The LC_CCI combined spectral data from 300 m full and 1000 m reduced resolution MERIS surface reflectance (and SPOT-VEGETATION for the pre-MERIS era) to classify land cover into 22 level 1 classes and 14 level 2 sub-classes following the UNLCCS legend (Di Gregorio and Jansen, 2000). The whole archive of full and reduced resolution MERIS data, 2003–2012, was first pre-processed in a series of steps that include radiometric and geometric corrections, cloud screening and atmospheric correction with aerosol retrieval before being merged to 7-day composites. An automated classification process, combining supervised and unsupervised algorithms, was then applied to the full time series to serve as a baseline to derive land cover maps that were representative of three 5-year periods, referred to as epochs, for 2000 (1998–2002), 2005 (2003–2007) and 2010 (2008–2012). The classification process was achieved through back- and up-dating methods using the full-resolution SPOT-VEGETATION and MERIS time series. The three global land cover maps described all the terrestrial areas by 22 land cover classes explicitly defined by a set of classifiers according to the UNLCCS, each classifier referring to vegetation life form, leaf type and leaf longevity, flooding regime, non-vegetated cover types and artificiality. Inland open-water bodies and coastlines were mapped using wide-swath mode, image mode at medium-resolution (150 m) and global monitoring image mode (1 km) acquired by the Advanced Synthetic Aperture Radar (ASAR) sensor aboard ENVISAT satellite for a single period (2005–2010).

In addition to the land cover classification, the land surface seasonality products describe, for 1 km2 rather than 300 m resolution, the average behavior and the inter-annual variability of the seasonal normalized difference vegetation index (NDVI), the burned area, and the snow occurrence, computed over the 1998–2012 period. These seasonality products were spatially coherent with the land cover classification and were provided at weekly intervals averaged over this 15-year period and were based on existing independent products: SPOT-VEGETATION NDVI daily time series, MODIS burned area (MCD64A1), and MODIS snow cover (MOD10A2). All products are provided to users in NetCDF and geotiff file format referenced to plate carrée projection using the World Geodetic System (WGS 84) and are available at http://maps.elie.ucl.ac.be/CCI/viewer/. Detailed descriptions of each component in the processing chain can be found on the European Space Agency Land Cover Climate Change Initiative website: http://www.esa-landcover-cci.org.

The conversion of land cover classes to PFTs is a non-trivial task that is made more complicated by the fact that the number and description of PFTs are not standardized across DGVMs. In the past, land cover (and other) information has been used to derive PFT maps based on individual model PFT descriptions. The method used to convert the land cover to PFTs has not always been documented in detail for each model. The aim of the approach taken here was to develop a general framework that could easily be adapted to the specific PFT description of any individual model. In consultation with the three climate modeling teams engaged in the LC_CCI project, Laboratoire des Sciences du Climat et de l'Environnement (LSCE), Met Office Hadley Centre (MOHC) and Max Planck Institute for Meteorology (MPI), 10 PFT groups were defined based on their phenology (needleleaf or broadleaf, evergreen or deciduous), physiognomy (tree, shrub, or grass) and grassland management status (natural or managed). Three additional non-PFT classes were added for bare soil, water and snow/ice. The cross-walking methodology is based on the approach of Poulter et al. (2011) and assumes that each UNLCCS category could be split into one or more PFT classes according to the LC class description at the per pixel level (Table 2). For example, the “cropland” UNLCCS land cover class was assigned as 100 % managed grass, whereas the UNLCCS “tree cover, needleleaf evergreen, open (15–30 %)” class was assigned to 30 % needleleaf evergreen, 5 % broadleaf deciduous shrub, 5 % needleleaf evergreen shrub and 15 % natural grass. Of note, wet tropical forest vegetation, mainly the UNLCCS class “tree cover, broadleaf evergreen, closed to open (> 15 %)”, was assigned to the PFT categories of “broadleaf evergreen” tree (90 %) and deciduous (5 %), evergreen shrub (5 %) following observations that moist tropical forests tend to have indeterminate phenology rather than distinct periods of onset and offset (Borchert et al., 2002; Fontes et al., 1995; Reich and Borchert, 1984). The derivation of Table 2 was the result of consultative process among the producers of the land cover map and the three modeling groups that reached a consensus on the PFT fractions for each LCCS-defined land cover class. The aim of this process was to gain a fuller understanding of the methods behind, and implications of, the respective vegetation classifications (LC and PFT). For example, previous LC class descriptions have included “semi-deciduous” in the description of broadleaf evergreen trees, as in tropical rainforests in particular, phenological strategies of certain species result in more pronounced seasonal leaf dynamics. However, such subtle differences in functionality are not currently incorporated into DGVMs, and tropical rainforests are considered to be 100 % evergreen. Thus, in the cross-walking table derived in this study, the relevant LC class was mapped only as evergreen trees and shrubs (see LC class 50 in Table 2). Other issues that were discussed included how different vegetation types are treated within a grid cell for DGVMs and the lack of representation of over- and understory canopies, which both had implications for how to deal with mosaic and open-cover classes.

