Improving evapotranspiration in a land surface model using biophysical variables derived from MSG/SEVIRI satellite

Monitoring evapotranspiration over land is highly dependent on the surface state and vegetation dynamics. Data from spaceborn platforms are desirable to complement estimations from land surface models. The success of daily evapotranspiration monitoring at continental scale relies on the availability, quality and continuity of such data. The biophysical variables derived from SEVIRI on board the geostationary satellite Meteosat Second Generation (MSG) and distributed by the Satellite Application Facility on Land surface Analysis (LSA-SAF) are particularly interesting for such applications, as they aimed at providing continuous and consistent daily time series in near-real time over Africa, Europe and South America. In this paper, we compare them to monthly vegetation parameters from a database commonly used in numerical weather predictions (ECOCLIMAP-I), showing the benefits of the new daily products in detecting the spatial and temporal (seasonal and inter-annual) variability of the vegetation, especially relevant over Africa. We propose a method to handle Leaf Area Index (LAI) and Fractional Vegetation Cover (FVC) products for evapotranspiration monitoring with a land surface model at 3–5 km spatial resolution. The method is conceived to be applicable for near-real time processes at continental scale and relies on the use of a land cover map. We assess the impact of using LSASAF biophysical variables compared to ECOCLIMAP-I on evapotranspiration estimated by the land surface model HTESSEL. Comparison with in-situ observations in Europe and Africa shows an improved estimation of the evapotranspiration, especially in semi-arid climates. Finally, the impact on the land surface modelled evapotranspiration is compared over a north–south transect with a large gradient of vegetation and climate in Western Africa using LSA-SAF radiation forcing derived from remote sensing. Differences are highlighted. An evaluation against remote sensing derived land surface temperature shows an improvement of the evapotranspiration simulations.


Introduction
In the past decades, an increasing number of models were developed to monitor evapotranspiration (ET) at different scales using remote sensing measurements.Simple empirical or statistical methods to fully detailed physical models have been developed, using a wealth of information provided by various satellites (e.g.Courault et al., 2005;Kalma et al., 2008;Li et al., 2009).In all methods, vegetation has been recognized to be a cornerstone of the evapotranspiration process since plants are the main medium for exchange of water between the soil and the atmosphere.In particular, land surface models are widely used for meteorological and climate studies.They are based on a conceptual and semi-empirical description of, respectively, the physical and physiological processes of heat and water exchanges between soil, plants and atmosphere media.Information usually needed for land surface models is 1) the exact coverage of plant functional types (PFT) (Bonan et al., 2002), information given by a land cover map, 2) the state of the vegetation development, N. Ghilain et al.: Benefits of LSA-SAF biophysical variables for evapotranspiration modelling mostly given by the variable Leaf Area Index (LAI).While most of those models require explicit external information on the vegetation status (e.g.Balsamo et al., 2009;Noilhan and Planton, 1989), a new generation of models considers it as a new model variable (e.g.Jarlan et al., 2008;Gibelin et al., 2006;Blyth et al., 2006).But, for both model classes, information on vegetation status is highly relevant.For the first class, it is necessary information to provide.For the second one, it is recommended for updating the model forecasts (e.g.Jarlan et al., 2008;Albergel et al., 2010), or at least evaluate their output (e.g.Brut et al., 2009).Past studies have used variety of land surface models or crop growth models and different vegetation products issued from remote sensing (e.g.Dorigo et al., 2006).In this way, the use of remote sensing vegetation indices has revealed an improved forecast of surface variables.
Different exploitations have become possible along with the ever growing length of time series available from remote sensing.Early usage of the vegetation indices in soilvegetation-atmosphere transfer (SVAT) models focused on a mean vegetation status by plant type throughout a year.New available long time series and reprocessing capabilities have given rise to vegetation databases, that include a mean monthly or weekly evolution of vegetation parameters, PFT and geographically dependent (e.g.ECOCLIMAP).Those two approaches have been widely used for operational purposes (e.g.van den Hurk et al., 2000van den Hurk et al., , 2003) ) for practical reasons: easy to handle, not dependent on necessary incoming information from an external source, and hence on timeliness.However, the direct use of remote sensing time series is an important step towards a closer monitoring of the land surface, because it captures better both spatial and temporal variations including intra-species and inter-annual variability.For long-term monitoring of ET, it is necessary to capture the inter-annual variability of the vegetation state signal, which can be caused, for example, by a different date of budbreak for deciduous forests, harvest timing shifts (Cooley et al., 2005), rain occurence in semi-arid areas where the vegetation growth is driven by the hydrological regime, or by land re-allocation or fires.For the short term monitoring, at a daily time scale, poor time sampling of the vegetation phenology can have a negative impact on ET, especially during the rapid development stage of the canopy (e.g.Sepulcre-Cantó et al., 2012).Moreover, even if the classification in ecosystems of ECOCLIMAP-I is quite fine, the assumption of equal properties across different spatial scales for one ecosystem is an oversimplification of the problem.
Previous efforts have been carried out mainly using time series from polar orbiters (e.g.Albergel et al., 2010;Rodell et al., 2004).However, most of those studies focus on assimilation in a reanalysis mode, and cannot be used for a near-real time monitoring of the evapotranspiration.Recently, the EUMETSAT Land Surface Analysis Satellite Application Facility (LSA-SAF) developed a coordinated service in the area of remote sensing of the land surface and proposes products based on the geosynchronous MSG satellites (http://landsaf.meteo.pt)(Trigo et al., 2011).Among these products, remotely sensed biophysical parameter products, i.e. leaf area index, LSA-SAF LAI, and fractional vegetation cover (Fveg), LSA-SAF FVC, are delivered daily at the spatial resolution of the SEVIRI instrument on board MSG (LSA-SAF PUM VEGA, 2008) 3.1 km sub-satellite, and have been available since 2007.This study is done in the framework of land surface modelling for the purpose of ET monitoring at MSG/SEVIRI spatial resolution.It focuses on the applicability and gain of using biophysical variables issued from geostationary satellites, compared to monthly variables provided by databases.
In this paper, we demonstrate the advantages of using LSA-SAF biophysical variables.We compare them to those from ECOCLIMAP-I, and describe a method to use them in a land surface model at the same spatial resolution.A set of validation results over Europe and Africa and an example of application over West Africa further evaluated by comparison with complementary remote sensing data are presented.
In Sect.2, we give a short description of the material for this study, which consists of a land surface model, the vegetation database used (ECOCLIMAP-I), and the LSA-SAF biophysical variables.In Sect.3, we present a method to use the LSA-SAF biophysical variables in the land surface model in view of operational implementation.In Sect.4, we compare the new vegetation dataset to ECOCLIMAP-I, highlighting its advantage.We analyze its characteristics and deduce a practical estimation mean to obtain robust estimates of biophysical variables.Then we evaluate the impact of using the new vegetation dataset on the modelled evapotranspiration against in-situ observations in Europe and Africa.Lastly, different aspects of the methodology and its application are discussed and conclusions from this work are outlined.

