Spatially Explicit Model of Deforestation in Bolivia

Preprint of: Vanclay, J.K., D. Kaimowitz, A. Puntodewo and P.Mendez, 1999. Spatially explicit model of
deforestation in Bolivia. In: Y. Laumonier, B.King, C.Legg and K. Rennolls (eds) Data Management and Modelling
using Remote Sensing and GIS for Tropical Forest Land Inventory. Rodeo, Jakarta, p. 371-382. ISBN 979-95696-0-5.

Spatially Explicit Model of Deforestation in Bolivia
Jerome Vanclay1, David Kaimowitz1, Atie Puntodewo1, Patricia Mendez2
1. Center for International Forestry Research, Bogor, Indonesia, Fax +62 251 622100,
cifor@cgiar.org, http://www.cgiar.org/cifor
2. Land Use Planning Technical Office (OT-PLUS), Departmental Government (Prefectura),
Santa Cruz, Bolivia
Abstract
A GIS compiled by the Departmental Government of Santa Cruz, Bolivia offers data that may
help to resolve some competing theories of tropical deforestation. The GIS contains many
attributes relating to land use at two points in time, 1989 and 1994, and allow us to address
questions like:
• What has been the impact of past road construction on deforestation and land use?
• What impacts might be expected from future road construction?
• What impact do zoning policies such as forest concessions and protected areas have?
• What influence do cultural factors have on forest clearing and fragmentation?
We discuss our methodology and report interim results. We seek to provoke discussion on
appropriate statistical procedures for such analyses.

Introduction
Tropical deforestation is topical and controversial, and many researchers and agencies would
like to better understand when, where, and more importantly, why it occurs. Kaimowitz and
Angelsen (1998) reported the existence of 150 deforestation models, most of which were
developed since 1990. They found that in all, more than115 different variables had been used in
these attempts to explain deforestation and that major uncertainties continue to exist about how
most of these influence deforestation. We take this to be an indication of the inherent
complexity of the task, the scarcity of decisive indicators, and the limitations of proxy variables
used in these studies.

Among different possible modeling approaches, Kaimowitz and Angelsen (1998) concluded
that household and regional-level studies are likely to be more productive than national and
global studies. They expressed particular enthusiasm for the potential of the growing availability
of spatial data bases providing insights into the role in deforestation processes of such spatial
variables as access to markets, land use zoning policies, and ecological conditions. They note
that such models use relatively reliable data, involve large sample sizes that give model makers
more degrees of freedom to work with, and are particularly suited for predicting where
deforestation is likely to occur. In addition, the model’s robustness can often be tested by
measuring what percentage of the time they accurately predict where deforestation will occur.

This paper presents a spatial econometric model of one particular Latin American region, the
department of Santa Cruz in eastern Bolivia. Several reasons inspired us to study deforestation
patterns in that region:

• Deforestation in the Bolivian tropics has historically been limited but has increased rapidly
in recent years, and it is important to understand why;

• Deforestation patterns in Bolivia differ significantly from other areas in Latin America in
that the expansion of large-scale mechanized agriculture has been more important in the
former;
• CIFOR has an on-going international project comparing the effects of different policies and
social trends on tropical forests in Bolivia, Cameroon, and Indonesia; and
• The Santa Cruz government (prefectura) had compiled a GIS with much of the data needed
to examine the influence of different geographic variables on deforestation trends.

Our objectives in pursuing the present study were three-fold. We wanted to
1) Test some established theories of deforestation,
2) Improve the capacity to formulate land use policies within the Department of Santa Cruz,
and
3) Contribute to a better understanding of factors that determine land use in locations similar to
those of study area.
Previous Spatial Econometric Deforestation Models
Spatial regression models measure the correlation between land use and other geo-referenced
variables such as:
• Transportion costs (distance from markets and road, railways, and rivers),
• Zoning categories (national parks, forest concessions, colonization areas, indigenous
territories), and
• Ecological conditions (topography, soil quality, precipitation, and forest fragmentation).

