Estimating changes in global vegetation cover (1850–2100) for use in climate models



[1] Historical changes in global cropland area based on estimates of Ramankutty and Foley (1999), and projections of future changes under IPCC SRES development scenarios taken from the IMAGE 2 model, were combined with a simple classification of present-day satellite data. These data were used to estimate annual changes in area fractions occupied by primary plant functional types (PFTs) between 1850 and 2100 using two different approaches. The linear interpolation approach assumed that natural vegetation area varies in inverse proportion to cropland area. The rule-based approach added simple transition rules to define the sequence by which natural PFTs are converted to agriculture (e.g., grassland before forest) and by which abandoned cropland reverts to natural vegetation. In both approaches, constraints were imposed to ensure the simulated PFT composition was consistent with available information. The resulting time series data can be used in coupled biosphere-atmosphere models, and in uncoupled global climate models, to represent time-varying land cover.

1. Introduction

[2] Human activities have altered 30–50% of natural resources [Vitousek et al., 1997] and have transformed global vegetation distribution through land use change, where human pressures or management have caused persistent alterations in land cover (e.g., from forest to agriculture). Anthropogenic land use changes contrast with natural disturbances (e.g., fires and pests) where the existing natural vegetation composition is generally unchanged and recovers following the disturbance. Changes in land cover have significant effects on the global carbon cycle and on the global climate system [e.g., Price and Apps, 1996a, 1996b; Houghton et al., 2000; Claussen et al., 2001; McGuire et al., 2001; Schimel et al., 2001; Brovkin et al., 2004]. Data of Ramankutty and Foley [1999] indicate that about 13% of Earth's land surface had been converted to permanent crop agriculture by 1992. Land use change is likely to continue rapidly in the near future according to most projections for the global economy and population growth [Tilman et al., 2001].

[3] Although there is considerable uncertainty in the terms of the global carbon budget, research over the last 15 years has shown that land use change has contributed significantly to observed increases in atmospheric carbon content, reaching perhaps 2.0 Pg C yr−1 during the 1990s [e.g., Houghton, 2003]. At the same time, terrestrial ecosystems evidently take up approximately 30% of global anthropogenic carbon emissions [Prentice et al., 2001], though the cause of this terrestrial C sink is highly uncertain [Intergovernmental Panel on Climate Change, 2000a]. Caspersen et al. [2000] suggested that increased terrestrial carbon uptake in the last 2 decades was due primarily to forest regrowth (e.g., on abandoned agricultural land) rather than to growth enhancements resulting from climatic feedbacks, possible CO2 fertilization or other environmental changes, and concluded that ecosystem models focusing exclusively on physiological processes omit the dominant factor, namely land use change history. Subsequent experiments with global ecosystem models appear to confirm the importance of land use change in the terrestrial carbon balance [e.g., McGuire et al., 2001; Hurtt et al., 2003].

[4] Land use change also affects several physical characteristics of the land surface, including albedo, roughness length, rooting depth and leaf area index. These changes lead to alterations in the partitioning of net radiation into latent and sensible heat fluxes and of precipitation into runoff and evapotranspiration, thereby providing forcing mechanisms that may lead to significant changes in regional and global climate [Charney, 1975; Bonan et al., 1992; Lean and Rowntree, 1993; Costa and Foley, 1997; Betts, 2000; Chase et al., 2000; Douville et al., 2000; Heck et al., 2001; Negri et al., 2004].

[5] Enhanced understanding of the interactions between global climate change and the carbon cycle requires spatially explicit assessments of the dynamics of human-caused land use change [Verburg et al., 2003], as well as of natural terrestrial ecosystems. The importance of vegetation cover change for regional as well as global scale terrestrial carbon budgets and climate is now well accepted, leading to recent attempts to represent terrestrial ecosystem processes incorporating human-caused land use change explicitly or implicitly in global climate simulations [e.g., Cox et al., 2000; Friedlingstein et al., 2003; Karoly et al., 2003; Brovkin et al., 2004; Betts et al., 2006; Sitch et al., 2005; Tett et al., 2006; Stott et al., 2006; Feddema et al., 2005]. In particular, Betts et al. [2006] incorporated historical land cover change into the UK Hadley Centre coupled climate model and found that the dominant effect of land cover change on climate may be a cooling resulting from deforestation in the temperate regions causing increased surface albedo, rather than from reduced forest carbon sinks inducing warming. Conversely, deforestation in tropical regions could result in a warming effect due to the combination of reduced carbon sequestration, reduced evapotranspiration from forests and reduced cloud cover. According to Betts et al. [2006], the influence on global climate resulting from land use change is relatively small compared to the anthropogenic CO2 forcing, but comparable to effects attributed to other greenhouse gases and aerosols. Nevertheless, it is clear that changes in land use can result in significant regional climate alteration, consistent with other observations, notably those of Feddema et al. [2005].

[6] The work reported here was carried out within the Canadian Global Coupled Climate Carbon Modeling (CGC3M) Research Network to address the need for a global data set of land use change suitable for coupled climate model simulations. The data product is to be used initially in the Canadian Terrestrial Ecosystem Model (CTEM) [Arora, 2003; Arora and Boer, 2005] coupled to the Canadian Centre for Climate Modeling and Analysis (CCCma) general circulation model (CGCM). CTEM is a dynamic vegetation model that includes process-based representations of photosynthesis, autotrophic and heterotrophic respiration, phenology, allocation, mortality, fire and land use change. Version 1 of CTEM does not model competition among PFTs, however, so the time-varying fractional coverages of PFTs will instead be prescribed using the data sets presented here.

