Farming the planet: 1. Geographic distribution of global agricultural lands in the year 2000

2.1. Satellite Data Sets

[9] We used two different high-resolution (1-km) satellite-based, global land cover classification data sets that are available for circa 2000: Boston University's Moderate resolution Imaging Spectrometer (MODIS) based global land cover product [Friedl et al., 2002] (BU-MODIS hereafter), and the Satellite Pour l'Observation de la Terre (SPOT) VEGETATION based Global Land Cover 2000 (GLC2000) data set [Bartholome and Belward, 2005]. The BU-MODIS land cover product used data acquired from 15 October 2000 to 15 October 2001 to derive 17 land cover classes using a supervised classification scheme (see Table 1 for legend). The GLC2000 data set utilized data acquired from 1 November 1999 to 31 December 2000 to derive 22 global land cover classes on the basis of regional classifications performed with the expertise of regional institutions (Table 1).

[10] We applied a simple set of climatic parameters to mask obviously nonagriculture areas within the satellite data sets, else we obtain some spurious results in the Northern Hemisphere high latitudes (e.g., land classified as open shrublands are found both in the western United States and boreal Canada; because open shrublands have a small fraction of cultivation in the former we end up with croplands in boreal Canada also). This mask included all regions north of 50°N with Growing Degree Day (GDD; base 5°C) less than 1000°C d (see Figure S1). This GDD cutoff approximates that empirically estimated by Ramankutty et al. [2002a] (using their previous cropland data set that does not include a mask, hence there is no circularity), and is conservative compared to the 2000°C d cutoff for agriculture used by Cramer and Solomon [1993]. GDD data were calculated according to Ramankutty et al. [2002a], and interpolated to 1km resolution. We also masked out protected/minimal use areas in the central part of the Australian continent (Bureau of Rural Sciences, Land Use of Australia, Version 3: 2000/2001, Department of Agriculture, Fisheries and Forestry, Australian Natural Resources Data Library, Canberra, Australia, available at http://www.nlwra.gov.au/Data_Library/index.aspx), which otherwise gets classified entirely as pasture.

2.2. Agricultural Inventory Data

[11] We extensively compiled cropland and pasture inventory data for the globe at the national and subnational level (Figures 1a–1c and Table 2; more details in Text S1) for circa year 2000. We compiled data for 15,990 different administrative units of the world, ranging from political units like countries, states and counties, which represents a 46-fold improvement in the richness of our inventory data compared to our previous effort (348 units in the work of Ramankutty and Foley [1998]). For 57 countries, we compiled census data at the subnational level (e.g., “Level 1” indicating states in the United States or India, provinces in Canada or Argentina, departments in Bolivia or Columbia, etc., and “Level 2” indicating smaller units like U.S. counties, Brazilian municipios, or Indian districts). For 159 countries, we used national level statistics from the Food and Agriculture Organization's (FAO) FAOSTAT database (http://faostat.fao.org); for these 159 countries, we calculated an average around the year 2000 using data from 1998 to 2002. For another 19 countries in our database, no FAOSTAT data was available, and we set the data to be missing (see Text S1 for details).

[12] We compiled the cropland and pasture data to be consistent with the FAO definition of “Arable lands and permanent crops” and “Permanent pastures” respectively. Arable land is defined by FAO (http://faostat.fao.org/site/375/default.aspx) as including “land under temporary crops (double-cropped areas are counted only once), temporary meadows for mowing or pasture, land under market and kitchen gardens and land temporarily fallow (less than 5 years). The abandoned land resulting from shifting cultivation is not included in this category. Data for arable land are not meant to indicate the amount of land that is potentially cultivable.” Permanent crops are defined as “land cultivated with crops that occupy the land for long periods and need not be replanted after each harvest, such as cocoa, coffee and rubber; this category includes land under flowering shrubs, fruit trees, nut trees and vines, but excludes land under trees grown for wood or timber.” Permanent pastures are defined as “land used permanently (5 years or more) for herbaceous forage crops, either cultivated or growing wild (wild prairie or grazing land). The dividing line between this category and the category ‘Forests and woodland’ is rather indefinite, especially in the case of shrubs, savannah, etc., which may have been reported under either of these two categories.”

[13] The agricultural inventory data were seldom available exactly for the year 2000 because agricultural censuses are only taken every 5–10 years a in most industrialized nations of the world, and less frequently in other countries. We collected inventory data between the years 1998 to 2002 where possible, but in several instances we relied on older data (see Table 2). For nations where the inventory data did not fall within the 1998–2002 period, or where cropland or pasture data were unavailable but some proxies were available (such as harvested area of individual crops or heads of livestock), we estimated cropland and pasture data for circa 2000 by calibrating the available information to national totals from FAOSTAT (see Text S1 for details).

