Below are the first 10 and last 10 pages of uncorrected machine-read text (when available) of this chapter, followed by the top 30 algorithmically extracted key phrases from the chapter as a whole.
Intended to provide our own search engines and external engines with highly rich, chapter-representative searchable text on the opening pages of each chapter.
Because it is UNCORRECTED material, please consider the following text as a useful but insufficient proxy for the authoritative book pages.
Do not use for reproduction, copying, pasting, or reading; exclusively for search engines.
OCR for page 70
4
Linking Satellite, Census, and Survey Data to
Study Deforestation in the Brazilian Amazon
Charles H. Wood and David Skole
Advances in remote sensing technology undoubtedly rank among the most
significant contributions to the study of environmental topics in recent decades. The
ability to use orbiting platforms to measure the magnitude, pace, and pattern of land-
cover change has been particularly relevant to the study of the Brazilian Amazon, a
region that has experienced one of the highest rates of deforestation worldwide.
High-resolution satellite data provide a firm empirical base for measuring
the amount and the spatial configuration of forest clearing, but they do not them-
selves explain the causes of deforestation. It is well understood that, beyond the
need for refined measurement, explanations and projections of land-cover change
depend critically on the ability to model the social determinants of deforestation.
When the concern is for large regions, such as the Amazon Basin, population and
agricultural censuses are virtually the only source of regionwide data on the
socioeconomic and demographic characteristics of the population. These consid-
erations suggest the prospect of modeling the causes of deforestation by using a
data set that links the estimates of land-cover change derived by satellite images
to the social indicators generated by the various censuses.
A regionwide research design based on satellite and census data was a natu-
ral vehicle for a productive collaboration between a systems ecologist with exper-
tise in remote sensing technologies (Skole), and a social demographer with field
experience in the Brazilian Amazon (Wood). Collaborations of this kind, al-
though hardly new, have been relatively rare, at least in the context of Amazonian
research. In the case of this particular collaboration, the joint effort can be traced
in large measure to trends internal to both research traditions, the implications of
which provided the impetus for the present partnership.
70
OCR for page 71
CHARLES H. WOOD AND DAVID SKOLE
71
The National Aeronautics and Space Administration (NASA)-funded
Pathfinder project achieved international recognition for its singular contri-
bution to the production of accurate estimates of deforestation for the Ama-
zon region as a whole (Skole and Tucker, 1993~. The estimates, which were
years in the making, represented a timely contribution to a controversial and
highly politicized debate regarding the amount of land clearing that had
taken place in northern Brazil (e.g., World Bank, 1992~. Although disputes
of this import are never fully resolved, it is safe to say that the publication of
the results went a long way toward settling some of the major controversies
in the field.
Ironically, perhaps, the findings also underscored a fundamental limitation
of the Pathfinder data on deforestation, namely the inability to explain the reasons
for the observed outcomes in land-cover change. Concern over the limited ability
to explain the social causes of deforestation has grown in recent years in propor-
tion to the priority accorded the so-called human dimensions of global change.
Attention to these human dimensions within both scholarly and funding institu-
tions, in turn, has compelled members of the remote sensing community to go
beyond the question of "how much?" to address the question of "why?" The
change, sometimes stated in terms of a shift in focus from "pattern to process"
(Skole, 1997), means remote sensors have increasingly been thrust from the
relative safety of the grid-cell maps to which they were accustomed into the
turbulent waters of economics, politics, and sociology and other disciplines within
the social sciences.
Wood arrived at the partnership by traveling in the opposite direction. After
completing a 15-year longitudinal study of a particular site within the Brazilian
Amazon, he had grown impatient with the "why?" question and wanted, instead,
to know "how much?" The in-depth study of frontier change in which he had
been engaged produced a detailed account of the events that led to the massive
deforestation of the southern region of the state of Para (Schmink and Wood,
1992~. Yet for all the advantages of the "thick description" produced by the case
study method, the findings were inherently limited to a single locale, leaving
unanswered whether the same degree of deforestation was under way in other
parts of the region. From the vantage point of the research site, it was impossible
to determine whether southern Para was a special case or was typical of what was
happening elsewhere.
Despite the disparate routes taken to arrive at the point of collaboration, it
was easy to agree on the main objective of the project to explore the feasibility
of constructing a regionwide model of the determinants of land-cover change in
the Brazilian Amazon. The modeling exercise has the potential to address two
different albeit related lines of inquiry. One, which is common to the questions
raised by social scientists, looks to empirical results for explanations of the defor-
estation in the region. Attention focuses on the covariates of land-cover change
as a means to identify and rank in importance the socioeconomic and demo
OCR for page 72
72
LINKING DATA TOSTUDYDEFORESTATIONINTHEBRA7;1LIANAMA:Z;ON
graphic variables associated with forest clearing. Another line of research, more
common to global modelers, looks to the statistical covariations as a tool for
projecting the probable future levels of deforestation that are likely to be associ-
ated with assumed changes in the socioeconomic indicators.)
At this stage in the project, it remains to be seen whether the joining of
satellite and census data can produce robust and valuable findings. It is worth
learning this for the simple reason that it is always more cost-effective to use
existing sources of data than to produce new data. Since both satellite images of
land-cover change and census-based indicators of sociodemographic change are
available for many parts of the world, the lessons learned from this effort are
potentially applicable to places beyond the Brazilian Amazon.2 Even if our
objective is met only partially, a careful assessment of the strengths and the
weaknesses of the research design can provide important insights for similar
projects in other regions of the world.
With such an assessment in mind, our purpose here is to summarize the
methods and present the preliminary results of our NASA-funded project.