[7] Changes in global area fractions of CTEM PFTs (i.e., the fractions of each 0.5° pixel's area occupied by the given PFT) were estimated as an annual time series for the period 1850–2100. The objective was to develop a method of simulating historical changes in the distribution of natural vegetation, reclassified into area fractions of CTEM PFTs, based on present-day satellite maps, combined with Ramankutty and Foley's [1999] historical time series of global cropland cover (here after referred to as “R&F”) to estimate changes during 1850–1992. Projections for changes in the period up to 2100 were obtained from the IMAGE 2 model of Alcamo et al. [1998].

2. Data

[8] The availability of historical global land cover data from satellite imagery is limited to a relatively recent period [Anderson et al., 1976], and some early satellite-based vegetation classifications were questionable. More recently, improved ground-based validation has enabled researchers to generate several credible classifications of present-day global land cover, including the International Geosphere–Biosphere Program (IGBP) Data and Information System (DIS) DISCover database [Belward et al., 1999; Loveland et al., 2000] and the University of Maryland (UMD) data set of Hansen et al. [Hansen et al., 2000; Hansen and Reed, 2000; Friedl et al., 2002]. One of the IGBP DISCover data sets is the Seasonal Land Cover Regional (SLCR) data set which provides detailed classifications of land cover in 1992 for each of five continental regions. In 2003, the Global Vegetation Monitoring Unit (GVMU) of the Joint Research Centre of the European Commission (EC/JRC) released the Global Land Cover 2000 (GLC2000) map. Compared to earlier mapping approaches, the GLC2000 adopted 19 regional windows, each of which was classified by local experts so that accuracy could be improved on the basis of local knowledge [e.g., see Latifovic et al., 2004].

[9] We adopted the GLC2000 (available at as our primary basis for classifying present-day global land cover characteristics into the CTEM PFTs listed in Table 1. Particular attractions of this data set were that it was the most recently observed, with only 22 classes, of which five compared directly to the tree/woody PFTs identified for CTEM, and another three could be grouped as “bare ground” (see Table 2). The classified forest PFTs and bare ground implied that all remaining land area should be allocated as crops and grasslands, which also were simply adjusted according to information obtained from reclassification of SLCR (Version 2) data for 1992 (see section 2.1).

Table 1. Plant Functional Types Identified for CTEM Version 1.0.
 Plant Functional TypeSymbol
PFT1needleleaf evergreen treeNE
PFT2needleleaf deciduous treeND
PFT3broadleaf evergreen treeBE
PFT4broadleaf cold deciduous treeBCD
PFT5broadleaf drought deciduous treeBDD
PFT6C3 cropC3C
PFT7C4 cropC4C
PFT8C3 grassC3G
PFT9C4 grassC4G
Table 2. Reclassification of GLC2000 Data Set and Ramankutty and Foley's [1999] Potential Vegetation Classes to CTEM Plant Functional Typesa
1234 + 56 + 78 + 91011
  • a

    GLC2000 data set available at Each vegetation class was assigned a mixture of CTEM PFT area fractions based on the class description and knowledge of global biomes. CTEM broadleaved deciduous trees were subsequently split into cold-deciduous (PFT 4) and drought-deciduous (PFT 5). Similarly, croplands (PFTs 6 and 7) and grasslands (PFTs 8 and 9) were later each split into C3- and C4-dominated types.

GLC2000 Class
1,tree cover, broadleaved, evergreen  1.00     
2,tree cover, broadleaved, deciduous, closed   1.00    
3,tree cover, broadleaved, deciduous, open   0.60
4,tree cover, needle-leaved, evergreen1.00       
5,tree cover, needle-leaved, deciduous 0.80   0.100.10 
6,tree cover, mixed leaf type0.40  0.50 0.10  
7,tree cover, regularly flooded, fresh water  0.50    0.50
8,tree cover, regularly flooded, saline water  0.50    0.50
9,mosaic: tree cover/other natural vegetation   0.60 0.200.20 
10,tree cover, burnt0.20  0.20 0.300.30 
11,shrub cover, closed-open, evergreen  0.60
12,shrub cover, closed-open, deciduous   0.40 0.300.30 
13,herbaceous cover, closed-open     0.700.30 
14,sparse herbaceous or sparse shrub cover   0.10 0.100.80 
15,regularly flooded shrub and/or herbaceous cover  0.50  0.300.100.10
16,cultivated and managed areas    0.500.400.10 
17,mosaic: cropland/tree cover/other natural veg  0.20 0.500.200.10 
18,mosaic: cropland/shrub and/or grass cover   0.100.500.300.10 
19,bare areas      1.00 
20,water bodies       1.00
21,snow and ice      1.00 
22,artificial surfaces and associated areas      1.00 
Ramankutty and Foley's Potential Vegetation Class
1,tropical evergreen forest/woodland  0.900.05 0.05  
2,tropical deciduous forest/woodland  0.050.90 0.05  
3,temperate broadleaf evergreen forest/woodland  0.800.10 0.10  
4,temperate needleleaf evergreen forest/woodland0.80  0.10 0.10  
5,temperate deciduous forest/woodland0.10  0.80 0.10  
6,boreal evergreen forest/woodland0.800.10 0.10    
7,boreal deciduous forest/woodland0.100.10 0.80    
8,mixed forest0.40 0.100.40 0.10  
9,savanna   0.35 0.400.25 
10,grassland/steppe     0.650.35 
11,dense shrubland  0.300.50 0.20  
12,open shrubland   0.25 0.350.40 
13,tundra0.050.05 0.10 0.450.250.10
14,desert      1.00 
15,polar desert/rock/ice      1.00 