[14] The quality of our census data is varied. Some regions of the world are not well represented in terms of the resolution of inventory data, with the African continent and the Former Soviet Union being the most underrepresented. Sometimes, the national level census statistics were inconsistent with the FAOSTAT data; in such cases we mostly relied on the national statistics (as recommended by FAO), except for a few cases where we believed that FAOSTAT data were more reliable (see Text S1 for details). Inconsistencies were a result of either unclear definitions of the category, or sometimes poor reporting by the national statistics agency. For example, the cropland census statistics for China has been noted by various studies to be particularly problematic [Crook, 1993; Frolking et al., 1999; Seto et al., 2000; G. K. Heilig, Can China Feed Itself?: A System for Evaluation of Policy Options, International Institute for Applied Systems Analysis, 1999, available at http://www.iiasa.ac.at/Research/LUC/ChinaFood/index_m.htm]. Here we have used the data for China from the study of Verburg and Chen [2000], which seems reliable. The definition of permanent pasture is particularly problematic, as acknowledged by the FAO (http://faostat.fao.org/site/375/default.aspx). Some countries (e.g., the United States) clearly distinguish between grassland pasture and range and forest use land grazed, while most countries do not. So it is not clear whether grazed forest land or semiarid grazing is included under the definition of pastures. One egregious example is Saudi Arabia for which FAOSTAT reports 1.7 million km² of permanent pasture in 2000, which is 80% of its total land area. However, most of Saudi Arabia is arid land and it is clear that much of the nomadic grazing areas are included under pasture. The Saudi Arabian subnational census data we obtained reports 486 km² of pasture, which is 3500 times smaller; we have chosen to rely on this lower value. Similarly, FAOSTAT reports 4 million km² of pasture in China. However, Verburg et al. [2000] report a total of only 2.6 million km² of grassland in China. In this case, we have used the FAOSTAT value for the total amount of pasture in the country (but our final estimates for China are lower; see section 4.1, paragraph 2). Similar inconsistencies exist in other countries and are reported in Text S1.

[15] In summary, our agricultural inventory database yields a global total of 15 million km² of cropland and 31.5 million km² of pasture. This compares to 15.3 million km² of cropland and 34.4 million km² of pasture reported by FAOSTAT; significantly, our census data compilation and interpretation yields about 8% less pasture than FAOSTAT. The difference between FAOSTAT and our own inventory for pasture can mainly be explained by differences in Saudi Arabia (as described earlier in section 2.2, paragraph 4), Australia, Nigeria, Brazil, Mexico, Indonesia, Argentina, Colombia, Russia, and Spain. There were very few countries where the national census pasture area was greater than FAOSTAT; the one exception is Iran, for which FAOSTAT reports 0.4 million km² of pasture while the national census reported 0.9 million km². For croplands, while the global total areas are comparable between FAOSTAT and our inventory, there are significant national differences. For example, the national inventories of Australia, Brazil, Canada, China and Turkey report lower cropland area compared to FAOSTAT, while Iran, Argentina, Nigeria, Mexico, and Indonesia report greater cropland area.

3.1. Combining the Two Satellite-Derived Land Cover Classification Data Sets

[17] Giri et al. [2005] compared the GLC2000 and BU-MODIS land cover data sets and found some consistency at the aggregate class level, but widespread disparities in the details. A few other studies have also tried to intercompare and harmonize these two and other global land cover data sets [Herold et al., 2005; Jung et al., 2006; See and Fritz, 2006]. The GLC2000 data set was developed using SPOT vegetation data with the assistance of regional experts and used a flexible classification scheme. The BU-MODIS data set, on the other hand, was developed using a globally consistent procedure with a fixed classification scheme, but without regional expertise. The two data sets therefore bring different kinds of expertise and information that are potentially complimentary. Here we decided to capitalize on both by combining them into a single land cover data set. To do so, we overlaid the BU-MODIS and GLC2000 data sets and developed new land cover classes that contain all possible combinations of the individual land cover classes (see Table 3 for an example). The resulting combined land cover data set consists of 391 possible land cover types and is shown later in section 3.3 to provide more accurate results when calibrated against the inventory data, as opposed to using either data set individually.

