1. Introduction
People in sub-Saharan Africa experience some of the most severe living conditions on the planet. They are ravaged by civil wars, droughts, famines, and rapid population growth coupled with increasing death rates from the AIDS epidemic. Indeed, the Food and Agriculture Organization (FAO)'s The State of Food Insecurity in the World 2000 report identified sub-Saharan Africa as one of the most food-deprived regions of the planet: 34% of the population there is estimated to be undernourished, with the depth of undernourishment being greater than 300 kcal per person per day in 46% of the countries (FAO 2000).
To meet the food demands of the growing population, croplands in sub-Saharan Africa have rapidly expanded over the last 50 yr. According to FAO (FAO 2004), cropland area increased from ∼120 Mha in 1961 to 140 Mha in 1980, and 162 Mha in 2000, representing a 35% increase in 40 yr. Such expansion, while necessary to meet food demands, can have detrimental environmental consequences in the long term. For example, it has been argued that the persistent Sahelian drought since the 1960s might be related to vegetation changes (Charney 1975; Wang and Eltahir 2000; Zeng et al. 1999), although this is still contentious (Giannini et al. 2003; Taylor et al. 2002). Land-use change can also have severe consequences for water resources of the region (Coe and Foley 2001), and has been implicated in the soil degradation seen in the Sahel (Oldeman et al. 1991; Scherr 1999), although this is disputed as well (Niemeijer and Mazzucato 2002).
To evaluate the potential environmental consequences of land-use and land-cover change, numerical models of terrestrial ecological and hydrological systems are being developed by several groups around the world (e.g., Bonan 1996; Coe 1998; Foley et al. 1996; Melillo et al. 1993; Sellers et al. 1996; Sitch et al. 2003; Vorosmarty et al. 1989). These models require the specification of land surface properties such as surface albedo, leaf area index, vegetation height, etc., that are normally derived from land-cover maps. However, as the geospatial ecosystem models almost always represent the land surface in a spatially explicit, gridded framework, the input datasets also need to be in a compatible format. Therefore, the critical need to develop spatially explicit datasets of the world's land cover, as modified by human activities, has been recognized (Lambin et al. 1999). In addition to ecosystem modeling activities, spatially explicit land-use data are also being sought after by the international policy community to enable more informative policy analysis (see online, e.g., at http://www.spatial-info.org/spatial-info/pages/ifpri.htm, by rapid land-cover change assessments (see online at http://www.geo.ucl.ac.be/LUCC/MA_Project/MA.html, and in various other applications such as land-degradation analysis, nutrient balance analysis, and farming systems research (George et al. 2004).
Spatially explicit datasets of the world's land use and land cover are now being developed using remotely sensed satellite data (DeFries et al. 2000; DeFries and Townshend 1994; Friedl et al. 2002; Skole and Tucker 1993). Recently, Ramankutty and Foley (Ramankutty and Foley 1998, hereafter RF98; Ramankutty and Foley 1999) used a combination of moderate-resolution satellite data and census data to derive the spatial distribution of contemporary and historical global cultivated lands. This global dataset is being widely used to examine the environmental consequences of cropland change around the world (Matthews et al. 2003; McGuire et al. 2001; Myhre and Myhre 2003; Tegen et al. 2004; Yang et al. 2003). However, the RF98 croplands dataset for 1992 compares poorly to an independent Landsat-derived dataset for the Sahel (Figure 1). As part of the Famine Early Warning System (FEWS), the Earth Resources Observation Systems (EROS) Data Center used Landsat data [Multispectral Scanner (MSS) composite from 1986–88] to create a dataset of cropland-use intensity1 (CUI, henceforth referred to as FEWS CUI), depicting five categories of cropland use: 0%–5%, 5%–30%, 30%–50%, 50%–70%, and 70%–100% cultivation [data can be downloaded from the Africa Data Dissemination Service (ADDS) online at http://edcsnw4.cr.usgs.gov/adds/index.php]. Both the RF98 and FEWS CUI dataset were qualitatively reviewed by regional experts at a workshop, “Trajectories of Land Change in Sub-Saharan Africa” (see online at http://www.indiana.edu/act/focus1/tlc/africa/Af_index.htm, and the FEWS CUI dataset was judged to be a good representation of cultivation in the Sahel, while the RF98 dataset was evaluated to be inaccurate in the Sahel and over the rest of West Africa. The RF98 dataset fails to represent cultivation in southern Mali, southern Burkina Faso, southern Niger (Maradi district), and southern Chad. RF98 shows a strip of croplands running east from Senegal and Gambia to Uganda along the Sahel, and along coastal West Africa, with little cultivation in between; in fact, Nigeria, with the highest population in the region, shows very little cultivation. Clearly, an improved representation of croplands in West Africa would be desirable.
2. Charting the way forward from a review of existing datasets
In this section, I review the datasets that are currently available either from remotely sensed or ground-based methods, which can be used to derive the spatial patterns of croplands across West Africa. I then discuss the specific approach taken by this study.
2.1. Satellite data
The best technology currently available to characterize land cover is data from spaceborne sensors. Landsat data, with fairly high resolution [79 m for Multispectral Scanner (MSS) and 30 m for Thematic Mapper (TM) and Enhanced Thematic Mapper Plus (ETM+)], have become the “workhorse” for developing regional land-use/cover datasets (see online at http://landsat7.usgs.gov/ for more information). The Landsat sensors measure reflectances from the land and ocean in multiple spectral bands of visible and infrared radiation (four bands for MSS and seven bands for TM & ETM+). These reflectances, or some indices derived from them, form the different satellite images, which are then interpreted using image classification algorithms to derive land-cover characteristics (Lillesand and Kiefer 2000).