To establish the substantive context for the discussion, the next section de-
fines the geographic scope of the study and presents a brief history of the
factors that led to the migration of people into Amazonia and to the clearing
of vast stretches of tropical forest. Next we summarize the rationale for
using satellite and census data to construct regionwide models of the social
and demographic determinants of deforestation. The following section de-
scribes the methods used to estimate land-cover change and to generate the
sociodemographic indicators. We then review the problems associated with
merging the two types of data. The final two sections present the findings of
our initial efforts to establish the covariates of deforestation in the Amazon
and describe a proposed method for using field work to establish ground
truth for the statistical models.
THE BRAZILIAN AMAZON
Defining the Region
The geographic boundaries of Amazonia can be defined in various ways.
The Amazon River drainage basin in the South American continent is an area of
approximately 6,600,000 km2 that includes land in Brazil, Colombia, Ecuador,
Peru, Bolivia, and Venezuela. Within Brazil, the states of Acre, Amapa,
Amazonas, Para, Rondonia, and Roraima an area of around 3,500,000 km2-
are referred to as "Classical Amazonia" or "North Region." The last and most
commonly used definition (and the one used here) is the "Legal Amazon," a
federal planning designation that conforms more or less to the watershed within
Brazil's national boundaries. It consists of the North Region, plus the states of
Mato Grosso, Tocantins, and Maranhao west of the 44th Meridian (Figure 4-1~.
OCR for page 73
CHARLES H. WOOD AND DAVID SKOLE
f
JO AMA~AS
\ C~
W: ~
~ LORAN (
_ ~
Waif::
I ~
~-~ ~ AlVIAPA ~
:' ~
i_
J
/
l
/
/ PARA
/
\
~( ~
~- ~
~_
~ROr.ON ~0
-
r
and,
, GOIAS
Into
) DO SUL
: J
1~l ~ .
1'
11 CEARA RIO GRANDE DO NORll:
~ _-, ~W~
ySlLRGIPE
r ~ ~+
ARIA
~ iP~
f ~ fys`Nro
~ ~f I
- .C_ GERus ~
IS RIO DE
- JANEIRO
PAWO
ItlO GRENADE /
W_ By/
FIGURE 4-1 States of the Legal Amazon, 1980.
Development Policy, Land Settlement, and Deforestation
73
The contemporary movement of people into the Brazilian Amazon began in
the 1970s when the agricultural frontier moved into the northern states of Para,
Tocantins, and Rondonia. Whereas earlier periods of expansion were relatively
spontaneous, the exploitation and settlement of Amazonia in the 1970s were
aggressively promoted by the federal government. Development policies de-
signed to populate the region included credit and tax incentives to attract private
capital to the region, construction of the Transamazon Highway, and the coloni-
zation of small farmers on 100-hectare plots along both sides of the new road
(Fearnside, 1986; Moran, 1981; Smith, 1982~. Colonization projects organized
by the Institute of Colonization and Agrarian Reform (INCRA) attracted mi-
grants from all parts of Brazil, who soon arrived in numbers that far exceeded
INCRA' s capacity to absorb in the planned communities. With few alternatives
available to them, newcomers to the region who could not find a place in the
colonization areas simply cleared whatever land they could find, often to be
OCR for page 74
74
LINKING DATA TOSTUDYDEFORESTATIONINTHEBRA7;1LIANAMAZON
dispossessed later by ranchers and land speculators (Wood and Schmink, 1978~.
In as little as 2 or 3 years, places that once held a handful of families exploded
into makeshift towns of 15,000 to 20,000 people. According to recent estimates
of population growth in the Amazon, the states in northern Brazil experienced a
net in-migration of nearly 1.6 million people between 1970 and 1991 (Wood and
Perz, 1996~.
At the same time that small farmers migrated northward in search of land,
well-financed investors took advantage of profitable tax and credit programs
offered by the Superintendency for the Development of the Amazon (SUDAM)
to convert huge tracts of land to pasture and to buy land to hold in investment
portfolios as a hedge against future inflation (Hecht, 1985; Mahar, 1979~. Capi-
talized investors came mostly from the southern part of the country, where land
values were high relative to the price of land in the Amazon. In the early 1980s,
for example, a rancher could obtain 15 hectares in the Amazon for every hectare
sold in the south (World Bank, 1992:12-13~. To increase the size of their hold-
ings substantially, ranchers sold out in the south and moved to the north, where
they cleared the forest for pasture. The tendency to deforest among large land-
holders was further stimulated by the progressive features of Brazil's Land Stat-
ute, which levied a 3.5 percent tax on the value of unused (i.e., forested) land
(Binswanger, 1991~.
Although much, if not most, of the deforestation that took place in the Ama-
zon was carried out by medium- and large-scale ranchers, small farmers were
also implicated in the process, as evidenced by the typical cycle of land use.
Small farmers commonly clear 2 to 3 hectares of land, which they cultivate for as
long as soil fertility remains high. In most areas, soil fertility is depleted in 2 to
3 years, necessitating the clearing of more land. Since there are approximately
500,000 small farmers in the region, these figures imply a demand for an addi-
tional 500,000 hectares of cleared land per year (Homma et al., 1992:9~. Crude as
these estimates may be, they nonetheless point to the magnitude of the existing
internal demand for land clearing, even if the migration of small farmers to the
Amazon were to stop altogether.
Beginning in the mid-1970s, violence became commonplace as cattle ranch-
ers, land speculators, peasant farmers, and Indian groups competed for control of
the newly accessible territories. In a rural context characterized by violent com-
petition for land and in the absence of clear property rights to guarantee owner-
ship, individuals asserted their land claims by clearing the forest cover, often to a
much greater degree than was economically necessary.