3.2. Step 1: Calibrating the Satellite Data Sets Against the Agricultural Inventory Data

[18] We followed methods developed in previous efforts [Ramankutty and Foley, 1998; Cardille et al., 2002; Ramankutty, 2004] to merge the satellite data sets and the agricultural inventory data to develop global maps of croplands and pastures for the year 2000. As mentioned in section 3 (paragraph 1), our statistical data fusion procedure is different from Ramankutty and Foley [1998] in a couple of major ways. First, to allow for potential misclassifications in the satellite-derived land cover data sets [e.g, Hurtt et al., 2001; Cardille et al., 2002], and also because pasture is not explicitly identified as a land cover class by the satellite-derived data sets, we utilized all land cover classes in our training procedure, as described below. Second, we used the two-step procedure of Ramankutty [2004] which assumes that the inventory data is the “truth” (except for identified outliers), and uses the satellite data to spatially disaggregate the census data within each administrative unit.

[19] For each administrative unit, i, the proportion of cropland and pasture area from the inventory data, cf_i and pf_i respectively, was calculated by dividing the inventory cropland and pasture areas by the total land area (A_i) for the administrative unit. Then, we determined λ_j,i, which is the proportion of each of the satellite-derived land cover classes, j, within administrative unit i.

[20] We have = 1.0, where n_λ is the number of land cover categories in the satellite data set.

[21] We formulated a linear model relating the satellite-derived data sets to the agricultural inventory data, as follows:

and

where α_j and β_j are unknown parameters associated with each land cover category j, and ɛ_c,i and ɛ_p,i are error terms that represent the residual difference between the inventory and linear model predicted cropland and pasture proportions respectively.

[22] Additionally, (1) and (2) were subject to the following constraints,

These constraints ensured that the cropland or pasture proportions in any pixel (when the model is later applied at pixel level) will be between 0 and 100%, and that the sum of cropland and pasture proportions will be less than or equal to 100%.

[23] We used a least squares minimization method to solve for the parameters α_j and β_j. In particular, we specified the weighted least squares error (LSE) to be minimized as:

where n_i is the number of administrative units, and ω_i is a term that weights the residuals ɛ_c,i and ɛ_p,i by the land area, A_i, of each administrative unit, normalized by the maximum value, i.e.,

To estimate the parameters in equations 1 and 2, we used a multiple linear regression model from the STARPAC package (http://www.cisl.ucar.edu/softlib/STARPAC.html). We developed three separate models, first using the BU-MODIS and GLC2000 data sets separately, and then using the combined land cover data set. In each case, we started with a complete model specifying all the land cover classes as potentially being cropland or pasture; the following classes: BU13 (urban), BU15 (snow and ice), BU16 (barren), BU17 (no data), GLC19 (bare), GLC20 (water), GLC21 (snow and ice), GLC22 (artificial surfaces), and GLC23 (no data), and their combinations in our combined land cover data set, were left out of the model. We then used stepwise regression using backward selection to estimate the parameters. The details of this procedure are outlined in Text S2.

[24] We applied our optimization procedure separately to six different regions of the world (Figure S2), similar to Ramankutty and Foley [1998]. These six regions were a compromise between selecting small enough regions with similar agricultural characteristics, but large enough regions to have enough observations within each to obtain robust parameter estimates. Clearly some of the regions extend across different types of agricultural land uses, but subdividing the world into smaller regions resulted in too few observations in some regions to get robust model estimates.

[25] The estimated parameter values (not shown) were used to make global cropland and pasture maps at 5 min spatial resolution in latitude by longitude for the BU-MODIS and GLC2000 data sets individually, and for the combined land cover data set (Figure 3). For each 5 min grid cell, x, y (latitude by longitude), we determined λ_j,x,y, which is the proportion of each of the satellite-derived land cover classes, j, within 5 min grid cell x, y. Using the estimates of α_j and β_j, we calculated cropland and pasture proportions in each 5 min grid cell, cf_x,y and pf_x,y respectively, using:

and

3.3. Comparison of the Performance of BU-MODIS, GLC2000, and Combined Data Sets

[26] We now present a comparison of the inventory agricultural land area to the predicted values for each data set from calibration Step 1 (from equations 6 and 7, aggregated from 5 min resolution to the administrative level) (Table 4). The cropland and pasture areas predicted using the combined satellite data set is better correlated to the inventory data compared to the models using either the BU-MODIS or GLC2000 data sets alone in every region except for a few exceptions (croplands in Africa/Middle East, and pastures in Europe and Russia and Australia/NZ). Further, neither the BU-MODIS nor the GLC2000 data set always performs better than the other; for example, in Africa and Middle East, GLC2000 predicts cropland much better than BU-MODIS, but the reverse is true for pastures. We next present a regional comparison of the inventory and predicted agricultural land areas from the combined data set (Figures 4a and Figures 4a and 4b). The most notable differences lie in the underestimation of pasture areas in Asia and Australia/New Zealand. We will discuss the significance of this underestimation in the next section.