The models focus on land use in a single time period or the change in land use over two or more
periods. The majority relate the state of the independent variables in the first period to the
probability that the forest in that location is removed between the first and second periods.

Unlike the Santa Cruz model presented below, most previous models have drawn their data
from a random sample of locations (points) within a selected region or country. Sample sizes are
typically several thousand points or more. Chomitz and Gray (1995) used a multinomial
maximum likelihood model with a random sample of 10,000 data points. Tom Tomich (pers
comm) examined deforestation rates within a study area of 4.9 million hectares in Jambi,
Sumatra, by sub-sampling with a 1 km square grid and using a binomial-probit transformation.
Gerald Nelson (pers comm) made a similar study with raster data by taking a systematic 1%
sample that yielded 25,000 sample points. He claimed results were "meaningful" because some
parameters were found to be statistically significant while others were not.

Some models include all types of locations, others just locations covered with forest during the
first time period. Typically, the land use information comes from national forest inventories,
remote sensing and aerial photographs.

The models show land holders are more likely to convert forest to agricultural use where good
access to markets and favorable conditions for farming make agriculture more profitable and the
government has not restricted forest conversion (Table 1). Forests close to roads in physical
distance and traveling time are more likely to be cleared (Chomitz and Gray, 1995; Liu et al.,
1993; Ludeke et al., 1990; Mertens and Lambin, 1997; Nelson and Hellerstein, 1995; Sader and
Joyce, 1988; Rosero-Bixby and Palloni, 1996). Most studies show that forest clearing declines
rapidly beyond distances of two or three kilometers from a road, although Liu et al. (1993)
report significant forest clearing up to around 15 km from the nearest road for the Philippines.
Similarly, Chomitz and Gray (1995) found that locations closer to urban markets have less
remaining forest in Belize and Mertens and Lambin (1997) reported that deforestation drops off
dramatically beyond 10 kilometers from the nearest town in Eastern Cameroon.

Forest fragments have a higher risk of being lost than forests in large compact areas, with
forests close to the forest edge especially likely to be cleared (Brown et al., 1993; Liu et al.,
1993; Ludeke et al., 1990; Mertens and Lambin, 1997; Rosero-Bixby and Palloni, 1996). In
addition, areas with higher quality soils (flat, adequate drainage, and high soil fertility) and drier
climates are also more likely to be cleared (Chomitz and Gray, 1995; Gastellu-Etchegorry and
Sinulingga, 1988; Sader and Joyce, 1988; Rosero-Bixby and Palloni, 1996).

The effect of roads and environmental conditions may interact. Thus roads may induce greater
deforestation in areas with good soils and favorable climatic conditions. In Belize, Chomitz and
Gray (1995) showed that the probability of an area being used for agriculture (rather than being
retained as natural vegetation) on high quality land next to a road was 50%, whereas lands next
to roads with marginal soils had only a 15% probability of being deforested.

Mertens and Lambin (1997) noted that variables affect forest clearing differently depending on
the type of deforestation process. In peri-urban deforestation, forest clearing exhibits a circular
pattern around the towns, and distance to towns and roads strongly affects forest clearing but
proximity to forest edge does not. Roads may exhibit a “corridor pattern” of deforestation where
proximity to roads and forest edges are significant determinants of forest clearing, but distance
to towns is not. Finally, in areas where diffuse shifting cultivation dominates, proximity to
forest edge increases the probability of forest clearing, whereas distance to roads and towns is
less important.
Deforestation in Santa Cruz, Bolivia
Department of Santa Cruz extends some 900 by 800 kilometers, and occupies some 35 million
hectares (Figure 1). Forest cover estimates based on Landsat data are available for 1989 and
1994 (Morales 1993 and 1996). The total accumulated area of forest cleared by humans prior to
1994 was about 2.1 million hectares or 6% of the land area, most of it is concentrated within
about 200 km of the capital city, Santa Cruz. In 1994, some 15 million hectares of forest
remained, along with some 1.9 million hectares of agriculture and 3.2 million hectares of
pasture or savanna (an increase of 281 thousand hectares since 1989). Much of the savanna and
pasture is natural, especially in areas of the Chiquitano Shield, the Pantanal, the Quimome area,
and in the sub-Andean zone.