The direct cause of deforestation in Amazonia was thus the change in land
use that came about as a consequence of the decline of fishing, forest extraction,
and shifting small-plot agriculture. These traditional forms of rural sustenance
were replaced in economic importance by the emergence of large peasant farm-
ing communities and the creation of pastures for cattle raising associated with the
influx of migrants into the region. Table 4-1 presents estimates of the magnitude
OCR for page 75
CHARLES H. WOOD AND DAVID SKOLE
TABLE 4-1 Area Deforested in Legal Amazon,
Brazil, 1978 and 1988
Area Deforested (in km2)
State19781988
Acre2,6126,369
Amapa182210
Amazonas2,30011,813
Maranhao9,42631,952
Mato Grosso21,13447,568
Para30,44995,075
Rondonia6,28123,998
Roraima1961,908
Tocantins5,68811,431
Total78,268230,324
SOURCE: Skole and Tucker (1993:1906).
75
of deforestation in various states in the Legal Amazon during 1978 and 1986.
The results indicate that the size of deforested areas rose from 78,268 km2 in
1978 to 230,324 km2 in 1988. These figures imply an annual average rate of
deforestation of 15,000 km2 per year during the period. The table shows that the
highest rates of land-cover change in both years took place in Para and Rondonia,
the primary destination of heaviest migration flows into the region.
RATIONALE FOR THE RESEARCH DESIGN
Analysts have applied different approaches to study the determinants of
deforestation. Numerous cross-country studies conclude that population growth
and land-cover change are strongly correlated (e.g., Allen and Barnes, 1985;
Rudel,1989~. Studies of this kind, however, based on highly aggregated units of
analysis (countries), generally offer limited insights into the dynamics of land-
cover change as compared, for example, with regional analyses that are carried
out within countries and make use of a wider range of independent variables (e.g.,
Reis and Guzman, 1992; Pfaff, 1997~. The most detailed results unsurprisingly
come from case studies, which have been especially valuable in producing highly
nuanced analyses of particular sites in the Brazilian Amazon. Examples include
studies of the history of land settlement in southern Para (Schmink and Wood,
1992), surveys of the agricultural practices of colonists in the Altamira region
(Moran, 1981; Walker et al., 1993), and economic assessments of public and
private colonization projects (Almeida, 1992~. Most studies carried out in this
tradition have relied on interviews and surveys, although more recent analyses
OCR for page 76
76
LINKING DATA TOSTUDYDEFORESTATIONINTHEBRA7;1LIANAMA:Z;ON
have sought to combine the data produced by conventional social science meth-
ods with the satellite-generated information on a particular scene (e.g., Brondizio,
1996; Moran et al., 1994a; Moran and Brondizio, in this volume). The case study
approach has been especially valuable in producing detailed, often historically
informed treatments of the events that are the cause of land-use and land-cover
change in a particular place.
Yet for all its advantages, the time and resource intensity of the case study
method precludes its application to large areas. Moreover, the conclusions gen-
erated by a case study approach do not go very far toward answering a broad
range of questions. For example, the availability of satellite images for large
areas in this case, the data for the entire Legal Amazon produced by the NASA
Pathfinder project opens up the possibility of constructing regionwide models
of the human dimensions of deforestation. Regional models, in turn, respond to
the call for empirical results that are comparable from one country to another and
can serve as inputs to improve the modeling and projection of various kinds of
global dynamics. Many of the global change models, including those dealing
with climate and trace-gas dynamics, rely on projections of land-cover change for
countries across the world.
The latter projections require a coordinated program of comparative studies
conducted at the regional level that specify the relationships between land-use
and land-cover change, and a common set of independent variables (and their
surrogates), such as changes in population size, distribution, and density, and
changes in economic structure and technology. With these objectives in mind, it
is worth noting that census data not only in Brazil, but in other countries as
well are nearly always the only source of comparable sociodemographic data
for large areas. By the same token, satellite images, which can be obtained for
almost any place on the globe, are virtually the only source of accurate and
georeferenced data on land cover for large geographic expanses.
Because of their potential contribution to global environmental modeling,
the production and testing of regional models is a goal that has been promoted by
a host of influential international institutions. The International Geosphere-Bio-
sphere Programme (IGBP) and the International Human Dimensions Programme
on Global Environmental Change (IHDP), through the Land-Use/Cover Change
(LUCC) project, have called for a common protocol for studies that make use of
existing data sources (Turner et al., 1994:93~. The goals set forth in the IGBP-
IHDP agenda are echoed in parallel documents produced by the Committee on
the Human Dimensions of Global Change of the International Social Science
Council (ISSC), prepared in cooperation with UNESCO (Jacobson and Price,
1991), and by the National Research Council (NRC) (1992~. Similar recommen-
dations have been put forth by the Social Science Research Council Committee
for Research on Global Environmental Change and the 1991 Global Change
Institute on Global Land Use Change, sponsored by the Office of Interdiscipli-
nary Earth Studies. A data set that merges satellite-based estimates of land-cover
OCR for page 77
CHARLES H. WOOD AND DAVID SKOLE
77
change with census-based indicators of socioeconomic and demographic struc-
ture thus has the potential to go a long way toward meeting the aims of the IGBP-
ISSC-NRC scientific agenda.
DATA AND DESIGN
Satellite Estimates of Deforestation
The measures of deforestation used in this study were generated by Skole as
part of the Landsat Pathfinder Tropical Deforestation Project (funded by NASA,
the Environmental Protection Agency, and the U.S. Geological Survey) at the
Institute for the Study of Earth, Oceans and Space of the University of New
Hampshire, in collaboration with NASA' s Goddard Space Flight Center and the
Department of Geography at the University of Maryland. Images were obtained
from the U.S. national archive at the Earth Resources Observation System (EROS)
Data Center, from foreign ground stations, and from programmed acquisitions.
The satellite data were preprocessed at the EROS Data Center to a standard
format and projection (Universal Transverse Mercator) and sent on 8 mm tape to
the University of New Hampshire for analysis. The image thresholding method
was used to identify seven thematic features: forest, deforestation, secondary
forests, water, clouds, cloud shadows, and cerrado (natural savanna). The data-
bases were compiled from 210 Landsat Multispectral Scanner (MSS) scenes at a
spatial resolution of 57 m. Because of both the methodology used and the nature
of digital remote sensing, the output classification was not entirely accurate.
