1. Introduction

Vegetation cover influences the land–atmosphere exchanges of water, energy, and carbon (e.g., Dickinson et al. 1986; Sellers et al. 1996; Bonan 1996; Dai et al. 2003). Green vegetation fraction (GVF; Deardorff 1978) is widely used in global models along with many other applications such as studies of land-cover (LC) change. Along with leaf area index (LAI; Myneni et al. 2002), GVF is used to describe the abundance of vegetation in most global models. Some models, such as the Community Land Model (CLM; Lawrence and Chase 2007) represent this abundance with a seasonally varying LAI and a maximum annual GVF at each grid cell (or subgrid tiles of a grid cell), and other models, such as the “Noah” land model (Chen and Dudhia 2001; Ek et al. 2003), represent this abundance with a constant maximum LAI and a seasonally varying GVF. As such, some GVF datasets represent seasonally varying vegetation cover (e.g., Gutman and Ignatov 1998; Miller et al. 2006), while others represent the maximum annual vegetation cover (e.g., DeFries et al. 1999; Zeng et al. 2000, hereinafter Z00; Hansen et al. 2003).

CLM currently uses estimates of maximum GVF from “Continuous Fields” (CF) data that are based on measurements from the Moderate Resolution Imaging Spectrometer (MODIS) from 2001 (Hansen et al. 2003). Basing land-cover data on one year of data is a potential drawback because anomalous conditions (e.g., flood/drought) can lead to unusual vegetation greenness in certain areas. This is especially problematic in semiarid environments, where vegetation greening is strongly linked to the availability of water (e.g., Hadley and Szarek 1981; Weiss et al. 2004). This problem might lead to inconsistencies with other vegetation data that are used in the model, especially if the data are from different years.

Another simpler method to derive GVF, which can be updated annually, is to use satellite-based vegetation indices directly. Considerable research has shown that the widely used normalized difference vegetation index (NDVI; Tucker 1979) is related to GVF (e.g., Leprieur et al. 1994; Carlson and Ripley 1997; Wittich 1997; Gutman and Ignatov 1998; Purevdorj et al. 1998). This relationship has been used to create maps of GVF for global studies (e.g., Gutman and Ignatov 1998; Z00; Zeng et al. 2003), and such implementations have been tested in global models (Barlage and Zeng 2004; Miller et al. 2006). The derivation of high-resolution datasets for maximum GVF using this method [which we refer to as maximum green vegetation fraction (MGVF)] has largely relied upon NDVI measurements from previous-generation Advanced Very High Resolution Radiometer (AVHRR) sensors, although more advanced satellite sensors (e.g., MODIS) have better spectral resolutions and so are more ideal for measuring vegetation indices such as NDVI (e.g., Fensholt and Sandholt 2005). Therefore these efforts should be updated with newer data.

Building on our prior efforts [Z00, based on 1-km AVHRR data from April 1992 to March 1993, and Zeng et al. (2003), based on 8-km AVHRR data from 1982 to 2000], here we develop an updated 1-km MGVF product for global applications that is based on 12 years of MODIS data. We both analyze the interannual variability in the new MGVF product and compare it with data from the CF Collection 3 that are currently used in CLM.

2. Developing global MGVF data

This study uses MODIS Collection-5 NDVI data (MOD13A2; Huete et al. 2002) and MODIS Collection-5.1 data for land-cover type (MC12Q1; Friedl et al. 2010) to define MGVF. The MOD13A2 product contains 16-day composites of NDVI from 2001 to 2012 at an ~1-km pixel resolution on the MODIS sinusoidal grid. Land-cover classes are defined using the International Geosphere–Biosphere Program (IGBP; Townshend 1992) LC classification in the Collection-5.1 MCD12Q1 dataset.

Our recent study (Broxton et al. 2014) demonstrated that there is considerable spurious interannual variability of the LC data (which have been generated annually since 2001). Therefore, instead of using the MCD12Q1 data directly, we generate a land-cover “climatology” that is based on 2001–10 LC data using the confidence scores in the MCD12Q1 data to weigh the LC classification for individual years (Broxton et al. 2014). The LC data have an ~0.5-km spatial resolution on the sinusoidal grid.

