gb001386 1..17

Document Sample
gb001386 1..17 Powered By Docstoc
                       GLOBAL BIOGEOCHEMICAL CYCLES, VOL. 16, NO. 3, 10.1029/2000GB001386, 2002

Characterizing patterns of agricultural land use in Amazonia by
merging satellite classifications and census data
Jeffrey A. Cardille and Jonathan A. Foley
Center for Sustainability and the Global Environment (SAGE), Institute for Environmental Studies, University of Wisconsin,
Madison, Wisconsin, USA

Marcos Heil Costa
                                                                ¸       ¸
Department of Agricultural Engineering, Federal University of Vicosa, Vicosa, Brazil
Received 3 January 2001; revised 29 October 2001; accepted 30 November 2001; published 24 August 2002.

[1] Amazonia has been under considerable development pressure as croplands and
pasture are established in areas formerly occupied by tropical forest and cerrado.
Although this region is an important part of several important planetary biogeochemical
cycles, the location and impact of human land use are not well understood. In particular,
there is no existing satellite-based map of agriculture across the Amazon or Tocantins
river drainage basins. Recent efforts have classified land cover across this vast region,
although they disagree on the location and amount of cropland and do not directly address
pasture, a land use that has grown in importance in the last 2 decades. Here we present an
analysis of land cover and land use practices over the Amazon and Tocantins basins of
South America. In this study, we demonstrate how satellite imagery and agricultural
censuses can be merged in order to provide a geographically explicit, fine-scale
description of land cover and land use practices. The result depicts the fraction of each 5-
min (9 Â 9 km) grid cell that was devoted to agricultural activity during the mid-1990s.
The resultant map retains many of the characteristics of the agricultural census data, but
with a much finer spatial resolution. During the mid-1990s, cultivated area is estimated to
have been 1.7 Â 107 ha (2.5% of the basin), natural pasture is estimated at 3.3 Â 107 ha
(4.9% of the basin), and planted pasture is estimated to cover 3.3 Â 107 ha (4.9% of the
basin). Perhaps more important than the quantities, however, is that these data sets
provide a new blend of ground-based and satellite-based spatially explicit data. This
snapshot can be used as a basis to project either forward or backward in time, as a new
check of finer scale land use classifications or as a driver of ecosystem models.      INDEX
TERMS: 1615 Global Change: Biogeochemical processes (4805); 1640 Global Change: Remote sensing; 1699
Global Change: General or miscellaneous; KEYWORDS: Amazon, land use, cover, agriculture, census

1. Introduction                                                 the late 1970s, deforestation rates have ranged between
                                                                roughly 10,000 and 30,000 km2 yrÀ1 [Fearnside, 1993;
  [2] Amazonia, the region of forest and cerrado occupying
                                                                Instituto Nacional de Pesquisas Espaciais, 1999].
the tropical belt of South America (Figure 1), is the focus of
                                                  8               [4] While tropical rain forests have garnered much of this
much recent attention. This huge (6.7 Â 10 ha) area is
                                                                recent attention, cerrado ecosystems are also undergoing
home to a unique collection of ecosystems, with some of the
                                                                great change in Amazonia. Cerrado, a regional name for the
greatest biological diversity on the planet. It is important to
                                                                savanna biome, covers $20% of the Amazon and Tocantins
understand how changes in land use affect the biological,
                                                                basins. Cerrado usually occurs in flat areas with deep, well-
chemical, and physical functions of Amazonia, including
                                                                drained soils and a favorable climate that in some cases
the influence of Amazonia on global climate [Nobre et al.,
                                  be harvested in 2 years. These con-
                                                                allows five crops to
                                                                ditions are nearly ideal for the establishment of mechanized,
  [3] Given Amazonia’s role in important Earth system
                                                                commercial agriculture, and many areas still covered by
functions, concern has mounted over the rapid acceleration
                                                                cerrado are likely to be converted to agricultural use in the
of human land use in recent decades. For example, Fearn-
                                                              2 near future. The development pressure on remaining cerrado
side [1993] estimated that by 1991, over 400,000 km
                                                                is so great that it has recently been named a ‘‘threatened hot
($10%) of Amazon forest had already been removed. Since
                                                                spot’’ by Conservation International [Mittermeier et al.,
Copyright 2002 by the American Geophysical Union.                 [5] Development of human agricultural systems is likely
0886-6236/02/2000GB001386$12.00                                 to affect the structure and functioning of extremely impor-

                                                                       18 - 1
18 - 2                                 CARDILLE ET AL.: LAND USE IN AMAZONIA

            Figure 1. Major administrative boundaries in the Amazon and Tocantins river drainage basins.

tant ecosystems in Amazonia. Fire-driven establishment and forest biomass could take 100 years or more [Uhl et al.,
management of agriculture may create a system that appears 1988]. Additionally, many areas still classified by satellites
to be a savanna or grassland to a satellite, but which as ‘‘forested’’ have already been degraded by accidental fires
functions differently than these natural ecosystems. Since and by logging, a major nonagricultural land use [Nepstad et
the typical land use history includes fires every 2 – 3 years, al., 1999]. Current human land use patterns will likely affect
many fields present a chronosequence of decreasing total forest and cerrado in Amazonia for many years.
biomass over time in which residual wood debris from the          [8] Although much progress has been made toward
converted primary forest is burned repeatedly during main- describing the location and effects of land use at regional
tenance fires [Guild et al., 1998]. In these changing systems and global scales [e.g., DeFries et al., 1999; Frolking et al.,
our understanding of the effects of land use on hydrological 1999; Houghton, 1991; Houghton et al., 1991a, 1991b,
budgets [e.g., Rosenfeld, 1999; Williams and Melack, 1997] 2000; Moran and Brondizio, 1998; Ramankutty and Foley,
and on CO2 emissions over time [e.g., Houghton et al., 2000] 1998, 1999; Verburg and Chen, 2000; Wood and Skole,
could be aided by a spatially explicit map of agricultural land 1998], there is no existing spatially explicit agricultural land
use activity.                                                   use map for Amazonia. Understanding the distribution and
  [6] Human land use can also affect adjacent land that abundance of land use, as distinct from land cover, appears
has not been directly converted. Forest fragments show a to offer modelers additional information about possible
decrease in animal biodiversity of mammals, birds, insects, effects on climatic, hydrological, and biogeochemical pro-
and marsupials [Turner, 1996], and fragmentation of Ama- cesses at regional and perhaps global scales.
zonian forests can increase tree mortality near edges             [9] Here we present an analysis of agricultural land use
through woody vine encroachment, microclimate changes, over the Amazon and Tocantins basins of South America
and increased wind turbulence, leading to a significant (Figure 1). We use a statistical method (based on regression
change in biomass [Laurance et al., 1998, 1997, 2001]. In tree analysis) to find the best fit between satellite and census
addition, fires that accidentally spread from pastures or information, using this relationship to estimate agricultural
logged forests can leave remaining forests more suscep- density at a 5-min ($9 km) spatial resolution. This fusion of
tible to damage from future fires, agricultural censuses provides a new
                                          creating a positive satellite imagery and
feedback loop [Cochrane et al., 1999; Uhl and Kauffman, geographically explicit, fine-grained description of culti-
1990].                                                          vated area, natural pasture, and planted pasture.
  [7] Even if agricultural conversion were to slow or stop in
the next few decades, there would still be a significant
amount of area undergoing succession along only partially 2. Satellite-Based Land Cover Images
understood pathways. In pastures that had seen heavy,             [10] In order to accurately record the geographic position
mechanized land management, there would likely be sig- and kinds of land cover that are found across the vast
nificant barriers to tree establishment [Nepstad et al., 1996, Amazonia region, we must rely on satellite-borne instru-
1990]; in more typical pastures the reattainment of mature- ments. One class of satellite data comes from the advanced
                                        CARDILLE ET AL.: LAND USE IN AMAZONIA                                               18 - 3

         Figure 2. International Geosphere-Biosphere Programme (IGBP) DISCover (left) and Global Land
         Cover Facility (right) land cover classifications for the Amazon and Tocantins basins. See color version
         of this figure at back of this issue.