Therefore, the classification was edited manually using the geographic informa-
tion system (GIS). The vector product was plotted at 1:250,000 scale on vellum
using an electrostatic plotter. The vellum plot was then overlaid on a 1:250,000
scale colorfire photoproduct of the Landsat scene, and misclassified polygons
were identified and corrected. The vector coverage was repeatedly plotted and
checked until the classification had been completed. Individual digitized scenes
were projected into geographic coordinates (latitude and longitude), edge-
matched, and merged into a sinusoidal equal-area projection to create a final
digital map from which all calculations of area were made.
The areas of the Amazon deforested by human activities were defined using
spectral characteristics of deforested sites. These characteristics were developed
through field measurements at five calibration sites in the basin. It is rather easy
to distinguish deforested areas from intact virgin forests since the spectral charac-
teristics of the two are very different. Because there are some problems in
differentiating cerrado, the study was confined to the closed-canopy forest re-
gion. Accuracy assessment was performed in the field using the Global Position-
ing System (GPS) and standard methods of assessment based on contingency
tables. Overall accuracy was better than 95 percent for more than 300 check-
points. Kappa and Tau statistics were also computed following the method of Ma
OCR for page 78
78
LINKING DATA TOSTUDYDEFORESTATIONINTHEBRA7;1LIANAMA:Z;ON
and Redmond (1995) (97 and 99 percent, respectively). In addition, sample
SPOT scenes at 20 m resolution were compared with our analysis, which used
Landsat MSS and Thematic Mapper (TM) data. These intersensor comparisons
were in agreement to within 6 percent. A complete description of the methods
used to process and analyze the deforestation data set can be found in Skole
(1992) and on the project Web site (http://pathfinder-www.sr.unh.edu/pathfinder).
Cloud cover is a serious problem in the tropics. However, using the entire
catalog of all Landsat scene acquisitions (several hundred thousand) contained in
the U.S. and Brazilian archives made it possible to select a data set for specific
years that was generally free of clouds. The data set reported here was almost
completely cloud free; less than 10 percent of the surface area was contaminated
by clouds, mostly in the state of Amapa (Skole and Tucker, 1993~.
A regional portrait of the spatial distribution of deforestation is given in Plate
4-1 (after page 150), for circa 1986. Identical estimates are forthcoming for circa
1992. The pattern of deforestation in 1986 clearly shows a crescent shape that
corresponds to the expansion of the agricultural frontier into the southern part of
the Amazon region. The land-cover change associated with the construction of
roads is similarly evident in the lines of deforestation that stretch across the
center of the map. The presence of large areas of savanna is also evident, as
indicated by the band across the bottom of the map and several patches to the far
north. It is important to eliminate the savanna regions in analyses of deforesta-
tion because these areas are not the outcome of deforestation, but were always
naturally unforested.
Census Estimates of Demographic and Economic Structure
Indicators of demographic and economic structure presented here were de-
rived from the 1980 population and agricultural censuses. Identical indicators are
forthcoming for 1991 (the date of the most recent demographic census). For the
demographic estimates, we used a micro data set that represented 25 percent of
the complete enumeration. The large sample size enabled us to disaggregate the
variables down to the municipio level. Because the data are available in the form
of individual records, we were able to generate a number of indicators that are not
present in the published materials. The census tapes, for example, contain infor-
mation on 86 variables, of which 26 refer to housing characteristics and the
remaining 60 to characteristics of individuals (with personal identifiers removed).
The latter variables include information on age, sex, relationship to head of
household, rural-urban location, place of birth, migration, length of residence,
education, marital status, occupation, industry, class of worker, and income.
The data on occupation and industry categories serve as indicators of the
proportion of the labor force engaged in various economic activities. Of special
significance in the frontier setting are the numbers of people working on ranches
and in agriculture. To these demographic characteristics we added additional vari
OCR for page 79
CHARLES H. WOOD AND DAVID SKOLE
79
ables drawn from the agricultural census, such as the number of cattle in a municipio
and the area of land devoted to ranching and the production of subsistence crops,
such as rice and beans. Annex 4-1 presents a list of the satellite- and census-based
indicators generated for each municipio in Brazil's Legal Amazon region.
The data set for this study will thus be constructed by merging in a GIS the
satellite- and census-based variables for each of the municipios that comprise the
Legal Amazon (353 in 1980, 482 in 1991~. The municipio boundaries in 1980 are
shown in Figure 4-2. A glance at this map is sufficient to appreciate the highly
irregular character of the municipios in the region. In the eastern region (in the
state of Para and in the states of Maranhao and Tocantins), population density is
high, and the municipios are small in size. This pattern contrasts with the western
and northern regions (especially the states of Amazonas, Roraima, and Amapa),
where population density is low, and the municipios are quite large. The irregular
shape of the geopolitical boundaries has important implications for the spatial
correspondence between the satellite and census data.
SPATIAL SCALES AND SPATIAL CORRESPONDENCE
To merge the satellite and census data into a single data set, the land-cover
data at 57 m resolution were aggregated to conform to the boundaries of each
municipio, the smallest spatial unit for which economic and demographic data
are available. In effect, this meant we had to reconfigure the detailed information
depicted in Plate 4-1 to conform to the much larger and highly irregular geopoliti-
cal boundaries depicted in Figure 4-2, which resulted in the pattern shown in
Figure 4-3. When the land-cover data were reconfigured to municipal bound-
aries, the crescent shape of the agricultural frontier remained plainly visible, yet
transforming the data to a coarser scale was done at the cost of spatial precision.