We use a linear mixing model with the assumption that the vegetation-covered fraction for each pixel can be obtained from measured NDVI by treating this NDVI as the area-weighted average of NDVI contributions from vegetated area and from bare soil within each pixel. Following Z00, we define MGVF between 0 and 1 on the basis of the annual maximum of NDVI:

View Expanded

Here, N_max is the annual maximum observed NDVI at each pixel, N_s represents the annual maximum NDVI of bare soil, and N_c,υ represents the annual maximum NDVI of the vegetated surface; N_s is a global constant and is based on the NDVI of “bare ground” pixels, and N_c,υ is computed separately for each vegetation type (described below). The values that N_s and N_c,υ take are based on histograms of N_c,υ for each LC category (which are generated by grouping all N_max pixels according to their respective LC type). To find the histograms, the ~0.5-km LC-type climatology is resized to ~1 km (to match the spatial resolution of the NDVI data) using nearest-neighbor resampling, because this method works with data that are discrete. We use the same rules (with a couple of exceptions, described below) to determine N_s and N_c,υ that Z00 do, who estimate N_s and N_c,υ on the basis of GVF estimates from 1- and 2-m satellite data. Here, N_s is taken to be a global constant and is defined to be the value of the 15th percentile of the N_max histogram of the barren IGBP type. The value of N_c,υ is unique to each IGBP land-cover type, and it takes the value of the 75th percentile of the N_max histogram for most IGBP types, with the exception of the closed-shrubland, open-shrubland, urban, and barren IGBP types. For closed shrubland, N_c,υ is the 95th percentile of the N_max histogram, for urban, it is the 90th percentile of the N_max histogram, and the value of N_c,υ for open shrubland and barren is the same as that for closed shrubland. These rules are summarized in Table 1, and more details on Eq. (1) and relevant parameters are provided in Z00. Note that the computed vegetation cover within each LC class is a function of NDVI, which is variable through time, and therefore it is possible to have bare ground for any LC class at some times and not others.

Table 1.

Percent of land area, N_c,υ used in Eq. (1), and the minimum and maximum annual MG values from 2001 to 2012 (in parentheses), MG₀₁, values for MG reported by Z00 (on the basis of AVHRR from April 1992 to March 1993), and MG_CF (from 2001) for each IGBP land-cover type provided in Fig. 3.

View Table

There are a few differences between these rules and those used by Z00. First, the N_max histogram percentile of N_s is adjusted (from the 5th percentile to 15th percentile) to ensure that known desert areas have near-zero vegetation cover in the new classification. This is possibly a reflection of better LC classification of barren areas in the MODIS-based classification than in the previous AVHRR-based classification (including fewer areas with vegetation). Nevertheless, we feel it is justified because N_max is higher than that given by the 5% threshold (0.07) for large portions of the Gobi Desert and almost all of the Atacama and Sahara Deserts, whereas it is only higher than that given by the 15% threshold (0.09) for limited portions of the Atacama and eastern Gobi Deserts and large portions of the Sahara Desert. In addition, the N_max histogram percentile of N_c,υ for closed shrublands is adjusted (from 0.90 to 0.95), partly because a much smaller proportion of the earth’s surface is covered by closed shrublands in the new MODIS LC classification.

MGVF maps are created using this method for each calendar year from 2001 to 2012 (annual values of MGVF will henceforth be referred to as MG). Values of MG are computed for each ~1-km pixel as the average of MG values using the four 0.5-km LC-type values in each 1-km N_max pixel because it visually softens hard edges created by the fact that different LC types have different values of N_c,υ. This has a minimal impact on the MG maps at the global scale, though, versus just taking a single LC type to be representative within an N_max pixel. In this study, comparisons are made between the 12-yr averaged annual MG (denoted as), MG from 2001 (denoted as MG₀₁), and the 2001 bare-ground percentage map [with 1 − (bare percentage) denoted as MG_CF] in the CF Collection-3 product (MOD44B.003; Hansen et al. 2002, 2003). We do not do a comparison with newer versions of the CF product (i.e., Collection 4v3 or Collection 5) because the newer versions only provide tree-cover fraction (without bare-soil fraction). The CF data (Collection 3) as used here were generated using a regression tree with MODIS surface-reflectance data, MOD09A1, as input from November 2000 to November 2001. The MG_CF is downloaded as tiles on the MODIS sinusoidal grid, and maps of MG₀₁, , and MG_CF are regridded using bilinear resampling to a 0.01° (or ~1 km) pixel resolution for comparison.