The land surface model
The land surface model is based on H-TESSEL, the ECMWF land surface model (Beljaars and Viterbo, 1994;Viterbo and Beljaars, 1995;van den Hurk et al., 2000;Balsamo et al., 2009).It follows a bulk resistance formulation for the estimation of the surface turbulent fluxes, with two resistances modulating the variation of the sensible and latent heat fluxes.The aerodynamical resistance accounts for the turbulence generated by air temperature or wind gradient, while the stomatal resistance includes the influence of vapor pressure deficit, fraction of absorbed energy and soil water availability on the opening of the leaves stomata via a simple parameterization (van den Hurk et al., 2000).In addition, vertical distribution of moisture and temperature in the soil is modeled as a solution to diffusion equations.Direct evaporation from interception is modelled.
The needed forcing of this model consists of short and long-wave radiation reaching the land surface, surface albedo, wind speed, air temperature, air humidity and precipitation.In this model, the spatial unit for the energy balance assessment is divided into different tiles, associated with PFTs.The estimation of the surface fluxes is performed for each PFT before averaging over the spatial unit.The formulation uses parameters describing the vegetation state for each PFT separately: LAI, Fveg, roughness lengths, a minimum stomatal resistance, and the vertical distribution of the roots in the soil.The PFT classes have been changed from the original formulation of H-TESSEL to match the classification of ECOCLIMAP-I (Ghilain et al., 2011) and the model parameters have been calibrated.
The land surface model can be used at different spatial scales from a single point simulation using as input local measurements up to coarse grids used by global circulation models.In this paper, the model is applied at meso-scale and investigates the utility of the LSA-SAF biophysical variables.We therefore work in the MSG/SEVIRI grid, with a pixel as the basic model spatial unit.
Meteorological forcing is provided by the Global Circulation Model re-analyses ERA-Interim (Simmons et al., 2006).The spatial resolution available is 0.5 • and temporal sampling is 3 h.Meteorological forcing is linearly interpolated in time to half-hourly values.Tri-hourly precipitation rates are distributed equally for each 30 min step.If radiative forcing is missing, ERA-Interim is used to fill the gaps.
Complementary information on vegetation is needed: a land cover map and biophysical variables.Since each pixel can be composed of different PFTs, biophysical parameters must be known for each PFT inside the pixel, as well as the exact percentage that the PFTs occupy.The land surface model outputs an ET estimate each 30 min.

The ECOCLIMAP-I vegetation database
ECOCLIMAP-I (Masson et al., 2003;Champeaux et al., 2005) is a land cover map at 1 km resolution.In addition, it provides a vegetation parameterization for use in SVAT models in the form of a synthetic global vegetation database (http://www.cnrm.meteo.fr/gmme/PROJETS/ECOCLIMAP/pageecoclimap.htm, 2006).It is widely used by numerical weather prediction models in European countries (e.g.ALADIN International Team, 1997); furthermore, it is also used for inter-comparison of land surface models in Western Africa (Boone et al., 2009;Grippa et al., 2011), and in specific applications of H-TESSEL (Wipfler et al., 2011).
Mixed ecosystems are decomposed into PFTs (defined in Table 1) and each PFT is associated with a 10 day or monthly varying set of biophysical variables (LAI, Fveg, roughness lengths) averaged spatially over all the grid cells occupied by the concerned ecosystem.ECOCLIMAP-I is used as in Ghilain et al. (2011).
The ecosystems are decomposed into PFTs compatible with the vegetation parameterization of the model, and the land cover map obtained at 1 km is projected onto the coarser model grid.The land cover map obtained gives the 5 dominant PFTs and their contribution to a satellite pixel.But, since one PFT can result from the combination of several ecosystems, LAI for each PFT is computed as the weighted mean contribution of the involved ecosystem's LAI.Monthly LAI maps at the satellite resolution are then created.