very high resolution radiometer (AVHRR) instrument,                60.7% were identified by the UMD group as wooded
which produces data at high temporal frequency (twice              grassland. Of those pixels identified as cropland in the
daily satellite passes), but at fairly coarse spatial resolution   UMD classification, conversely, only 8.6% were also clas-
(roughly 1 Â 1 km image pixels). Another class of satellite        sified as IGBP cropland; 19.4% of them were labeled as
data comes from the Landsat platform, which produces data          grasslands, while 40.8% were labeled cropland/natural
at lower temporal frequency (once every 2 weeks), but at           vegetation mosaic, a category not directly comparable with
higher (30 Â 30 m) spatial resolution.                             any in the UMD classification [Hansen and Reed, 2000].
                                                                   This general disagreement was characteristic of categories
2.1. Using AVHRR-Based Land Cover Imagery in the                   other than cropland as well; when evergreen broadleaf forest
Amazon                                                             and water were excluded from consideration, only 13.9% of
  [11] In response to the need for large-extent assessments        the remaining part of the study area was classified as the
of land cover, numerous global AVHRR-based images of               same category in the two classifications. There were also
land cover have recently been developed. Here we examine           substantial differences in proportions of major land covers,
two of them over the Amazon and Tocantins basins: the              which was seen throughout each of the countries of the
International Geosphere-Biosphere Programme (IGBP)                 study area (Table 1). The absolute accuracy of either
DISCover classification [Belward and Loveland, 1996]               classification is difficult to establish for the labeled catego-
(Figure 2 (left)) and the University of Maryland (UMD)             ries: Although a validation phase was recently undertaken
Global Land Cover Facility classification [Hansen et al.,          for the global IGBP classification [Loveland et al., 1999],
2000] (Figure 2 (right)). These land cover data sets were
developed from monthly values of the normalized differ-
ence vegetation index (NDVI) and reflectance values in the
five AVHRR channels for 1992– 1993 and represent the two
most widely used efforts to categorize global land cover
during the early to middle 1990s. Focused on identifying
land cover in each 1-km pixel, these classifications are
clearly a major step forward in understanding the distribu-
tion and abundance of major biomes and a few easily
identified land use categories. However, there were limita-
tions that prevented their direct use for the identification of
cropland and pasture in the study area.
  [12] A pixel-by-pixel comparison of the two classifica-
tions in the study region found widespread disagreement
about land cover in areas other than evergreen broadleaf
forest or water (Figure 3). This is consistent with a global
comparison of the two data sets that found agreement on
broad vegetation categories but low per-pixel agreement on
individual classes and significant regional variability [Han-
sen and Reed, 2000]. In particular, only 3.2% of the pixels
labeled as cropland in the IGBP classification were also Figure 3. Vegetation class agreement between Global
labeled as cropland in the UMD classification; the greatest Land Cover Facility classification and IGBP-DISCover
source of disagreement of IGBP cropland pixels was that classification in the Amazon and Tocantins basins.
18 - 4                                         CARDILLE ET AL.: LAND USE IN AMAZONIA

Table 1. Percentages of Major Land Covers in University of Maryland (UMD) and International Geosphere-Biosphere Programme
(IGBP) DISCover Data Sets in Amazoniaa
                Evergreen        Woody Savanna       Savanna (Wooded                                            Cropland/Vegetation
             Broadleaf Forest     (Woodland)            Grassland)            Cropland          Grassland       Mosaic (Not in UMD)           Other
             IGBP      UMD        IGBP      UMD        IGBP       UMD      IGBP     UMD        IGBP    UMD        IGBP        UMD        IGBP      UMD
Bolivia      52.8      62.5         5.9      11.3         5.2      15.8       1.7        0.9    10.1     4.1         11.3     –            13.0       5.4
Brazil       64.5      70.5         2.8       8.0         5.0      14.6       3.2        0.8     0.4     1.9         21.1     –             3.0       4.3
Colombia     89.8      89.9         1.5       1.7         1.6       6.2       0.5        0.1     0.6     0.5          4.5     –             1.5       1.7
Ecuador      61.4      71.3         0.3       9.7         1.9       5.0       3.1        0.4     8.5     6.4         15.2     –             9.5       7.2
Peru         69.7      73.6         0.3       7.3         1.6       5.2       1.2        1.2    12.3     7.5          9.0     –             5.9       5.3
Region       64.3      69.9         2.7       8.0         4.5      13.0       2.7        1.0     3.6     3.4         17.3     –             4.9       4.7
    Where land cover names were not identical in the two classifications, comparable categories following Hansen et al. [2000] were used. In those cases,
column headers are of the form IGBP class name (UMD class name).

the number of sampled points within Amazonia was too                            [16] Though data from this project could not be directly
small to judge the quality of that classification in these river              used to determine land use, we expected that these data
basins.                                                                       could be used to assess land use maps of Amazonia. We
  [13] Identification of the extent of pasture within these                   believed that there should be substantial correlation between
basins presented a more fundamental problem. Since the                        fine-scale land cover data and agricultural activity in Ama-
spectral and spatial resolution of AVHRR imagery prohib-                      zonia, because these classifications looked at the same
its the direct identification of most agricultural land uses in               region in the same time period.
Amazonia, the UMD and IGBP classifications did not
include pasture as a mapped class. There were categories                      3. Agriculture and Census Data
whose descriptions were similar to pasture (e.g., savanna:
lands with forest cover between 10 and 30% and herba-                           [17] Agriculture in Amazonia takes two main forms:
ceous cover elsewhere). However, the disagreement                             pasture and cropland; of these two, pasture dominates
between the two classifications in these categories was                       [Kauffman et al., 1998]. Pastoral practices vary widely
extreme enough to discourage simply adopting one or the                       across Amazonia, from the mechanized, fertilized planted
other. Furthermore, without additional information it                         pasture and machete-weeded pastures of Eastern Para         ´
would have been difficult to determine, for example, what                     [Fearnside, 1990; Uhl et al., 1988] to low-productivity
proportion of the savanna area is used as pasture and                         pastures formed either to gain title to unclaimed land or
whether this proportion varies spatially across the study                     as a response to declining crop yields [Fearnside, 1990] to
region.                                                                       extensive natural pastures in Bolivia [Instituto Geogra ´fico
                                                                              Militar and Bolivia Ejercito Comando General, 1997;
2.2. Using Landsat-Based Land Cover Imagery in the                            International Association of Agricultural Economists,
Amazon                                                                        1969] to sheep and cattle ranges in Peru [Auge S. A.
  [14] The Humid Tropical Deforestation Project (for                          Editores, 1994].
Colombia, Peru, Ecuador, and Bolivia) [Townshend et al.,                        [18] To characterize the patterns of agriculture across the
1995] and the Tropical Rain Forest Information Center (for                    Amazon and Tocantins basins, we assembled detailed
Brazil) [Skole and Tucker, 1993] used Landsat data from the
1970s, 1980s, and 1990s to classify land cover for much of
Amazonia. The Brazil data set contains categories of forest,
nonforest, deforestation, cerrado, and secondary vegetation,
while the classifications for the other countries principally
distinguish forest from nonforest.
  [15] Although these data sets provide valuable, high-
resolution information about the locations of major land
covers during the mid-1990s, they do not explicitly include
croplands or pastures, the major agricultural land uses in
Amazonia. These data sets, like the AVHRR classifications,
thus could not be directly used to identify the distribution of
these important agricultural land uses in the region. In
addition, these classifications do not cover the entire study
area (particularly Goias and the Peruvian Andes), although
together they do cover much of the interior of the Amazon
and Tocantins basins (Figure 4). Finally, areas with very
high known land use (e.g., in Tocantins and Mato Grosso)
were often classified simply as ‘‘cerrado,’’ which makes the
use of these maps to quantify differences in land use Figure 4. Location of forest, nonforest, and missing data
amounts in these areas difficult.                               for Landsat tropical rain forest classification project.
                                              CARDILLE ET AL.: LAND USE IN AMAZONIA                                                            18 - 5

