ECOCLIMAP-II / Europe : a twofold database of ecosystems and surface parameters at 1km resolution based on satellite information for use in land surface , meteorological and climate models

Introduction Conclusions References


Introduction
Land cover regulates the surface energy budget and hydrological cycle, which are essential inputs for climate and weather prediction models.It is strictly defined as the observed physical layer that covers the surface of the Earth, including natural and planted vegetation, and man-made constructions.Actually, land cover is one of the most crucial properties of the Earth system for many areas of benefit to society (GEOSS, 2005).Information on land cover is essential for the protection of environment quality and biotic diversity worldwide (Sutherland et al., 2009).It is also of primary importance for sustainable management of natural resources (EEA, 2005) and human needs (Vitousek et al., 1997).
In climate modelling, originally 1-degree global land cover databases were derived that combined pre-existing land cover maps and other atlases (Matthews, 1983;Olson et al., 1983;Wilson and Henderson-Sellers, 1985).Clearly, the coarse resolution of the grid mesh of a climate model led to mixing of vegetation species while focusing on broad scale natural ecosystems.This meant that climate modellers needed to reclassify pre-existing information in order to accurately model the land surface processes on the basis of a mosaic of individual ecosystems that were homogeneous from the functional point of view.The conversion of land covers -of ecosystems -into a suitable number of plant functional types (PFTs) is a matter of great concern as PFTs allow vegetation models to capture most variations of defined plant traits that seem to be better represented by state variables than by fixed parameter values (Gitay and Noble, 1997;Kattge et al., 2011).
During recent decades, the advent of satellite observations has fostered the development of land cover products compatible with landscape units.In this respect, vegetation indices that combine spectral measurements in the visible and near infrared spectral wavebands have been widely used to discriminate vegetation species.The normalized fifference vegetation index (NDVI), defined as the difference between the near infrared and red reflectance divided by the sum of the two, is undoubtedly the most widely used of the many indices available as it responds clearly to change in the amount of green biomass (Tucker, 1979;Hill and Donald, 2003), chlorophyll content (Dawson et al., 2003), fire (Telesca and Lasaponara, 2006) and climate variability (Gong and Shi, 2003).A pioneering global study using satellite-based information has identified vegetation species with the objective of creating a coherent worldwide 8 km land cover map (Defries et al., 1995).Since the beginning of the 2000s, with the advent of a new generation of onboard sensors having increased radiometric and geometric resolutions, numerous land cover maps have been developed within the frameworks of national and international initiatives.At the scale of the European Union member states, there is the CORINE mapping initiative, which has produced two classifications at 30 m resolution: one for the year 2000, based on single-satellite images (Landsat 7 ETM), and the other one for the year 2006, based on images from two satellites (SPOT-4 or IRS LISS).
The other most popular land cover maps use the International Geosphere-Biosphere Programme Data and Information System (IGBP DISCover) (Loveland et al., 2000), University of Maryland (UMD) (Hansen et al., 2000), Moderate Resolution Imaging Spectroradiometer (MODIS) (Friedl et al., 2002), ECOCLIMAP-I database (Masson et al., 2003), Global Land Cover (GLC2000) (Bartholomé and Belward, 2005) and GlobCover (Bicheron et al., 2006), which were produced using data from global observation systems such as NOAA/AVHRR, MODIS, SPOT/Vegetation and Envisat/MERIS at a spatial resolution of few hundred metres to 1 km.In addition to land cover classifications, the ECOCLIMAP product provides sets of surface parameters that are primarily useful in meteorology, notably surface albedo and leaf area index (LAI).However, stratification of land surface in Europe is permanently under investigation as knowledge of the geographic extent and dynamics of land cover is still incomplete, and broad levels of disagreement exist among current land cover maps due to the high level of landscape fragmentation at mid-latitudes (Herold et al., 2008;Fritz and See, 2008).
The aim of the present study is to update ECOCLIMAP-I at the European continental scale.This database was specifically designed to answer the needs of the meteorological community in investigating natural and managed ecosystems in connection with weather forecasting and climate change modelling (Masson et al., 2003).For example, the COSMO CLM initiative uses the ECOCLIMAP database (http://www.clm-community.eu).The rationale for building the new classification map (ECOCLIMAP-II/Europe hereafter) is to better discriminate the land cover classes over Europe than is done by the existing continental maps such as ECOCLIMAP-I, GLC2000, and MODIS products.The latter products refer to single annual cycles of satellite data and may capture some undesirable anomalies that could be avoided with a multi-annual time series analysis of consistent remote sensing observations, as has been clearly demonstrated over Africa (Kaptué et al., 2011a, b).Hence, ECOCLIMAP-II/Europe will take advantage of the improvements provided by SPOT/Vegetation (acquired during the 7yr period [1999][2000][2001][2002][2003][2004][2005] in regard to radiometry, calibration monitoring, atmospheric correction, and normalization of surface directional effects compared to the NOAA/AVHRR datasets (acquired between April 1992 and March 1993) that were used to produce ECOCLIMAP-I.
In this paper we describe the methods and datasets used to produce ECOCLIMAP-II/Europe.Section 2 first recalls the main characteristics of the ECOCLIMAP database and gives some technical information.In Sect.3, we detail the satellite information used as input for the classification (SPOT/VGT NDVI) and the aggregation tool (MODIS LAI), the existing land cover products, and finally the elements of validation.The method of K-means employed for implementing ECOCLIMAP-II/Europe is thoroughly described in Sect.4.1, along with the strategy for maintaining a minimum number of clusters and the technique of aggregation, i.e. limiting the division of land cover classes into a reduced number of PFT.The description of the method is illustrated in Sect.4.2 by looking over the 273 ECOCLIMAP-II land covers and combining them with 103 upper categories for a more synthetic view.Section 5 first draws a comparison between ECOCLIMAP-I and ECOCLIMAP-II in terms of functional types and LAI mean annual profiles, which should be relevant for the interpretation of further meteorological simulations.The validation procedure, which is the comparison with independent datasets at finer or coarser resolution, is detailed in Sect. 5.The last part, Sect.6, summarizes the study and presents some perspectives for exploiting the results.