For consistency check, we also compare these products with Collection-5 MODIS LAI product (MOD15A2; Myneni et al. 2002) that is retrieved from the same MODIS surface-reflectance data used in CF and our MGVF development. The methods for the development of these three products, however, are largely independent of each other. The LAI data, whose native resolution is ~1 km on the sinusoidal grid, are also regridded (using bilinear resampling) to a 0.01° pixel resolution for analysis. In this study, we compare MG₀₁ and MG_CF with the maximum annual LAI from 2001 (LAI_max,01; computed for each pixel by finding the maximum LAI value from the 8-day MOD15A2 LAI composites for 2001).

3. Results

a. Characteristics of our MGVF

Overall, the values computed in this study are similar to MGVF values obtained by Z00 for most LC classes (Table 1), despite significant differences between LC abundances and N_c,υ values. Values of N_c,υ and N_s from this study are significantly higher than those reported by Z00. The average of N_c,υ values for all LC types in this study is 0.84 as compared with 0.62 in Z00, and our N_s value here is 0.09 as compared with 0.05. Some LC types are much more abundant in our newer LC classification: on the basis of the MODIS data used in this study, types 5, 10, and 13 (mixed forest, grassland, and urban) make up 6.15%, 14.13%, and 0.50% of global land areas, whereas Z00, who used an AVHRR-based classification, reported that these classes cover 4.86%, 8.53%, and 0.2% of land areas. Other MODIS LC classes are much less abundant: we find that types 1, 3, 6, and 14 (evergreen needleleaf, deciduous broadleaf, closed shrubland, and cropland/natural vegetation) make up 2.25%, 1.09%, 0.11%, and 6.71% of global land areas, whereas Z00 found their percentages to be 5.03%, 2.50%, 2.01%, and 10.80%, respectively. The average values for most LC classes are within 0.07 of those found by Z00, however, with the exception of type 7 (open shrubland), which we find to have an average of 0.58 (as compared with 0.39). The larger difference for open shrubland is partially caused by the more abundant MODIS open-shrubland areas (by including some closed-shrubland pixels with higher MGVF) than in the original AVHRR LC data.

The global distribution of our product is shown in Fig. 1a. In many places is similar from year to year, but in some places there is a substantial amount of interannual variability. Over the globe, the standard deviation of MG (Fig. 1b) from 2001 to 2012 is less than 0.05 for 64% of land areas, is between 0.05 and 0.15 for 34% of land areas, and is greater than 0.15 for 2% of land areas. Certain LC classes are also prone to more or less interannual variability of MG than others (Table 2). Types 1–5, 8, 14, and 16 (the five forest classes, woody savanna, cropland/natural vegetation, and barren) have a lower amount of interannual variability than do the remaining classes as the median (95th percentile) of the standard deviation maps for these classes is less than 0.04 (0.08). Types 7, 9, 10, and 12 (open shrubland, grassland, savanna, and cropland) have a high degree of interannual variability and are globally abundant, and so much of the interannual variability shown in Fig. 1b is found in these classes. Despite the sometimes substantial amount of interannual variability in some land-cover types, we find that MG for each land-cover class is (averaged across all pixels of a given type) consistent from year to year (the range never exceeds 0.03; Table 1).

View Full Size

(a) Average annual MG from 2001 to 2012 (), (b) standard deviation of MG from 2001 to 2012, (c) MG₀₁ − MG_CF (from 2001), and (d) difference between the absolute difference of − MG_CF and the absolute difference of MG₀₁ − MG_CF, along with the locations of the zonal cross sections that are referred to later in Table 3. In (d), redder colors indicate better agreement between MG₀₁ and MG_CF than between and MG_CF, and bluer colors indicate better agreement between and MG_CF than between MG₀₁ and MG_CF. To reduce data volume, the original data on an ~1-km sinusoidal grid in (a) and (b) and a 0.01° grid in (c) and (d) are projected to geographic coordinates at a 0.1° resolution from 60°S to 80°N.

Citation: Journal of Applied Meteorology and Climatology 53, 8; 10.1175/JAMC-D-13-0356.1

• Download Figure
• Download figure as PowerPoint slide

Table 2.

Median and 95th-percentile values for the standard deviation of MG map (Fig. 1b) and the absolute value of the MG₀₁ − MG_CF map (Fig. 1c) for each IGBP land-cover class.