The implications of reconfiguring the finely graded raster data to the coarser
scale of municipio boundaries can be appreciated by contrasting the present data
set with the ideal case. In the best of worlds, there would be a perfect correspon-
dence between the spatial definition of the dependent and independent variables
used in the analysis. In other words, the classification of land cover derived from
the satellite images (dependent) would refer precisely to the land-cover charac-
teristics of the area lying within the boundaries of each rural establishment. By
the same token, the sociodemographic variables generated from the census data
(independent) would refer precisely to the characteristics of the actor~s) respon-
sible for making land-use decisions within the corresponding spatial unit. Such
congruence would ensure that the dependent/independent variables were linked
at the level of the decision unit involved. In this way, the analysis would maxi-
mally exploit the fine tuning made possible by advances in the production and
analysis of satellite images and, by virtue of corresponding to the behavioral unit
involved in land-use decisions, would avoid problems of interpretation associ-
ated with "ecological correlations" (described below).
OCR for page 80
80
Hi,
.
- I'm ~
-, ~
': ~
:. .- ' .
~An,
= , ~-~
an' ~ ; an'
N ~-
To
00
o
Ct
o
Cot
o
·_.
·_.
.=
OCR for page 83
CHARLES H. WOOD AND DAVID SKOLE
TABLE 4-2 Variable Names, Descriptions, and Definitions
83
Name Description Definition
AREA Size of municipio Total number of square kilometers within the
geographic boundaries of the municipio.
DEF Deforestation Number of kilometers within the municipio
classified as 'deforested' in 1986. Excludes
naturally unforested areas (cerrado).
DEFECT Percent deforested DEF/AREA.
TOTPOP Total population Total number of people enumerated in the
municipio in the 1980 demographic census.
POPDNS Population density TOTPOP/AREA.
CLOUDPCT Percent of area under Percent of the total municipio that was under
clouds cloud cover at the time of the satellite
image.
SAVANPCT Percent in savanna Percent of the total municipio classified as
naturally unforested areas (cerrado).
RMIGDNS Rural migration Total number of migrants in rural areas/AREA.
density
RMIGSQR Rural migration RMIGDNS squared.
density squared
FARMDNS Farm density Total number of heads of household
classified as farmers in the 1980
demographic census/AREA.
RANCHDNS Ranch density Total number of heads of household
classified as ranchers in the 1980
demographic census/AREA.
LT50HA Less than 50 hectares Percent of rural establishments less than
50 hectares in size.
GT1OOOHA More than 1,000 Percent of rural establishments greater than
hectares 1,000 hectares in size.
CONFLICT Conflict proxy Density of cattle times density of area
devoted to foodcrops.
NOTE: AREA, DEF, DEFPCT, CLOUDPCT, and SAVANPCT are from satellite images; the re
mainder are from demographic and agricultural censuses; see Annex 4.1.
The positive sign on the population density variable and the negative sign on its
square are findings consistent with this expectation.
Total population density is, however, a variable that is subject to potential
bias because the total number persons (numerator in the density ratio) includes
people living in both urban and rural areas. Although the size of the urban
population is not irrelevant to the study of land-cover change, it is plausible to
argue that land-cover change is more closely related to the number of new arriv
als in rural places. With this notion in mind, we selected the density of migrants
in a rural area. We did so on the assumption that a relatively large number of
OCR for page 84
84
LINKING DATA TOSTUDYDEFORESTATIONINTHEBRA7;1LIANAMA:Z;ON
TABLE 4-3 Various Measures of Population Regressed on Percent
Deforestation (unstandardized ordinary least squares coefficients)
Variable Model 1 Model 2 Model 3 Model 4 Model 5Model 6
CLOUDPCT -.302 -.191 -.063 .009 .001.005
SAVANPCT -.120* -.098* -.075* -.108* -.113*-.118*
POPDNS .311 * .609*
POPSQR -.002*
RMIGDNS 8.586* 8.644* 8.594*7.206*
RMIGSQR -.318* -.318* -.317-.348*
RANCHDNS 31.09* 30.68*27.22
FARMDNS 4.650 4.4352.831
FISHDNS -11.96 -14.65-6.149
MINEDNS -12.60 -14.76-17.492
LT50HA .031.013
GT1OOOHA .171.145
CONFLICT .028*
R2 .192 .237 .373 .387 .389.453
*Coefficient is statistically significant (p < .05).
newcomers would correspond to areas undergoing an expansion in the agricul-
tural frontier, which would therefore be more likely to experience a high rate of
deforestation. This hypothesis is borne out by the results of Model 3, which
shows that the rural migration density variable is statistically significant, as is its
square, indicating diminishing effects on deforestation with a rise in density.
Moreover, the R2 (.373) is considerably higher as compared with the model based
on total population density (Model 2, R2 = .237~.
The next step in the analysis (Model 4) took into account the economic
characteristics of the population. The latter are especially valuable in the study of
land-cover change because different forms of land use have different conse-
quences for land-cover change. A municipio that has experienced the in-migra-
tion of, say, 1,000 ranchers is apt to have a higher degree of deforestation than a
place that was the destination of 1,000 farmers or fishermen. Census data on the
number of ranchers, farmers, miners, and fishermen thus provide proxy measures
of land use. As expected, the results indicate that the percentage of land defor-
ested is strongly associated with the density of ranchers, but not with the density
of farmers, miners, and fishermen.
Additional indicators available in the 1980 agricultural census suggest the
structure of land tenure in the region. Of particular interest are the percent of
rural establishments smaller than 50 hectares (ha) and the percent larger than
1,000 ha. The indicators of land tenure are of potential significance in light of the
widely held conviction that the majority of deforestation in the region occurs at
OCR for page 85
CHARLES H. WOOD AND DAVID SKOLE
85
the hands of large landholders. However, as noted in Model 5, neither variable is
statistically significant, and the inclusion of these two indicators added little to
the explanatory power of the model.
Finally, numerous studies of the process of frontier expansion in the Amazon
have called attention to the relationship between land conflict and deforestation.