Principles of the ECOCLIMAP database
The objectives of the ECOCLIMAP classification are first to perform a stratification of the landscape into land cover units.ECOCLIMAP-I consisted of a global land cover map of 215 ecosystems (or classes or covers) at 1 arc s resolution,  with a dataset of surface parameters associated for each grid mesh in tabular form: albedo, leaf area index (LAI), fraction of vegetation cover, fraction of photosynthetically active radiation (FPAR), roughness length, minimum stomatal resistance, and root zone.Figure 1 gives a schematic description of the organization of the ECOCLIMAP database, with the detailed terminology.This set of surface parameters is highly suitable for initializing Soil-Vegetation-Atmosphere Transfer (SVAT) models (Boone et al., 2009).As in other models investigating vegetation dynamics (Bonan et al., 2003;Rodell et al., 2004;Krinner et al., 2005), landscape scenes in ECOCLIMAP are organized by surface types or tiles (at first level, see Fig. 1) and PFTs or vegetation types (at second level within the tile nature, see Fig. 1).Sets of surface parameters are assigned mostly at the level of the PFTs representing generic vegetation types.The tile approach has been widely employed (Avissar and Pielke, 1989;Molod and Salmun, 2002).It consists of assigning surface parameters (albedo, LAI, and emissivity, for instance) to parts of the grid mesh within which these parameters vary as little as possible.The exercise in fact attempts to describe each land cover as a combination of possible fractions of surface types and PFTs.Hence, the spatial distribution of the vegetation within a given cover is crucial as it ascertains the subsequent aggregation of the energy, water, and carbon fluxes, which are calculated separately for each surface type and possibly for each PFT.Average fluxes over the entire grid cell are returned to the atmospheric model and are used as the lower boundary condition.In the SVAT model ISBA (Interactions Surface Biosphere Atmosphere) used at Météo France (Noilhan and Mahfouf, 1996), the content of each ecosystem is formulated as a linear combination of 4 main surface types or tiles: sea, water bodies, nature and urban areas.The nature tile is composed of 12 plant functional types (PFT; see Fig. 1): bare soil, bare rock, permanent snow and ice, deciduous broadleaf forest, evergreen broadleaf forest, needleleaf forest, C3 crops, C4 crops, irrigated crops, C3 herbaceous, C4 herbaceous, wetlands.C3 crops are winter crops (wheat, barley), C4 crops are summer crops (maize, sorghum).Requirements coming from land surface modelling establish the set of parameters that are needed.For each land cover and each PFT present in this land cover, ECOCLIMAP-I defined an annual profile of 10-day averaged LAI values, a root depth, a soil depth and a height for tree stands.Other surface parameters (notably fraction of vegetation, vegetation albedo, and roughness length) were only determined per PFT, regardless of the cover.They were assigned values that were constant or calculated with formulas relying on LAI, soil depth or tree height (Masson et al., 2003).Worth noting here is that the 10-day periods of ECOCLIMAP-I were defined from the 1st to the 10th; from the 11th to the 20th; and from the 21st to the end of each month.Known limitations of ECOCLIMAP-I are the obsolete character of the input datasets and the artificial delineation of the boundaries of the Koeppe and De Long (1958) climate zoning.Because of the need for the highest possible resolution over land, some initiatives for updating the ECOCLIMAP-I database have already been implemented.For instance, Han et al. (2005) updated the land cover over France.An example is the improvement of the description of biomes for south-western France, with which winter and summer crops could be separated, thereby leading to relevant detailed simulations of the atmospheric carbon dioxide in the CarboEurope Regional Experiment Strategy (CERES) (Sarrat et al., 2007).More recently, Kaptué et al. (2010) developed a new ecosystem classification within the ECOCLIMAP-II programme, with 37 distinct types over West Africa.This database was developed over the AMMA (African Monsoon Multidisciplinary Analysis) zone to provide upgraded information on the land surface properties of the West Africa region.In this study, GLC2000 classes were split using ECOCLIMAP-I classes.Then the MODIS LAI temporal profiles were used to group together the classes obtained.

Input and validation datasets
The update of ECOCLIMAP database was limited to Europe -defined here as the region comprised between longitudes 11 • W and 62 • E and latitudes 25 • N and 75 • N (corresponding to the area represented in Fig. 2) -in order to serve as testbed prior to achieving the global extension.The implementation was performed according to three steps: gathering input datasets to produce a new land cover map, delineating the definition and characteristics of land cover types, and, finally, consolidating the database through validation exercises.The characteristics of information sources used either as input or for validation are indicated in Table 1.

Synthesis of pre-existing land cover map products
The starting point was pre-existing land cover maps, which were used to delineate the boundaries and mark out the content of the new ecosystems.The first one was the popular Corine Land Cover map for the year 2000 (CLC2000), produced by the European Environment Agency and covering the 25 European Union member states (EEA, 2005).This map is available for the EU at a spatial resolution of 30 m (for vector data) and 100 m (for raster data used here) in the Lambert Azimuthal Equal Area (LAEA) projection.It mirrors land cover in Europe for the years 1999 to 2001.Based on photo-interpretation of high-resolution satellite imagery from SPOT (Système Pour l'Observation de la Terre) and Landsat Earth observation satellites, and embedding other sources of data (aerial photographs, topographic and thematic maps), national land cover maps were produced by each EU member state.CLC2000 consists of 44 classes with a fine breakdown into categories obtained by merging the consistent national products into one dataset.As CLC2000 can be deemed a fully dependable product, the European study area covered by this project appears trustworthy.
The second land cover map used in this study was the Global Land Cover 2000 (GLC2000) database produced by the Joint Research Centre (JRC) (Bartholomé and Belward, 2005).It yields a global product derived from the analysis of 14 months (November 1999 to December 2000) of daily global data acquired by the SPOT/Vegetation (VGT) sensor at a spatial resolution of 1/112 • (around 1 km).The product was developed on the basis of regional classifications made with the aid of regional expertise.Using a bottomup approach of 19 regional windows, the regional legends were based on the FAO (Food and Agricultural Organization) classification scheme (Di Gregorio and Jansen, 2000), which consists of 22 land cover classes globally.The GLC2000 whole package also contains a mosaic of five regional maps for Europe, including main land units with more detailed categories than the global one.These two land cover maps were combined, so that GLC2000 filled areas not covered by CLC2000.In this respect, the respective legends were simplified to depict the more widespread surface types forming the 14 categories: crop, needleleaf forest, broadleaf forest, herbaceous forest, rock, water body, bare soil, mixed forest, crop/natural vegetation mosaic, wetland, urban area, forest/other vegetation mosaic, irrigated crop, snow and ice.The resulting map, named C14, is shown in Fig. 2.This review of pre-existing map products will be of great support for a splitting of ECOCLIMAP-II new ecosystems into generic PFTs.This will also allow to better control the boundaries of the ecosystems only based on an analysis of NDVI time series.

