. rimswm

533m.

3

...:‘ll

. . 2.. r.
. , . 54% ’1 rah. I...»
Vfififlflvwgfi.‘?? Wm. LARK. A. I .

13.1,: 1‘. .

. pn {If

grumvr‘ 9.“ .. , . ”usury.

. v.‘

. 2...? u .fixfiztd. fithWJnoéJr 1-...

 

 

 

. 3
21““;
CD
.23
4.: >~.
Era-*5
< .. L
c: "D
{11.93
:41; C
.93
E

 

 

This is to certify that the
dissertation entitled

SPATIOTEMPORAL DYNAMICS OF FOREST
DEGRADATION BY SELECTIVE LOGGING AND FOREST
FIRE IN THE BRAZILIAN AMAZON

presented by

ERALDO A. T. MATRICARDI

has been accepted towards fulfillment
of the requirements for the

degree In Geoggphy

 

mwg (LL

Major Professor’s Signature

v

'27 (JWL‘ Zoo?

 

Date

MSU is an affinnative-action, equal-opportunity employer

 

 

PLACE IN RETURN BOX to remove this checkout from your record.
TO AVOID FINES return on or before date due.
MAY BE RECALLED with earlier due date if requested.

 

DATE DUE

DATE DUE

DATE DUE

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

6/07 p:/CIRC/DateDue.indd-p.1

 

 

SPATIOTEMPORAL DYNAMICS OF FOREST
DEGRADATION BY SELECTIVE LOGGING AND FOREST

FIRE IN THE BRAZILIAN AMAZON
By

Eraldo A.T. Matricardi

A DISSERTATION

Submitted to
Michigan State University
in partial fulfillment of the requirements
for the degree of

DOCTOR OF PHILOSOPHY
Department of Geography

2007

 

ABSTRACT
SPATIOTEMPORAL DYNAMICS OF FOREST DEGRADATION BY SELECTIVE
LOGGING AND FOREST FIRE IN THE BRAZILIAN AMAZON
By
Eraldo A T Matricardi

Selective logging and forest fires have increased at a rapid pace in tropical
regions in recent decades. Forest disturbances caused by selective logging and
forest fires may vary in scale, ranging from local damage to forest canopy,
habitats, soils, and biodiversity, to global changes caused by logging-related
release of carbon into the atmosphere. This study provides a regional
assessment of forest impacts by selective logging and forest fires for 1992, 1996,
and 1999. Multivariate statistical models, remote sensing approaches,
Geographic Information System (GIS), and remotely sensed imagery combined
with field data were applied to verify the scale of environmental changes
associated with these processes of forest disturbance. In this regard, the study
widens the current knowledge on land use and land cover classifications to
include selectively logged and burned forests as additional thematic classes.
These classes have not yet been properly accounted for by conventional remote
sensing approaches of deforestation assessment, despite their relevance for the
understanding of the changes affecting tropical forests. This study is the first
multi-temporal and spatial assessment of the selective logging and forest fire
impacts in the Brazilian Amazon. The resulting estimates show that at least

11800 kmz, 16500 kmz, and 35600 km2 of natural forests were selectively logged

and/or burned by 1992, 1996, and 1999, respectively. More than 60% of these
forest disturbances observed in the Brazilian Amazon during those years were
due to selective logging activities. However, forest fires were responsible for the
greatest impacts on natural forests, causing an estimated loss of 18.8% of forest
canopy in the study region. I also estimated that approximately 5467 kmz, 7618
kmz, and 17437 km2 were active areas of selective logging and/or forest fires in
1992, 1996, and 1999, respectively. In addition, approximately 4% of total forest
disturbed by selective logging and forest fires was geographically located within
protected areas. Areas affected by selective logging and forest fire
corresponded to 2.7%, 3.2%, and 6.3% of total deforestation in the Brazilian
Amazon by 1992, 1996, and 1999, respectively. Altogether, the present study
demonstrated not only the importance of the selective logging and forest fires as
important drivers of forest change in the tropics, but improved the existing
knowledge of their combined impacts in forested lands in the Brazilian Amazon.
Furthermore, the results of this research are expected to support and enhance
the scope of global climate change studies associated with environmental
changes caused by economic activities based on the exploitation of natural
resources available in tropical forests. In terms of its applications to normative
efforts, this study could be used to enlighten authorities and staff of
environmental agencies working to develop sustainable management programs

and environmental policies in Brazil and other tropical countries.

DEDICATION

Dedico esta tese a minha esposa, Cleusa, pelo seu amor incondicional,
companherismo e dedicagao durante todos estes anos em terras
estrangeiras, as minhas filhas, Camila e Marcela, pelo estimulo de
fazer e viver sempre mais, aos meus pais, Mauro e Adelina, pela
referéncia de vida pessoal, trabalho incansavel, amor e fé, aos meus
irma’ios, Wagner, Hamilton e Mauro L0cio, por toda amizade, suporte e
apoio incondicional, a Dona Marli, pelo amor, carinho e compreensao,
ao Senhor Pedro Inocéncio (in memoriam), pela bengao e luz sobre
toda minha familia, aos meus inumeros amigos no Brasil, Estados
Unidos, india, Filipinas, Tailéndia, Dinamarca, Japao, etc., que me
apoiaram e suportaram pacientemente durante meus estudos. N50
teria realizado todo este trabalho sem estas pessoas fazendo parte na
minha vida.

iv

ACKNOWLEDMENTS

It is very difficult to acknowledge all those who directly or indirectly
supported and helped me during this academic journey. And to begin, I have no
words to express my gratitude to Dr. Marcos Pedlowski for trusting, encouraging,
and supporting me since the beginning of my PhD program. Thank you very
much Dr. Pedlowski. I also would like to thank Dr. David Skole, my faculty
advisor, for this opportunity to earn a PhD degree in the USA and for his
guidance, patience, and financial support. Many thanks to Drs. Jiaguo Qi, Ashton
Shortridge, and Runsheng Yin, my committee members, for the helpful lectures
and advices during my PhD program and this manuscript preparation. I also
thank very much the CAPES Fellowship program and the Rondonia
Environmental State Secretariat for this opportunity and financial support in
pursuing my PhD program at Michigan State University. I thank Dr. Mark
Cochrane and Narendran Kodandapani for their help and support to accomplish
my field studies in the Brazilian Amazon. Special thanks are given to Walter
Chomentowski and Oscar Castafieda for their friendship and support in data
processing and analysis. I would like to thank Stephen Cameron for his
friendship and help during my PhD program. I would also recognize the help
provided by my brothers, Wagner, Hamilton, and Mauro Lucio Matricardi, and
friends Wilson Soares Abdala, Vilmar Ferreira, Daniel Gomes de Oliveira, during
the fieldwork in the States of Rond6nia, Mato Grosso, and Acre. I would like to

deeply recognize and thank many mateinos and picadeiros that worked very hard

and helped me to successfully conduct field measurements in the Brazilian
Amazon. I thank Dr. Carlos Castro, Luiz Claudio Fernandes, Antonio Lisboa, and
Israel Xavier Baptista for providing data and comments on my research. I thank
very much Deana Hanner and Jay Sammek for their friendship and support at
Global Observatory for Ecosystem Services. I would also like to recognize the
help and support in my academic pursuits by the Department of Geography,
especially Dr. Antoinette Winklerprins and Dr. Richard Groop. I am especially
thankful for the expertise and contribution from Drs. Asthon Shortridge, Jiaguo
Qi, and Eugenio Arima, during the coursework and this manuscript preparation.
Many thanks are given to Sharon Ruggles, Judy Reginek, and Marilyn Bria for
their academic and administrative support at Department of Geography. I also
would like to express my gratitude to Beth Weisenborn and Juliegh Bookout for
the help, expertise, and good time shared during Virtual Courses. Last but not
least, many thanks to my colleagues, friends, and classmates from the Global
Observatory for Ecosystem Services and Department of Geography for helping

me and sharing good moments during this period of time in the USA.

vi

TABLE OF CONTENTS

LIST OF FIGURES ................................................................................ x
LIST OF TABLES ....................................................... . .......................... xii
LIST OF APPENDICES ......................................................................... xiv
CHAPTER I
Problem Statement ............................................................................................... 1
1.1 . Statement of purpose ..................................................................................... 2
1.2. Hypothesis testing .......................................................................................... 2
1.3. Background .................................................................................................... 7
1.3.1. Forest degradation concepts ................................................................ 9
1.3.2. Forest fires ......................................................................................... 11
1.3.3. Defining selective logging .................................................................. 12
1.3.3.1. Forest impacts by selective logging ............................................. 14
1.3.3.2. Impacts on carbon cycle .............................................................. 15
1.3.3.3. Impacts on biodiversity ................................................................ 16
1.3.3.4. Socioeconomic impacts ............................................................... 17
1.4. General comments ....................................................................................... 20
1.5. Overview of the study ................................................................................... 21

CHAPTER II: Methods for assessment of impacts by selective logging and forest
fire: A case study in Mato Grosso, Brazil

2.1. Abstract research ......................................................................................... 23
2.2. Introduction .................................................................................................. 24
2.3. Methodology ................................................................................................. 27
2.3.1. Study site ........................................................................................... 27
2.3.2. Dataset ............................................................................................... 28
2.3.2.1. Land use and land cover dataset ................................................ 28
2.3.2.3. Selective logging dataset ............................................................. 29
2.3.2.4. Satellite imagery .......................................................................... 30
2.3.2.4.1. Radiometric and atmospheric corrections of Landsat imagery
.............................................................................................................. 31
2.3.2.4.2. Normalization of seasonality effect of Landsat imagery ........ 33
2.3.2.4.3. Geometric correction ............................................................. 34
2.4. Methods ....................................................................................................... 34
2.4.1. Field measurements of forest canopy ................................................ 34
2.4.2. Remote sensing approaches ............................................................. 36
2.4.2.1. Vegetation lndices ....................................................................... 37
2.4.2.2. Modified Vegetation Index approaches ....................................... 41
2.4.2.3. Spectral Mixture Analysis ............................................................ 42
2.4.2.3.1. Endmember selection ........................................................... 44
2.4.2.4. Fractional cover ........................................................................... 46
2.4.2.5. Burned forest detection ............................................................... 47
2.4.2.5.1. Burned forest detection approach ......................................... 47
2.4.3. Forest canopy degradation assessment ............................................ 49

vii

2.4.3.1. Testing Hypotheses ..................................................................... 50

2.4.3.2. Data sampling ............................................................................. 53

2.5. Results and Discussions .............................................................................. 53
2.5.1. Relationships among various Vegetation lndices ............................... 54
2.5.2. Optimum Vegetation Index ................................................................. 56
2.5.3. Accuracy assessment of burned forest detection technique .............. 58
2.5.4. Multi-annual land use and land cover change .................................... 59
2.5.5. Assessment of forest canopy cover impacts ...................................... 63

2.6. Conclusions .................................................................................................. 67

CHAPTER III: Basin-wide assessment of forest disturbances by selective logging
and forest fires

3.1 . Abstract ........................................................................................................ 71
3.2. Introduction .................................................................................................. 73
3.3. Methods ....................................................................................................... 76
3.3.1. Regional setting ................................................................................. 76
3.3.2. Dataset ............................................................................................... 77
3.3.2.1. Deforestation dataset .................................................................. 77
3.3.2.2. Selective logging dataset ............................................................. 78
3.3.2.3. Landsat imagery .......................................................................... 80
3.3.2.3.1. Radiometric correction .......................................................... 80
3.3.2.3.2. Landsat imagery normalization ............................................. 80
3.3.2.3.3. Geometric correction ............................................................. 82

3.3.3. Field study .......................................................................................... 82
3.3.3.1. Field sampling and measurement ................................................ 83
3.3.3.2. Identification of species ............................................................... 84
3.3.3.3. Data processing for biodiversity .................................................. 85
3.3.3.4. Field validation for remote sensing approaches .......................... 86

3.3.4. Remote sensing approaches ............................................................. 87
3.3.4.1. Burned forests detection .............................................................. 87
3.3.4.2. Fractional coverage ..................................................................... 88

3.3.5. Forest canopy impact assessment ..................................................... 89
3.3.5.1. Testing l-Iypotheses ..................................................................... 90
3.3.5.2. Data Sampling ............................................................................. 92

3.4. Results and Discussion ................................................................................ 93
3.4.1. Multi-temporal assessment of selective logging and forest fires ........ 93
3.4.2. Optimum vegetation index selection .................................................. 96
3.4.3. Forest canopy impact assessment based on remotely sensed data..97
3.4.4. Field measurements ........................................................................... 99
3.4.5. Links between forest canopy and further ecological indicators ........ 103

3.5. Conclusions ................................................................................................ 105

viii

CHAPTER IV: Spatially explicit probabilistic models of forest fire and selective
logging in the Brazilian Amazon

4.1. Abstract ...................................................................................................... 11 1
4.2. Introduction ................................................................................................ 112
4.3. Methods ..................................................................................................... 114
4.3.1. The setting ....................................................................................... 114
4.3.2. Datasets ........................................................................................... 114
4.3.2.1. Deforestation dataset ................................................................ 114
4.3.2.2. Selective logging dataset ........................................................... 1 15
4.3.2.3. Thematic maps .......................................................................... 116
4.3.2.4. Timber volume in the Brazilian Amazon .................................... 118
4.3.3. Forest Fragmentation ....................................................................... 120
4.3.4. Modeling forest fie and selective logging probabilities in the Amazon
................................................................................................................... 121
4.3.4.1. Forest fire conceptual model ..................................................... 122
4.3.4.2. Selective logging conceptual model .......................................... 127
4.3.4.3. Data sampling ........................................................................... 130
4.4. Results and Discussion .............................................................................. 132
4.4.1. Forest fire probabilistic model .......................................................... 132
4.4.2. Selective logging model ................................................................... 139
4.5. Conclusions ................................................................................................ 145
CHAPTER V: Concluding Remarks
5.1. Seating the current study in the context of global change research ........... 150
5.2. Research questions revisited ..................................................................... 154
5.3. Hypotheses revisited .................................................................................. 160
5.4. Opportunities for further studies ................................................................. 162
6. References .................................................................................................... 246

ix

Figure 1.1.

Figure 2.1.

Figure 2.2.
Figure 2.3.

Figure 2.4.

Figure 2.5.

Figure 3.1.

Figure 3.2.

LIST OF FIGURES

Forest production (round wood) from 1990 to 2003 in the Brazilian
Amazon (adapted from IBGE, 2006) ................................................ 19
Stepwise procedure for radiometric and geometric corrections of
Landsat imagery ............................................................................... 33
Procedures for calculating fractional cover ....................................... 46

Flow diagram for burned forest automated detection technique ....... 48

Land use and land cover change from 1992 to 2004 in the study site
......................................................................................................... 61

Partial effects of various forest uses on fractional coverage values in
the study area estimated using linear multi-regression model and
multi-annual remotely sensed derived dataset ................................. 66

Partial effects of various anthropogenic activities on forest canopy
cover in the Brazilian Amazon using linear multiple-regression model
and multi-temporal remotely sensed data ........................................ 98

Number of tree species, Simpson’s index of diversity, and Simpson’s
reciprocal index against number of sampling (transect) ................. 101

Figure 4.1 Stratified random sampling used to collect spatial data for the probit

Figure 4.2.
Figure 4.3.

Figure 4.4.

Figure 4.5.

Figure 4.6.