View Table

The interannual variability of MG also changes systematically with : it is more pronounced at lower to intermediate MG values. In general, the highest variability of MG occurs where is 30%–60% (Fig. 2a). At lower MG, this interannual variability is limited by the fact that is low. For higher (primarily for forests), it is limited as well. In fact, when viewed relative to , the variability of MG decreases nearly monotonically (Fig. 2b). The amount of interannual variability for a given varies widely for different LC classes. For example, the barren class (type 16) exhibits a high amount of interannual variability (relative to the value) throughout the entire range of values that are common for barren, whereas the forest classes (types 1–5) exhibit a low amount of interannual variability throughout the entire range of values. The interannual variability in other LC classes, such as shrublands (types 6–7) and grasslands (type 10), decreases relative to , albeit at different rates. As a result of this higher interannual variability at lower , we find that, even on a 0.5° climate-model grid (e.g., that used by CLM), the average absolute difference between MG for a particular year and is over 20% of the value of for 11% of global land areas.

View Full Size

Box plots showing (a) the relationship between and the standard deviation of [SD(MG)] and (b) the relationship between and the ratio of SD(MG) over. Each box plot is broken into 100 bins from 0 to 1, and the shaded areas capture the 25th–75th percentile range for each bin. Also shown at 0.05 increments are median SD(MG) and SD(MG)/ for grassland (IGBP type 10) pixels (green circles), shrubland (types 6–7) pixels (yellow circles), barren (type 16) pixels (red circles), and forest (types 1–5) pixels (blue circles).

Citation: Journal of Applied Meteorology and Climatology 53, 8; 10.1175/JAMC-D-13-0356.1

• Download Figure
• Download figure as PowerPoint slide

b. Comparison with bare-ground continuous fields

Comparison between and MG_CF [i.e., 1 − (bare-soil fraction)] is complicated by two factors. First, is based on data from 2001 to 2012, whereas MG_CF is based on data from 2001 only. In some places, the difference between and MG_CF reduces dramatically by comparing MG₀₁ and MG_CF. For example, in a region in the Great Plains of southern Canada (50°–55°N, 105°–110°W), MG_CF of 0.76 is more similar to MG₀₁ (0.75) than to (0.84). In another instance, in eastern Saudi Arabia (27°–29°N, 44°–46°E), MG_CF of 0.20 is different from of 0.07 but is very close to MG₀₁ of 0.18. The fact that some of the differences between and MG_CF are related to interannual variability also partly explains some of the codependence between the standard deviation of MG (Fig. 1b) and differences between MG₀₁ and MG_CF (Fig. 1c). After all, many of the LC classes with larger differences between MG_CF and are the same classes with large variability of MG (Table 3).

Table 3.

The R² values (the coefficient of determination, or squared correlation coefficient) that show the strength of the regression between LAI_max,01 and MG₀₁ or MG_CF for the indicated land-cover type in each 0.5° grid box over the zonal cross sections in Fig. 1d.

View Table

In many places, though, the fact that MG_CF is based on 2001 MODIS data and is based on 2001–12 MODIS data cannot explain the differences between and MG_CF. Not only are there many differences between data for a comparable time period (shown by the comparison of MG₀₁ and MG_CF in Fig. 1c), but these differences are not much smaller than those between and MG_CF in many areas in Fig. 1d. In addition, there are many global systematic differences between MG₀₁ and MG_CF. Consistent with the conclusion from Fig. 1c and Table 2 that MG_CF is, in general, larger than MG₀₁, MG_CF is usually higher for a given N_max (from 2001; referred to as N_max,01) than MG₀₁ in Fig. 3. In fact, for the forest types, MG_CF is close to 1 (100% cover) for nearly the entire range of N_max,01 that occurs in these classes. The relationship between N_max,01 and MG_CF is similar to that between N_max,01 and MG₀₁ only in types 10 and 16 (grasslands and barren).

View Full Size

Comparison between N_max,01 and MG_CF (the ordinate on the left shows the scale) in each IGBP land-cover class. Black dots denote the median value of MG_CF data for each 0.01 N_max,01 bin, and the shaded areas capture the 25th–75th percentile range. Gray lines show N_max,01 histograms (the ordinate on the right shows the scale). Comparisons between N_max,01 and MG_CF are only shown where the histogram of N_max,01 is greater than 0.0025. Also shown are values of N_s and N_c,υ (vertical dotted lines) and the conversion between N_max,01 and MG₀₁ (red lines).

