Assessing the extent of agriculture/pasture and secondary succession forest in the Brazilian Legal Amazon using SPOT VEGETATION data

1

Assessing the extent of agriculture/pasture and secondary succession forest in the

Brazilian Legal Amazon using SPOT VEGETATION data

João M. B. Carreiras a,*, José M. C. Pereira a, Manuel L. Campagnolo b, and Yosio E. Shimabukuro c

a Department of Forestry, Instituto Superior de Agronomia, Tapada da Ajuda, 1349-017 Lisboa,

Portugal.

b Department of Mathematics, Instituto Superior de Agronomia, Tapada da Ajuda, 1349-017 Lisboa,

Portugal, and Center for Logic and Computation, Instituto Superior Técnico, Lisboa, Portugal.

c Remote Sensing Department, Instituto Nacional de Pesquisas Espaciais, Av. dos Astronautas, 1758,

CEP 12227-010, São José dos Campos, SP, Brazil.

*Corresponding author. Tel.: +351-213653387.

E-mail address: [email protected] (J.M.B. Carreiras).

Manuscript

2

Abstract

There has been growing concern about land use/land cover change in tropical regions, as there is

evidence of its influence on the observed increase in atmospheric carbon dioxide concentration and

consequent climatic changes. Mapping of deforestation by the Brazilian Institute for Space Research

(INPE) in areas of primary tropical forest using satellite data indicates a value of 587,727 km2 up to the

year 2000. Although most of the efforts have been concentrated in mapping primary tropical forest

deforestation, there is also evidence of large-scale deforestation in the cerrado savanna, the second

most important biome in the region.

The main purpose of this work was to assess the extent of agriculture/pasture and secondary

succession forest in the Brazilian Legal Amazon (BLA) in 2000, using a set of multitemporal images

from the 1-km SPOT-4 VEGETATION (VGT) sensor. Additionally, we discriminated primary tropical

forest, cerrado savanna, and natural/artificial waterbodies. Four classification algorithms were tested:

quadratic discriminant analysis (QDA), simple classification trees (SCT), probability-bagging

classification trees (PBCT), and k-nearest neighbors (K-NN). The agriculture/pasture class is a

surrogate for those areas cleared of its original vegetation cover in the past, acting as a source of

carbon. On the contrary, the secondary succession forest class behaves as a sink of carbon.

We used a time series of 12 monthly composite images of the year 2000, derived from the

SPOT-4 VGT sensor. A set of 19 Landsat scenes was used to select training and testing data. A 10-

fold cross validation procedure rated PBCT as the best classification algorithm, with an overall sample

accuracy of 0.92. High omission and commission errors occurred in the secondary succession forest

class, due to confusion with agriculture/pasture and primary tropical forest classes. However, the

PBCT algorithm generated the lower misclassification error in this class. Besides, this algorithm yields

information about class membership probability, with ~ 80% of the pixels with class membership

probability greater or equal than 0.8. The estimated total area of agriculture/pasture and secondary

succession forest in 2000 in the BLA was 966 x 103 km2 and 140 x 103 km2, respectively. Comparison

with an existing land cover map indicates that agriculture/pasture occurred primarily in areas

previously occupied by primary tropical forest (46%) and cerrado savanna (33%), and also in transition

3

forest (19%), and other vegetation types (2%). This further confirms the existing evidence of extensive

cerrado savanna conversion.

This study also concludes that SPOT-4 VGT data are adequate for discriminating several

major land cover types in tropical regions. Agriculture/pasture was mapped with errors of about 5%.

Very high classification errors were associated with secondary succession forest, suggesting that a

different methodology/sensor has to be used to address this difficult land cover class (namely with the

inclusion of ancillary data). For the other classes, we consider that accurate maps can be derived from

SPOT-4 VGT data with errors lower than 20% for the cerrado savanna, and errors lower than 10% for

the other land cover classes. These estimates may be useful to evaluate impacts of land use/land

cover change on the carbon and water cycles, biotic diversity, and soil degradation.

Keywords: Brazilian Legal Amazon (BLA); agriculture/pasture; secondary succession forest; SPOT-4

VEGETATION (VGT); supervised classification

1. Introduction

There is clear evidence of increased carbon dioxide (CO2) concentration in the atmosphere, as careful

and systematic measurements since 1950s at the reference station in Hawaii (Mauna Loa) indicated a

steadily increase from ~315 ppm to ~380 ppm of CO2 (Keeling & Whorf, 2005), an increase of

approximately 20% in less than 50 years. Other studies, namely from air bubble analysis of ice caps

(e.g. Petit et al., 1999), showed that the CO2 concentration has been more or less stable around 280

ppm for thousands of years until 1800. Fossil fuel burning and forest depletion are quoted as the main

activities responsible for the observed increase in atmospheric CO2 concentration (Korner, 2000;

Vitousek et al., 1997). This increase has lead to an enhanced concern with land use/ land cover

change in tropical regions (Cochrane et al., 1999; Gedney & Valdes, 2000; Houghton et al., 2000;

Lambin et al., 2003). There has been growing concern with deforestation and its influence on the

carbon cycle, in particular in the Brazilian Legal Amazon (BLA), home of the major tropical forest

4

ecosystem in the world, (Achard et al., 2002; Cardille & Foley, 2003; Houghton et al., 2000, 2001;

Skole & Tucker, 1993; Souza Jr. et al., 2003). According to the Brazilian Institute of Geography and

Statistic (IBGE), the major biomes in the BLA are primary tropical forest (ombrophyllous and seasonal

forest) (~ 65%) and cerrado savanna (~ 15%), the remainder being transition forest and other

vegetation types (IBGE, 1988). The potential distribution area of primary tropical forest is in central,

northern, and western BLA, whilst cerrado savanna is mostly concentrated in the states of Mato

Grosso and Tocantins, the southern and eastern rims of the BLA (IBGE, 1988). Therefore, it is

important to know the main land use land cover changes in these two main biomes, particularly in the

less studied cerrado savanna. Myers et al. (2000) have identified the South America cerrado as one of

the 25 most important terrestrial biodiversity hotspots in the world, with thousands of plant species and

hundreds of mammal, bird, reptile, and amphibian species. Identification of areas undergoing forest

regeneration is also important, as they play an important role as carbon sink (Brown & Lugo, 1990;

Curran & Foody, 1994). Recent international and Brazilian led studies focused on the Amazon Basin:

the Tropical Ecosystem Environment observation by Satellite (TREES) project of the Joint Research

Centre (JRC) aimed at mapping and monitoring the pan-tropical humid forest belt using National

Oceanic and Atmospheric Administration (NOAA) Advanced Very High Resolution Radiometer

(AVHRR) data (e.g. Achard et al., 2001); the JRC’s Global Land Cover 2000 (GLC2000) project is

providing a harmonized land cover database over the whole globe for the year 2000, with some

emphasis in areas considered as deforestation hotspots (e.g. Bartholomé & Belward, 2005); the Large

Scale Biosphere Atmosphere Experiment in Amazonia (LBA), an international research program led

by Brazil, aims at studying the role of the Amazon region as part of the Earth system, and at

evaluating the influence of land-use and climate change in biological, chemical, and physical

processes (e.g. Roberts et al., 2003). Due to the extent of the BLA, approximately 5 x 106 km2, as well

as restricted accessibility throughout most of the area, remote sensing data have an important role in

characterizing land use/land cover changes. Several studies have addressed the problem of land

use/land cover change, deforestation, and forest regeneration mapping in the BLA using high

resolution (Achard et al., 2002; Adams et al., 1995; Foody et al., 1996; Kimes et al., 1999; Lucas et al.,

2002; Skole & Tucker, 1993) and coarse resolution (Achard et al., 2001; Carreiras et al., 2002; Lucas

et al., 2000; Rodriguez-Yi et al., 2000) optical remote sensing data.

5

The spatial and temporal patterns of deforestation in the BLA closely follow the construction of

highways (Fearnside, 1993; Lambin et al., 2003; Moran et al., 1994; Nepstad et al., 1997). Large-scale

deforestation in the BLA began with the construction of the Belém-Brasilia Highway (BR-010) in 1958

(Moran et al., 1994). Subsequently, the BR-364 in Mato Grosso, Rondônia and Acre, the

Transamazon highway in central Pará and Amazonas, and the PA-150 in east-central Pará, promoted

human settlement and forest conversion (Nepstad et al., 1997). Most of the deforested areas were

occupied by agricultural crops and cattle pasture, but also by mining and logging activities (Asner et

al., 2005; Moran et al., 1994; Nepstad et al., 1997). Although most of the effort in mapping

deforestation has been concentrated in areas of primary tropical forest (e.g. INPE, 2002; Skole &

Tucker, 1993), large-scale deforestation in the cerrado savanna biome has received some attention

(Fearnside, 1993; Cardille & Foley, 2003; Ferreira et al., 2003; Kaimowitz & Smith, 2001; Nepstad et

al., 1997; Sano et al., 2001). The Brazilian Institute for Space Research (INPE) performs nearly

annual mapping of BLA deforestation since the end of the 1980’s using data from the Landsat

Thematic Mapper (TM) sensor. INPE (2002) has reported that BLA deforestation in areas of primary

tropical forest reached a value of 587,727 km2 by the end of 2000, including 97,000 km2 of old

deforestation (prior to 1960) in Pará and Maranhão. Native vegetation clearing in regions dominated

by cerrado savanna is not well documented, mainly because spectral differences between natural

vegetation and agricultural crops are more subtle than those observed when primary tropical forest is

converted to agriculture (Nepstad et al., 1997). Nonetheless, there is evidence from integrated satellite

and census data that extensive conversion has occurred in regions of cerrado savanna between the

early 1980s and the mid 1990s (Cardille & Foley, 2003).