NDVI data from SPOT/Vegetation
The overall objective was to build up a consistent map product at the continental scale privileging satellite information.The orbital configuration combined with the viewing geometry of the Vegetation (VGT) sensor, which has been on board SPOT-4 since 1998 and SPOT-5 since 2002, ensures daily global Earth coverage.Based on the maximum value composite (MVC) (Holben, 1986), 10-day composite products (S10) of the normalized difference vegetation index (NDVI) are produced at 1/112 • spatial resolution in a Plate-Carrée projection (WGS84 ellipsoid) (Hagolle et al., 2004;Maisongrande et al., 2004).The compositing period was defined from the 1st to the 10th, from the 11th to the 20th, and from the 21st until the last day of the month.The choice of the compositing period was a trade-off between the expected frequency of changes in vegetation and the minimum length of time necessary to produce cloud-free images.The period investigated spans seven years, from 1 January 1999 to 31 December 2005.This seven-year-long archive captures the mean annual vegetation cycle on a nearly climatic scale but can also be used to depict the inter-annual variability.S10 data composites also provide per-pixel cloud condition information, allowing most cloud contamination in the NDVI signal to be removed.If less than 4 unrealistic NDVI values occur successively, a linear interpolation is applied to fill the gaps.Otherwise, the gaps are kept as missing data.To fill in the gaps caused by cloud contamination, a 4-degree polynomial function is used.This approach is similar to that previously used by Mayaux et al. (2004) and Kaptué et al. (2010).NDVI time series are the central parameter used to process the automatic classification, as will be seen later (Set.4).

LAI data from MODIS
The Leaf Area Index (LAI) is defined as the ratio of onesided green foliage area per unit of horizontal ground area in broadleaf canopies, or the projected needle-leaf area per unit of ground area in conifer canopies, and is given in m 2 m −2 (Yang et al., 2006).In the original version, ECOCLIMAP-I, the seasonality of LAI was scaled on the annual dynamics of NDVI obtained from the Earth observing satellite system NOAA-AVHRR.The maximum value of LAI corresponded to the annual maximum green vegetation.In the new version, ECOCLIMAP-II, we consider collection 5 of MODIS LAI, the algorithm of which employs a look-up table (LUT) approach using the MODIS 8-biome land cover classification with the radiative transfer approach of Myneni et al. (1997).MODIS LAI is available at a spatial resolution of 30 arc-seconds in an Integerized Sinusoidal Grid (ISG) and at a temporal resolution of 8 days.It was re-projected on a Plate-Carrée grid to match VGT NDVI.MODIS LAI was also linearly interpolated for the sake of synchronicity with the ECOCLIMAP 10-day temporal resolution.The data were smoothed using a 4-degree polynomial fit, following the same procedure as described for the VGT NDVI.Unclassified and missing data, including urban areas, wetlands, snow, bare soil and water bodies, were excluded from the procedure.
These LAI data will allow to initialize LAI time series for each cover and PFT in ECOCLIMAP-II.

Climatic datasets
A climate database is used in order to avoid grouping classes pertaining to different climates.Two climate maps were actually used.The first, proposed by Koeppe and De Long (1958), is global with 16 climate classes.The second, produced by the FIRS project (EC, 1995), covers Europe and suggests 23 classes.It is leveraged by geo-factors such as climate, soil and topography.A combined map covering all areas of interest was built by assigning values of the FIRS classes to Koeppe and De Long's (1958) classes, which extended the former out of their area of definition (Fig. 3).Slight evidence of the coupling at the boundaries between the two original maps was notable in southern and eastern parts of Europe, which reveals conspicuous agreement between the two climate maps.In Fig. 3, the change in contrast for a same color aims at highlighting the transition between the two climate maps, thereby showing the good continuity between them.

Validation data
Multi-scale land cover verification benefits from multi-scale ancillary information in order to consolidate the strategy of distribution of the fractions of PFTs within the land covers.In the procedure of validation, the comparative products must adopt same common projection, resolution and quantity although the downstream strategy may be more specific to a dataset.
Statistics from the French Ministry of Agriculture and Forests can be found in the AGRESTE (http://agreste.agriculture.gouv.fr/)database.This database is composed of annual land use expressed in hectares for each French department, distinguishing different types of crops and nonagricultural areas.The available years are from 1999 to 2005, which are common with the available NDVI datasets.Another means of verification was the 1 • global map of percentage of C4 vegetation produced within the framework of the International Satellite Land Surface Climatology Project (ISLSCP-II) Initiative II Data Collection (Still et al.,  2003).We also considered a classification product prepared by CESBIO (Centre d'Etudes Spatiales de la BIOsphere) at 20 m resolution for an area of 3600 km 2 located near the city of Toulouse (France) where crop and forest types are notably encountered.The method to derive the CES-BIO product was a supervised maximum likelihood combining multi-date (6 dates) and multi-spectral (21 bands) data from FORMOSAT and SPOT imagery.The resulting land cover map was updated each year from 2002 to 2005.Only the 4 yr average was considered for the purposes of this study, which concerned the percentage of FORMOSAT classes in ECOCLIMAP-II 1 km pixels.Note that all land cover maps were finally re-projected on a Plate Carré grid with the WGS84 geoid system.
This list of products of validation cannot be too exhaustive.The ECOCLIMAP-II area is vast and actually many various datasets could be considered.Because the same method to derive the ECOCLIMAP-II database was applied over the whole domain (Fig. 2), we expect that the outcomes of the validation considering France are also representative for other areas.

Description of the methods
Prior to stratifying the domain of interest into land cover types, we used the C14 map (Fig. 2) to mask out pixels identified as water (either inland or oceanic) or urban areas.Then, the classification was performed according to 3 steps: (1) use of the K-means algorithm on the NDVI dataset to build a large number of clusters of homogeneous vegetation, (2) reduction of the number of clusters, and (3) integration of the information from the existing land cover and climate maps.
After the classification process, the last step of our procedure was (4) to define, for each cover, the percentage of the 4 main surface types (land, sea, inland water and urban areas) according to the "tile" approach, and to determine the 12 fractions of PFTs inside the nature tile, and then the LAI profiles, the root and soil depths, and the heights of trees for each of the 12 functional types of natural land areas represented in the cover (i.e. the only parameters that directly depend on the cover, see Fig. 1).The complete procedure is summarized in Fig. 4.