Citation: Journal of Applied Meteorology and Climatology 53, 8; 10.1175/JAMC-D-13-0356.1

• Download Figure
• Download figure as PowerPoint slide

In a similar way, for most LC classes (again except for grasslands and barren), MG_CF is higher than corresponding MG₀₁ at a given LAI_max,01 (Fig. 4). In fact, for most LC classes, MG_CF is close to 1 for LAI_max,01 values as low as 2, and in some cases (e.g., types 1 and 11; evergreen needleleaf and permanent wetland), MG_CF is above 0.9 for LAI_max,01 values as low as 1. At the same time, MG₀₁ is below 1 for all except very high LAI_max,01 values for some LC classes (e.g., 6 for types 4 and 5; evergreen broadleaf and mixed forest), although usually MG₀₁ is greater than 0.9 above LAI_max,01 = 4. The relationships between and the average annual N_max from 2001 to 2012 (), and between and the average value of LAI_max from 2001 to 2012 () are almost identical to those between MG₀₁ and N_max,01 and between MG₀₁ and LAI_max,01.

View Full Size

Comparison between LAI_max,01 and MG_CF (black dots and blue bars) and MG₀₁ (black dots and red bars) for each IGBP land-cover class (the ordinate on the left shows the scale). Black dots denote the median MG_CF or MG₀₁ value for each 0.01 LAI_max,01 bin, and the blue and red bars capture the 25th–75th percentile range. Gray lines show histograms of maximum LAI_max,01 in each category (the ordinate on the right shows the scale). Comparisons between MG_CF or MG₀₁ and LAI_max,01 are only shown for LAI_max,01 categories where the histogram of LAI_max,01 is greater than 0.0025.

Citation: Journal of Applied Meteorology and Climatology 53, 8; 10.1175/JAMC-D-13-0356.1

• Download Figure
• Download figure as PowerPoint slide

There is better spatial agreement between MG₀₁ and LAI_max,01 than between MG_CF and LAI_max,01. Table 3 presents the coefficient of determination (or squared correlation coefficient R²) between the spatial variability of LAI_max,01 and MG_CF/MG₀₁ along zonal cross sections shown in Fig. 1d. Cross sections are 0.5° wide, have a variable length, and are located to measure spatial variability in areas with homogenous LC classes but with poor agreement between MG₀₁ and MG_CF. Each cross section is divided into 0.5° increments (to match the spatial resolution of the land-cover data used in CLM), and values of LAI_max,01, MG_CF, and MG₀₁ are averaged for each 0.5° grid cell. Linear regression is performed on the collection of averaged values for each 0.5° grid cell for each cross section, because most relationships are linear or only weakly nonlinear.

The agreement between MG₀₁ and LAI_max,01 is higher than that between MG_CF and LAI_max,01 for almost all zonal cross sections, and the average R² of 0.87 is 0.20 higher for the former than the latter (Table 3). The R² differences are usually significant, as the 95th confidence intervals for each trend line (which are found using a Fisher Z transformation) do not overlap with the other trend line except for cross sections B, H, I, M, and N. For comparison, the R² value between MG₀₁ and is 0.85, and the value between MG_CF and is 0.59 (a difference of 0.26).

4. Discussion and conclusions

We have created a MODIS-based MGVF product for global applications. This product uses 12 years of MODIS satellite data of NDVI and land-cover type, which have a higher spectral resolution and an overall better quality than previous-generation AVHRR-based data. We find that, for most LC classes, values of are similar to those found by Z00 (upon which this work is based), despite significant differences between LC abundances and N_c,υ values, demonstrating the robustness of the method used in Z00. The exception is type 7 (open shrubland), for which MGVF is substantially higher than that found in Z00. This is probably a reflection of the fact that the MODIS open-shrubland areas are more abundant (by including some closed-shrubland pixels with higher MGVF) than in the original AVHRR LC data. In general, N_c,υ values are lower in Z00 because the AVHRR NDVI used in Z00 is lower than MODIS-based NDVI. In addition, the LC type products that are based on AVHRR data (e.g., the version-2 Global Land Cover Classification dataset that is widely used in regional and global models) are substantially different, and they “validate” worse than the MODIS LC-type data (Broxton et al. 2014).