As described earlier, the FEWS CUI dataset derived from Landsat MSS was judged by regional experts to be a good representation of cultivation in the Sahel. However, it has only partial coverage of West Africa; wall-to-wall Landsat-based land-cover classification for West Africa is as yet unavailable and is an expensive proposition that requires a massive undertaking. The Africover project (see online at http://www.africover.org/) began such a task in 1994, but only eastern Africa has been completed to date.
Wall-to-wall global coverage is, however, available from moderate-resolution satellite-borne sensors—AVHRR (Advanced Very High Resolution Radiometer), and more recently, MODIS (Moderate Resolution Imaging Spectroradiometer) and VEGETATION. AVHRR is an instrument on the National Ocean and Atmospheric Administration's polar-orbiting satellite. MODIS is on board the National Aeronautics and Space Administration's Terra satellite. VEGETATION is an instrument on the European Systeme Pour l'observation de la Terre (SPOT) satellite. All of these instruments also measure reflectances in multiple spectral bands and used them to derive land-cover characteristics (Lillesand and Kiefer 2000).
Global land-cover classification datasets have been derived from these various instruments for application to land-cover change research (Friedl et al. 2002; Hansen et al. 2000; Loveland et al. 2000; Mayaux et al. 2004). A comparison of these moderate-resolution land-cover datasets in West Africa to the high-resolution Landsat-derived FEWS CUI dataset shows a poor depiction of the spatial distribution of cultivated lands (Figure 2). While the 2000 Global Landcover Classification (GLC2000) dataset (Mayaux et al. 2004) does show more croplands in West Africa compared to the other three products, it also statistically compares poorly to the FEWS CUI dataset (statistical results not shown). Also note that the AVHRR DISCover dataset was used by RF98 to generate their croplands dataset; clearly, that satellite-derived product is not very accurate in West Africa.
It is understandable that moderate-resolution sensors have difficulty in characterizing croplands in West Africa. First, these datasets are continental/global products, and thus their classification algorithms were not fine-tuned for West Africa, and moreover, croplands were only two of the several categories under consideration. Second, cultivation in West Africa is of low intensity and part of a heterogeneous landscape and, therefore, “land-cover classification” datasets at 1-km resolution are unable to depict cultivation. A “continuous fields” approach (DeFries et al. 2000) may be necessary to depict low-intensity cultivated lands. Moreover, such a continuous fields representation would be beneficial for ecosystem modeling (DeFries et al. 1995). Finally, the satellite data seem to have a difficult time characterizing agricultural lands in the humid zones of coastal West Africa, probably because of the lack of sufficient images due to cloud cover contamination; strategies to deal with this need to be developed, including ground-based monitoring.
2.2. Cropland inventory data
The Food and Agriculture Organization of the United Nations compiles data annually on various agricultural production statistics. The data are collected, primarily with the use of questionnaires sent to member nations. In addition, FAO also collects data from national and international publications (including agricultural yearbooks, periodicals, reports issued by boards and associations, etc.) and country visits and discussion with national experts. The compiled data are further scrutinized for quality and consistency through cross-referencing. Finally, missing data are filled using various methods, including making rough estimates based on the little information that can be gleaned from household surveys, trade reports, food balance sheets, etc., and also sometimes based on the trend from previous years as well as observations of the behavior of the commodity in neighboring countries with similar agroeconomic conditions and food habits. (A detailed discussion of the methodology can be obtained online at http://www.fao.org/WAICENT/FAOINFO/ECONOMIC/ESS/index_en.asp.)
As described above, the FAO data are a compilation of agricultural statistics reported by member nations, rather than more direct observations. Indeed, such data can often be skewed because of incentives to under- or overreport cropland statistics, or because poor nations do not have the resources or infrastructure to conduct rigorous surveys. For example, it is now well established that agricultural statistics from the State Statistical Bureau of China are underestimated by roughly 30% (Crook 1993; Heilig 1999; Xiao et al. 2003). West Africa is particularly hampered by poor infrastructure and lack of resources, as well as frequently being ravaged by civil wars, and the quality of the statistical data are therefore even more suspect.
In spite of all the shortcomings of the FAO statistics, they are nevertheless the standard data source for cross-national and global analysis of the status of agriculture, and indeed the only source available for large-scale studies (Wood et al. 2000; WRI 2004). The FAO online database, called FAOSTAT (http://apps.fao.org, contains annual agricultural statistics from 1961 to 2003. For the nations in West Africa within the domain of this study, I extracted FAOSTAT data for 1992 on “arable land and permanent crops,” which is the adopted definition of croplands for this study (defined in greater detail later). I also extracted national statistics on “harvested area” in 1992 for all the primary crop commodities in FAOSTAT. While FAOSTAT provides a standard estimate of cropland area across West Africa, the national level of aggregation is a major stumbling block for the purpose of mapping croplands across the region.