(Step 1) The K-means algorithm applied to the NDVI time series
Clusters of pixels at 1 km resolution were formed using the K-means clustering algorithm of Hartigan and Wong (1979).This is an unsupervised learning algorithm that is suitable for clustering multidimensional datasets.The K-means method seeks to partition all points into k clusters such that, in a multivariate attribute space, the total sum of squares (or squared deviations) from a set of individual points -represented here by the pixels -is minimized with respect to an optimum number of cluster centroids.The algorithm can be parsed as follows: (i) a number, k, of points is randomly placed in the space represented by the objects to be clustered (here, objects are NDVI time series); (ii) using the Euclidian distance measure each object is assigned to the closest centroid; and (iii) once all objects have been assigned, the positions of the k centroids are recalculated.The process is repeated until the position of the centroids is stable.The final step is then to minimize the metric of the objects with respect to the k clusters.
The K-means algorithm is sensitive to the initial configuration of cluster seeds and does not necessarily find the optimal configuration corresponding to the global objective function minimum (Kanungo et al., 2002).In practice, the K-means algorithm can be run several times to reduce this effect.We start with randomly chosen centres of clusters.Then, after a few iterations, rapid convergence is obtained for the centres of clusters based on the criterion of stability to decide that an optimal distribution of clusters has been reached.The K-means algorithm is well-known for its efficiency and robustness in handling the bulk of datasets.However, some constraints and limitations seem to exist for this approach.Firstly, the number of clusters ascertains the precision for obtaining a sharp description of the whole dataset.Secondly, Jain et al. ( 2003) calculated the sensitivity of the algorithm to the initial positions of cluster centres because, in some situations, important populations may be misrepresented.Thirdly, the use of the Euclidian distance as a unique criterion may be too restrictive to describe the dynamics of NDVI time profiles.Nevertheless, in our approach, we did try to circumvent such difficulties by using first a large number of clusters k (around 2000), and then refining the selection using other criteria such as optimizing the combination between the correlation and the distance (Sect.4.1.2,Eq. 1).Moreover, due to the initially large number of clusters, it is expected that no specific patterns were buried in clusters that were too big.

(Step 2) Reducing the number of clusters
In order to reduce the number of clusters to the target number (set between 200 and 300), several criteria were tested on the centres of the clusters obtained.Finally a resemblance criterion (referred to as RC in the rest of the paper) was selected: where d and r refer to the Euclidean distance and the Pearson's correlation, respectively, between the NDVI time series averaged for all 2000 clusters.This criterion is a trade-off that serves to account for both the dynamics and intensity of the NDVI signal.The use of a squared correlation gave stronger weight to the correlation at this stage of the process.Families of clusters were formed by grouping NDVI mean profiles of clusters if the RC criterion computed over their NDVI profiles relative to all the other clusters in the family was below a fixed threshold.This latter threshold was empirically fixed to finally reach the expected total number of This provided a supervised step and the new classes were thoroughly examined at each step, in particular their spatial organization, NDVI mean profiles and the total number of clusters.Visual inspection was ultimately the main criterion at this stage of the classification.The value of the threshold was modulated according to the size of the clusters, a less demanding threshold being assigned to smaller clusters in order not to multiply poorly representative covers.Also, the maximum value of NDVI was examined: the threshold was increased if the NDVI value was considered negligible in order not to segregate clusters having a low vegetation rate.The interest of this supervised stage is both to analyse and strengthen automatic results from the K-means method.Besides, it also serves to calibrate number of clusters and precision of the final map to specific convenience.The above operations revealed that clusters built with only NDVI information were geographically consistent.Inspection of the distribution of clusters within the study domain generally revealed bundles of pixels for which a justification could be found by looking either at the orography, or at a coarse climatic zoning (latitude, proximity of sea and even of country boundaries for arable land).Such outcomes validated the choice of the NDVI as the main classifier.Even if NDVI did not succeed in describing all surface types characteristics, it was at least able to capture most of the variability of the land cover at the continental scale, thereby resulting in 270 final clusters for the classification product based on NDVI alone.

(Step 3) Integration of information derived from existing land cover maps
At the third step, to strengthen the coherency between the 270 NDVI-based classification map (described in Sect.4.2) and the C14 map (Fig. 2, described in Sect.3), for each NDVI-based cluster, the mean NDVI time profile per main land cover of C14 was calculated.Pixels that did not belong to the two C14 land covers that were dominant within the ECOCLIMAP-II cluster were moved to another cluster where their type would be more appropriate.This option was activated based on RC again, with an adapted threshold, but only if the Pearson's correlation was above 0.9.This operation was supervised and minimized the discrepancy between the NDVI-based clusters and the C14 map.It is worth noting here that this criterion would disregard the geographical distance.Actually, it was not even necessary to introduce it as geographical coherency of the grouped pixels was obtained by construction.
Climate maps were finally used to avoid the mixing of bounds of pixels belonging to different climatic areas.But this actually concerned very few situations.For instance, three classes were split into two to better fit the land cover maps, which means that there were three more classes after this step.A dedicated treatment was nonetheless provided for urban covers.Suburban areas based on both GLC2000 and CLC2000 information were classified using the above method based solely on NDVI, while the other urban covers were directly inherited from the land cover maps.The new, final ECOCLIMAP-II classification includes 273 covers.

(Step 4) Defining the surface parameters
A correspondence was established between the 273 covers, the 4 surface types and the 12 functional types of natural land areas.This step only focused on suburban areas (that contain a part of urban areas and a consistent part of natural land areas) and natural land areas, while no further parameters were needed for the other surfaces (pure urban classes, inland water bodies and sea).A synthetic interpretation was made of the CLC2000/GLC2000 classes appearing in a given cover in terms of PFTs.The crossing information between our classification and other land cover maps was found to be of great benefit to convert the covers into functional types.The percentages of presence of CLC2000/GLC2000 classes in ECOCLIMAP-II covers were used to fix the percentages of functional and surface types.
At the beginning of this step, only the LAI profile P (j ) inherited from MODIS satellite data and the previously established functional type fractions F i(j ) of a given cover j are known.An iterative technique based only on these two sources of information was implemented to determine the LAI profiles of functional types i, i = 1, .., 12, inside the cover j , P i(j ). Figure 5 describes this process in a diagram.
First, an approximation identified the LAI profile of the main functional type of a cover j : P 1(j ) with the mean LAI profile of this cover: P (j ).Then, the principle of disentanglement was to search another cover k in the vicinity, in which one minority functional type 2, 3,. . . of the initial cover yielded the major functional type of the new cover k.The LAI profile of the minority functional type of the first cover P 2(j ), P 3(j ), . . .was then associated with the LAI of the selected cover k, P (k).
The selective criteria for the procedure were, in order of decreasing importance: preponderance of the type in the cover, geographic proximity (with reference to the climate map), maximum correlation between the two associated classes with respect to the NDVI temporal profile.At this stage of the process, the LAI temporal profiles of secondary PFTs P 2(j ), P 3(j ),. . .were known for a given cover j .Then, the LAI temporal profile of the major PFT P 1(j ) was re-calculated by subtracting the LAI temporal profiles of secondary PFTs P 2(j ), P 3(j ),. . ., weighted according to their fractions F 2(j ), F 3(j ). . ., from the initial LAI P (j ).In a very small number of cases where negative LAI values were identified, we selected other covers of reference for the minor PFTs.Then, a new iteration was performed using the re-calculated LAI P 1(j ) as a guess.This algorithm showed rapid convergence after the second step was reached.The determination of root depth, soil depth and tree heights for the functional types was inherited from ECOCLIMAP-I as it was judged that no reliable additional source was known available to improve the values of these parameters within the framework of ECOCLIMAP-II development.