Figure 4.7.

model of forest fire and selective logging in the study region. ........ 131
Probability of forest fire with respect to distance to deforestation ...134
Probability of forest fire with respect to distance to roads ............... 135
Probability of forest fire with respect to water deficit from May to
September ...................................................................................... 136
Probability of forest fire with respect to forest patch size and shape1 37
Probability of forest fire with respect to forest types, states, selective
logging, protected areas, and year of analysis ............................... 139
Probability of selective logging with respect to distance to roads and

timber centers ................................................................................. 141

Figure 4.8. Probability of selective logging with respect to timber volume ........ 143
Figure 4.9. Probability of selective logging with respect to states, vegetation
types, protected areas, and year of analysis .................................. 144

"Images in this Dissertation are presented in color”.

xi

LIST OF TABLES

Table 2.1. Land Use and Land Cover dataset characteristics ............................. 29
Table 2.2. Selective Logging dataset characteristics .......................................... 30

Table 2.3. Correlation matrix among various vegetation indices under smoke-free
atmospheric condition for the study site .............................................. 55

Table 2.4. Correlation matrix among various vegetation indices for areas in the
presence of heavy smoke in the study site. ........................................ 56

Table 2.5. Empirical linear relationship between various vegetation indices and
forest canopy coverage based on field measurements in the study site
............................................................................................................ 57

Table 2.6. Accuracy assessment results of detecting burned forests using non-
photosynthetic vegetation (NPV) fraction image derived from Spectral
Mixture Analysis (SMA) for the study area in 2004. ............................ 59

Table 2.7. Cumulative deforestation and undisturbed and disturbed forests in the
study area from 1992 to 2004. ............................................................ 60

Table 2.8. Annual increase of selectively logged forest, burned and logged
forests combined, and burned forest only, from 1992 to 2004 in the
study area. .......................................................................................... 62

Table 2.9. Multi-annual land use dynamics in the study site from 1992 to 2004.63

Table 3.1 Deforestation dataset characteristics .................................................. 78

Table 3.2 Selective Logging dataset characteristics ........................................... 79

Table 3.3 Detected selective logging, burned forests, and dataset characteristics
............................................................................................................ 93

Table 3.4.Total area of selectively logged and burned forests detected in 1992,
1996, and 1999 within protected areas (indigenous land, conservation
unit, and military area) ......................................................................... 95

Table 3.5 Correlation matrix among various forest parameters measured in the

study sites in the State of Acre, Rondonia, and Mato Grosso. .......... 104
Table 3.6. Estimated increase of new selectively logged, forests burned, and
selectively logged and burned forests in the Brazilian Amazon. ....... 109

xii

Table 4.1 Deforestation dataset characteristics ................................................ 115

Table 4.2 Selective Logging dataset characteristics ......................................... 116
Table 4.3. Description of the thematic maps ..................................................... 117
Table 4.4. Probit regression analysis results of forest fire ................................. 133
Table 4.5. Probit regression analysis results of selective logging ..................... 140

xiii

LIST OF APPENDICES
Appendix A.1. Total deforestation in the Brazilian Amazon from 1978 to 2004. 165

Appendix 8.1. Study site location (Landsat path 226 row 068) in the Amazon
State of Mato Grosso, Brazil. .......................................................... 167

Appendix 8.2. Satellite imagery used in the case study in the Amazon State of
Mato Grosso, Brazil. Landsat Path 226 and Row 068 ................ 169

Appendix 8.3. Slope and intercept for atmospheric correction of the Landsat
imagery using as 58 model. ............................................................ 171

Appendix 8.4. Slope and intercept for normalization of the Landsat imagery
using as a reference the path 226 row 068, acquired in August 8,
2001 ..................................................................................................... 174

Appendix 8.5. Fieldwork locations within the study area ...................................... 177

Appendix 8.6. Canopy openness estimation using gap light analyzer and
hemispherical photos from forest canopy cover ........................... 179

Appendix 8.7. Visible and SWIR (at 2.1 pm) reflectance, and MSAVI and
MSAVIaf retrieved from Landsat image acquired in 2003. .......... 181

Appendix 8.8. Burned forests detection based on non-photosynthetic vegetation
fraction image .................................................................................... 183

Appendix 8.9. Cumulative deforestation and persistence of selective logging and
forest fire on remotely sensed data ................................................ 185

Appendix 8.10. Forest canopy losses estimated using multi-regression model for
the study area .................................................................................... 187

Appendix C.1. Study region location and Landsat scenes used in the
spatiotemporal assessment of selective logging and forest fire in

the Legal Amazon ............................................................................. 189
Appendix 02. Acquisition dates of the Landsat imagery for 1992, 1996, and

1 999 ..................................................................................................... 191
Appendix C.3. Case study (fieldwork) locations in the Brazilian Amazon ........... 194
Appendix 04 Case study and transect locations and characteristics ................ 196
Appendix C.5. Transect design for fieldwork measurements ................................ 198

xiv

Appendix C.6. List of tree species identified and recorded during the field work in
the States of Acre, Rond6nia, and Mato Grosso. ........................ 200

Appendix C.7. Estimates of indicators of forest disturbances by selective logging
and forest fires based on field measurements in the States of
Acre, Rond6nia, and Mato Grosso ................................................. 204

Appendix 08 Empirical linear relationship between various vegetation indices
and forest canopy coverage based on field measurements for the
study sites in Mato Grosso, Rondonia, and Acre. ........................ 207

Appendix C.9. Basin-wide results of selective logging and burned forest detection
for 1992, 1996, and 1999 ................................................................. 209

Appendix C.10. Selectively logged and burned forests detected by State in the
Brazilian Amazon .............................................................................. 212

Appendix C.11. Forest canopy losses estimated using multi-regression model for
the study region ................................................................................. 214

Appendix C.12. Spatial distribution and intensity (logged area /625 km2) of
selectively logged forests detected in 1992 .................................. 216

Appendix C.13. Spatial distribution and intensity (logged area / 625 km2) of
selectively logged forests detected in 1996 .................................. 218

Appendix C.14. Spatial distribution and intensity (logged area /625 km2) of
selectively logged forests detected in 1999 .................................. 220

Appendix C.15. Spatial distribution and intensity (burned area I 625 km2) of
burned forests detected in 1992 ..................................................... 222

Appendix C.16. Spatial distribution and intensity (burned area / 625 km2) of
burned forests detected in 1996 ..................................................... 224

Appendix C.17. Spatial distribution and intensity (burned area / 625 km2) of
burned forests detected in 1999 ..................................................... 226

Appendix C.18. Fractional coverage derived from Landsat imagery acquired in
1992 for the study region ................................................................. 228

Appendix C.19. Fractional coverage derived from Landsat imagery acquired in
1996 for the study region ................................................................. 230

Appendix C.20. Fractional coverage derived from Landsat imagery acquired in
1999 for the study region ................................................................. 232

XV

Appendix D.1.

Appendix 02.

Appendix D.3.

Appendix 04.

Appendix D.5.

Appendix D6.

The study region location within the Brazilian Amazon ............... 234

Deforestation map for the Brazilian Amazon in 1992, 1996, and
1999 ..................................................................................................... 236

Vegetation map of the Brazilian Amazon in 1999. ...................... 238

Map of road network, protected areas, and timber centers in the
Amazon ............................................................................................... 240

Map of timber volume generated using ordinary kriging
interpolation and the dataset derived from forest inventory

conducted by RADAMBRASIL project ........................................... 242
Water deficit (mm) for the Brazilian Amazon during the dry
season (May to September) in 1999. ............................................. 244

xvi

CHAPTER I
Problem Statement

The highest rates of tropical deforestation (conversion of natural forests to
non-forest uses) have been observed in the Brazilian Amazon in recent decades.
In addition to outright deforestation, forest degradation by selective logging and
forest fire has become an important issue due to its potential impacts on global
environmental processes, natural ecosystems, biodiversity, and economic
development. As a result, several studies have been conducted to assess
impacts caused by selective logging and forest fire on tropical ecosystems.
These studies have provided important direct and indirect estimates of the
impacts brought by selective logging and forest fires for particular study sites in
which forests have been partially damaged and not removed as in cases of
deforestation. However, only a few attempts have been made to measure both
the extent of selective logging and forest fires and their impacts on tropical
forests using remotely sensed data gathered for the entire Brazilian Amazon.
Therefore, relevant questions have yet to be addressed to provide a better
understanding of selective logging and forest fire dynamics, their impacts on
tropical forests, and their interactions with other land use types. For example: ls
selective logging and forest fire causing more forest degradation than otherwise
would occur? Can we accurately measure these anthropogenic activities in the
Brazilian Amazon basin? Where are the most severe selective logging and forest
fire events occurring in that region? Why are selective logging and forest fires

occurring there and not somewhere else? Is there a particular spatial pattern to

these activities in that region? If so, are those patterns changing over time? Is
there an interaction between selective logging and fire that may be synergistically
causing even more forest disturbances?
1 .1. Statement of purpose

The purpose of this study is to investigate the dynamics of selective
logging and forest fire processes to better understand the factors that might
control these processes at various scales and times, their impacts on natural
forests, and their interactions with other factors that may contribute to increase
forest degradation in the Brazilian Amazon basin, using field and remotely
sensed data, geographic information system, remote sensing techniques, and
spatial analysis.
1.2. Hypothesis testing
Hypothesis 1: Forest imficts by selective loggiggjctivities incremver time
and severely disturb natu_ral forests in the Brazilian Amazon.
Natural forests encompass around 3.15 million square kilometers and are the
main source of raw material for timber industries in Brazil (Lentini et al., 2003).
Amazonian timber production averaged 51 million cubic meters annually from
1990 to 1996, decreasing to an annual average of 17 million cubic meters from
1997 to 2003 (IBGE, 2006). Amazonian timber production and its relative
contribution to national markets in Brazil have been quite steady in this period of
time (see Figure 1.1). These numbers indicate that most accessible natural
forests in the Brazilian Amazon have been under increasing pressure by timber

industry in the last decade. Uhl et al. (1997) observed that loggers are advancing

into undisturbed forests and are re-logging previously disturbed forests searching
for second tier economic species. Matricardi et al. (2007) estimated that the total
area of selective logging detected with satellite imagery increased about 436%
between 1992 and 1999 in the Amazon. Based on a case study in the state of
Mato Grosso, Matricardi et al. (2005) estimated that areas of formerly logged
forests that were re-logged increased 1442% between 1993 and 2002.
Consequently, it is expected that forest degradation in the Brazilian Amazon has
also increased as a consequence of the growth in area and intensity of selective
logging activities. If this hypothesis is true, both selective logging impacts on
natural forests and degraded forests have significantly increased over time. For
the present study, this hypothesis was tested at local and regional scales,
focusing on areas where selective logging has been detected using remotely
sensed data. At the local scale, a multi-temporal assessment of selective logging
impacts was conducted in the Southeastern Amazon using 13 years (1992 to
2004) time-series remotely sensed data and field measurements (i.e. forest
regeneration, number of tree species and mortality, and fuel load). This case
study sought to contribute to a better understanding of the annual change in
degradation rates and the capacity for forest regeneration in areas affected by
selectively logging. Additionally, fieldwork activities were carried out to validate
remote sensing techniques in different study sites in the Amazon region. At the
regional scale, fractional coverage was estimated for areas where selective
logging has been detected using Landsat imagery for three different years (1992,

1996, and 1999). These results provided estimates of impacts and temporal

changes in canopy degradation and regeneration of selectively logged forests in
the Brazilian Amazon. Multiple regression analysis was used to statistically infer
the contribution of each land use type on forest disturbances at local and regional
scales. Additional statistical tests were applied to assess forest disturbances
based on field measurements for different land use types.

Hypothesis 2: Forest fires mostly occur on fire-prone selectively logged forests
and further increage forest degrfllation in the Brazilian Amazon.

Selective logging activities can severely impact tropical forests by damaging
individual trees and under-story vegetation, compacting and eroding soils
(Verissimo et al., 1992, Johns et al., 1996), reducing canopy cover (Uhl and
Vieira, 1989, Verissimo et al., 1992), and, potentially, increasing forest fire
susceptibility (Holdsworth and Uhl, 1997), especially during years of severe
droughts (Uhl and Kauffman, 1990). Uhl and Buschbacher (1985) were the first
to observe a synergistic interaction between selective logging and fire-maintained
pastures where fire spreads into logged forests. It was then hypothesized that
most forest fires were occurring on previously logged forests, and forest fires
were spatially associated with logging and deforestation areas; forest
degradation would be significantly higher when selective logging was combined
with forest fire. This hypothesis was tested at local and regional scales. At the
local scale, a study site in Mato Grosso where fire hotspots have been more
frequently observed using NOAA and MODIS imagery (INPE, 2007) was studied.
First, burned forests were detected using a Landsat time-series imagery (1992 to

2004) and a new remote sensing method based in part on Cochrane and Souza

(1998) was applied. This method involved masking out non-forest areas and
applying linear mixture model classification to detect burned forests. Linkage
between selective logging and forest fires was demonstrated by juxtaposing
selectively logged and burned forests. Moreover, fractional coverage was
calculated for both selectively logged only and selectively logged/burned forests.
This procedure helped to estimate forest canopy losses and time required for
forest regeneration under different land use characteristics (i.e. burned and non-
burned forests). Temporal changes were assessed by comparing data from 1992
to 2004. Finally, field measurements (i.e. forest regeneration, number of tree
species, trees mortality, and fuel load) were also conducted on disturbed and
undisturbed forest sites to provide additional information on forest degradation.
At the regional scale, burned forests were detected for three years (1992, 1996,
and 1999) using the same methodological approaches applied and validated at
the local scale analysis. Fractional coverage was estimated for areas where
selective logging has been detected using Landsat imagery. These datasets
were mostly derived from remotely sensed data and focused on areas where
evidences of selective logging and forest fire had been previously observed in
the study region.

Hypothesis 3: Variables that reflect landscape and land use chiacten'stics (ea.

 

vegetation type. soil. precipitation. deforestation. forest fragmentation, proximity
to ad network. etc.) explain spatial distribution and probaLiIitv of selective

logging and forest fire to occur in the Brazilian Amazon.

Links between selective logging and various socioeconomic and biophysical
factors have been studied by many authors. (Uhl and Vieira, 1989) affirrned that
access roads left behind by selective logging activities contribute to increase
deforestation in new frontiers. Furthermore, fire-maintained agricultural systems
may spread out into adjacent logged forests (Uhl and Buschbacher, 1985). More
recently, (Cochrane et al., 2004) observed that there are synergistic interactions
between different anthropogenic activities (eg. selective logging and fire,
deforestation and forest fragmentation, etc.) potentially increase forest
degradation. Particularly, this analysis focused on existing variables derived from
remotely sensed data for a sub-region within the Amazon basin where selective
logging and forest fire were more evident on satellite imagery, and excluded
those areas where this type of activity were improbable. Therefore, it was
expected that variables such as forest and soil types, proximity to roads,
precipitation regimes, land tenure (private and public lands), proximity to urban
areas, and proximity to previous deforestation events could play important roles
in explaining the spatial patterns of selective logging and forest fire in the study
area. It was also expected that temporal changes in spatial patterns of selective
logging and forest fire in the study area could be explained by the use of a multi-
temporal dataset (i.e. 1992, 1996, and 1999). Multivariate probit models were
applied to assess which factors were influencing the spatial distribution of
selective logging and forest fire. This analysis, however, required several
attempts and different variables to estimate probability of selective logging and

forest fire in the study region. This procedure was intended to select statistically

significant variables that represented factors influencing selective logging and
forest fire occurrences in the study area. The final statistical models were
expected to significantly explain both the selective logging and forest fire
probabilities throughout the study region.

1.3. Background

The Brazilian Amazon contains 40% of the Earth’s remaining tropical
rainforests and plays an important role in sustaining biodiversity, regional
hydrological regimes, climatic patterns (Fearnside, 1999), and terrestrial carbon
pools (Fearnside, 1999, Houghton et al., 2000).

Consequently, the increasing rates of deforestation in the Brazilian
Amazon has become an environmental concern of global dimension (Laurance et
al., 2001, Nepstad et al., 2002). Multi-annual estimates of deforestation in the
Brazilian Amazon provided by the National Institute for Space Research - INPE,
show that total deforestation was 152,200 kmz, 377,500 sz, and 551,782 km2
by 1978, 1988, and 1998, respectively. Such estimates represent an annual
increment of 19,979 kmz, an increase of 362% since 20 years ago (the world’s
highest absolute rate of deforestation). Most recent measurements based on
remotely sensed data show that 703,198 sz were deforested by 2005
(Appendix A.1), equivalent to 14.06% of the total legal Brazilian Amazon (INPE,
2007).

Satellite-based remote sensing has been used for several decades to
monitor and to evaluate land use and land cover changes in the Amazon,

focusing most on deforestation assessments (Fearnside et al., 1990, Skole and

Tucker, 1993). Data from these studies (eg. rate of deforestation in Brazil’s
Legal Amazon) are often used to estimate human effects on the global carbon
cycle (Fearnside, 1997, Houghton, 1997, Houghton et al. 2000). However, there
is evidence that approximately half of the total area of forest damaged or
impoverished annually by human activities has not been detected under
conventional deforestation classification schemes and, consequently, the total
area of disturbed forests has been greatly underestimated. Most of those non-
detected areas are related to forest degradation caused by selective logging and
forest fires (Cochrane et al., 1999, Nepstad et al., 1999).