Another inventory of croplands that became recently available, and covering the entire West African region, is statistical data at the subnational level from the AgroMAPS project (online at http://www.fao.org/ag/agl/agll/agromaps/default.stm; George et al. 2004). AgroMAPS, a consortium comprised of the FAO, the International Food Policy Research Institute (IFPRI), and the Center for Sustainability and the Global Environment (SAGE), is compiling subnational agricultural census data on crop-harvested area, yield, and production from major agricultural nations of the world. The first CD-ROM with Africa data is now available, which contains data from the first administrative level below the nation for most of the West African nations (see Table 1 for details on data used in this study). All the caveats discussed for the FAOSTAT data also apply to the AgroMAPS dataset, because it uses similar sources of information, that is, reporting by member nations, national yearbooks, etc. AgroMAPS provides agricultural statistics over subnational administrative units and therefore has improved spatial resolution compared to FAOSTAT (see the administrative outlines in Figure 3, top). However, it is also not a geographically explicit, gridded dataset and needs to be disaggregated to be useful.
A second source of subnational cropland inventory data is from the West African Long-Term Perspective Studies (WALTPS), which undertook a survey of population, agriculture, and urbanization in the region (Brunner et al. 1995). The WALTPS study also provides subnational inventory data on major crops for all the nations in West Africa, representative of the year 1990.
2.3. The way forward
As discussed earlier, wall-to-wall land-cover datasets derived from moderate-resolution satellite-borne sensors represent cultivation in the Sahel very poorly. Direct observations of croplands from the ground using rigorous surveys are not available at the scale of West Africa (although the WALTPS study is a good example of a broad survey of the region). The only regional-scale datasets available are the AgroMAPS and WALTPS inventory data at the subnational administrative unit level. However, the inventory data, as described in the previous section, are not geographically explicit and need to be disaggregated. In this study, I used population density data to disaggregate the cropland inventory dataset.
I synthesized three different sources of information—the Landsat-based FEWS CUI data, the AgroMAPS/WALTPS subnational cropland inventory data, and a spatially explicit population density data—to derive a spatial dataset of croplands for West Africa (Figure 4). Here, the three steps in the synthesis process are briefly described (more details are in sections 3 and 4).
A linear model was established between the population density dataset and the cropland inventory dataset. This model was used to simulate a spatially explicit map of croplands for West Africa, using population density as an independent variable. The simulated distribution of croplands was evaluated against the independent FEWS CUI dataset over the Sahel.
For the categorical FEWS CUI dataset, “best-guess” single values for each category were determined by calibrating it against the cropland inventory data.
The linear-model-simulated cropland distribution over West Africa was merged with the FEWS CUI dataset over the Sahel. The merged product was further adjusted to match the cropland inventory data over each administrative unit.
The new dataset covers the West African domain stretching from the west coast (−17.5°W) to the Sudan in the east, and from the coastal nations bordering the Gulf of Guinea to 21.3°N latitude in the north (the exact domain can be seen by the extent of the population density data in Figure 3, bottom). This new dataset has a resolution of 5 min in latitude by longitude, and has a continuous description of the land cover depicting the fraction of cultivated land within each 5-min grid cell. Such a continuous description of the landscape is preferable to a categorical description (such as FEWS) or a Boolean description (such as the land-cover classification products) because it can be directly used within a numerical ecosystem or climate model to examine the environmental consequences of cultivation. However, I hasten to add a word of caution that this dataset should not be used in studies that try to understand the drivers of cropland change; as population density is already used to derive this dataset, spurious results will be obtained. This dataset is simply meant to be a provisional product for use by ecosystem modelers; in the future improved datasets will hopefully be derived from remote sensing or other direct observational methods.
The definition of croplands in this study follows the FAO definition of arable land [“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 yr). The abandoned land resulting from shifting cultivation is not included in this category”] and permanent crops (“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”).
There are many potential problems associated with the interpretation of the definition of croplands, as well as with the measurement of cultivated areas. Here I list several potential problems, which also form major caveats for our final product (see online at http://www.fao.org/es/ess/rmcrops.asp for more details).
Cadastral mapping of crop area often includes ditches, headlands, and other noncropped areas, as opposed to the actual areas where crops are sown (the terms “gross area” and “net area” are often used to distinguish the two). Furthermore, crops may not actually be harvested over the entire sown area because of loss due to weather, pests, etc. Therefore, a further distinction needs to be made between “sown area” and “harvested area.”
The FAO definition allows for croplands that are temporarily used for grazing and temporarily fallow to be included. But changes in the cropping–grazing rotation cycle as well as in the cropping–fallow cycle can confound the statistics.
Another issue is related to the scale at which cultivation is practiced. In some countries, land holdings are used as the unit of enumeration and a certain minimum size is used as a cutoff for enumeration. Therefore, small-scale subsistence agriculture, kitchen gardens, etc., are often not enumerated.
Often two associated crops are intercropped; that is, they are interplanted (e.g., maize and beans). FAO recommends that estimates be made for each crop of the area that would be occupied if each crop were grown alone; however, this is a difficult proposition, and the statistics are likely prone to errors. A similar issue arises with crop rotations that happen within a single year (multicropping).
3. Modeling the spatial distribution of croplands
The AgroMAPS/WALTPS cropland inventory data have an estimate of crop-harvested area by subnational administrative unit. The inventory data was first adjusted (by uniformly scaling the values) so that the national totals match the national statistics from FAOSTAT (see the appendix). While FAOSTAT also reports data on cropland area (the sum of arable lands and permanent crops), only crop-harvested area data are available from AgroMAPS/WALPTS. To derive subnational cropland area data, I disaggregated the FAOSTAT national cropland area using the subnational proportions of total harvested area from AgroMAPS/WALTPS (Figure 3; see the appendix for more details). Therefore, an assumption is made that the total harvested area is roughly proportional to cropland area.