                                                                             was calculated from the area ‘‘covered by pastures which
                                                                             have grown by natural means,’’ [Instituto Nacional de
                                                                                    ´                 ´            ´n ´
                                                                             Estadıstica e Informatica Direccio Tecnica de Censos y
                                                                             Encuestas, 1996], that is, those pastures in which relatively
                                                                             little effort is required to graze animals. The area in planted
                                                                             pasture was calculated from the area ‘‘dedicated for grazing
                                                                             and formed through planting’’ [Fundacao Instituto Brasi-
                                                                             leiro de Geografia e Estatistıca, 1997] and was similarly
                                                                             converted to the fraction of planted pasture for each unit.
                                                                                [20] Although the accuracy of census data is impossible to
                                                                             quantify over such a large region, indications are that these
                                                                             particular variables are well reported for these censuses.
                                                                             Planted area, the census variable used for this study, is one
                                                                             of the most reliable variables in agricultural censuses,
                                                                             according to census officials (Fundacao Instituto Brasileiro
                                                                             de Geografia e Estatıstica, personal communication, 2001).
                                                                             In order to secure financing, for example, farmers typically
                                                                             need to know an exact measurement of area planted in
Figure 5. Administrative unit boundaries for development                     crops. Interannual variation in cropland can be large, how-
of spatially explicit agricultural census data.                              ever, owing to changes in crop prices, interest and inflation
                                                                             rates, and other financial factors. Natural and planted
                                                                             pasture amount is more stable from year to year and thus
agricultural censuses from as many of the relevant states                    is more reliable than the cropland area value. Since most of
and countries as possible (Figure 5 and Table 2). This                       the agricultural activity in Amazonia is for pasture, the
collection of census data is quite detailed and attempts to                  greater stability of these variables raises our confidence that
describe land use practices within small political units,                    the census-based estimates of agricultural area are reason-
including municipios in Brazil and distritos in Peru, roughly                able. Additionally, Brazilian census officials note that their
the equivalent of counties in the United States.                             most recent census was well funded, increasing the reli-
  [19] This compilation of census data contains three major                  ability of values, since more agricultural units were directly
categories: cultivated area, natural pasture, and planted                    sampled than in previous censuses.
pasture. Each category represents the fraction of the given                     [21] Although substantial effort was made to find mid-
administrative unit estimated, from agricultural census data,                1990s agricultural census data for the entire study region,
to have been used for the specified land use in the mid-                     we could find such information only for Brazil and Peru. In
1990s. In particular, the cultivated area in each administra-                Bolivia, no detailed agricultural censuses (below state level)
tive unit was calculated by summing the area dedicated to                    have been published since 1984. We estimated the condition
annual crops (e.g., corn and rice), perennial crops (e.g.,                   of mid-1990s agricultural data from information contained
avocados and oranges), fallow land (i.e., land dedicated to                  in the district-level 1984 Bolivian census [Republica de
cultivation but temporarily at rest at the time of the census),                                                       ´
                                                                             Bolivia Instituto Nacional de Estadıstica, 1990] and the
and land that had been harvested but not yet replanted at the                mid-1990s Food and Agriculture Organization (FAO) state-
time of the census. The fraction of cultivated area of each                  level data [U.N. Food and Agriculture Organization, 2000].
unit was then calculated by dividing the cultivated area by                  The fractions of cultivated area, natural pasture, and planted
the area of that unit, as determined by the ARC/INFO                         pasture from the 1984 census were adjusted using the ratio
Geographic Information System (Environmental Systems                         of state-level FAO estimates of those quantities in 1995 to
Research Institute, Redlands, California) after projecting                   that of 1984. Additionally, the 1984 census did not include
administrative boundary files into an area-preserving coor-                  the state of La Paz. Ecuador, which comprises 2% of the
dinate system. The fraction of natural pasture in the county                 study region (Figure 1), also did not conduct an agricultural

Table 2. Agricultural Census Data Sources and Attributesa
                                                        Administrative                 Source of Administrative                 Median Administrative
Country                      Source 
                                                      Level                Borders                                                    Area, ha
Bolivia                   INE 1984,                    provincia                       IGM 1980                                 266,680
                            FAO 1995
Brazil                    IBGE 1996                    municipio                       IBGE 1994                                63,195
Colombia                  N/A                          N/A                             ESRI 1995                                N/A
Ecuador                   INEC 1995                    provincia                       ESRI 1995                                56,563
Peru                      INEI 1995                    distrito,                       INEI 1998                                14,784
    Abbreviations used: ESRI, Environmental Systems Research Institute; FAO, United Nations Food and Agriculture Organization; IBGE, Fundacao       ¸˜
                                         ´                            ´
Instituto Brasileiro de Geografia e Estatıstica; IGM, Instituto Geografico Militar; INE, Instituto Nacional de Estadistica; INEC, Instituto Nacional de
Estadistica y Censos; INEI, Instituto Nacional de Estadistica e Informatica.
18 - 6                                           CARDILLE ET AL.: LAND USE IN AMAZONIA

Table 3. Mid-1990s Census-Derived Agricultural Area for Amazon Basin Study Region
         Country                         State                  Cultivated              Natural             Planted            Total, ha
                                                                 Area, ha              Pasture, ha         Pasture, ha
Bolivia (except La Paza)                                         787,717              12,279,462              242,645         13,309,824
Brazil                               Totalb                    6,014,226              15,869,468           33,404,530         55,288,224
                                     Amazonasc                   303,931                 320,101              207,406            831,438
                                     Goias                       668,251               3,354,511            8,118,471         12,141,232
                                     Mato Grosso               2,642,069               3,561,833           10,915,143         17,119,045
                                     Para                        776,398                 746,147            4,729,820          6,252,365
                                     Rondonia                    509,540                 350,406            2,622,694          3,482,640
                                     Tocantins                   631,384               5,868,675            5,340,404         11,840,463
Colombiad                                                           N/A                     N/A                   N/A               N/A
Ecuador                                                          410,402                 420,328              718,370          1,549,100
Peru                                                           2,578,224               7,309,279              266,414         10,153,917
Other countries                                                     N/A                     N/A                   N/A                N/A
Totale                                                         9,790,569              35,878,537          34,631,9959         80,301,065
Relative fraction                                                      0.12                    0.45                 0.43               1.00
    State of La Paz was not included in 1984 census. See text for details.
    All municipios are included in study area. Totals for major states are shown.
    State totals include only those municipios within study area.
    No agricultural census data was found for Colombia. See text for details.
    Total includes only those areas for which agriculture data were collected.

census in the 1990s; values were estimated using 1992 state                   census data derived from mid-1990s county-level agricul-
data [Sistema Estadıstico Agropecuario Nacional, Instituto                    tural censuses and the remainder of the data derived from
Nacional de Estadıstica y Censos, 1995].                                      either older censuses or state-level estimates. Across the
  [22] This compilation of agricultural census data permits                   basin in the mid-1990s, only 12% of agricultural land use
us to quantify agricultural activity across most of Amazonia                  was for cultivated area (Figure 6b), 45% was natural pasture
(Table 3 and Figure 6). Across the portion of the basin                       (Figure 6c), and 43% was planted pasture (Figure 6d and
where census information was available, agricultural activ-                   Table 3). This result is consistent with other analyses that
ity covered a total of 8.0 Â 107 ha (Table 3), with 82% of                    have found pasture to be the ‘‘most important agent of


