., Ineichen, P., Moore, K., Kmiecik, M., Chain, C., George, R. and Vignola, F., 2002,A New Operational Satellite-to-Irradiance Model. Solar Energy, 73, 307–317.

Pinker, R.T., Frouin, R. and Li, Z., 1995, A Review of Satellite Methods to Derive Surface ShortwaveIrradiance. Remote Sens. Environ., 51, 108–124.

Remund, J., Wald, L., Lefevre, M., Ranchin, T. and Page, J., 2003, Worldwide Linke turbidityinformation. In Proceedings of the ISES Solar World Congress, 16-19 June, Goteborg, Sweden.

Renne, D.S., Perez, R., Zelenka, A., Whitlock, C. and DiPasquale, R., 1999, Use of Weatherand Climate Research Satellites for Estimating Solar Resources. In Advances in Solar Energy, D.Y.Goswami and K.W. Boer (Eds), 13, ASES, Boulder, Co., chap. 5.

Rigollier, C., Lefevre, M. and Wald, L., 2004, The method Heliosat-2 for deriving shortwave solarradiation from satellite images. Solar Energy, 77, 159–169.

Schmetz, J., 1989, Towards a surface radiation climatology: Retrieval of downward irradiances fromsatellites. Atmospheric Research, 23, 287–321.

Schmetz, J., Pili, P., Tjemkes, S., Just, D., Kerkmann, J., Rota, S. and Ratier, A., 2002, Anintroduction to METEOSAT second generation (MSG). Bull. Amer. Meteorol. Soc., 87, 977–992.

Schulz, J., Albert, P., Behr, H.D., Dewitte, S., Durr, B., Gratzki, A., Hollmann, R., Karls-son, K.G., Manninen, T., Muller, R., Roebeling, R., Selbach, N., Sheldon, S., Tetzlaff,A., Thomas, W., Werscheck, M. and Zelenka, A., 2005, Operational climate monitoring fromspace: the satellite application facility on climate monitoring. In Proceedings of the 2005 EUMETSATMeteorological Satellite Conference, Dubrovnik, Croatia.

Skartveit, A. and Olseth, J.A., 1987, A model for the diffuse fraction of hourly global radiation. SolarEnergy, 38, 271–274.

WMO, 2006, Guide to Meteorological Instruments and Methods of Observation. Technical report 8, WorldMeteorological Organization, 41 avenue Giuseppe Motta, CP-2300, CH-1211 Geneva, Switzerland.

Zelenka, A., 2001, Estimating insolation over snow covered mountains with METEOSAT’s vis-channel: Atime series approach. In Proceedings of the EUMETSAT Meteorological Satellite Data Users’ Conference,EUMETSAT Publ. EUM P 33, Antalya, Turkey, pp. 346–352.

Zelenka, A., Perez, R., Seals, R. and Renne, D., 1999, Effective accuracy of satellite-derived hourlyirradiances. Theor. Appl. Climatol., 62, 199–207.

Table 1. Surface global irradiance data from the Alpine Surface Radiation Budget (ASRB) network sites, shortname, altitude,

three different parameters used for calculation of the minimum atmospheric transmittance τatm for cloud-free skies and validation

results for SDRflag. The percentage of manually observed total cloud-cover (N) observations above two octas at synoptic times

is given, when SDRflag is indicating cloud-free conditions for different sites where manual observations were available from 1996

– 1998.

Altitude τatm,avg τatm,ci τatm,amp Cases with N > 2/8Site Abbr. [m asl] 9 UTC [%] 12 UTC [%] 15 UTC [%]

Locarno-Monti LOM 370 0.821 0.039 0.049 4.7 6.8 1.7Payerne PAY 490 0.831 0.016 0.042 2.2 9.4 6.0Davos DAV 1610 0.873 0.032 0.050 - 7.3 -Cimetta CIM 1670 0.859 0.031 0.045 - - -Mannlichen MAE 2230 0.885 0.027 0.038 - - -SLF-Versuchsfeld VSF 2540 0.898 0.034 0.043 - - -Weissfluhjoch WFJ 2690 0.897 0.033 0.039 - 3.3 -Eggishorn EGH 2890 0.904 0.028 0.037 - - -Gornergrat GOR 3110 0.913 0.019 0.035 - - -Jungfraujoch JFJ 3580 0.919 0.028 0.028 1.6 7.4 4.9

Table 2. Overview of surface global irradiance retrieval out of Meteosat Second Generation (MSG) satellite data. Improvements of the

HELIOSAT scheme suggested in this paper are highlighted.