The LSA-SAF vegetation products
LSA-SAF LAI and FVC products (García-Haro et al., 2005a) are derived from MSG/SEVIRI using a two components Spectral Mixture Analysis method (García-Haro et al., 2005b;Verger et al., 2009).LSA-SAF operational center generates both variables daily, based on the past 5-days estimations, at the sensor spatial resolution.Retrieval uncertainty as well as a quality flag indicating the quality and reason for possible no production (e.g.snow, traces of snow, traces of inland water) is also provided (LSA-SAF PUM VEGA, 2008).
LSA-SAF products represent an interesting alternative to SPOT-VGT or MODIS for meso-scale applications because of its relatively high generation frequency, its near-real time availability and its improved time stability (LSA-SAF PUM VEGA, 2008).LSA-SAF LAI and FVC have been produced routinely since August 2005 with successive algorithm (VEGA) improvements related to methodology (VEGA v2.0, September 2006), input data quality (December 2006), and post-processing (VEGA v2.1, May 2008).

Exploitation of the LSA-SAF biophysical variables
LSA-SAF LAI and FVC are provided daily at the MSG/SEVIRI spatial scale over the whole field of view.At that scale, landscapes can be very heterogeneous, especially over Europe, and one pixel can represent a mixed signal of different ecosystems or PFTs.This information cannot be used as such in SVAT-type models because more details on each PFT is necessary to compute surface fluxes.In addition, there can be some gaps in the products or pixels in the images could be unprocessed, because the models require continuous, stable and consistent LAI time series.It is therefore needed to implement a procedure able to retrieve LAI at sub-pixel (or PFT level), and to cope with continuity and stability issues.A two-steps procedure is implemented For every pixel, it is necessary to build continuous and consistent LAI time series as close as possible to the actual LSA-SAF LAI product, LAI SAF .First, an annual LAI climatology is built pixel-wise using daily LSA-SAF LAI images from 2007 to 2010.The climatological year is divided in periods of 10 days.For each 10 days period, a mean LAI, LAI c , is calculated pixel-wise along with an error bar, σ c , corresponding to the standard deviation of the LSA-SAF LAI sample used.
Even if snow cover is detected, some spurious data, occurring especially during the 2007-2008 winter in northern latitudes and over coniferous forests, still persist.Similar problems have been encountered by Jiang et al. (2010) with the NOAA satellite.In May 2008, LSA-SAF changed the algorithm to VEGA v2.1 to screen those data by applying a postprocessing on the problematic areas.However, in order to take advantage of the longest time series available, the spurious data are filtered out using a threshold.Winter LAI higher than 85 % of the mean summer value are filtered out.The extent of the filtering period is based on the collection of the spurious data: it is longer for Northern European countries (e.g.Sweden, Finland) and shorter for the European midlatitude countries (e.g.Germany, France).
In order to get continuous time series needed by operational models (Jiang et al., 2010), a linear interpolation is applied to both LAI values and standard deviations to fill the gaps in the climatological series.
Finally, for each actual LSA-SAF LAI, an optimal interpolation scheme (Gu et al., 2006) is applied.In our case, LSA-SAF LAI time series are more stable in time than the MODIS LAI (LSA-SAF PUM VEGA, 2008), and the LAI climatology is created from 4 yr of data.Appropriate weighting factors have been tuned and the error on the climatology, σ c , is multiplied by 2.5 and noted σc .Consequently, the weight Flowchart of the two-steps procedure to obtain LAI for each PFT in a given MSG/SEVIRI pixel.First, continuous and consistent LAI series are created at pixel level, LAI a , using a 4-yr climatology from LSA-SAF LAI, LAI SAF , actual LAI SAF , and an optimal interpolation (OI) method.Then, the pixel LAI is decomposed into LAI PFT for each PFT contributing to the pixel area, using a land cover map, LAI a and an ordinary least square method (OLS).
of the climatology is decreased to better follow the daily remote sensing product.The result is a LAI for each pixel (LAI a ); corrected using a background climatological information (Eq.1).Therefore, most dubious data are removed, the time series are smooth and gaps are filled.
As in Gu et al. (2006), when , LAI SAF is unusable and LAI a = LAI c .An ordinary least-square algorithm (OLS) is then applied to derive LAI for each PFT, using LSA-SAF LAI and PFT percentages over a neighbourhod of 9 (3×3) or 25 (5×5) pixels (see Appendix A).Two assumptions are needed: (1) the percentage of each PFT in one pixel represents exactly the real cover of the land surface, (2) the vegetation parameters, e.g.LAI, in close neighbourhood are assumed to be homogeneous.This hypothesis is reasonable if we consider that, in a close neighbourhood, meteorological and climatic conditions are similar, and one can assume the growth of the natural vegetation will be the same for one specific plant type.For human-forced vegetation, like for crops, the assumption can still hold if we assume that agricultural practices are homogeneous in the region of interest, which is often the case for large areas.
Repeating the operation by moving the size-defined neighbourhood by one pixel at a time, we obtain a smoothed spatial average estimate of LAI for each PFT.
Special attention has been paid to the correct handling of remote sensing derived vegetation products in the land surface model, avoiding a simultaneous variation of LAI and Fveg (Oleson and Bonan, 2000;van den Hurk et al., 2003;Jiang et al., 2010;Miller et al., 2006) and defining the LAI for the vegetation needed by the model from the remote sensing derived product (Ge, 2009).Fveg is therefore constant and is built pixel-wise using the complement to the absolute minimum LSA-SAF FVC over 2007 and 2008.Assuming that the bare soil fraction has a nul LAI, consistent handling of LAI for the land surface model is ensured through OLS.
Because of the uncertainty of land cover maps due to misclassification (e.g.Jung et al., 2006), some erroneous LAI could be estimated from the least-square algorithm.We propose to assess the impact of the uncertainty on the retrieved LAI PFT at 120 pixels selected randomly over the Euro window.The uncertainty is evaluated by applying the LAI decomposition using 3 different land cover maps for 2007.ECOCLIMAP-I, MODIS (MCD12Q1) and GlobCover (Bicheron et al., 2006(Bicheron et al., , 2008) ) land cover maps, available at 1000, 500 and 300 m resolution respectively are considered here.For each PFT in each selected pixel, the PFT percentage differences (rescaled to Fveg) are computed and the time correlation and root-mean square difference (RMS) between the LAI PFT series obtained from the 3 land covers are estimated, by PFT and by pixel.Results are presented in Sect.4.2.
Since a PFT is a class of vegetation (Alton, 2011;Williams et al., 2009), pixel-to-pixel variation of LAI PFT may occur, and could affect the retrieval.In Sect.4.2, we analyze those variations.For a given PFT, only the LAI value of homogeneous pixels (PFT covers more than 99 % of the pixel surface), LAI Hg , are taken into account.A reference sample of LAI Hg is selected randomly from the total population.The difference between the reference sample and the remaining LAI Hg gives a probability density function of LAI PFT variability as a function of the distance to the reference, up to 3 pixels, being the maximum size of the neighbourhood window used in OLS (3 × 3, 5 × 5 and 7 × 7).As the probability density functions are nearly symmetric and unimodal, very close to a gaussian, we study the normalized standard deviation as a proxy for the uncertainty.Results are given in Sect.4.2.