[27] We now consider the spatial patterns of predicted agricultural land from the combined data set versus the individual data sets, using South American pastures as an example (Figure 5). In Figure 5a the inventory data shows large extent of pasture along the arc of deforestation, along coastal and southern Brazil, in Uruguay, and in the Patagonia region of Argentina. Figures 5b and 5c are the BU-MODIS and GLC2000 based pasture maps. Both do a reasonable job of reproducing the inventory data in Figure 5a. The most glaring difference is in the Nordeste (northeast) region of Brazil, where both data sets overestimate the distribution of pasture as compared to the inventory data in Figure 5a. The combined land cover data was able to best reproduce the inventory data, especially in the Brazilian Nordeste region and Patagonia (Figure 5d).

[28] Thus by using the combined data set we are able to capitalize on whichever satellite-based land cover data set is best suited to each region. If we were to use either the BU-MODIS or GLC2000 data sets alone, we would get reasonably good global results, but would lose accuracy in some regions. The use of the combined land cover data set especially yields marked improvements over Asia and South America (Table 4). Therefore we use the combined data set in the next step of this study.

3.4. Bootstrap Procedure to Estimate Uncertainty

[29] We further used a bootstrap technique in order to estimate the uncertainty in our parameter estimates. This procedure was applied at this stage for only the combined land cover data set. We performed 1000 bootstrap runs, where the entire census data was sampled with replacement each time, and reestimated our regression model using the combined satellite-based land cover data set (Sampling with replacement used the standard statistical procedure wherein the census data formed the population (of n_i values, the total number of administrative units), and for each of the 1000 sample sets we randomly selected n_i values from the population, with the sampled value being replaced back into the population. Each sampling outcome is therefore independent of the previous outcome; that is, every census data value has equal probability of being chosen during every sampling event, no matter whether it has been chosen before). This yielded a distribution of values for each parameter α_j and β_j, and we report the mean and 90% confidence intervals here (see Data Set S1).

[30] Using the 1000 estimates of α_j and β_j, we calculated 1000 estimates of cropland and pasture proportions in each 5 min grid cell, cf_x,y and pf_x,y respectively, using:

and

From these 1000 estimates, we calculated the mean, 5th percentile and 95th percentile values of the cropland and pasture values, represented as cf_x,y^mean, cf_x,y^5th%;ile, and cf_x,y^95th%;ile, and as pf_x,y^mean, pf_x,y^5th%;ile, and pf_x,y^95th%;ile, respectively (figures not shown from this stage of analysis).

3.5. Step 2: Adjusting the Predicted Cropland and Pasture Data to Match Inventory Data

[31] In this final step, we followed the methods of Ramankutty [2004], to adjust our spatially explicit predictions from Step 1 (the bootstrap model estimates using the combined data set) to match the inventory data at the administrative level where available. To do so, we first aggregated our 5 min resolution cropland and pasture data sets to the administrative level. We then compared them to the inventory data to derive a correction factor for each administrative unit. The correction factors were, however, constrained to be within a factor of 5 (i.e., to lie between 0.2 and 5.0) for administrative units that were considered outliers in the regression, thereby trusting the satellite-based land cover data more than the inventory data in those cases. Outliers were determined to be those administrative units with residuals (predicted cropland area from the Step 1 calibration procedure minus inventory cropland area) that were greater than 2 standard deviations from the mean. Correction factors were set to 1.0 for administrative units with missing data, thereby relying on the satellite-estimated spatial patterns from Step 1 in these units. We then applied Pycnophylactic Interpolation [Tobler, 1979] to the administrative level correction factors to obtain a smooth surface of correction factors at 5 min resolution (without this smoothing, artificial boundaries between administrative units might appear in the final product; note, however, that only the correction factors were smoothed, so any real boundaries in the original satellite data will remain). The spatial correction factors were then applied to our results from Step 1 to derive our final maps of cropland and pastures at 5 min resolution (Figure 6) and respective confidence intervals (Figure 7). The final equations for cropland and pasture proportions can be represented as

and,

where μcf_x,y and μpf_x,y are the spatially explicit correction factors for cropland and pasture respectively.