[10] Several approaches to creating global maps of land cover/land use change at global and regional scale from a combination of remote sensing data and inventory statistics have emerged [e.g., Ramankutty and Foley, 1998, 1999; Klein Goldewijk, 2001; Hurtt et al., 2001]. Of these researchers, R&F reconstructed global 0.5-degree resolution data on permanent cropland areas for 1700–1992 on the basis of satellite observation, a land cover change model and census data, while Klein Goldewijk produced a time series of cropland and grazing lands at similar spatial resolution (HYDE), but instead simply reported presence/absence of crops. The R&F and HYDE data sets are not independent, being based to some extent on the same historical records. Major differences include (1) the manner in which cropland areas are allocated: the HYDE data set is based on historical population densities, while R&F uses crop area data; (2) the HYDE classification explicitly includes grazing land, while R&F does not; and (3) the R&F cropland database is an annual time series, while the HYDE temporal resolution is only 50 years for the period 1700–1950 and 20 years for 1950–1990. Pixel-wise comparison of the two data sets for the cropland areas in key years shows worsening agreement with increasing time into the past, that becomes rather poor by 1850 [Klein Goldewijk and Ramankutty, 2004]. Some of this divergence can be explained as an overlap in the two classifications: pixels classified as predominantly cropland by R&F are sometimes reported to be dominated by pastureland in the HYDE data set. As we had no further evidence to determine which data set gave better estimates of cropland areas in the past, we chose to use R&F because of its higher time resolution and because it reported areas in fractional terms. Then we used 1992 as a reference year to cross-reference the GLC2000 data sets with the final year in the R&F historical cropland area time series (see section 2.1).

[11] For 1992, we assumed that all pasture and grazing land were included in the grassland areas identified from the present-day satellite classification, because the satellite imagery does not discriminate between natural grasslands and managed pasture. Moreover, with the simplistic representation of grass physiology used in CTEM (and probably most comparable global vegetation models), it is not practical to identify separate PFTs for natural and managed grasslands. Klein Goldewijk and Ramankutty [2004] also indicated that the historical expansion of pastureland was mostly due to conversion of natural grassland. For these reasons, we assumed that all grassland in 1850 was unmanaged. This point will be further addressed in section 5.

2.1. Reclassification for the Reference Years 1850 and 1992

[12] Figure 1 summarizes the steps followed to estimate the area fractions of CTEM PFTs in each 0.5° pixel of the global maps. The reclassifications for GLC2000, SLCR and potential natural vegetation were carried out independently at 1 km, 1 km and 5 min resolutions, respectively, with each then aggregated to 0.5 degrees. Of the 22 land cover classes identified in the GLC2000 data set, five forest classes (needle-leaved evergreen, needle-leaved deciduous, broadleaved evergreen and broadleaved deciduous open and closed) correspond directly to CTEM's tree PFTs (Table 1). In addition, three other GLC2000 classes (snow and ice, bare areas, and urban) could conveniently be considered as bare ground. Table 2 shows the area fractions of each CTEM PFT assigned to each GLC2000 class, noting that the distinctions between C3 and C4 pathways and between cold- and drought-deciduous trees were not made at this step. Reclassifications for the SLCR and the natural potential vegetation databases were carried out in a manner similar to those for GLC2000 (Table 2).

Figure 1.

Flow diagram illustrating the steps for estimating the time series of changes in area fractions of CTEM plant functional types (PFTs).

[13] There are some differences in the reported areas of inland and coastal water between the R&F map of potential natural vegetation (as estimated using BIOME3 for 1700) and the SLCR 2.2 and GLC2000 data sets representing the late 20th century. To overcome this discrepancy, all grid cells of 1850 data were assumed to have the same water fractions as those found in the 1992 data.

[14] Data from the GLC2000 and SLCR data sets were then merged to represent global land cover in 1992, on the basis of the assumption that the forested areas did not change greatly between 1992 and 2000. This assumption was made because: (1) the IMAGE 2 scenarios of future land-use change do not indicate much change in cropland areas between 1992 and 2000 (as seen in Figures 7 and 8 in section 4); and (2) more recent data reported by Hirsch et al. [2004] show that rates of tropical forest clearance and regrowth were in near-equilibrium over this period. This assumption will be discussed further in section 5.

[15] Combining the different data sets required several steps. First, grid cells in 1992 were adjusted to have the same fractions of trees, bare ground and inland water as the GLC2000 data set and to have the same cropland and grassland fractions as SLCR. If the total of those PFTs exceeded 100% (of grid cell area), then the grassland area was reduced first. If trees plus cropland area exceeded 100%, the grassland area was set to zero and the cropland fraction reduced. Conversely, if the total PFT area was less than 100%, then the shortfall was made up by increasing the grassland fraction. For the second step, the grid cell areas for 1992 were further adjusted so that cropland area agreed with R&F 1992 while the area fractions of natural PFTs (i.e., forest and grassland) remained proportional to those determined in step 1.