           Figure 6. Mid-1990s agricultural census data for the Amazon and Tocantins basins. In Bolivia, Peru,
           and Brazil, data are at a spatial scale analogous to a U.S. county; in Ecuador, state-level data are
           presented. See color version of this figure at back of this issue.
                                      CARDILLE ET AL.: LAND USE IN AMAZONIA                                            18 - 7

deforestation’’ [Skole, 1994] and ‘‘the dominant land use       in that same county, of some category or combination of
following forest clearing’’ [Fearnside, 1996] in Amazonia.      categories found in the finer scale UMD classification, the
The dominance of pasture among these land uses suggests         IGBP classification, or both. This relationship between land
that data sets mapping only croplands [Belward and Love-        use and land cover data, which we hypothesized to be
land, 1996; Hansen et al., 2000; Ramankutty and Foley,          consistent across the county scale and the 5-min scale,
1998] are likely to underrepresent most agricultural activity   would allow us to estimate the proportion of cropland and
in these basins.                                                pasture in any part of the study region. Derived from a
  [23] The censuses show heavy pasture density in the           fusion of the two types of data sources, this relationship can
Brazilian states of Tocantins, Goias, and southeastern Para,
                                  ´                         ´   also be used as a statistical basis for an estimate in those
planted pasture and crop cultivation in Rondonia, cropland      areas with no agricultural census data.
and planted pasture in Mato Grosso, extensive natural             [27] Here we have used a set of statistical techniques to
pasture in northern Bolivia, and cropland and natural           create a regular gridded map of agricultural activity, depict-
pasture along the eastern slopes of the Andes in Ecuador        ing the fraction of each grid cell devoted to croplands,
and Peru (Figure 6). The locations of this activity are         natural pasture, and planted pasture during the mid-1990s.
consistent both with single-site studies and with relatively    We blended the two data sets to create a regular, latitude-
coarse representations of land use in Amazonia [Auge S. A.      longitude grid of 5-min resolution ($9 km at the equator).
Editores, 1994; Conselho Nacional de Geografia Divisa      ˜o   The 5-min resolution, an intermediate scale between that of
de Geografia, 1960; Guild et al., 1998; International           the AVHRR data and the average-sized political unit of
Association of Agricultural Economists, 1969; Neill et al.,     census data, is well suited for most ongoing large-scale
1996; Rand McNally et al., 1994; U.S. Department of             modeling studies.
Agriculture, 1994; Uhl et al., 1988]. The analysis of census
data underscores the critical need to accurately map the       4.1. Regression Tree Analysis
natural and planted pasture of Amazonia.                         [28] We captured the relationship between census data
                                                               and satellite land cover classifications using regression tree
                                                               analysis (RTA). RTA is a technique most often used to
4. Methods for Merging Censuses and Satellite                  explore relationships (represented as binary trees) among
                                                               various candidate ‘‘predictor’’ variables to a given response
Imagery                                                        variable. Unlike traditional regression analysis, RTA does
  [24] Although agricultural census data can generally not assume normal distributions of the training data; trees
depict the broad patterns of agricultural activity in Ama- are well suited to nonlinear relationships since predictor
zonia during the mid-1990s, there are limitations to the variables can appear more than once along the path to a
direct use of Figure 6 as a cropland-pasture map. First, terminal node. RTA has seen increasing use as a method of
several countries did not conduct agricultural censuses in exploration of the variation among variables [e.g., Lamon
the mid-1990s (e.g., Bolivia, Colombia, and Ecuador), and and Stow, 1999; Michaelsen et al., 1994] or prediction
we needed a reasonable way to assess crop and pasture based on the regression-like relationship [Iverson and
distributions there. Second, the land area of many political Prasad, 1998]. RTA has been used in a small number of
units is very large (some are 105 km2 in size), and it is studies of Amazonia: Prince and Steininger [1999] revealed
implausible to assume that the density of cropland and relationships among a large number of spectral variables to
pasture is uniform within areas of this size. Finally, com- maximize differences among suggested study sites, and a
puter models used to examine environmental processes in method closely related to RTA has been successfully used to
the Amazon (including climate, hydrological, and ecolog- determine which land cover class was most likely, given
ical models) typically need a complete, and geometrically certain spectral characteristics [Hansen et al., 2000].
regular, depiction of land characteristics.                      [29] Regression tree analysis recursively partitions the
  [25] Satellite-based land cover maps, with their ability to variable space into a nested series of subspaces such that
represent characteristics of an entire region, provide a the remaining variability is minimized at each split [Brei-
possible solution to this problem. Recent work [Ramankutty man, 1984; Venables and Ripley, 1997]. A binary ‘‘regres-
and Foley, 1998] has indicated that it is possible to ‘‘pull’’ sion tree’’ is produced during the process of ‘‘growing’’ the
land use information from a satellite land cover classifica- tree, and each split is based on a single predictor variable. A
tion by exploring the statistical relationship between the new case’s predicted value is determined by following the
satellite classifications and the ground-derived census data splits of the tree to the appropriate terminal node, which
and using it to derive an improved land use/land cover contains the estimate of the response variable, given the
classification. Given census data and
                                        the IGBP and UMD new data for the case.
land cover classifications for Amazonia, we reasoned that        [30] To merge the agricultural census data with existing
although these satellite classifications do not directly pro- satellite-based classifications, we developed a new applica-
vide clear representations of major agricultural activities in tion of RTA that allowed us to estimate cropland and pasture
the Amazon, there should be a systematic relationship information in each 5-min cell within the study region.
between agricultural census information and classified land Since Peru and Brazil together cover more than 80% of the
cover categories.                                              study region, and since agricultural censuses were con-
  [26] Specifically, we reasoned that the density of cropland ducted there during the mid-1990s, information from only
and pasture within a county should be related to the density, those two countries was used to grow the trees. Since we
18 - 8                                 CARDILLE ET AL.: LAND USE IN AMAZONIA

believed the satellite land cover classifications would have
trouble distinguishing cropland, natural pasture, and planted
pasture from each other due to the likely similarity of their
spectral signatures, the total proportion of agricultural
activity (defined as the sum of cropland and pasture
proportions) from the census data was computed for com-
parison with the land cover classifications. The borders of
each administrative unit (Figure 5) in Peru and Brazil were
passed through the UMD and IGBP land cover classifica-
tions. For each administrative unit, we determined the
proportion of that unit labeled as each land cover category
in the two classifications (Figure 7). From this, we created a
set of more than 1000 county-based observations in which
the cropland-pasture value was the response variable and in
which the category proportions were the predictors.
  [31] We randomly divided the county-based observations
into two equal-sized sets: Half of the counties were used for
creating regression trees, while the remaining half were kept
in reserve for tree evaluation. Using the area of each unit as
the weight of its case in the tree growth algorithm, we then
used the R statistical package [Ihaka and Gentleman, 1996]
to grow a regression tree in which the split criteria were the
proportions of land cover categories. Terminal nodes were
the estimated cropland-pasture proportion under those con-
  [32] We explored the fit of each land cover classification     Figure 7. Conceptual diagram of relationship between
with census data by developing trees using predictors from       proportion of cropland-pasture of categories derived from
only the IGBP classification, from only the UMD classi-          land cover classifications.
fication, and from both classifications. This resulted in a
suite of three candidate trees.                                  istrative unit of Brazil, Peru, Bolivia, and Ecuador, the
                                                                 relative proportions of cropland, natural pasture, and
4.2. Five-Minute Gridded Prediction and Candidate                planted pasture were computed using the agricultural census
Tree Evaluation                                                  data from Table 3. Within a given unit, these proportions
   [33] After growing the candidate trees, the entire study      were then multiplied by the fitted cropland-pasture densities
region was partitioned into a set of 5-min cells, and the        for all 5-min cells whose centers were inside the borders of
proportion of each land cover category computed for each         that unit. In the southern part of the study region, propor-
cell. For a given tree, the land cover proportions within        tions calculated for Bolivia were used to distribute the fitted
each 5-min cell were used to produce a spatially explicit        data for La Paz. For land area in the northern part of the
map fitted for the entire study region. This represented the     study region that did not have agricultural census data
best merged fit, in each 5-min cell, of the census and           (Figure 1 and Table 3) the area-weighted average propor-
satellite data used to create the tree. In this way, three       tions from census data of Peru and Brazil were used to
spatially explicit fitted maps were produced for further         apportion the cropland-pasture data.
   [34] To evaluate the behavior of each regression tree, the
                                                                5. Results
half of the administrative units kept in reserve during tree
growth was used to examine the properties of the resulting 5.1. Evaluating and Selecting the Best Fit Between
fitted map. For a given tree, the fitted map’s total agricul- Census Data and Satellite Data
tural area was computed in each evaluation unit; a tree that      [36] The tree built using only data from the UMD
fit census data perfectly in a unit would have the same classification produced the best fit with agricultural census
amount of agricultural area in both the census and the fitted data. When the land use map derived from the UMD
map. For each tree, the relationship between the fitted data classification was overlaid with boundaries from the eval-
and agricultural census data across
                                        evaluation units was uation units, the correlation coefficient computed between
computed in an area-weighted linear regression.                 the census results and the average cropland-pasture value
                                                                within each evaluation unit was 0.81. When the IGBP
4.3. Estimating Cultivated Area, Natural Pasture, and DISCover classification was considered for its power to
Planted Pasture                                                 fit land use data, correlation of the resulting map with
   [35] Since the relationship presented by a regression tree census results (r = 0.69) suggested that it did not capture
was based on the combined cropland-pasture map, we land use as well as the UMD classification. Although this
apportioned the fitted map into three components, using lack of fit could have resulted from widespread errors in the
the relative relationships from spatially explicit agricultural census information, the success of the UMD classification at
census data to the greatest extent possible. For each admin- this same task suggests that it did not. It appears more likely
                                               CARDILLE ET AL.: LAND USE IN AMAZONIA                                                  18 - 9