To be useful for application within numerical models or for spatially explicit analysis, the cropland inventory data need to be disaggregated (i.e., gridded) within each administrative unit using a geographically explicit proxy. In this study, I tested the use of population density as a proxy2. A dataset of population density for the African continent for 1990 was obtained from the United Nations Environment Programme (UNED/GRID-Sioux Falls) (available online at http://www.na.unep.net/; Deichmann 1997). This product was derived by compiling subnational population census data and gridding them based on accessibility as derived from transportation network and the location of cities (Deichmann 1997). A visual comparison of this dataset to both the FEWS CUI dataset and to the cropland inventory dataset showed that population density may be a reasonable proxy for cultivation in West Africa (Figure 3). In this study, I reconstructed the spatial distribution of croplands in West Africa by calibrating the population density dataset against the AgroMAPS/WALPTS cropland inventory data. This assumes that food is produced close to where it is consumed. This is likely true for the most part in West Africa characterized by subsistence agriculture; however, with the increasing importance of urbanization, cash cropping for export, and mechanization of agricultural production replacing labor, this assumption would be invalid in many places (Cour 2001). Nevertheless, as will be shown later, population density is well correlated with cropland area over the large scale.
The population density dataset is available at 2.5-min resolution for the year 1990; this was aggregated to a 5-min resolution. The FEWS CUI dataset is provided as a vector dataset. This was converted to a raster grid with a sampling cell size of 5-min resolution to match the resolution of the population density dataset. I also developed a land-water mask at 5-min resolution that is consistent with our existing global land-cover products (RF98), as well as a mask of built-up areas where the area of croplands is set to zero. The mask of built-up areas was created using the Nighttime Lights of the World dataset (Elvidge et al. 1997; available online at http://dmsp.ngdc.noaa.gov/html/download.html). A comparison of the Nighttime Lights data to the polygons of settlements from CIESIN (CIESIN 2004) showed that setting all light values of more than 10 to built-up yielded the best match. It should be noted that this technique is used to try to mask out only the largest urban settlements.
To disaggregate the AgroMAPS/WALTPS inventory data within each administrative unit, I developed a linear regression model relating population density to the inventory data (Figure 4). As the census data are collected by administrative unit, while the population density dataset is spatially explicit and has many values within each unit, the linear model necessarily had to be constructed at the administrative unit level. To do this, a “typical” estimate of population density for each administrative unit was required to build the linear model. Because the population density dataset is heavily skewed, with a few values of very high-population density, estimating an average population density for each administrative unit was not satisfactory. I therefore chose the median population density within each administrative unit as the representative value. (I also ignored anomalously high population densities—values above 594 people per square kilometer representing only 0.1% of the data. Furthermore, medians were calculated only if a minimum of 10 samples were available for each administrative unit; otherwise, the administrative unit was ignored while building the linear model. Also note that this typical median estimate was used only to build the linear model. When the model was applied to the spatially explicit population density data, all values of population density were used, excepting the upper threshold of 594 people per square kilometer, as explained in more detail below).
For West Africa, an independent measure of the spatial distribution of croplands is available from the FEWS CUI dataset. I compared the simulated croplands and the earlier RF98 dataset against the FEWS CUI dataset using a contingency table (not shown; final results shown in Table 3). The fraction of agreement of the simulated dataset to FEWS CUI is slightly smaller than that obtained using the RF98 dataset. However, this agreement is more likely due to chance as indicated by the kappa statistic (Monserud 1990). The resulting kappa value of 0.19 for the new dataset is superior to the kappa value of 0.13 between RF98 and FEWS CUI. Indeed, visual inspection shows that the RF98 dataset and FEWS CUI dataset match over central Mali and Burkina Faso, where the east–west strip in RF98 overlaps by chance with FEWS CUI. Even the kappa statistic is not a sufficient estimate because it measures cell-by-cell agreement and ignores agreement at coarser scales. Indeed, when the datasets were aggregated to a half-degree resolution, the linear model was clearly superior (Table 3; for the FEWS CUI dataset, optimum values were determined for each CUI category before aggregation, as described in the next section). I also aggregated the datasets by subnational administrative units and found the new dataset to be an improvement over RF98 (Table 3). Finally, the new simulated data compares very well to the cropland inventory data, whereas RF98 shows very poor agreement (however, this is not a fair comparison because the linear model was built by training against the census data and is thus not independent of it).
4. Integrating the simulated cropland distribution with FEWS CUI and census data
In a final step, I integrated the linear-model-simulated cropland distribution with the FEWS CUI dataset and also adjusted the data to exactly match the cropland inventory dataset (Figure 4). In the previous step, I used the FEWS CUI dataset as an independent dataset to evaluate the quality of the linear-model simulation. Having done that, as the FEWS CUI dataset has a good representation of croplands in the Sahel, it could be valuably integrated into the final product. Before doing so, however, the categorical CUI dataset first needed to be assigned best-guess single values for each category. To determine an optimum single value for each category, the CUI dataset was calibrated against the cropland inventory data. I picked a single cropland intensity value within the permissible range for each CUI category that provided the best match between the CUI dataset and the cropland inventory dataset (figure not shown); the best-guess values occurred at the lower end of the range for all five CUI categories. I calculated a new cropland intensity map as the average of the linear-model simulation and the calibrated FEWS CUI data over the Sahel.
