The annual ESA CCI LC maps cover a period of 24 years from 1992 to 2015 at a spatial resolution of 300 m (ESA, 2017). These maps describe the Earth's terrestrial surface in 37 original LC classes based on the United Nations Land Cover Classification System (UN-LCCS; Di Gregorio, 2005).

This unique long-term land cover time series was achieved by combining the global daily surface reflectance of five different observation systems while aiming to maintain a good consistency over time. This was identified as a key requirement from the modeling community (Bontemps et al., 2012). Each of these global daily measurements of multispectral radiance recorded from 1992 to 2015 have been pre-processed to complete radiometric calibration, geometric and atmospheric correction, and clouds and cloud shadow screening. The full archive of MERIS (2003–2012) providing 15 spectral bands at 300 m resolution was classified to establish a baseline by fusing the outputs of machine learning and unsupervised algorithms (ESA, 2017). The 1 km time series recorded respectively by AVHRR from 1992 to 1999, SPOT-VGT from 1999 to 2013 and PROBA-V from 2014 and 2015 were used to detect and confirm the change which was eventually delineated more precisely at the 300 m spatial resolution whenever possible, i.e. later than 2004. This last step results in both back- and forward-dating the 10 year baseline LC map to produce the 24 annual LC maps from 1992 to 2015. In order to avoid false change detections due to the interannual variability in classifications, each change has to persist over more than 2 successive years in the classification time series to be confirmed (for more information see Sect. 3.1.2 of the ESA CCI LC Product User Guide; ESA, 2017). The resulting series of consistent 300 m annual LC maps from 1992 to 2015 is delivered with a pixel-based uncertainty value indicating the confidence at which a LC class was assigned for each pixel.

The accuracy of ESA CCI LC products was evaluated on a global scale according to international standards, using an independent validation dataset to produce a confusion matrix and derive overall an accuracy figure. An object-based validation database of 2600 primary sampling units was built by a panel of international experts to specifically assess the accuracy of both the LC classes and changes (ESA, 2017). Research is currently ongoing to find how to address the new challenges underlying this database, i.e. following a per-object approach and interpreting not just a unique land cover class but also a distribution of land cover classes within a primary sampling unit. The uniqueness of these two concepts in the framework of global land cover validation requires more time to derive reliable figures about LC classes and LC change accuracy. It will also prevent any comparison with previous validation figures.

In this respect, for the sake of comparison, the accuracy of the ESA CCI LC product from 2010 was assessed using the GlobCover 2009 validation database (Bontemps et al., 2011). Using all the points interpreted as “certain” by the experts, whether “homogeneous” (i.e. made of a single LC class) or “heterogeneous” (i.e. made of several or mosaic LC classes), the overall accuracy was found to be 71.5 %. Accounting for only the “homogeneous” and “certain” points, the overall accuracy raised to 75.4 % (ESA, 2017). The highest user accuracy values were found for the classes of rainfed cropland, irrigated cropland, broadleaved evergreen forest, urban areas, bare areas, water bodies and permanent snow and ice. Conversely, mosaic classes of natural vegetation were associated with the lowest user accuracy values, as well as the three classes of lichens and mosses, sparse vegetation and flooded forest with fresh water.

The overall accuracy of the ESA CCI LC products was also assessed by independent studies over specific regions (e.g. Tsendbazar et al., 2015, over Africa and Yang et al., 2017, over China), which can give valuable insights for specific applications.

The original 37 ESA CCI LC classes were first aggregated into 0.5∘×0.5∘ resolution and then translated into 14 different PFTs based on the cross-walking table (Table S1 in the Supplement) from the ESA Land Cover Product User Guide (ESA, 2017). This table originated from Poulter et al. (2015) and was further adjusted for some classes due to improved understanding of how the LC class descriptions can be interpreted to estimate fractional cover of PFTs from each LC class, in particular for mosaic classes and sparsely vegetated regions. PFTs were grouped into major vegetation types: forest, shrub, grassland and cropland. The tree PFTs and shrub PFTs (Table S1) were summed to obtain the forest and shrub area respectively; thus, the shrub PFTs are excluded from tree PFTs in our analyses. The net area change was calculated by comparing two annual PFT maps at 0.5∘×0.5∘ resolution.