Land abandonment is common in the BLA, resulting in a mosaic of regenerating vegetation of

different ages (Brown & Lugo, 1990; Perz & Skole, 2003). Several studies have shown that secondary

succession forest establishes rapidly in abandoned areas (Buschbacher et al., 1988; Mesquita et al.,

2001; Uhl, 1987; Uhl et al., 1988), mainly as the result of the high cost of re-clearance for most

farmers (Moran et al., 1994). Mapping and quantification of the area under forest regeneration in the

BLA, using remote sensing data, has been carried out by several authors (Foody et al., 1996; Kimes

et al., 1999; Lu et al., 2003; Lucas et al., 2000, 2002; Mausel et al., 1993; Skole et al., 1994;

Steininger, 2000; Vieira et al., 2003). However, most of the forest regeneration mapping in the entire

BLA was carried out in the late 1980s-early 1990s. Lucas et al. (2000), used data from the NOAA

6

AVHRR sensor, and estimated that 157,973 km2 supported forest regeneration during 1991-1994, with

approximately 50% being less than five years old. Therefore, a large amount of the mapped

secondary forest could possibly be temporary fallows, and not regenerating forest. Schroeder &

Winjum (1995), used a land cover map of South America, obtained from 1-km NOAA AVHRR data for

the period 1988-1991 (Stone et al., 1994), and estimated the extent of forest regeneration in the BLA

at 151,000 km2. Fearnside (1996), using a Markov matrix of annual transition probabilities, estimated

that approximately 48% (195,170 km2) of the deforested landscape in 1990 supported forest

regeneration. Recently, Cardille & Foley (2003), also using a Markov matrix of annual transition

probabilities, estimated that 36% (90,600 km2) of the deforested area between 1980 and 1995 in the

entire Amazon river drainage basin were in some stage of secondary succession forest. The work of

Lucas et al. (2000) was the only one to map several stages of secondary succession forest directly

from remote sensing data in the entire BLA (adding up to 157,973 km2), using an unsupervised

classification algorithm. Other studies (Cardille & Foley, 2003; Fearnside, 1996) estimated the total

area of forest regeneration through Markov modeling of existing land cover maps.

The main objective of this study was to map agriculture/pasture and secondary succession

forest in the BLA, using a time series of monthly composite images of the 1-km SPOT-4

VEGETATION (VGT) sensor, for the year 2000. A special emphasis was placed on agriculture/pasture

and secondary succession forest classes due to their influence on the carbon cycle. The

agriculture/pasture class included all the areas that were cleared of its original vegetation cover (i.e.

primary tropical forest and cerrado savanna) in the past, and that in the year 2000 still supported some

kind of agricultural or pasture use. Secondary succession forest refers to areas that were deforested in

the past, and that in 2000 exhibited some kind of regenerated vegetation. Additionally, we were

interested in discriminating primary tropical forest, cerrado savanna, and natural/artificial waterbodies.

We tested four supervised classification algorithms for land cover mapping: (i) quadratic discriminant

analysis, (ii) simple classification trees, (iii) probability-bagging classification trees, and (iv) k-nearest

neighbors. The dataset used in this study, the first daily 1-km annual coverage of the BLA spanning a

full year, has the potential for detecting cerrado savanna conversion, as phenological differences may

be more evident on a seasonal basis. Accuracy assessment of the resulting land cover map was

performed both quantitatively (i.e. with an error matrix and information about class membership

probability) and qualitatively (i.e. via comparison of class extent with land cover maps available from

7

the literature). Estimates of total areas were obtained through the application of the inverse method

calibration technique (e.g. Gallego, 2004; Tenenbein, 1972; Walsh & Burk, 1993).

2. Study area

The BLA is a politically defined region of Brazil and encompasses the states of Acre, Amapá,

Amazonas, Mato Grosso, Pará, Rondônia, Roraima, Tocantins, and a part of Maranhão (west of 44º

W) (Fig. 1). Most of the BLA is included in the Amazon River basin, with the exception of southern

Mato Grosso and western Maranhão, included in the Paraguai and Parnaíba river basins, respectively.

This region covers an area of approximately 5 x 106 km2, consisting primarily of closed primary tropical

forest (ombrophyllous and seasonal forest), but also including large areas of flooded forest and

cerrado savanna (Alves, 2002). Mean temperature within the region varies from 25.8 ºC during the wet

season (October to April) to 27.9 ºC during the dry season (May to September) (Junk & Firch, 1985).

Mean annual rainfall is approximately 2250 mm, but varies considerably, ranging from 1500 mm in the

north and south to over 3000 mm in the northwest. (Goulding et al., 2003).

[Insert Fig. 1 about here]

3. Data

The original dataset used in this study was a set of daily 1-km SPOT-4 VGT images spanning the

entire year 2000, and covering the BLA (3360 x 2800 pixels) (images from Julian days 39, 45, 58, 77,

134, 178, 213, 220, 230, 252, 280, 292, and 294 were not available or had unrecoverable errors). The

dataset over the BLA ranged from 45º W to 75º W and from 5º N to 20º S, thus missing the portion of

the state of Maranhão between 44º and 45º W (~ 64 x 103 km2), and a small area of Roraima above 5º

N (~ 2 x 103 km2). We used the S1 product, consisting of 1-km georeferenced, calibrated,

8

atmospherically corrected surface reflectance data (Passot, 2000). Among other characteristics, the

VGT sensor onboard the SPOT 4 and 5 satellites provides a better imagery acquisition geometry and

additional reflective bands [blue 0.43 – 0.47 µm, red 0.61 – 0.68 µm, near infrared (NIR) 0.78 – 0.89

µm, and short wave infrared (SWIR) 1.58 – 1.75 µm], when compared with the commonly used NOAA

AVHRR, which only has the red band (0.58 – 0.68 µm) and the NIR band (0.72 – 1.10 µm) in the

reflective part of the electromagnetic spectrum.

We have further combined SPOT-4 VGT S1 daily images into 12-monthly composite images,

from January to December 2000. The compositing algorithm is described in Carreiras & Pereira

(2005). Basically, each monthly composite image was produced by combining a monthly compositing

criterion and a transformation of the principal components analysis, with the objective of minimizing

cloud cover, while preserving as much as possible the original spatial structure of the data. The

monthly compositing criterion uses different conditions according to the nature of the land cover,

defined by a given vegetation index (VI), in this case the Soil Adjusted Vegetation Index (Huete,

1988). If a pixel is considered to represent a vegetated surface (VI greater than a given threshold)

then the date of the third lowest value of the SWIR band is selected. Otherwise, (VI lower than a given

threshold), it indicates a bare ground/sparsely vegetated pixel, and the third lowest value of the red

band is selected. This algorithm assumes that a pixel is unlikely to be covered by cloud shadows more

than two times per month. Furthermore, the Maximum Noise Fraction (MNF) (Green et al., 1988), a

principal components transformation that rescales the noise in the image, was applied to the

multitemporal dataset of monthly composite images and tested as a method of additional signal-to-

noise ratio improvement. The back-transformed dataset using the first few MNF eigenimages yielded

an accurate reconstruction of monthly composite images from the dry season (May-September) and

enhanced spatial coherence from wet season images (October-April).

Our main focus was on the correct evaluation of the spatial distribution of agriculture/pasture

and secondary succession forest. Wall-to-wall mapping was obtained by considering areas of primary

tropical forest, cerrado savanna, and natural/artificial waterbodies. Landsat TM and Enhanced TM+

(ETM+) data from the years 2000 and 2001 were used to collect training data for the SPOT-4 VGT

imagery (Fig. 1 and Table 2). We first defined a set of scenes through a stratified sampling procedure.

Then, in each scene, training polygons were chosen through a quasi-systematic sampling design.

Each polygon corresponds to two to four pixels of the VGT imagery. The number of polygons per class

9

exceeds the minimum recommended by Congalton & Green (1999), for a significance level of 95%

and an error of 5%.

More specifically, a total of 19 scenes were selected to represent the various

agriculture/pasture hotspots, existing secondary succession forest areas, and remaining land cover

types (Table 2). The selection of scenes representing agriculture/pasture hotspots was made through

the inspection of deforestation maps produced by INPE in the scope of the Prodes Digital Project (e.g.

INPE, 2002). Since INPE does not account for cerrado savanna conversion, we selected scenes in

areas where there is evidence of that process (e.g. Mato Grosso, Tocantins). Areas that were

deforested and currently support agriculture or pasture were easily identified in Landsat data, due to

their spectral characteristics and regular field geometry. The choice of scenes with secondary

succession forest was accomplished essentially by looking at the aforementioned literature, which list

several areas of the states of Pará, Maranhão, Amazonas, and Mato Grosso as having progressed to

forest regeneration. Areas of secondary succession forest were selected using two dates per scene;

an area was considered to support secondary succession forest if in the first date it had been

subjected to deforestation, and in the second date (2000/2001) it no longer supported agriculture or

pasture. Stratification of the remaining land cover classes was done according to an original 59-class

vegetation map of the BLA, which was further generalized into a 13-class map (IBGE, 1988). Landsat

data were obtained from INPE, in the scope of the Prodes Digital Project

(http://www.obt.inpe.br/prodes), and from the Global Land Cover Facility (GLCF), University of

Maryland (USA) (http://glcf.umiacs.umd.edu).

[Insert Table 1 and Table 2 about here]

In the late 1980s IBGE produced a map of potential vegetation types of Brazil, with 59 classes

(Mapa de Vegetação do Brasil (IBGE, 1988)). Later, it was generalized into a 13-class map. We

further aggregated it into four major classes: primary tropical forest, cerrado savanna, transition forest,

and other vegetation types. The latest is mostly composed of alluvial riparian vegetation (IBGE, 1988).

We used this last map to assess the extent of agriculture/pasture and secondary succession forest.