Next, I aggregated the integrated dataset over the same administrative units that are present in the cropland inventory data. By comparing the two, a correction factor was estimated for each administrative unit. The correction factor was set to unity for the five outliers identified during the calibration of population density to inventory data, and it was limited to the range of 0.9–1.1 for the Sahel because of higher confidence in the FEWS CUI dataset. I then used pycnophylactic interpolation (Tobler 1979) to convert the correction factor at the administrative level to a spatial, 5-min resolution surface. The spatially resolved correction factor was applied to the croplands dataset to obtain the final cropland intensity map, which now matches the cropland inventory data (Figure 6, bottom).
5. Discussion and conclusions
I integrated population density data, agricultural census/inventory data, and a Landsat-derived cultivation intensity dataset over the Sahel to derive a new dataset of croplands for West Africa (Figure 6, bottom). The new dataset appears to be a significant improvement over the RF98 dataset. A statistical comparison of the linear-model-simulated result (which calibrated a map of population density against the inventory data) and the RF98 dataset to the independent FEWS CUI dataset over the Sahel (Table 3) shows marked improvement. The final dataset of croplands over West Africa was a blend of the linear-model simulation and the FEWS CUI dataset, adjusted to match the cropland inventory data.
I realize that using population density as a sole predictor of cropland distribution can raise serious objections. However, this study was an attempt to derive a provisional product for use in numerical models and was not meant to develop a methodology or dataset for wide application. Moreover, additional independent variables—climate (growing degree days, ratio of actual to potential evapotranspiration), soil quality (soil carbon density, soil pH), satellite-derived land-cover dataset, GLC2000 (Mayaux et al. 2004), and University of Maryland (UMD) fractional tree cover (DeFries et al. 2000)—were tested for predictive power against the cropland inventory data using both linear models as well as regression tree models. In all cases, very little improvement in explained variance was gained with the use of additional variables and, moreover, the additional variables did not yield statistically significant results. It appears as if population density is indeed a very good proxy for the distribution of croplands in West Africa, at least on a coarse scale. This also seems reasonable intuitively considering the low-intensity subsistence type of cultivation practiced over much of West Africa. However, with the increasing importance of export agriculture, urbanization, and mechanization, this assumption will not remain valid for many parts of West Africa. Furthermore, while population density works as a good proxy in West Africa, this is not necessarily the case over the rest of the world. In general, the pattern of croplands is determined by a complex mix of biophysical, socioeconomic, and political factors. Therefore, applying the methods from this paper to other parts of the world is not recommended; direct observations from remote sensing or other ground-based survey methods should always be the first choice.
The new croplands dataset shows the highest intensities of cultivation in Nigeria, in the arid Sudano–Sahel region surrounding Kano (growing millet, sorghum, cowpeas, and maize), and along the southern coast (with root crops and tree crops). High cultivation intensities can also be found in the groundnut- and millet-growing regions of Senegal, parts of Cote d'Ivoire (growing cocoa beans, coffee, maize, and rice), Togo (maize, sorghum, cotton), portions of Ghana (cocoa beans, maize, and cassava), in northern Cameroon (sorghum, maize), and parts of southern Cameroon (cocoa beans). In general, millet and sorghum are the major crops grown in the arid Sahel, while maize, rice, cassava, and tree crops are grown in the coastal regions.
This new dataset of cropland distribution for West Africa is at 5-min resolution in latitude by longitude and is representative of the early 1990s. It will be made digitally available on our Web site (online at http://www.sage.wisc.edu/) upon publication of this paper.
Acknowledgments
I thank Bill McConnell and Mike Mortimore for very constructive reviews and excellent suggestions. The manuscript is much improved with their input. Bill also provided me with the WALTPS database that was used to fill gaps in the cropland inventory data. I owe many thanks to Jeanine Rhemtulla for help with the statistics. Although the final paper uses very little statistics, more complex statistical procedures were tried and rejected, thus providing me with a sense that the eventual simplicity is not due to laziness in application of statistics. Dr. Michael T. Coe provided much needed counseling during the many twists and turns this paper took along its way. He was a great sounding board for much of this work. Hubert George from the Food and Agricultural Organization (FAO) provided me with the Africa data from AgroMAPS long before it came available online. My collaboration with his group has been extremely valuable. I would like to acknowledge my gratitude to the Land-Use and Land-Cover Change (LUCC) Focus 1 office at Indiana University (led by Emilio Moran and Bill McConnell) for inviting me to participate in a regional workshop on land change in Africa; thanks also to the various scientists who attended that workshop and provided me with feedback on my work. Dr. Ademola K. Braimoh gave me the FORMECU manuscript from which I derived the statistics for Nigeria. Kristin Tenwinkel and Tristan Wagner spent many hours photocopying statistical data and entering them into spreadsheets—they have been invaluable to my research. This research was funded by NASA LCLUC Grant NAG5-11085, MOD. 04.
References
Bonan, G. B. 1996. A land surface model (LSM version 1.0) for ecological, hydrological, and atmospheric studies: Technical description and user's guide. NCAR Tech. Note NCAR/TN-417+STR, National Center for Atmospheric Research, 150 pp.
Brunner, J., N. Henninger, U. Deichmann, and B. Ninnin. 1995. West Africa Long-Term Perspective Study (WALTPS). Database and User's Guide, World Resources Institute and Club du Sahel, Washington, D.C., and Paris, France, 26 pp.
Charney, J. G. 1975. Dynamics of deserts and drought in the Sahel. Quart. J. Roy. Meteor. Soc 101:193–202.
