Here we assess established algorithms and a newly developed scheme for the estimation of downward long-wave radiation flux at the surface (DLR), i.e., the irradiance reaching the surface within 4 and 100 μm. These different methods correspond to bulk parameterization schemes, which merge the signature of clouds on Meteosat second-generation (MSG) data with information on atmosphere water content and near-surface air temperature available from numerical weather prediction (NWP) fields. The new formulation consists of a generalization of a method first developed for clear sky cases and now fine-tuned for a wider range of atmospheric conditions. The performance of this and three other parameterization schemes is compared with independent ground observations. Such a validation exercise is extended also to European Centre for Medium-Range Weather Forecast (ECMWF) flux forecasts, since the ECMWF model is the main source of information on air temperature and water vapor content, and to surface fluxes obtained from the Clouds and the Earth's Radiant Energy System (CERES). It is shown that the new parameterization scheme performs well when compared to other methods, with root mean square errors within 20 Wm−2. The overall good matching between parameterized values and in situ data suggests a good performance of a relatively simple bulk scheme and also of the use of MSG-based cloud identification.
 The downward long-wave radiation at the surface (DLR hereafter) is an important component of the heat exchange across the surface-atmosphere interface. Accurate values of DLR are essential in determining the surface radiation budget, which controls to a large extent the surface energy balance [e.g., Kalma et al., 2008; Jiménez et al., 2009]. Over large areas, DLR is generally inferred from atmospheric models [e.g., Wild et al., 1995; Morcrette, 2002] or from remotely sensed data [e.g., Gupta, 1989; Gupta et al., 1992; Zhang et al., 1995; Diak et al., 2000; Wang and Liang, 2009]. However, the modeling of surface radiation depends crucially on an accurate description of clouds. Known deficiencies in the spatial and temporal representation of clouds derived from models [e.g., Crewell et al., 2002] may be overcome by satellite observations [Meetschen et al., 2004]. Over the past two decades, there has been a significant increase in the use of satellite data to identify clouds [e.g., Rossow and Garder, 1993; Feijt et al., 2000; Derrien and Le Gléau, 2005], as well as to retrieve top cloud properties [e.g., Gupta, 1989; Derrien and Le Gléau, 2005]. Although cloud microphysics plays an important role in radiative processes within the cloud, both solar and thermal fluxes reaching the surface are largely determined by local cloud cover [e.g., Dürr and Philipona, 2004; Meetschen et al., 2004]. Here we demonstrate how a relatively simple formulation relying on a limited set of atmospheric variables (essentially near-surface temperature and column water vapor), along with a reliable information on the presence of clouds (differentiating between clear sky, partially cloudy and overcast), is able to reproduce reasonably well DLR in situ observations. The methodology combines the signature of clouds on infrared and visible channels with information on atmosphere water vapor content and near-surface temperature from numerical weather prediction (NWP) models. This is based on the assumption that the latter, which also includes the assimilation of measurements from sounding instruments along with other remotely sensed and conventional observations, provides the best knowledge of atmospheric profiles at any given point.
 Below we describe a bulk parameterization scheme to be applicable under all-sky conditions. The adjustment of semi-empirical formulations to estimate DLR from near-surface data has been the subject of many studies before. Most of these are applicable to clear-sky cases only, e.g., Brunt , Idso and Jackson , Brutsaert , and, more recently, Prata  and Dilley and O'Brien . The restriction of these studies to clear conditions strongly limits their utility; however, there have been fewer attempts to derive all-sky parameterizations, e.g., Crawford and Duchon , Diak et al. , Josey et al. , and Bilbao and de Miguel . The formulation proposed here follows that first developed by Prata  for clear-sky cases and now fine-tuned for a wider range of atmospheric conditions. The calibration of this semi-empirical method is based on downward infrared flux simulations obtained with the moderate spectral resolution atmospheric transmittance algorithm [MODTRAN4; Berk et al., 2000]. The algorithm described here is the current baseline for the DLR product, generated operationally by the Satellite Application Facility on Land Surface Analysis [LSA SAF; Trigo et al., 2010]. It makes use of the cloud mask developed by the SAF on support to Nowcasting and Very Short-Range Forecasting for [NWC SAF; http://nwcsaf.inm.es/; Derrien and Le Gléau, 2005] to take advantage of the spectral characteristics of the Spinning Enhanced Visible and Infrared Imager (SEVIRI) onboard Meteosat. Information on precipitable water and near-surface air temperature is obtained from forecasts of the European Centre for Medium-Range Weather Forecasts (ECMWF).