Earlier studies (Stone and Lefebvre, 1998, Souza and Barreto, 2000,
Asner et al., 2002) in the Brazilian Amazon state of Para have shown that it is
possible, with some limitations, to detect selective logging using remotely sensed
data. Matricardi (2003) expanded the scope of these studies by identifying and
mapping areas of selective logging in tropical “terra-finne” (high land) forests for
the entire Amazon basin for three different years. Matricardi showed that at least
3,689 sz, 5,107 sz, and 11,638 sz of natural forests had being actively
logged in 1992, 1996, and 1999, respectively.

More recently, additional attempts to automatically detect selective logging
and its impacts on tropical forests using remote sensing techniques have been
conducted in different parts of the Brazilian Amazon. Souza et al. (2003) and
Asner et al. (2004) provided valuable information based on different case studies
in the state of Para. Matricardi et al. (2005) also conducted a multi-temporal

analysis, from 1992 to 2002, in the state of Mato Grosso. Although these studies

have significantly improved detection techniques and current understanding of
selective logging for specific sites, its impacts, spatial patterns, and temporal
dynamics over the entire region have yet to be assessed.

This research sought to enhance the understanding of the spatial patterns
and temporal changes of forest degradation associated with selective logging
and forest fires, the interactions between selective logging, forest fires, and other
land use types, and how forests ecosystems respond to these anthropogenic
activities. A final goal was to provide support for prognostics of selective logging
and forest fire spatial trends throughout the Brazilian Amazon region.

1.3.1. Forest degradation concepts

The term forest degradation has been defined as a ‘temporary or
permanent deterioration in the density or structure of vegetation cover or its
species composition’ (Grainger, 1993). Unlike deforestation, this term has been
used to describe disturbances in forest characteristics that do not involve land
cover conversion into other land cover types (Grainger, 1999, Lambin, 1999) and
that may occur in a short or long-run time-scale (Lambin, 1999). In the Amazon,
anthropogenic forest degradation is often related to forest fragmentation (Skole
and Tucker, 1993, Cochrane, 2001), selective logging (Pinard and Putz, 1996,
Lambin, 1999, Nepstad et al., 1999, Huth and Ditzer, 2001), and forest fires (Uhl
and Buschbacher, 1985, Cochrane and Schulze, 1998, Cochrane et al., 1999).

More specifically, forest fragmentation is normally associated to different
forms of deforestation or clear-cutting, whether involving conversion of forest for

agriculture or infrastructure development, because it increases forest

disturbances along the edges of those remaining forests within deforested areas
(Skole and Tucker, 1993, Lambin and Ehrlich, 1997, Laurance et al., 1997).
Forest disturbances tend to be higher at the edges because of the increased
amounts of sunlight and wind that penetrate the forest, causing, among other
things, increased tree mortality (Laurance et al., 1997) and increased drying
rates of leaves and wood debris on the forest floor (Nascimento and Laurance,
2002). Ultimately, such effects increase forest fire susceptibility. Concomitantly,
frequent exposure of fragmented forest edges to agricultural fires increases the
risk of forest fires (Laurance et al., 1997).

As a result, forest fire is another important proximate cause of forest
degradation. Although tropical forests were until recently deemed as immune to
fire (Uhl and Kauffman, 1990, Cochrane and Schulze, 1998), forest fires have
been increasing as the deforestation and forest fragmentation increase.
Deforested areas display greater risk to fire and often act as ignition sources for
fragmented forest sites that have lost their immunity to fire infiltration (Uhl and
Kauffman, 1990, Cochrane and Schulze, 1998, Cochrane, 2001, Cochrane et al.,
2004).

Selective logging also increases forest degradation. Although this
economic activity involves cutting down only the few most marketable trees in
tropical forests, it can severely impact undisturbed forests during the harvesting
operations (Pinard and Putz, 1996, Uhl et al., 1997, Nepstad et al., 1999, Huth
and Ditzer, 2001). Thus, the combined effects of selective logging, forest

fragmentation, and forest fire can dramatically increase forest degradation rates

10

in tropical regions (Cochrane et al., 2004). A focus on degradation thus expands
the previous discussions on deforestation alone. Further details on selective
logging impacts on natural forests are presented subsequently.

1.3.2. Forest fires

Previous studies had shown that forest fires rarely occur in intact tropical
forests. Fire events should be unexpressive in most of the Brazilian Amazon (Uhl
and Kauffman, 1990, Cochrane and Schulze, 1998) because of moist
microclimates provided by tropical forests that make it difficult for the propagation
of fires, even during dry season periods (Uhl and Kauffman, 1990). In spite of this
expected resistance to fire, the spread of forest fire is a growing concern in many
tropical countries. Uhl and Buschbacher (1985) suggested that there is a direct
relationship between agricultural practices that use fire for land maintenance,
such as clearing woody debris and weed control, and increased forest fire
susceptibility particularly on logged sites close to agriculture. Accordingly,
agricultural plots have been designated as the most common fire triggers in the
Brazilian Amazon (Uhl and Buschbacher, 1985, Uhl and Kauffman, 1990,
Cochrane, 2001).

In fact, there is a growing probability of fires in tropical forest as the forest
become more fragmented by selective logging and the edge formation along
forest and deforested areas (Cochrane, 2001). Forest fires are devastating for
tropical ecosystems because of their natural low resistance to fire, which allows
dominance by thin bark trees (Uhl and Kauffman, 1990). Moreover, once a site is

burned for a first time, increase chances for recurring fire and more intense fires

11

because of increased fuel loads and drier microclimate created by the previous
fire (Cochrane and Schulze, 1998). Therefore, sequential fires can degrade
forests rapidly by drastically destroying their structure and composition and,
ultimately, creating a new land cover type whose structural features resembles a
secondary regrowth and not an actual pristine tropical forest.

1.3.3. Defining selective logging

The term selective logging is used to define the targeted extraction of a
group of tree species in which, usually, the most economically valuable trees are
removed from the forest (Verissimo et al., 1995, Uhl et al., 1997).

While deforestation in tropical areas affects directly the carbon cycle and
causes massive destruction of plants and animals by clear-cutting and burning
natural forests (Skole and Tucker, 1993), selective logging activities do not clear-
cut and burn forests, but harvest only a portion of trees. In spite of being a less
aggressive form of forest intrusion than clear-cutting, selective logging can
damage a large portion of forest and affect many trees during logging operations
(Pinard and Putz, 1996, Nepstad et al., 1999, Huth and Ditzer, 2001). For
instance, when traditional techniques are used, the harvesting of a single tree
alone can directly result in the death of thirteen other trees as result of its fall
direction, and, in addition, ordinary logging operations such trail opening, acess
road construction, and the establishment of log storage patios also contribute to
increase forest degradation (Verissimo et al., 1992).

Meanwhile, selective logging activities have been occurring in Brazil’s

tropical forests for several decades (Uhl and Kauffman, 1990, Stone and

12

Lefebvre, 1998, Nepstad et al., 1999, Alvarado and Sandberg, 2001). These
activities have been traditionally restricted to flood plains (vérzeas) due the pre-
existent access through the rivers (fluvial transportation). However, the
expansion of road networks during 1960s and 19705 in the Brazilian Amazon
allowed the expansion of selective logging to the interfluvial (terra finne) forest
areas (Uhl and Vieira, 1989, Uhl et al., 1997). Moreover, the Avanca Brasil1
program set by the Federal government to strengthen transportation networks in
the Brazilian Amazon as a mean to further accelerate economic development
(eg. roads, timber and mining industry) in the region is expected to have major
impact on the diffusion of logging operations (Laurance et al., 2001, Fearnside,
2002)

Selective logging can be classified in terms of spatial distribution and
severity. Spatially, selective logging can occur in ‘varzeas’ and ‘tene firme’ in the
Brazilian Amazon. Moreover, selective logging varies in terms of impacts (i.e.
low, moderate, and high), depending on the volume and number of harvested
trees (Uhl et al., 1997). Low impact logging (i.e. highly selective logging) involve
the extraction of one or two high valuable tree species, such as Virola (Virola
sun'namensis (Rol.) Warb ) (Uhl et al., 1997) and Mahogany (Swietenia

macrophylla, King) (Verissimo et al., 1995, Uhl et al., 1997). On the other hand,

 

1 “Avanca Brasil" (Fonrvard Brazil) is a development program that is being conducted by the
Brazilian government in the Brazilian Amazon, involving a total of US$ 43 billion over the 2000-
2007 period of which US$20 billion is for infrastructure such as transportation corridor expansion
(roads, water ways, and railways) with direct impacts on the (Laurance et al. 2001, Feamside

2002).

13

high impact logging encompasses a hundred or more tree species, and requires
the use of heavy machinery and the construction of greater infrastructure (i.e.
patios, forest roads, etc.) for tree harvesting and transportation. The later form of
selective logging is the most common form in the more developed parts of the
Brazilian Amazon (Uhl et al., 1997).

Finally, (Uhl et al., 1997) described an additional logging type of logging
known as forest ‘mining’, actually a sequential form of selective logging. Forest
mining can completely impoverish a seemingly intact forest tract over a period up
to thirty year of successive and ever more intensive selective logging activities
that at the end may lead to an acute form of forest degradation or even outright
deforestation.
1.3.3.1. Forest impacts by selective logging

Selective logging activities leave behind a mixture of intact forest, treefall
gaps, roads, log-loading patios, and damaged forests, increasing amounts of
dead slash or dried biomass (fuel) and, consequently, greater risks of forest fire
events (Uhl and Buschbacher, 1985, Stone and Lefebvre, 1998, Nepstad et al.,
1999, Souza and Barreto, 2000). Estimates of the total of damages arising from
logging operations point out that about 40% of the remaining trees in a given
forest patch affected by selective logging may end either killed or severely
damaged (Uhl et al., 1991). Such impacts can be devastating even when a small
volume of timber is harvested (Frumhoff, 1995). The potential impacts of

selective logging in tropical forest ecosystem are further discussed bellow.

14

1.3.3.2. Impacts on carbon cycle

(Skole and Tucker, 1993) have argued that “tropical deforestation is a
major component of the carbon cycle and has profound implications for biological
diversity”. The greatest carbon and biodiversity losses happen when a forested
area of high biomass content is converted into a low biomass system such as
pasture and agriculture, releasing carbon directly to the atmosphere and
destroying natural habitats (Skole and Tucker, 1993).

Nevertheless, Amazonian forests are increasingly being exposed to
logging activities (Stone and Lefebvre, 1998, Nepstad et al., 1999, Souza and
Barreto, 2000). Although the impacts of selective logging are known to vary
according to extraction intensity, their overall impacts are usually substantial on
the complex and fragile tropical environments. Meanwhile, selectively logged
forests are expected to accumulate carbon over time and return to pre-harvest
levels of biomass if left undisturbed. However, many forests are revisited several
times by loggers seeking to harvest additional tree species as regional timber
markets evolve (Verissimo et al., 1995, Uhl et al., 1997). Re-Iogged forests
become highly degraded and may have 40 — 50% of the canopy cover is
destroyed during these re-logging operations (Uhl and Vieira, 1989, Verissimo et
al., 1992). In addition, forest heavily impacted by selective logging are expected
to face an increase in fire susceptibility (Holdsworth and Uhl, 1997) and,
consequently, more emissions of carbon to the atmosphere (Houghton, 1997).
Intentionally or not, uncontrolled forest exploitation by loggers also serve to

catalyze deforestation by opening roads inside idle public lands and protected

15

areas that are often squandered by large ranchers and small farmers. This
process results in increased rates of deforestation and higher carbon emissions
to the atmosphere (Verissimo et al., 1995).

1.3.3.3. Impacts on biodiversity

Upper Amazonian forests are considered to be one of the world’s richest
biomes in terms of plant and animal diversity (Gentry, 1988). Although during
selective logging only a few valuable tree species are harvested, high intensity
logging can heavily impact a whole forest ecosystem (Frumhoff, 1995) including
damage to nearby trees and soils (Uhl and Buschbacher, 1985, Johns et al.,
1996) and increased risk of local species extirpation (Martini et al., 1994).
Furthermore, hunters gain greater access through the logging road network,
which increases both commercial and subsistence hunting, usually aimed at
primates and duikers (Frumhoff, 1995).

Therefore, impacts by selective logging may affect both plant and animal
species, especially those associated with the timber species (eg. mammals and
birds that eat timber specie fruits) (Martini et al., 1994). (Frumhoff, 1995)
affirrned that the ecosystem changes caused by selective logging result in loss of
equilibrium in animal population in terms of number of species and species
diversity, with some species tending to become more abundant than others
because of hunting. In terms of plant populations, selective logging eliminates the
most valuable timber species and, systematically, reduces the number of most fit
individuals and, therefore, preclude sustainable species reproduction (Uhl and

Vieira, 1989).

16

Selective logging activities also cause losses in the forest structural
integrity and lead to further degradation, especially in open canopy forest
patches with presence of vines (Uhl and Buschbacher, 1985). Therefore,
although selective logging in the Amazon floodplain concentrates on very few
species, it results in profound ecological changes. (Macedo and Anderson, 1993)
observed a dramatic transformation in the under-story vegetation as a result of
selective logging, with the development of secondary vegetation predominantly
formed by vines and herbs. Other environmental changes commonly associated
to selective logging include the development of drier conditions within logged
patches (Martini et al., 1994) and soil compaction (Pinard and Putz, 1996,
Fredericksen and Pariona, 2002). Meanwhile, habitat modifications directly affect
the distribution and abundance of amphibians and reptiles, seed dispersion, and
forest regeneration (Martini et al., 1994). The extended effects of selective
logging, including the decrease in forest yield, can lead to outright forest
conversion to other land use types (Pinard and Putz, 1996, Huth and Ditzer,
2001), which implies the complete destruction of a given natural forest patch.
1.3.3.4. Socioeconomic impacts

The Brazilian economy was placed among the 10 largest economies in the
world in 2000, with a Gross Domestic Product (GDP) of US$ 596 billion. Alone,
the timber sector contributed with approximately 5.5 billion to the Brazilian GDP
(ABIMCI, 2002). In the Brazilian Amazon, the timber industries are responsible
for approximately 15% of the regional GDP and employ 5% of the regional

workforce (World Bank, 1999). It is estimated that 61.1% of the total of effective

17

productive native forests are located in the Brazilian Legal Amazon, placing the
region as the main source of raw materials for the Brazilian timber industry
(ABIMCI, 2002).

In addition, according to (IBGE, 2006), from the country’s reported2
timber production between 1990 and 2003 a total around of 445 million cubic
meters of round wood was extracted in the Brazilian Amazon states, an average
around of 34.2 million m3 yr". The Brazilian Amazon alone produced
approximately 80% of national round wood production in Brazil over this time
(Figure 1.1).

Based on the census data provided by lBGE (2006), approximately 74%
of the timber production in the Brazilian Amazon from 1990 to 2003 was
extracted in the state of Para, 10% in the state of Amazonas, and 8% in the state
of Mato Grosso. Although this overall round wood production results place the
State of Amazonas as the second timber producer in the region, the State of
Mato Grosso has occupied this place from 1991 to 2003, immediately followed by
the states of Rondonia and Amazonas (IBGE, 2006).

As a result of an abundant timber resource pool, selective logging related
activities enhance the local economy. (Verissimo et al., 1992) observed that 112
sawmills in the vicinity of the city of Paragominas in Para, generated
approximately 5,700 jobs, including forestry workers (e.g. forest timber

extractors, truck drivers, machinery operators, etc.) and industrial workers.

2 Total of round wood production reported by timber industries to the Brazilian Institute for

Environment and Natural Resources - IBAMA (IBGE, 2005).

18

Verissimo and collaborators observed that more than 50% of the urban
population of Paragominas was directly dependent of the timber industry for
income generation. A major indirect impact of the timber industry is the creation
of significant tax revenues, which can be used for the benefit of the local and
regional population.