Annual deforestation rates have been increasing rapidly since the mid-1980s. Between 1986 and
1990, CUMAT (1992) found that 38,000 hectares of forest were cleared annually in the
Amazonian portion of Santa Cruz. That region covers only 61% of Santa Cruz, but accounts for
a much higher percentage of forest clearing. Approximately 78,000 hectares were cleared
annually in all of Santa Cruz between 1989 and 1992, rising to 117,000 hectares annually
between 1992 and 1994 (Morales 1993 and 1996).

Most deforestation in Santa Cruz is by large mechanized soybean and wheat farmers, small
agricultural colonists who practice mainly slash and burn rice and maize cultivation, and large
cattle ranchers (Pacheco 1998). The mechanized farm sector has grown rapidly over the last
fifteen years, and now accounts for a majority of forest clearing. Most of this growth has been in
the area east of the Rio Grande River, known as the “expansion zone”. Small agricultural colonists
have expanded into moister forest areas suitable for rice growing to the north and west of the city
of Santa Cruz. Forest clearing for pastures is concentrated in northeastern Santa Cruz.
Data
Our data were drawn from a GIS produced by the ‘Santa Cruz National Resource Protection
Project’ implemented by the Government of Santa Cruz with funding from KFW and technical
assistance from a consortium composed of the IP, SCG, and KWC consulting companies. The
initial objective of that GIS was to develop a land use plan (PLUS) for the entire department of
Santa Cruz. Hence forth, we will refer to it as the PLUS GIS.

The PLUS GIS was compiled from several sources. Most data were digitized from 1:250,000
maps, but some layers were captured at other scales and obtained from other sources. Many
layers obtained were based on the UTM elipsoid IU661967.

GIS layers of particular significance for our study included:
• Land use in 1989, 1992, and 1994 (i.e., several classes of urban, agriculture, forest, etc)
• Vegetation, soil types and rainfall data using a standard classification,
• Details of the road and rail network (including logging/mining roads),
• Administrative data including urban areas, forest concessions, colonization areas,
indigenous territories, protected areas, etc.

The land use, vegetation type and soils data were provided in raster form, and were converted to
vector format. The forest concession boundaries were obtained from the Sustainable Forestry
Management (BOLFOR) Project. The 1989 land use data were compiled from Earthsat Data
analyzed by the CUMAT consulting company, and were considering "quite reliable" by Ivan
Morales (pers comm), the expert who analyzed the 1992 and 1994 Landsat data.

The 1989 land use data delineate forests, deforested areas, savanna and pastures, areas with little
or no vegetation, water, and urban areas. The 1994 land use data further sub-divides the
deforested areas into traditional agriculture, commercial agriculture, mixed agriculture,
agriculture with forests and forests with agriculture. The 1989 data had a resolution of 1 x 1 km
(100 ha), whereas the resolution in 1994 was 250 x 250 m (about 6 ha). Some areas were
omitted from the north-west in the 1989 data and from the east in the 1994 data and these areas
have been excluded from our study. The cloud cover was minimal in the 1994 images used to
assess land use, so this assessment is considered more comprehensive than the previous
assessments.

Despite considerable care and attention to detail in compiling the GIS, there were some
anomalies that we could not reconcile. Deforestation estimates obtained by calculating the area
in agricultural land in 1994 that had been forest in 1989 inexplicably provided different
estimates than when we combined all agricultural lands in 1989, 1992, 1993 and 1994, and then
subtracted the land already agriculture in 1989. Although the latter approach provided estimates
consistent with independent estimates by BOLFOR (namely 552,985 ha), the discrepancy is
unsettling.
Methods
An analysis of deforestation of this kind poses many interrelated questions:
• What should we try to predict: deforestation rate 1989-94 or total deforestation to 1994?
• Should we use a binomial (forest, non-forest) or a multinomial model that considers the
various end-uses of former forest land?
• How should we transform the dependent variable to make analyses tractable and results
meaningful: is it better to use a logarithm, logistic or probit transformation?
• What explanatory variables should we consider in our analysis, and how can we minimize
the correlation between these variables?
• How can we efficiently transfer the data between the GIS and the statistics package, while
minimizing spatial autocorrelation?1