4.1. Total Global Area of Croplands and Pastures in 2000

[32] Our final results (Figure 6) indicate that there were 15.0 (95% confidence range of 12.2–17.1) million km² of cropland and 28.0 (95% confidence range of 23.6–30.0) million km² ofpasture in the world in the year 2000. This compares to 15.3 million km² of cropland and 34.4 million km² of pasture reported by the FAOSTAT database. Thus we predict significantly lower extent of pasture (by 6.1 million km² or ∼18% lower) than reported by FAO. Our own inventory data reports 15.0 million km² of cropland and 31.5 million km² of pasture. Thus our inventory data for pasture is already lower than FAO statistics; this difference was explained earlier in section 2.2. Our cropland extent for the year 2000 of 15.3 million km² is lower than the 18.0 million km² of cropland in 1992 from our previous study [Ramankutty and Foley, 1998]. This does not mean that cropland extent has decreased from 1992 to 2000; rather the difference is a result of changes in methodology: We plan to reconstruct a consistent historical time series; until then a comparison of our 1992 and 2000 data sets is precluded.

[33] Our final predicted pasture area is even lower than our inventory data, especially in Asia and Australia/New Zealand (as already evidenced in Figures 4a and 4b). We anticipated that this problem would be overcome in Step 2, when we scaled our spatial cropland and pasture data to match our inventory data, but despite this our final predicted extent of pasture differs significantly from our inventory data. This is because we did not allow pixels with predicted agriculture proportions of 0% in Step 1 to be scaled, and each pixel was limited to have a maximum agriculture proportion of 100%. The biggest difference is in China, where the inventory reports 4 million km² of pasture, but we estimate a final area of only 2.9 million km². Our estimate, however, is similar to the 2.6 million km² of grassland reported by Verburg et al. [2000]. Verburg and van Keulen [1999] discuss that much nontraditional grazing occurs in China, with goats grazing along roads, dry river beds, and in fields after harvest, and pigs in back yards; it is not clear how these landscapes should be characterized. It is also likely that grazed forestland and semiarid land are included under the pasture category in the Chinese inventory. Other big differences in pasture area arise in Australia (inventory area of 3.2 million km² versus our prediction of 2.7 million km²), Mongolia (inventory of 1.3 million km² versus our estimate of 0.9 million km²), Mauritania (inventory of 0.4 million km² versus our 0.1 million km²), Iran (inventory of 0.9 million km² versus our 0.6 million km²), and the United States (inventory of 2.3 million km² versus our 2.1 million km²). It is interesting to note that all of these countries have significant amount of semiarid land. One potential explanation could be the variability in grazing associated with the interannual changes in climate that can be quite extreme in semiarid regions; the satellite data set only captures a single year of land cover condition, which can vary from year to year. For croplands, while the global totals agree, there are compensating national level differences but these differences are less significant compared to the differences in pasture areas.

[34] Our final predicted cropland extent of 15.0 million km² in year 2000 amounts to roughly 12% of the global land area (excluding Greenland and Antarctica), and pasture extent of 28.0 million km² amounts to 22% of global land area. Thus humans are using 34% of the global land area for their agricultural needs.

4.2. Geographic Distribution of Croplands and Pastures in Year 2000

[35] We analyzed the distribution of agricultural regions by 14 different regions of the world (Table 5). The greatest proportion of croplands in the world are found in South Asia, Europe, Southeast Asia, and United States, east of the Mississippi, while the greatest proportion of pastures are found in Argentina, Uruguay and Chile, Pacific developed countries, China, Mexico and Central America, United States, west of the Mississippi, and tropical Africa. The smallest proportion of croplands are found in Canada, the Pacific developed Countries, and northern South America, while the smallest proportion of pastures are found in Southeast Asia, Canada, and United States, east of the Mississippi.

[36] We also examined which potential natural vegetation types of the world have been most affected by agriculture (Table 6). We overlaid our agricultural maps over the global map of potential natural vegetation developed by Ramankutty and Foley [1999]. We find that croplands have mostly replaced temperate deciduous forests (in Europe and eastern United States), and tropical deciduous forests (in South Asia), while, pastures have mostly replaced grasslands, savannas, and shrubland. Roughly 30% of temperate deciduous forests have been converted to cropland, while 50% of grasslands have been converted to pasture. However, this global picture varies regionally (Table 7). While forests have been cleared for croplands predominantly in Asia, a substantial amount of savanna and grasslands have been converted to croplands in North America, Africa, and the Former Soviet Union. Also, a significant amount of forests in South America have been cleared for pastures, even though globally most pastures have replaced savanna/grasslands.