Table 4. Correlation Between Fitted Cropland-Pasture Data and                  merged map was strongly negatively correlated with ever-
UMD Classificationa                                                            green broadleaf forest (Table 4), as would be expected in a
         Category                                               Correlation    region where the process of deforestation typically replaces
Water                                                             À0.12
                                                                               forest with cropland and pasture. The amount of fitted
Evergreen needleleaf forest                                        0.06        cropland and pasture within a 5-min cell was highly pos-
Evergreen broadleaf forest                                        À0.81        itively correlated in the UMD classification with woodland
Deciduous broadleaf forest                                         0.01        and wooded grassland (Table 4). These two classes were
Mixed forest                                                       0.00        noted by Hansen et al. [2000] as being particularly difficult
Woodland                                                           0.52
Wooded grassland                                                   0.81        to identify correctly from the original satellite data used to
Closed shrubland                                                   0.05        produce the UMD classification. This positive correlation is
Open shrubland                                                     0.07        extremely sensible, given that woodland and wooded grass-
Grassland                                                          0.26        land together encompass lands with herbaceous or woody
Cropland                                                           0.30
Bare ground                                                        0.05
                                                                               understories with tree canopy cover between 10 and 60%
                                                                               [Hansen et al., 2000]. It seems likely that many agricultural
    For each 5-min cell, the fraction of each classified land cover type was
calculated and compared with fused data. Absolute values above 0.5 are
                                                                               areas were categorized as woodland or wooded grassland
shown in bold.                                                                 during the production of the UMD classification.
                                                                               5.2. Map of Cropland and Pasture for Mid-1990s
that the high amount of land classified as ‘‘cropland/natural                  Amazonia
vegetation mosaic,’’ which was not highly correlated with                        [38] The fusion of satellite-based and agricultural census
cropland and/or pasture (Figures 2 and 6), diminished the                      data indicates that in the mid-1990s there was a total of 8.3
effectiveness of the IGBP classification. The merger of                        Â 107 ha of cropland and pasture in Amazonia. This
census data with the combined predictors from both the                         represents a total agricultural land use fraction of 12.3%
IGBP and UMD classification was substantially worse (r =                       of the entire basin’s area. Agriculture was geographically
0.75) than the map produced by the categories from the                         distributed such that 70.5% was in Brazil, 13.6% was in
UMD data alone.                                                                Bolivia, 11.5% was in Peru, 1.6% was in Ecuador, 2.3%
  [37] The cropland-pasture map appears to be related real-                    was in Colombia, and 0.5% was in remaining countries. The
istically with the UMD classification. Land use on the                         map of the fused data set (Figure 8) clearly has the same


           Figure 8. Merged satellite and agricultural census data for the mid-1990s. Each 5-min cell in Figure 8a
           contains the estimated fraction of cropland and pasture estimated through the fusion technique merging
           land use and land cover data. Relative proportions of land use types from agricultural census data are
           used to estimate amounts of (b) cropland, (c) natural pasture, and (d) planted pasture. See color version of
           this figure at back of this issue.
18 - 10                                       CARDILLE ET AL.: LAND USE IN AMAZONIA

Table 5. Statistics for Fused Data Derived From Fusion of Agricultural Census Data and Satellite Classifications
           Country                    State               Total Administrative        Cultivated      Natural         Planted        Agriculture
                                                                Area, ha               Area, ha      Pasture, ha     Pasture, ha      Total, ha
Boliviaa                                                      68,944,066              2,108,883       8,773,454         324,374      11,206,711
Brazil                            Totalb                     454,282,180             10,183,358      17,598,612      30,461,521      58,243,491
                                  Amazonas                   157,771,452              3,887,990       1,168,982         597,406       5,654,379
                                  Goias                       19,401,286                549,320       2,989,709       6,594,446      10,133,475
                                  Mato Grosso                 72,808,175              2,461,056       2,938,761       8,684,726      14,084,543
                                  Para                       107,131,152              1,592,771       1,356,758       6,069,609       9,019,138
                                  Rondonia                    23,854,180                213,582         215,452       1,199,993       1,629,027
                                  Tocantins                   27,678,589                738,083       6,970,203       5,089,076      12,797,362
Colombia                                                      33,097,051                771,857         430,956         735,750       1,938,563
Ecuador                                                       12,717,504                382,021         349,986         589,025       1,321,032
Peru                                                          94,266,193              3,415,138       5,681,341         387,762       9,484,241
Other Countries                                                6,771,955                158,657          89,966         147,151         395,774
Subtotal: BL, BR, EC, PEc                                    630,209,943             16,089,400      32,403,393      31,762,681      80,255,475
Grand Totald                                                 670,078,949             17,019,914      32,924,315      32,645,583      82,589,812
Fraction                                                          –                        0.21            0.40            0.40          –
    Country totals include only land within study area.
    State totals include only land within study area.
    Subtotal is for comparison to totals from Table 3.
    Grand total includes all land within study area.

general appearance as the cropland-pasture totals from                           with Landsat data revealed a correlation of 0.71 between the
census data (Figure 6), as well as finer spatial detail and                      fitted cropland-pasture data and the fraction of each 5-min
properties derived from its merger with the UMD classi-                          grid cell categorized as nonforest. A visual inspection of the
fication (Figure 2 (right)).                                                     two sets indicates the strong agreement (Figure 9).
  [39] The distributed cropland and pasture information
(Figures 8a – 8d) reveals a richly varied distribution of     5.4. Brazil
cultivated area, natural pasture, and planted pasture within    [43] In the mid-1990s, agricultural activity covered 12.8%
Amazonia. During the mid-1990s, cultivated area is esti-      of the part of Brazil within the Amazon and Tocantins
mated to be 1.7 Â 107 ha (2.5% of the basin), natural         basins, with 2.2% of the area in cropland, 3.9% of it in
pasture is estimated at 3.3 Â 107 ha (4.9% of the basin), and natural pasture, and 6.7% of it in planted pasture (Table 5
planted pasture is estimated to cover 3.3 Â 107 ha (4.9% of   and Figure 8).
the basin) (Table 5).                                           [44] Agricultural activity in Mato Grosso, Brazil’s largest
                                                              producer of soybeans [U.S. Department of Agriculture,
5.3. Bolivia                                                  1994], is split between cropland and planted pasture, with
  [40] In the mid-1990s, agricultural activity covered 16.3% crops seen primarily in the municipios near the tributaries of
of that part of Bolivia within the Amazon basin, with 3.1% Rio Saueruina and with planted pasture in the western,
of the land area in cropland, 12.7% of it in natural pasture, northern, and eastern extremities of the state. Cropland is
and 0.5% of it in planted pasture (Table 5).                  extensive in the BR 364 area and along Highway BR 163.
  [41] The density of agriculture in the mid-1990s neared Additionally, Alta Floresta can be clearly seen in the north
78% along the Mamore River, near the city of Trinidad in El of the state, along with heavy pasture density near the
Beni, and was nearly as high throughout Santa Cruz, near intersection of BR 080 and BR 163.
the provincial capital of Santa Cruz de la Sierra. The vast     [45] Roraima’s agricultural activity is primarily pastoral,
majority of this activity was natural pasture used for cattle with planted pasture to the south and natural pasture to the
grazing, a system that has been in place for at least several north. Pasture appears to be concentrated to the east of
decades [International Association of Agricultural Econo- the Rio Branco in the southern part of the state, between the
mists, 1969]. In the state of La Paz, where there was no river and Highway BR 174. In the northern half of the state,
agricultural census information, the fitted model predicts land use density was highest to the southeast of Boa Vista
substantial levels of agricultural activity around and to the and along the west side of Highway BR 174 between Boa
northeast of Ixiamas. Additional activity was seen in La Paz Vista and the Venezuelan border. Cropland and pasture
and Cochabamba along the eastern slopes of the Andes that appear to have developed up to the eastern border of the
defined the southern border of the study area. Cultivation is Reserva Florestal Parima, but have not significantly pene-
evident in the northern tip and central region of Cocha- trated the Reserva, the Parque Indigena Yanomami, or the
bamba, as well as in western Santa Cruz, near Santa Cruz de nearby Reserva Biologica de Mucajaı, which together
la Sierra. Very little cropland or pasture was seen in Pando, dominate Roraima’s northwest quadrant. Additional signifi-
which is low lying, sparsely settled, and mainly forested.    cant land use centers appear near Contao and along the
  [42] Since Landsat information was not used in the slopes of the Sierra Pacaraima in the northern tip of the
merger of the agricultural census data and satellite classi- state.
fications, it can provide an independent qualitative check on               ´
                                                                [46] In Para, where census data (Figure 6) reported only
the fusion process. Comparison at the 5-min spatial scale minimal land use, this fusion of satellite and census data
                                      CARDILLE ET AL.: LAND USE IN AMAZONIA                                          18 - 11