Figure 1.1. Forest production (round wood) from 1990 to 2003 in the Brazilian
Amazon (adapted from IBGE, 2006)

Timber-based local economies are subject to collapse, however. Pinedo-

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

I 90,000 100
80,000
§ 70,000
2 60,000 - ~~ - - -- -~~ 70 Q
.2. 50,000 ~ I, --/ r" ‘3‘ j :3 9:;
, ,.
f; 40,000 ~— «’7‘ ~ ~ —~—: 40 g
2 30,000 ~—-- - ~ ~ A— 2- 30 >
> 20,000 *..‘_'_I, ,,. 9:731” 20
10,000 — — ~=L 10
aasssssaasssss
e—It—uv—nt—nv—cv—r—n—r—Iv—tNNNN
+ Amazon round wood production
+ Amazon/Brazil round wood production (%)

 

 

Vasquez et al. (2001) conducted a study case in a floodplain forest in the State of
Amazonas, following a logging “boom” . In this case, the local timber industry
collapsed after the high value timber species were exhausted, leading industries
and individuals to migrate to other regions seeking new sources of raw material

and job opportunities, respectively.

19

Uhl and Vieira (1989) observed some indirect socio-economical impacts
of selective logging. Uhl and Vieira suggested that forest access roads left
behind after selective logging activities lead to the occupation of new areas by
landless peasants, increasing deforestation and threatening indigenous
communities. Furthermore, Uhl and Buschbacher (1985) observed that fire used
by farmers in the Amazon region to manage crops and pastures often spread into
adjacent forests. In contrast to non-logged forest, fire spreads readily in exploited
forest causing extensive damage.

1.4. General comments

A great number of articles available in the literature have been used to
support analyzes regarding the impacts by selective logging and forest fires in
the Brazilian Amazon. This body of literature has contributed to create a
conventional wisdom that characterizes these anthropogenic activities as major
drivers of forest degradation in tropical regions. For example, Feamside (1997)
and Uhl and Vieira (1997) affirmed that logging often leads to significant losses in
canopy and biomass content. Furthermore, Holdsworth and Uhl et al. (1997),
Cochrane (1999), and Cochrane and Schulze (1999) observed that selective
logging leads to a significant increase in the number of forest fires in tropical ever
green forests. Meanwhile, Nepstad et al. (1999) have estimated areas impacted
by logging based on records of sawmills’ production because they believed that
logging was cryptic and, therefore, could not be adequately resolved by the
usage of remotely sensed data. In addition, these authors had never directly

measured the extent of selective logging at the regional scale. Nonetheless,

20

Nepstad et al. (1999) combined these studies and concluded that selective
logging is highly significant, equal or more than deforestation to explain forest
degradation in tropical countries. Another major conclusion of Nepstad and his
colleagues was the lack of reliable measurements of the combined impacts of
selective logging and forest fires in the Brazilian Amazon.

In my research, one of my main goals was to investigate the validity of the
existent conventional wisdom about the extent and impacts of selective logging
and forest fires in the Amazon region. To achieve this goal I created a
methodological framework to quantify forest degradation at local and regional
scales by examining the spatial characteristics and the temporal changes of
selective logging and forest fires. I also investigated the factors controlling the
processes of forest degradation at various geographical and temporal scales in
the Amazon basin.

1.5. Overview of the study

This manuscript was divided into five chapters to address the previously
stated research purposes. Chapter I presents the research design and describes
the general scientific development of the study. Chapter II focuses on the
development of remote sensing and statistical analysis approaches used to
assess extent and impacts by selective logging and forest fires in a case study in
the southwestern Amazonian state of Mato Grosso. Chapter III shows the
application of the scientific approaches developed in Chapter II to a basin wide
assessment of both the extent of selective logging and forest fires and their

impacts on Amazonian forests. Chapter IV shows the development and

21

implementation of spatiotemporal analyzes of selective logging and forest fires
using probabilistic probit models at the regional scale in the Brazilian Amazon.
Finally, Chapter V summarized analyses presented in the previous chapters by
revisiting all research questions and hypotheses, and by connecting all research
findings to the global change context. Finally, I concluded by presenting the
potentials raised by this study for further inquiries on the processes controlling

the pace of environmental changes in the Brazilian Amazon.

22

CHAPTER II
Methods for Assessment of Impacts by Selective Logging and Forest Fire:
A Case Study in Mato Grosso, Brazil
2.1. Abstract research
Many studies have been conducted to assess the process of forest degradation
in the Brazilian Amazon. Several of those researches relied on remote sensing
approaches to estimate the extent and impact by selective logging and forest
fires on tropical rain forest. However, only a few studies assessed the combined
impacts of those anthropogenic activities. I conducted a detailed analysis of
selective logging and forest fire impacts on natural forests in the Southern
Amazon State of Mato Grosso, Brazil, one of the key logging centers in Brazil. I
used a 13-year series of annual Landsat images (1992-2004) to test different
remote sensing techniques for measuring the extent of selective logging and
forest fires and their impacts and interactions with other land use types in the
study region. I also assessed forest canopy regeneration following these
disturbances. Field measurements and observations were conducted to validate
remote sensing approaches. The results indicated that the Modified Soil Adjusted
Vegetation Index aerosol free (MSAVIaf) is a reliable estimator of fractional
coverage under both clear sky and smoke conditions in this study region. Multi-
temporal analysis of land use and land cover change from 1992 to 2004
indicated that around of 29% of the study area had been converted into
agricultural land by 2004. In terms of extent, selective logging activity was the

most important forest disturbance factor, impacting more than 31% of the total

23

study area by 2004. Approximately 28%, 5%, and 3% of the study area was
impacted by deforestation, selective logging and forest fire combined, and forest
fire only, respectively. Although selective logging impacted the largest extent of
natural forest in the period of analysis, more than 35% and 28% of the observed
forest canopy losses were due to forest fire and selective logging combined and
to forest fire only, respectively. These study results also indicated that
approximately 70% of total area of forest disturbed by logging and fire had
sufficiently recovered to become undetectable on satellite imagery in 2004, which
implies that selectively logged and burned forests may transition from a carbon
sources to carbon sinks in the subsequent years due to strong forest
regeneration. This type of forest degradation is an addition to outright
deforestation analysis and has yet to be accounted for. Finally, the results of this
research could be useful to global climate change analysis, sustainable forest
management programs, and environmental policy development in Brazil.

2.2. Introduction

Selective logging activities have been occurring in Brazil’s tropical forests

for several decades (Stone and Lefebvre, 1998, Nepstad et al., 1999, Alvarado
and Sandberg, 2001). Logging activities were restricted to flood plains (véizeas)
until the 1960s due to the pre-existent access though rivers (fluvial
transportation) in areas flooded annually, but the widespread construction of
roads during the 1960s and 1970s allowed the expansion of selective logging
into the inter-fluvial (terra-firme) forest in the Brazilian Amazon (Uhl and Vieira,

1989, Uhl et al., 1997).

24

As a result of increased access provided by roads, selective logging has
become a major environmental concern in the Brazilian Amazon due to its
potential negative effects on natural forests. Selective logging is a form of timber
extraction of a select group of tree species where only the most valuable tree
species are removed from the forest (Verissimo et al., 1995, Uhl et al., 1997).
Selective logging activities leave behind a complex landscape comprised of intact
forest, treefall gaps, roads, log-loading patios, and damaged forest. These
activities increase the amount of dead slash or dried biomass (fuel) and,
consequently, forest fire susceptibility is substantially increased (Uhl and
Buschbacher, 1985, Stone and Lefebvre, 1998, Nepstad et al., 1999, Souza and
Barreto, 2000).

The impacts caused by logging in tropical forests are significant in terms
of forest degradation and fire susceptibility (Uhl and Buschbacher, 1985, Stone
and Lefebvre, 1998, Nepstad et al., 1999, Souza and Barreto, 2000). Selectively
logged forests become highly degraded and usually have 40 — 50% of the
canopy cover destroyed during logging operations (Uhl and Vieira, 1989,
Verissimo et al., 1992). Forests heavily impacted by selective logging will have a
significant increase in the level of forest degradation and fire susceptibility
(Holdsworth and Uhl, 1997, Stone and Lefebvre, 1998, Cochrane, 2001). Not
surprisingly, forest fires have been increasing in the Brazilian Amazon, and are
more common in fragmented forests located next to deforested areas, which
serve as ignition sources of forest fire (Uhl and Kauffman, 1990, Cochrane and

Schulze, 1998, Cochrane, 2001, Cochrane et al., 2004).

25

Most studies on the impacts of selective logging that relied on remotely
sensed data quantified forests affected by selective logging were conducted for
study sites in Eastern Para, Brazil, using Landsat images and visual
interpretation. For example, Souza and Barreto (2000) conducted an
investigation using a linear mixture model and Landsat imagery to detect logging
patios within selectively logged forests in a study site in the State of Para, Brazil.
Cochrane and Souza Jr (1998) developed a remote sensing technique to detect
and classify burned forests using non-photosynthetic vegetation derived from
Linear Mixture Analysis for a study site in Tailandia, Amazon State of Para.
Cochrane and Sousa Jr (1998) and Cochrane et al. (1999) conducted field
studies and a multi-temporal analysis of remotely sensed imagery to understand
forest fire dynamics in a study site in the Amazon state of Para, Brazil. Souza et
al. (2003) developed a methodology to map classes of degraded forest for a
study case in the State of Para using fraction images (vegetation, non-
photosynthetic vegetation, soil, and shade) derived from spectral mixture models.

More recently, Souza et al. (2005) conducted an evaluation of different
vegetation and infrared indices and fraction images derived from spectral mixture
analysis in assessing multi-temporal forest degradation within nineteen transects
in the eastern Amazon. Matricardi et al. (2005) estimated areas of selectively
logged forests throughout an entire Landsat scene in the State of Mato Grosso,
Brazil, by combining fieldwork and remote sensing approaches.

Although these previous studies significantly improved remote sensing

approaches to assess both extent and impacts caused by selectively logging and

26

forest fires using remotely sensed data in tropical forests, they have focused on
particular study sites, a more comprehensive assessment of forest disturbances
by forest fire and selective logging and their interactions with other land use and
land cover processes have yet to be conducted. Moreover, atmospheric changes
caused by smoke derived from deforested and forested areas have limited multi-
temporal analysis using remotely sensed data (Karnieli et al., 2001, De Moura
and Galvao, 2003).

To address these scientific gaps, I conducted field observations and
tested performance of different vegetation indices and a green vegetation fraction
(GV) derived from Spectral Mixture Analysis (SMA) to assess forest canopy
degradation in the presence of smoke. I also improved the Modified Soil-Adjusted
Vegetation Index (MSAVI) to MSAVI_af (MSAVI aerosol-free). In addition, I
developed an automated remote sensing approach to map burned forest using
non-photosynthetic vegetation (NPV) fraction derived from SMA. Finally, I
conducted a multi-annual analysis of forest canopy degradation in the study area
that provided useful information to assess forest impact and forest regeneration
for years following selective logging and forest fire events. Spatial interactions
between burned and selectively logged forests were also examined.

2.3. Methodology
2.3.1. Study site

This research was conducted using one Landsat scene (path 226 and row

068) that encompassed approximately 30,000 km2 in the State of Mato Grosso,

Southern Brazilian Amazon (see Appendix 8.1). The study area includes the

27

municipios of Santa Carmem and Uniao do Sul and parts of the municipios of
Colider, Feliz Natal, ltaI'Jba, Marcelandia, Nova Ubirata, Paranatinga, Sinop,
Sorriso, and Vera, which form a territory known as the Sinop region.

The climate in the study area is humid tropical with very distinct dry and
wet seasons that extend from June through September and from December
through March, respectively. The average annual precipitation and temperature
is 2000 mm and 26° C, respectively (RADAMBRASIL, 1980). Prior to modern
colonization the study area was mostly covered by transitional forests, a semi-
deciduous forest type, with emergent can0py. However, by 2004 approximately
8,500 km2 of native forest had been converted into extensive soybean plantations
and pasture. Nonetheless, the Sinop region remains a major timber center in the
Amazon in spite of prevailing high deforestation rates.

2.3.2. Dataset
2.3.2.1. Land use and land cover dataset

Multi-annual land use and land cover GIS (Geographic lnforrnation
System) layers for 1992 through 2004, produced at GOES for the Landsat path
226 and row 068, were used in this analysis. These layers were generated via
standard unsupervised classification of Landsat TM and ETM+ imagery and
subsequent manual editing using GIS. Each layer included seven classes of land
use (forest, deforestation, secondary regrowth, savannah, cloud, shadow, and
water body) (see Table 2.1). Multiple non-forest land use and land cover classes
were lumped together to create a forest and non-forest mask. This masking

procedure was performed to separate forest only, degraded or not, on each

28

Landsat image. Forest masked images were used to detect selectively logged
and burned forests.

Table 2.1. Land Use and Land Cover dataset characteristics

 

 

Dataset Features Format Spatial Period of Projection
resolution Analysis
Land use Forest, Arc/Info 30m x 30 m 1992 UTM,
and land deforestation, grid to Zone 21,
cover GIS secondary 2004 Datum and
layers regrowth, cerrado, Spheroid
clouds, and WGS84
shadows.

 

Moreover, cumulative cloud and shadow layers for the period of analysis
also were masked out of all Landsat images, which provided a common area of
analysis. However, forested areas affected by smoke, visible on bands 4, 5, and
7, were not masked out.
2.3.2.3. Selective logging dataset

Selective logging GIS layers prepared by Matricardi et al. (2005) were
used. Matricardi et al. (2005) used semi-automated (texture algorithm) analysis
of Landsat band 5 to detect log landings (log storage patios). Subsequently,
Matricardi et al. (2005) applied variable buffer zones (180 and 450 meters)
around patios to estimate areas affected by selective logging. In addition,
selective logged forests showing obvious forest canopy degradation on Landsat
images were mapped by digitally circumscribing them on a computer screen. The
combination of these techniques allowed the detection of selective logging with

92.9%, 91.2%, and 92.9% of user, producer, and overall accuracy, respectively.

29

Common areas between these two techniques were around 50% without double
counting. Areas eventually missed by using automatic analysis were added later
by visual interpretation.

In this study, the dataset created by Matricardi et al. (2005) was used.
This dataset involves multi-annual measurements (1992 to 2002) of selective
logging in the study area (path 226 and row 068), located in State of Mato
Grosso (Table 2.2).

Table 2.2. Selective Logging dataset characteristics

 

Dataset Features Format Spatial Period of Projection
resolution Analysis

 

Selectively Binary classes Arc/Info 30m x 30 m 1992 UTM,

logged (1 = logging, grid to Zone 21,

forest GIS 0 = no data) 2002 Datum and

layers Spheroid
WGS84

 

Selective logging layers for two additional years (2003 and 2004) were
prepared for this study area using the same methodological approach developed
by (Matricardi et al., 2005). Subsequently, selective logging layers were
incorporated as part of the land use and land cover change dataset the present
analytical effort.
2.3.2.4. Satellite imagery

The satellite imagery used in this study was drawn from the archive
maintained by the Tropical Rain Forest lnforrnation Center
(mpzllwwwtrficmsuedul) at the Global Observatory for Ecosystem Services

(GOES), Michigan State University, and by the Brazilian National Institute for

30

Space Research (MH/wwwinpebr). The archive holds Landsat TM and ETM+
images from 1992 to 2004, bands 1 to 5, and 7. The Appendix 8.2 shows further
details of the Landsat imagery used in this analysis.
2.3.2.4.1. Radiometric and atmospheric corrections of Landsat imagery
Radiometric and atmospheric corrections can significantly improve image
quality by reducing atmospheric effects of the absorption by atmospheric gases
and the atmospheric scattering processes (Vermote et al., 1997). In this study,
radiometric correction was performed to normalize satellite images for factors
such as sensor degradation, earth-sun distance variation, incidence angle, view
angle, and time of data gathering. This correction involved converting digital
number (DN) into radiance and, subsequently, radiance into reflectance using
different calibration coefficients provided by the metadata files that accompanied
the image files. Coefficients provided by Chamder and Makham (2003) were
used for those Landsat TM images in which coefficients were not provided.
According to Chamder and Markham (2003), radiometric corrections of
Landsat imagery can be performed, first, by converting digital numbers (DN) to

radiance:

 

(LMAX/t — LMIN/t
L). =

Qcalmax )Qcal + LMIN/t. (1)

where LA = spectral radiance at the sensor’s aperture in Wl(m2 . sr .p) and Goal =

Quantized calibrated pixel value in units of digital numbers (DNs).

Then, by converting radiance values to top of atmosphere reflectance

using the following equation:

31

2
p? = [ngcar + LMIN}. . (2)
where pP = unit less planetary reflectance, LA = spectral radiance at the sensor’s
aperture, d = earth-sun distance in astronomical units, ESUN/l = mean solar
exoatmospheric irradiances, and Os = solar zenith angle in degrees.