[16] We combined the 1992 PFT maps with Ramankutty and Foley's [1999] global map of potential natural vegetation (PNV) in 1700 to create a global land cover map for 1850 (referred to hence forward as “the 1850 data”). Ramankutty and Foley derived their PNV map mainly from the DISCover (SLCR 1.2) data set, but grid cells with more than 50% crop cover or less than 20% of dominant potential vegetation in 1992 were replaced with simulated natural vegetation data from Haxeltine and Prentice [1996]. The map has 15 classes which were classified similarly to the GLC2000 data (see Table 2), and then used to assign natural PFT area fractions to 0.5-degree grid cells where cropland area evidently exceeded 50% at some time between 1850 and 1992. It should be noted that our 1850 data set makes the implicit assumption that there were no systematic changes in the composition of natural vegetation in any grid cell before 1992, for example, due to climate change or increasing CO2 concentration.

[17] A major problem was how to assign area fractions of natural PFTs in 1850 to pixels that had clearly been subject to major land-use change since then. A practical method of resolving this was to track the changes in R&F's cropland area data between 1850 and 1992 for each half-degree grid cell. If the maximum cropland fraction exceeded 50% at some time between 1850 and 1992, then the natural PFT area fractions in 1850 were based on those reclassified from R&F's map of potential natural vegetation. These were combined with R&F cropland areas reported for 1850, but adjusted proportionately to conserve area. For example, if a grid cell reported to contain 100% cropland in 1992 contained 50% needleleaf evergreen (NE), 30% broadleaf deciduous (BD) and 20% grassland in the potential vegetation map, but 20% cropland in 1850 according to R&F, then the PFT fractions for 1850 were set to 20% cropland, 40% NE, 24% BD and 16% grassland. Conversely, if the maximum cropland fraction during 1850–1992 never exceeded 50% of the vegetated area, then we assumed that the proportions of natural PFTs reported for 1992 were unchanged from those in 1850. For example, a grid cell estimated to contain 20% cropland in 1850, increasing to a maximum of 45% in 1903, with 40% cropland, 30% BD and 30% grassland in 1992, was estimated to contain 40% BD and 40% grassland in 1850.

2.2. Separation of C3/C4 and Cold/Drought-Deciduous Vegetation

[18] To separate the cropland and grassland area fractions into C3 and C4 components, we used the modeled global fractional distribution of C4 vegetation developed by Still and Berry [2003]. This 1° data set was produced from continuous fields of vegetation characteristics [DeFries et al., 2000], the R&F cropland fractions, and harvest area estimates from the US Department of Agriculture and the UN Food and Agriculture Organization (FAO), as well as global climate data. Where Still and Berry [2003] found contradictions between R&F and the remotely sensed continuous vegetation fields, they used the latter data so their cropland area estimates are not actually consistent with our 1992 data. The discrepancies were generally small, however, and in any case, we used only Still and Berry's C4 fractions rather than their absolute area estimates.

[19] For each 0.5° pixel where the herbaceous fraction of the total natural vegetation area was greater than Still and Berry's C4 vegetation fraction, the C4 cropland and grassland fractions were estimated by multiplying the C4 vegetation fraction by the total cropland area and total grassland area for that pixel, respectively. Conversely, if the herbaceous fraction was less than (or equal to) Still and Berry's estimate of the C4 fraction, then all cropland and grassland was assumed to be C4, and the C3 cropland and grassland fractions were set to zero. In all pixels where Still and Berry [2003] predicted no C4 vegetation, all vegetation was assumed to be C3 only.

[20] The broadleaf deciduous tree cover obtained from GLC2000 was split into two PFTs: broadleaf cold deciduous (BCD) and broadleaf drought deciduous (BDD). At latitudes above 34°, the broadleaf deciduous habit was assumed to be defined entirely by temperature (“cold-deciduous”) and at latitudes below 24°, entirely by drought (“drought-deciduous”). A linear transition between BCD and BDD was assumed within the two 24° to 34° latitudinal ranges. This empirical approach is supported by observations, which indicate that the primary leaf phenological control gradually changes from soil moisture in the tropics to temperature in the temperate regions [Arora and Boer, 2005].

2.3. Reclassification of IMAGE 2 Data

[21] For projections of future changes in global land use (up to 2100), simulation results for 1970–2100 were obtained from the Integrated Model to Assess the Global Environment (IMAGE 2) [Alcamo et al., 1998]. The IMAGE 2 model uses the Intergovernmental Panel on Climate Change (IPCC) SRES emission scenarios [Intergovernmental Panel on Climate Change, 2000b] to compute geographically explicit changes in global land cover by taking into account regional needs for agricultural land and other land use pressures resulting from regional economic and population growth [Alcamo et al., 1998]. It should be noted that changes in natural vegetation due to climate were simulated using the BIOME equilibrium projection model of Prentice et al. [1992], and not by a dynamic vegetation model. Changes in land cover projected for the A1B, A2, and B1 SRES scenarios from the IMAGE 2 model were used in this study.

[22] Differences between the IMAGE 2 vegetation classes and the PFT maps obtained from remote sensing sources (GLC2000 and SLCR 2.0) were reconciled using a simple “moving window” algorithm. The IMAGE 2 data for each 10-year period were overlaid on the PFT maps created for 1990. Working through each target pixel in turn, the algorithm first identifies the 10 closest pixels with the same IMAGE 2 classification in 1990. The area fractions of each PFT for these ten pixels estimated for 1990 were then averaged, and these average fractions applied to the target pixel. After all the IMAGE 2 maps were reclassified in this way, the linear interpolation approach was used to interpolate annual changes in PFT area fractions for the period 1993 to 2100.