Qualitative description of the new ECOCLIMAP-II product
The method for building the ECOCLIMAP-II land cover map is essentially based on a trimmed analysis of the spatial distribution of the NDVI time profiles over the domain of interest.Actually, this highlights how suitably such variations of NDVI seasonal patterns are represented, as can be confirmed by a dependable comparison, with correlations up to 0.9, between SPOT/VGT NDVI and MODIS LAI calculated for each cover (except for urban areas, wetlands, snow, bare soil and water bodies) (see Table 2).The map was first realized at 1/112 • resolution and NDVI mean profiles for covers were calculated at this resolution.Then, the map was resampled at 30 arc-seconds resolution to fit ECO-CLIMAP and MODIS LAI resolutions.The following subsections are devoted to a description of land cover characteristics in terms of NDVI intensity and evolution with respect to their geographic location.This emphasis on NDVI patterns for the extended European domain will complete the information provided by GLC2000 and CLC2000 to characterize the new land covers of ECOCLIMAP-II.Major cover types are reviewed along with illustrations of their NDVI time profiles (Fig. 6).

Forests
The types of forests seem to evolve from north-east to southwest.In northern Russia, NDVI values reach their peak during summertime and fall to spurious values in wintertime due to snow contamination for most scenarios.Hence, the narrow seasonal peak mirrors the short warm season and activity (Fig. 6a).Approaching central Europe, the NDVI annual cycles of forests take on square shape, i.e. high NDVI values last over a longer time, with a reduced amplitude, signifying less variability in climatic conditions, although with various degrees of severity (Fig. 6b and c).Near the Mediterranean Sea, the annual amplitude of NDVI time series decreases further, and observed profiles sometimes even become flat (Fig. 6d).Clearly, permanently mild temperatures coupled with an increasing number of sunny days and proximity to the sea seems to support quiescent periods of thriving vegetation.

Herbaceous plants and shrubs
Over northern, central and western Europe, NDVI time profiles for herbaceous plants and shrubs resemble those of forest with a strong annual amplitude and sharp peak in the north and east (Fig. 6e).Moving towards the southwest, the annual NDVI variations become broader and more squareshaped again.In particular, Atlantic meadows can be distinguished by a regular, smooth, rounded NDVI time profile (Fig. 6f).Other noteworthy profiles are those of grassland located in the Massif Central (France) (Fig. 6g) and the mosaic of grassland and crops in the Vendée region (France) (Fig. 6h).Their inter-annual variability and annual cycles have no equivalent within the study area.Around the Mediterranean basin, about 5 types of herbaceous plants and shrubs can be catalogued: -The first has a "triangular" profile, with a first peak triggered in spring and a second, weaker one in autumn.
Classes inside this type are notably distinguished by the position of the time-shifted moderate second peak relative to the summer peak.This kind of herbaceous vegetation and shrub can be found in central Asia and Turkey (Fig. 6i and j).
-A single NDVI peak starting during springtime characterizes a second noticeable type.It is rather similar to the previous type except that the second peak is flattened.This type is located exclusively in North African and north Arabian regions with smooth NDVI variations (Fig. 6k).
-In some cases, the peak of the second type extends towards the wintertime, which yields a third type that is also present in North Africa and northern Arabia (Fig. 6l).
-A secondary peak occurring in wintertime, which, in contrast to the first type, is associated with a triangular profile, forms a fourth type.This type principally occurs in North Africa, Spain and Portugal (Fig. 6m and n).

573
-A final singular type also has a square-shaped NDVI present in wintertime.A somewhat similar NDVI pattern is noticeable in spring and summer for forested areas of west-central Europe and is also significant in areas surrounding the Mediterranean Sea (Fig. 6o).
So far, these five types show a rather sparse distribution around the Mediterranean basin, in poorly delimited locations, rather than forming a complete set of Mediterranean ecosystems.It is likely that different botanical properties in connection with different environments (climatic and/or geological) may be able to explain such apparent differences.

Crops
Crop areas are well managed and thus generally well delimited in space, having characteristics that form an integral part of a climatic region.Nonetheless, an exception has to be made for the Mediterranean basin, where the organization of covers is complex with a high level of mixed plots.For this region, NDVI profiles for crops resemble those of herbaceous plants and shrubs.The fifth type of profile corresponds to Spanish Estremadura agro-forestry areas (according to CLC2000).Crops and herbaceous plants are probably quite mixed there.
In the eastern part of Europe (Russia, Kazakhstan), NDVI for crops show stretched profiles, also triangular-shaped with a peak in the middle of summer (Fig. 6p).The most smooth, rounded NDVI profiles are found in southern Europe, notably along the Po plain (Italy) and in southern France.A tiny difference between winter and summer crops is noticeable within this type as the peaks are about 3 months apart (Fig. 6q and r).
In western Europe, for example in the Paris Basin, a typical NDVI profile consists of a first high peak during spring followed by a secondary, very small peak in early winter.This is characteristic of winter crops being sowed in early autumn immediately after the harvest and showing some growth before the dormant period that precedes further growth in the following spring (Fig. 6s).
Some regions are very well delimited and are marked by compact areas of crops: Bulgaria (Fig. 6t), Hungary (Fig. 6u), Turkey at the Bosphorus, Poland, Germany, south-west England, French Brittany, French Vendée, the Po plain, Spanish Castile.Nile delta crops have specific NDVI profiles with 2 peaks of equal amplitude (Fig. 6v).These profiles are structured in very small clusters and extend, although very locally, from Turkey to Syria.In contrast, several classes of crops are quite scattered and their NDVI profiles resemble those of forests and herbaceous plants.In such situations, it is believed that crops grow alongside other types of vegetation.

Areas of sparse vegetation and bare land
NDVI profiles for sparse vegetation look like those of herbaceous plants and shrubs but have a lower intensity.This points out the effect of density of vegetation on the intensity of the NDVI signal.Bare land areas show NDVI profiles that can be directly related to the height of the vegetation.