In a social context in which tenure is highly insecure, landholders tend to clear
large amounts of land (often far more than they can cultivate), primarily to
strengthen de facto control over their land claims. Net of the effects of other
factors, a high degree of deforestation can therefore be anticipated in places
characterized by a high level of conflict. Conflicts over land, in turn, occur most
often between ranchers and small farmers (Schmink and Wood, 1992~. This
observation suggests that a proxy measure of land conflict could be obtained by
including in the equation an interaction term generated by multiplying the density
of cattle by the percent of land devoted to foodcrops (rice and beans, which are
characteristic of peasant production). Model 4 shows that, net of the other vari-
ables in the equation, the proxy for social conflict has a statistically significant
association with the percent of land deforested.
FINDING ANSWERS IN THE ERRORS
The results presented above are highly schematic, based as they are on a
limited number of variables and on information presently available at only one
point in time. Our goal is to expand the analysis by including additional
sociodemographic indicators from the population and agricultural censuses, and
by generating a merged satellite/census data set for two points in time (circa 1980
and 1991~. The expanded data set will allow a cross-sectional analysis of the
determinants of deforestation in both years, as well as an analysis of the changes
that took place over the period.
With the inclusion of additional variables, we expect to increase the explana-
tory power of the statistical models well beyond what has been presented here.
At the same time, statistical models of real-world processes even when they do
include a much wider range of independent variables will always contain error
terms. The errors reflect the difference between the actual amount of deforesta-
tion measured in a particular municipio and the value predicted by the least-
squares regression model. The errors can be visualized by tracing the least-
squares regression line through the array of deforestation values the model is
intended to predict. We can anticipate that some cases will fall well above the
regression line (+ outliers), indicating a much higher level of deforestation than
the amount anticipated by the model, while others will fall well below the regres-
sion line (- outliers), indicating a much lower level of deforestation than the
model predicts.
The positive and negative statistical outliers most likely result from the
failure to include in the model one or more variables highly correlated with
OCR for page 86
86
LINKING DATA TOSTUDYDEFORESTATIONINTHEBRA7;1LIANAMAZON
deforestation, or what is sometimes called "specification error." If the problem is
due to specification error, the only way to shed further light on the relationship is
to visit the municipios in question in order to collect additional information. The
goal of field work would be to identify the factors not already included in the
equation that account for much higher/lower levels of deforestation than the
model predicts. In other words, the strategy we have planned is to exploit all the
available sources of data to construct as robust a statistical model as possible of
the sociodemographic covariates of deforestation for the municipios in the Bra-
zilian Amazon. We will then use the model to identify a handful of extreme
outlier municipios (both + and -), which will become the targets of field investi-
gation.
Field work can also be used to address another type of potential error. In
addition to the problem of model specification (which produces statistical outli-
ers), it is necessary to pay attention to those municipios that fall along (or close
to) the regression line. In such cases, we are tempted to conclude that the model
"works" that the correlations we find in the model reflect true causality. The
problem with drawing such a hasty conclusion is the possibility that other unmea-
sured factors, correlated with the independent variables) in the equation, are the
true causes of the relationship. In this situation, the associations produced by the
statistical model are said to be "spurious." As in the case of specification error,
the only way to be sure that the relationships depicted in the equation are faithful
representations of real events is to visit the municipios in question in order to
collect additional information.6
Note that the research design for the field component of the project is sys-
tematically derived from the results of the modeling exercise. Research sites will
be selected by identifying a subset of municipios that are statistical outliers (to
address the issue of specification error) and a subset of municipios that are not
outliers (to reduce the potential for misinterpretation due to spurious associa-
tions). Similarly, the content of the questions to be asked in the field will be
tailored to the different research sites: in the case of positive (negative) outliers,
the purpose of the inquiry will be to determine what factors beyond those in-
cluded in the statistical model account for the higher (lower) level of observed
deforestation; in the case of municipios that are not outliers, the purpose will be
to explore the possibility that the associations found in the statistical model may
be due to other, unmeasured factors.
What we propose to do in the field work stage of the project can be thought
of as a variant of conventional ground trothing. The term generally refers to a
process by which one verifies the interpretation of a remotely produced image.
When the analyst assumes that a given pixel is, say, a deforested area, field work
is carried out to ground truth the image, making sure that the interpretation is, in
fact, correct. For the most part, the task is limited to establishing the correspon-
dence between the signature of a given pixel and what is actually observed on the
OCR for page 87
CHARLES H. WOOD AND DAVID SKOLE
87
ground. In this sense, ground truthing is a procedure that is arguably more
straightforward than what we have in mind.
In the context of the present study, the objective of field work is to verify a
relationship established in a statistical model.7 For example, when we determine
that deforestation is highly associated with some variable in a regression equa-
tion, the question becomes whether that relationship is really what one observes
on the ground. In effect, we are proposing to carry out what might usefully be
called "relational ground trothing." Although we have not yet put this method
into practice, it would appear to be substantially more complex than ordinary
ground trothing. Among other things, it is far from clear how one designs a field
project to address issues such as specification error (in the case of statistical
outliers) and spuriousness (in the case of municipios on or close to the regression
line). Indeed, the task of developing a relational ground truthing methodology is
one of the challenges we confront in the coming year. If we are successful in
doing so, our results have the potential to advance the process of integrating
satellite, census, and field data in the study of deforestation in the Amazon and
elsewhere in the world.
ACKNOWLEDGMENT
Thanks are due to Stephen Perz for his contribution to the construction of the
demographic and agricultural indicators used here and his help in the analysis of
the data.
NOTES
1 When data are available at two points in time, as in this study, it is possible to assess the
results by using the model at time 1 to predict the values at time 2, and then compare the projected
values with what is actually observed at time 2.