The resultant at satellite apparent reflectance was then corrected for
atmospheric effects providing at-ground surface reflectance. For converting
apparent into surface reflectance the SS (Simulation of the Satellite Signal in the
Solar Spectrum) model was used. It required, however, the input of different
parameters from the site and sensor as input such as site elevation, sensor
altitude, visibility, date of image acquisition, and solar and sensor zenith and
azimuth angles. Landsat images are provided with sensor parameters available
in a corresponding metadata file for each Landsat image. The site altitude (340
meters) at the center of the Landsat path 226 row 068, and visibility of 20
Kilometers were applied. Based on these parameters, this atmospheric
correction model provided linear equation coefficients (Appendix 8.3.) for each

Landsat scene. Figure 2.1 shows the procedure for performing these image

corrections.

32

 

 

 

 

Radiometric Digital Radiance TOA Surface
correctIon Number Reflectance Reflectance

Geometric [ Geometric Registration I
correction f

 

 

 

 

 

 

 

Figure 2.1. Stepwise procedure for radiometric and geometric corrections of
Landsat imagery

Finally, the atmospherically corrected Landsat images were geometrically
corrected. The entire procedure for geometric registration is described bellow.
2.3.2.4.2. Normalization of seasonality effect of Landsat imagery

Seasonality effects on soil moisture, vegetation phenology, sun angle, etc.
occur due to different date of acquisition of remotely sensed data. These effects
may affect performance of remote sensing procedures on multi-temporal analysis
(Mas, 1999). Therefore, in order to minimize these effects in this study, most
images were acquired between May and August. Additionally, normalization of
seasonality effects was applied to enable inter-comparison of fourteen Landsat
images from different dates in this analysis. Normalization also may minimize
atmospheric and other sensor and image variation Elvidge et al. (1995) due to
the study area characteristics during the image acquisition period.

Normalization was carried out on each Landsat scene. First, a Landsat
scene acquired in 2001 was selected as the reference image (a good quality
image) after a careful spectral and visual examination of all Landsat scenes
included in this study. Then, the brightest and darkest pixels were selected on

the reference image and every other image used in the analysis (Elvidge et al.,

33

1995). These sampled pixel values consisted of areas covered by undisturbed
forests, soil exposure (brightest pixels), and deep water (darkest pixels). A linear
regression model was calculated using surface reflectance values from the
reference (dependent variable) and every other Landsat image (independent
variable), which produced the normalization coefficients (slope and intercept) for
each band and Landsat scene. Based on these coefficients (Appendix 8.4), each
band and Landsat image was normalized accordingly.
2.3.2.4.3. Geometric correction

The Landsat images were previously system-corrected at the TRFIC using
the sensor calibration data. Geometric correction accuracy was calculated by
comparing common points on system-corrected images to ground collected using
GPS from several locations within the study area. Image rectification using 3 to
10 point coordinates and nearest neighbor re-sampling technique was applied.
Individual geometric correction was deemed acceptable only if the root-mean-
square (RMS) was less than 0.5 pixels.
2.4. Methods

This research was based on remote sensing approaches and fieldwork
measurements and observations to investigate the processes of forest
degradation due to selective logging and forest fire and their temporal changes in
the study area in the Mato Grosso state.
2.4.1. Field measurements of forest canopy

Field measurements intended to support the assessment of remote

sensing approaches and analysis of forest disturbances were conducted from

34

June 20 to July 10, 2004. Different types of selective logging (e.g. ongoing
logging and previously logged forests), logged and burned forests, and
undisturbed forests were sampled within the study area. Eight transects of 500
meter length were set in these sites. Each transect was appropriately located
using Geographic Positioning System (GPS) with x and y coordinates at the start
and finish locations (see Appendix 8.5 for further details). Hemispherical photos
were acquired for each transect. Canopy openness was measured for both
disturbed and undisturbed forests using hemispherical photos (further details are
described below). Various types of forest disturbances and intensities were also
observed.

Hemispherical photos acquired with “fish eye” lens and digital camera (3.2
mega pixels resolution) beneath undisturbed and disturbed forest canopies were
used to estimate canopy openness. The photo-shots were taken at 10-meter
intervals along each transect. Canopy openness measurement for each
individual location was calculated using the Gap Light Analyzer (GLA) software.
Fraction canopy openness was estimated using a semi-automated technique
because each individual photo requires a threshold empirically defined by the
photo interpreter (Appendix 8.6). Subsequently, the canopy fraction was
obtained by calculating the difference between 100% and canopy openness (%).

In addition, 120 sampling points randomly distributed over forested areas
within the study site were visited. Sampling points located too far from the road
network were re-located to a closer distance, within a similar forest pattern

visually observed on Landsat image. Based on visual evidences of forest fires

35

observed in the field, each sampling point was classified either as burned or not
burned forest and appropriately located with x and y coordinates using GPS.
2.4.2. Remote sensing approaches

As previously indicated, changes in forest attributes (degradation) are
usually too subtle to be accurately assessed using conventional remote sensing
approaches and land use classification schemes (Lambin, 1999, Nepstad et al.,
1999). Lambin (1999) suggested that biophysical attributes derived from remotely
sensed data (e.g. vegetation indices and land surface temperature) at fine spatial
resolution should be used to assess changes in forest canopy and identify
process of forest degradation throughout a study region. Landscape attributes
such as type, size, distribution, and shape, may reveal important spatial
components and their interactions that may help to explain process of land cover
modification. And, finally, the multi-temporal information can explain processes of
land cover changes over time by detecting natural, anthropogenic, seasonal or
permanent modifications on the landscape.

Analysis forest degradation due to selective logging and forest fire impacts
were assessed in this study at local and multi-annual scales, using information
from the previously described field measurements, remotely sensed data, and
statistical analysis. Moreover, as suggested by Lambin (1999), spectral, spatial,
and temporal indicators of forest degradation and regeneration associated with
selective logging and forest fire were assessed. Firstly, vegetation indices
retrieved from Landsat imagery were tested against canopy cover measurements

derived from hemispherical photos acquired in the field, which helped to select

36

an optimum vegetation index, which was used to estimate fractional cover using
Landsat imagery. Secondly, time-scale variation of these forest attributes was
assessed through multi-temporal analysis of canopy degradation derived from
remotely sensed data. Finally, spatial pattern of forest degradation due to
selective logging and forest fire and their relationship with other land uses and
land cover was analyzed.

Further details on remote sensing approaches to assess forest canopy
degradation are described as following.
2.4.2.1 . Vegetation lndices

Different vegetation indices have been developed and used to better
understand the vegetation dynamics in different regions (Qi et al., 1994, Karnieli
et al., 2001, Huete et al., 2003). Vegetation lndices (Vls) are mathematical
transformations of multiple spectral reflectances employed to estimate the
amount of photosynthetically active vegetation at the land surface (Huete et al.,
2003) and may be used as indicators of vegetation growth (Wiegand et al., 1991)
and forest disturbances (Qi et al., 2002). Vl’s principles are based on the
variation of reflected near infrared and red energies by vegetation, and takes
advantage of contrasts between soil and vegetation reflectances. Consequently,
near infrared and red are the most commonly used spectral bands to calculate
vegetation index (Qi et al., 1995).

Distance and slope based are the two most important vegetation index
developing approaches. Differential between near infrared and red reflectances

and the ratio of these two spectral bands are used by the distance and slope

37

based approaches, respectively (Qi et al., 1995, Qi, 2001). Both approaches aim
to be more sensitive to the vegetation component, while minimizing external
effects due to reflectance of red and near infrared variation at the sensor caused
by solar irradiance, atmospheric conditions, canopy background, and vegetation
canopy structure and composition (Qi, 2001).

MSAVI (Modified Soil Adjusted Vegetation Index), NDVI (Normalized
Difference Vegetation Index), and GEMI (Global Environmental Monitoring Index)
were tested against canopy cover derived from hemispherical photos to support
the selection of the optimum vegetation index. NDVI is one of the earliest and
widely used indexes in environmental studies using remotely sensed data.
Meanwhile, MSAVI and GEMI are newer approaches aiming to adjust to soil
factor and to minimize atmospheric effects, respectively (Qi, 2001).

MSAVI is a ratio-based vegetation index developed by (Qi et al.,
1994).The basic idea of MSAVI is to provide a variable correction factor “L”. This
adjustment factor “L” depends on the level of vegetation cover being observed
and on the product of NDVI and WDVI (Weighted Difference Vegetation Index).
In this case, the isovegetation lines do not converge to a single point, leading to a
circular problem because vegetation covers must be known before calculating
the vegetation index, which will generate the vegetation cover. The soil line has

arbitrary slope and passes through origin, ranging from —1 to +1.

MSAVI: [MR—pm“! (1+L), (3)
pN1R+pred+L

 

38

where leR = reflectance at near infrared band, pred = reflectance at red

band,

2
L =I(.0N1R ‘ Pred) * S +1 + Ple + PredI — 8.0 * S * (PNIR ‘ Pred ). (4)

where s = 1.2 (slope of the soil line calculated from surface reflectance at

non-forested areas in the study site).

The NDVI, developed by Tucker (1979), is the most applied index in
vegetation studies. NDVI is a ratio-based vegetation index calculated from the
difference divided by the sum of two spectral bands yielding a normalized (-1 to
+1) difference. NDVI preceded and offered theoretical basis for many other
vegetation indices and it still is widely used in remote sensing applications (Qi,

2001). It is defined as follows:

 

NDVI = pm' pm! . (5)
pNIR + pred

where pNIR = reflectance at near infrared band and pred= reflectance at red

band.

The GEMI was developed by (Pinty and Verstraete, 1992). It is a distance-
based approach and non-linear vegetation index developed to be less sensitive
to atmospheric effects. Although GEMI was initially intended to minimize
atmospheric effects, it also is more sensitive index to soil background (Qi et al.,

1994).

39

770— 0.257;) — (pred — 0.125)
(1 — pred) '

where pNIR = reflectance at near infrared band and pred = reflectance at red

GEMI = (6)

 

band,and,

2( p2N1R -p2red) + 1.5 pNIR + 0.5 pred
77 :

leR + pred + 0.5 (7)

 

Vegetation index applications, however, have been limited by their
sensitivity to atmospheric effects, especially on visible bands which are strongly
affected by aerosols such as smoke and other types of aerosols (Karnieli et al.,
2001, Huete et al., 2003, Moura and Galvao, 2003). Shortwave bands are
insensitive to atmospheric effects and are also highly correlated with visible
bands under aerosol-free atmospheric conditions. Therefore, shortwave bands
have been favored as an alternative approach to replace the most sensitive
bands when atmospheric effects thwart multi-temporal analysis (Karnieli et al.,
2001, Moura and Galvéo, 2003).

Based on these characteristics of the shortwave bands, Karnieli et
al.(2001) tested four modified vegetation indices (NDVI, Soil-Adjusted Vegetation
Index — SAVI; Atmospheric Resistant Vegetation Index - ARVI; Soil-Adjusted
Atmospheric Resistant Vegetation Index - SARVI) for a study site in the Amazon
State of Mato Grosso, Brazil. Karnieli and his colleagues developed a modified
NDVI, named the Aerosol Free Vegetation Index (AFRI), which uses an empirical
relationship to estimate red band reflectance from shortwave infrared band at 2.1

pm. Equation 8 defines this vegetation index:

40

pNIR — 05,0st (8)

AFR12 1 = .
I leR 'I- O'SPSWIR

 

where IONIR= reflectance at near infrared and pom: reflectance at 2.1 pm

(Shortwave infrared).

This vegetation index has proven to be useful to estimate biomass burned
in the presence of smoke and other atmospheric pollutants. However, AFRI may
not produce good results if atmospheric particles are of equal or larger sizes than
the wavelength at 2.1 pm (Karnieli et al. 2001).
2.4.2.2. Modified Vegetation Index approaches

The acquisition of remotely sensed data affected by smoke and other
aerosols was unavoidable due to the prevailing land use practices in the study
region. Consequently, I had to assess performance of various vegetation indices
to estimate forest canopy degradation under adverse atmospheric conditions. I
tested four vegetation indices (AFRI, NDVI, MSAVI, and GEMI) and the Green
Vegetation (GV) derived from SMA to estimate fractional coverage in the study
site under smoke conditions. Additionally, I also tested a modified GEMI and
MSAVI using the same empirical linear relationship estimated by Karnieli et
al.(2001) to predict red band reflectance based on shortwave infrared at 2.1 pm.

The equations 9 and 10 define the GEMI“ model.

 

770- 0.2527) — (0.5pswm —0.125)

GEMI =
2'1 (1 — 0.5pSW1R)

(9)

where pNIR = reflectance at near infrared band and pswm = reflectance at

41

shortwave infrared band (2.1 pm), and,

_ 2(p2NIR - 0.5p2SWIR)+1.5leR + 0.5(0.5pSll'lR)
_ leR + OSPSWIR + 0.5 '

 

(10)

The MSAVI using wavelength at 2.1 pm was named MSAVIaf (Modified

Soil-Adjusted Vegetation Index aerosol resistant). The equations ‘1 1’ and ‘12’

 

describe the MSAVIaf model.
1R — 0.5 WlR
MSAVIaf= pN ’05 (1+L), (11)
prime + 0.5pSWIR + L

where leR = reflectance at near infrared band, pswm = reflectance at

Shortwave infrared band (2.1 pm).

L = [(PNIR -0-5DSWIR)* S + 1 + PNIR + 0-5PSWIR12 -8-0 * 5* (PNIR -0-5PSWIR) .(12)

where s = 1.2 (slope of the soil line calculated from surface reflectance at

deforestation areas in the study site).

2.4.2.3. Spectral Mixture Analysis

Multi-spectral categories may be represented by each pixel, and mixing
proportions (class fractions or abundances) change from pixel to pixel.
Furthermore, real world scenes always have spatial details with smaller
dimension than any Ground Instantaneous Field of View (GIFOV) which
ultimately represent the spatial resolution of a given sensor or satellite image.
Moreover, spatial mixing can be also derived from a image resampling process

(Rencz, 1999).

42

The number of spectral bands available in the dataset plus 1 defines the
total of possible end members. In practice, the number of end members ranges
from three to seven, depending on the number of bands of the data set and the
spectral variability of the scene components (Rencz, 1999).

The selected endmembers are used to solve the following mixture
modeling equation:

N N
DNb =i§1FiDNib +Eb or DNb = ERR“, + ab, (13)
where DN,> = intensity of a given pixel in bandpass ‘b’, DN,- = intensity of image
endmember ‘r’ at wavelength ‘b’, Rib = reflectance of end member ‘i’ in bandpass

‘b’ (if an image is calibrated to reflectance, Rib must be used instead), N =

number of endmembers, and ab = error of the fit for bandpass ‘b’.

A second equation may be used to constrain the sum of fractions to 1:

Fi = 1.0 , (14)

'M2

I I

where F,-= fractional abundance of endmember ‘i’

Since spectral endmembers may be the result of mixing with different
materials, fractional abundance of endmember ‘I’ may be less than 0 and greater
than 1, although the overall fraction will be 1. It indicates that two materials may
be combined, which will result a negative or greater than 1 fraction. It does not

necessarily mean that there are method errors (Rencz, 1999).

43

2.4.2.3.1 . Endmember selection

Spectral mixture analysis approach relies on endmember selection and,
therefore, accuracy depends on their accurate selection of endmembers.
Spectral signatures or spectral endmembers correspond to the physical
endmembers that show maximum abundance. Spectral endmember identification
for each piece of terrain is the first stage Rencz (1999) and main objective in
mixture modeling (Adams et al., 1995, Bateson and Curtiss, 1996, Rencz, 1999).
Theoretically, the data dimension and constraints of the mixture inversion define
the number of end members. Consequently, the number of spectral bands in the
dataset plus 1 defines the total of possible end members. However, empirically,
the number of end members ranges from three to seven, depending on the
number of bands of the data set and the spectral variability of the scene
components (Rencz, 1999).

As previously indicated, the available dataset (spatial and spectral
resolution, bands, etc.) and classes of surface materials (soil, shade, green
vegetation, non-photosynthetic vegetation, etc.) of a given area affect the
definition of endmembers in SMA. Thus, image endmembers can be identified
based on some criteria (spectral library acquired from a specific site) or using
other techniques (Rencz, 1999).