10

4. Methods

4.1. Classification algorithms

Several classification techniques and algorithms have been used to characterize land use/land cover

change in the BLA, using optical remote sensing data. Skole & Tucker (1993) used visual

interpretation of Landsat TM band 5 black and white photographic images of the entire BLA, to map

deforestation over closed tropical forests in 1988. Lucas et al. (2000) used NOAA AVHRR data for the

period 1991-1994, to map the extent of secondary succession forest stages in the entire BLA, relying

on an unsupervised classification algorithm (isodata). Other studies focused in particular regions of the

BLA. Adams et al. (1995) mapped land-cover change near Manaus (Amazonas), between 1988 and

1991, based on fractions of spectral endmembers over Landsat TM data. Foody et al. (1996) mapped

different forest regeneration stages, using a maximum likelihood classifier with Landsat TM data from

1991 in an area corresponding approximately to that of Adams et al. (1995). Kimes et al. (1999)

mapped primary forest, forest regeneration age classes, and deforested areas, using SPOT High

Resolution Visible (HRV) 1994 data for Rondônia, with neural networks and linear discriminant

functions. Rodríguez-Yi et al. (2000) used supervised classification by regions (Bhattacharya distance)

followed by image segmentation of 1993 NOAA AVHRR data, to map eight vegetation classes in the

state of Mato Grosso. Achard et al. (2001), in the scope of the TREES project, mapped several land

cover classes in the pan-tropical humid forest belt, using unsupervised classification algorithms and

NOAA AVHRR data from the early 1990s. Lu et al. (2003) used Landsat TM data of 1998, over

Rondônia, to classify secondary succession forest stages and primary forest, with a maximum

likelihood classifier and threshold in image fractions derived from a linear mixture model. Vieira et al.

(2003) relied on Landsat ETM+ data of 1999 from northeast Pará, to classify forest regeneration, with

a supervised classification algorithm. Souza Jr. et al. (2003) using a decision tree classifier, mapped

intact forest, logged forest, degraded forest, and secondary succession forest, with 20-m SPOT 4

multispectral data of 1999 in northeast Pará. Carreiras et al. (2004) mapped nine land cover classes in

the state of Mato Grosso, using a probability bagging classification tree and SPOT-4 VGT data from

2000.

11

Next, we describe briefly the four supervised classification algorithms used in this study.

Quadratic discriminant analysis (QDA), usually known as the maximum likelihood classifier in remote

sensing literature, is a parametric supervised classification method. Assuming that spectral signatures

in each class are Gaussian, and that the (maximum likelihood) estimates of the mean vectors (µi) and

the variance-covariance matrices (Σi) in each class are accurate, the quadratic discriminant rule is an

optimum classification rule. The training sample has to be large enough to yield good estimates of µi

and Σi for each class (Richards, 1986).

Classification trees are a non-parametric method. A classification tree partitions the space of

all possible spectral signatures x, starting with the whole spectral space (at the root of the tree) and

successively splitting that space in subsets, such that each subset is more likely to be assigned to one

of the land cover classes than the subset from which it is split (Breiman et al., 1984; Safavian &

Landgrebe, 1991). All subsets of the spectral space are represented by nodes in a tree, and each split

corresponding to the descendents of a node. Since each node of the tree represents an element Pj of

a partition of the space of all possible x, one can estimate P(classi /x∈ Pj) for all terminal nodes and all

classes, and assign the node to the class with the highest probability. The estimate of P(classi /x∈ Pj)

is simply the proportion of pixels that belong to classi among all the training sample pixels that are in

Pj. The aim of the procedure is to find a tree that better approximates the actual conditional

probabilities P(classi /x). In the Classification and Regression Trees (CART) algorithm of Breiman et al.

(1984) heuristic techniques are used to find a tree structure that discriminates the classes (i.e. which

terminal nodes have a high proportion of sample individuals of some class) but is not overfitted to the

training sample (i.e. the tree should not be “too large”). Hereafter, this type of classification algorithm

will be referred as simple classification tree (SCT). However, classification trees are sensitive to small

perturbations in the training set, which may originate large changes in the resulting classifiers

(Breiman, 1996). Therefore, these unstable methods can have their accuracy improved with a

perturbing and combining technique, that is, by generating multiple perturbed versions of the classifier

(a.k.a. ensemble, or committee) and combining those into a single predictor (Breiman, 1998). Methods

that use ensemble of classifiers have demonstrated to be very successful at improving the accuracy of

classification trees (Bauer & Kohavi, 1998; Maclin & Opitz, 1997). These methods can be divided in

two types: those that adaptively change the distribution of the training set based on the performance of

previous classifiers (e.g. boosting) and those that do not (e.g. bagging) (Bauer & Kohavi, 1998). In this

12

study we will only focus on the bagging algorithm applied to classification trees. In bagging, each sub-

classifier ci (i=1…n)(in this case a SCT) is run on n different bi bootstrap samples of the original m

training set observations. Each bi is generated by uniformly sampling m observations from the training

set with replacement. The final classifier C is built from ci, whose output is the class most frequently

predicted by its sub-classifiers, with ties broken arbitrarily (Bauer & Kohavi, 1998; Breiman, 1996).

Although the main purpose of bagging was to build a strong classifier by means of variance reduction

(Breiman, 1996), some variants of bagging have also proven adequate for the estimation of class

membership probability (Bauer & Kohavi, 1998; Breiman, 1996; Perlich et al., 2003; Provost &

Domingos, 2003). Probability-bagging classification trees (PBCT) is one of such variants, so that

instead of returning a classification, each sub-classifier returns a probability distribution for the classes

in each terminal node (Bauer & Kohavi, 1998; Provost & Domingos, 2003). Subsequently, the PBCT

algorithm averages the probability for each class over all sub-classifiers, and predicts the class with

the highest probability. However, Provost & Domingos (2003) note that these probability estimators of

class membership are not unbiased. Nevertheless, those estimates can be useful in land cover

mapping, by assigning to each pixel a relative degree of classification confidence. In this study we

used 25 bootstrap replicates to build a PBCT, evaluated with a 10-fold cross validation approach.

Breiman (1996) suggests that a higher number of replicates tend not to produce a significant test set

error reduction.

The k-nearest neighbors (K-NN) is a non-parametric classification method. In the machine

learning terminology it is called an instance-based method (Mitchell, 1997). The K-NN classifier

estimates, for each single x in the image, the probability P(classi /x) from the training sample. Let

{x1,...,xk} be the k elements of the training sample which are closest to x in the spectral space with

respect to the Euclidean distance. Those are called the nearest neighbors of x. The estimate of

P(classi /x) is simply the proportion of pixels that belong to classi among the k nearest neighbors of x. If

k=1, then the method is named the nearest neighbor classifier and it simply assigns each pixel in the

image to the class of its training sample nearest neighbor. This can lead to overfitting the classification

to the training data. Choosing a larger k can overcome that problem. So, one crucial parameter of the

K-NN classifier is the number of neighbors, k. In this work, we looked for the optimum k by minimizing

the overall (for all classes) test set error. To estimate k we ran the method for k=1,2,…,10, plotted the

corresponding cross-validation estimate of the overall test set error, and chose the best k.

13

4.2. Accuracy assessment

Accuracy assessment of the land cover map obtained with each classifier was evaluated both

quantitatively and qualitatively. The accuracy of each classifier was assessed by a 10-fold cross

validation procedure, with associated error matrices and measures of overall and class-specific

accuracy (Congalton & Green, 1999; Foody, 2002; Smits et al., 1999). Accuracy assessment was

performed on spatial independent testing subsets, to avoid overestimation of classification accuracy

results (Friedl et al., 2000, 2002). The error matrix of each classifier is used as an approximation for

the accuracy of the land cover map. The difficulty in obtaining up-to-date independent information of

spatial distribution of land cover classes in the BLA led us to choose a complementary approach. As

mentioned before, the application of the PBCT algorithm can provide information of class membership

probability. Therefore, application of this classifier to the entire BLA can provide a map of class

membership probability, that is, in each pixel, the highest averaged probability over the sub-classifiers

derived from the 25 replicates of the training samples. Friedl et al. (2002) and McIver & Friedl (2001)

and used a similar approach to build maps of classification confidence, based on recent results

explaining boosting as a form of additive logistic regression (Friedman et al., 2000).

The most straightforward way of estimating total areas for each class is to count the number of

pixels in each one after classification. This is called “naïve estimation” in the remote sensing literature

(Gallego, 2004). As a matter of fact, areal estimates obtained this way are biased (e.g. Gallego, 2004;

Walsh & Burk, 1993). Some post-classification methods have been used to improve these estimates,

namely the so called calibration techniques (e.g. Krutchkoff, 1967). They are basically divided into

classical and inverse methods, and they utilize the information contained in the error matrix to correct

for misclassification bias (Walsh & Burk, 1993). Walsh & Burk (1993) carried out a simulation study

and concluded that the inverse method consistently performed better than the classical method. The

idea behind that calibration technique is to use the misclassification probabilities among classes to

revise the proportions given by pixel counting (pi). In particular, the error matrix allows us to estimate

the probabilities of classification of a pixel of class i in class j, as nij/n.j (where i represents the

observed class, j the predicted class, and n.j the sum of all training observations predicted as class j).

14

Then, the revised proportion for class i is given by ∑j

jijj nnp ./ . It has been shown that the inverse

method leads to unbiased estimates for total areas (Tenenbein, 1972, Walsh & Burk, 1993).

A more qualitative comparison with existing land cover maps was also performed, by

assessing overall class extent in the BLA. The comparison was made with three known available

global land cover datasets (Eva et al., 2004; Friedl et al., 2002; Hansen et al., 2000). Hansen et al.