The MGVF values computed in this study are highly variable from year to year in some parts of the world. It is unlikely that uncertainty of NDVI values can lead to the amount of interannual variability that characterizes MG, suggesting that the variability is largely real, perhaps reflecting variability in climatic conditions for a given year. We also find that in many parts of the world much of the interannual variability of MG occurs in the open-shrubland, grassland, savanna, and cropland classes. This is consistent with Zeng et al. (2003), who found that these four classes are among the most variable classes from 1982 to 2000.

Furthermore, areas with lower are more likely to have higher interannual variability of MG. The standard deviation of MG is highest for that is close to 0.5. For low values of , interannual variability of MG is limited by the fact that there is not much vegetation cover. There is also less variability of MG for more-vegetated regions (e.g., forests), however. In fact, relative to , the standard deviation of MG decreases nearly monotonically with . This result is probably a reflection of the fact that vegetation is more responsive to climatic factors (especially precipitation) in drier, water-limited environments (e.g., Hadley and Szarek 1981; Weiss et al. 2004). More vegetated regions, which are not water limited, are less affected by interannual variability in precipitation (unless, e.g., a severe drought or forest fire leads to a large-scale forest die-off). Some of the regions with higher interannual variability of MG are the same as those that are thought to have strong coupling with the atmosphere (Koster et al. 2004; Zeng et al. 2010), further underscoring the need to have vegetation products that encompass multiple years (rather than just a single year).

This interannual variability can explain some, but not all, differences between and MG_CF. In some parts of the world, the differences between and MG_CF are much larger than those between MG₀₁ and MG_CF, suggesting that these differences are more likely to reflect real differences between 2001–12 “average” conditions and those during 2001. A majority of the differences between and MG_CF remain in the comparison between MG₀₁ and MG_CF, however, and so cannot be explained by the different years of data used. In most regions, MG_CF is higher than MG₀₁, and, for most LC classes, MG_CF is higher than MG₀₁ for a given value of N_max. It is difficult to judge whether the “greener” or “less green” representation of MGVF is more accurate because of a lack of ground truth. For example, MG_CF for most LC classes is close to 1 for unrealistically low LAI_max,01 values (as low as 1–2); at the same time, however, for some classes MG₀₁ does not approach 1 until very high LAI values (as high as 5–6), although most of the time MG₀₁ is higher than 0.9 for LAI values from 3 to 4. In addition, there is uncertainty in the way that both products relate N_max to MGVF.

On the other hand, the method that we use (which is based directly on NDVI) has one clear advantage over the CF data: its spatial variability is significantly more consistent with LAI. Along zonal cross sections that are located in areas with a homogenous land-cover type but where there is poor agreement between MG₀₁ and MG_CF, we find that the spatial variability of MG₀₁ is more consistent with that of LAI_max,01 than is that of MG_CF. This consistency is important because both MGVF and LAI are used by global models to describe vegetation abundance. In the ideal case, the spatial variability of MGVF and LAI_max should be relatively consistent within a given land-cover type, as they are both measures of vegetation abundance. For example, for grasslands, it does not make sense to have LAI_max decrease as MGVF increases (across space) and vice versa because vegetation structure is probably not highly variable at the 1-km scale. For other LC classes, such as open shrublands, where one would expect more diversity in terms of vegetation structure, it is a little more realistic to have differing spatial variability for MGVF and LAI_max, but the two should still probably be very closely related. It is not surprising that there are high correlations between MG₀₁ and LAI_max,01 (with an average R² of 0.80); it is more surprising that the correlations between MG_CF and LAI_max,01 are significantly lower. Note that MG_CF from an older version of the MODIS CF data is used because the bare-ground fraction field is not available in more recent versions. Such fields in future CF data might show differences with the previous CF data.

In this study, we develop a global 1-km MGVF product based on a climatology of MODIS NDVI and LC-type data. The MGVF climatology is better than the product for a single year because it removes biases associated with unusual greenness and inaccurate land-cover classification for a specific year. Its spatial variability is also more consistent with that of annual maximum LAI (which is also used to describe vegetation abundance in models and therefore should be relatively consistent with MGVF) than are the CF data. Therefore, we believe that it is a good candidate for global-modeling applications that use MGVF to describe vegetation abundance. The data are freely available from the authors.