Gross changes need to be considered differently because it is only possible to derive the net change by comparing the annual 300 m maps sequentially. Gross changes may be far larger than the net changes, and thus may show different magnitudes or even directions of LULCC carbon fluxes when simulated in LSMs. To document all the bidirectional LC transitions at 0.5∘×0.5∘ resolution, high-resolution LC transition data are needed. Therefore, the annual ESA CCI LC maps are compared year by year at 300 m resolution to record the gross loss and gain of each original LC class over the whole period from 1992 to 2015. There are 23 original LC classes that experienced gross changes (classes with stars in Table S1).

In order to derive the gross transitions, all possible transitions (506 in total) between the 23 original LC classes with gross changes were calculated at 300 m resolution. There are a total of 422 gross transitions between these 23 original LC classes. These gross changes in the original classes were then translated into gross changes in PFTs using the LC-to-PFT cross-walking table (Table S1) and grouped into the major vegetation types (forest, shrub, grassland, cropland). For example, a LC transition from class “50”, corresponding to 90 % tree PFT in Table S1, to class “30” (10 % tree PFT) is taken as a forest loss of 80 % in that 300 m grid cell. Finally, the converted transitions were aggregated into fractions in each 0.5∘×0.5∘ grid cell.

Three land-use and land-cover datasets (Table 1) were used for comparison: (i) forest, grassland and cropland area from the Land Use Harmonization (LUH2v2h) dataset (Hurtt et al., 2011); (ii) forest cover data from Hansen et al. (2013) and (iii) national forest area data from Houghton and Nassikas (2017). The cropland and pasture areas in the LUH2v2h dataset are from HYDE 3.2 (Klein Goldewijk et al., 2016), in which the ESA CCI epoch LC map in 2010 (representing 2008–2012) was used as a spatial reference map for the area allocation and the national cropland and grazing land were adjusted to match the FAO STAT data (FAOSTAT, 2015) as close as possible. The national forest areas from Houghton and Nassikas (2017) are based on FAO Forest Resources Assessment (FRA) data (FAO, 2015; also see Keenan et al. ,2015, for the main findings of FAO FRA 2015). Thus, these two additional sources of data, HYDE 3.2 (Klein Goldewijk et al., 2016) and FAO FRA (FAO, 2015) were not shown in the figures.

It should be noted that land use data are not necessarily the same as land cover data, and the exact definitions and categorization of forest (cropland and grassland) are different for each dataset (see details in Discussion). Nevertheless, these represent the best datasets available for the use in LSMs for comparison, and we have tried to harmonize the definitions where possible (see below), but to some degree this is an ongoing discussion between the modeling and data communities. Furthermore, all the LSMs have to use these datasets for deriving PFT changes back through time, so it is a very worthwhile exercise to determine if the broad groupings differ, and to what extent.

Absolute areas, net changes and gross transitions from 1992 to 2015 in the LUH2v2h dataset (Hurtt et al., 2011) were used for comparison. Forest used in this study from LUH2v2h (Hurtt et al., 2011) refers to the total of primary and secondary forest, cropland refers to all crop types and grassland refers to the total of pasture and rangeland. Because LUH2v2h data use cropland and grazing land areas from HYDE 3.2 as an input (Hurtt et al., 2011), the spatial distributions are mainly determined by HYDE 3.2. The gross transitions in LUH2v2h data are calculated from the Global Land Use Model (Hurtt et al., 2006) that tracks sub-grid cell loss and gain in land use categories. They first determined the urban area in each grid cell proportionally from cropland, pasture and secondary lands, and if these areas cannot fulfill the urban increase, primary lands were cleared. The minimum transition rates between cropland, pasture and other (sum of primary and secondary) lands were then calculated to identify the gross transitions between these land use categories (Hurtt et al., 2011). Transitions related to shifting cultivation and wood harvest were determined last (Hurtt et al., 2011).

Only annual gross forest loss each year during 2000–2014 and total gross forest gain during 2000–2012 are available in the dataset of Hansen et al. (2013). Thus, the net forest area change from this dataset only refers to the period of 2000–2012. The national forest area data from 1992 to 2015 in the dataset of Houghton and Nassikas (2017) were used to calculate the forest area changes.

A land mask with nine regions (Fig. 1) defined by Houghton (1999) was used to derive the regional values.