CIESIN 2004. Global Rural-Urban Mapping Project (GRUMP) alpha development data—Africa and Asia urban masks. CIESIN, IFPRI, and CIAT. [Available online at http://www.ifpri.org/data/gs_grump.htm.].
Coe, M. T. 1998. A linked global model of terrestrial hydrologic processes: Simulation of modern rivers, lakes, and wetlands. J. Geophys. Res 103:(D8). 8885–8899.
Coe, M. T. and J. Foley. 2001. Human and natural impacts on the water resources of the Lake Chad Basin. J. Geophys. Res 106:(D4). 3349–3356.
Cour, J. M. 2001. The Sahel in West Africa: Countries in transition to a full market economy. Global Environ. Change 11A:31–47.
Crook, F. 1993. Underreporting of China's cultivated land area: Implications for world agricultural trade. China International Agriculture and Trade Report, Situation and Outlook Series RS-93-4, U.S. Department of Agriculture, 33–39.
Defries, R. S. and J. R. G. Townshend. 1994. NDVI-derived land cover classifications at a global scale. Int. J. Remote Sens 15:3567–3586.
Defries, R. S. and Coauthors. 1995. Mapping the land surface for global atmosphere–biosphere models—Toward continuous distributions of vegetation's functional properties. J. Geophys. Res 100 D10:20867–20882.
Defries, R. S., M. Hansen, J. Townshend, A. Janetos, and T. Loveland. 2000. A new global data set of percent tree cover derived from remote sensing. Global Change Biol 6:247–254.
Deichmann, U. cited 1997. African population database documentation. National Center for Geographic Information and Analysis. [Available online at http://www.na.unep.net/globalpop/africa/index.php3.].
Elvidge, C. D., K. E. Baugh, E. A. Kihn, H. W. Kroehl, and E. R. Davis. 1997. Mapping of city lights using DMSP Operational Linescan System data. Photo. Eng. Remote Sens 63:727–734.
FAO 2000. The state of food insecurity in the world. Food and Agriculture Organization of the United Nations, Rome, Italy, 21 pp.
FAO cited 2004. FAOSTAT, Food and Agriculture Organization of the United Nations. [Available online at http://apps.fao.org.].
Foley, J. A., C. I. Prentice, N. Ramankutty, S. Levis, D. Pollard, S. Sitch, and A. Haxeltine. 1996. An integrated biosphere model of land surface processes, terrestrial carbon balance, and vegetation dynamics. Global Biogeochem. Cycles 10:603–628.
FORMECU 1995. An assessment of land use and vegetation changes in Nigeria between 1978–1993. Geomatis, 75 pp.
Friedl, M. A. and Coauthors. 2002. Global land cover mapping from MODIS: Algorithms and early results. Remote Sens. Environ 83:(1–2). 287–302.
George, H., F. Nachtergaele, and P. Koohafkan. 2004. Agro-MAPS: A global spatial database of sub-national agricultural land-use statistics. Food and Agriculture Organization, Rome, Italy, 16 pp.
Giannini, A., R. Saravanan, and P. Chang. 2003. Oceanic forcing of Sahel rainfall on interannual to interdecadal time scales. Science 302:1027–1030.
Hansen, M., R. DeFries, J. R. G. Townshend, and R. Sohlberg. 2000. Global land cover classification at 1km resolution using a decision tree classifier. Int. J. Remote Sens 21:1331–1365.
Heilig, G. K. cited 1999. Can China feed itself?—A system for evaluation of policy options. International Institute for Applied Systems Analysis. [Available online at http://www.iiasa.ac.at/Research/LUC/ChinaFood/index.htm].
Lambin, E. F. and Coauthors. 1999. Land-Use and Land-Cover Change (LUCC): Implementation strategy. IGBP Rep. 48, IHDP Rep. 10, International Geosphere–Biosphere Programme, International Human Dimensions on Global Environmental Change Programme, 125 pp.
Lillesand, T. M. and R. W. Kiefer. 2000. Remote Sensing and Image Interpretation. John Wiley and Sons, 724 pp.
Loveland, T. R., B. C. Reed, J. F. Brown, D. O. Ohlen, J. Zhu, L. Yang, and J. W. Merchant. 2000. Development of a Global Land Cover Characteristics Database and IGBP DISCover from 1-km AVHRR Data. Int. J. Remote Sens 21:1303–1330.
Matthews, H. D., A. J. Weaver, M. Eby, and K. J. Meissner. 2003. Radiative forcing of climate by historical land cover change. Geophys. Res. Lett 30:1055. doi:10.1029/2002GL016098.
Mayaux, P., E. Bartholomé, S. Fritz, and A. Belward. 2004. A new land cover map of Africa for the year 2000. J. Biogeogr 31:861–877.
McGuire, A. D. and Coauthors. 2001. Carbon balance of the terrestrial biosphere in the twentieth century: Analyses of CO2, climate and land-use effects with four process-based ecosystem models. Global Biogeochem. Cycles 15:183–206.
Melillo, J. M., A. D. McGuire, D. W. Kicklighter, B. Moore III, C. J. Vorosmarty, and A. L. Schloss. 1993. Global climate change and terrestrial net primary production. Nature 363:234–240.
Monserud, R. A. 1990. Methods for Comparing Global Vegetation Maps. International Institute for Applied Systems Analysis, 31 pp.