 The LSA SAF [Trigo et al., 2010] is currently retrieving and distributing, in near real-time or off-line, a set of radiation budget parameters over land surfaces, all obtained from SEVIRI/Meteosat data, namely, albedo, land surface temperature, and downward long- and short-wave fluxes. These form a consistent data set derived from the same source data and, in the case of the fluxes, the same cloud information.
 The validation of the algorithm presented here for downward long-wave fluxes is performed using independent data and through the comparison of DLR estimates with in situ measurements. The validation exercise is extended to (1) other bulk parameterizations for benchmarking; (2) ECMWF flux forecasts, since the model is the main source of information on air temperature and water vapor content; and also (3) downward long-wave surface fluxes obtained from the Clouds and the Earth's Radiant Energy System [CERES; Wielicki et al., 1996; Gupta et al., 1992].
2. Downward Long-Wave Radiation at the Surface
2.1. Parameterization Schemes
 Downward long-wave radiation (DLR) is defined as the total irradiance within the infrared part of the spectrum (4–100 μm). Within this range, atmospheric scattering may be neglected and DLR corresponds essentially to radiation emitted by the lowest hundred meters of the atmosphere [Zhao et al., 1994]. DLR is often estimated as a bulk parameterization, where the thermal radiative flux reaching the surface is emitted by an atmospheric layer with emissivity ɛsky and temperature Tsky:
where σ is the Stefan-Boltzmann constant. Among the greenhouse gases, water vapor is the most important contributor to DLR and also the most variable [e.g., Niemelä et al., 2004]. Thus both ɛsky and Tsky are generally estimated as a function of near-surface atmospheric temperature and/or water vapor content. Many of developed bulk parameterizations are applicable under clear-sky conditions only. The presence of clouds contributes to a significant increase in DLR, since they “close” the infrared atmospheric window [e.g., Takara and Ellingson, 2000]. To take such an effect into account, all-sky bulk parameterizations generally introduce correcting factors as a function of estimated cloudiness and, in some cases, of cloud base height [e.g., Gupta et al., 1992]. Several formulations estimate cloud cover from the fraction between actually observed solar irradiance and that which would be observed under identical conditions but with clear sky [e.g., Crawford and Duchon, 1999; Bilbao and de Miguel, 2007].
 A number of studies compared the performance of different methodologies [e.g., Prata, 1996; Crawford and Duchon, 1999; Niemelä et al., 2004; Bilbao and de Miguel, 2007], but the results have often been inconclusive and highly dependent on the season and geographical distribution of the data used for model verification. Table 1 summarizes the bulk parameterizations analyzed here, comprising two relatively recent schemes valid for clear-sky conditions and one applicable for all situations.
Table 1. Formulations for Effective Sky Emissivity and Temperature, Used by Different Bulk Parameterization Schemesa
To is the screen temperature (K), ΔTo is the dew point depression (K), w is the total column water vapor (mm), and n is the fraction of cloud cover.
Prata  developed a clear-sky emissivity model assuming a continuum absorption correction. Sky effective emissivity is a function of total column water vapor (see Table 1), while the sky effective temperature is the screen temperature; model parameters were calibrated using in situ measurements. Here we develop a similar formulation for DLR where ɛsky and Tsky in equation (1) are given by:
ɛsky is again a function of total column water vapor (TCWV), w (mm), and Tsky equals 2m temperature, To (K), corrected by observed dew point depression at 2m (ΔTdo). The parameters in equations (2) and (3), α, β, m, γ, and δ are fitted for clear-sky and overcast conditions separately. We assume that for remote sensing retrievals, DLR at the pixel scale results from the contribution of clear, Fclear↓, and cloudy, Fcloudy↓, portions of atmosphere:
where n is the cloud fraction obtained from visible and infrared imagers.