After translating the original ESA CCI LC classes into PFTs using the cross-walking table (Table S1), the global and regional areas of forest, cropland and grassland PFTs in the year 2000 are shown in Fig. 1. Global areas of forest (excluding shrub), cropland and grassland PFTs are 30.4, 19.2 and 35.7 million km2, respectively Global forest area is 6.7, 11.2 or 10.1 million km2 lower than that from LUH2v2h (Hurtt et al., 2011), Hansen et al. (2013) or Houghton and Nassikas (2017), respectively. It is also much lower than the recently reported global forest area of 43.3 million km2 with increased forest area estimate in dryland biomes using Google Earth images (Bastin et al., 2017). Global cropland area from ESA CCI is 4.2 million km2 larger than that from LUH2v2h, while the difference in global grassland area is relatively small.

Forest area from ESA CCI is lower than that from Hansen et al. (2013) in all regions except North Africa and Middle East and the Pacific developed region. Here, the regions refer to the defined regions in Fig. 1. Forest area from LUH2v2h (Hurtt et al., 2011) is larger than that from ESA CCI in most regions except in South and Central America, tropical African and the Pacific developed region. Forest area from Houghton and Nassikas (2017), however, is systematically higher than that from ESA CCI in all regions. Cropland area from ESA CCI matches that from LUH2v2h (Hurtt et al., 2011) in North America but is higher in all the other regions. Although the global grassland area is similar between ESA CCI and LUH2v2h (Hurtt et al., 2011), larger differences are seen on a regional scale. Grassland area from ESA CCI was found to be much higher than that from LUH2v2h (Hurtt et al., 2011) in North America and the former Soviet Union (4.0 and 3.5 million km2 higher, respectively) but much lower (2.4 million km2) in North Africa and Middle East.

The forest, cropland and grassland areas from different datasets do not match on global or regional scales (Fig. 1), mainly caused by the differences in land cover definitions and data sources (Table 1), as well as the uncertainties in the cross-walking table used for translating original ESA CCI LC classes into PFTs. The canopy cover of forest varies in different ESA CCI LC classes with defined ranges such as >15 %, 15–40 % and >40 % depending on the “openness” of the canopy and according to the UN-LCCS framework provided by the FAO (Di Gregorio, 2005). Although continuous tree cover fractions are provided in data from Hansen et al. (2013), the forest cover is defined as >25 % canopy closure for trees higher than 5 m (Hansen et al., 2010). Forest areas in Hansen et al. (2013) are obtained from NASA's Landsat instruments with a spatial resolution of 30 m that can capture the small-scale forest areas. This partly explains the larger forest extent in Hansen et al. (2013) than ESA CCI PFT maps. The definition of forest by FAO, which is the data source of Houghton and Nassikas (2017), is a canopy cover >10 %. FAO's forest areas are based on reports from the member countries (FAO, 2015) and the methods of compiling data in each country may vary largely, e.g. from field survey or from satellite-imagery-based estimation (Grainger, 2008; Harris et al., 2012). Furthermore, in the definition of forest by FAO, natural disturbance suppressing forests do not change the land remaining a forest, but from satellite, they are not detected as forest cover. Keenan et al. (2015) also compared the forest area from FAO FRA 2015 with remote sensing data and attributed their differences to five factors, the major one of which is the different definitions of “forest”. It seems that forest area from ESA CCI PFTs is lower than that from Hansen et al. (2013) in humid regions (Fig. S2). Additionally, in the drylands like tropical Africa, it is difficult to map and estimate forest area using medium- (e.g. Landsat; Hansen et al., 2013) or coarse-resolution satellite data (e.g. ESA CCI LC; Achard et al., 2014; Bastin et al., 2017; Gross et al., 2017), in particular when tree cover is below 30 % (Achard et al., 2014). Bastin et al. (2017) recently reported a forest estimate in drylands using very high spatial resolution satellite imagery, which is 40–47 % more than previous forest assessments. The difficulty of detecting forest in these sparse tree cover regions could partly be responsible for the lower forest area from ESA CCI PFT maps than those from Hansen et al. (2013) and Houghton and Nassikas (2017) in tropical Africa (Fig. 1).

Forest area estimates in LUH2v2h (Hurtt et al., 2011) are based on aboveground biomass density from the Miami-LU ecosystem model (Hurtt et al., 2006), and cropland and pasture areas are based on HYDE 3.2 (Klein Goldewijk et al., 2016). HYDE 3.2 uses the cropland and pasture areas from FAO STAT (FAOSTAT, 2015) as the main land-use input data and the ESA CCI epoch LC map of 2010 as a spatial reference map (Klein Goldewijk et al., 2016). Thus, the grasslands in LUH2v2h refer to the sum of intensively managed pastures and less intensively used rangelands (Klein Goldewijk et al., 2016), while the grassland PFT from ESA CCI maps also includes natural grassland, which may be the reasons for less grassland in LUH2v2h (Hurtt et al., 2011) than ESA CCI, especially in the former Soviet Union, western Europe and North America (Fig. 1).

Similarly, the underestimate in cropland area in ESA CCI maps (Fig. 1) is likely due to differences in definitions of what constitutes a cropland based on remote sensing datasets used to derive the ESA maps vs. land use statistics and country-dependent reporting used to derive FAO statistics that are used to define croplands in HYDE 3.2 (Klein Goldewijk et al., 2016), in addition to differences in spatial resolution. For example, the attribution of oil palm plantations is an important factor for the differences in area changes between different datasets, especially in Indonesia. Oil palm is taken as cropland rather than forest in the FAO definitions (FAOSTAT, 2015) but detected as tree covers from the remote sensing (Carlson et al., 2012, 2013; Hansen et al., 2013; Koh et al., 2011; Tropek et al., 2014), including in the CCI LC products. This partly explains that the larger cropland increase in LUH2v2h (Hurtt et al., 2011) and larger forest decrease in Houghton and Nassikas (2017) than those in ESA CCI PFTs and Hansen et al. (2013) in Indonesia (Fig. 4). Furthermore, the classification of cropland in ESA CCI is also based on remote sensing temporal analysis. In the ESA CCI algorithm, for example, spectral features at key moments during the year were used to optimize the discriminations between all major crop classes: differentiating between cropland and natural vegetation (typically harvesting dates). Cropland in LUH2v2h that is essentially from FAO statistics (Klein Goldewijk et al., 2016), on the other hand, depends on country reporting and therefore comprises different definitions and data sources.

Pérez-Hoyos et al. (2017) provide an extensive comparison of multiple cropland datasets, including ESA CCI epoch and annual maps, for the purposes of cropland monitoring, and they found that the ESA CCI 2015 annual map is more suitable for cropland monitoring than the epoch map because of the reduction in cropland area over the Congo basin. They also showed that spatial resolution is a key driver of product suitability for agriculture monitoring (Pérez-Hoyos et al., 2017). However, the specific framework of their study is the suitability for agriculture monitoring for early warning and they focus on a limited number of countries selected to be “with high risk of food insecurity” (Pérez-Hoyos et al., 2017). The issues of cropland area from the ESA CCI LC maps discussed in their study is fully justified for a study addressing the challenge of cropland monitoring, but it does not allow generalizing the conclusion to all domains (e.g. to derive PFT maps for use in global land surface modeling in this study). As Pérez-Hoyos et al. (2017) and our study show, the agreement (or lack thereof) is country-dependent, further implying that more consistent definitions of LC classes are required and/or regional LC satellite mapping algorithms (or cross-walking table, see below) are needed. Cropland mapping issues, including those discussed in Pérez-Hoyos et al. (2017) are being addressed in upcoming versions of the ESA CCI maps. Additionally, Waldner et al. (2016) have produced a product that aims to combine the “fittest” LC maps on the country level into a unified 250 m resolution cropland product, but again this is dependent upon a specific definition (the JECAM, Joint Experiment of Crop Assessment and Monitoring, cropland definition for the purposes of cropland monitoring).