Myhre, G. and A. Myhre. 2003. Uncertainties in radiative forcing due to surface albedo changes caused by land-use changes. J. Climate 16:1511–1524.
Niemeijer, D. and V. Mazzucato. 2002. Soil degradation in the West African Sahel—How serious is it? Environment 44:(2). 20–31.
Oldeman, L. R., R. T. A. Hakkeling, and W. G. Sombroek. 1991. World map of the status of human induced soil degradation: An explanatory note. International Soil Reference and Information Center, Wageningen, Netherlands, 34 pp.
Ramankutty, N. and J. A. Foley. 1998. Characterizing patterns of global land use: An analysis of global croplands data. Global Biogeochem. Cycles 12:667–685.
Ramankutty, N. and J. A. Foley. 1999. Estimating historical changes in global land cover: Croplands from 1700 to 1992. Global Biogeochem. Cycles 13:997–1027.
Scherr, S. J. 1999. Soil degradation: A threat to developing-country food security by 2020? Food, Agriculture, and the Environment Discussion Paper 27, IFPRI, Washington, D.C., 17 pp.
Sellers, P. J. and Coauthors. 1996. A revised land surface parameterization (SiB2) for atmospheric GCMs. Part I: Model formulation. J. Climate 9:676–705.
Sitch, S. and Coauthors. 2003. Evaluation of ecosystem dynamics, plant geography and terrestrial carbon cycling in the LPJ dynamic global vegetation model. Global Change Biol 9:161–185.
Skole, D. and C. Tucker. 1993. Tropical deforestation and habitat fragmentation in the Amazon: Satellite data from 1978 to 1988. Science 260:1905–1910.
Taylor, C. M., E. F. Lambin, N. Stephenne, R. J. Harding, and R. L. H. Essery. 2002. The influence of land use change on climate in the Sahel. J. Climate 15:3615–3629.
Tegen, I., M. Werner, S. P. Harrison, and K. E. Kohfeld. 2004. Relative importance of climate and land use in determining present and future global soil dust emission. Geophys. Res. Lett 31.L05105, doi:10.1029/2003GL019216.
Tobler, W. R. 1979. Smooth pycnophylactic interpolation for geographical regions. J. Amer. Stat. Assoc 74:519–530.
Vorosmarty, C. J., B. Moore, A. L. Grace, M. P. Gildea, J. M. Melillo, B. J. Peterson, E. B. Rastetter, and P. A. Steudler. 1989. Continental scale models of water balance and fluvial transport: An application to South America. Global Biogeochem. Cycles 3:241–265.
Wang, G. and E. A. B. Eltahir. 2000. Ecosystem dynamics and the Sahel drought. Geophys. Res. Lett 27:795–798.
Wood, S., K. Sebastian, and S. J. Scherr. 2000. Pilot Analysis of Global Ecosystems: Agroecosystems. International Food Policy Research Institute and World Resources Institute, 110 pp.
WRI 2004. World Resources 2002–2004. World Resources Institute, 328 pp.
Xiao, X. and Coauthors. 2003. Uncertainties in estimates of cropland area in China: A comparison between an AVHRR-derived dataset and a Landsat TM-derived dataset. Global Planet. Change 37:297–306.
Yang, D. W., S. Kanae, T. Oki, T. Koike, and K. Musiake. 2003. Global potential soil erosion with reference to land use and climate changes. Hydrol. Process 17:2913–2928.
Zeng, N., J. D. Neelin, K. M. Lau, and C. J. Tucker. 1999. Enhancement of interdecadal climate variability in the Sahel by vegetation interaction. Science 286:1537–1540.
APPENDIX: Cropland Inventory Data Preparation
The AgroMAPS and WALTPS projects provided subnational data on the crop-harvested area for all of the countries in West Africa. From these databases, the harvested area of crops was extracted at one level below the nation, from the year with available data closest to 1992 (Table 1); 1992 was chosen as the base year to make this new dataset comparable to Ramankutty and Foley (Ramankutty and Foley 1998). Where duplicate data were available, the AgroMAPS data were used first, and the WALTPS data were used only to fill gaps. Next, I aggregated the subnational data for each crop to the national level and compared the data to FAOSTAT national crop-harvested area data for 1992 (FAO 2004). In most of the cases where subnational data were available for 1992, the totals matched the FAO national data. If they did not match, the subnational data for each crop were calibrated to the FAOSTAT harvested area data by multiplying by a constant factor. Next, the total harvested area over all the crops was calculated from the subnational inventory data and FAOSTAT. Because some crops are missing data in AgroMAPS/WALTPS, these totals also do not match (Table 1 shows the ratio of total AgroMAPS/WALTPS harvested area to FAOSTAT harvested area). For most countries, sufficient data were available, and once again the AgroMAPS/WALTPS harvested areas were adjusted to match FAOSTAT. This procedure yielded subnational total harvested areas for most of the countries in West Africa, adjusted to match FAOSTAT data.
From the statistical data, I wanted to obtain subnational cropland area. The subnational harvested area data do not match cropland area because of issues related to multiple cropping, as well as not accounting for fallow areas, cropland areas temporarily used for pastures, etc. However, it can probably be assumed that cropland area is distributed among the subnational units in roughly the same proportion as a harvested area. Therefore, I obtained national statistics on cropland area from FAOSTAT and redistributed it among the subnational units using the subnational proportions obtained from the AgroMAPS/WALPTS harvested area data.