Evapotranspiration modelling and validation strategy
The impact of using LSA-SAF LAI instead of ECOCLIMAP-I database in the land surface model is evaluated at two different scales: at local scale and over The European sites are situated in mediterranean (Puéchabon and Tojal), temperate (Vielsalm and Wetzstein) and boreal (Sodankylä) regions.Four European stations monitor the exchanges between a forest stand and the atmosphere: coniferous forest (Wetzstein and Sodankylä), evergreen forest (Puéchabon) and mixed forest (Vielsalm).The station of Tojal monitors exchanges over a C3/C4 grassland site.Most of the sites selected have a large homogeneous fetch, that allows a good representativity of the measurements.
At the selected sites, the model is forced as described in Sect.2.1.However, in order to better assess the impact of LAI on the simulations, in-situ measurements are used for radiation and precipitation input when available.ERA-Interim precipitation rates are used for Agoufou and Vielsalm sites.Precipitation is especially important in semi-arid environments (Merbold et al., 2009), and the more accurate the input is the better is the model output.Soil moisture and soil temperature for the four soil layers are initialized with ERA-Interim analysis, and the model is run over one year before the analyzed run.Model parameters are not tuned for the selected sites, and the general parameterization for a global simulation is set.
At large scale, we evaluate the impact on ET by comparing the daily ET rates over a latitudinal transect crossing a vegetation density gradient.As ECOCLIMAP-I gives an oversimplification of LAI spatial variability over Africa (Kaptué Tchuente et al., 2010); we compare the simulations along a north-south transect in West Africa, 15 • N to 6 • N, 0 • E, for the year 2007 that samples a full range of vegetation dynamics from an evergreen humid forest to desert.This exercise N. Ghilain et al.: Benefits of LSA-SAF biophysical variables for evapotranspiration modelling is especially relevant in the context of the growing interest in the land surface monitoring in Western Africa, where quantitative remote sensing is expected to play an important role (Stisen et al., 2008;Boone et al., 2009).The land surface model is forced as described in Sect.2.1 and half-hourly simulated LE are converted to ET rates and daily cumulated.
As there is no ET observation dataset to validate/verify the improvements at large scale over West Africa, other remote sensing products are of interest for output verification.We can assess the impact of using LSA-SAF LAI in the land surface model by comparing other modeled variables to complementary remote sensing derived data.The modelled skin temperature from H-TESSEL can be compared to the land surface temperature derived from infra-red sensors.Even if the model skin temperature has not the exact same physical meaning as the radiative land surface temperature measured by the satellite (LST), it is often considered to be comparable (e.g.Ghent et al., 2011) and LST is used for indirect model's verification (e.g.Edwards, 2009;Jimenez et al., 2008).A particular interest of LST for ET verification is its link with soil moisture.For example, during observation campaigns in Western Africa, LST was linked to the soil moisture retrieved with passive microwave sensors (Kergoat et al., 2011).In addition, LST from MSG has been used to derive evaporative fraction (Stisen et al., 2008), which should likely be mainly driven by soil moisture in that region.LSA-SAF provides land surface temperature images (LSA-SAF LST) every 15 min, based on MSG/SEVIRI at the sensor resolution (Trigo et al., 2008).LST product is suitable for comparison as it is provided with the same spatial resolution.
To avoid the effect of short-term fluctuation of the temperature or cloud contamination, daily surface heating rates, expressed in K h −1 , are calculated from LSA-SAF 15 min LST data, and from the simulated 30 min skin temperature separately.The heating rates are the slope of the linear fit through the land surface temperature data between sunrise and noon LT.