 As mentioned before, and particularly under clear-sky conditions, DLR arises almost entirely from the lowest several hundred meters of the atmosphere [e.g., Schmetz, 1989; Zhao et al., 1994; Ellingson, 1995; Prata, 1996; Diak et al., 2000]. This layer can be adequately characterized for the purpose of DLR estimations by near-surface air temperature and humidity, which explains the generally good performance of clear-sky formulations based on 2m air temperature and humidity or TCWV [Prata, 1996; Dilley and O'Brien, 1998; Diak et al., 2000]. In the presence of clouds, cloud base emission contributes to the enhancement of downward long-wave flux at the surface when compared to that of clear-sky conditions by filling the atmospheric window part of the spectrum [e.g., Ellingson, 1995; Takara and Ellingson, 2000; Niemelä et al., 2004]. Both effective sky emissivity and temperature depend on parameters such as cloud base height (via cloud base temperature) and liquid water path [Gupta et al., 1992; Niemelä et al., 2004; Zhou and Cess, 2001; Hollmann and Gratzi, 2002; Zhou et al., 2007]. However, such parameters are difficult to obtain from an imager with the characteristics of SEVIRI [e.g., Hollmann and Gratzi, 2002]. Moreover, satellite retrievals of cloud characteristics such as cloud phase, cloud base, or even parameters more closely related to the satellite observations, such as cloud top height, present large uncertainties, particularly for multilayers or broken clouds [Thoss, 2009; Marchand et al., 2010], which then propagate to higher-level retrievals such as surface flux estimations. The characteristics of the sensor are therefore crucial for the choice of cloud parameters to be used in DLR estimates. Under cloudy conditions, the parameterization of sky emissivity in equation (2) assumes that liquid water path correlates with TCWV [e.g., Hack, 1998]. The empirical parameters in equation (2) are adjusted to different sets of clear and cloudy conditions, as described in the next section, to mimic the respective impact of TCWV (or TCWV and liquid water path) on sky emissivity. On the other hand, the introduction of a correction to the screen temperature as a function of dew point depression in equation (3) compensates, to some extent, for the difference between cloud base temperature and surface temperature [Josey et al., 2003].
2.2. Calibration of the New DLR Formulation
 The calibration of parameters in equations (2) and (3) relies on radiative transfer simulations of downward long-wave fluxes at the surface. These simulations are obtained from the moderate spectral resolution atmospheric transmittance algorithm (MODTRAN4; Berk et al. ) applied to a database of 13,495 atmospheric profiles described in Chevallier et al. . Downward surface radiances at five different zenith angles are computed for wave numbers ranging between 100 and 2500 cm−1 at a resolution of 1 cm−1. Such values are then integrated to provide total downward fluxes within the long-wave domain of the spectrum. The model is configured to use ozone and trace gases from a climatological data set and the “Rural” aerosol profile, all available from MODTRAN. Temperature and humidity values, as well as cloud liquid and ice water content and cloud cover, are taken from the atmospheric profiles database, available at 60 vertical levels between the surface and 10 Pa. These profiles are sampled from ECMWF 40-year reanalyses to be globally distributed and representative of a wide range of atmospheric conditions [Chevallier et al., 2000].
 The calibration of parameters in equations (2) and (3) was performed separately for clear-sky cases (total cloud cover, TCC = 0) and overcast conditions (TCC >0.9) following a piecewise regression approach. The calibration domain is split into overlapping temperature and humidity classes. Parameters α and β were adjusted by least squares fitting of equation (2), with m = 0.5 and m = 1 for clear and overcast conditions, respectively, to MODTRAN flux simulations. The m values are those that best adjust equation (2) to sky emissivity simulated by MODTRAN as a function of total column water vapor (TCWV). The following ranges of 2m temperature (T2m) and TCWV were considered: (1) “dry and cold” with TCWV ≤ 10 mm and T2m < 270 K; (2) “dry and warm” with TCWV ≤ 10 mm and T2m ≥ 270 K; and (3) “moist” with TCWV > 8 mm. The next step consisted of fitting coefficients γ and δ in equation (3) to an optimal Tsky obtained by introducing ɛsky from equation (2) into equation (1) for the same ranges of TCWV and T2m. The resulting parameters are described in Table 2. It is worth noting that, as expected, most overcast cases in the calibration database (∼83%) fall into the “moist” category, while about 15% correspond to high-latitude “dry and cold” atmospheres. Since only about 2% of overcast atmospheric profiles fall into the “dry and warm” class and many of these overlap with the “moist” category, we opted to ignore that class and to consider the “moist” parameters when applying the scheme. Moreover, when estimating DLR for TCWV falling into two classes, the resulting value is a linear combination of F↓ obtained with both sets of valid coefficients to avoid discontinuity among classes.