[37] Next, we examined the amount of spatial overlap between croplands and pastures (Figure 8). Our analysis shows that cropland and pastures are distinct biomes over much of our planet's land surface. The regions of the world with significant overlap lie along the western edge of cultivation in the midwestern United States and in Texas, northeast Brazil, in parts of West Africa, eastern China, Europe, and the former Soviet Union, scattered regions elsewhere. This is not to say that grazing does not occur in cultivated regions of the world; many regions of the world are characterized by multifunctional agricultural lands, subject to different uses during different parts of the year (e.g., grazing occurs following the harvest of a crop) [Reenberg and Fog, 1995]. Therefore this lack of overlap partly reflects the inability of global monitoring systems, including satellite data and agriculture inventory data, to characterize multiple uses, and land is often classified as a single category. Our final estimates likely underestimate the real overlap between cultivation and grazing, especially the multiple uses that occur within a year.

[38] We also investigated the frequency distribution of croplands and pastures globally, and across the 14 different regions of the world. We calculated probability distributions using all pixels with nonzero cropland or pasture values, and estimated what proportion of total cropland or pasture area can be attributed to different categories of cropland or pasture proportions (Figure 9). Globally, we find that a large proportion of the total cropland area comes from land that is between 60% and 80% cultivated. This pattern is observed in most regions of the world, except in northern South America, tropical Africa, and North Africa and Middle East (where larger number of pixels with low proportion of cultivation contribute most to the total cropland area, although in the latter two regions, there are a few pockets of greater than 90% cropland). With pastures, the total area is dominated by pixels with greater than 90% pasture. This pattern is true over most regions of the world, but is reversed in the eastern United States, Europe, and Southeast Asia (and South Asia, to a lesser extent). However, it is to be noted that the proportion of pasture in a grid cell provides no information on the grazing intensity in that grid cell; an area of grassland with very low stocking density of livestock would have the same proportion of pasture as one with very high stocking densities.

4.3. Evaluation Against Other Independent Data Sets

[39] While there are no consistent, global spatial data sets of agriculture in the year 2000 to evaluate our products against, there are numerous regional products against which we can compare our global products (see Figures S3–S12). While many of these regional data are not for the year 2000 (and are sometimes a decade older), they nevertheless provide an independent measure of the large-scale spatial patterns of agriculture in these regions. These data were not always available in a consistent digital format; they were often in vector format that is difficult to quantitatively compare to our raster data and sometimes only available as images in publications. Therefore rather than making quantitative comparisons we present detailed regional maps from our data set compared to these independent regional maps, as online supplements (Figures S3–S12).

[40] Our regional comparisons are mostly for croplands because there are few spatial data sets depicting pastures (likely because it is difficult to distinguish between natural and grazed grasslands). Visual comparison (Figures S3–S12) suggests that our distribution of croplands is reasonable in North America (United States, Canada, and Mexico), and so is the distribution of pastures in the United States (except for noticeable differences in northeastern Texas, and eastern Oklahoma, arising from lower pasture census values). In the Amazon Basin, our distribution of both croplands and pastures shows similar geographic patterns but greater intensity compared to Cardille et al. [2002]. This may be a result of our data being representative of 2000, while Cardille et al. [2002] data is for the mid 1990s (changes in the southern Amazon are rapid) or because of differences in statistical methods with the regression tree method of Cardille et al. [2002] not able to deal well with extremes. In China also, our geographic patterns of croplands matches well with that of Liu et al. [2005], but our intensity is lower (the differences are greater than can be explained by the different spatial resolutions of the two data sets). In West Africa, our cropland distribution from this study is a significant improvement over our earlier work [Ramankutty and Foley, 1998], but consistent with our more recent effort [Ramankutty, 2004]. Our distribution of croplands compares reasonably well with Africover data in East Africa [Latham, 2001] (http://www.africover.org/) but is not a significant improvement over our earlier effort, while in South Africa, our estimated patterns are an improvement over our previous effort but of much lower intensity. In Europe, our data set of croplands compares well to the CORINE land cover database. We seem to underestimate croplands in eastern Australia, while our distribution of pastures is significantly underestimated in the Northern Territory. The problem with pastures in Australia actually emerges from Step 2, with the census data in the Northern Territory suggesting very little pastures; our predicted pattern from Step 1 compares better to the data from the Australian Natural Resources Atlas.