For Nigeria, a large country that is heavily cultivated, the data quality was poor, and indeed, the final calibrated statistics had more cropland area in one administrative unit than land area. This suggested that the Nigerian census statistics could not be reliably used. Therefore, I derived subnational cropland area from a study that used Landsat, SPOT, ERS-1 radar, and JERS-1 radar data to obtain estimates for 1993–95 (FORMECU 1995).
Comparison of the (bottom) RF98 global croplands dataset over West Africa to a (top) Landsat-derived CUI map developed by the EROS Data Center as part of FEWS. The FEWS CUI dataset is categorical, representing five different ranges of CUI values, while the RF98 dataset is continuous (but is shown in this figure over the same categories as FEWS CUI for ease of comparison)
Citation: Earth Interactions 8, 23; 10.1175/1087-3562(2004)8<1:CIWAAG>2.0.CO;2
Land-cover classification datasets over West Africa derived from moderate-resolution satellite sensors. The data shown in the top panels (DISCover and University of Maryland) were derived from the AVHRR dataset and are representative of land cover in 1992. The lower panels are datasets representative of the year 2000, derived from MODIS and SPOT VEGETATION sensors
Citation: Earth Interactions 8, 23; 10.1175/1087-3562(2004)8<1:CIWAAG>2.0.CO;2
(top) Subnational cropland inventory data derived from the AgroMAPS/WALTPS subnational crop-harvested area (calibrated to FAOSTAT national cropland area statistics)—see the appendix for details. Note that statistics for Nigeria were based on satellite data. (bottom) Population density in West Africa from Deichmann (Deichmann 1997)
Citation: Earth Interactions 8, 23; 10.1175/1087-3562(2004)8<1:CIWAAG>2.0.CO;2
Flowchart depicting the three steps used in this study to derive a cropland dataset for West Africa by merging population density, cropland inventory data, and the Landsat-based FEWS CUI dataset
Citation: Earth Interactions 8, 23; 10.1175/1087-3562(2004)8<1:CIWAAG>2.0.CO;2
A scatterplot comparing the AgroMAPS/WALTPS-derived subnational cropland inventory data (y axis) to the median population density within the same administrative units (population densities > 594 people km−2 were ignored in calculating the median; also a minimum of 10 values were required within each administrative unit to estimate a median, otherwise that unit was ignored). The line (y = 0.0053 x) shows the best-fit linear-regression model estimated by S-PLUS. Also shown are the five outliers identified while fitting the model—Ouest, Cameroon (CMR-OUE); Akwa Ibom, Nigeria (NGA-AKI); Rivers, Nigeria (NGA-RVR); Kogi, Nigeria (NGA-KOG); and Anambra, Nigeria (NGA-ANM)
Citation: Earth Interactions 8, 23; 10.1175/1087-3562(2004)8<1:CIWAAG>2.0.CO;2
(top) Cropland intensity for West Africa estimated using the linear-regression model from Figure 5, and using population density as the independent variable. The predicted values are shown here using the same categories as the FEWS CUI data in order to facilitate comparison with Figure 1. (bottom) The final map of cultivation intensity in West Africa was obtained by merging the linear-model simulation from the top panel, the FEWS CUI dataset over the Sahel, and the AgroMAPS/WALTPS cropland inventory data
Citation: Earth Interactions 8, 23; 10.1175/1087-3562(2004)8<1:CIWAAG>2.0.CO;2
Subnational crop-harvested area data obtained from the AgroMAPS project and the WALTPS study. The various crops included in each country are listed, as well as the period from which the data were extracted from AgroMAPS or WALTPS (the range indicates that data for different crops were available from different years). When data from multiple years were available, I extracted data from the year closest to the baseline of 1992. The ratio of total national harvested area from the subnational data to the FAOSTAT national totals is also shown to indicate the completeness of the subnational data compared to the FAOSTAT national data
Continued
Results from the S-PLUS linear model calibrating population density against AgroMAPS/WALTPS inventory data. The linear model was initially built using 80% of the data and then tested against the withheld 20% of the data. After concluding that the model was robust, we built it using the full dataset for final use
Intercomparison of simulated cropland data to observations. The original RF98 croplands dataset and the linear-model-generated output from this study (output of step 1 in Figure 4) were compared to various observations. A cell-by-cell comparison was performed against the categorical FEWS CUI dataset, and a fraction of cell agreement, as well as a kappa statistic, was calculated. Next, the data were compared at coarser resolutions (at 0.5° resolution in latitude by longitude, and aggregated over administrative units) to the calibrated FEWS CUI data (output of step 2 in Figure 4). Finally, the data were compared to the AgroMAPS/WALTPS inventory data (although the linear model was built using the inventory data, and therefore such a comparison is biased). In step 3, we integrated the linear-model output, FEWS CUI data, and the AgroMAPS/WALTPS inventory data (i.e., all available data were merged into one final product), and therefore a similar statistical comparison is meaningless for the final product
The term “intensity” is used here, and in the rest of this paper, to represent the percentage of a grid cell or pixel that is covered by croplands. It is not to be confused with the more common use of the term intensity (or intensification) in the land-use literature where it is often used to indicate the amount of inputs (fertilizer, irrigation, etc.) or land-use practices (choice of crop, multiple cropping, etc.) that are used to increase crop yields or revenue.
The term “proxy” is used here as a synonym for “correlate.” In the other words, I simply use the fact that population density and cropland distribution are well correlated to use the spatial pattern of the former as a surrogate for the pattern of the latter. No assumption about the causal relationship between population density and croplands is implied.