         Figure 9. Comparison for Bolivia between Landsat-based land cover data and merged cropland-pasture
         data. See color version of this figure at back of this issue.

allows us to identify widespread agricultural activity. Para particularly near Pimenta Bueno and Rolim de Moura.
contains a mix of cultivated area and planted pasture, with Pasture can be seen along the road between Ariquemes
cultivation seen near Belem and near the river and with and Pombal, as well as along the road between Porto Velho
planted pasture occurring in eastern Para and near major and Abuna. Although significant development was seen
roads in the state. Most striking is the linear feature along most major roads of the state, there was little
corresponding to the Transamazonica Highway (BR 230) agricultural activity in the Parque Nacional de Pacaas                ´
in the center of the state. Land use density is as high as 48% Novos or in the many indigenous areas of Rondonia. Some
near the road between Itaituba and Maraba. Altamira, the agricultural activity was seen in the southern part of the
northernmost part of the highway in this state, is surrounded state, which may indicate that the regression tree model is
by land use, especially to its west along the highway. too sensitive to the presence of the wetlands there. A
Southeastern Para was dominated by pasture, particularly comparison with Landsat data indicates that the fitted land
in the region of Highway PA 150 and near the road use is highly correlated (r = 0.64) with the sum of the
connecting Xinguara and Sao Felix do Xingu. The only deforestation, secondary vegetation, and cerrado land cov-
substantial limits appear to have been imposed in the areas ers at the 5-min scale, and it is possible to visually locate
                  ´            ´
Indigena Kayapo and Catete, both of which had relatively many of the same roads and features in the two data sets
lower land use density. In northern Para the area along the (Figure 10).
                       ´                  ´
Amazon near Santarem, Alenquer and Obidos had land use            [49] The majority of the state of Acre saw little or no
densities of around 25%.                                        agricultural activity during this time period. However, there
  [47] Tocantins and Goias contain a dense mix of natural was substantial planted pasture around Rio Branco, with
and planted pasture throughout the two heavily agricultural additional land use along the road between Rio Branco and
states. Cropland and pasture density is at its highest level Brasileia.´
within Amazonia: between 50 and 80% along the north-              [50] Amazonas, with its mix of forest and flooding-related
south Highway BR 153 running through the center of the nonforest areas, appears to have presented a problem for the
states. This is consistent with census data (Figure 6) and modeling technique. Since cropland and pasture in the basin
with government descriptions of the states (Infrastructure were so strongly negatively correlated with evergreen
Brazil, available from the Brazilian Ministry of Planning broadleaf forest (Table 4), areas with extensive nonforested
Budget and Management at http://www.infraestruturabrasil. wetlands were often incorrectly presumed by the model to Land use density is lower only in the sparsely be agriculture. As a result, there appears to be a substantial
populated, relatively inaccessible eastern tip of Tocantins overestimate of cropland and pasture in the swampy areas
and in the Parque Nacional do Araguaia in the west.             near the Amazon and its tributaries. One notable exception
  [48] The dominant human land use in Rondonia in the is the municipio of Humaita at the extreme southern edge of
mid-1990s was planted pasture, which constituted 74% of Amazonas, a well-known center of soybean production.
agricultural activity within the state (Figure 6). The crop-
land-pasture map indicates heavy pasture density along 5.5. Peru
Highway BR 364 in the Ji-Parana area. Significant pasture         [51] In the mid-1990s, agricultural activity covered 10.1%
can be seen along that road throughout most of its length, of the part of Peru within the Amazon basin, with 3.6% of
18 - 12                               CARDILLE ET AL.: LAND USE IN AMAZONIA

          Figure 10. Comparison for Rondonia, Brazil, between Landsat-based land cover data and merged
          cropland-pasture data. See color version of this figure at back of this issue.

the area in cropland, 6.0% of it in natural pasture, and 0.4%   sets (Figure 11) shows the agreement, both in the heavily
of it in planted pasture (Table 5 and Figure 8). Agriculture    nonforested areas near Florencia and in the mostly forested
was centered along the Cordillera Oriental, the eastern         lowlands.
slopes of the Andes. In states with both mountainous and
relatively flat areas, such as Pasco, Junin, and Cusco,         6. Discussion and Conclusions
agricultural activity was focused in mountain areas. While
pasture existed in the highlands from Pasco to Cusco,             [54] In the effort to determine the spatial pattern of land
cultivation was seen in the lower lying areas, particularly     use in the Amazon basin, it is clear that existing satellite
in San Martin, Loreto, and the lowlands of Cusco.               information alone is not sufficient. Satellite-based classifi-
                                                                cations that cover the entire study region have been limited
5.6. Colombia and Ecuador                                       mainly to classifying land cover, which meant that we could
  [52] Using the relationship between census and satellite      not simply use existing maps to identify cropland and
data, we used satellite imagery to determine agricultural       pasture. Although there were no existing land use maps
patterns in Colombia, which comprises 6% of the study           for the entire basin for the mid-1990s, we reasoned that this
area. Agricultural activity was highest along the eastern       important data source, if it existed, could be statistically
edge of the Cordillera Oriental, centered on Florencia.         blended with satellite imagery showing land cover. This
Smaller areas of agriculture existed along the Ajaju and
                                                      ´         technique appears to be successful at quantifying land use
Yari rivers in the forested lowlands.                           practices in Amazonia.
  [53] A similar pattern spatial pattern was seen in Ecua-        [55] Merging satellite-based land cover images with agri-
dor, where nearly all cropland and pasture was along the        cultural censuses appears to have been extremely success-
eastern slopes of the Andes, with limited lowland defor-        ful. The resulting cropland-pasture data set clearly captures
estation seen near Nueva Loja in Sucumbios and Tena in          the fine-scale variation of existing land cover classifications,
Napo. When compared to Landsat land cover data, this            while preserving the broad patterns of land use from census
land use information was highly correlated with nonforest       data. This new method for merging land use and land cover
(r = 0.79), as expected. A visual inspection of the two data    data sets was able to pull those categories and conditions