Miscellaneous
A clear difference is noticeable in soil occupancy between the land surface surrounding the Mediterranean basin and the rest of Europe.Generally speaking, classes that are located outside the Mediterranean region are geographically well outlined and rather compact.For these classes, the changes in NDVI time profiles can be ranked according to latitude but also depend on the presence of a sea or ocean nearby.In this case, land cover types can be referred to as pure.On the other hand, ecosystems bordering the Mediterranean region are made up of mosaics of vegetation kinds spread over broad geographic areas and are often referred to as mixed land cover types.In this respect, and unlike the rest of the domain, regions bordering the Mediterranean Sea do not permit a straightforward analysis of the spatial distribution of the land covers because of possible overlapping between vegetation units.In this case, the partitioning of land cover types into PFTs (or patches) is rather challenging.
Figure 7 proposes a simplified visualization of the map obtained.The 273 classes kept for modelling are grouped by proximity of content into 103 named classes.

Comparison with ECOCLIMAP-I map
With the elaboration of the new ECOCLIMAP-II database, a key objective is to update the former, rather obsolete ECOCLIMAP-I product.Another shortcoming with ECOCLIMAP-I was the use of AVHRR images having a pixel resolution between about 1 km (nadir) and 6 km (skewed scanning), which enhanced the existence of mixed pixels and overlap between classes.Also, the lack of highquality initial products like GLC2000 or CLC2000 at the time of implementing ECOCLIMAP-I may be perceived as a serious issue.Clearly, ECOCLIMAP-I was a first brick in the wall, and its reliability as well as interest were proven thanks to a wide utilization in weather forecast models.There is certainly a value here to quantify the consequences of the modifications and also input datasets used by performing a comparison between the LAI values of the two ECOCLIMAP versions in respect to the fractions of PFTs.This verification exercise is performed at 1 km resolution on a pixel-bypixel basis independently of land covers.The outcomes are composed of raster maps of the differences, which provide a critical tool for the application upgrade and the maintenance strategy of ECOCLIMAP versions.It will certainly be of great aid to identify in the future the impact of changing  the ECOCLIMAP physiography on the results of numerical simulations.

Comparison of LAI temporal profiles (Fig. 8)
The evolution of LAI with time is analysed by pixel of 1 km, which already embraces the landscape aggregation of PFTs.The following three quantities are considered for further analysis as explained below: -The correlation between LAI temporal profiles of the two ECOCLIMAP versions, in order to assess the potential changes in the dynamics of LAI.This correlation is given only when two ECOCLIMAP PFTs other than just bare land, rock and snow are present (Fig. 8a).High correlation indicates none or small differences between the two ECOCLIMAP databases, and low or even negative correlation means large differences.
-The relative difference, diffrel (%), in maximum and minimum values of LAI (Fig. 8b and c), defined as where V is the maximum or the minimum value of LAI on a given grid point and the subscripts ec1 and ec2 denote the ECOCLIMAP-I and ECOCLIMAP-II databases, respectively.These two quantities give information on the detection of changes in the LAI. Figure 8 displays maps for these three quantities, i.e. the correlation and the two cases of maximum and minimum of LAI applied to the formula in Eq. ( 3).
The LAI correlations (Fig. 8a) are dependable (values > 90 %) over the whole north-eastern part of the domain.The correlation values decrease towards the Mediterranean sea, even falling below 50 % along the Mediterranean coastline.Negative correlation over semi-deserts but this is not drastic because LAI remains low in these regions.The representation of LAI in the Mediterranean region is expected to be improved with ECOCLIMAP-II because the discrimination between land covers relies on a method that offers refinement and focuses, above all, on homogeneity for a given land cover with respect to NDVI and LAI.
The value of LAI max (Fig. 8b) in ECOCLIMAP-II is higher for the main forests and herbaceous areas, and lower for main crops and semi-desert regions.It has already been mentioned that MODIS LAI maximum values, on which the ECOCLIMAP-II LAI is built, are higher for forested areas than, for instance, the LAI values from CYCLOPES, which made use of SPOT/Vegetation observations (Baret et al., 2007).Otherwise, and perhaps incidentally, maximum heights were found to be generally equivalent between MODIS and CYCLOPES LAI values (Weiss et al., 2007), and the reason why MODIS was selected was the availability of longer time series of datasets.The results of this comparison in terms of correlation are summarized in Table 2.  Median correlation is always higher than 0.9, showing the good correlation between the three datasets.

Comparison of the fractions of PFTs (Fig. 9)
This comparison is complementary to the LAI temporal profiles because the quality of the update of ECOCLIMAP should also be judged through the new distribution of PFTs within the land covers.Incidentally, note that LAI has repercussions on some other parameters like the root zone, soil depth and aerodynamic roughness (for low vegetation) but this is not important enough to deserve a dedicated analysis.The parameter investigated in this section is the simple difference F ec2 − F ec1 , where F is the fraction (expressed in %) of the PFT considered in a given grid point.Figure 9 shows these differences on a map for the 12 PFTs.
Bare soil, bare rock, snow (Fig. 9a, b and c) The percentage of bare land (Fig. 9a) increased almost everywhere and particularly on the marked topography near the Mediterranean Sea (close to +50 %) and in pre-desert zones north of the Caspian Sea.Different elements of an explanation can be found, like an increased number of burn scars for the Mediterranean region but also a better pixel resolution in the case of ECOCLIMAP-II.However, the greater area of bare soil in place of rock zones, notably at high altitudes in Norway and in the Alps, seems to be an artefact, rather than being inherent in the differences in method and data quality between the two versions of ECOCLIMAP.The fraction of bare land is lower along the coasts of the Mediterranean (−10 %), the explanation for which may be found in management policy.In the southern deserts of the domain, some bare soil has become rock and conversely so, which is thought to be due to the method.Concerning snow targets (Fig. 9c), some snow plots in ECOCLIMAP-I are replaced by land cover with 10 % of snow and 85 % of bare land in ECOCLIMAP-II.
These changes are noticeable because of the will to include a large range of nuances between the pure land cover types in ECOCLIMAP-II in order to access a continuity between land covers that did not exist in ECOCLIMAP-I.On the other hand, the choice to focus on the NDVI homogeneity rather than on the pureness of the land covers could be accompanied by some imprecision.The mean distribution of PFTs inside certain land covers may not be totally exact, even if this point has been verified, e.g. by checking that CLC2000 and GLC2000 classes are well-blended in ECOCLIMAP-II covers where they appear.
Deciduous broadleaf trees, needleleaf trees, evergreen broadleaf trees (Fig. 9d, e and f) Broadleaf trees (Fig. 9d) are now more present in central Russia (+40 %, +70 %) while they have tended to disappear from the Mediterranean region, especially near the coastline (−25 %).It is worth emphasizing that needleleaf trees (Fig. 9e) are notably less represented in northern and western European mountains.They have been replaced by broadleaf trees in northern Europe and grassland in western Europe.The few evergreen trees initially found in Mesopotamia in ECOCLIMAP-I have vanished in ECOCLIMAP-II.
For the forest scenario, as for bare land areas, an effort was made to adequately reproduce the complexity of mosaics inside the land covers, hence discrepancies are naturally observed.Here again, this is at the expense of a clear discrimination of dominant land cover types.