2 The potential relevance of this study to other regions in the world does not rest on the
assumption that the statistical patterns observed in Brazil will apply to other places, which is un-
likely. Instead, the relevance lies in testing the feasibility of merging census and satellite data in the
study of the social determinants of deforestation. The findings have the potential to generate insights
and caveats valuable to others wishing to apply the same or similar methods in other locations.
3 We eliminated municipios that are state capital cities, which are urban centers not relevant to
the present analysis.
4 The ecological fallacy can be thought of as a special case of spuriousness in which the
relationships found in the regression analyses are due to a shared spatial location, rather than a causal
connection.
5 Future analyses of these data will account for spatial effects, which are important in two
instances: (1) when the processes under study are intrinsically spatial, e.g., when they follow a
spatial diffusion pattern or incorporate adjacency effects; and (2) when models are estimated using
spatial (i.e., geographic) data for which the scale and unit of observation do not necessarily match the
scale and unit of the process. In Anselin (1988), these two types of spatial effects are referred to as
substantive spatial dependence and nuisance spatial dependence, respectively. Both are relevant to
regression models of deforestation processes. On the one hand, substantive spatial dependence
OCR for page 88
88
LINKING DATA TOSTUDYDEFORESTATIONINTHEBRA7;1LIANAMA2;ON
allows explicit consideration of the effects of adjacency in the model. That is, it provides a way to
model both forest dynamics and socioeconomic change as spatial (and/or space-time) processes. On
the other hand, given the nature of the data used to estimate deforestation models for example,
census variables collected at an administrative unit level and indicators of forest dynamics aggre-
gated to these administrative units it is highly unrealistic to assume that the scale of the observa-
tional units matches that of the processes under consideration. In both instances, ignoring the spatial
nature of the dependence causes problems of model misspecification.
6 Field work can also address the possibility of the ecological fallacy noted earlier.
7 Similar efforts to go beyond traditional ground truthing through extensive field work include
those that attempt to develop new age classes of secondary growth (Moran et al., 1994b) and to
understand management practices and intensification (Brondizio, 1996).
REFERENCES
Allen, J.C., and D.F. Barnes
1985 The causes of deforestation in developing countries. Annals of the Association of Ameri-
can Geographers 75(2): 163- 184.
Almeida, Ana Luiza Osorio de
1992 The Colonization of the Amazon. Austin: University of Texas Press.
Anselin, L.
1988 Spatial Econometrics: Methods and Models. Dordrecht, The Netherlands: Kluwer Aca-
demic.
Binswanger, H.P.
1991 Brazilian policies that encourage deforestation in the Amazon. World Development
19(7):821-829.
Brondizio, E.
1996 Land cover in the Amazon estuary: Linking the thematic mapper with botanical and
historical data. Photogrammetric Engineering and Remote Sensing 62(Aug.):921-929.
Fearnside, P.M.
1986 Human Carrying Capacity of the Brazilian Rainforest. New York: Columbia University
Press.
Hecht, S.B.
1985 Environment, development and politics: Capital accumulation and the livestock sector in
Eastern Amazonia. World Development 13(6):663-684.
Homma, A.K.O., R.T. Walker, F.N. Scatena, A.J. de Conto, R. de A. Carvalho, A.C.P.N. da Rocha,
C.A.P. Ferreira, and A.I.M. dos Santos
1992 A Dinamica dos Desmatamentos e das Queimadas na Amazonia: Uma Analise
Microeconomica. Unpublished manuscript, EMBRAPA, Belem, Para, Brazil.
Jacobson, H.K., and M.F. Price
1991 A Framework for Research on the Human Dimensions of Global Environmental Change.
Paris: International Social Science Council with UNESCO.
Ma, Z., and R. Redmond
1995 Tau coeffficients for accuracy assessments of classification and remote sensing data.
Photogrammetric Engineering and Remote Sensing 61(4):435-439.
Mahar, D.J.
1979 Frontier Development Policy in Brazil: A Study of Amazonia. New York: Praeger.
Moran, E.F.
1981 Developing the Amazon. Bloomington: Indiana University Press.
Moran, E.F., E. Brondizio, and P. Mausel
1994a "Secondary Succession." Research and Exploration 10(4, Autumn): 458-466.
OCR for page 89
CHARLES H. WOOD AND DAVID SKOLE
89
Moran, E. F., E. Brondizio, P. Mausel, and W. You
1994b Integrating Amazon vegetation, land-use and satellite data. BioScience 44(5, May):329-
338.
National Research Council
1992 Global Environmental Change: Understanding the Human Dimensions. Committee on
the Human Dimensions of Global Change. P.C. Stern, O.R. Young, and D. Druckman,
eds. Washington, D.C.: National Academy Press.
Pfaff, A.
1997 Spatial Perspectives on Deforestation in the Brazilian Amazon: First Results and a Spa-
tial Research Agenda. Paper presented in conference on Research Transformations in
Environmental Economics: Policy Design in Responses to Global Change, Durham,
N.C., May 5-6. Department of Economics, Columbia University.
Reis, E., and R.M. Guzman
1992 "An econometric model of Amazon deforestation." IPEA/Rio de Janeiro, Working Paper
265.
Rudel, T.K.
1989 Population, development, and tropical deforestation: A cross-national study. Rural Soci-
ology 54(3):327-338.
Schmink, M., and C.H. Wood
1992 Contested Frontiers in Amazonia. New York: Columbia University Press.
Skole, D.L.
1992 Measurement of deforestation in the Brazilian Amazon using satellite remote sensing.
Ph.D. dissertation, University of New Hampshire.
1997 From Pattern to Process. Presentation at the Open Meeting of the Human Dimensions of
Global Environmental Change Research Community, IIASA, Laxenburg, Austria, June
12-14.
Skole, D.L., and C.J. Tucker
1993 Tropical deforestation, fragmented habitat, and adversely affected habitat in the Brazilian
Amazon: 1978-1988. Science 260:1905-1910.
Smith, N.J.H.