          Figure 11. Comparison for Colombia between Landsat-based land cover data and merged cropland-
          pasture data. See color version of this figure at back of this issue.
                                                CARDILLE ET AL.: LAND USE IN AMAZONIA                                                               18 - 13

from the satellite data that most closely coincided with                      Fearnside, P., Deforestation in Brazilian Amazonia: The effect of population
                                                                                and land-tenure, Ambio, 22(8), 537 – 545, 1993.
agricultural activity.                                                        Fearnside, P., Amazonian deforestation and global warming: Carbon stocks
  [56] The method described here has several advantages                         in vegetation replacing Brazil’s Amazon forest, For. Ecol. Manage.,
over using raw satellite-based land cover maps for the                          80(1 – 3), 21 – 34, 1996.
                                                                              Frolking, S., X. M. Xiao, Y. H. Zhuang, W. Salas, and C. S. Li, Agricultural
region. Because merged values are partly governed by                            land-use in China: A comparison of area estimates from ground-based
satellite imagery, it provides a counterbalance to the possi-                   census and satellite-borne remote sensing, Global Ecol. Biogeogr., 8(5),
bility that individual farmers or census organizations incor-                   407 – 416, 1999.
rectly underreport or overreport agricultural activity for                           ¸˜                                            ´
                                                                              Fundacao Instituto Brasileiro de Geografia e Estatıstica, Censo Agropecuar-   ´
                                                                                io: 1995 – 1996, Rio de Janeiro, 1997.
economic or political reasons. The method further allows                      Guild, L. S., J. B. Kauffman, L. J. Ellingson, D. L. Cummings, and E. A.
the density of agriculture to be estimated in areas without                     Castro, Dynamics associated with total aboveground biomass, C, nutrient
reliable census information, using flexible, region-based,                      pools, and biomass burning of primary forest and pasture in Rondonia,
                                                                                Brazil during SCAR-B, J. Geophys. Res., 103(D24), 32,091 – 32,100,
statistical relationships. The resulting fused data sets are at a               1998.
scale suitable for incorporation into models and are pre-                     Hansen, M. C., and B. Reed, A comparison of the IGBP DISCover and
sented on a regular grid, in which each cell represents the                     University of Maryland 1 km global land cover products, Int. J. Remote
                                                                                Sens., 21(6 – 7), 1365 – 1373, 2000.
best regional fit between census and satellite information.                   Hansen, M. C., R. S. Defries, J. R. G. Townshend, and R. Sohlberg, Global
  [57] The method described here is not limited to merging                      land cover classification at 1 km spatial resolution using a classification
census information with classified satellite imagery. For                       tree approach, Int. J. Remote Sens., 21(6 – 7), 1331 – 1364, 2000.
example, it is possible, though outside the scope of this                     Houghton, R. A., Tropical deforestation and atmospheric carbon dioxide,
                                                                                Clim. Change, 19(1 – 2), 99 – 118, 1991.
study, to return to the original AVHRR satellite reflectance                  Houghton, R. A., D. S. Lefkowitz, and D. L. Skole, Changes in the land-
data and use agricultural census information as training data                   scape of Latin America between 1850 and 1985, 1, Progressive loss of
to classify land use in the region.                                             forests, For. Ecol. Manage., 38(3 – 4), 143 – 172, 1991a.
                                                                              Houghton, R. A., D. L. Skole, and D. S. Lefkowitz, Changes in the land-
  [58] Finally, it is important to stress the need for more                     scape of Latin America between 1850 and 1985, 2, Net release of Co2 to
complete and accurate descriptions of land use and land                         the atmosphere, For. Ecol. Manage., 38(3 – 4), 173 – 199, 1991b.
cover change across the planet. Here we have shown how                        Houghton, R. A., D. L. Skole, C. A. Nobre, J. L. Hackler, K. T. Lawrence,
merging satellite imagery and agricultural censuses can                         and W. H. Chomentowski, Annual fluxes of carbon from deforestation
                                                                                and regrowth in the Brazilian Amazon, Nature, 403(6767), 301 – 304,
provide a useful means of characterizing the geographic                         2000.
patterns of land use within the Amazon and Tocantins                          Ihaka, R., and R. Gentleman, A language for data analysis and graphics, J.
basins. This technique can easily be adapted to other regions                   Comput. Graphical Stat., 5, 299 – 314, 1996.
                                                                              Instituto Geografico Militar and Bolivia Ejercito Comando General, Atlas
of the world, provided that detailed agricultural census                        de Bolivia, 272 pp., La Paz, Bolivia, 1997.
information and satellite imagery are available.                                                           ´                ´           ´ ´
                                                                              Instituto Nacional de Estadıstica e Informatica Direccion Tecnica de Censos
                                                                                y Encuestas, III Censo Nacional Agropecuario, Lima, 1996.
                                                                              Instituto Nacional de Pesquisas Espaciais, Amazonia Deforestation: 1997 –
   [59] Acknowledgments. This work was supported by the NASA                                     ˜      ´
                                                                                1998, report, Sao Jose dos Campos, Brazil, 1999.
LBA-Ecology program and an EPA STAR graduate fellowship. Murray               International Association of Agricultural Economists, World Atlas of Agri-
Clayton and Pete Hoff provided crucial statistical guidance and help. Karin     culture: Under the Aegis of the International Association of Agricultural
Swanson provided extremely valuable help with gathering census data.            Economists, Istituto Geografico De Agostini, Novara, Italy, 1969.
Aurelie Botta, Mike Coe, and Navin Ramankutty gave good advice along          Iverson, L. R., and A. M. Prasad, Predicting abundance of 80 tree species
the way. Lisa Dent, Matthias Burgi, and Scott Mackay provided early             following climate change in the eastern United States, Ecol. Monogr.,
helpful comments on the merging technique. Many thanks are owed to the          68(4), 465 – 485, 1998.
various government agencies for their efforts to make this data public. A     Kauffman, J. B., D. L. Cummings, and D. E. Ward, Fire in the Brazilian
final thanks is owed to the thousands of anonymous census takers in the         Amazon, 2, Biomass, nutrient pools and losses in cattle pastures, Oeco-
Amazon basin; this project would not exist without them. The agricultural       logia, 113(3), 415 – 427, 1998.
census data and the gridded cropland/pasture data sets will be made           Lamon, E. C., and C. A. Stow, Sources of variability in microcontaminant
available on our web site ( upon publication.              data for Lake Michigan salmonids: Statistical models and implications for
                                                                                trend detection, Can. J. Fish. Aquat. Sci., 56, 71 – 85, 1999.
                                                                              Laurance, W. F., S. G. Laurance, L. V. Ferreira, J. M. Rankin-De Merona,
                                                                                C. Gascon, and T. E. Lovejoy, Biomass collapse in Amazonian forest
References                                                                      fragments, Science, 278(5340), 1117 – 1118, 1997.
                              ´                           ´           ´
Auge S. A. Editores, El Peru y Sus Recursos: Atlas Geografico y Economico,    Laurance, W. F., L. V. Ferreira, J. M. Rankin-De Merona, and S. G. Laur-
   235 pp., Lima, 1994.                                                         ance, Rain forest fragmentation and the dynamics of Amazonian tree
Belward, A. S., and T. R. Loveland, The DIS 1 km land cover data set,           communities, Ecology, 79(6), 2032 – 2040, 1998.
   Global Change Newsl. 27, pp. 7 – 8, Int. Geosphere-Biosphere Pro- Laurance, W. F., D. Perez-Salicrup, P. Delamonica, P. M. Fearnside, S.
   gramme, Toulouse, France, 1996.                                              D’Angelo, A. Jerozolinski, L. Pohl, and T. E. Lovejoy, Rain forest frag-
Breiman, L., Classification and Regression Trees, 358 pp., Wadsworth,           mentation and the structure of Amazonian liana communities, Ecology,
   Belmont, Calif., 1984.                                                       82(1), 105 – 116, 2001.
Cochrane, M. A., A. Alencar, M. D. Schulze, C. M. Souza, D. C. Nepstad, Loveland, T. R., Z. L. Zhu, D. O. Ohlen, J. F. Brown, B. C. Reed, and L. M.
   P. Lefebvre, and E. A. Davidson, Positive feedbacks in the fire dynamic      Yang, An analysis of the IGBP global land-cover characterization pro-
                                         Remote Sens., 65(9), 1021 – 1032, 1999.
   of closed canopy tropical forests, Science, 284(5421), 1832 – 1835, 1999.    cess, Photogramm. Eng.
Conselho Nacional de Geografia Divisao de Geografia, Atlas do Brasil: Michaelsen, J., D. S. Schimel, M. A. Friedl, F. W. Davis, and R. C. Du-
   Geral e Regional, 705 pp., Conselho Nacional de Geografia, Rio de            bayah, Regression tree analysis of satellite and terrain data to guide
   Janeiro, 1960.                                                               vegetation sampling and surveys, J. Veg. Sci., 5(5), 673 – 686, 1994.
DeFries, R. S., C. B. Field, I. Fung, G. J. Collatz, and L. Bounoua, Com-     Mittermeier, R. A., C. G. Mittermeier, N. Myers, and P. Robles Gil, Hot-
   bining satellite data and biogeochemical models to estimate global effects   spots: Earth’s Biologically Richest and Most Endangered Terrestrial
   of human-induced land cover change on carbon emissions and primary           Ecoregions, 430 pp., Cemex, Mexico City, 1999.
   productivity, Global Biogeochem. Cycles, 13(3), 803 – 815, 1999.           Moran, E., and E. Brondizio, Land-use change after deforestation in Ama-
Fearnside, P. M., Predominant land uses in Brazilian Amazonia, in Alter-        zonia, in People and Pixels: Linking Remote Sensing and Social Science,
   natives to Deforestation: Steps Toward Sustainable Use of the Amazon         edited by D. M. Liverman, pp. 94 – 120, National Academy, Washington,
   Rain Forest, edited by A. B. Anderson, pp. 233 – 251, Columbia Univ.         D. C., 1998.
   Press, New York, 1990.                                                     Neill, C., B. Fry, J. Melillo, P. Steudler, J. Moraes, and C. Cerri, Forest- and
18 - 14                                         CARDILLE ET AL.: LAND USE IN AMAZONIA