3. Interpolation Approaches

[23] To calculate the annual time series for the period 1850–1992, the 1850 and 1992 PFT data sets were used as boundary conditions and the R&F cropland time series was used to model changes in the remaining fractions of eight “natural” PFTs (five forest and two grassland types, plus “bare”). Projections of land use change for 1993–2100 were based on the simulations from the IMAGE 2 model with results provided at 10-year intervals. Two approaches are used: the Linear Interpolation Cover Change Scheme (LICCS) and the Rule based Interpolation Cover Change Scheme (RICCS).

3.1. Linear Interpolation Cover Change Scheme (LICCS)

[24] The linear interpolation scheme assumes that the conversion of natural PFTs can be estimated by simple proportional adjustments in the area fractions of each of the natural PFTs between the starting and ending reference years. The area fraction occupied by any PFT is assumed to change linearly, as “natural” (noncrop) PFTs are replaced by croplands. Further, the total area of the natural PFTs, with the exception of water surfaces, is assumed to vary inversely with changes in cropland area. Further details are available in the online auxiliary material. The area fraction of a pixel covered by water was assumed to remain constant (although this is questionable given both the construction of many large dams and many land-drainage projects during the 19th and 20th centuries).

[25] The linear interpolation approach provides a simple, explicit and robust solution to the problem of estimating changes in natural vegetation areas as land is converted to (and from) cropland over time, where only the changes in cropland area are known. Figure 2 shows the time series of area fractions of CTEM PFTs at selected pixels during 1850 to 1992, estimated using LICCS. Although robust, a key weakness of this approach is that it does not account for differences in the timing and rate of conversions of different types of natural vegetation to agriculture and ignores vegetation succession that follows after croplands are abandoned. We therefore attempted to improve the linear interpolation approach by assigning a simple set of rules that are followed when converting natural vegetation types to cropland and vice versa.

Figure 2.

Area fractions of different CTEM PFTs for 1850–1992 estimated by the LICCS model for different locations. Symbols are as in Table 1.

3.2. Rule-Based Interpolation Cover Change Scheme (RICCS)

[26] The rule-based approach is based on LICCS but also incorporates a set of rules for conversion of natural forest types to croplands and reversion of croplands to natural vegetation types following their abandonment. It is assumed that an increase in cropland area within a single pixel is first met by converting the area covered by herbaceous (i.e., grass) PFTs. Further increases in crop area that cannot be met from this area are then obtained by removal of woody PFTs. This assumption seems reasonable because grasslands are generally easier to convert and often more productive than woody ecosystems in the same region [e.g., Wang and Hsieh, 2002]. In addition, grasslands are generally characterized by higher soil carbon amounts, because of their greater belowground allocation compared to woody PFTs [Raich and Tufekcioglu, 2000], which makes them relatively more suitable for agricultural use. Amongst the tree PFTs, broadleaf forests are assumed to occupy more productive land than needleleaf forests, and are therefore generally converted first. If present, the “bare” fraction in a grid cell is assumed to be the least productive and thus utilized last for expansion of crop area. These assumptions lead to a specific order in which natural PFTs are converted to satisfy the demand for croplands, as follows:

equation image

(see Table 1 for list of PFT symbols). The rule-based model starts with the distribution of PFTs in 1850, and works forwards in time, converting natural PFTs to cropland in this order until the area fraction of each PFT reaches that observed in 1992. In practice, not all natural PFTs are likely to occur in any given pixel, but the order implied imposes a logical progression on the sequence of land conversion.

[27] Where cropland areas are reported to have decreased with time, conversions of agricultural land are assumed to be the source of new “bare” land. The rationale here is that in many cases, agricultural land is often the initial source of land used for urban expansion. In other cases, mainly in tropical and temperate regions, bare land may result from simple abandonment of agriculture. Grasses and other herbaceous PFTs are generally faster colonizers than tree/woody PFTs [e.g., Arora and Boer, 2006] and therefore invade the bare ground first. In some cases of course, they will be later supplemented or replaced by woody PFTs. Among the tree PFTs, we assume that broadleaf trees colonize first, but in those regions where conifer forests are found, needleleaf PFTs will follow. This provides a very simple representation of succession, as seen in the coastal and boreal conifer forests of North America, for example, where broadleaf deciduous species often colonize rapidly following disturbance, but needleleaf species are often present, or even dominant, in the typical natural forest. For those regions where needleleaf forests are not found, however, broadleaf PFTs remain dominant. These assumptions lead to the following order in which the abandoned land is sequentially converted to other natural PFTs:

equation image

[28] For a pixel with no change in cropland area during 1850 to 1992, the area fraction of a PFT in year t is estimated using the linear interpolation approach. Elsewhere, changes in area fractions of natural PFTs are constrained by the changes in cropland area between 1850 and 1992. There are, however, two “special cases” where the adjustments in natural PFTs cannot be constrained in this way, because the trajectory of cropland area fraction may pass through a local maximum or minimum that either exceeds, or falls below, the values observed in both 1850 and 1992.

[29] Figure 3 shows an example of the changes in area fractions of different PFTs for the same 0.5° pixels seen in Figure 2, for the period 1850–1992, estimated using the rule-based approach. In Figure 3a, compared to the proportional adjustments seen in Figure 2a, the rule-based approach converts the natural PFTs in the specified order, with C3 grass maintaining the minimum fraction until no other natural PFT area is available. The continual presence of C3 grass is simulated because it is one of only two natural PFTs still observed in 1992. The selected grid cell used in Figures 2a and 3a is also an example of a special case in that all the natural vegetation was converted to cropland by 1890, after which some croplands were gradually abandoned. The linear interpolation approach handles this by allowing all the PFTs observed in either 1850 or 1992 to have a proportional presence after 1890. In comparison, the rule-based approach, finding that the main natural vegetation in 1992 is mainly C3 grassland with a small fraction of bare ground, first reverts to bare land, which expands until it reaches the fraction seen in 1992, after which only C3 grass is allowed to establish and, in turn, increases toward its final area in 1992. This implies that most of the bare ground in this pixel is converted to grassland relatively quickly, with only a short time delay. In the absence of any a priori knowledge of the historical changes in natural vegetation cover fractions, the rule-based approach seems more logical. See the online auxiliary material for further discussion of RICCS.