Comparison of ECOCLIMAP-I and LSA-SAF LAI
To show the advantages of the new LSA-SAF products, we illustrate comparison of LAI with ECOCLIMAP-I database.The mean monthly LAI from ECOCLIMAP-I and LSA-SAF are compared by PFT (Fig. 2) in Europe for March to November 2007.Each monthly LAI is based on homogeneous pixels, i.e. pixels occupied by exclusively one PFT, and six LSA-SAF LAI images per month.In addition, the monthly spatial and temporal LAI variability is estimated from LSA-SAF LAI.In Fig. 2, it is shown that ECOCLIMAP-I LAI has a bias of about 1 m 2 m −2 compared to LSA-SAF LAI.It is especially visible for evergreen coniferous forests during the summer months and crops for the beginning of autumn.This bias had also been reported in comparison with other products and in-situ data (Garrigues et al., 2008).A larger variability is found in LSA-SAF LAI for Swamp areas, forests and partly for grasslands.For crops, the variability is comparable.
For Europe, small differences are observed in average, probably because the number of defined ecosystems in the land cover map is sufficiently high.However, over Africa, LSA-SAF LAI provides more spatial variability compared to the database.As an example, we compare the LAI variability of PFTs from 3 different homogeneous ECOCLIMAP-I ecosystems (tropical African grassland, semi-arid African grassland, and Nile Valley and desertic crops).Three LSA-SAF LAI images are selected for comparison on three different dates (15 April, 15 August and 15 December), and the LAI distribution is compared to the single value given by ECOCLIMAP-I monthly estimates (Fig. 3).For tropical African grassland, LAI given by ECOCLIMAP-I ranges between 2.0 and 3.0, while LSA-SAF LAI gives in one image a distribution ranging between 0 and 4. For the ecosystem Nile Valley and desertic crops, LSA-SAF LAI analysis shows a double peaked distribution poorly represented by the LAI database climatology.The difference is especially striking for April.The database gives a LAI of 0.5, while the peak reaches its maximum around 2.8.Those differences will inevitably impact on land surface model applications at MSG/SEVIRI scale.
LSA-SAF LAI product allows a better temporal representation of the inter-annual LAI evolution.As an example, we compare the LAI time series from ECOCLIMAP-I and LSA-SAF over a pixel in Western Africa corresponding to the

Retrieval of sub-pixel leaf area index
As presented in Sect.3, LSA-SAF LAI needs to be processed to obtain continuous and coherent time series.We illustrate that process on one specific LAI time series over Wetzstein, a forested site in Europe (Fig. 5).The MSG/SEVIRI pixel is mainly composed of coniferous forest.LSA-SAF LAI (red points) presents unrealistic high values in the 2007-2008 winter.During the next winters, most of those unrealistic estimates have been screened out in the product.The LAI climatology (black points) follows the trend given by LSA-SAF LAI, with very high values in winter.Using the filter, we remove all the spurious data, and a linear interpolation is applied.After correction, the optimal interpolation procedure is applied to produce the LAI analysed, LAI a .
With the new pixel estimates LAI a , the method to decompose at sub-pixel scale is applied.However, as mentionned, the method is subject to two sources of error.The impact of the uncertainty of the land cover map and of the pixel-topixel LAI variability of PFTs is assessed here.
The impact of the land cover map uncertainty on the LAI retrieval is represented in Fig. 6, where the relation between the mean PFT percentage obtained from the three land cover maps and the RMS of the LAI time series for the 120 pixels defined in Sect.3.1 is shown.The mean correlation decreases with a smaller PFT percentage, and RMS increases.It implies that land cover differences have little impact on the retrieved LAI for the dominant PFTs, but that it has a big impact on the less dominant PFTs.In addition, we observe less uncertainty for crops class than for evergreen forests compared to the mean.
As shown in Fig. 7, effect of enlarging the neighbourhood to 5 × 5 and 7 × 7 pixels for OLS lessens the impact of land cover uncertainty: the correlation decreases only to 0.5 for very small PFT percentages.The effect of enlarging to 7 × 7 has no big impact compared to 5 × 5.
The impact of the pixel-to-pixel LAI variability by PFT is represented in Fig. 8.For each homogeneous PFT represented in Europe and Africa, the normalized standard deviation increases with the OLS neighbourhood size (Fig. 8).For Europe, the variability inside a 3 × 3 neighbourhood ranges between 13 % and 19 %.Increasing the pixel number to 5×5 degrades the quality of the assumption for all PFTs.The variability increases gradually from 3 × 3 to 7 × 7 for C3 crops and grasslands.For Africa, the uncertainty is much less than for Europe, ranging between 5 % and 9 % for all classes in the case of the 3 × 3 neighbourhood.In addition, enlarging the neighbourhood to 5 × 5 or 7 × 7 has no big impact for most PFTs in Africa, except for grassland in Northern Africa.Variability could not be assessed for deciduous broadleaved forests over Africa because no african ecosystem defined in ECOCLIMAP-I is homogeneously covered by that PFT.The method proposed in Sect. 3 leads to a more robust estimate of LAI PFT when applied over a larger neighbourhood (Fig. 7), since impact of land cover map uncertainty is reduced.However, at least for Europe, LAI variability of a given PFT increases with the size of the neighbourhood (Fig. 8).Therefore, depending on the application, a compromise has to be found between accuracy and robustness.In applications involving spatial resolution equal to MSG/SEVIRI, the smallest size is obviously preferred since it implies less uncertainty.However, it has been shown that the quality of the OLS algorithm is decreasing along with PFT occupation fraction, due to the land cover uncertainty.Therefore, in waiting for a better agreement between land cover maps, a correction has to be applied, to estimate a reliable LAI PFT even for the least contributing PFT to the total pixel ET.A corrected estimate, LAI PFT,PostP , is derived from a weighted average of LAI PFT and a reference estimate, LAI R (Eq. 2).The weight α (Eq. 3) is proportional to the PFT percentage, and we choose it to be the linear regression of Fig. 6.
LAI R could be the LAI pixel value, LAI a , or the LAI value corresponding to the nearest pixel homogeneous of that PFT, LAI Hg .Since African ecosystems are mostly a mix of PFTs, LAI a is adopted.For Europe, clear homogeneous PFTs can most of the time be identified and isolated, and LAI Hg is preferred.This choice is further discussed in Sect. 5.