The final spatial area of each PFT in this study is derived from a combination of the ESA LC map and the cross-walking table (Table S1) used for translating original ESA LC classes into PFTs. The range in tree cover canopy openness (as discussed above) and percent of each type of vegetation for mosaic LC classes in the LC description contributes to uncertainty in the conversion fractions used to translate the LC classes into PFTs in the cross-walking table. Thus, uncertainties in the cross-walking table contribute to the differences in forest, cropland and grassland PFT areas when comparing with other datasets. Only one value is used to prescribe the fraction of each PFT for a given class, e.g. class “50” corresponds to 90 % of broadleaf evergreen trees in Table S1. This hinders an explicit representation of spatially heterogeneous tree cover fractions. In the absence of other information, the approximate mid-point of the range in the LC class description is used when calculating the fraction of forest PFT from a given LC class. For example, class “61” represents a closed canopy (>40 %) and therefore we use a LC-to-tree PFT conversion fraction of 70 % (0.7) as the mid-point between 40 and 100 % (Table S1). Class “62”, on the other hand, is an open canopy (15–40 % cover) and therefore we use a LC-to-tree PFT conversion factor of 30 %. Some exceptions to this general rule are made when we have a better understanding of the species or biomes included in a given LC class. For example, class “50” (broadleaved evergreen trees) encompasses tropical rainforests. Although the class description states that the canopies in this class can be closed to open (>15 %), we know that the tree cover fraction is much higher than a mid-point of ∼60 %, therefore we use a conversion factor of 90 %. However, this level of knowledge is not available for all LC classes. This is particularly true for mosaic and sparse vegetation classes (e.g. classes “100”, “110” and “150”) that span different regions/biomes that may contain different fractional coverages of vegetation.

Likewise, an explicit regional classification is required for cropland. For example, class “10” (cropland, rainfed) is separated well in North America, i.e. mainly partitioning into class “11” (herbaceous cover), and thus the cropland area in this region is highly consistent with LUH2v2h data (Hurtt et al., 2011; Fig. 1). In tropical Africa where class “10” is not separated into a more detailed classification, the difference in cropland areas between these two datasets are large (Fig. 1). This is because if most of the cropland in this region belongs to class “12”, using the corresponding value for class “10” in the cross-walking table (90 % for class “10” vs. 30 % for class “12”, Table S1) overestimates cropland areas.

Hartley et al. (2017) also investigated the uncertainty in simulations of carbon, water and energy fluxes from three LSMs as a result of cross-walking table uncertainty. This study found that the spread in model outputs due to cross-walking uncertainty was higher than uncertainty due to the underlying LC maps (mapping algorithm; Hartley et al., 2017). Despite these uncertainties, satellites provide the only plausible way to derive the global maps of vegetation distribution needed to drive LSMs and validate dynamic global vegetation models. Future efforts by the ESA CCI LC project and collaborators will focus on reducing the uncertainty introduced when translating from LC to PFT, including using optimized and regionally based cross-walking tables.

The ESA CCI LC magnitudes of gross changes for all PFTs are lower than those of all three products considered. This is explained by the effect of spatial resolution combined with a change consolidation approach. Using Earth observation time series of 1 km spatial resolution to annually detect the land cover change for the ESA CCI maps does not allow capturing small-scale LC changes, which is part of the reason for smaller gross and net forest changes than those in Hansen et al. (2013). On the other hand, this is the only way to have a consistent method of LC change detection over the whole period. In spite of the consolidation strategy, confirming the change over several years, ESA CCI LC trends of area change mitigate only partly the impact of the heterogeneous quality of the data acquired by the various sensors. For instance, larger change variations for forest and cropland in the 1990s result from poorer radiometric and spectral quality of the AVHRR input data. This instrument, first designed for meteorological observation, is, however, the only one recording the land surface systematically before 1999.