(2000) produced a 1-km 12-class global land cover map (a.k.a. University of Maryland global land

cover map), using AVHRR data from 1992-1993. Friedl et al. (2002) produced a 1-km 17-class global

land cover map (a.k.a. Boston University global land cover map), using Terra/Aqua Moderate

Resolution Imaging Spectroradiometer (MODIS) data from 2001, and updated at quarterly (96 days)

intervals. Eva et al. (2004) produced a 1-km 42-class vegetation map of South America for the year

2000, as part of the GLC2000 project, using four sets of satellite data: Along Track Scanning

Radiometer (ATSR-2) onboard the ERS-2 satellite, SPOT-4 VGT, JERS-1 radar data, and Defense

Meteorological Satellite Program (DMSP) Operational Linescan System (OLS) data. To establish a

comparison we had to aggregate primary tropical forest and secondary succession forest classes, as

the abovementioned studies do not discriminate between those two classes.

5. Results

5.1. Image classification

A total of 8386 pixels of known land cover, corresponding to 2264 sampling polygons, were identified

in the Landsat imagery and overlaid on the SPOT-4 VGT monthly composites. The distribution per

land cover class is shown in Table 1. The precision of each classification algorithm was evaluated

using 10-fold cross validation. The criterion used for the K-NN algorithm resulted in the selection of

k=7. Comparison with this relatively large number of neighbors reduces the danger of overfitting. The

15

comparison of the four algorithms, using overall and class-specific accuracy is shown in Table 3. All

algorithms yielded similar results for the natural/artificial waterbodies, agriculture/pasture, and primary

tropical forest classes, displaying commission and omission errors always below 0.10, but lower for

the K-NN and PBCT classifiers. The cerrado savanna class had commission and omission errors

between 0.166-0.247 in all algorithms, but lower in the K-NN algorithm. The secondary succession

forest class was the most problematic, with high commission and omission errors. Nevertheless, the

PBCT classifier produced the lower commission (0.432) and omission (0.662) errors for this class. In

this algorithm, the major confusion occurred between the secondary succession forest class and the

primary tropical forest and agriculture/pasture classes (Table 4). The cerrado savanna class had also

considerable commission (0.190) and omission (0.219) errors, resulting from confusion with primary

tropical forest and agriculture/pasture classes.

[Insert Table 3 and Table 4 about here]

Comparison of the overall class extent in the BLA produced by the four tested algorithms,

using the inverse method calibration technique, is shown in Table 5. Although there are considerable

variations in some of the classes, the main discrepancy was in the secondary succession forest class.

The difference from the lowest extent, generated by the PBCT classifier (140 x 103 km2) to the highest,

produced by the QDA (233 x 103 km2) is quite large. However, the most accurate algorithms, namely

PBCT and K-NN, produced a very similar overall extent of the mapped land cover classes (e.g. 140 x

103 km2 and 156 x 103 km2 in secondary succession forest, respectively). Of the four algorithms, the

PBCT yielded the greatest areal estimate of primary tropical forest and the lowest estimate of cerrado

savanna and secondary succession forest.

[Insert Table 5 about here]

16

Application of the PBCT algorithm to the entire dataset resulted in the land cover map of the

BLA for the year 2000 shown in Fig. 2A. The class represented in each pixel is that with the highest

probability, averaged over the 25 sub-classifiers. Pixels with class membership probability lower than

0.6 are shown in white; the corresponding color intensity increases according to the value of class

membership probability. Fig. 2B shows a detailed area over the state of Pará, highlighting several

regions with lower class membership probability. The calibrated area and percentage of each land

cover class per state is presented in Table 6, as well as the mean class membership probability. The

state of Maranhão had the highest proportion of agriculture/pasture, with 61.0% (122 x 103 km2),

followed by Tocantins with 39.4% (110 x 103 km2), and Mato Grosso with 36.9% (334 x 103 km2). The

state of Rondônia with 28.3% (68 x 103 km2), Pará with 17.5% (219 x 103 km2), and Roraima with

14.3% (32 x 103 km2) also had a significant proportion of their territory occupied with

agriculture/pasture in 2000. The states of Amazonas (3.1%), Amapá (9.8%), and Acre (11.3%),

displayed a relatively low proportion of agriculture/pasture. The mean membership probability for this

class ranged from 0.70 (Amazonas and Roraima) to 0.90 (Mato Grosso). Secondary succession forest

was concentrated in Pará with 49 x 103 km2 (3.9%), Amazonas with 42 x 103 km2 (2.6%), Mato Grosso

with 17 x 103 km2 (1.9%), and Maranhão with 9 x 103 km2 (4.5%). The highest mean membership

probability for this class occured in Maranhão (0.64) and the lowest in Roraima (0.49).

[Insert Fig. 2 and Table 6 about here]

5.2. Accuracy assessment

The information about class membership probability (Fig. 2) was derived from the output of the PBCT

algorithm. It gives useful information about the relative degree of membership of the most probable

class. Natural/artificial waterbodies and primary tropical forest had the highest degree of class

membership, i.e. they were the classes mapped with the highest degree of confidence. Some areas

associated with cerrado savanna and agriculture/pasture displayed lower class membership

probability. The class with highest commission and omission errors, secondary succession forest, also

17

had low class membership probability, meaning that allocation of a pixel to this class was done with a

relatively low degree of confidence.

6. Discussion

6.1. Image classification

The misclassification among secondary succession forest, primary tropical forest, and

agriculture/pasture (Table 4) is understandable, since from a spectral standpoint, secondary

succession forest is a transitional class between agriculture/pasture and primary tropical forest. After

deforestation, and following land abandonment, the initial stages of forest regeneration are spectrally

similar to agriculture/pasture; conversely, final stages of forest regeneration are more related to

primary tropical forest. Lucas et al. (2000), using NOAA AVHRR data to map secondary succession

forest in the BLA during early 1990s, reported a similar error pattern, with omission and commission

error ranging between 0.55-0.75 and 0.51-0.76, respectively. The problem can be better understood

by looking at the distribution of the spectral signatures in the plane of the two main discriminant axes,

which maximize the spectral distance between classes (Fig. 3). It is particularly evident some

confusion of the secondary succession forest class (5) with agriculture/pasture (3) and primary tropical

forest classes (1). The same difficulty exists in accurately mapping cerrado savanna (Table 4),

although at a lesser extent and not visible in Fig. 3 (only showing 20 randomly selected spectral

signatures per class). Regions of cerrado savanna with high tree density (cerradão) may have

structural and spectral characteristics similar to those of primary tropical forest, and areas of cerrado

savanna dominated by shrubs and herbaceous vegetation (campos) may be spectrally/structurally

similar to areas currently under agriculture/pasture use. These results suggest that a different

methodology should be used to help discrimination of the two major problematic classes, secondary

succession forest and cerrado savanna. Using ancillary data to improve the discrimination of classes

that are difficult to discriminate has been proven to be helpful in remote sensing studies (e.g. Brown et

al., 1993; Eva et al., 2004; Foody & Hill, 1996; Friedl et al., 2002; McIver & Friedl, 2002; Strahler,

18

1980). This information could be particularly useful in large scale land cover mapping (i.e. continental

or global), characterized by high within-class variability (McIver & Friedl, 2002). Some methods use

ancillary data to help labeling cluster classes resulting from unsupervised classification algorithms

(Brown et al., 1993; Eva et al., 2004); these can include elevation, soil, and land use/land cover data.

Other methods rely on ancillary data to improve the ability of supervised classification algorithms to

discriminate land cover classes (Friedl et al., 2002; McIver & Friedl, 2002; Strahler, 1980). McIver &

Friedl (2002) showed that the incorporation of prior probabilities of class membership in a supervised

non-parametric classification algorithm successfully improved land cover classification results; these

are being used by the MODIS land cover classification algorithm (MLCCA) to produce the Boston

University land cover map (Friedl et al., 2002).

[Insert Fig. 3 about here]

QDA, SCT, and PBCT/K-NN yielded a distinct overall class extent using the same training set,

especially for the secondary succession forest class. The QDA and SCT algorithms seem to behave

consistently worse than K-NN or PBCT (Table 3). Both PBCT and K-NN produce a very similar class

extent, and they are also the two algorithms with higher overall accuracy. Furthermore, the PBCT

classifier proved to be the one with lower commission and omission errors for the secondary

succession forest class, as well as producing the lower omission error and the second lower

commission error for the agriculture/pasture class. Therefore, this algorithm is the one that better

discriminated these two classes.

Deforestation started long ago in the states of Maranhão and Pará (Nepstad et al., 1997) and

a part of this area appears to have been abandoned, allowing for the regrowth of secondary

succession forest. If the combined area of agriculture/pasture and secondary succession forest in

2000 (1106 x 103 km2), mapped with the PBCT algorithm, can be viewed as a proxy for the total

deforested area in the BLA by 2000, then this number is almost double the cumulative deforestation

value of 587,727 km2 reported by INPE (2002) up to the year 2000. Discrepancies between this study

and INPE’s estimate are more evident in those states where the proportion of cerrado savanna is

19

higher (e.g. Tocantins, Maranhão, Mato Grosso, Pará, Amapá, and Roraima) (Fig. 4). The objective of

INPE’s methodology is to map deforestation that took place in areas previously occupied by primary

tropical forest. Comparison with an existing vegetation map (IBGE, 1988), indicates that establishment

of agriculture/pasture in Amapá, Maranhão, Mato Grosso, Roraima, and Tocantins, in regions of

cerrado savanna, has been as significant as in areas of primary tropical forest (Fig. 5). Our analysis

indicates that only 46% of the area in the BLA with agriculture/pasture in 2000 has established in

areas previously occupied by primary tropical forest. A large amount of the area with

agriculture/pasture in 2000 is located in regions formerly occupied by cerrado savanna (33%); the

remaining was established in areas of transition between the previous classes (19%), and, to a much

lesser extent, in other vegetation types (2%). Consequently, our results bring further evidence of large-

scale conversion in cerrado savanna, supporting the findings of Cardille & Foley (2003), and the

remarks in Fearnside (1993), Kaimowitz & Smith (2001), and Nepstad et al. (1997). The

abovementioned discrepancies between this study and the values of INPE (2002) are largely

explained by the extent of agriculture/pasture established in areas of cerrado savanna, not considered

in their analysis.