Figure 3.

Area fractions of different CTEM PFTs for 1850–1992 estimated by the RICCS model for different locations. Symbols are as in Table 1.

4. Results

[30] Figure 4 presents the global area fractions of needleleaf evergreen, broadleaf evergreen, and grass PFTs and bare areas in 1992 reclassified from the GLC2000 and SLCR data sets, combined with R&F cropland area data. Although it is difficult to assess the quality of the reclassification results at this scale, at least qualitatively, the global distribution of land cover for each biome appears to be realistic and the coarse-scale distribution of natural vegetation is in good agreement with expectation. Needleleaf evergreen trees (Figure 4a) dominate the regions of the boreal forest and contribute to forest cover in some temperate regions, while broadleaf evergreen trees (Figure 4b) dominate in the tropical forest regions. Bare ground (Figure 4d) is characteristic of the world's major desert regions. Many of the major grassland ecosystems (grass-covered plains) in central North America, Europe and Asia have been converted to croplands from 1850 onward, as shown in Figure 4c.

Figure 4.

Global area distribution of CTEM Plant Functional Types and “bare ground” estimated for 1992 (expressed as percentage of total land area). (a) Needleleaf evergreen trees. (b) Broadleaf evergreen trees. (c) Grasses. (d) Bare ground.

[31] Figures 5 and 6show the global distribution of grasses and crops, split into C3 and C4 PFTs, according to the global distribution of C4 vegetation provided by Still and Berry [2003]. Figure 6 shows global distributions of crop area fractions, split into C3 (e.g., wheat, cotton, rice and soybeans) and C4 (e.g., corn and sugar cane) crop types. In particular, Figure 6b clearly shows the corn belts of North America and significant proportions of C4 crops in India and China. The 1992 ratios of C3 and C4 crop and grass PFTs were calculated for each grid cell and then assumed to apply to the crop and grass areas for 1850, and hence for all years in between these dates. A similar assumption was made for the ratios of C3 and C4 crops and grasses under the different IMAGE 2 scenarios of future vegetation change. Simulated changes in climate would likely invalidate this assumption but it is impossible to predict a priori how these ratios would change in a fully coupled model.

Figure 5.

Global area distributions of CTEM grass PFTs estimated for 1992 (expressed as percentage of total land area). (a) C3 grasses. (b) C4 grasses.

Figure 6.

Global area distributions of CTEM crop PFTs estimated for 1992 (expressed as percentage of total land area). (a) C3 crops. (b) C4 crops.

[32] Figure 7 presents the changes in estimated global extent of all natural CTEM PFTs due to cropland changes during 1850–1992, estimated using (Figure 7a) the linear interpolation approach (LICCS) and (Figure 7b) the rule-based interpolation approach (RICCS). When compared to the linear interpolation approach in Figure 7a, the assumptions built into the rule-based approach imply that the estimated global area of grass cover shown in Figure 7b decreased more rapidly in the early part of the time series before stabilizing and eventually increasing slightly. Conversely, changes in coverage of other PFTs were comparatively slow until the latter part of the 20th century according to the rule-based approach. However, the overall differences in the results between the two approaches are relatively small at the global scale. Estimates of global extent of all CTEM PFTs at 50-year intervals, including the projections for 1993–2100 derived from the IMAGE 2 scenario data, are summarized in Table 3.

Figure 7.

Changes in global distribution of CTEM PFTs due to changes in cropland area [from Ramankutty and Foley, 1999], estimated by (a) the LICCS algorithm and (b) the RICCS algorithm for the period 1850–1992. Symbols are as in Table 1.

Table 3. Changes in Global Extents of CTEM Plant Functional Types for the Period 1850–2000 Estimated Using LICCS From Historical Changes in Cropland Area and Compared to Projections Under IPCC A1B, A2, and B1 Scenarios Derived From the IMAGE 2 Simulation for 2000–2100a
  • a

    Values are in 106 km2.

Needleleaf evergreen10.9410.419.9510.1110.1010.5110.1110.049.9610.1010.4310.64
Needleleaf deciduous3.573.553.533.463.483.663.463.453.523.463.493.66
Broadleaf evergreen17.0516.6415.9715.4115.0715.2115.4014.6112.9715.4015.4816.04
Broadleaf deciduous (cold)11.2610.329.679.619.589.709.609.569.309.629.9610.19
Broadleaf deciduous (dry)10.049.679.198.698.388.528.678.368.558.708.558.76
C3 crop6.179.1512.2314.2615.0114.4614.3015.3016.5814.2613.8812.88
C4 crop2.272.613.474.144.253.884.144.564.604.143.973.60
C3 grass14.9314.3013.4212.9212.9013.2312.9012.9813.3012.9112.9013.09
C4 grass7.867.567.257.036.977.