Impact on modelled evapotranspiration
Statistical results for the 30 min model estimates of LE are presented in Table 3. Statistical scores show an improvement in the modelling of the surface latent heat flux, or equivalently ET, especially for semi-arid regions where the signal Table 3.Comparison of simulations with ECOCLIMAP-I (ECO) or LSA-SAF LAI (LSA) against 30 min observed LE (W m −2 ): statistical scores: correlation coefficient (ρ), root mean square error (RMSE), bias and NS (Nash and Suttcliffe, 1970) shows an interannual variability in amplitude and phase.The largest improvements are observed for Skukuza and Agoufou sites, with a significant decrease of the global bias compared to data (−37 to −6 W m −2 for Skukuza, and −32 to −18 W m −2 for Agoufou) as well as an improved Nash-Sutcliffe (NS) index (0.46 to 0.60, and 0.26 to 0.45, respectively) (Nash and Suttcliffe, 1970).Even if the scores are largely improved in the Agoufou comparison, most of the remaining bias is due to non ideal precipitation forcing, i.e.ERA-Interim in that simulation.The other simulations over African sites, i.e.Demokeya and Tchizalamou, also show an improvement of the statistical scores when compared to data (NS: 0.53 to 0.62, and 0.51 to 0.53, respectively).Those results confirm the findings of Kahan et al. (2006) over the Sahel region in Western Africa.
As an example, we present in Fig. 9, the time series of the simulated and observed LE in Demokeya for the year 2007 to 2009.The simulation is clearly improved with the use of LSA-SAF LAI, with a better match during the wet season.Also, the simulations of LE during the dry season is closer to the observations.
Results over Europe show an improvement for 3 sites over the 5.In Puéchabon, the scores are well improved, with a global bias reduction, from −9 to 3 W m −2 , and a better NS, from 0.61 to 0.67.As well, the results are improved for 2 forested sites in different climates, i.e.Wetzstein and Sodankylä.For Tojal and Vielsalm, the model performs well with ECOCLIMAP-I and the statistics are either slightly degraded or equal.
The evolution of the LAI on the latitude gradient is shown in Fig. 10 for monthly ECOCLIMAP-I and daily LSA-SAF.The northern bound of the transect is completely arid with no vegetation.Southwards, vegetation is seasonal, with longer periods of vegetation cover, and higher amplitudes.The global pattern for ECOCLIMAP-I and LSA-SAF is similar, however, the most striking difference is seen on the length of the vegetation period that is noticeably shorter for LSA-SAF at highest latitudes.Between 6 • N and 8 • N, the transect meets lakes, resulting in zero LAI as observed from the dark blue stripes in the figures.
The evolution for 2007 along the transect is shown in Fig. 11, when using ECOCLIMAP-I and LSA-SAF LAI, respectively.The global pattern is mostly the same, as ET is driven by precipitation.However, large differences are found between September and December (also in April), with higher ET produced using ECOCLIMAP-I database.
An indirect evaluation of the changes is based on the comparison of the daily surface heating rates modelled and derived from the satellite.The heating rates are shown in Fig. 12. Vegetated areas are characterized by lower mean surface heating rates than bare soil.As well, surface heating rates are said to be increased with decreasing soil moisture (Stisen et al., 2008).The global pattern with the contrast low/high heating rate is seen in the three images.Good correspondence is found in the patterns of both simulations.That is probably due to the same precipitation forcing for the model.The absolute differences between simulated heating rates and those derived from LSA-SAF LST are represented in Fig. 13.A closer agreement, especially with a global lower bias, is found with the skin temperature simulated using LSA-SAF LAI than with ECOCLIMAP-I when compared to LSA-SAF derived heating rates, suggesting than simulation is improved using LSA-SAF LAI.

Discussion
Different aspects of the method to use LSA-SAF biophysical variables in H-TESSEL and its application for ET monitoring are discussed, especially in the perspective of continuous operational monitoring.