Table 2. Parameters for the New Flux Parameterization Scheme as Defined by Equations (1)–(3)a
Clear Sky (m = 0.5)
Overcast (m = 1)
Parameters calibrated for the following ranges of temperature and water vapor: dry cold when TCWV ≤ 10 mm and T2m < 270 K; dry warm when TCWV ≤ 10 mm and T2m ≥ 270 K; and moist when TCWV > 8 mm.
Figure 1 presents scatterplots of different parameterized DLR values versus MODTRAN simulations. When compared with the remaining schemes described in Table 1, the new parameterization (equations (1)–(3)) is able to reproduce MODTRAN DLR, with negligible bias and lower root mean square differences (RMSD). Most formulations analyzed here for DLR exhibit conditional biases, i.e., they generally overestimate the lower DLR values (e.g., Prata , Dilley and O'Brien , Josey et al.  schemes in Figure 1) and/or underestimate those within the higher ranges (Dilley and O'Brien , Josey et al.  schemes in Figure 1). Parameterizations of downward long-wave fluxes are often strongly tight to the calibration data. This is clearly the case for the Josey et al.  (Table 1) scheme, which relied on data collected during oceanographic campaigns. Although these observations were taken over a wide latitudinal range, from the subtropics to the Arctic [Josey et al., 2003], the formulation is unable to reproduce the extremes of DLR distribution. These are likely to correspond either to dry and extremely warm or very dry and cold conditions only likely to be observed inland.
 The comparison of different parameterization schemes with MODTRAN simulations does not constitute a proper validation of the former. Moreover, the new version of Prata's  formulation was adjusted to the MODTRAN estimations, and thus its better performance when compared with the remaining methodologies was expected. In the next section, all schemes are tested against a set of independent data obtained from in situ observations.
3. Comparison Against In Situ Observations
 DLR is then estimated using the formulations detailed in Table 1, as well as the new parameterization scheme corresponding to equations (1)–(4), with a 30-minute frequency and for the SEVIRI/Meteosat disk. Atmospheric parameters TCWV, To, and ΔTdo are provided by ECMWF 3-hourly forecasts (steps between 12 and 36 h), linearly interpolated in time and bilinearly interpolated to the projection of SEVIRI level 1.5 data (sampling distance of 3 km at nadir). To is corrected for the difference between model orography and the local pixel height, using a constant lapse rate of 0.67K/100m, while ΔTdo is kept unchanged. This is a first approach that needs to be readdressed, particularly when temperature inversions are likely to be present. The cloud fraction is obtained from the cloud mask developed by Derrien and Le Gléau  for SEVIRI/Meteosat data within the context of the NWC SAF. The algorithm is based on a multispectral threshold technique designed to optimize the use of the rich spectral characteristics of SEVIRI [Derrien and Le Gléau, 2005]. The Prata  and Dilley and O'Brien  schemes (Table 1) are tested for clear-sky conditions only, according to the classification provided by the cloud mask. For pixels classified as “totally cloudy” by the NWC SAF cloud mask, n in equation (4) and in the Josey et al.  scheme is assigned to 1, while for “partially cloudy” pixels, n is assigned to 0.5. The time series of DLR based on the blending of ECMWF forecasts and satellite clouds, used as input for the various formulations described above, have been available since 2005.
Figure 2 shows an example of DLR obtained with the new parameterization scheme for 25 May 2009 at 1200 UTC. Such fields are generated on an operational basis by the LSA SAF and are available in near real time (with a maximum difference of 1 hour between observation time and distribution to users) or off-line, at the original spatial and temporal resolution (further information is available at http://landsaf.meteo.pt/). Together with the remaining surface radiation products derived from SEVIRI/Meteosat by the LSA SAF, DLR targets essentially applications in land surface or hydrological modeling [see, e.g., Boone et al., 2009].