Section Summary Input Output

2.1 Retrieval of MSG level 1.5 native data - EUMETSAT U-MARF archive - HRV- VIS0.6, NIR1.6- IR3.9, IR10.8

3.1 Georeferencing of MSG-HRV channel bylake-shore information

- MSG VIS/IR channels- LM

- HRV geo

- VIS |IRgeo

3.2 Orthorectification of MSG data with the helpof SRTM DEM information

- HRV geo

- VIS |IRgeo

- z

- HRV geo|ortho

- VIS |IRgeo|ortho

3.3 Sun zenith angle normalisation and applicationof MSG calibration information

- HRV geo|ortho- θz

- αtoa

VIS: bidirectional reflectance factor (BRF) andIR: brightness temperature (BT)

- VIS |IRgeo|ortho- MSG calibration

- VISBRF

- IRBT

3.4 Heuristic MSG scene classification:threshold tests using spectral andtemporal information; definition ofautomatic αtoa,max,t0 determinationscheme

- αtoa

- αtoa,min,s, αtoa,max,s

- SPM t0−1

- VISBRF, IRBT

- µ

- αtoa,min,t0, αtoa,max,t0

- SPM t0

- CPM t0

3.5 Extension of HELIOSAT cloud-indexn formulation for clouds above snow-covered pixels

- αtoa,min,t0, αtoa,max,t0

- CPM t0

- SPM t0

- n

3.6 Improved altitude-dependence of clima-tological cloud-free shortwave downwardradiation (SDRcfr) model

- θz

- TL

- SDRcfr

3.7 Application of upgraded HELIOSAT schemeto calculate the shortwave downward radiation(SDR)

- SDRcfr

- n- SDR

3.8 Correction of shadow casting and terraineffects on SDR to obtain SDRcor

- SDR- z

- SDRcor

15

Table 3. Classified pixel map (CPM ) classes used in this paper with the corresponding test

numbers from tables 4 and 5.

Class Name of CPM class Test

0 Undefined pixel, without underlying snow-cover 8, 141 Undefined pixel, snow-covered 8, 142 Cloud-free, and snow-free pixel 4, 113 Cloud, or terrain shadow casting, without underlying snow-cover 3, 104 Cloud, or terrain shadow casting, snow-covered pixel 35 Snow-covered pixel without cloud-cover 1, 2, 96 Fog, without underlying snow-cover 67 Opaque cloud, without underlying snow-cover 7, 138 Fog, snow-covered pixel 69 Transparent cloud, snow-covered pixel 5, 1210 Opaque cloud, snow-covered pixel 7, 13

Table 4. Cascade of MSG pixel classification tests using spectral and temporal information to detect pixels with/without snow-cover,

clouds and terrain/cloud shadows. Generally pixel information from the current slot t0 is applied. Fields from the same slot which are

one or more days old are indicated by s, and the latest fields are indicated by t0 − 1, t0 − 2, . . .. Updating of some fields is depending on

georeferencing success (see subsection 3.1 and Appendix C), otherwise default values are used instead: αtoa,min,s for αtoa,min, SPM t0−1

for SPM and 0 for CPM . The minimum snow surface brightness temperature (Tsnow−min) is defined in equation (5).