[Insert Fig. 4 and Fig. 5 about here]

The estimated value of 140 x 103 km2 for the overall extent of secondary succession forest in

the BLA is lower than the 157,973 km2 from Lucas et al. (2000) for early 1990s. These authors

mapped forest regeneration with an unsupervised classification algorithm and NOAA AVHRR data,

relying in ancillary information to label the resulting clusters. The type of classification algorithm used

can generate substantially different estimates of this class, as shown in Table 5. The fact that our

study does not consider the part of Maranhão between 44º and 45º W explain a part of that difference,

as this state has the highest incidence of forest regeneration (4.5%, Table 6); the secondary

succession forest map produced by Lucas et al. (2000) shows a large incidence of this class between

those two meridians. The small size of some regenerating areas could also result in the

underestimation by a sensor with 1-km spatial resolution, like the SPOT VGT. Perhaps most important

is the fact that mitigation measures implemented by some state governments to abate deforestation of

20

primary tropical forest and cerrado savanna (e.g. Fearnside, 2003; Fearnside & Barbosa, 2003) could

result in the re-deforestation of secondary succession forest, thus reducing its extent. The distribution

of the areas of secondary succession forest in the BLA in 2000 indicates that the majority of this class

occurred in deforested areas previously occupied by primary tropical forest (88%); the exceptions are

the states of Mato Grosso, Tocantins, and Amapá, where forest regeneration appears in areas

originally covered by transition forest, cerrado savanna, and other vegetation types, respectively (Fig.

6). This is not surprising, as some farmers often abandon deforested areas due to intense vegetation

regrowth in areas of primary tropical forest (Fearnside, 1993; Moran et al., 1994; Nepstad et al., 1997;

Skole et al., 1994).

[Insert Fig. 6 about here]

Our combined areal estimates of agriculture/pasture and secondary succession forest for the

entire BLA in 2000 (comparable with INPE’s deforestation data by 2000) added up to 1106 x 103 km2,

and occurred mainly in primary tropical forest (531 x 103 km2), cerrado savanna (343 x 103 km2),

transition forest (199 x 103 km2), and other vegetation types (33 x 103 km2).

6.2. Accuracy assessment

The lower probability of class membership in the state of Roraima may be due to residual cloud

contamination in the monthly composite images. These class membership probability estimates were

compared with class-specific accuracy (i.e. commission error) and used to extend and complete

conventional accuracy assessment (Fig. 7). It can be seen that the class-specific mean membership

probability is highly inversely correlated with classification errors, i.e. higher classification errors tend

to have lower class membership probability. The lower class membership probability is associated with

the secondary succession forest class, which is in agreement with the higher omission and

commission errors. The agriculture/pasture class is the only one with a larger deviance from a linear

21

relation, represented by the y=1-x line. This class represents different types of landscape structures,

ranging from bare soil to fully developed agricultural crops, originating a higher dispersion of class

membership probability, as indicated by the highest standard deviation. A higher standard deviation is

also associated with cerrado savanna, a heterogeneous class as well, ranging from herbaceous

vegetation to closed-canopy forest. Also, the overall mean probability of class membership for the

entire BLA was 0.90 (0.15 standard deviation). This outcome corroborates the results of McIver &

Friedl (2001), who used a similar approach to derive estimates of classification confidence, and

showed that incorrectly classified areas (pixels) tend to have low classification confidence.

Consequently, the estimates of class membership probability add supplementary information

regarding map quality, mainly by defining the degree of classification confidence on a pixel-by-pixel

basis.

[Insert Fig. 7 about here]

A more qualitative analysis was performed by comparing overall estimates of class extent with

those from other studies, namely Eva et al. (2004), Friedl et al. (2002), and Hansen et al. (2000) (Fig.

8). In these studies, secondary succession forest class was not mapped separately, but was included

in various forest classes. Thus, for results to be comparable, we aggregated our estimates of primary

tropical forest and secondary succession forest. The main discrepancies between our study and

Hansen et al. (2000) are the extent of cerrado savanna and agriculture/pasture; the images used by

Hansen et al. (2000) are from the early 1990s, which may contribute to explain the low amount of

agriculture/pasture mapped; another explanation could be that they did not accurately mapped

agriculture/pasture in areas previously occupied by cerrado savanna; the poor spectral and geometric

characteristics of the NOAA AVHRR sensor compared to SPOT VGT may also explain that difference.

Friedl et al. (2002) used MODIS imagery from 2001 and a boosted decision tree, with an overall

classifier confidence of 0.90 (0.29 standard deviation), comparable to our estimated overall class

membership probability. The main difference with our results is in the agriculture/pasture class. They

mapped a substantially lower extent of agriculture/pasture, with a corresponding increase in cerrado

savanna, and aggregated primary tropical forest with secondary succession forest. The extent of

22

cropland, derived from their work (comparable to our agriculture/pasture class), reached a value

equivalent to the cumulative deforestation up to 1988 (377,500 km2), estimated by INPE (INPE, 2002).

Eva et al. (2004) utilized the same original SPOT VGT dataset used in this study, but used a different

image compositing procedure and an unsupervised classification algorithm (isodata). The overall

estimates of land cover classes between these two studies are very similar, the major differences

occurring in the aggregated primary tropical forest with secondary succession forest class, where our

analysis points to a smaller area, and in the agriculture/pasture class, which we mapped to a larger

extent. The type of classification algorithm used can explain that difference, as we have shown that

different classification algorithms produce different estimates of class extent.

[Insert Fig. 8 about here]

7. Conclusions

The higher amount of aggregated agriculture/pasture and secondary succession forest classes in

2000 obtained in this study (1106 x 103 km2), when compared with deforestation estimates up to 2000

from INPE (2002) (587,727 km2), is mostly due to agriculture/pasture occurring in areas previously

occupied by cerrado savanna, not considered in INPE’s estimates. It appears therefore that

agriculture/pasture establishment in areas previously occupied by cerrado savanna was almost as

important as in areas of primary tropical forest. The comparison of four classification algorithms

showed that the PBCT and the K-NN classifiers tend to perform better than the QDA and SCT. The

PBCT performs better especially over the secondary succession forest class, with lower commission

and omission errors. However, this class is problematical as it is extensively confused with primary

tropical forest and agriculture/pasture. The possibility of obtaining maps of class membership

probability derived from PBCT further improved the spatial characterization of agriculture/pasture and

secondary succession forest classes. We have shown that higher class membership probability is

associated with lower classification errors. This approach adds supplementary information for

accuracy assessment of land cover maps, useful, for instance, by assigning a given uncertainty to

23

each pixel. This study demonstrated the utility of the SPOT-4 VGT sensor in predicting the extent of

agriculture/pasture in the BLA with a reasonable degree of accuracy, as well as for the primary tropical

forest, cerrado savanna, and natural/artificial waterbodies classes. The lower accuracy in the

secondary succession forest class indicates that a different sensor/methodology should be used to

assess that land cover class (e.g. multi-year analysis, hyperspectral data, ancillary data). Overall class

areal estimates derived from the land cover map may be useful for analysis of regional carbon and

water fluxes, and for evaluating impacts of land use/land cover change on biotic diversity and soil

degradation.

Acknowledgments

The authors gratefully acknowledge the three anonymous reviewers for their helpful comments and

suggestions. J.M.B. Carreiras’s work was partially developed at Instituto Nacional de Pesquisas

Espaciais (INPE, Brazil), as a contribution to the GLC2000 and LBA projects (doctoral grant Ref.

PRAXIS XXI/BD/21507/99). J.M.C. Pereira’s participated in this study under the scope of research

project “Fire in the Brazilian Amazon: multi-year mapping of area burned and estimation of pyrogenic

emissions using remotely sensed data” (Ref. POCTI/CTA/45126/2002). M.L. Campagnolo’s

contribution was partly done while visiting the Department of Geography at the University of Maryland

(post-doctoral grant Ref. SFRH/BPD/21012/2004). Funding was obtained from the Ministério da

Ciência e Tecnologia, Fundação para a Ciência e a Tecnologia, Portugal. VEGETATION images were

made available in the framework of the GLC 2000 and GBA 2000 projects of the Joint Research

Centre (JRC) of the European Commission. We acknowledge Instituto Brasileiro de Geografia e

Estatística (IBGE) for the Mapa de Vegetação do Brasil (Vegetation Map of Brazil), distributed by the

University of New Hampshire, EOS-WEBSTER Earth Science Information Partner (ESIP). We

acknowledge INPE and the GLCF (University of Maryland) for the Landsat data used in this study.

24

References

Achard, F., Eva, H.D., & Mayaux, P. (2001). Tropical forest mapping from coarse spatial resolution

satellite data: production and accuracy assessment issues. International Journal of Remote Sensing,

22(14), 2741-2762.

Achard, F., Eva, H.D., Stibig, H.-J., Mayaux, P., Gallego, J., Richards, T., & Malingreau, J.-P. (2002).

Determination of deforestation rates of the world’s humid tropical forests. Science, 297, 999-1002.

Adams, J.B., Sabol, D.E., Kapos, V., Almeida Filho, R., Roberts, D.A., Smith, M.O., & Gillespie, A.R.

(1995). Classification of multispectral images based on fractions of endmembers: application to land-

cover change in the Brazilian Amazon. Remote Sensing of Environment, 52(2), 137-154.