 The estimated values of DLR are compared with the in situ observations detailed in Table 3. Most of these are provided by the Baseline Surface Radiation Network [BSRN; Ohmura et al., 1998], while data at Roissy and Carpentras were obtained directly from Météo-France. The choice of ground observations was constrained by the area coverage of the Meteosat disk and the beginning of the DLR time series (January 2005). Since most BSRN ground stations are located in Europe, the set is complemented with data collected during 2006 at Niamey (Niger) within the RADAGAST experiment [Slingo et al., 2009].
Table 3. Stations With In Situ Measurements of Downward Long-Wave Radiation Used in This Study
Latitude and Longitude
May 2005 – Dez 2006
May 2005 – Dec 2007
May 2005 – Oct 2006
Jan 2005 – Dec 2005
May 2005 – Aug 2007
Jan 2005 – Dec 2007
Jan 2005 – Nov 2007
Jul 2005 – Dec 2007
Jul 2005 – Dec 2007
Jan 2006 – Dec 2006
 The next section discusses the comparison among DLR estimations obtained from different parameterization schemes, DLR forecasts provided by ECMWF, and in situ observations. The data under analysis correspond to averages over 3-hourly intervals (0–3, 3–6, …, 21–24 UTC) to match ECMWF flux data. To correct for the difference in height between model and observations, the ECMWF forecasts of downward long-wave flux are further adjusted using a height gradient of 2.8 Wm−2/100 m [Wild et al., 1995].
 The performance of the new parameterization scheme is also assessed with respect to downward long-wave surface flux obtained from Clouds and the Earth's Radiant Energy System [CERES; Wielicki et al., 1996] data flying onboard Terra. For this purpose, we use downward long-wave surface fluxes available from the Single Scanner Footprint TOA/Surface Fluxes and Clouds (SSF; Editions 2B and 2F) product from CERES. The SSF-CERES product contains instantaneous (all-sky) surface long-wave radiation estimated using the parameterization described in Gupta et al. . In this model, downward long-wave fluxes are a function of the atmosphere “effective emitting temperature” (a weighted average of skin and lower troposphere temperatures), TCWV, the cloud fraction within the pixel, and the cloud base height. Here we assess the simultaneous comparison of the new parameterization scheme (with cloud information from SEVIRI) and SSF-CERES estimations with observations at the ground stations listed in Table 3. We use instantaneous fluxes, allowing for a maximum time difference to in situ observations of 10 minutes, and a maximum distance between the pixel center and the station of 10 km. It should be noted that a reduction to 5 minutes or less in the time mismatch between ground observations and parameterized fluxes does not change the results significantly, but greatly reduces the sample size. This analysis is performed for SSF-CERES retrievals for the morning and nighttime overpasses of Terra satellite (around 10 h and 22 h local time, respectively) for data collected in the period between January 2006 and April 2007.
4. Results and Discussion
Figures 3–5 present scatterplots of modeled DLR values using different formulations and the ECMWF model for three stations characteristic of middle latitudes, high latitudes, and arid regions, respectively. These diagrams generally confirm the results obtained with the comparison between parameterization schemes and MODTRAN simulations. Overall, the validation against in situ measurements indicates that the modified version of the algorithm initially proposed by Prata  performs better than the remaining formulations for both clear and cloudy conditions. The deviations to the 1:1 line in the scatterplots of Figures 3–5 are generally smaller and suggest lower conditional biases, particularly when compared with Josey et al. .
 The plots in Figures 3–5 show that dispersion around the 1:1 line is smaller for clear-sky cases than for cloudy conditions when the thermal radiation reaching the surface depends on factors such as cloud base height and cloud microphysics. However, the misclassification of cloudy scenes as clear sky may lead to large underestimation of the observations. Toravere presents a set of such points clearly lying below the 1:1 line in Figures 4b and 4c (see arrows), most of which were obtained during the winter months, when low solar zenith angles combined with high viewing angles make the pixel classification more difficult.