For the most part, the cross-walking approach followed the definitions of the UNLCCS classes, where fixed proportions of land cover were split using a one-to-one rule for the respective PFT categories, as described above. In cases where the UNLCCS class was defined by a large range of tree cover and with no upper bound, i.e., “> 15 %” (Table 2), the uncertainties in this conversion can be considered larger than compared with other categories. In these cases, the land cover remote sensing team of experts provided the criteria for the conversion approach, taking into account their improved understanding of the constraints of DGVMs. The impact of these uncertainties on the final PFT fractions, and on the simulated variables, is beyond the scope of this study. Here we purely aim to properly document a new, generic method for mapping between LC classes and PFT fractions that can be used for all DGVMs. However, the issue of uncertainty in the cross-walking procedure is currently being investigated in phase 2 of the LC_CCI project.

The LC_CCI land cover and seasonality products are initially downloaded in full spatial resolution, i.e., 300 m grid cells for land cover, and 1 km grid cells for the seasonality products, at global extent in plate carrée projection. In order to fulfill a range of ESM requirements, the LC_CCI project team developed the LC_CCI user tool to allow users to adjust parameters of the LC products in a way that is suitable to their model setup, including modifying the spatial resolution and converting the LC_CCI classes to fractional PFT area. The BEAM Earth Observation Toolbox and Development Platform, designed for visualization and analysis of ENVISAT products, was selected to provide the basis of the conversion software. A list of resampling resolution and coordinate system options is provided in Table 3. The coordinate re-projection and aggregation of the LC_CCI data uses slightly different resampling algorithms depending on whether the tool is used on the land cover or seasonality products. The tool converts the original LC_CCI geotiff file to target files produced in NetCDF-4 format and following CF (Climate and Forecast) conventions, more commonly used in numerical modeling. The open-source BEAM tool (source code at https://github.com/bcdev) can be run independently using either Windows or Unix-based operating systems and the compiled operational tool can be downloaded from http://maps.elie.ucl.ac.be/CCI/viewer/download.php.

For the land cover classes, the resampling algorithm produces an aggregated LC_CCI data set that in addition to the fractional area of each PFT, also includes the fractional area of each LC_CCI UNLCCS class, the majority (dominant) LC_CCI UNLCCS class and the overall accuracy of the aggregated classification. The majority class n is defined as the LC_CCI class which has the rank n for the sorted list of LC_CCI classes by fractional area in the target cell (see Fig. 1). The number of majority classes computed is a parameter, which can be defined by the user, so that the full number of LCCS classes can be reduced to a user-defined subset, i.e., the top 3. Each original valid land, water, snow or ice pixel contributes to the final target cell according to its area percentage contribution. The accuracy is calculated by the median of the land cover classification probability values weighted by the fractional area.