The end members are idealized pure signature for a class of a particular
location and are the extreme pixels of the scatter plot when spectral libraries are
used. Endmembers do not actually exist in the dataset, but their convex hull will

enclose all data in the scattered plot. Endmembers represent 100% pure pixels in

44

their respective classes. For example, the shade endmember is a result of both
shade and shading highly affected by the local topography, vegetation, and
seasonal and daily solar variation. Therefore, shade endmember may compose
with other endmembers in a given satellite scene and proportionally vary with
these illumination characteristics (Rencz, 1999).

The Pixel Purity Index (PPI) also is presented as an option to select end
members. This technique was based upon the principle that spectral signature of
a specific material is ‘pure’ (not compounded) Adams et al. (1995) and, therefore,
PPI was designed to identify the most spectrally extreme, different, or ‘pure’
pixels, which commonly correspond to mixing endmembers (Adams et al., 1995,
Boardman et al., 1995). The PPI is calculated by repeatedly projecting n-
dimensional scatterplots onto a random unit vector. To find the PPI value, the
extreme pixels in each projection are identified and their frequencies as extreme
are determined. A PPI image is generated accordingly with the resultant
histogram of the frequency of each pixel value as extreme, and thus making
evident the purest pixels with extremity-scores (Boardman et al., 1995).

For the present study, a four-endmember mixing model was constructed
to generate fraction images of shade, soil, non-photosynthetic vegetation, and
green vegetation. The end-member candidates were selected by using the
previously describe PPI technique. Six image subsets of 502 x 502 pixels
representing the variety of land use and land cover types within the study site
were used as input for the PPI algorithm for each year of analysis. Purest pixel

candidates were identified based on the PPI results and located on the original

45

Landsat imagery within the study area. A final visual and spectral examination of
spectral shape and image context was conducted accordingly to select the final
endmembers for the SMA.
2.4.2.4. Fractional cover

Fractional cover or green fractional percentage can be estimated by using
a simple linear mixture model (LMM) and two spectral end members (plant

canopy and bare soil) as following (Maas, 2000, Qi et al., 2000):

fc- VI—Vlopen *100 (15)

— VIcanopy — VIopen

 

where fc = green fractional percentage, VI = vegetation index value, VI canopy =
vegetation index value of the tree green canopy, and VI open = vegetation index
value of senescent open areas (Figure 2.2).

Figure 6 shows the procedure to calculate fractional cover.

[ Surface reflectance H Vegetation index ]
l
t
[ End member selection H Fractional cover (%) ]

Figure 2.2. Procedure for calculating fractional cover

 

 

 

 

 

 

 

 

 

Forest canopy disturbances and regeneration were estimated using
fractional cover derived from the discussed vegetation indices previously (NDVI,
MSAVI, GEMI, AFRI, MSAVIaf, and GEMI“, and GV), all retrieved from remotely
sensed data. The performance of each vegetation index was observed based on

the statistical results of Pearson correlations among the vegetation indices

46

retrieved from a Landsat image acquired in August 6, 2003, under clear sky and
in the presence of smoke for a variety of land cover types. The optimum
vegetation index was defined based on its performance under smoke conditions
and on the results of the empirical linear relationship between fractional cover
retrieved from a Landsat image acquired in June 21, 2004 and fractional cover
calculated from hemispherical photos also acquired during June, 2004, for
different sites within the study area. Forest canopy degradation of selectively
logged and burned forests, selectively logged only, and undisturbed forests were
also compared.
2.4.2.5. Burned forest detection

Burned forests were detected using non-photosynthetic vegetation fraction
images derived from the linear mixing model. This technique was first developed
by Cochrane and Souza Jr. (1998) to detect selective logging sites and burned
forests using Spot images in a case study in Para, Brazil. I improved this
approach to a fully automated technique to detect and map burned forests using
non-photosynthetic vegetation fraction image derived from Landsat imagery.
Therefore, burned forests were detected and mapped in the study area from
1992 to 2004. Additional details of this remote sensing approach for detecting
burned forests are described bellow.
2.4.2.5.1. Burned forest detection approach

This approach consisted in: 1) masking Landsat images in forest and non-
forest, and 2) unmixing multispectral images using shade, green vegetation, and

non-photosynthetic vegetation (NPV) endmembers. In this case, multispectral

47

images were unmixed based on Adams et al (1995) who observed that fraction
images could be generated considering particular endmembers (spectral pure
signature) that represent specific components such as green vegetation, soil,
non-photosynthetic vegetation, and shade. The Landsat imagery was processed

as described on Figure 2.3.

 

 

 

Landsat Image
[(Surface ReflectancJ [ LULC Layer

 

 

 

 

V

Masking out
non-forest

 

Forest and ’non-forest
binary classification

 

l

Lorest, NPV, Shade
E

 

 

 

 

ndmember Selection

T’T

 

 

 

 

 

L__J

[NPV FractIon Image (289 P'Xels)

 

 

 

7

Binary classification
NPV>0.12=1, 0else

 

Low paSSI filter
(5 x 5 pixels)

Masking out
non-forest
l Eliminate clumps ]

.83

 

 

 

 

 

 

 

 

,
Pixel clumping
L (4 pixels) J [Burned Forest Map]

 

 

 

 

Figure 2.3. Flow diagram for burned forest automated detection technique

The NPV fraction images were classified following a binary labeling
(burned forests = 1 if NPV value was greater or equal 0.12; and unburned forests
= 0 othenivise). Subsequently, these classified images (burned and non-bumed

forests) were 4 pixels clumped using the clump algorithm from Erdas Image

48

software. Clumps lesser than or equal 289 pixels were eliminated to remove
small pixel clusters (noise) of misclassified unburned forests. Finally, a low pass
filter was applied on clump eliminated images to smooth out limits of the detected
burned forests.
2.4.3. Forest canopy degradation assessment

Forest canopy degradation by selective logging, forest fire, logging and
fire combined and forest regeneration following each of these forest disturbances
in the study area was assessed using multiple regression analysis. Multiple
regression analysis is useful technique to estimate partial effects of the
independent variable by controlling many other factors that simultaneously affect
the dependent variable (Wooldridge, 2000).

By using the pixel as the unit of observation and based on (Wooldridge,
2000), the multiple linear regression model for green fractional percentage was

defined as:

FC% = ,60 + fl1x1+ ,Bzxz +...+,624x24 + u , (16)
where FC% is the observed continuous variable (green fractional percentage
derived from remotely sensed data), and ,80 is the intercept or constant, ,6] is
the parameter associated with x1 (independent variable 1), ,62 is the parameter

associated with x2 (independent variable 2) and so on. The variable u is the error
term, which contains factors other than the independent variables included in this
model that affect FC%.

The vector of independent variables (x) tested are the following:

49

x1 = Deforestation areas

x2 = Newly logged forests

x3 = Old logged forest detectable

x4 = Old logged forest non-detectable

x5 = Newly logged and burned forests

x5 = Newly burned forests

x7 = Old logged and burned forest detectable
x8 = Old logged and burned forest non-detectable
x9 = Old burned forest detectable

x10 = Old burned forest non-detectable
X11= Year Of1993

X12 = Year Of1994

x13 = Year of 1995

x14 = Year of 1996

x15 = Year of 1997

x16 = Year of 1998

x17 = Year of 1999

x13 = Year of 2000

x19 = Year of 2001

x20 = Year of 2002

x21 = Year of 2003

X22 = Year Of 2004

x23: Longitude or x location (kilometer)
x24 = Latitude or y location (kilometer)

OOOOOOOOOOOOOOOOOOOOOOOO

Note: Undisturbed forest and the year of 1992 were used as omitted variables for
the land use classes and annual changes, respectively.

By using this multiple regression analysis, I was able to measure the effect
of each land use or year of analysis on forest canopy cover in the study area,
holding the other independent variables fixed.
2.4.3.1 . Testing Hypotheses

Based on field observations, I expected that impacts by any anthropogenic
activities should have effects on forest canopy coverage. Therefore, any
particular land use and land cover type should affect (increase or decrease)
values of fractional coverage when compared to undisturbed forest. I also

expected that effects of non-controlled seasonal and natural variations (e.g. El

50

Nino effects, extreme droughts, etc.) would impact values of fractional coverage.
Finally, I expected that values of fractional coverage could be affected by
uncontrolled spatial variability of natural forests (e.g. response of vegetation
phenology to precipitation) and human activities in the study area. I hypothesized

about each value of twenty four slope coefficients ( ,6, , 62, ,624), once the

effect of other coefficients had been account for, and used statistical inferences
to test these hypotheses as following:

Ho: Null hypothesis:
Land use and land cover change variables ( ,6, to 6,0):

,6] = 0: Deforestation has no effect on fractional coverage.

,62 = 0: Newly logged forest has no effect on fractional coverage.

,63 = 0: Old logged forest detectable has no effect on fractional coverage.

,64 = 0: Old logged forest non-detectable has no effect on fractional
coverage.

,65 = 0: Newly logged and burned forest has no effect on fractional coverage.

,66 = 0: Newly burned forest has no effect on fractional coverage.

,67 = 0: Old logged and burned detectable forest has no effect on fractional
coverage.

,68 = 0: Old logged and burned non-detectable forest has no effect on
fractional coverage.

,69 = 0: Old burned forest detectable has no effect on fractional coverage.

.310 = 0: Old burned forest non-detectable has no effect on fractional
coverage.

Temporal variables (,6l1 to ,822 ):

,8”: 0: Non-controlled seasonal or temporal variations for 1993 have no
effect on fractional coverage.

,612= 0: Non-controlled seasonal or temporal variations for 1994 have no
effect on fractional coverage.

,613 = 0: Non-controlled seasonal or temporal variations for 1995 have no
effect on fractional coverage.

fl” = 0: Non-controlled seasonal or temporal variations for 1996 have no
effect on fractional coverage.

,615 = 0: Non-controlled seasonal or temporal variations for 1997 have no

effect on fractional coverage.

51

H1:

.316: 0: Non-controlled seasonal or temporal variations for 1998 have no

effect on fractional coverage.
,6” = 0: Non-controlled seasonal or temporal variations for 1999 have no

effect on fractional coverage.
,6”; = 0: Non-controlled seasonal or temporal variations for 2000 have no

effect on fractional coverage.
[319: 0: Non-controlled seasonal or temporal variations for 2001 have no

effect on fractional coverage.
,820= 0: Non—controlled seasonal or temporal variations for 2002 have no

effect on fractional coverage.

621: 0: Non-controlled seasonal or temporal variations for 2003 have no
effect on fractional coverage.

,322= 0: Non-controlled seasonal or temporal variations for 2004 have no

effect on fractional coverage.
Spatial variables ( ,623 to ,624):
,623 = 0: Anthropogenic and natural spatial variations over longitude (x)

direction have no effect on fractional coverage.
,624= 0: Anthropogenic and natural spatial variations over latitude (y)
direction have no effect on fractional coverage.

Alternative hypothesis:
Land use and land cover change variables ( ,6] to ,610 ):

,6] ¢ 0: Deforestation has effect on fractional coverage.

,62 t6 0: Newly logged forest has effect on fractional coverage.

,63 at 0: Old logged forest detectable has effect on fractional coverage.

,64 ¢ 0: Old logged forest non-detectable has effect on fractional coverage.
,65 ¢ 0: Newly logged and burned forest has effect on fractional coverage.
,66 at 0: Newly burned forest has effect on fractional coverage.

fl7 ¢ 0: Old logged and burned detectable forest has effect on fractional

coverage.
,68 ¢ 0: Old logged and burned non-detectable forest has effect on fractional
coverage.
fig at 0: Old burned forest detectable has effect on fractional coverage.

filo ¢ 0: Old burned forest non-detectable has effect on fractional coverage.

Temporal variables (,611 to ,622):
fl11¢ 0: Non-controlled seasonal and temporal variations for 1993 have effect

on fractional coverage.
,612 at 0: Non-controlled seasonal and temporal variations for 1994 have effect

on fractional coverage.

52

613150:
614560:
615% 0:
616;! 0:
61759 0:
613:5 0:
,619¢0:
flzoaéo:
1321’”):

fl22¢ OI

Non-controlled seasonal and temporal variations for 1995 have effect
on fractional coverage.
Non-controlled seasonal and temporal variations for 1996 have effect
on fractional coverage.
Non-controlled seasonal and temporal variations for 1997 have effect
on fractional coverage.
Non-controlled seasonal and temporal variations for 1998 have effect
on fractional coverage.
Non-controlled seasonal and temporal variations for 1999 have effect
on fractional coverage.
Non-controlled seasonal and temporal variations for 2000 have effect
on fractional coverage.
Non-controlled seasonal and temporal variations for 2001 have effect
on fractional coverage.
Non-controlled seasonal and temporal variations for 2002 have effect
on fractional coverage.
Non-controlled seasonal and temporal variations for 2003 have effect

on fractional coverage.
Non-controlled seasonal and temporal variations for 2004 have effect

on fractional coverage.

Spatial variables ( ,623 to ,624):
,823 at 0: Anthropogenic and natural spatial variations over longitude (x)

direction have effect on fractional coverage.

624:6 0: Anthropogenic and natural spatial variations over latitude (y)

direction have effect on fractional coverage.

2.4.3.2. Data sampling

Spatial autocorrelation can substantially increase the error term in the

at al. 2007)

regression model. Anselin (2002) suggested that these effects of spatial
autocorrelation can be minimized by systematically sampling the dataset.
Therefore, in this analysis, a vector file of 3 km x 3 km sample points distributed
over the study area was generated. 8y sampling each point at a fixed distance

from the others, potential contagion between neighboring cells is isolated (Arima

and the error term in the regression model due to spatial

53

autocorrelation in the dataset is reduced (Anselin 2002, Arima 2007). The study
area was systematically sampled by acquiring unique values for the dependent
and independent variables at each sampling point location, all derived from
remotely sensed data.

In addition, each cell location was identified by its longitude (x) and
latitude (y) locations in the study area. The cell x and y locations were used in the
regression model as an additional set of independent variables. This procedure
helps to identify spatial trends and also to mitigate potential effects of spatial
autocorrelation in the data population (Chomitz and Gray, 1996).

2.5. Results and Discussions
2.5.1. Relationships among various Vegetation lndices

In this analysis, the relationship among NDVI, MSAVI, GEMI, AFRI,
MSAVIaf, GEMI“, and GV retrieved from a Landsat image acquired in August 6,
2003 were tested under smoke and smoke free atmospheric conditions. Table
2.3 shows the relationship among these vegetation indices under clear sky for
the study area.

This correlation matrix demonstrates high correlation (r > 0.90) among all
vegetation indices under clear sky atmospheric condition. Although a perfect
correlation (r= 1) was not observed among different vegetation indices, as
observed by (Karnieli et al., 2001), these research results showed high
correlations between NDVI and AFRI (r = 0.99), GEMI and GEMI2_1(r= 0.98), and
MSAVI and MSAVIaf (r = 0.98), which are the modified vegetation indices that

use an empirical linear relationship observed by (Karnieli et al., 2001) to predict

54

red band reflectance from shortwave infrared at 2.1. High correlation (r2 0.98)
also was observed between GV and GEMI, CV and GEM|2.1, GV and MSAVI, and
GV and MSAVIaf.

Table 2.3. Correlation matrix among various vegetation indices under smoke-free

atmospheric condition for the study site

 

 

Vegetation
Index NDVI AFRI GEMI GEMIM MSAVI MSAVIg GV

NDVI 1
AFRI 0.99 1
GEMI 0.97 0.96 1
GEMI“ 0.93 0.94 0.98 1
MSAVI 0.96 0.94 0.98 0.96 1
MSAVIaf 0.93 0.94 0.97 0.97 0.98 1
GV 0.95 0.96 0.99 0.99 0.98 0.98 1

 

Correlation is significant at the 0.01 level (1-tailed). N= 2000. Bold values
represent special interest in this analysis.

Vegetation indices, however, are very sensitive to atmospheric conditions,
especially on visible bands, which can strongly affect their results (Karnieli et al.,
2001, Huete et al., 2003, Moura and Galvao, 2003). In this study, sensitivity of
each vegetation index to presence of aerosol was tested. Table 2.4
demonstrates the performance of various vegetation indices under smoke
condition in the study site.