 The average differences (bias) and RMSE between DLR estimations and in situ observations are shown in Figures 6 and 7, for DJF, MAM, JJA, and SON; the standard deviation of the observations taken at each station is also displayed in Figures 6b and 7b. Overall, the new parameterization scheme presents systematic errors of the order of 10 Wm−2 or lower, with the exception of the cases discussed below. The new scheme also tends to perform better than ECMWF simulations, suggesting that despite its simplicity, it partially corrects for deficiencies in ECMWF cloud modeling [e.g., Crewel et al., 2002; Meetschen et al., 2004], benefiting as well from the finer spatial representation of the remote sensing cloud mask. The new cloudy conditions scheme outperforms the Josey et al.  simulations for most stations. For clear-sky conditions, the scores obtained by the original and modified version of Prata's  algorithm are fairly similar.
 Results obtained using the scheme developed by Josey et al.  are the most variable: DLR estimations present low systematic errors (10 Wm−2 or less) for stations located in the middle to high latitudes during winter, but have large negative biases during the warm season in Europe and all year round in the most southern stations (Sde Boquer in Israel, Tamanrasset in Algeria, and Niamey in Niger). As mentioned before, the calibration data set used by Josey et al. , consisting of observations collected over the Atlantic Ocean, limits its applicability to inland regions, particularly under extreme cold/warm and dry conditions. Accordingly, for coastal stations such as Lerwick or Camborne, the flux estimations by that scheme present relatively stable scores throughout the year, with results comparable (or better in some cases) to those of the remaining models (Figures 6 and 7).
 The new parameterization scheme exhibits rather poor performances for Toravere during the winter months, where it underestimates local observations by over 20 Wm−2 in both clear and cloudy sky conditions. In that station, during winter, the new scheme is outperformed by both ECMWF flux forecasts and the Josey et al.  estimations. The clear sky results may be explained by an underclassification of cloudy scenes. A deeper analysis of those cases reveals that 25% of winter clear-sky fluxes, obtained using the new parameterization scheme, are underestimated by 40–90 Wm−2 (cluster of points indicated by the arrows in Figures 4b and 4c). These are likely cloudy pixels misclassified as clear sky. Accordingly, ECMWF TCC averaged over those 25% cases is about 94%, significantly larger than the 44% for the remaining points. In high latitudes, winter carries a higher uncertainty in scene classification due to the combination of large SEVIRI viewing angles, low solar illumination, and frequently snow covered surfaces. When those values are ignored, the winter bias and RMSE for the modified Prata  scheme are reduced to −5.5 and 10 Wm−2, respectively. The reasons behind the large negative bias obtained for cloudy cases in the same station and season are harder to identify. It is possible that the radiation emitted by low clouds occurring frequently in high-latitude winter [e.g., Keevallik and Russak, 2001] is not well reproduced by the current parameterization schemes, e.g., because those situations are not sufficiently represented in the calibration database. On top of that, temperature inversions are frequent in high latitudes during winter. Under such conditions, the use of 2m temperature would contribute to the underestimation of surface fluxes for all parameterization schemes under analysis [e.g., Niemelä et al., 2004].
Figure 8 presents monthly means of the flux diurnal cycle, obtained by averaging 3-hourly observations/ flux estimations. Toravere (Figure 8a) shows a relatively weak diurnal cycle during winter months. By then, the underestimation of the new parameterization does not present any significant daily dependency. During spring and summer, the new scheme follows well the observed diurnal cycle, which is underestimated by ECMWF. For Carpentras, a station located in central Europe, flux estimates obtained with the new formulation reveal very good agreement with the in situ diurnal cycle (Figure 8b). In the same station, the overall underestimation of ECMWF fluxes tends to be smaller during the morning for most of the year. A similar pattern is also observed for Tamaransset (Figure 8c), where ECMWF generally underestimates the observations, particularly during the peak of the day and afternoon during the spring and summer months. The estimations with the new scheme are able to follow the observed diurnal cycle more accurately, although underestimation also occurs toward the end of the day, particularly between May and August.