Retrieval stability of the biophysical variables
With one observation every 15 min over the same location, a higher level of stability of biophysical variables is reached when derived from MSG/SEVIRI comparatively to polar orbiters.Indeed, it has been shown that LSA-SAF LAI was more reliable than SPOT-VGT over central Africa where cloud cover is most frequent (Pekel et al., 2010).However, stability problems, that have been addressed by the methodolody proposed in Sect.time series for Northern European regions covered by coniferous forests for 2007 and 2008, while further production by LSA-SAF made used of a filter to screen out those spurious data.Similar problems are encountered for the generation of other biophysical products.GLOBCARBON LAI products show unexpected variations for boreal forests, with possible overestimated LAI (Garrigues et al., 2008).The MODIS Collection 4 and 5 LAI products present a spurious seasonal variability over boreal forests due to the low data availability with weak illumination, extreme solar zenith angles conditions, snow and cloud contamination (Yang et al., 2006a,b;Tian et al., 2006;Garrigues et al., 2008) and low spatial continuity due to regular failures from the main retrieval algorithm (Garrigues et al., 2008).Retrieval problems are also encountered during peak seasons over deciduous broadleaf forests (Garrigues et al., 2008).Corrections and smoothings are therefore needed for an optimal use in land surface models (Gu et al., 2006).

Towards an operational use of LSA-SAF biophysical products
In Sect.4.2, it has been shown that the land cover map uncertainty has an effect on the LAI retrieval at sub-pixel scale, and that a correction need to be applied to obtain reliable estimates using a reference.It has been proposed, for Europe, to use as reference the LAI of the nearest homogeneous pixel, LAI Hg .However, the success of it depends on the relative invariability of LAI for one PFT in a neighbourhood, and therefore on finding a LAI Hg within a close area.The analysis of (1) its variability within an increasing area (Fig. 14) and (2) the distribution of distance to the nearest LAI Hg (Fig. 15) leads to PFT specific conclusions in using it as a correction for the retrieved LAI PFT .As expected, the relative variability increases with distance for every PFT, with the sharpest difference within the first five pixels.In addition, the total increase is most important for grasslands (up to 55 % variability 30 pixels away) and crops (45 %), while it is less for forests (32 % and 37 %).
Figure 15 represents the distribution, based on ECOCLIMAP-I land cover map, of the distance from each pixel to the closest homogeneous pixel for the same PFTs in Europe.For every PFT, the number of occurence decreases with the distance.The decrease is sharp for crops, with more than 70 % of homogeneous pixels no further than six distance units.For the other types, there are less homogeneous pixels and the decrease is quasi linear.In 90 % of the cases, the nearest homogeneous pixel is found at a distance of 16 pixels (Crops), 30 (ENF), 31 (G), and 33 (DBF).
From Figs. 14 and 15, we conclude that (1) the farther the homogeneous pixel used for correcting the OLS solution comes from, the larger is the introduced error, especially for crops and grasslands; (2) the most uncertain correction in the post-processing will be for grasslands in Europe.LAI a is used as reference in that case, as well as for PFTs not represented by an homogeneous pixel in the neighbourhood.
We therefore propose a post-processing scheme, summarized by a flowchart in Fig. 16, integrating the results discussed here.If the area to process is Europe, C3 crops, and forests (DBF and ENF), LAIs are corrected with LAI Hg (option 2).For grassland, if the distance is higher than 10 pixels, the pixel LAI is used as surrogate (option 1).For other PFT classes, i.e.C4 crops, irrigated crops and swamp areas, and for other areas, correction follows option 1.

A decreased uncertainty associated to LAI
While a thorough analysis of the uncertainty is out of the scope of the present paper, we compare estimates of LAI uncertainty for ECOCLIMAP-I (σ LAI ECO ), and retrieved from LSA-SAF (σ LAI SAF ).σ LAI ECO can be evaluated with the spatial and temporal variability of LSA-SAF LAI.σ LAI SAF is the combination of the error on OLS, σ OLS , and the error on assumptions, σ dist (Fig. 8).
For Europe, LAI variability for a given PFT increases with the distance from a reference pixel (previous section).But, while for forests the variability remains almost steady with the distance, crops and grasslands are the most affected PFTs, giving a clear methodological advantage in using LSA-SAF LAI and a lower uncertainty.However, since ECOCLIMAP-I proposes a very fine classification into 92 ecosystems, e.g.27 different types of crops and 21 types of grasslands, the difference between σ LAI SAF and σ LAI ECO is not expected to be large.
For Africa, Fig. 3 shows examples representative of wide areas in the NAfr region for a given year (2007).Consequently, the deduced σ LAI ECO can be very large.Contrarily, σ LAI SAF is expected to be much less, because σ dist is less than 10 %, and σ OLS is very small due to the relative homogeneity of the landscape over the OLS neighbourhood.In particular, σ LAI SAF is expected to be much smaller than σ LAI ECO in semi-arid to arid environments in Africa, giving a more reliable LAI estimate for Sahel.That difference should obviously be exacerbated if we take into account the interannual variability not represented in ECOCLIMAP-I, as in Fig. 4.