The aggregation of LC_CCI seasonality products is specific for NDVI (i.e., greenness), burned areas and snow cover. In the case of the LC_CCI NDVI condition, the mean NDVI over all valid NDVI observations is included in the aggregated product. The burned area and snow cover LC_CCI products also contain three different layers: the proportion of area (in %) covered by burned or snow area, the average frequency of the burned area or snow area detected over the aggregated zone and the sum of all valid observations of burned or snow area. Similar to aggregation rules for land cover, each original pixel contributes to the target cell according to its area percentage, but the value of a pixel will only be considered if its value falls within its valid range, i.e., zero to one for NDVI.

The LC_CCI tool provides users with a zero-order classification; that is, the PFT classes are defined as broadly as possible so that users have the advantage to continue to aggregate to the requirements of their model (Fig. 2). For example, models that do not include shrub PFTs can merge shrub and tree categories together to create a single woody PFT category. Modeling groups that require climatic distinctions for PFTs; for example, temperate versus tropical versus boreal types can use their own climate or biome data sets, such as Köppen–Geiger or Trewartha ecological zones (Baker et al., 2010; Kottek et al., 2006; Peel et al., 2007), and define classification rules based on temperature thresholds, for example (Poulter et al., 2011). Most models also require a distinction between the C3 and C4 photosynthetic pathways for different grass species, where C4 is more common in warm and dry climates (Edwards et al., 2010; Still et al., 2003). The photosynthetic biochemistry of C4 grasses is very different to C3 grasses and their distribution can be mapped either according to climate (Poulter et al., 2011) or to some combination of remote sensing, ground-based observations and ecosystem modeling (Still et al., 2003). The LC_CCI managed grassland PFT category represents all non-irrigated, irrigated and pasture lands, and therefore drawing finer thematic distinctions between these must come from country or sub-country statistics similar to downscaling work made by Hurtt et al. (2006), Klein Goldewijck and Batjes (1997) and others (Monfreda et al., 2008; Ramankutty and Foley, 1998).

For analysis and demonstration of the tool, we compare the LC_CCI PFTs with the original PFTs used by the land surface model (LSM) components of the ESMs from the three modeling centers developing ORCHIDEE at LSCE (Krinner et al., 2005), JULES at MOHC (Clark et al., 2011; Cox et al., 2000; Pacifico et al., 2011), and JSBACH at MPI (Knorr, 2000; Pongratz et al., 2009; Reick et al., 2013). The original ORCHIDEE PFT map, based on 12 PFTs plus bare soil, has its origins in the Olson land cover data set from the 1980s (Olson et al., 1983) and the International Geosphere Biosphere Program (IGBP) DISCover data set for the period 1992–1993 (Loveland and Belward, 1997). This was implemented within ORCHIDEE using a look-up table approach to estimate PFT fractions (Verant et al., 2004). The JULES model also uses PFT distributions derived from the IGBP DISCover data set to estimate fractional coverage of five PFTs and four non-vegetated surfaces (water, urban, snow/ice and bare soil). JSBACH uses original data from Wilson and Henderson-Sellers (1985) and continuous tree fractions from DeFries et al. (1999) to represent the distribution and abundance of 12 PFTs. The LC_CCI Epoch 2010 was converted to 0.5 degree resolution using the LC_CCI user tool and compared with the individual default model PFT maps to illustrate regional differences and biases between products and to provide a baseline of how the LC_CCI products may improve LSM performance.

The global land areas covered by the aggregated 0.5∘ LC_CCI PFT equivalents (Fig. 3) are dominated by barren and bare soil (39 Mkm2), followed by forests (30 Mkm2), managed grasslands, croplands and pasture (25 Mkm2), natural grasslands (18 Mkm2), and shrublands (14 Mkm2). For comparison, the MODIS collection 5 land cover product developed by Friedl et al. (2010) covers barren area 18 Mkm2, forest and savanna at 49 Mkm2, a shrubland area of 22 Mkm2, and 12 Mkm2 for croplands. With reference to the Food and Agriculture Organization (FAO) statistics, forest area is reported as 38 Mkm2 (FAO and JRC, 2012), cropland area as approximately 15 Mkm2 (Monfreda et al., 2008) and pasture lands of 28 Mkm2 (Ramankutty et al., 2008). While part of the areal differences are explained by the spatial resolution between the moderate-resolution MODIS data (500 m) in comparison to the 0.5∘ LC_CCI data, thematic differences introducing uncertainty in aggregating to forest, grassland, classes and factors stemming from different definitions of forest cover thresholds are used to categorize forest land between the UNLCCS approach (10 % cover) and the IGBP (60 %) approach used for MODIS. In addition, the UNLCCS to PFT conversion approach considers assumptions related to plant community level variability, and so a bare soil fraction is introduced during the conversion (see Table 3) increasing its global area and partially explaining the difference with MODIS land cover.