C3 crops, C4 crops, irrigated crops (Fig. 9g, h and i)
In ECOCLIMAP-II, there are obviously less irrigated crops (Fig. 9i) in the north of the Mediterranean region, except on the Turkish west coast and in the south-west of France.These differences are related to the mixing of vegetation types in ECOCLIMAP-II.Nile valley crops now appear irrigated in ECOCLIMAP-II.The fractions of C3 crops (Fig. 9g) are often lower in ECOCLIMAP-II because there are more mosaics of grassland with crops in the new classification.The category of C4 crops (Fig. 9h) is evenly more present in the Po plain, French Alsace and Hungary, thanks to the dating of maximum of NDVI profiles for the corresponding land covers.The fractions are slightly lower (−10 %) for the majority of the domain.
Temperate grassland, tropical grassland, wetlands (Fig. 9j, k, and l) The area of temperate grasslands (Fig. 9j) has shrunk in Russia, north-western Europe and in arid south-eastern areas (−20 % to −50 %).They are more present in a great part of central and western Europe, especially continentally (around +30 %), which counterbalances the tree loss as seen previously.Tropical grasslands (Fig. 9k) present in the Maghreb area and also in Mesopotamia in the case of ECOCLIMAP-I are no longer conserved in ECOCLIMAP-II.It is probable that the way fractions of vegetation types are decided tends to disregard this distinction, bearing in mind that Europe is not really an appropriate place to observe tropical grasslands.The distribution of wetlands (Fig. 9l) is completely modified in ECOCLIMAP-II.What seems to have happened here is that wetlands in ECOCLIMAP-II often show a high NDVI signal now, leading trees to be added in these areas.Conversely, the merging of CLC2000 and GLC2000 classes into ECOCLIMAP-II results in new areas of wetlands for some land covers.

Elements of validation for ECOCLIMAP-II
In this section, the goal is to carry out a validation exercise concerning the fractions of PFTs to be assessed for all pixels of a given land cover of the ECOCLIMAP-II database.
For this, we use independent sources of information to build maps of new spatial resolution, which then serve as references.to compare with the ECOCLIMAP-II fractions of PFTs computed for each of the 95 administrative divisions (departments) of metropolitan France.The fractions of PFTs are first weighted by the representative fractions of the covers in each department and then summed.These fractions are also reduced to 6 common distinctive PFTs: forests, grassland, C3 crops, C4 crops, permanent crops, plus all the other PFTs grouped into one.

Comparison with AGRESTE for France
Further, the information contained in AGRESTE hectares and ECOCLIMAP-II kilometre pixels is converted into percentages of land use in the administrative divisions, and compared at department level for these 6 types.The findings of the comparison are displayed in Fig. 10.An error estimate is given per department, and consists of the root mean square error (rmse) between the representative fractions of the 6 PFTs for ECOCLIMAP-II and AGRESTE: (3) It can be seen that this error falls below 15 % for the great majority of departments, which validates the approach.It should be stressed that it represents a cumulated error for all 6 PFTs and that, for a single PFT, it would come down to only 2.5 % on average.Nevertheless, it can be observed that a few departments, well spread over France (Loire-Atlantique, Cantal, Hérault and Ile-de-France) have errors higher than the average estimates, for instance up to 35 % for Ile-de-France.But this department is strongly urbanized, which brings out a somewhat thorny problem of representation.

Comparison with ISLSCP2 C4 map
The International Satellite Land Surface Climatology Project, Initiative II (ISLSCP2) notably addresses landatmosphere interactions focusing on land cover, hydrometeorology, radiation, and soils.In this respect, ISLSCP2 The Food and Agricultural Organization (FAO) (http://www.fao.org/es/ess/top/country.html) holds freely available yearly yield statistics per country for several categories of agricultural products.Over our domain of study, the first producers of maize in 2005 were France, Italy, Romania, Hungary, Ukraine and Egypt.This is in full agreement with the consistent results between the ECOCLIMAP-II and ISLSCP2 maps.For instance, higher fractions of C4 for ECOCLIMAP-II are observed in Italy, Hungary and Egypt.The interpretation of the discrepancies between the two maps can be oriented in two directions.First, owing to a better spatial resolution, the ECOCLIMAP-II products seem superior for the inventory of C4 crops.Moreover, the landscape stratification performed in the framework of ECOCLIMAP-II suggests a land cover classification based purely on a multi-temporal analysis of the surrogate variable NDVI.Therefore, caution is advisable as to the homogeneity of the land covers concerning C4 fraction representation because a low fraction of C4 NDVI would be drowned in the pixel integration of the NDVI signal.Unfortunately, the attempt to adjust C4 fractions in France with the aid of AGRESTE information could not be duplicated elsewhere.Clearly, the extrapolation of these C4 fractions outside France through the land covers must contain a part of uncertainty.
Generally, the spatial attributions of C4 crops in ECOCLIMAP-II are in agreement with ISLSCP2 and FAO statistics (FAO, 2008) whereas the method for implementing ECOCLIMAP-II leads to some imprecision on the depiction of C4 fractions.Certainly, an equivalent of AGRESTE at European scale would bring new insights but information extrapolated from AGRESTE already helps in the setting up of a continental-scale map of C4 crops with a level of reliability that allows it to be fully exploited further.