 Niamey, particularly during DJF and MAM, is another critical site, where the modified Prata  formulation underestimates local observations of clear and cloudy fluxes (Figures 6 and 7). The Niamey region was characterized by relatively high aerosol loads during those two periods and also suffered severe dust storms in March 2006 [Slingo et al., 2006]. The diurnal cycle of in situ observations during that period (Figure 8d) also seems to be smoother when compared to flux estimations, leading to higher underestimation of both the new formulation and ECMWF values during nighttime. Under such high aerosol conditions, there are several factors that may contribute to the poorer performance of remote sensing estimates, namely, (1) calibration with a set of simulations using a single type of aerosol profile (rural aerosol model available in MODTRAN; section 2.2); and (2) a possible degradation of input variables, including a reduced accuracy of the cloud mask and of ECMWF forecasts. Accordingly, ECMWF presents large negative biases of flux values in DJF and MAM, for both clear and cloudy cases. The Niamey values obtained with Prata , available for clear sky only (Figure 6), present slightly better results and seem to be less affected by the conditions in DJF and MAM. Despite such deficiencies, it is worth noting that, for cloudy conditions, the new scheme outperforms Josey et al.  and ECMWF forecast results (Figure 7). However, further work is needed to fully understand the variability of downward long-wave flux measurements at that site.
 The results obtained by the new parameterization scheme, where cloud information from SEVIRI/Meteosat is combined with 2m temperature and total column water vapor fields from ECMWF, are also compared with downward long-wave fluxes available from SSF-CERES product. For that purpose, both instantaneous fluxes are compared to the same in situ observations (stations in Table 3). The results are summarized in Table 4 and Figure 9, aggregated for Northern Europe (Cambourne, Lerwick, Toravere), Central Europe (Palaiseau, Payerne, Carpentras), and Semi-Arid/Desert (Tamanrasset, Sde Boqer, Niamey) sites. The clear-sky statistics in Table 4 are obtained for a different subset of SEVIRI- and CERES-based fluxes, according to the cloud mask provided by each data set.
Table 4. Aggregated Bias (Wm−2) and Root Mean Square Error (RMSE; Wm−2) of Instantaneous Flux Estimations Obtained From the SSF-CERES Product and From the New Parameterization Scheme (Based on SEVIRI Cloud Information), for Clear and All Sky, as Classified in the CERES and SEVIRI Products, Respectively
Northern Europe: Cambourne, Lerwick, Toravere; Central Europe: Palaiseau, Payerne, Carpentras; and Semi-Arid/Desert: Tamanrasset, Sde Boqer, Niamey.
New Param (SEVIRI)
 SSF-CERES data outperform the new parameterization scheme for Northern Europe sites, where the latter shows a set of clear sky/cloudy events that underestimate/overestimate the observations (Figure 9b and Table 4). These results, with RMSE of the order of 22–25 Wm−2, are in agreement with those presented separately for each station in Figures 6 and 7, which also indicate relatively large RMSE for Lerwick, Camborne, and Toravere. SEVIRI view angles for these stations, all north of 50°N, are relatively high, which introduces further uncertainty in cloud identification. It is worth noting that the conditions for CERES pixels to be considered “clear sky” are fairly restrictive, demanding a clear area fraction at subpixel resolution of 99.9% or more, and therefore leading to a relatively low number of clear-sky cases.
 For Central Europe, the statistics obtained by SSF-CERES and the new parameterization values are very similar, for both clear- and all-sky conditions (Table 4, Figures 9c and 9d). For semi-arid/desert stations, representative of a large portion of land pixels within the Meteosat disk, the new formulation presents smaller biases and lower RMSE when compared to SSF-CERES fluxes (Table 4). The Data Quality Summary of Surface Fluxes published by the CERES team (http://eosweb.larc.nasa.gov/PRODOCS/ceres/SSF/Quality_Summaries/ssf_surface_flux_terra_ed2B.html) also presents larger RMSE for desert sites, but with values around 22 W−2 (20 Wm−2) for all-sky (clear-sky) conditions. The sample analyzed here is smaller than that used by the CERES team; nevertheless, Figure 9e seems to indicate that SSF-CERES fluxes overestimate the higher values of in situ observations (obtained at Tamanrasset) and therefore lead to positive biases and relatively large RMSE (nearly 30 Wm−2 for all-sky conditions).