Differences between the LC_CCI PFT data sets and the original PFT data sets were specific for each ESM (Fig. 4) largely because the original reference data were different per modeling group. Another challenge was that different PFT classification schemes were used for each model (Table 1), introducing further aggregation uncertainties in the comparison between LC_CCI and the original PFT data.

For all modeling teams, grassland PFT distributions showed the largest changes, with significant reductions in northern latitudes for ORCHIDEE and JULES (Fig. 6). For ORCHIDEE, the grassland PFT reductions were associated with an increase in bare soil, together with a shift from C3 grasses to (boreal) forest in the mid-to-high latitudes (Fig. 5). Agricultural PFTs, not included in JULES, were similar for the original ORCHIDEE and LC_CCI inputs at regional scales, but showed increases in tropical regions where deforestation activities were high, e.g., the Brazilian arc of deforestation region. JSBACH generally had a reduction in cropland area, especially over North America and the North African arid regions.

Over arid regions, in comparison to the original PFT map, JULES decreased in C4 grasses over Australia, with an associated increase in the fractional cover of shrubs and bare soil. In the Sahel, apparent differences in the definition of natural and managed C4 grass account for differences found between ORCHIDEE and JSBACH. The inclusion of the LC_CCI product resulted in a large increase in the C4 grass fraction over the Sahel in ORCHIDEE, whereas no significant change in the C4 grass fraction has been found over these areas for JSBACH. Instead, an increase in C4 crops was found over the Sahel for JSBACH. Since the JSBACH conversion also accounts for pasture, this difference may be well the result of the pasture definition, which is a weighted part of all herbaceous PFTs. This also partly explains why the JSBACH C4 pasture PFT decreases exactly in the same areas where the C4 crops increase due to the use of the LC_CCI data. In JULES, the C4 types over Sahel shift to bare soil.

In the tropics, reductions in broadleaf tropical tree cover were largely consistent across all three ESMs, although increases in broadleaf forest area were found for some parts of the African Congo Basin for JULES (Fig. 6). Needleleaf forest area increased compared to the reference data set for both JULES and JSBACH for boreal Europe and Australia (shrubland PFTs). The increase in needleleaf PFTs in boreal Europe was partially associated with a decrease in broadleaves (Fig. 6a and b) for all three models, but also a decrease in natural grassland cover.

The LC_CCI approach provides the ESM modeling community with a flexible tool for using up-to-date land cover information consistently provided over time. Following the requests of the user survey, the land cover data set is available across multiple spatial domains, conforms to standard file formats used in numerical models, and includes information on classification confidence levels for the land cover classes and resulting PFT fractions. The standardized conversion tool provides users with a consistent documented approach for aggregating land cover classes and thus overcomes limitations associated with consensus approaches (e.g., Tuanmu and Jetz 2014). Of particular importance is that the multi-temporal LC_CCI mapping approach facilitated more accurate mapping leading to improved remote sensing observations of deforested areas in the tropics, the tree line–tundra boundary in the high latitudes, and better distinctions between managed and non-managed grasslands in Africa. Additionally, the ASAR-based water bodies and coastline delineation helped to standardize the physical boundaries between terrestrial and water systems for all models. Using this standardized PFT mapping approach for ESMs can be expected to reduce model ensemble uncertainty as attempted by recent inter-model comparison efforts (Huntzinger et al., 2013).