Alves, D.S. (2002). An analysis of the geographical patterns of deforestation in the Brazilian Amazon

in the period 1991-1996. In C.H. Wood, & R. Porro (Eds.), Deforestation and Land Use in the Amazon

(pp. 95-106). Gainsville, Florida: University Press of Florida.

Asner, G.P., Knapp, D.E., Broadbent, E.N., Oliveira, P.J.C., Keller, M., & Silva, J.N. (2005). Selective

logging in the Brazilian Amazon. Science, 310, 480-482.

Bartholomé, E., & Belward, A.S. (2005). GLC2000: a new approach to global land cover mapping from

Earth observation data. International Journal of Remote Sensing, 26(9), 1959-1977.

Bauer, E., & Kohavi, R. (1998). An empirical comparison of voting classification algorithms: bagging,

boosting, and variants. Machine Learning, 36, 105-142.

Breiman, L., Friedman, J., Olshen, R., & Stone, C. (1984). Classification and Regression Trees.

Pacific Grove, California: Wadsworth.

25

Breiman, L. (1996). Bagging predictors. Machine Learning, 24(2), 123-140.

Breiman, L. (1998). Arcing classifiers. The Annals of Statistics, 26(3), 801-849.

Brown, J.F., Loveland, T.R., Merchant, J.W., Reed, B.C., & Ohlen, D.O. (1993). Using multisource

data in global land-cover characterization: concepts, requirements, and methods. Photogrammetric

Engineering & Remote Sensing, 59(6), 977-987.

Brown, S., & Lugo, A.E. (1990). Tropical secondary forests. Journal of Tropical Ecology, 6(1), 1-32.

Buschbacher, R., Uhl, C., & Serrão, E.A.S. (1988). Abandoned pastures in eastern Amazonia. II.

Nutrient stocks in the soil and vegetation. Journal of Ecology, 76(3), 682-699.

Cardille, J.A., & Foley, J.A. (2003). Agricultural land-use change in Brazilian Amazônia between 1980

and 1995: evidence from integrated satellite and census data. Remote Sensing of Environment, 87,

551-562.

Carreiras, J.M.B., Shimabukuro, Y.E., & Pereira, J.M.C. (2002). Fraction images derived from SPOT-4

VEGETATION data to assess land-cover change over the state of Mato Grosso, Brazil. International

Journal of Remote Sensing, 23(23), 4979-4983.

Carreiras, J.M.B., Pereira, J.M.C., & Shimabukuro, Y.E. (2004). Land cover mapping in the Brazilian

Amazon using SPOT-4 VEGETATION data and machine learning classification methods.

Photogrammetric Engineering & Remote Sensing (accepted).

Carreiras, J.M.B., & Pereira, J.M.C. (2005). SPOT-4 VEGETATION multitemporal compositing for land

cover change studies over tropical regions. International Journal of Remote Sensing, 26(7), 1323-

1346.

26

Cochrane, M.A., Alencar, A., Schulze, M.D., Souza Jr., C.M., Nepstad, D.C., Lefebvre, P., & Davidson,

E.A. (1999). Positive Feedbacks on the fire dynamic of closed canopy tropical forests. Science, 284,

1832-1835.

Congalton, R.G., & Green, K. (1999). Assessing the Accuracy of Remotely Sensed Data: Principles

and Practices. Boca Raton, FL: Lewis Publishers.

Curran, P.J., & Foody, G.M. (1994). The use of remote sensing to characterise the regenerative states

of tropical forests. In G.M. Foody, & P.J. Curran (Eds.), Environmental Remote Sensing from Regional

to Global Scales (pp. 44-83). Chichester, UK: John Wiley & Sons.

Eva, H.D., Belward, A.S., De Miranda, E.E., Di Bella, C.M., Gonds, V., Huber, O., Jones, S.,

Sgrenzaroli, M., & Fritz, S. (2004). A land cover map of South America. Global Change Biology, 10,

731-744.

Fearnside, P.M. (1993). Deforestation in Brazilian Amazonia: the effect of population and land tenure.

Ambio, 22(8), 537-545.

Fearnside, P.M. (1996). Amazonian deforestation and global warming: carbon stocks in vegetation

replacing Brazil’s Amazon forest. Forest Ecology and Management, 80, 21-34.

Fearnside, P.M. (2003). Deforestation control in Mato Grosso: a new model for slowing the loss of

Brazil’s Amazon forest. Ambio, 32(5), 343-345.

Fearnside, P.M., & Barbosa, R.I. (2003). Avoided deforestation in Amazonia as a global warming

mitigation measure: the case of Mato Grosso. World Resource Review, 15(3), 352-361.

27

Ferreira, L.G., Yoshioka, H., Huete, A., & Sano, E.E. (2003). Seasonal landscape and spectral

vegetation índex dynamics in the Brazilian Cerrado: An analysis within the Large-Scale Biosphere-

Atmosphere Experiment in Amazonia (LBA). Remote Sensing of Environment, 87, 534-550.

Foody, G.M., & Hill, R.A. (1996). Classification of tropical forest classes from Landsat TM data.

International Journal of Remote Sensing, 17(12), 2353-2367.

Foody, G.M., Palubinskas, G., Lucas, R.M., Curran, P.J., & Honzák, M. (1996). Identifying terrestrial

carbon sinks: classification of successional stages in regenerating tropical forest from Landsat TM

data. Remote Sensing of Environment, 55, 205-216.

Foody, G.M. (2002). Status of land cover classification accuracy assessment. Remote Sensing of

Environment, 80, 185-201.

Friedl, M.A., Woodcock, C., Gopal, S., Muchoney, D., Strahler, A.H., & Barker-Schaaf, C. (2000). A

note on procedures used for accuracy assessment in land cover maps derived from AVHRR data.

International Journal of Remote Sensing, 21(5), 1073-1077.

Friedl, M.A., McIver, D.K., Hodges, J.C.F., Zhang, X.Y., Muchoney, D., Strahler, A.H., Woodcock,

C.E., Gopal, S., Schneider, A., Cooper, A., Baccini, A., Gao, F., & Schaaf, C. (2002). Global land

cover mapping from MODIS: algorithms and early results. Remote Sensing of Environment, 83(1/2),

287-302.

Friedman, J., Hastie, T., & Tibshirani, R. (2000). Additive logistic regression: a statistical view of

boosting. The Annals of Statistics, 28(2), 337-407.

Gallego, F.J. (2004). Remote sensing and land cover area estimation. International Journal of Remote

Sensing, 25(15), 3019-1047.

28

Gedney, N., & Valdes, P.J. (2000). The effect of Amazonian deforestation on the northern hemisphere

circulation and climate. Geophysical Research Letters, 27(19), 3053-3056.

Goulding, M., Barthem, R., & Ferreira, E. (2003). The Smithsonian Atlas of the Amazon. Washington:

Smithsonian Books.

Green, A.A., Berman, M., Switzer, P., & Craig, M.D. (1988). A transformation for ordering multispectral

data in terms of image quality with implications for noise removal. IEEE Transactions on Geoscience

and Remote Sensing, 26(1), 65-74.

Hansen, M., DeFries, R., Townshend, J.R.G., & Sohlberg, R. (2000). Global land cover classification

at 1km resolution using a decision tree classifier. International Journal of Remote Sensing, 21(6/7),

1331-1364.

Houghton, R.A., Skole, D.L., Nobre, C.A., Hackler, J.L., Lawrence, K.T., & Chomentowski, W.H.

(2000). Annual fluxes of carbon from deforestation and regrowth in the Brazilian Amazon. Nature, 403,

301-304.

Houghton, R.A., Lawrence, K.T., Hackler, J.L., & Brown, S. (2001). The spatial distribution of forest

biomass in the Brazilian Amazon: a comparison of estimates. Global Change Biology, 7, 731-746.

Huete, A.R. (1988). A soil-adjusted vegetation index (SAVI). Remote Sensing of Environment, 25,

295-309.

IBGE (1988). Mapa de Vegetação do Brasil, escala 1:5000000. Brasília, DF, Brazil: Ministério da

Agricultura, Instituto Brasileiro de Desenvolvimento Florestal, Secretária de Planejamento e

Coordenação da Presidência da Republica, Fundação Instituto Brasileiro de Geografia e Estatística.

29

INPE (2002). Monitoring of the Brazilian Amazonian forest by satellite, 2000-2001. São José dos

Campos, SP, Brazil: Instituto Nacional de Pesquisas Espaciais.

Junk, W.J., & Firch, K. (1985). The physical and chemical properties of Amazonian waters and their

relationship with the biota. In G.T. Prance, & T.E. Lovejoy (Eds.), Amazonia (pp. 3-17). New York:

Pergamon.

Kaimowitz, D., & Smith, J. (2001). Soybean technology and the loss of natural vegetation in Brazil and

Bolivia. In A. Angelsen, & D. Kaimowitz (Eds.), Agricultural Technologies and Tropical Deforestation

(pp. 195-211). Oxon, UK: CAB International Publishing.

Keeling, C.D., & Whorf, T.P. (2005). Atmospheric CO2 records from sites in the SIO air sampling

network. In Trends: A Compendium of Data on Global Change. Oak Ridge, Tennessee, USA: Carbon

Dioxide Information Analysis Center, Oak Ridge National Laboratory, U.S. Department of Energy.

Kimes, D.S., Nelson, R.F., Salas, W.A., & Skole, D.L. (1999). Mapping secondary tropical forest and

forest age from SPOT HRV data. International Journal of Remote Sensing, 20(18), 3625-3640.

Korner, C. (2000). Biosphere responses to CO2 enrichment. Ecological Applications, 10(6), 1590-

1619.

Krutchkoff, R.G. (1967). Classical and inverse regression methods of calibration. Technometrics, 9,

425-440.