The large magnitude of gross changes in forest and cropland in LUH2v2h (Hurtt et al., 2011; Fig. 2) mostly distributes in the tropical regions (Figs. S3 and S4) where gross changes reflect shifting agriculture (Heinimann et al., 2017). The gross gain and loss of forest (or cropland) in the tropics from LUH2v2h maintains a similar constant rate with other small variations (Figs. S3 and S4). This is because the gross changes in LUH2v2h are mainly generated from the shifting cultivation in the tropics by assuming a turnover rate of 6.7 %year-1 (i.e. a residence time of 15 year) of all agricultural lands (Hurtt et al., 2011) and based on a spatial distribution map from Butler (1980). The Butler (1980) map is a hand-drawn map indicating the presence or absence (no precise area or fraction) of both shifting cultivation and other non-shifting farming systems based on some regional studies and “general knowledge” (Heinimann et al., 2017). The estimate of shifting cultivation extent from LUHv2h (Hurtt et al., 2011) is thus highly uncertain because of the simple assumptions and the old reference map (representing 1960s–1970s) but strongly affects the gross land use change areas. Heinimann et al. (2017) recently estimated the global extent of shifting cultivation visually using Landsat 30 m forest data and very-high-resolution satellite imagery from Bing and Google. They found that shifting cultivation area decreases over the last 40 to 50 years, in particular in Southeast Asia (Heinimann et al., 2017). This, however, is not reflected in the LUH2v2h dataset (Figs. S3 and S4). From LUH1 to LUH2v2h, The area of shifting agriculture is reduced (see an example in tropical Africa; Fig. S5) because of the separation of forest from natural vegetation in LUH2 (Hurtt et al., 2011). However, the gross forest changes in LUH2v2h (Hurtt et al., 2011) are still much higher than those in ESA CCI PFTs and Hansen et al. (2013). Especially in the ESA 300 m resolution data, the gross change area seems very small (Figs. S3 to S5). Therefore, the shifting cultivation area in LUH2v2h may be overestimated due to (1) the binary (presence/absence) indication rather than a precise extent of shifting cultivation in the Butler (1980) map and (2) no temporal change (missing the decreasing trend) of the reference map. Still, it should be noted that the coarse spatial resolution of ESA CCI products cannot detect small-scale LC changes, resulting in an underestimation of gross changes. The shifting cultivation today remains extensive and is very important for land carbon modeling, but there are only very limited studies on the regional or national extent estimates (Heinimann et al., 2017). More research and developments in the mapping and change detection of shifting cultivation are urgently desired.

The discrepancies in temporal PFT net area changes between ESA CCI maps and FAO data (cropland and pasture area changes in LUH2v2h, Hurtt et al., 2011, and forest area changes in Houghton and Nassikas, 2017; Figs. 4–6) are mainly caused by the different approaches for estimating LC change used by different countries in FAO reports (FAO, 2015; FAOSTAT, 2015). Some countries like Canada distinguish land use and land cover when compiling forest statistics. For example, a forest cleared for wood harvest is not taken as a forest loss because new secondary forest will be planted on this land, thus no change in land use (Keenan et al., 2015). However, remote sensing can easily detect such land cover change and treat it as forest loss. Cropland and pasture in HYDE 3.2 (Klein Goldewijk et al., 2016) adopted the FAO categories for “Arable land and permanent crops” and “Permanent meadows and pastures”, respectively, as the main data source. In the ESA CCI LC maps, pastures are mapped as grassland and translated into 100 % “Natural Grass” PFT (Table S1). The different trajectories of temporal cropland changes between ESA CCI and LUH2v2h (the former shows increasing from 1992 to 2004 while the latter increases after 2007; Fig. 4) are probably caused by the time lag between the real changes and country reporting to FAO. Finally, the trends of cropland area change from FAO STAT data may contradict those from national statistics (Li et al., 2016), e.g. comparing FAO STAT data (FAOSTAT, 2015) with USDA estimates (Nickerson et al., 2011) for the United States or with NBSC estimates (NBSC, 2015) for China.

There are also many other land cover and land use datasets that can be used for comparisons to assess the accuracy of land cover or land cover change in ESA CCI LC products. However, they are either regional maps (e.g. the maps for Europe from Fuchs et al., 2015) or global epoch maps (e.g. the GlobeLand30 maps for 2000 and 2010; Chen et al., 2014) and not suitable for the application in LSMs. Thus, we did not include them in this study. In fact, there have already been studies on the detailed comparisons of different datasets in a region (e.g. Fuchs et al., 2015, for Europe; Yang et al., 2017, for China; and Achard et al., 2014, for the tropics). In addition to the accuracy assessments conducted in the ESA CCI project (ESA, 2017), a systematic comparison with all other land cover datasets in future will help to validate the land cover classification and land cover change detection in the ESA CCI LC products. Instead of using a single dataset, combining a sample of several datasets is reported to be considerably more efficient and accurate to estimate land cover area and change (Olofsson et al., 2014; Sannier et al., 2016) and has been adopted as technical guidelines (GOFC-GOLD, 2016; GFOI, 2016) in the remote sensing community, especially for forest monitoring to reduce emissions from deforestation and forest degradation in developing countries (the REDD+ programme).