Towards a continuous ET monitoring
The scheme presented in this paper and assessed with the land surface model has been developed to comply with requirements for near-real time applications.Applied to LSA-SAF LAI and FVC, it is convenient for ET monitoring through land surface modelling at spatial scales equal or coarser than MSG/SEVIRI.However, for a correct evaluation of LAI at PFT level, it is necessary to use a land cover map that proposes a sufficient accuracy in locating, classifying and decomposing into PFTs accepted by the land surface models.The gain in spatial resolution (ECOCLIMAP-I: 1 km; GlobCover: 300 m) and satellite overpass frequency Hydrol.Earth Syst.Sci., 16, 2567-2583, 2012 www.hydrol-earth-syst-sci.net/16/2567/2012/ are clear advantages to obtain more accurate LAI at PFT level, reducing σ OLS , and therefore providing an optimal input for land surface models.
In line with the monitoring concept, it would be necessary to have a clear assessment of the land cover changes, as it can affect our method.Regular updates of land cover would therefore be appreciated, to address land conversion issues like deforestation or fires.

Conclusions
In the present contribution, we show the benefits of using the daily biophysical variables derived from SEVIRI sensor aboard the geostationary satellite MSG for the evapotranspiration modelling and monitoring using the model H-TESSEL over Europe and Africa, compared to the use of the semistatic database ECOCLIMAP-I.
LAI maps from LSA-SAF and ECOCLIMAP-I are compared at different time and spatial scales, showing that ECOCLIMAP-I LAI presents less variability than LSA-SAF LAI, with respect to spatial coverage and temporal frequency.A rough analysis of the uncertainty on ECOCLIMAP-I LAI and LSA-SAF LAI over Africa suggests that LSA-SAF LAI is beneficial for land surface models intended for short time scales, e.g.half-hourly and daily, at MSG/SEVIRI resolution.
We propose a practical methodology to use the biophysical products LSA-SAF LAI and FVC in land surface models, correcting spurious data, filling gaps, and deriving sub-pixel LAI estimates needed by the models using a land cover map.Effects of the spatial scale and land cover map uncertainty on the method have been investigated, and results show that land cover uncertainty has an increasing impact on the less represented PFTs.The proposed method is shown to be more robust as the spatial scale considered is coarser.While waiting for land cover maps with reduced uncertainty, corrections of the errors on LAI by PFT are proposed for monitoring at the finest possible scale.
Given the respective features of both LAI sources investigated here, we have assessed the impact of using LSA-SAF LAI instead of ECOCLIMAP-I in H-TESSEL evapotranspiration simulations at MSG/SEVIRI resolution, both at local scale and at large scale.
Statistical scores of the comparison of simulated latent heat flux against local observations from ground measurement sites are improved in most cases, and especially in semi-arid climates.Amplitudes of anual cycles, as well as phase shifts are more correctly taken into account in the model compared to the use of the ECOCLIMAP-I database.
A 9-months simulation over a north-south transect in Western Africa using LSA-SAF radiation products and ERA-Interim meteorological variables, illustrates the impact of using LSA-SAF vegetation parameters compared to ECOCLIMAP-I database.An evaluation of the improvement is based on the similarity of the simulated skin temperature daily heating rates and the observed land surface temperature heating rates from MSG/SEVIRI, showing a better global agreement by using LSA-SAF biophysical variables.
Use of LSA-SAF LAI and FVC is therefore recommended for near-real time applications of land surface models working at the MSG/SEVIRI scale or coarser and intending at closely monitoring evapotranspiration.Moreover, it will improve the land surface model applications over Africa, capturing the full spatial, seasonal and inter-annual variability.
Fig.1.Flowchart of the two-steps procedure to obtain LAI for each PFT in a given MSG/SEVIRI pixel.First, continuous and consistent LAI series are created at pixel level, LAI a , using a 4-yr climatology from LSA-SAF LAI, LAI SAF , actual LAI SAF , and an optimal interpolation (OI) method.Then, the pixel LAI is decomposed into LAI PFT for each PFT contributing to the pixel area, using a land cover map, LAI a and an ordinary least square method (OLS).

Fig. 2 .
Fig. 2. Comparison of monthly mean and variability of LAI between ECOCLIMAP-I and LSA-SAF, for March to November 2007 over Europe.March to May ( ), June to August (•), September to November ( )

Fig. 5 .Fig. 6 .Fig. 7 .Fig. 8 .
Fig. 5. Illustration of the pre-processing of LSA-SAF LAI images on a test case over a forest (Wetzstein) in Germany.(a) A 10-days composite (black) is derived from LSA-SAF LAI daily data (red).(b) Spurious data are filtered out, and linear interpolation is applied to create the climatology, LAI c (black).(c) At last, an optimal interpolation scheme (OI) is applied using the climatology and the actual LSA-SAF LAI, LAI SAF , to provide the analysed LAI, LAI a .

Fig. 16 .
Fig.16.Flowchart with post-processing steps of PFT derived LSA-SAF LAI, LAI PFT , to obtain a consistent LAI, LAI PFT,PostP .Option 1 refer to the correction of the retrieved LAI with the pixel estimate, while option 2 points to the nearest homogeneous pixel as the reference LAI.

Table 1 .
List of PFTs defined in ECOCLIMAP-I and used in the land surface model.

Table 2 .
In-situ eddy covariance observations used in this study.Results of point scale simulations are compared to in-situ observations of latent heat flux, LE (list of stations in Table2), equivalent to ET.Four sites are situated in Africa and five in Europe, with a good sample of different climates.The four sites in Africa are installed in a savannah landscape, but with different precipitation regimes ranging from mostly dry (Agoufou) to mostly wet (Tchizalamou).
. Improved NS with LSA compared to ECO are in bold.