Lambin, E.F., Geist, H.J., & Lepers, E. (2003). Dynamics of land-use and land-cover change in tropical

regions. Annual Review of Environment and Resources, 28, 205-241.

30

Lu, D., Moran, E., & Batistella, M. (2003). Linear mixture model applied to Amazonian vegetation

classification. Remote Sensing of Environment, 87, 456-469.

Lucas, R.M., Honzák, M., Curran, P.J., Foody, G.M., Milne, R., Brown, T., & Amaral, S. (2000).

Mapping the regional extent of tropical forest regeneration stages in the Brazilian legal Amazon using

NOAA AVHRR data. International Journal of Remote Sensing, 21(15), 2855-2881.

Lucas, R.M., Honzák, M., Do Amaral, I., Curran, P.J., & Foody, G.M. (2002). Forest regeneration on

abandoned clearances in central Amazonia. International Journal of Remote Sensing, 23(5), 965-988.

Maclin, R., & Opitz, D. (1997). An empirical evaluation of bagging and boosting. Proceedings of the

Fourteenth National Conference on Artificial Intelligence (pp. 546-551). 27-31 July 1997, Providence,

Rhode Island: American Association for Artificial Intelligence Press.

Mausel, P., Wu, Y., Li, Y., Moran, E.F., & Brondizio, E.S. (1993). Spectral identification of successional

stages following deforestation in the Amazon. Geocarto International, 4, 61-71.

McIver, D.K., & Friedl, M.A. (2001). Estimating pixel-scale land cover classification confidence using

nonparametric machine learning methods. IEEE Transactions on Geoscience and Remote Sensing,

39(9), 1959-1968.

McIver, D.K., & Friedl, M.A. (2002). Using prior probabilities in decision-tree classification of remotely

sensed data. Remote Sensing of Environment, 81, 253-261.

Mesquita, R.C.G., Ickes, K., Ganade, G., & Williamson, G.B. (2001). Alternative successional

pathways in the Amazon Basin. Journal of Ecology, 89, 528-537.

31

Mitchell, T.M. (1997). Machine Learning. Singapore: McGraw-Hill.

Moran, E.F., Brondizio, E., Mausel, P., & Wu, Y. (1994). Integrating Amazonian vegetation, land-use,

and satellite data. Bioscience, 44(5), 329-338.

Nepstad, D.C., Klink, C.A., Uhl, C., Vieira, I.C., Lefebvre, P., Pedlowski, M., Matricardi, E., Negreiros,

G., Brown, I.F., Amaral, E., Homma, A., & Walker, R. (1997). Land-use in Amazonia and the Cerrado

of Brazil. Ciência e Cultura Journal of the Brazilian Association for the Advancement of Science,

49(1/2), 73-86.

Passot, X. (2000). VEGETATION image processing methods in the CTIV. Proceedings of

VEGETATION 2000 - 2 years of operation to prepare the future (pp. 15-21). 3-6 April 2000, Ispra,

Italy: SAI/JRC.

Perlich, C., Provost, F., & Simonoff, J.S. (2003). Tree induction vs. logistic regression: a learning-curve

analysis. Journal of Machine Learning Research, 4, 211-255.

Perz, S.G., & Skole, D.L. (2003). Social determinants of secondary forests in the Brazilian Amazon.

Social Science Research, 32, 25-60.

Petit, J.R., Jouzel, J., Raynaud, D., Barkov, N.I., Barnola, J.-M., Basile, I., Bender, M., Chappellaz, J.,

Davis, M., Delaygue, G., Delmotte, M., Kotlyakov, V.M., Legrand, M., Lipenkov, V.Y., Lorius, C.,

Pépin, L., Ritz, C., Saltzman, E., & Stievenard, M. (1999). Climate and atmospheric history of the past

420,000 years from the Vostok ice core, Antarctica. Nature, 399, 429-436.

Provost, F., & Domingos, P. (2003). Tree induction for probability-based ranking. Machine Learning,

52(3), 199-215.

32

Richards, J.A. (1986). Remote Sensing Digital Image Analysis. Berlin: Springer-Verlag.

Roberts, D.A., Keller, M., & Soares, J.V. (2003). Studies of land-cover, land-use, and biophysical

properties of vegetation in the large Scale Biosphere Atmosphere experiment in Amazônia. Remote

Sensing of Environment, 87, 377-388.

Rodríguez-Yi, J.L., Shimabukuro, Y.E., & Rudorff, B.F.T. (2000). Image segmentation for classification

of vegetation using NOAA AVHRR data. International Journal of Remote Sensing, 21(1), 167-172.

Safavian, S.R., & Landgrebe, D. (1991). A survey of decision tree classifier methodology. IEEE

Transactions on Systems, Man, and Cybernetics, 21(3), 660-674.

Sano, E.E., Barcellos, A.O., & Bezerra, H.S (2001). Assessing the spatial distribution of cultivated

pastures in the Brazilian savanna. Pasturas Tropicales, 22(3), 2-15.

Schroeder, P.E., & Winjum, J.K. (1995). Assessing Brazil’s carbon budget: II. Biotic fluxes and net

carbon balance. Forest Ecology and Management, 75, 87-99.

Skole, D., & Tucker, C. (1993). Tropical deforestation and habitat fragmentation in the Amazon:

Satellite data from 1978 to 1988. Science, 260, 1905– 1910.

Skole, D.L., Chomentowski, W.H., Salas, W.A., & Nobre, A.D. (1994). Physical and human dimensions

of deforestation in Amazonia. Bioscience, 44(5), 314-322.

Smits, P.C., Dellepiane, S.G., & Schowengerdt, R.A. (1999). Quality assessment of image

classification algorithms for land cover mapping: a review and a proposal for a cost-based approach.

International Journal of Remote Sensing, 20(8), 1461-1486.

33

Steininger, M.K. (2000). Satellite estimation of tropical secondary forest above-ground biomass: data

from Brazil and Bolivia. International Journal of Remote Sensing, 21(6/7), 1139-1157.

Stone, T.A., Schlesinger, P., Houghton, R.A., & Woodwell, G.M. (1994). A map of the vegetation of

South America based on satellite imagery. Photogrammetric Engineering & Remote Sensing, 60, 541-

551.

Strahler, A.H. (1980). The use of prior probabilities in maximum likelihood classification of remotely

sensed data. Remote Sensing of Environment, 10, 135-163.

Souza Jr., C., Firestone, L., Moreira Silva, L., & Roberts, D. (2003). Mapping forest degradation in the

Eastern Amazon from SPOT 4 through spectral mixture models. Remote Sensing of Environment, 87,

494-506.

Tenenbein, A. (1972). A double sampling scheme for estimating from misclassified multinomial data

with applications to sampling inspection. Technometrics, 14(1), 187-202.

Uhl. C. (1987). Factors controlling succession following slash-and-burn agriculture in Amazonia.

Journal of Ecology, 75(2), 377-407.

Uhl, C., Buschbacher, R., & Serrão, E.A.S. (1988). Abandoned pastures in eastern Amazonia. I.

Patterns of plant succession. Journal of Ecology, 76(3), 663-681.

34

Vieira, I.C.G., Almeida, A.S., Davidson, E.A., Stone, T.A., Carvalho, C.J.R., & Guerrero, J.B. (2003).

Classifying successional forests using Landsat spectral properties and ecological characteristics in

eastern Amazônia. Remote Sensing of Environment, 87, 470-481.

Vitousek, P.M., Mooney, H.A., Lubchenco, J., & Melillo, J.M. (1997). Human domination of Earth’s

ecosystems. Science, 277, 494-499.

Walsh, T.A., & Burk, T.E. (1993). Calibration of satellite classifications of land area. Remote Sensing

of Environment, 46, 281-290.

35

Table 1

Distribution per land cover class of the training dataset collected over SPOT-4 VGT data.

Land cover class # pixels # sample

polygons

Primary tropical forest 3578 902

Cerrado savanna 886 222

Agriculture/pasture 3337 908

Natural/artificial waterbodies 351 119

Secondary succession forest 234 113

Total 8386 2264

36

Table 2

Landsat scenes used to select training data over the SPOT-4 VGT monthly composites of 2000 (1 * INPE, ‡

GLCF; 2 PTF – primary tropical forest, TF – transition forest, CER – cerrado savanna; 3 These statistics does not

include cerrado savanna deforestation, na – not available).

Landsat path/row

(WRS 2)

Location (state)

Date(s) 1 (dd/mm/yyyy)

1st 2nd

Main

vegetation

types 2

(IBGE, 1988)

Scene

deforestation

by 2000 (%)

(INPE) 3

002/067 Acre 06/06/1993 ‡, 28/08/2000 * PTF 23.22

002/062 Amazonas 28/08/2000 * PTF 0.63

005/065 Amazonas 02/09/2000 * PTF 5.38

230/062 Amazonas 13/11/2000 * PTF 10.76

231/062 Amazonas 17/07/1992 ‡, 10/07/2001 ‡ PTF na

232/065 Amazonas 22/07/2000 * PTF 1.83

221/066 Maranhão/Tocantins 25/07/2000 * CER 0.05

224/067 Mato Grosso 01/08/1992 ‡, 12/06/2000 * PTF and TF 34.73

224/068 Mato Grosso 31/08/2000 * TF and CER 25.09

226/069 Mato Grosso 11/07/1988 ‡, 13/08/2000 * TF 23.08

227/067 Mato Grosso 06/08/1992 ‡, 17/06/2000 * PTF and TF 44.85

228/069 Mato Grosso 26/07/2000 * TF and CER 4.39

223/062 Pará 27/07/1984 ‡, 03/08/2001 ‡ PTF 36.63

224/063 Pará 14/08/1988 ‡, 09/07/2001 ‡ PTF 23.43

227/062 Pará 14/08/1986 ‡, 20/08/2000 * PTF 12.41

227/063 Pará 13/07/1986 ‡, 20/08/2000 * PTF 9.64

231/068 Rondônia 16/08/2000 * PTF 51.56

232/058 Roraima 13/02/2000 * CER and TF 8.28

222/067 Tocantins 18/09/2000 * CER 0.39

37

Table 3

Accuracy assessment of the four algorithms tested, using a 10-fold cross-validation approach (QDA – quadratic

discriminant analysis; SCT – simple classification trees; PBCT – probability-bagging classification trees; K-NN –

k-nearest neighbors). The land cover classes are primary tropical forest (PTF), cerrado savanna (CER),

agriculture/pasture (AP), natural/artificial waterbodies (NAW), and secondary succession forest (SSF).