[33] Figure 8 shows projected changes in the global extent of CTEM PFTs estimated from the IMAGE 2 modeled data for the IPCC A1B, A2 and B1 scenarios using the linear interpolation approach. The historical data generally merge with the future projections quite smoothly although some small steps appear in one or two cases. Figure 8a shows the A1B scenario where agricultural land cover gradually increases between 1992 and 2050 and then decreases from around 2050 onward. There is a relatively rapid increase in land conversion to agriculture according to the A2 scenario results shown in Figure 8b, whereas the B1 scenario shown in Figure 8c projects stabilization around 2020 and a gradual decrease in crop cover from thereon. These results also demonstrate that according to the IMAGE 2 model, most land conversion is expected to occur in tropical regions, because the area of broadleaf evergreen trees (PFT3) that dominate tropical forests, decreases much more dramatically compared to other natural PFTs as seen in Figure 8b. Although it is not possible to validate future deforestation, the results from the IMAGE 2 model are consistent with a recent synthesis of land cover data by Lepers et al. [2005], which identifies Amazonian and South East Asian regions as hot spots for land cover change over the period 1980–2000.

Figure 8.

Changes in global area distribution of CTEM Plant Functional Types due to changes in cropland area [from Ramankutty and Foley, 1999], estimated using the linear interpolation algorithm, with alternative IMAGE 2 modeled emissions scenario data for the period 1993–2100. (a) IPCC SRES A1B. (b) IPCC SRES A2. (c) IPCC SRES B1. Symbols are as in Table 1.

[34] There are some important regional differences in the global cropland time series, particularly in rates of increase during the 1950s. Figure 9 shows the consequent changes in total areas of tree and grass PFTs compared to crops for five continents and nine regions, as estimated by RICCS for the period 1850–1992, followed by LICCS applied to the IMAGE 2 results for the SRES A1B scenario. In North America (Figure 9a), total crop cover did not increase appreciably during the 1950s, and shows slight decreases after this period. This trend is also reflected at regional scale: in the western United States (Figure 9f), cropland areas increased smoothly until 1950, accelerated somewhat in conversion from forests during the 1950s and remained steady thereafter; in the southeastern United States, cropland conversion and abandonment occurred alternately over the period 1950–1990, reflecting some major increases in forest plantation area in that region.

Figure 9.

Changes in total area of forests, grassland (including pastures and grazing land), and cropland, for the period 1850–1992 as estimated by RICCS and for 1993–2100 as estimated using LICCS applied to the IPCC SRES A1B development scenario. Graphs for each region follow those indicated in the map at top left, where the rectangular areas indicate the geographic extents of the subregion sampled for these data. Regions are as follows: (a) North America, (b) South America, (c) Africa, (d) Eurasia, (e) Oceania, (f) western United States, (g) southeastern United States, (h) the Amazon Basin, (i) eastern Brazil, (j) the Sahel, (k) eastern Europe/western Russia, (l) India, (m) South East Asia, and (n) eastern China.

[35] Most global cropland expansion in the 1950s occurred in Russia and Asia. As shown in Figure 9d, some 2 million km2 of forest were converted to cropland during this period, accounting for much of the global total (see Figure 7). Among the four regions shown in Figures 9k–9n, western Russia underwent a dramatic increase in crop cover area in the 1950s, resulting mainly from deforestation, while the other three regions show smaller increases. This compares to a more gradual and persistent conversion of natural ecosystems in the tropical regions since the 1850s, notably of natural grasslands in eastern Brazil (Figure 9i) rather than deforestation in the Amazon region (Figure 9h).

[36] As shown in Figure 9, the estimates for the two time periods are generally very similar at the transition between historical data in 1992 and the IMAGE 2 simulations for 1993 onwards (see Figures 9a–9e and 9g–9n). The only exception is for the western United States (Figure 9f), where the discrepancies in PFT area are approximately 10%. This is likely due to the difficulties of classifying the fractions among natural vegetation and bare land in that region. Elsewhere, according to the A1B scenario, cropland area will expand rapidly, notably in southern India around 2025–2075 causing extensive deforestation. Conversely, cropland is likely to be abandoned in eastern China with replacement by forests. Results from the A2 and B1 scenarios are provided in the online auxiliary material.

[37] Figures 10a–10c show changes in the areas of tree, grass and crop PFTs in 1992 relative to 1850. Some reductions in tree cover have evidently occurred in the central United States, central Europe, southeast Asia, eastern China and southern Australia. Conversely, forest increases can be seen in the eastern United States, Europe, eastern China and southern South America, which evidently correspond to cropland abandonment shown in Figure 9. Grassland areas have also decreased in the central United States, Europe and southern South America. Figures 10d–10l show the future changes for the same combination of PFTs at 2100 relative to 1992, based on the A1B, B1 and A2 scenarios from IMAGE 2. The most notable feature of the A2 scenario are that deforestation is projected to increase in tropical regions, with conversion to grazing land in the Amazon and to cropland in southeast Asia. The A1B, and particularly the B1, show less deforestation and both project significant abandonment of cropland with reversion to forest, particularly in eastern China.

Figure 10.

Global distribution of changes in forest, grassland and cropland PFT areas. (top row) Changes in area from 1850 to 1992: (a) forest, (b) grassland, and (c) cropland. (second, third and fourth rows) Changes in area of forest, grassland, and cropland, respectively, from 1992 to 2100 under three IPCC SRES scenarios: (d, e, f) A1B, (g, h, i): B1; and (j, k, l): A2.