In this case, high correlation among aerosols resistant vegetation indices
(AF RI, GEMI”, MSAVIaf, GV) and those more sensitive vegetation indices
(NDVI, GEMI, MSAVI) to atmospheric effects indicated low performance in
penetrating through the aerosols and depicting the vegetation cover under heavy
smoke atmospheric condition. Therefore, low correlation between them was

desired. Lowest correlation was observed for GEMI2,1and NDVI (r= 0.87),

55

MSAVIaf and NDVI (r= 0.87), and AFRI and NDVI (R=0.89). Relatively low
correlation also was observed between MSAVIaf and MSAVI (r= 0.90).
Table 2.4. Correlation matrix among various vegetation indices for areas in the

presence of heavy smoke in the study site.

 

 

Vegetation
Index NDVI AFRI GEMI GEMI“ MSAVI MSAVlg GV

NDVI 1
AFRI 0.89 1
GEMI 0.95 0.84 1
GEM|2,1 0.87 0.88 0.96 1
MSAVI 0.96 0.78 0.98 0.90 1
MSAVI,“ 0.87 0.87 0.94 0.97 0.90 1
GV 0.93 0.88 0.99 0.99 0.95 0.97 1

 

Pearson correlation is significant at the 0.01 level (1-tailed). N: 2000. Bold
values represent special interest in this analysis.

The Appendix 8.7 demonstrates MSAVI and MSAVIaf performances under
smoke condition. Although GV derived from SMA was highly correlated to
GEMI2_1 (r= 0.99) and MSAVIaf (r= 0.97), it was also highly correlated (r> 0.90)
with those more sensitive vegetation indices (NDVI, GEMI, and MSAVI) to
atmospheric influences, which indicated that GV was more sensitive than AFRI
and MSAVIaf to the presence of heavy smoke in this study area.

Low correlation (r < 0.90) also was observed between AFRI and all other
vegetation indices, which may indicate that each index has a different response
to atmospheric effects when compared to AFRI.

2.5.2. Optimum Vegetation Index

In addition to the previously discussed relationships among various

vegetation indices, empirical linear relationship between fractional coverage

values derived from each vegetation index and from forest canopy coverage

56

calculated using hemispherical photos acquired in the study site. A total of 423
hemispherical photos were used to calculate 47-canopy fraction points on the
ground, estimated based on the average of 9 hemispherical photos each point.
The hemispherical photos were located over the study area using GPS. Likewise,
47-canopy fraction points were calculated from NDVI, MSAVI, GEMI, GV, AFRI,
GEMI“, and MSAVIaf fractional coverage derived from Landsat image, based on
the average of a group of 9 pixels (window 3 x 3) each point.

The table 2.5 shows the statistical results of the linear relationship
between each fractional coverage image derived from various vegetation indices
and canopy coverage estimated using hemispherical photos acquired in the
study area.

Table 2.5. Empirical linear relationship between various vegetation indices and

forest canopy coverage based on field measurements in the study site

 

 

Vegetation R-Square Prob > F F(1, 45)
Index
MSAVI 0.84 0.000 236.83
GEMI 0.80 0.000 182.21
NDVI 0.67 0.000 92.26
MSAVI,“ 0.81 0.000 191.90
GEMI“ 0.78 0.000 155.79
AFRI 0.60 0.000 68.25
GV 0.76 0.000 144.24

 

N= 47. Bold values represent special interest in this analysis

MSAVI showed the highest R2 (0.84) indicating that 84% of the sample
variation in fractional coverage derived from ground measurements can be
explained by fractional coverage derived from MSAVI model. MSAVIaf and GEMI

showed slightly lower R2 (0.81 and 0.80, respectivelY). and NDVI, GEMI2_1,AFRI,

57

and GV showed R2 lower than 0.80. Indeed, a good performance by MSAVI had
been observed for this study region by (Wang et al., 2005). In that study, the
authors selected MSAVI as the optimum vegetation index to estimate fractional
coverage derived from ETM+ image using lkonos image as a ground truth for
validation. Any abnormal atmospheric condition, however, was not reported by
(Wang et al., 2005).

AFRI, GEM12_1, and MSAVlafpresented great potential to estimate
fractional coverage from Landsat imagery for this study area, if considered that
these indices fitted fairly well the fractional coverage derived from hemispherical
photos acquired in the field and showed good capability to penetrate the
atmosphere when aerosols is present.

Based on these previously discussed limitation and potential of various
vegetation indices, I selected MSAVIaf as the optimum vegetation index for this
study site given its good performance under clear sky for the study area, high
correlation with its non-modified version (MSAVI), slightly better fitting to field
measurements compared to other modified vegetation indices, and, most
importantly, its good performance under smoke conditions.

2.5.3. Accuracy assessment of burned forest detection technique

The accuracy assessment was conducted by comparing binary
classification (1 = burned forest, 0 = non-burned forest) from the burned forest
map derived from Landsat TM5 image with 120 randomly distributed points in
forested area, checked during the field observation within the study site. Both

remotely sensed and fieldwork data were acquired in June 2004. Kappa statistics

58

and standard Producer’s, Users, and overall accuracies of this burned forest
detection technique are presented in Table 2.6.

Table 2.6. Accuracy assessment results of detecting burned forests using non-
photosynthetic vegetation (N PV) fraction image derived from Spectral Mixture

Analysis (SMA) for the study area in 2004.

 

 

 

 

Land Use Reference Classified Number Accuracy
Points (*) points Correct Producer User
Burned forests 79 80 77
non-burned forests 41 40 38 92.7% 95.0%
Overall accuracy = 96.0%

Overall Kapa Statistic = 0.91
* 120 randomly distributed sampling points within forested area

 

The commission and omission errors were 7.3% and 5%, respectively.
The accuracy assessment results demonstrated that this remote sensing
approach could correctly classify 96% of burned forest if the source of error
(commission and omission errors) is not accounted for. Appendix 8.8 shows a
sample of burned forests mapped using this remote sensing approach.
2.5.4. Multi-annual land use and land cover change

Based on these research results, it was observed that deforestation
increased from approximately 3400 km2 by 1992 to 8500 km2 by 2004, an
average increase of 426 km2 yr" (:1: 202 kmz, standard deviation). The greatest
deforestation increase, however, was observed between 2003 and 2004, at 980
km2. The total disturbed forests within the study region (path 226 and row 068)
increased from 5.4% to 40.1%, by 1992 and by 2004, respectively. These forest
disturbances correspond to an annual increase of 1,270 km2 yr’1 (i482 km2) of

forests selectively logged and/or burned. Only 30.5% of the study area was

59

considered undisturbed forest by 2004. Table 2.7 shows the annual land use and
land cover change over the study site.
Table 2.7. Cumulative deforestation and undisturbed and disturbed forests in the

study area from 1992 to 2004.

 

Logged &

 

Undisturbed Logged Burned forest
Year forest forest only Burned forest only Deforestation
(km2) (km2) (km2) (km2) (km2)
1992 24,1315 1,120.9 45.1 395.7 3,403.6
1993 22,4721 2,243.7 56.1 424.5 3,900.3
1994 21 ,441 .7 2,911.7 70.6 426.5 4,246.4
1995 20,3781 3,441.9 144.9 468.6 4,663.3
1996 19,2849 4,142.8 176.5 455.8 5,036.8
1997 17,9098 5,089.3 181.3 495.6 5,420.8
1998 16,5548 6,069.1 246.1 493.3 5,733.4
1999 ' 14,1606 7,522.0 486.9 705.8 6,221.4
2000 12,5159 7,560.2 1,491.7 1,171.3 6,357.7
2001 11,1287 8,553.0 1,508.5 1,062.6 6,844.0
2002 10,4222 8,975.8 1,553.2 1,028.1 7,117.5
2003 9,606.4 9,440.2 1,570.0 939.2 7,541.0
2004 8,894.1 9,274.7 1,549.3 857.8 8,520.7

 

A mask of 972 km2 of the cumulative clouds and shadows was applied.
Approximately 36.3 km2 of water bodies that were annually detected on all
scenes were not included.

In this given period, selective logging alone was responsible for disturbing
a large extent of natural forests, equivalent to more than 31 % of this study area
by 2004. An additional of 5.5% and 2.9% of the study site forest was disturbed by
selective logging and fire combined and forest fire only, respectively. By 2002,
the total area of disturbed forests already had surpassed by more than 1,100 km2
the total area of undisturbed forest in this study site. Figure 2.4 shows the annual
change of undisturbed into disturbed forests and the forest conversion into

agricultural lands (deforestation) in the study region.

60

 

30.000

25,000

 

20,000 ' ,

 

15.000 -

 

Area (km2)

10,000 «

5,000 ,

 

 

 

1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004

Year

 

 

I; Undisturbed forest I Disturbed forest I Deforestatiofl

 

 

Figure 2.4. Land use and land cover change from 1992 to 2004 in the study site

These results also show that more than 10,000 km2 of remnant logged
and burned forests had recovered sufficiently to be undetected (non-persistent)
by 2004 using Landsat imagery. Appendix 8.9 shows annual changes of
persistent and not persistent disturbed forests by selective logging and forest fire
in this study area.

New areas of selective logging only increased around of 904 km2 yr"
(i354 km2) from 1992 to 2004 in this study area. An annual average of 422 km2
yr" (i440 km2) of logged forests were burned, and around of 426 (1:202 km2) and
102 km2 yr" (i152 km2) of undisturbed forests were deforested and burned,
respectively.

The greatest annual increase (around of 1728 km2) of selective logging

only occurred in 1999. Immediately following this year, it was observed a

61

substantial increase (around of 1370 km2) of newly burned forests over logged
forest by 2000. Table 2.8 shows the annual increase of selectively logged only,
selectively logged and burned forests combined, and burned forest only, from
1992 to 2004 in the study site.

Table 2.8. Annual increase of selectively logged forest, burned and logged

forests combined, and burned forest only, from 1992 to 2004 in the study area.

 

Annual Increase (km2)

 

 

Logged Burned & Burned & Burned Deforestation
Year forest newly logged Old Logged forest
hma mma

1992 1,120.9 45.1 - 395.7

1993 1,158.1 14.0 0.8 88.7 496.7
1994 708.9 11.1 1.6 22.5 346.0
1995 634.2 42.0 25.9 70.2 416.9
1996 785.2 17.2 23.1 23.6 373.5
1997 1,022.3 11.6 0.5 66.9 384.0
1998 1,072.0 41.0 11.9 38.9 312.6
1999 1,727.9 111.0 145.4 249.4 488.0
2000 873.1 296.8 721.0 539.5 136.3
2001 1,116.6 64.4 6.3 10.5 486.3
2002 540.5 34.9 14.5 50.2 273.5
2003 834.8 22.2 37.2 33.9 423.5
2004 375.7 10.2 63.2 24.1 979.7

 

The analysis of the land use dynamics over the study site showed that
previously logged and burned sites could be revisited by loggers and recurrent
forest fires, respectively. Revisited logging accounted for more than 3300 km2 by
2004, an average of 276 km2 yr" (1197 km?) from 1992 to 2004. Recurrent fires
accounted for 866 kmz, an average of 72 km2 yr" (:I:50 km2) in the same interval.
It was also observed that more than 1756 km2 of selectively logged forests and

an additional 368 km2 of burned forests had been deforested by 2004.

62

Table 2.9. Multi-annual land use dynamics in the study site from 1992 to 2004

 

 

 

 

Revisited Recurrent Disturbed forests deforested by 2004

Year logging fire 8y logging By logging By Fire only
(km2) (km2) only (km2) & Fire (km ) (km2)

1992 372.9 25.1 211.5
1993 0 0 337.2 1.4 46.4
1994 32 61.8 307.2 0.8 9.2
1995 51.4 54.95 191.2 15.6 44.5
1996 101.6 37.2 154.8 3.3 12.8
1997 148.9 82.58 135.9 3.0 19.2
1998 319.3 74.78 175.5 10.1 20.6
1999 431.7 181.52 137.4 16.2 68.1
2000 491.6 92.43 85.6 44.5 134.9
2001 394.5 23.34 70.4 2.2 2.6
2002 455.3 73.71 - - -
2003 537.3 142.09 62.4 1.9 10.2
2004 350.1 41.63 - - -
Total 3,313.7 866.0 1,657.7 99.0 368.3

 

2.5.5. Assessment of forest canopy cover impacts

In this analysis, I was primarily concerned in assessing impacts of
temporal and land use and land cover changes on forest canopy cover.
Therefore, multiple regression analysis was applied to measure partial effect of
each of factor of interest, holding fixed all other factors affecting the fractional
coverage. The results of this analysis showed that the multiple regression
equation (16) can significantly explain 63.07% of the dependent variable
(fractional coverage) at a = 0.01 (probability of 99%). Additionally, the hypothesis
testing for measuring partial effect of each land use and land cover on fractional
coverage was conducted as following.

I could not reject the Null Hypothesis for the old burned forest non-
detectable, year of 2004, and longitude (x) direction variables. More specifically,

the fractional coverage values for old burned forest non—detectable and

63

undisturbed forest (the land use omitted variable) derived from Landsat imagery
were not significantly different, at q = 0.01. Similarly, fractional coverage values
for the years of 2004 and 1992 (the temporal omitted variable) also were not
significantly different, at a = 0.01. Fractional coverage values did not significantly
change according to longitude (x) location, at a = 0.01, which means that there is
no spatial trend in east-west direction. I could reject, however, the null
hypotheses for all other tested coefficients and accepted the alternative
hypotheses that there were statistically significant differences between fractional
coverage values of each coefficient and its corresponding omitted variable
(undisturbed forest or year of 1992).

Based on these accepted alternative hypotheses, partial effect of each
factor of interest was assessed. Water body significantly decreased 61.9% (:I:
3.6), at a = 0.01, of fractional coverage values compared to fractional coverage
for undisturbed forests (the land use omitted variable). Indeed, vegetation indices
are very low for water bodies due to low reflectance on infrared wavelengths
(Wang et al., 2004). The effects of deforestation were the second greatest
contributor for forest canopy loss, decreasing 53.5% (:I: 0.8) of forest canopy
values, at a = 0.01, compared to fractional coverage values for undisturbed
forests, holding all other factors (independent variable) fixed. Indeed,
deforestation implies in cutting down, damaging, and burning forests, converting
natural forests into other land use type Skole and Tucker (1993) and FRA (2005)

where forest canopy cover is drastically reduced (FRA, 2005).

64

In terms of forest uses, newly logged forest only significantly decreased
5.0% (:I: 0.8), at q = 0.01, forest canopy in the study area compared to fractional
coverage for undisturbed forests. On the other hand, newly burned forest
significantly decreased 28.9% (:I: 2.0), at d = 0.01, forest canopy and, when
combined selective logging and forest fires, decreased 35.6% (1 2.1) forest
canopy compared to fractional coverage for undisturbed forests. Figure 2.5
shows the estimated losses and gains of canopy coverage based on the multiple
regression model of fractional coverage for various forest uses classified in the
study site using Landsat imagery acquired from 1992 to 2004.

The alternative hypothesis that old (previously) logged forests detectable
using remotely sensed data has effect on fractional coverage also was accepted.
In this case, old logged forest significantly contributed to decrease only 1.6% (i
0.7), at a = 0.01, of fractional coverage values compared to those for undisturbed
forest. On the other hand, old logged forests that recovered sufficiently to
become undetectable using Landsat images, accounted for a significant increase
of 3.1 (:I: 0.5), at a = 0.01, of fractional coverage values, compared to undisturbed
forests. These results indicate that strong forest regeneration occurred and
contributed to increase vegetation greenness as well as fractional coverage
values, surpassing even fractional coverage value for undisturbed forests in this
study area.

Old burned forests detectable using remotely sensed data showed a
significant decrease of 12.6% (:I: 1.5), at a = 0.01, of fractional coverage values

and, when not detectable, showed no significant change in canopy cover. In this

65

last case, the null hypothesis that old burned forest has no effect on fractional
coverage could not be rejected. And again, old areas of detectable selective
logging and forest fire combined showed significant decrease of 25.6% (1 2.0), at
d = 0.01, of fractional coverage values, and even when not detectable, still
showed significant decrease of 4.7% (:t 1.1), at a = 0.01, of fractional coverage

values compared to undisturbed forests.