1982 Rainforest Corridors: The Transamazon Colonization Scheme. Berkeley: University of
California Press.
Turner II, B.L., W.B. Meyer, and D.L. Skole
1994 Global Land-Use/Land-Cover Change: Toward an Integrated Study. Ambio 23(1):91-95.
Walker, R.T., A. Homma, F. Scatena, A. Conto, R. Carvalho, A. Rocha, C. Ferreira, A. Santos, and
R. Oliveira
1993 Sustainable farm management in the Amazon piedmont. Congresso Brasileiro de
Economia e Sociologia Rural 31 :706-720.
Wood, C.H., and M. Schmink
1978 Blaming the victim: Small farmer production in an Amazon colonization project. Studies
in Third World Societies 7:77-93.
Wood, C.H., and S. Perz
1996 Population and land use change in the Brazilian Amazon. Pp. 95-108 in Population
Growth and Environmental Issues, S. Ramphal and S. Sindig, eds. Westport, Conn.:
Praeger.
World Bank
1992 Brazil: An Analysis of Environmental Problems in the Amazon. World Bank Report No.
9104-BR. Washington, D.C.: World Bank.
OCR for page 90
9o
LINKING DATA TOSTUDYDEFORESTATIONINTHEBRA7;1LIANAMA:Z;ON
ANNEX 4-1
SOCIAL INDICATORS FOR MUNCIPIOS IN THE
BRAZILIAN LEGAL AMAZON, 1980
Items 1-11 are from the 1980 demographic census; items 12-20 are from
the 1980 agricultural census; items 21-22 are from satellite images, circa 1986.
Geographic Identifiers
2. Base Variables
3. Migration
4. Labor Force Composition
5. Age, Sex Composition
6. Child Survival
7. Fertility
Subregion
State
Microregion
Municipio
Deforestation Analysis Code
Total Population
Rural Population
Economically Active Population
Total Households
Rural Households
Total Population Aged 5+
Rural Population Aged 5+
Total Migrants
Rural Area Migrants
Northeast Origin Migrants
Number of Farmers
Number of Ranchers
Number of Forest Product Extractors
Number of Fishers
Number of Miners
Number of Day Laborers in Agriculture
Males and Females, from Ages 0-4 to 75+
at 4-Year Intervals
For Women Aged 15-19, 20-24, 25-29,
and 30-34:
Children Ever Born
Children Surviving
For Women Aged 15-19, ..., 45-49:
Infants Born During the Previous Year
OCR for page 91
CHARLES H. WOOD AND DAVID SKOLE
8. Income
9. Housing Quality
10. Literacy
11. Agricultural Producers
12. Land Use
13. Land Distribution
14. Agricultural Inputs
15. Agricultural Outputs
91
For Total and Rural Heads of Households:
Income in Minimum Wages: <1, 1-<2,
2-<3, 3+
For Total and Rural Households:
Housing Units with Mud Walls
Housing Units with Electricity
For Total and Rural Populations Aged 5+:
Literate Persons
Number of Owners, Renters, Tenants,
Occupants
Total Number of Rural Properties
Rural Land Area Claimed in Properties
Land Area Under Different Uses:
Annual and Perennial Crops
Fallow
Natural and Cultivated Pasture
Natural and Cultivated Forest
Land Not in Use
Number of Properties and Land Area in
Properties of <1 to 100,000+ ha
Number of Properties with <1 to 1,000+
ha of Cultivated Land
Use of Fertilizers
Number of Tractors
Value of Productive Goods
Value of Investments During Previous Year
Value of Credit During Previous Year
Value of Fuels Consumed
Amount of Various Fuels Consumed
Amount of Electricity Produced and
Consumed
Land Area, Production Yields for Annuals:
Sugar Cane
Rice
Beans
Manioc
OCR for page 92
92
LINKING DATA TOSTUDYDEFORESTATIONINTHEBRA7;1LIANAMAZON
Corn
Soybeans
Land Area, Production Yields, and
Number of Plants for Perennials:
Bananas
Rubber
Cacau
Coffee
Black Pepper
Number of Cattle
Cattle Sold or Slaughtered During
Previous Year
Milk Production
16. Extractive Products
17. Silviculture Products
18. Rural Industries
A~cai
Babassu Nuts
Rubber
Biomass Charcoal
Babassu Nut Charcoal
Castanha do Para
Firewood
Timber
Palm Heart
Firewood
Timber
Paper Pulp
Plantation Trees: Andiroba, Cedro,
Eucalyptus, Gmelina, Ipe, American
Pine, Ucuubeira
Sugar Cane Transformed:
For Sugar (Production Volume and Value)
For Cane Liquor (Production Volume and
Value)
For Molasses (Production Volume and
Value)
For Brown Sugar (Production Volume and
Value)
OCR for page 93
CHARLES H. WOOD AND DAVID SKOLE
19. Other Industrial Activity
20. Land Area
21. Land Cover
22. Roads
93
Milk Transformed:
For Cream (Production Volume and Value)
For Doce de Leite (Production Volume
and Value)
For Butter (Production Volume and Value)
For Cheese (Production Volume and Value)
Manioc Transformed:
For Manioc Meal (Production Volume
and Value)
For Tapioca Powder (Production Volume
and Value)
For Tapioca (Production Volume and Value)
Total Industrial Establishments:
Number of Mineral Extraction
Establishments
Number of Mineral Processing
Establishments
Number of Metallurgy Establishments
Number of Logging Establishments
Number of Rubber Product Establishments
Number of Rubber Processing Establishments
Square Kilometers 1980 Demographic
Census Estimate
Square Kilometers Satellite-Based
Estimate
Land Area Under Forest
Land Area Under Savanna
Land Area Deforested
Land Area Under Secondary Growth
Land Area Under Water
Land Area Under Clouds
Land Area Under Shadow
Kilometers of Roads
Representative terms from entire chapter:
david skole