5. Discussion and Conclusions

[38] The two interpolation approaches provide alternative simple and robust estimates of changes in the composition of natural vegetation, based on the assumption that human pressures are the dominant cause of change during the 19th to 21st centuries. It is recognized that these estimates can not be completely accurate, but with the constraints imposed by relatively detailed spatial data sets derived from recent satellite observations and the obvious requirement that total area of all PFTs must be conserved within each grid cell at all times, the estimated trajectories in historical changes of vegetation area fractions appear reasonable. The small differences in global-scale results obtained using the different assumptions contained in the two models corroborate this argument.

[39] The intended application for these data is to provide a simple estimate of changes in vegetation cover for global climate simulations. We think it is unlikely that there would be significant differences if compared to more rigorous estimates of past vegetation change, particularly when they are aggregated up to the coarse spatial resolutions typical of current generation climate models, such as the operational 3.75° or 2.8° resolutions of the CCCma CGCMs for which these data sets are designed. For other applications, such as a regional climate model with a grid resolution closer to that of the data set, inaccuracies in vegetation composition could be a problem. Hence, the use of these data for regional studies should be undertaken with caution.

[40] Nevertheless, we propose to invest some further effort in testing the results reported here. While the global distribution of natural vegetation PFTs seems reasonable, quantitative evaluation of the classifications has not been performed. There appears to be little information available that can be used for a pixel-by-pixel validation of PFTs in all regions, but it may well be possible to carry out tests on relatively small regions where present-day vegetation and past trends in land use change are relatively well documented. Such validation would require access to information from land-use statistics as well as remote sensing data.

[41] One particular problem not yet addressed adequately is the expansion of grazing land and pasture, due to forest conversions in the recent past. Historically, grazing land area expanded mostly at the expense of natural grasslands [Klein Goldewijk and Ramankutty, 2004], which implies no change in the dominant PFTs in our simplified representation of land cover change. However, a comparison of the HYDE classification with GLC2000 revealed that some areas reported as grazing land in the HYDE data set were classified as cropland or bare in the satellite product. Interpreting the HYDE classes as area fractions at 0.5 degree resolution is difficult. We assumed that the classes described as “marginal cropland/used for grazing” and “intensive cropland” implied grid cell area fractions of 60% pastureland and 60% cropland, respectively. With this approximation in mind, our results indicate that the area of natural grass PFTs assumed to be present in 1850 was large enough to supply the demand for pastureland since then (Figure S1 in the auxiliary material/data supplements). Total areas of crop and grass PFTs are generally greater in our classification than the areas of cropland and pastureland in HYDE, with the exception of a region in southern and western Australia where HYDE indicates greater coverage during 1950–1990 (see Figure S1e). In the latter case, much of the land was classified as “bare” in our data set, given that it was taken directly from GLC2000 data (classes 12 and 14 in Table 2). To some extent this inconsistency is unimportant because a coupled biosphere model such as CTEM should predict very low productivity and low leaf area index, so that the simulated surface biophysical attributes will not be very different from bare ground. Nevertheless, it may be possible to make better use of the HYDE estimates of grazing land in an updated version of the PFT data set.

[42] In the initial stages of constructing the data set, we attempted to capture the distribution of grass cover and barren areas using SLCR data for 1992, but subsequently discovered that this resulted in a significant overestimate of vegetation cover in regions generally considered barren (such as the Sahara). We then turned to the newer GLC2000 classification to allocate the dominant tree PFTs and barren land, making the assumption that forest areas did not change appreciably over the 8-year period. Subsequently, it was found that some tropical forest areas reported in the GLC2000 were appreciably smaller than those obtained from the 1992 SLCR data, consistent with reports of deforestation rates between 1992 and 2000 [e.g., Eva et al., 2004; Mayaux et al., 2004; Stibig et al., 2004]. Globally, it has been estimated that forest area decreased by about 2% during this period, mostly in the tropics, with reductions of about 6% in Africa and 3% in South America [Food and Agriculture Organization, 2001]. Given the imprecision in our own classification of the major vegetation classes (certainly within the range of ±5%), and the errors inherent in the different satellite classifications (e.g., incorrect classification of mixed 1 km pixels, and the problem of forest regrowth mentioned in section 2.1), we do not think this discrepancy creates a serious problem, though it is something that should be addressed in any future revision of the data set.

[43] The projected future changes in natural PFT areas for the different IPCC emissions scenarios are intended to be used in models that do not simulate the competition among PFTs in response to changes in climatic forcing and instead require the fractional coverage of their PFTs to be prescribed, such as in CTEM 1.0. In a fully dynamic model, where the vegetation is able to respond changes in climatic conditions, the changes in area fractions would be prognostic rather than prescribed. The IMAGE 2 projections of natural vegetation change were themselves modeled as responses to projections of future climate simulated by the atmospheric component of the IMAGE 2 model and may not be consistent with climatic changes simulated by another climate model. As such their use in other coupled biosphere-atmosphere simulations may lead to inconsistencies. This is a legitimate concern, but our view is that the data set is an interim attempt to capture first-order projected changes in vegetation distributions according to different IPCC emissions scenarios using limited resources. The next version of CTEM will simulate the competition among its non-crop PFTs explicitly according to a modified form of the Lotka-Volterra equations [Arora and Boer, 2006] and therefore will require only the area fractions of crop PFTs to be prescribed.


[44] We would like to thank Navin Ramankutty for providing the global historical cropland data sets. We would like to thank Richard Betts of the U.K. Hadley Centre and one anonymous reviewer for their very helpful comments. We thank Marty Siltanen of Canadian Forest Service for assistance with GIS. This work was supported by funding from the Canadian Foundation for Climate and Atmospheric Sciences (CFCAS).