Nr. Test Field update

1 Snow (strict test)[γ < (γlim − 0.10)

]&

[(αtoa > 0.48αtoa,max)|(αtoa < 0.24αtoa,max)

]&[

VIS0 .6BRF > 0.40]&

[NDSI > 0.30

]&

[LM == 0

]&

[CLI < 5

]&

[θz < 80

]&[

IR10 .8BT > (Tsnow−min + 5)]

αtoa,min = α∗toa

CPM ∗ = 5SPM ∗ = 1

2 Snow (relaxed test)[γ < γlim

]&

[(αtoa > 0.39αtoa,max)|(αtoa < 0.24αtoa,max)

]&[

VIS0 .6BRF > 0.25]&

[NDSI > 0.15

]&

[LM == 0

]&

[CLI < 10

]&

[θz < 80

]&[

IR10 .8BT > Tsnow−min

]&[

(SPM t0−1 == 1)| ((CPM t0−1 == 7)&(CPM t0−2 == 7)&(CPM t0−3 == 7))]

αtoa,min = α∗toa

CPM ∗ = 5SPM ∗ = 1

3 Cloud- and terrain shadow without/with underlying snow-cover[γ ≥ (0.16 − 0.02µ/0.2)

]&

[(αtoa − αtoa,min,s) < −10

]&

[(αtoa ≤ 0.22αtoa,max)

] CPM ∗ = 3|4†

4 Cloud- and snowfree[γ < (0.16 − 0.02µ/0.2)

]&

[αtoa < 0.48αtoa,max

]&

[VIS0 .6BRF ≤ 0.20

]&[

(NDSI ≤ 0)|(αtoa,min,s < 0.24αtoa,max)]&

[CLI < 4

]αtoa,min = α∗

toaCPM ∗ = 2SPM ∗ = 0

5 Transparent clouds with underlying snow-cover[γ < (0.16 − 0.02µ/0.2)

]&

[SPM == 1

]&

[VIS0 .6BRF ≤ 0.20

]&

[NDSI > 0.20

]&[

IR10 .8BT > Tsnow−min

]&

[(αtoa − αtoa,min,s) < 50

]&

[CLI ≥ 10

]&

[CLI < 30

]CPM = 9

6 Fog without/with underlying snow-cover[γ < (0.21 − 0.02µ/0.2)

]&

[(αtoa > 0.39αtoa,max)

]&

[(αtoa ≤ 0.96αtoa,max)

]&[

VIS0 .6BRF > 0.30]&

[CLI > 15

]&

[CLI < 70

]&

[IR10 .8BT > 265.15

]CPM = 6|8†

7 Opaque clouds without/with underlying snow-cover[(CLI ≥ 10)|(IR10 .8BT ≤ Tsnow−min)

] CPM = 7|10†

8 Undefined MSG-HRV pixel without/with underlying snow-cover CPM = 0|1†

∗Successful georeferencing of current slot required for update, otherwise default values are used

†if SPM == 1

16

Table 5. Cascade of MSG pixel classification tests using spectral information only to detect pixels with/without snow-cover, clouds and

terrain/cloud shadows. Generally pixel information from the current slot t0 is applied. Fields from the same slot which are one or more

days old are indicated by s, and the latest fields are indicated by t0− 1, t0− 2, . . .. Updating of some fields is depending on georeferencing

success (see subsection 3.1 and Appendix C), otherwise default values are used instead: αtoa,min,s for αtoa,min, SPM t0−1 for SPM and 0

for CPM . The minimum snow surface brightness temperature (Tsnow−min) is defined in equation (5).

Nr. Test Field update

9 Snow (strict test)[(αtoa > 0.48αtoa,max)

]&

[NDSI > 0.30

]&

[LM == 0

]&

[CLI < 5

]&

[θz < 80

] αtoa,min = α∗toa

CPM ∗ = 5SPM ∗ = 1

10 Cloud- and terrain shadow without underlying snow-cover[(αtoa − αtoa,min,s) < −20

]&

[SPM == 0

]&

[(αtoa ≤ 0.19αtoa,max)

] CPM ∗ = 3

11 Cloud- and snowfree[αtoa < 0.39αtoa,max

]&

[CLI < −2

] αtoa,min = α∗toa

CPM ∗ = 2SPM ∗ = 0

12 Transparent clouds with underlying snow-cover[SPM == 1

]&

[NDSI > 0.20

]&

[IR10 .8BT > Tsnow−min

]&[

(αtoa − αtoa,min,s) < 50]&

[CLI ≥ 10

]&

[CLI < 30

]CPM = 9

13 Opaque clouds without/with underlying snow-cover[(CLI ≥ 10)|(IR10 .8BT ≤ Tsnow−min)

] CPM = 7|10†

14 Undefined MSG-HRV pixel without/with underlying snow-cover CPM = 0|1†

∗Successful georeferencing of current slot required for update, otherwise default values are used