Commission error Omission error Overall accuracy

Classes QDA SCT PBCT K-NN QDA SCT PBCT K-NN QDA SCT PBCT K-NN

PTF 0.091 0.091 0.070 0.076 0.042 0.045 0.034 0.026CER 0.247 0.184 0.190 0.166 0.208 0.245 0.219 0.187AP 0.068 0.083 0.057 0.050 0.091 0.065 0.061 0.064NAW 0.000 0.003 0.003 0.000 0.000 0.003 0.000 0.000SSF 0.663 0.593 0.432 0.468 0.880 0.906 0.662 0.752

0.899 0.903 0.920 0.923

38

Table 4

Confusion matrix for the PBCT algorithm, using the 10-fold cross validation approach.

Predicted class (# pixels) Observed class (# pixels) (1) (2) (3) (4) (5)

Total

Omission error

Primary tropical forest (1) 3457 60 37 0 24 3578 0.034

Cerrado savanna(2) 94 692 99 1 0 886 0.219

Agriculture/pasture (3) 70 98 3133 0 36 3337 0.061

Natural/artificial waterbodies (4) 0 0 0 351 0 351 0.000

Secondary succession forest (5) 97 4 54 0 79 234 0.662

Total 3718 854 3323 352 139 8386

Commission error

0.070 0.190 0.057 0.003 0.432

39

Table 5

Overall class extent generated by the four tested algorithms, using the inverse method calibration technique; QDA

– quadratic discriminant analysis, SCT – simple classification tree, PBCT – probability-bagging classification tree,

K-NN – k-nearest neighbors.

Class overall extent (x 103 km2)

Classes QDA SCT PBCT K-NN

Primary tropical forest 3,054 3,262 3,293 3,283

Cerrado savanna 610 642 528 539

Agriculture/pasture 1,041 831 966 956

Natural/artificial waterbodies 63 75 74 67

Secondary succession forest 233 191 140 156

Total 5,001 5,001 5,001 5,001

40

Table 6

Distribution of land cover classes (x 103 km2) obtained from the PBCT algorithm after calibration (inverse method),

per state of the BLA, for the year 2000. Total/state land cover class percentage (in parentheses), and mean class

membership probability (below) are also indicated (1 the states of Maranhão and Roraima are only included west

of 45º W and south of 5º N, respectively).

Land cover classes

State 1 Primary tropical forest

Cerrado savanna

Agriculture/pasture

Natural/artificial waterbodies

Secondary succession

forest Total

Acre 132 (83.1)

0.92

4 (2.5)

0.57

18 (11.3)

0.87

0 (0.0)

-

5 (3.1)

0.52

159

0.91

Amapá 117 (81.8)

0.92

6 (4.2)

0.60

14 (9.8)

0.73

2 (1.4)

0.98

4 (2.8)

0.54

143

0.90

Amazonas 1432 (89.4)

0.95

50 (3.1)

0.67

49 (3.1)

0.70

29 (1.8)

0.98

42 (2.6)

0.53

1602

0.94

Maranhão 25 (12.5)

0.78

43 (21.5)

0.79

122 (61.0)

0.85

1 (0.5)

0.97

9 (4.5)

0.64

200

0.82

Mato Grosso 362 (40.0)

0.92

188 (20.8)

0.79

334 (36.9)

0.90

4 (0.4)

0.98

17 (1.9)

0.51

905

0.88

Pará 899 (72.0)

0.93

50 (4.0)

0.70

219 (17.5)

0.82

32 (2.6)

0.99

49 (3.9)

0.61

1249

0.90

Rondônia 146 (60.9)

0.92

19 (7.9)

0.74

68 (28.3)

0.88

1 (0.4)

0.96

6 (2.5)

0.50

240

0.90

Roraima 153 (68.3)

0.88

30 (13.4)

0.83

32 (14.3)

0.70

4 (1.8)

0.96

5 (2.2)

0.49

224

0.86

Tocantins 27 (9.7)

0.74

138 (49.4)

0.83

110 (39.4)

0.83

1 (0.4)

0.96

3 (1.1)

0.50

279

0.82

Total 3293 (65.8)

0.93

528 (10.6)

0.80

966 (19.3)

0.85

74 (1.5)

0.98

140 (2.8)

0.60

5001

0.90

41

Figure captions

Fig. 1. State boundaries of the politically defined BLA. It is also shown the location of the 19 Landsat

scenes used for collecting training data.

Fig. 2. (A) Land cover map of the BLA for the year 2000, derived from the PBCT algorithm. The class

color intensity is proportional to the class membership probability (CMP). Areas of the BLA not

included in this study are shown by the grey regions between 45º W and 44º W, and above 5º N (PTF

– primary tropical forest, CS – cerrado savanna, AP – agriculture/pasture, NAW – natural/artificial

waterbodies, SSF – secondary succession forest). (B) Detailed area over the state of Pará.

Fig. 3. Distribution of spectral signatures on the plane of the two main discriminant axes. We have

randomly selected 20 training pixels of each class; the natural/artificial waterbodies class is not shown

as it is well separated from the other classes. (1 – primary tropical forest; 2 – cerrado savanna; 3 –

agriculture/pasture; 5 – secondary succession forest).

Fig. 4. Deforestation by 2000 (INPE, 2002), and combined agriculture/pasture and secondary

succession forest in 2000 from this study, per state of the BLA. (1 the states of Maranhão and Roraima

are only included west of 45º W and south of 5º N, respectively).

Fig. 5. State and overall distribution of agriculture/pasture per vegetation type, using an aggregated

four-class vegetation map, derived from an original 59-class vegetation map of Brazil (IBGE, 1988). (1

the states of Maranhão and Roraima are only included west of 45º W and south of 5º N, respectively).

42

Fig. 6. State and overall distribution of secondary succession forest per vegetation type, using an

aggregated four-class vegetation map, derived from an original 59-class vegetation map of Brazil

(IBGE, 1988). (1 the states of Maranhão and Roraima are only included west of 45º W and south of 5º

N, respectively).

Fig. 7. Comparison between class-specific mean class membership probability and commission error.

The land cover classes are primary tropical forest (PTF), cerrado savanna (CER), agriculture/pasture

(AP), natural/artificial waterbodies (NAW), and secondary succession forest (SSF). The values in

parentheses represent the mean class membership probability and the respective standard error. It is

also represented the perfect agreement y=1-x line.

Fig. 8. Comparison of overall estimates of class extent in the BLA, derived from Hansen et al. (2000),

Eva et al. (2004), this study, and Friedl et al. (2002).

Fig.1

70°W

70°W

65°W

65°W

60°W

60°W

55°W

55°W

50°W

50°W

45°W

45°W

15°S 15°S

10°S 10°S

5°S 5°S

0° 0°

5°N 5°N

km0 250 500

km0 20 40

PTF CS AP NAW SSF< 0.60.6 - 0.8> 0.8

CMP

A

B

Fig.2

Fig.3Click here to download high resolution image

0

50

100

150

200

250

300

350

400

Acre Amapá Amazonas Maranhão Mato

Grosso

Pará Rondônia Roraima Tocantins

BLA states

Co

mb

ine

da

gricu

ltu

re/p

astu

rea

nd

se

co

nd

ary

su

cce

ssio

nfo

rest

(th

isstu

dy)

;d

efo

reste

da

rea

(IN

PE

,2

00

2)

(x1

03

km

2) this study

INPE (2002)

1

Fig.4

0.0

0.2

0.4

0.6

0.8

1.0

Acre

Am

apá

Am

azonas

Mara

nhão

Ma

toG

rosso

Pará

Rondônia

Rora

ima

Tocantins

overa

ll

BLA states/overall

Pro

port

ion

of

agriculture

/pastu

reper

vegeta

tion

type

Other vegetation types

Cerrado savanna

Transition forest

Primary tropical forest

1

Fig.5

0.0

0.2

0.4

0.6

0.8

1.0

Acre

Amapá

Amazonas

Maranhão

Mato Grosso

Pará

Rondônia

Roraima

Tocantins

overall

BLA

sta

tes/o

vera

ll

Proportion of secondary succession forest per vegetation type

Oth

er

vegeta

tion

types

Cerra

do

savanna

Tra

nsitio

nfo

rest

Prim

ary

tropic

alfo

rest

1

Fig.6

PTF (0.93, 0.12)

CER (0.80, 0.17)

AP (0.85, 0.18)

NAW (0.98, 0.07)

SSF (0.60, 0.14)

0.5

0.6

0.7

0.8

0.9

1.0

0.0 0.1 0.2 0.3 0.4 0.5

Commission error

Me

an

pro

ba

bili

tyo

fcla

ss

me

mb

ers

hip

Fig.7

0

10

20

30

40

50

60

70

80

Primary tropical forest +

Secondary succession

forest

Cerrado savanna Agriculture/pasture Natural/artificial

waterbodies

Land cover classes

Exte

nt

(%)

Hansen et al. (2000), 1992-1993 NOAA AVHRR

Eva et al. (2004), 2000 SPOT-4 VGT

this study

Friedl et al. (2002), 2001 Terra/Aqua MODIS

Fig.8