5. Concluding Remarks
 To determine an adequate formulation to estimate downward long-wave radiation (DLR) from geostationary satellite data, we performed an assessment of different bulk parameterizations, including (1) two schemes applicable to clear-sky conditions developed by Prata  and Dilley and O'Brien , respectively; (2) the scheme first proposed by Josey et al. , applicable to all sky conditions; and (3) a generalized version of the formulation first proposed by Prata , valid for all-sky conditions. The latter was calibrated using data simulated MODTRAN for a database of over 13,400 globally distributed atmospheric profiles [Chevallier et al., 2000].
 The performance of the above-mentioned methods is verified against in situ data collected in several stations in Europe, one in the Middle East, and two in Africa (Table 3). It is shown that the scheme developed by Prata  and its modified version presented here are able to reproduce well the observations for clear-sky cases, often obtaining biases lower than ECMWF modeled fluxes. This is clear from the statistics obtained considering all available stations together (“ALL” in Figure 6), which show that both Prata  and the modified version outperform ECMWF, and particularly Josey et al. , for all seasons. The latter is heavily penalized by the results obtained for semi-arid/desert stations.
 DLR estimated with Prata  and the modified formulation exhibit larger departures from observations for Toravere during the winter months (Figure 6a). These are likely to be associated to the misclassification of (partially) cloudy pixels as cloud-free. Such errors tend to occur for regions with high view angles (such as Toravere), being more frequent during nighttime when visible channels are not available, or for low solar zenith angles, such as in winter, early morning, or close to sunset [Derrien and Le Gléau, 2005]. The results obtained for Niamey during the first 6 months of 2006 are also worth noting. Particularly during February and March, the area was characterized by large aerosol loads and a few dust storm events [Slingo et al., 2006, 2009], leading to the underestimation of DLR by all models analyzed here. The contribution of aerosol to DLR, often neglected, clearly needs to be addressed, particularly in regions under high loads. However, the degradation of ECMWF forecasts and/or of the accuracy of the cloud mask in such conditions may further contribute to a poorer performance of the parameterization developed here. Josey et al.  present generally poorer results for clear-sky conditions, particularly under warm conditions. As mentioned before, the performance of this methodology is limited, both under clear and cloudy sky conditions, by the observations used for training obtained in maritime environments.
 The comparison between modeled and observed DLR for cloudy pixels reveals, as expected, higher error dispersion than under clear sky. However, the summary of statistics obtained for all in situ stations is less uneven among different estimations (Figure 7) than that obtained for clear-sky cases. The fewer number of cloudy cases in stations such as Tamanrasset, when compared to Central and Northern Europe sites, prevents the degradation of Josey et al.  results. Overall, the new parameterization scheme presents better statistics for most seasons.
 Cloud characteristics such as cloud base or cloud microphysics play a role in the thermal flux that reaches the surface and are difficult to model or determine from remote sensing data. Nevertheless, the new parameterization scheme presents RMSE of the order of 20 Wm−2 for most stations under cloudy conditions and, despite its simplicity, performs better than ECMWF in most cases analyzed here. Also, when compared to SSF-CERES downward long-wave product, the new parameterization scheme presents similar results for sites in Central Europe and improved ones in semi-arid/desert regions; only for Northern Europe stations do the SSF-CERES fluxes present better statistics. Next, we will test the introduction of new cloud parameters from SEVIRI/MSG (most notably cloud classification and a finer estimation of cloud fraction, n), taking into account their uncertainty versus the gain in the accuracy of DLR formulation. The results presented here on the validation of DLR estimates (using the new scheme) suggest a good performance of the cloud identification method applied to SEVIRI and confirm the key role of correct scene classification for the estimation of long-wave fluxes at the surface.
 This work was carried out within the context of the LSA SAF project (http://landsaf.meteo.pt), funded by EUMETSAT. Most ground measurements were obtained from BSRN (http://www.bsrn.awi.de/). The in situ data for Roissy and Carpentras were kindly provided by Jean-Louis Roujean (CNRM, Météo-France). SSF-CERES data were obtained from the Atmospheric Science Data Center at the NASA Langley Research Center.