†if SPM == 1

17

Figure 1. Topography of investigation area over the Alpine region centered over Switzerland in satellite projection. HRV pixelsinvisible to the MSG satellites (more than 20% of pixel area invisible) are coloured in white.

Figure 2. Example for georeferencing error of HRV image. The black lake areas show the lakemask LM used for automatic slot-wisegeoreferencing of the HRV channel.

18

010

020

030

040

050

060

0

Nor

mal

ized

pla

neta

ry a

lbed

o [c

ount

s]

Jan Mar May Jul Sep Nov Jan

αtoa over snowαtoa without snowαtoa,min

αtoa,max

αtoa,p90

Daily normalized planetary albedo values at Versuchsfeld for 2006

Figure 3. Different normalised planetary albedo values at solar noon (1127 UTC) used for determination of cloud-index n at ASRBsite SLF-Versuchsfeld. The original n formulation (equation (2)) is not valid anymore for periods with snow-cover, and the modified n

formulation (equation (11)) is used instead.

LOM PAY DAV CIM MAE VSF WFJ EGH GOR JFJ

Comparison of hourly mean SDR from MSG and ASRB for 2006

MB

D a

nd R

MS

D [%

]

−40

−20

010

3050

7090

LOM PAY DAV CIM MAE VSF WFJ EGH GOR JFJ

−40

−20

020

4060

8010

0

1.3 2.4 −0.8 0.7 2.7 −1.6 −2.0 −3.0 −6.1 1.4 %4.5 7.8 −2.9 2.7 9.5 −5.9 −7.7 −11.7 −27.0 5.0 Wm−2

3779 3882 3871 3863 3872 3876 3872 3869 3872 2888 #

Figure 4. Comparison of hourly mean SDR estimates based on MSG data with ASRB measurements in 2006. The image shows themean bias difference (MBD) as bars for experiment (a) in black (mistaking snow as clouds), experiment (b) in white (detection of

snow-covered pixels, but without usage of special cloud index n above snow) and experiment (c) in grey (usage of special cloud index nabove snow). The numbers in the first row indicate MBD for experiment (c), the second row shows MBD for experiment (c) in absolute

units (Wm−2) and the third row the total number of hours involved. The root mean square difference (RMSD) is plotted as linesstarting from zero.

19

5 6 7 8 9 10 11 12 13 14 15 16 17 18

Comparison of mean SDR daily cycle from MSG and ASRB SLF−Versuchsfeld for 2006

UTC

MB

D a

nd R

MS

D [%

]

−50

−30

−10

1030

5070

90

5 6 7 8 9 10 11 12 13 14 15 16 17 18

−40

−20

020

4060

8010

00.8 3.5 2.9 0.3 −2.4 −3.2 −5.4 −5.3 −6.2 −2.7 0.8 7.8 23.0 39.0 %0.8 5.8 7.2 0.9 −9.8 −15.7 −28.9 −28.9 −31.8 −11.4 2.5 20.3 38.4 33.2 Wm−2

52 152 224 294 360 360 360 360 360 359 357 268 206 137 #

Figure 5. Comparison of mean SDR daily cycle based on hourly mean MSG data with measurements at ASRB site SLF-Versuchsfeldin 2006. The image shows the mean bias difference (MBD) as bars for experiment (a) in black (mistaking snow as clouds), experiment

(b) in white (detection of snow-covered pixels, but without usage of special cloud index n above snow) and experiment (c) in grey(usage of special cloud index n above snow). The numbers in the first row indicate MBD for experiment (c), the second row showsMBD for experiment (c) in absolute units (Wm−2) and the third row the total number of hours involved. The root mean square

difference (RMSD) is plotted as lines starting from zero.

20