During phase 1 of the LC_CCI project (2011–2014) several limitations of the conversion scheme and tool were recognized and have been targeted for improvement in phase 2, where improvements to the land cover thematic classes and to the conversion scheme will be made. For example, in the high latitudes, a reduction in grassland fractional cover was observed with the LC_CCI product for all models, and on further investigation, it was recognized that a better representation of lichens and moss vegetation (Class 140, Table 3) would be an improvement for the sparse vegetation category (Class 150), especially in the high latitudes. Conversion of high-latitude land cover classes to PFT equivalents has been a challenge in several recent regional studies (Ottlé et al., 2013; Wullschleger et al., 2014) where discriminating spectrally between shrubs and trees, or grass and non-vascular plant species, remains difficult. Accurate mapping of high-latitude vegetation can be particularly important for modeling wildfires (Yue et al., 2014), where the spread of tundra fire is sensitive to fuel loading. In the tropics, the seasonal cycle of forest canopies continues to be a contentious issue (Morton et al., 2014; Myneni et al., 2007; Poulter and Cramer, 2009; Ryan et al., 2014) with the binary distinction between evergreen and deciduous phenology proving to be overly simplistic where semi-deciduous traits are perhaps more appropriate (Borchert et al., 2002), and thus the development of tropical phenology traits that correspond to recent observations is a high priority (Bi et al., 2015). More specifically, phase 2 will (i) target improved thematic accuracy with a specific focus on transition areas (e.g., grassland-sparse vegetation-bare soil, tree–shrub–grassland) and the distinction between C3 and C4 grasses, (ii) create a historical land cover time series to cover the 1990s using 1 km Advanced Very High Resolution Radiometer (AVHRR) NDVI surface reflectances, (iii) include more detailed change detection, with more classes, i.e., IPCC land categories (forests, agriculture, grassland, settlement, wetland, other land) as targets, and (iv) deliver an albedo and/or LAI seasonality product.

Physiological traits such as nitrogen fixation and different photosynthetic pathways, C3, C4 or crassulacean acid metabolism (CAM) are presently not detectable from surface reflectance values, and so broad climate-based assumptions must be made to split into these groups. These assumptions can lead to large uncertainties that can impact a chain of ecosystem processes and land surface properties. While the LC_CCI data set provides updated information on inland water bodies, the seasonality of water bodies and wetlands is yet to be represented and only considered in radar-based surveys (Schroeder et al., 2015). Finally, the existing 22 UNLCCS land cover classes currently do not include pastures, whereas the importance of grazing on biogeochemical cycling is becoming increasingly recognized (Foley et al., 2005). Instead, pastures are currently mapped as croplands or grasslands according to their degree of management. Better thematic discrimination between these three classes would clearly improve the carbon cycle modeling as agriculture, in the broadest sense, is a significant contributor to land degradation and anthropogenic global greenhouse gas emissions (Haberl et al., 2007). Earth observation products are generally limited to mapping land surface structural properties rather than functional properties, and model–data fusion approaches can help reconcile problems that might arise from this limitation, especially in the case of grassland systems which may be managed or unmanaged, or may have different photosynthetic pathways. Nevertheless, remote sensing of land “management” categories remains a challenging task since existing classification approaches have yet to demonstrate an ability to capture the whole range of rangelands and crop diversity at global scale.

Updating PFT data sets used in ESMs will clearly lead to improvements in the realism of the patterns of biogeography and have important feedbacks on simulating ecosystem processes and interactions with the atmosphere. Available PFT data sets used in ESMs remain outdated, using land cover information from the 1980s mainly because of a lack of tools available for cross-walking land cover to PFTs. The LC_CCI scheme and tool fills a critical data need for improving the representation of carbon, water and energy cycles being developed by the modeling community; however, extensive model benchmarking and calibration activities may now be necessary before the new PFT data sets result in model improvement. For example, model processes may be calibrated to some extent to produce performance metrics under outdated land cover information, and thus a range of benchmarks should be considered when transitioning to new PFT information.