  pasture-derived carbon contributions to carbon stocks and microbial re-        in the Amazon: Satellite data from 1978 to 1988, Science, 260(5116),
  spiration of tropical pasture soils, Oecologia, 107(1), 113 – 119, 1996.       1905 – 1910, 1993.
Nepstad, D. C., C. Uhl, and E. A. S. Serrao, Surmounting barriers to forest    Townshend, J. R. G., et al., The NASA Landsat Pathfinder Humid Tropical
  regeneration in abandoned, highly degraded pastures: A case study from         Deforestation Project, paper presented at ASPRS Conference: Land Sa-
  Paragominas, Para, Brazil, in Alternatives to Deforestation: Steps Toward      tellite Information in the Next Decade, Am. Soc. for Photogramm. and
  Sustainable Use of the Amazon Rain Forest, edited by A. B. Anderson,           Remote Sens., Vienna, Va, 1995.
  pp. 215 – 229, Columbia Univ. Press, New York, 1990.                         Turner, I. M., Species loss in fragments of tropical rain forest: A review of
Nepstad, D. C., C. Uhl, C. A. Pereira, and J. M. C. daSilva, A comparative       the evidence, J. Appl. Ecol., 33(2), 200 – 209, 1996.
  study of tree establishment in abandoned pasture and mature forest of        Uhl, C., and J. B. Kauffman, Deforestation, fire susceptibility, and potential
  eastern Amazonia, Oikos, 76(1), 25 – 39, 1996.                                 tree responses to fire in the eastern Amazon, Ecology, 71(2), 437 – 449,
Nepstad, D., et al., Large-scale impoverishment of Amazonian forests by          1990.
  logging and fire, Nature, 398(6727), 505 – 508, 1999.                        Uhl, C., R. Buschbacher, and E. A. Serrao, Abandoned pastures in Eastern
Nobre, C. A., et al., The Large Scale Biosphere-Atmosphere Experiment in         Amazonia, 1, Patterns of plant succession, J. Ecol., 76(3), 663 – 681,
  Amazonia (LBA): Concise Experimental Plan, Winand Staring Cent. for            1988.
  Integrated Land, Soil and Water Res., Wageningen, Netherlands, 1996.         U.N. Food and Agriculture Organization, FAOSTAT Agricultural Data,
Prince, S. D., and M. K. Steininger, Biophysical stratification of the Ama-, Rome, 2000.
  zon basin, Global Change Biol., 5(1), 1 – 22, 1999.                          U.S. Department of Agriculture, Major World Crop areas and climatic
Ramankutty, N., and J. A. Foley, Characterizing patterns of global land use:     profiles, USDA Agric. Handb. 664, 279 pp., Washington, D. C., 1994.
  An analysis of global croplands data, Global Biogeochem. Cycles, 12(4),      Venables, W. N., and B. D. Ripley, Modern Applied Statistics With S-PLUS,
  667 – 685, 1998.                                                               548 pp., Springer-Verlag, New York, 1997.
Ramankutty, N., and J. A. Foley, Estimating historical changes in global       Verburg, P. H., and Y. Q. Chen, Multiscale characterization of land-use
  land cover: Croplands from 1700 to 1992, Global Biogeochem. Cycles,            patterns in China, Ecosystems, 3(4), 369 – 385, 2000.
  13(4), 997 – 1027, 1999.                                                     Williams, M. R., and J. M. Melack, Solute export from forested and par-
Rand McNally, et al., The New International Atlas, Chicago, Ill., 1994.          tially deforested catchments in the central Amazon, Biogeochemistry,
Republica de Bolivia Instituto Nacional de Estadıstica, II Censo Nacional        38(1), 67 – 102, 1997.
  Agropecuario (1984): Resultados Departamentales, 162 pp., La Paz,            Wood, C. H., and D. Skole, Linking satellite, census, and survey data to
  1990.                                                                          study deforestation in the Brazilian Amazon, in People and Pixels: Link-
Rosenfeld, D., TRMM observed first direct evidence of smoke from forest          ing Remote Sensing and Social Science, edited by D. M. Liverman, pp.
  fires inhibiting rainfall, Geophys. Res. Lett., 26(20), 3105 – 3108, 1999.     94 – 120, National Academy, Washington, D. C., 1998.
Sistema Estadıstico Agropecuario Nacional, Instituto Nacional de Estadıs- ´
  tica y Censos, Encuesta Nacional de Superficie y Produccion Agrope-
  cuarias de Areas, Quito, 1995.                                                À À ÀÀÀ ÀÀÀÀÀÀÀ
                                                                               À À À À À À À À À À À ÀÀ
Skole, D. L., Data on global land-cover change: acquisition, assessment,         J. A. Cardille and J. A. Foley, Center for Sustainability and the Global
  and analysis, in Changes in Land Use and Land Cover: A Global Per-           Environment (SAGE), Institute for Environmental Studies, University of
  spective: Papers Arising From the 1991 OIES Global Change Institute,         Wisconsin, 1710 University Avenue, Madison, WI 53705, USA. (cardille@
  edited by W. B. Meyer and B. L. Turner, 537 pp., Cambridge Univ. Press,;
  New York, 1994.                                                                M. H. Costa, Department of Agricultural Engineering, Federal University
Skole, D., and C. Tucker, Tropical deforestation and habitat fragmentation          ¸       ¸
                                                                               of Vicosa, Vicosa, MG 36571-000, Brazil. (

                           CARDILLE ET AL.: LAND USE IN AMAZONIA


Figure 2. International Geosphere-Biosphere Programme (IGBP) DISCover (left) and Global Land
Cover Facility (right) land cover classifications for the Amazon and Tocantins basins.


Figure 6. Mid-1990s agricultural census data for the Amazon and Tocantins basins. In Bolivia, Peru,
and Brazil, data are at a spatial scale analogous to a U.S. county; in Ecuador, state-level data are

                                        18 - 3 and 18 - 6
                            CARDILLE ET AL.: LAND USE IN AMAZONIA


Figure 8. Merged satellite and agricultural census data for the mid-1990s. Each 5-min cell in Figure 8a
contains the estimated fraction of cropland and pasture estimated through the fusion technique merging
land use and land cover data. Relative proportions of land use types from agricultural census data are
used to estimate amounts of (b) cropland, (c) natural pasture, and (d) planted pasture.


Figure 9. Comparison for Bolivia between Landsat-based land cover data and merged cropland-pasture

                                         18 - 9 and 18 - 11
                         CARDILLE ET AL.: LAND USE IN AMAZONIA

Figure 10. Comparison for Rondonia, Brazil, between Landsat-based land cover data and merged
cropland-pasture data.

Figure 11. Comparison for Colombia between Landsat-based land cover data and merged cropland-
pasture data.

                                          18 - 12

Shared By: