Validation of a Meteosat Second Generation solar radiation dataset over the northeastern Iberian Peninsula

Solar radiation plays a key role in the Earth’s energy balance and is used as an essential input data in radiation-based evapotranspiration (ET) models. Accurate gridded solar radiation data at high spatial and temporal resolution are needed to retrieve ET over large domains. In this work we present an evaluation at hourly, daily and monthly time steps and regional scale (Catalonia, NE Iberian Peninsula) of a satellite-based solar radiation product developed by the Land Surface Analysis Satellite Application Facility (LSA SAF) using data from the Meteosat Second Generation (MSG) Spinning Enhanced Visible and Infrared Imager (SEVIRI). Product performance and accuracy were evaluated for datasets segmented into two terrain classes (flat and hilly areas) and two atmospheric conditions (clear and cloudy sky), as well as for the full dataset as a whole. Evaluation against measurements made with ground-based pyranometers yielded good results in flat areas with an averaged model RMSE of 65 W m−2 (19 %), 34 W m−2 (9.7 %) and 21 W m−2 (5.6 %), for hourly, daily and monthly-averaged solar radiation and including clear and cloudy sky conditions and snow or ice cover. Hilly areas yielded intermediate results with an averaged model RMSE (root mean square error) of 89 W m−2 (27 %), 48 W m−2 (14.5 %) and 32 W m−2 (9.3 %), for hourly, daily and monthly time steps, suggesting the need of further improvements (e.g., terrain corrections) required for retrieving localized variability in solar radiation in these areas. According to the literature, the LSA SAF solar radiation product appears to have sufficient accuracy to serve as a useful and operative input to evaporative flux retrieval models.


Introduction
Knowledge of spatiotemporal distributions in solar radiation (R s ) is essential in many disciplines such as ecology, agronomy and hydrology, and plays a key role in the modeling of evapotranspiration (ET), both actual and potential, as well as air temperature.These variables are of high importance in monitoring and understanding the ecohydrological properties of terrestrial ecosystems and for agricultural support (Pons et al., 2012).Together with precipitation, ET is an essential variable in the hydrological cycle, and its modeling has been a research challenge over the last several decades (Dickinson, 1984;Manabe, 1969;Monteith, 1965).Currently, there is a wide variety of remote sensing models for calculating ET at regional or global scales that require R s as an input (Allen et al., 1998(Allen et al., , 2007;;Anderson et al., 2004;Bastiaanssen et al., 1998;Cristóbal et al., 2011;Jackson et al., 1977;Kalma et al., 2008;Kustas and Norman, 2000;Priestley and Taylor, 1972;Roerink et al., 2000;Seguin and Itier, 1983).For operational applications, most of these methods try to minimize use of data from ground-based meteorological stations.Therefore, ET algorithms operating at regional to global scales can benefit from R s surfaces retrieved using satellite imaging.Most of these ET methods have been validated in homogeneous covers (crops or natural vegetation) and flat areas, using a single value of R s from a meteorological station record to describe a large area.However, in more complex terrain conditions, a single meteorological record may not be accurate enough to reasonably estimate ET spatially, considering gradients in the spatial distribution of R s due to variable topography and cloud cover.
Published by Copernicus Publications on behalf of the European Geosciences Union.R s is typically estimated using one of three different methodologies: empirical models, based on statistical correlations between R s and other parameters; parametric models, based on the physics of interactions of R s with the atmosphere (Martínez-Durbarán et al., 2009); and hybrid models that combine both approaches.Some of these models use GIS-based techniques and a digital elevation model, DEM, (Pons and Ninyerola, 2008) to compute R s at regional and global scales in both simple and complex areas offering high accuracy and high spatial resolution, but relying on a well developed meteorological station network.In many regions, the density of meteorological stations is sparse and only satellites can realistically provide R s data, especially at global scales (Journée and Bertrand, 2010;Olseth and Skartveit, 2001;Pinker et al., 2005).
Operational satellite systems provide valuable information on atmospheric parameters at regular intervals on a global scale.This satellite-based information greatly enhances our knowledge and understanding of the processes and dynamics within the Earth-atmosphere system.Nowadays, there is a wide variety of satellites, both geostationary and sunsynchronous, from which R s can be retrieved regionally or globally such as Terra/Aqua Moderate Resolution Imaging Spectroradiometer (MODIS), the National Oceanic and Atmospheric Administration's (NOAA) Advanced Very High Resolution Radiometer (AVHRR), the Geostationary Operational Environmental Satellites (GOES) or the Meteosat Second Generation (MSG) Spinning Enhanced Visible and Infrared Imager sensor (SEVIRI).Unlike sun-synchronous sensors, geostationary sensors are especially interesting because of their high temporal resolution, which facilitates mapping of R s at intervals of 15-30 min over large areas.In the case of Europe, there are currently three facilities that produce and offer R s products from 30-min to monthly time steps derived from MSG SEVIRI data that can be used as input data in ET modeling: the Satellite Application Facility on Climate Monitoring, CM-SAF (http://wui.cmsaf.eu/), the Ocean and Sea Ice Satellite Application Facility, OSI SAF (http://www.osi-saf.org/),and the Land Surface Analysis Satellite Application Facility, LSA SAF (http://landsaf.meteo.pt/).
In this work we present a regional-scale evaluation of the LSA SAF R s product, generated using MSG SEVIRI images from 2008 to 2011.The product dataset is evaluated at hourly, daily and monthly time steps, both as a whole and as subsets depending on terrain class (flat and hilly areas) and atmospheric conditions (clear and cloudy skies).In addition, the R s product use as an input to evaporative flux retrievals is briefly discussed, as reported in the current literature.

Solar radiation product and model overview
Since 2007, the LSA SAF has offered an operative product describing the down-welling surface short-wave radiation flux (DSSF), obtained by means of the SEVIRI sensor.The DSSF product preserves the projection and spatial resolution of the MSG SEVIRI images, using the ellipsoid normalized geostationary projection with a nominal spatial resolution of 3 km at nadir.This product is generated at a 30-min time step using data from the three solar spectrum channels of the SE-VIRI sensor (centered on 0.6, 0.8 and 1.6 µm) and is encapsulated in an HDF5 file format.Each product file includes a set of three quality flag images (see Table 1): a land and sea mask, a cloud mask also including snow and ice cover; and the DSSF algorithm that was applied (clear or cloudy sky algorithm).
The model used to retrieve R s for the DSSF product is based on the framework of the OSI SAF (Brisson et al., 1999) using three short-wave SEVIRI channels, 0.6 µm, 0.8 µm, and 1.6 µm (LSA SAF, 2010).The model is designed to compute the effective atmospheric transmittance, applying a clear or cloudy sky retrieval method depending on cloud cover.Cloud cover estimates are provided by the cloud mask developed by the nowcasting and very short-range forecasting, which is integrated in the LSA SAF operational system (Geiger et al., 2008b).
In the case of the clear sky method, the atmospheric transmittance and the spherical albedo of the atmosphere are calculated according to the methodology of Frouin et al. (1989).The water vapour used to estimate the atmospheric transmittance is obtained from the European Center for Medium-Range Weather Forecasts and the ozone amount is specified according to the Total Ozone Mapping Spectrometer climatology, while the visibility is currently kept at a fixed value of 20 km.The surface albedo is taken from the LSA SAF near-real-time albedo product (Geiger, 2008a).
In the case of the cloudy sky method, a simplified physical description of the radiation transfer in the cloudatmosphere-surface system according to Gautier et al. (1980) and Brisson et al. (1999) is used.The cloud transmittance and albedo may be highly variable on small time scales depending on the daily evolution of the clouds.For this purpose the measured spectral reflectances in the 0.6 µm, 0.8 µm, and 1.6 µm SEVIRI are first transformed to broad-band top-ofatmosphere albedo by applying the spectral conversion relations proposed by Clerbaux et al. (2005) and the angular reflectance model of Manalo-Smith et al. (1998).
More information about the DSSF method can be found in Geiger et al. (2008b)  From the XEMA network, 140 meteorological stations measuring R s were selected, applying a filter criterion consisting of stations that have been in service for at least 5 yr (see Fig. 1).For each of these stations SMC applies a data quality process and produces a R s quality flag.The selected meteorological stations are located in different land uses and span a range in altitude (see Table 2), providing a broad basis for comparison with satellite retrievals under different circumstances.
In order to analyze the performance of the DSSF product in different terrain conditions, the meteorological stations were separated into two classes (see Fig. 1): those situated in flat and hilly terrain.This separation was based on a slope surface derived from a 30 m spatial resolution DEM from the Institut Cartogràfic de Catalunya (Cartographic Institute of Catalonia).The standard deviation in topographic slope was computed in a 3-km buffer area around each meteorological station, simulating the resolution of MSG SEVIRI.Slope standard deviation gives information about terrain heterogeneity, and whether or not the meteorological station is surrounded by mountains that might influence shading of the R s sensor.Based on these analyses, a threshold in slope standard deviation was selected to partition the network stations into two sets: 100 meteorological stations in relatively flat terrain, and 40 in hilly terrain.

DSSF product
A total of 1096 days of DSSF products for 2008 to 2010 were downloaded from the LSA SAF web site.A standard day consists of 48 files in HDF5 format, one image every 30 min, although there are days that have fewer files.In total, 52 608 files were downloaded and processed.To minimize impacts of data re-sampling due to reprojection, the analysis was carried out in the original projection and spatial resolution of the DSSF product.

Solar radiation extraction and evaluation criteria
Once the DSSF product was imported, data extraction was performed using bilinear interpolation in time between images and in space to meteorological station locations.Recent work in the literature suggests that averaging over a block of pixels centered on the location of a pyranometer significantly decreases the error compared to use of a single pixel, although there is no agreement on what is the optimal block size (Pinker and Laszlo, 1991;Rigollier et al., 2004;Journée and Bertrand, 2010).Nevertheless, in this work, we are interested in a pixel-based analysis to better capture effects of heterogeneity in the mountainous areas with narrow valleys found in our study area.
In order to manage data efficiently through the use of SQL statements, a database was built for product evaluation.This database consists of two parts: a DSSF record every 30 min, which incorporates both R s and quality flags; and 1 h meteorological records that include measured R s , data quality from the SMC and meteorological station terrain classes (flat or hilly).
DSSF evaluation was conducted using only pixels flagged as clear or cloudy (contaminated and cloud filled conditions) in the DSSF cloud mask and processed by clear and cloudy methods in the DSSF algorithm (see Table 1).Data under undefined and unprocessed categories in the DSSF cloud mask as well as algorithm failed, beyond specified view angle limit and not processed (cloud mask undefined) categories in the DSSF algorithm (see Table 1), representing less that 0.7 % for the whole dataset, were excluded from the analysis to avoid introducing errors in the evaluation analysis due to unreliable data.
In addition, images outside the interval between dawn and dusk (zero insolation) were also excluded from analysis in order not to magnify accuracy statistics.The calculation of dawn and dusk for each day and each meteorological station was carried out using the methodology proposed by Orús et al. (2007).
Based on these criteria, the DSSF product was evaluated with respect to ground observations at hourly, daily, and monthly timescales.This required aggregation of the DSSF product, available at 30 minute intervals, and the SMC R s data, available only at hourly intervals.For hourly evaluation, DSSF algorithm performance was analyzed under both clear and cloudy sky conditions and snow and ice cover.To focus on pixels where cloud conditions were relatively sta-ble over the hourly sampling interval of the SMC dataset, the hourly evaluation was conducted using only pixels reporting the same quality flags during each hourly interval in terms of cloud mask and DSSF algorithm (i.e., a pixel masked as clear both at 12:00 UTC and 12:30 UTC).On the other hand, pixels with different quality flags for a specific hour were excluded in the analysis (i.e., a pixel masked as clear at 12:00 UTC and cloudy at 12:30 UTC).
The relative performance of the clear and cloudy sky algorithms used in the DSSF process was also explored.Currently, there is no agreement on how to best define a clearsky day in terms of amount of cloud-free time.In this study, a clear-sky day at a given pixel was defined such that ≥ 80 % of the time samples between dawn and dusk were cloud free.Moreover, as criteria for computation of daily average R s from the DSSF product, we specified that at least 90 % of the potential images within a day must be available.A similar criterion of completeness was applied to the pyranometer data, but requiring all data samples from dawn to dusk to have a good quality flag.Finally, in the case of monthly aggregation, no distinction was made between clear or cloudy conditions given the length of the averaging interval.A criterion of having daily aggregates in both satellite product and pyranometer datasets for more than 25 days per month was enforced to ensure that these data were representative of monthly conditions.

Accuracy and error estimation
The performance of the DSSF product was evaluated using several statistical indices and measures of error.The coefficient of determination (R 2 ) indicates the precision of the estimates in relation to measured R s , the root mean square error (RMSE, Eq. 1) is used to measure the differences between values predicted by a model or an estimator and the values actually observed and is a measure of accuracy, the mean absolute error (MAE, Eq. 2) indicates the magnitude of the average error, the mean bias error (MBE, Eq. 3) indicates cumulative offsets between measured and observed values, and the percentage of error (PE, Eq. 4) expresses the magnitude of the error between observed and estimated values relative to the observed mean value: where e i refers to the estimated value of the variable in question (satellite-derived R s ), o i is the observed value (in situ R s measurement provided by the meteorological station), n is the number of datapoints, and X is the average of the n o i values.A literature review reveals only three comparative studies that clearly address the issue of terrain conditions on R s product evaluation.The use of different accuracy and error estimators, as well as differences in temporal extent of analysis and thresholding criteria, complicates a detailed comparison of the results presented here with those in other references.Nevertheless, a qualitative comparison with prior results provides some useful context for the current study.
In the case of flat conditions (column 1 in Table 3), the evaluation results presented here are in agreement with those found in the literature.In the DSSF product validation performed by Geiger et al. (2008b), with data of 6 meteorolog- found for clear and cloudy conditions, with MBE of 5 W m −2 in both cases.Improvement in RMSE in the current study may be due to further product improvement in 2010, or to differences in data rejection criteria.Comparing all satellite application facility irradiance products (CM-SAF, OSI SAF and LSA SAF) at 15 km of spatial resolution, including also the DSSF product, Ineichen et al. (2009) found a RMSE between 80 W m −2 and 100 W m −2 and a PE ranging from 15 % to 32 % using 8 meteorological stations over Europe.According to Rigollier et al. (2004), the obtained hourly results are within the error displayed by Heliosat-1 and Heliosat-2 models, also designed for Meteosat images, and reported RMSE errors from 64 W m −2 to 120 W m −2 and a PE from 7 % to 16 % in the case of Heliosat-1 hourly irradiation.In the same work, they reported an hourly irradiation RMSE and PE from 62 W m −2 to 103 W m −2 and from 18 % to 45 %, respectively, using the Heliosat-2 algorithm in three months from 1994 to 1995 and using 35 stations in flat areas, ranging the bias from −31 W m −2 to 1 W m −2 .Using GOES-W and GOES-E data, Otkin et al. (2005) report a PE, a RMSE and a MBE of 19 %, 62 W m −2 and −2 W m −2 , respectively, using observations from 11 meteorological stations from the US Climate Reference Network over a continuous 15-month period at 20 km of spatial resolution; and Garautza-Payan et al. ( 2001) reported similar results in northern Mexico of about 13 % of PE and 69 W m −2 of RMSE in a one year experiment using data from 2 flux towers.
In general terms, the presence of snow or ice yielded in higher MBE, RMSE and lower correlation, especially in hilly sites.Dürr et al. (2010) also found a large negative MBE on the order of 15 W m −2 to 25 W m −2 over the alpine region using the CM-SAF solar radiation product.In this case, the snow and ice surface albedo may be difficult to define, leading to higher errors compared to other land covers and according to Dürr et al. (2010) a dynamic snow-albedo map could improve the solar radiation retrieval in snow covered areas.Still, R 2 values of ∼ 0.8 indicate that useful data are being generated even for these difficult land cover situations.
It is worth remarking that the mean PE obtained in this study at hourly time steps for all atmospheric conditions in flat and hilly classes was 19 % and 27 %, respectively, and according to Zelenka et al. (1999) this value compares favorably with the value from 20 % to 25 % reported world-wide.While PE during cloudy sky conditions exceeds this range, with values of 36 % and 42 % respectively, this is mostly a function of lower mean observed R s during periods of cloud cover.Finally, DSSF values over snow and ice cover also yield a PE in this interval, ranging from 23 % to 27 %.

Daily evaluation
Table 4 shows  Figure 3 shows examples of daily R s dynamics from dawn to dusk at two meteorological stations located in flat (V8) and hilly terrain (UI) during a clear sky day (both on 6 April 2008) and a cloudy sky day (24 September 2008).The upper left plot shows the best-case scenario, for the site in flat terrain under clear sky, while the upper right shows data for the same clear day at the site in hilly terrain.The hilly site shows evidence of topographic shadows between 16:00 UTC and 19:00 UTC, while the 3-km average does not show a strong diurnal shadowing effect.According to a sensitivity analysis by Oliphant et al. (2003), to isolate the role of spatial variability of surface characteristics in generating variance in the radiation budget, one of most important characteristics was found to be slope aspect.This fact suggests that in hilly sites, the DSSF algorithm could be enhanced to reproduce sub-pixel variability in shadowing effects by accounting for topography using a DEM.
The lower plots in Fig. 3 show daily R s dynamics under cloudy sky conditions for both flat and hilly sites.As seen in Table 4, the accuracy of the DSSF algorithm is lower under cloud cover relative to the clear-sky case.Still, in this case the DSSF algorithm reproduces the meteorological station R s dynamics with reasonable fidelity at the flat site.
As in the hourly evaluation case, the MBE is negative in almost all cases, meaning that on average the DSSF algorithm underestimates R s at the daily timescale, although the bias determined for both terrain classes for the averaged 2008-2010 period does not exceed −6 W m −2 .
Results presented here at the daily time step are consistent with those found in the literature.Bois et al. (2008) reported a RMSE of 2.16 MJ m −2 and a PE of 14 % using a Meteosat R s product obtained by means of the Heliosat-2 method in comparison with daily data from 19 meteorological stations in flat areas from 2000 to 2004.With the same method applied  to Meteosat data, Rigollier et al. (2004) found a PE between 9 % and 20 % in his dataset as well as for other works using the same method.Using GOES, Otkin et al. (2005)

Monthly evaluation
Table 5 and Fig. 4 show results from comparison of satellite retrievals and pyranometer data aggregated to monthly time steps.Comparisons at sites in flat terrain yield an averaged RMSE, MAE, PE and R 2 for the 2008-2010 period of 21 W m −2 and 17 W m −2 , 5.6 % and 0.99, respectively; while hilly sites yield an averaged RMSE, MAE, PE and R 2 of 32 W m −2 and 23 W m −2 , 9.3 % and 0.97, respectively.As with the hourly and daily results, better agreement was obtained at sites in flat terrain, although dependency of accuracy on terrain condition was not as marked at the monthly time step.In both cases, MBE is negative in all cases meaning that on average the DSSF algorithm underestimates R s at the monthly timescale, although the bias determined for both terrain classes for the averaged 2008-2010 period does not exceed −5 W m −2 .Using the DSSF product, Geiger et al. (2008a) found no clear seasonal bias dependence in the results.However, seasonal trends in MBE show that MBE is generally more positive during summer months, from June to September, and negative for the rest of the year in flat and hilly sites (see Fig. 5).Pinker et al. (2003) and Otkin et al. (2005) also found similar seasonal trends in MBE using GOES to model R s .According to Geiger et al. (2008a) and Ineichen et al. (2009) this bias may be related to the atmospheric transmission inputs such as the atmospheric turbidity that could be addressed by considering the temporal and spatial variability of the aerosol concentration in more detail.The removal of this bias would further decrease the RMSE and the MBE in model estimates for both terrain classes.Figure 6 shows an example of the monthly DSSF solar radiation (R s DSSF) and monthly meteorological station solar radiation (R s DSSF) cycle from 2008 to 2010 at two meteorological stations located in flat conditions (DP) and hilly conditions (WQ).In general, the DSSF product reasonably reproduces the seasonal variability measured at these two meteorological stations, located in both a flat and hilly landscape.
These monthly results are in good agreement with previous work, although there are few studies that have aggregated R s on a monthly basis.Rigollier et al. (2004) reported a PE from 5 % to 24 % based on their dataset as well as for other studies using the same retrieval method.Using Meteosat retrievals, Pereira et al. (1996) reported a PE of 13 % during a 2-yr period (1985)(1986) and using 22 meteorological stations.Finally, using data from GOES, GMS and MT-SAT from 1995 to 2008, Janjai et al. (2011) found a PE of 6.3 % with 5 meteorological stations in flat areas in Cambodia.According to Pinker et al. (2005), several attempts to compute R s with remote sensing data at a monthly time step and at a global scale yielded RMSE between 11.7 W m −2 and 31.5 W m −2 .
It is interesting to remark that similar results were found by Pons and Ninyerola (2008) using a hybrid model, applying DEM-based corrections to R s retrievals and comparing to 5 yr series of monthly data from meteorological stations.They found a PE ranging from 7.3 % to 13.1 % in four months, with a RMSE from 20 W m −2 to 24 W m −2 .If we take also into account that 77 % of the meteorological stations analyzed present a complete 3-yr monthly R s record from the DSSF, then this means that this product can be also useful for mapping R s from a climatic perspective (Cristóbal et al., 2008;Ninyerola et al., 2000).

Using DSSF as input data in ET modeling
While few analyses of ET model sensitivity to R s accuracy have been published, Diak et al. (2004)     for reasonable model performance.When retrieving net radiation, an essential variable for estimating ET, Kustas et al. (1994), found that daily GOES R s data with a RMSE of 23 W m −2 led to acceptable basin-scale estimates.In the work of Diak et al. (1998), R s derived from GOES was applied to routinely estimate daily crop ET for irrigation scheduling in Wisconsin, USA.Stewart et al. (1999) used GOES data to retrieve hourly and daily R s in a comparison among three evapotranspiration formulations applied over an agricultural area in northwest of Mexico.They found that hourly PE of 9.1 % and 16.8 % for clear sky and all averaged conditions, respectively, and daily PE of 4.7 % and 9.2 % for clear sky and all averaged conditions, respectively, produced reasonable estimates of ET.Garautza-Payan et al. ( 2001) and Garautza-Payan and Watts (2005) estimated crop water requirements of irrigated vegetation, cotton and wheat, in the same area also using GOES data as the R s input for ET modeling.The GOES derived R s data displayed a PE from 9 % to 14 % and from 7 % to 14 % for hourly and daily periods, respectively, allowing retrieval of daily cotton and wheat ET with a RMSE of 23 W m −2 and 7 W m −2 , respectively.In the work conducted by Jacobs et al. (2002) Su et al. (2005) reported good agreement between modeled ET using GOES R s data and in situ observations of instantaneous ET, with a RMSE of 60 W m −2 in 8 corn and soybean plots.In evaluating crop reference ET estimation at a daily time step in two study areas in southern France, Bois et al. (2008), found better results from methods using Meteosat R s compared to methods only using air temperature, ranging the relative annual RMSE from 22 % to 28 %, according to the method and the type of climate, humid-Oceanic or semi-arid Mediterranean.Using a calibrated R s product derived from GOES data with a PE of 10 % (1.7 MJ m −2 ) for daily reference and potential ET in Florida, USA, Paech et al. (2009) estimated an error from 5 % to 6 % in potential ET retrieval, generating a product useful for routine water management-related activities.
In general terms, evaluation results derived from this study show that the DSSF product has a relative error of about 10 % in comparison with pyranometer data at daily time steps, and according to the literature, can be used as a functional input to radiation-based ET models.However, in order to fully understand the sensitivity of modeled ET to satellite-derived Rs uncertainty, an error analysis should be conducted for each ET model.
It is important to remark that DSSF products may produce errors in models running at finer spatial scales (sub 3-km) over regions of hilly terrain unless topographic corrections are applied.This may be important for ET modeling applied to Landsat thermal imagery from 60 m to 100 m spatial resolution.Errors are significantly higher over regions of snow cover, and this could affect studies monitoring energy fluxes and snow melt in cold land regions.At daily time steps, the DSSF product performs within the 10 % error, except for the most difficult modeling scenario involving hilly terrain under cloudy skies.Satellite-based insolation retrievals can therefore be of significant utility in extrapolating instantaneous clear-sky ET retrievals to daily, monthly and seasonal estimates (Anderson et al., 2012).

Conclusions
Hourly, daily and monthly solar radiation estimates derived from the DSSF product produced by LSA SAF using MSG SEVIRI imagery were compared to pyranometer data in two terrain classes (flat and hilly) and for two atmospheric conditions (clear and cloudy sky), as well as for snow and ice cover.In general terms, hourly results compared favorably with the RMSE value from 20 % to 25 % reported previously for global evaluation studies of satellite-based R s retrievals.
Evaluation yielded good results in flat areas with an averaged model RMSE of 65 W m −2 (19 %), 34 W m −2 (9.7 %) and 21 W m −2 (5.6 %), and good R 2 of 0.95, 0.96 and 0.99, for hourly, daily and monthly-averaged solar radiation and including clear and cloudy sky conditions and snow or ice covers.Sites in hilly terrain also yielded reasonable R 2 of 0.91, 0.93 and 0.96 for hourly, daily and monthly time steps, and averaged model RMSE of 89 W m −2 (27 %), 48 W m −2 (14.5 %) and 32 W m −2 (9.3 %), respectfully.Comparisons at these sites could be improved by applying terrain-based corrections for topographic shadowing at sub-pixel levels.Hourly solar radiation overestimation in cloudy sky conditions and especially over snow and ice cover, could lead to high errors in energy fluxes monitoring in snow-melting related studies.Finally, according to the literature, the LSA SAF solar radiation product can be used as an operative input to calculate evaporative fluxes.

Figure 3 .
Figure 3. Examples of daily solar radiation cycle from dawn to dusk at two

Fig. 3 .
Fig. 3. Examples of daily solar radiation cycle from dawn to dusk at two meteorological stations located in flat conditions (V8) and hilly conditions (UI) during a clear sky day (6 April 2008) and cloudy sky day (24 September 2008).Time in UTC.
evaluation results of daily time steps, depending on terrain class and sky conditions, averaged by year and from 2008 to 2010.Scatter plot comparisons for 2008-2010 are shown in Fig. 2. Results at the daily interval show similar general behavior with the hourly results.DSSF data retrieved over flat sites in both clear and cloudy sky conditions show better agreement with observations than retrievals over hilly sites.At flat sites, clear sky conditions yield an averaged RMSE, MAE, PE and R 2 for the 2008-2010 period of 26 W m −2 and 21 W m −2 , 5.7 % and 0.98, respectively; and cloudy sky conditions an averaged RMSE, MAE, PE and R 2 for the 2008-2010 period of 36 W m −2 and 26 W m −2 , 11.9 % and 0.95, respectively.For hilly sites, clear sky conditions yield an averaged RMSE, MAE, PE and R 2 for the 2008-2010 period of 40 W m −2 and 28 W m −2 , 8.6 % and 0.95, respectively; and cloudy sky conditions an averaged RMSE, MAE, PE and R 2 for the 2008-2010 period of 50 W m −2 and 36 W m −2 , 16.8 % and 0.91, respectively. 1

Figure 5 .
Figure 5. Monthly MBE from 2008 to 2010 in flat and hilly terrain classes.

3 Material and study area 3.1 Meteorological data
and LSA SAF (2010).

Table 3 .
Hourly solar radiation error and accuracy statistics depending on flat or hilly terrain, clear and cloudy sky conditions, and presence of snow or ice cover from 2008 to 2010.RMSE, MBE and MAE in W m −2 , PE in percentage and n is the number of samples.

Table 4 .
Daily solar radiation error and accuracy statistics depending on flat or hilly terrain and clear sky or cloudy sky conditions from 2008 to 2010.RMSE, MBE and MAE in W m −2 , PE in percentage and n is the number of samples.
s measurements.When datasets are segmented based on both atmospheric conditions and terrain classes, clear sky conditions show better measurement agreement than cloudy conditions(2008)(2009)(2010)in both flat and hilly sites.Averaged over the period 2008-2010, for flat terrain differences between cloudy and clear sky conditions in terms of RMSE, MAE and PE are 44 W m −2 , 28 W m −2 and 26 %, respectively.In the case of hilly terrain these differences are 44 W m −2 , 32 W m −2 and 37 %, respectively.Finally, for snow and ice covers these differences between hilly and flat classes in terms of RMSE, MAE and PE are 21 W m −2 , 17 W m −2 and 4 %, respectively.It is interesting to note that in most of the cases MBE is negative, meaning that the DSSF algorithm underestimates R s measured at the pyranometer, although mean MBE values for the averaged 2008-2010 period and for all conditions do not exceed −6 W m −2 .

Table 5 .
Monthly solar radiation error and accuracy statistics for 2008 to 2010 by terrain class.RMSE, MBE and MAE in W m −2 , PE in percentage and n is the number of samples.
Jacobs et al. (2004) to estimate wetland potential ET in Florida, USA, during a growing season under non-water-limited conditions.R s evaluation with ground data showed a PE of 28.3 % and 9.9 % for 30-min and daily R s time steps, yielding an ET error at 30-min time steps of around 30 % and a R 2 of 0.67 but lower error of 3.1 % and higher R 2 of 0.90 in daily ET retrievals.When comparing four potential ET methods at a daily time step over a wetland area in Florida, USA,Jacobs et al. (2004)found dramatic improvements in the efficiency of ET-radiation based models using GOES R s , with a RMSE of 19.4 W m −2 .During the soil-moisture atmospheric coupling experiment (SMACEX) carried out in Iowa, USA, in 2002,