1 Spatial autocorrelation is a common problem with geographic data, since nearby locations are
more likely to be similar than distant ones. This can lead to inaccurate measures of statistical
significance. Several methods exist for partially correcting for spatial autocorrelation, although
none is fully satisfactory (Rosero-Bixby and Palloni, 1996; Chomitz and Gray, 1995).

• How can we tell if we have a problem with multiple or spatial autocorrelation?
• How can we discriminate endogeneous and exogeneous variables?

Somewhat surprisingly, prior studies offer little guidance on these issues.
Theory, initial hypotheses and response variable
Based on economic theory and previous modeling exercises, we hypothesized that the more
productive land (i.e., Soil type I, with rainfall exceeding 1000 mm) and land with better access
to markets (lower transportation costs) would be cleared first. We anticipated that zoning an
area as a forest concession or protected area would impede deforestation, while zoning it as a
colonization areas would encourage deforestation. In addition, we hypothesized that indigenous
people have cultural attributes that lead to less deforestation.

Although both the total deforestation to date and the recent deforestation rate (1989-94) are of
interest, it is the latter that is of most interest, as it is the best indicator of current trends and
responses to existing policies. Similarly, although the end-use of deforested land is of interest,
the binomial model is more tractable and simplifies analyses. We examined a simple binomial
model that considered only land forested in 1989: if deforested during 1989-94 the response
variable was coded 0, otherwise it was coded 1. This provides a model that could be used to
predict deforestation during 1994-99, and could be checked by making empirical tests of
predictions for 1999. To ensure unambiguous regarding deforestation during 1989-94, we
deleted from our data set all areas that were not forest in 1989, including areas that were
ambiguously defined in the 1989 classification (e.g., cloud, “no data”, etc.).

Economists often favour the use of logarithmic transformations, as parameter estimates can then
be interpreted directly as elasticities (i.e., predictor variables are multiplicative, so that a unit
change in a predictor variable always causes the same percentage change in the response
variable). This may be helpful when all predictor variables are expressed in the same units, but
becomes less relevant when the nature of the predictors varies greatly. Statisticians tend to
prefer logistic and probit transformations for binomial data because standard assumptions are
better satisfied, and predictions are constrained correctly. The probit and logistic
transformations are similar in many respects, but my previous experience (Vanclay 1994)
inclines me to favour the logistic transformation (weighted for polygon area). Fortunately for
economists, the logistic is very similar to the logarithmic transformation if rates of change do
not exceed 0.25, so provided deforestation rates remain modest, parameter estimates may still be
interpreted as elasticities.
Sampling
Although some statistics packages claim to be able to interface directly with GIS, it is
convenient to extract data from the GIS as a simple text file, so that it can be used with any
statistics package. However, this raises the question of how best to extract the data: should a
sample of selected points be taken, should polygons form the basis for analysis, or should some
other alternative be adopted (Figure 2)?

Systematic selection of sample points ensures a compact data set and simplifies analyses, but
fails to make full use of the available information. A small sample may be statistically
inefficient, but computationally convenient. A larger sample size makes more efficient use of
information, but also increases the potential for spatial autocorrelation. Some researchers prefer
to sample tiles rather than points, arguing that these are more representative in fragmented
landscapes where correct alignment of the various GIS layers may be problematic. If these tiles
tessellate the study area (e.g., square tiles rather than circular plots), a complete census may be
analysed, but this again introduces the possibility of spatial autocorrelation. One way to make
better use of information while minimizing autocorrelation is to stratify (e.g., forested versus
deforested, and close to versus distant from town/road), sample strata with different intensities