Comparison with FORMOSAT-2 products
FORMOSAT-2 is an NSPO (Taiwan National Space Program Office) Earth imaging satellite with the objective of collecting high-resolution panchromatic (2 m) and multispectral (8 m) imagery for a wide variety of applications, such as land use, agriculture and forestry.FORMOSAT-2 is able to revisit the same point on the globe every day in the same viewing conditions.FORMOSAT observations of a very small area (60 × 60 km south-west of Toulouse, France) at very high resolution (about 2 m) have been acquired to establish a classification product at a resolution of 20 m with 21 land covers.The study area is composed as follows: 23 % of wheat crop, 21 % of grassland, 17 % of sunflower, 10 % fallow, 9 % of man-made material, 6 % of deciduous forest and 6 % of maize, the rest being a mosaic of different landscape units (sorghum, soybean, barley, rapeseed, conifers, river).The strategy for verifying ECOCLIMAP-II is different here than previously.
The surface cover information from FORMOSAT was crossed with boundaries of ECOCLIMAP-II covers, allowing for the estimation of percentages of the FORMOSAT classes inside each ECOCLIMAP-II land cover present in the area of interest.Moreover, FORMOSAT classes were grouped together in order to match the ECOCLIMAP surface types and PFTs nomenclature as far as possible, in order to be able to compare the content of ECOCLIMAP-II in terms of PFTs and in terms of FORMOSAT covers.
Figure 12 shows the percentages of presence for the 8 most common surface types or PFTs (abscissas: bare/waste land (1), broadleaf forest (2), needleleaf forest (3), C3 crops (4), C4 and irrgated crops (5), grassland (6), water bodies (7) and urban areas (8)) for each of the 12 most frequent (in terms of number of pixels) ECOCLIMAP-II land covers in the FORMOSAT area (plots).The comparison was done following the description of the content of covers coming from ECOCLIMAP-II (blue curves) and from FORMOSAT (red curves).
For first cover 450, meaningful agreement was obtained except for the fraction of bare land (1), higher in ECOCLIMAP-II and seemingly offset by grassland (6) and urban areas (8) higher fractions in FORMOSAT.The confusion between bare land and urban areas is not shocking.Concerning this between bare land and grassland, in ECOCLIMAP-II, small areas of bare land scattering grassland, crops or forests were distinguished and assessed parts of the bare land PFT, what can explain the observed differences.These variations of weighting between bare soil and other PFTs is also visible for covers 446, 451, 459 and 376.For covers 457, 456, 452 and 393, agreement is quite good with sensible (but not affecting the general proportions) differences in the distribution between broadleaf forest (2), C3 crops (4) and grassland (6), that are the main PFTs present in the area.
For cover 492, there is a great discrepancy for the fractions of broadleaf forest (2) (higher in ECOCLIMAP-II) and grassland (6) (higher in FORMOSAT).
Finally, for covers 326 and 561, the agreement is very good except for slight differences in the fractions of C3 crops (4) and broadleaf forest (2), respectively.
As a conclusion of this comparison, it can be said that moderate differences are unavoidable because of definitions of PFTs being independent in the two cases, ECOCLIMAP-II and FORMOSAT.From this view, obtained results are very satisfying.
Considering the comparisons with finer (FORMOSAT) and coarser (ISLSCP2) land cover classifications, along with the satisfactory level of consistency with both of them, ECOCLIMAP-II can, at this stage, be deemed successful in properly aggregating small scale information and harmonizing broad scale information.This is a necessary condition for the efficient use of ECOCLIMAP-II in seamless climate models.

Summary and conclusion
The together with a thorough description and a verification exercises.More insight regarding the validation is expected in the near future owing to the use of ECOCLIMAP-II for running numerical meteorological scenarios.In this respect, the added value of ECOCLIMAP-II versus ECOCLIMAP-I should be outlined in the near future.This upgraded information concerning the physiography is intended to improve the representation of the continental surface in the SURFEX model (http://www.cnrm.meteo.fr/surfex/)and also to foster advanced investigations related to the carbon and water cycles.The spatial distribution and association of the land surface properties as previously defined within ECOCLIMAP-I has been revisited using enhanced, consistent long-term series of moderate resolution maps of NDVI and LAI originating from the new generation of onboard remote sensing instruments.The natural evolution of the landscape in connection with hazards (floods, fires) and human impacts (high concentration of habitat and population) mean that the European domain needs to be regularly redrawn.The popular method of K-means has once again proved its ability to help respect the main features of the landscapes assembled into a rather limited number of clusters (classes).Interestingly, the conversion of these clusters into PFTs, as is required for many applications and also for validation purposes, was possible without being detrimental to the fine quality of information.The LAI and NDVI tools of discrimination have proven that information initially compiled from land cover maps remained trustworthy.Incidentally, it is worth underlining the commendable coherence between the SPOT/VGT NDVI products used for the classification and MODIS LAI used for aggregation, without which the study would not have been possible.Combining two sources of information has no doubt strengthened the reliability of the ECOCLIMAP-II classification product, as it is less sensors dependent.The given details in ECOCLIMAP-I classification foster its consideration in some popular Internet portals (http://www.geo-wiki.org/)(e.g.Fritz et al., 2009Fritz et al., , 2011)).
The way to perform the NDVI classification was largely supervised, the main interest being to really learn how the K-means algorithm performs.The improvement of weather forecasting in Europe was the main driver of this work.The high level of fragmented landscapes in Europe can be deemed interesting to appraise the reliability of the method.In this regard, a future work will be in a direction of a global extension of ECOCLIMAP-II with the intent of harmonization with the product existing for Europe.When implementing ECOCLIMAP-II over other continents, this classification process should probably be simplified and made more automatic.On the contrary, the strategy to assign fractions of PFTs and LAI time series to covers should be rather similar.
It emerges that, at moderate resolution, typically of one kilometre, the majority of pixels are mixed at the scale of measurement, particularly near the Mediterranean.Since the genuine asset of the ECOCLIMAP-II database is to include coherent sets of biophysical variables primarily used in meteorology, the consistency of variables like LAI and fraction of vegetation was verified after broad-scale aggregation.The uncertainty given on ECOCLIMAP-II surface parameters according to their intra-class and inter-annual variability favours their use in data assimilation systems of carbon and water budget models.

Figure 2 .
Figure 2. Classification map with 14 main land covers (C14) resulting from the combination 10 of CLC2000 and GLC2000.11

Figure 3 .Fig. 3 .
Figure 3. Climate map built using the FIRS climate map completed by the climate map of 13 Koeppe and De Long (1958) for the eastern part of the domain.14 15

Figure 4 .Fig. 4 .
Figure 4. Schematics of the different steps for the construction of ECOCLIMAP land cover 18 classification.19 20

Figure 8 .
Figure 8.Comparison of ECOCLIMAP-II and ECOCLIMAP-I LAI: Pearson's cor coefficient (a), normalized difference (%) of maximum LAI (b), and minimum LAI (c the normalized difference is defined as the difference between ECOCLIMAP ECOCLIMAP-I divided by the average of the two.

Fig. 8 .
Fig. 8.Comparison of ECOCLIMAP-II and ECOCLIMAP-I LAI: Pearson's correlation coefficient (a), normalized difference (%) of maximum LAI (b), and minimum LAI (c) where the normalized difference is defined as the difference between ECOCLIMAP-II and ECOCLIMAP-I divided by the average of the two.

Figure 10 .Fig. 10 .
Figure 10.Results of the comparison of ECOCLIMAP-II and AGRESTE for each French 9 administrative department.Quantities plotted are cumulated errors (in percentage) for the six 10 land use types examined according to Eq. (2).11

Table 1 .
Summary of the main characteristics of the data used in this study.

582, 2013 570 S. Faroux et al.: ECOCLIMAP-II/Europe clusters
, and to obtain a balance between big and small clusters.Several tests were performed on the threshold value.