 

 

Forest canopy change (%)

 

j {i Newly” logged 7' Dad logged forestidetectable A "7 7 ‘
1 l Old logged forest non-detectable I Newly logged & bLmed forest 1

I I Newly buried l Old buned forest detectable
l l Old bumed forest non-detectable Old logged & bu'ned forest detectable
II Old logged 8r buned forest non-detectable

 

 

Figure 2.5. Partial effects of various forest uses on fractional coverage values in
the study area estimated using linear multi-regression model and multi-annual
remotely sensed derived dataset

As previously stated, in terms of temporal change, all years from 1993 to
2003 showed significant change of fractional coverage values at a = 0.01
compared to 1992 (omitted variable) and the alternative hypotheses that non-

controlled seasonal and temporal variations significantly affected fractional

66

coverage values. The greatest annual losses of fractional coverage, 8.8% (:I: 0.7)
and 8.4% (:I: 0.7), were observed for 1993 and 1999, respectively. Based on
these results, the average around of 3.6% and 6.4% of forest canopy losses
were observed for 1994-1998 and for 1999-2003 time intervals, respectively.

The alternative hypothesis that anthropogenic and natural spatial
variations over latitude (y direction) have effect on fractional coverage was
accepted. These variations significantly increased 3.7% (:I: 0.3), at a = 0.01, of
fractional coverage values for each 100 kilometer towards North. On the other
hand, the null hypothesis could not be rejected for changes in longitude (x
direction), which means that there is no significant change, at a = 0.01, of
fractional coverage values when moving towards east or west direction, holding
fixed all other factors affecting fractional coverage values. Appendix 8.10 shows
further results of the linear multi-regression analysis.

It was also observed that forest regeneration increases fractional
coverage in the year immediately following selective logging activities and forest
fires. The peak of this forest regeneration was observed 3 years following the
harvesting year for logging and 5 years following forest fire.

2.6. Conclusions

Modified vegetation indices AFRI and MSAVIaf showed good performance
in predicting forest canopy cover under smoke condition. AF RI and MSAVIaf also
showed high correlation with NDVI and MSAVI, respectively. However, it was
observed that fractional coverage derived from MSAVI and MSAVIaf predicted

fractional coverage slightly better than the other tested vegetation indices when

67

compared to field measurements in this study area. Additionally, fractional
coverage, as a linear unmixing model in vegetation index domain retrieved from
Landsat TM and ETM+ imagery, showed to be useful in measuring forest canopy
degradation and regeneration in tropical evergreen forests.

The land use and land cover dramatically changed from 1992 to 2004 in
this study area. More than 29% of the study area, approximately 8,521 kmz, was
converted into agricultural land by 2004. More than 19%, 4%, and 1% of the
deforested area was previously detected as logged forest, burned forest, and
logged and burned forest, respectively, in the 1992-04 time period. In addition to
outright deforestation, forests disturbed by selective logging and forest fires
represented approximately 40% of this study area by that year. Assuming that
this observed trend of land use and land cover conversion continues, there will
be no undisturbed forests remaining by 2011.

In this context, selective logging only is by far the most important factor of
forest disturbances of large extent in the study site. Selective logging only was
responsible for the disturbance of more than 9,200 km2 of natural forests by
2004, which corresponds to more than 31% of the total study area. Indeed, large
areas of tropical forests have been damaged or impoverished each year by
selective logging activities in the Brazilian Amazon (Uhl and Vieira, 1989,
Verissimo et al., 1995, Houghton, 1997, Nepstad et al., 1999). Roughly 2350 km2
were degraded by forest fire or a combination of forest fires and selective logging
by 2004. This corresponds to less than 26% of the total forest disturbed by

selective logging only.

68

During this given period of analysis, there was a large increase of revisited
logging and recurrent forest fires. In this study area, more than 3,300 km2 and
800 km2 of cumulative area of revisited logging and recurrent fire was observed
by 2004, respectively. I defined revisit logging and recurrent fire as logging and
burned forest detected in areas of previous logging and forest fire, respectively,
that had recovered sufficiently to become undetectable for one or more years
prior to the return of the loggers or forest fire. An annual average of 27.8 km2 (1:
26), 255.4 km2 (:I: 152.8), and 445.8 km2 (1 74.8) of revisit logging was observed
for 1992-95, 1996-99, and 2000-04, respectively. An annual average of 38.9 km2
(3: 33.9), 38.5 km2 (s 61.6), and 59.1 km2 (s 46.3) of recurrent fire was observed
for 1992-95, 1996-99, and 2000-04, respectively. Recurrent forest fire is likely to
increase even more in the next years in this region. According to (Cochrane and
Schulze, 1998), once burned for the first time, chances of recurring fires are
increased as well.

Based on the results of this multi-temporal analysis of fractional coverage
derived from Landsat TM and ETM+ imagery, new areas of selective logging only
decreased fractional coverage by an average of 4.8% in the study area. Newly
burned forests alone and selectively logged and burned forests combined
decreased an additional 24% and 31 % of forest canopy cover, respectively,
compared to newly logged forests only. Old areas of selective logging and forest
fire combined that were still detectable, also showed a substantial decrease of
25.1% of fractional coverage values compared to undisturbed forests. A

fractional coverage decrease of 4.3% was observed for old selectively logged

69

and burned forest even after they become undetectable on Landsat images.
Although forest fires disturbed a smaller extent of natural forest than selective
logging in this study site, when it occurs, as observed by Uhl and Kauffman
(1990), its impacts can be devastating for tropical rainforests.

Fractional coverage values for old selectively logged forests were 3.3%
greater than fractional coverage values for undisturbed forests. These previously
selectively logged forests accounted for more than 60% of the total disturbed
forest in the study site by 2004. This finding implies that selective logging may be
a carbon source immediately following its activities but it may become a carbon
sink in the subsequent years. Therefore, further research is required to
accurately infer how much selective logging has been contributing to the carbon
cycle either as source or as sink in tropical areas. This amount may vary
according to forest characteristics, logging and fire intensity, and forest capacity
of regeneration of each site.

These research results indicate that forests are capable of a strong
canopy regeneration following logging activities and forest fires. Approximately
70% of the total area of disturbed forests by logging and fire had sufficiently
recovered to become undetectable on satellite imagery by 2004. Forest
regeneration, however, may vary according to logging and fire intensity, but in
general it takes more time if burned. Selectively logged forests can be detected
using Landsat imagery from 3 to 5 years following logging activities and burned
forests may persist detectable from 3 to 10 years, depending on fire intensity and

forest regeneration capacity.

70

CHAPTER III
Basin-Wide Assessment of Forest Disturbances
by Selective Logging and Forest Fires
3.1. Abstract
The rapid and drastic environmental changes happening in the Brazilian Amazon
because of widespread deforestation have attracted the attention of the scientific
community for several decades. An area of particular interest involves the
assessment of the impacts of selective logging and forest fires. Forest
disturbances by selective logging and forest fires may vary in scale, from local
damages to the local environment (microclimate, habitats destruction, species
extirpation, etc.), to global changes, mostly related to the increase of atmospheric
carbon dioxide released into the atmosphere. Selective logging activities and
forest fires have been reported by many studies as important agents of land use
and land cover changes but yet have to be properly addressed, especially at
large scale in tropical regions. Previous studies have focused on a few study
sites or were limited to estimate the extent of selective logging. Therefore, this
study involved a more comprehensive study of multi-temporal and basin-wide
changes of forest disturbances by selective logging and forest fires using
remotely sensed data acquired in 1992, 1996, and 1999. Methods for detecting
burned forests and estimating forest canopy cover were researched. Finally, I
conducted rigorous ground measurements and observations for assessing local
impacts of selective logging and forest fires in three different states in the

Brazilian Amazon. The results of this study showed a substantial increase in total

71

areas of selectively logged and burned forests, changing from approximately
11800 km2 to 35600 km2 by 1992 and 1999, respectively. Selective logging only
was responsible for 60.4% of this forest disturbance in the studied period.
Approximately 33% and 7% of forest disturbances detected in the same period
were due to impacts of forest fire only and selective logging and forest fire
combined, respectively. Most of the degraded forests (~90%) were detected in
the States of Mato Grosso and Para. My estimates indicate that approximately
5467 kmz, 7618 kmz, and 17437 km2 were new areas of selective logging and/or
forest fires in 1992, 1996, and 1999, respectively. Protected areas seemed to be
very effective in constraining these types of forest degradation. Approximately
2.4% and 1.3% of the total detected selectively logged and burned forests,
respectively, were geographically located within protected areas. I observed,
however, an increasing trend for these anthropogenic activities to occur within
the limits of protected areas from 1992 to 1999. Although forest fires impacted
the least extent of natural forests, new areas of burned forests detected in 1996
and 1999 were responsible for the greatest impact on canopy cover, with an
estimated loss of 18.8% of forest canopy when compared to undisturbed forests.
Selective logging and forest fire combined impacted even more those canopies,
with an estimated loss of 27.5% of forest canopy. Selectively logged forest only
showed the least impact on canopy cover, with an estimated loss of 5% of forest
canopy. Field measurements indicated a significant positive correlation between
forest canopy openness and tree mortality rate and dead basal area, and a

negative correlation between canopy openness and tree diversity. Selectively

72

logged forests had lesser impacts on tree species diversity than did those burned
forests. Selectively logged forests also showed stronger and faster forest
regeneration than burned forests, following these forest impact events.
3.2. Introduction

Natural forests in the Brazilian Amazon have been increasingly exposed to
selective logging activities and forest fires (Nepstad et al., 1999, Cochrane et al.,
2004, Asner et al., 2005, Matricardi et al., 2005, Matricardi et al., 2007). In
contrast to deforestation that converts large extents of natural forests into
agricultural lands, selective logging targets cutting down a few most marketable
tree species and leave behind of its activities highly disturbed forest patches in
which tree fall gaps, access roads, logging patios, and damaged trees are
abundant (Stone and Lefebvre, 1998, Nepstad et al., 1999, Souza and Barreto,
2000). Similarly to forest fragmentation that frequently occurs between the edges
of deforestation and forested areas, selective logging can increase forest fire
susceptibility as tropical forest becomes more fragmented (Cochrane, 2001).
Agricultural plots that frequently use fire as land management technique serve as
ignition sources and fragmented forests increase the chance of spreading forest
fire throughout the Amazon region (Uhl and Buschbacher, 1985, Uhl and
Kauffman, 1990, Cochrane, 2001). The consequences of forest fires, however,
can be devastating for those natural forests (Uhl and Kauffman, 1990).

Although selective logging and forest fires are considered important
anthropogenic activities in tropical regions, only a few attempts towards a

spatially explicit classification of both activities using remotely sensed data have

73

been conducted in the Brazilian Amazon. The first basin-wide selective logging
estimate was conducted by Nepstad et al. (1999) and it was based on a survey
of sawmills’ production in the Amazon. These authors estimated that 9,000 to
15,000 km2 of natural forest had been selectively logged in 1996-97 period.
Matricardi et al. (2007) detected and mapped selectively logged areas in the
Brazilian Amazon using Landsat imagery acquired in 1992, 1996, and 1999. In
that study, a total of 5980 kmz, 10064 kmz, and 28085 km2 of selectively logged
forests were detected by 1992, 1996, and 1999, respectively. Matricardi et al.
(2007) estimated an increase of new areas of selective logging in the Brazilian
Amazon around of 2547 kmz, 4381 kmz, 11889 km2 in 1992, 1996, and 1999,
respectively. A more recent multi-annual analysis of selective logging from 1999
to 2002 was conducted by Asner et al. (2005) in five major timber centers of the
Brazilian Amazon, using remotely sensed data. These authors observed an
increase around of 19800 kmz, 14200 kmz, and 12000 km2 of selectively logged
forests in 1999-00, 2000-01, 2001-02 periods, respectively.

In addition to selective logging activities, forest fires are increasingly
contributing to forest degradation in the Amazon region. Although fire is an
important agent of land use and land cover change, it has yet to be properly
assessed (Schroeder et al., 2005). Most large scale analyses of fire in the
Amazon region have been based on general information, manly focusing on the
geographic location of fires based on thermal bands and course spatial resolution
satellite products. Maps of fire locations or fire pixels have been generated in a

daily basis by the National Institute for Space Research (INPE) in Brazil. Most of

74

those maps are based on NOAA3 and MODIS4 products
(http://www.cptec.inpebr/queimadasl) and are available for the entire South
America continent. Schroeder et al. (2005) used the INPE products to analyze
spatiotemporal dynamics of fire in the Brazilian Amazon. These authors had to
integrate various datasets to improve spatial scale and reduce analytical
uncertainties caused by the use of a single sensor. Complementary and more
detailed researches have been restricted to small study areas (e.g. Cochrane
and Schulze, 1998, Souza et al., 2005) and are not sufficient for a more
comprehensive analysis of forest fires (Schroeder et al., 2005) and selective
logging in the Amazon region (Matricardi et al., 2007).

Based on the existing knowledge gaps, I conducted a comprehensive
research of forest degradation due to selective logging activities and forest fires
based on remotely sensed data and ground measurements in the Brazilian
Amazon. In addition to employing the selective logging dataset created by
Matricardi et al. (2007), I developed and applied a fully automated remote
sensing approach based on SMA (Spectral Mixture Analysis) to detect and map
burned forests basin-wide. Fractional coverage was retrieved from Landsat
imagery and used to assess forest canopy impacts by selective logging and
forest fires. This analysis was conducted on a time-series of Landsat TM and
ETM+ imagery acquired in 1992, 1996, and 1999. Field measurements and

observations included three different regions in the Amazon States of Acre,

 

3 NOAA: A series of Geostationary and Polar-Orbiting Weather Satellites operated by the
National Oceanic and Atmospheric Administration, USA;

‘ MODIS: Moderate Resolution Imaging Spectroradiometer sensor aboard of Terra and Aqua
satellites, both operated by the National Aeronautics and Space Administration, USA.

75

Rond6nia, and Mato Grosso. The field information was used to assess forest
impacts caused by selective logging and forest fires on forested areas and to
validate remote sensing approaches. The results of this study will contribute to
improve land use classification maps and to expand current understanding of the
characteristics and interactions between forest fire and selective logging, and fire
and other land management practices in the study region. The applied use of
these results by governmental and non-governmental organizations may
contribute to improve sustainable forest management, environmental monitoring,
and law enforcement activities in the Brazilian Amazon.
3.3. Methods
3.3.1. Regional setting

The study region was Brazil’s Legal Amazon, a territory with
approximately 5 million kmz. The so-called Legal Amazon includes the totality of
Acre, Amapa, Amazonas, Para, Rondcnia, and Roraima States, plus parts of
Mato Grosso, Maranhao, and Tocantins (see Appendix C1). The Legal Amazon
is a political-administrative region in Brazil, created by the Brazilian Federal Law
1806 in 1953. This analysis was conducted using more than 600 Landsat images
of the Legal Amazon, previously used for deforestation analysis at the Tropical
Rain Forest Information Center (T RF IC), of the Michigan State University. Signs
of selective logging and forest fire were searched throughout the study region.
Matricardi et al. (2007) found evidences of selective logging in 31 Landsat TM
scenes acquired in 1992 and 1996, and in 38 Landsat ETM+ scenes acquired in

1999. Evidence of forest fires also was found on these scenes and,

76

complementarily, for 1992 and 1996, additional 6 scenes acquired in 1992 and
1996 were used and for 1999, additional 7 scenes acquired in 1999 were used.
These Landsat scenes defined the subsequent effective areas of study used in
this analysis (Appendix 01).
3.3.2. Dataset
3.3.2.1. Deforestation dataset

The Brazilian Amazon is one of the regions covered by the deforestation
monitoring project implemented by the Tropical Rain Forest lnforrnation Center
(httpzllwwwtrficmsuedu) at Michigan State University, which involved a multi-
annual classification of 229 Landsat scenes. In that study, Landsat images were
individually classified into seven thematic classes: forest, deforestation, regrowth,
cenedo 5, cloud, cloud shadow, and water using unsupervised image
classification procedures. Classified Landsat scenes (path/row) were available
individually and placed together as deforestation mosaics for the entire Brazilian
Amazon. These deforestation mosaics preserved the original spatial resolution
(30 meters). Additional details of the deforestation dataset are presented in Table
3.1.

Selectively logged and burned forests were not included in these land use
and land cover classes. Both individual Landsat scenes and deforestation
mosaics for 1992, 1996, and 1999, produced by the TRFIC at Michigan State

University, were used to perform the intended analytical procedures.

 

5 Cerrado is a vegetation type similar to African Savannah.

77

Table 3.1 Deforestation dataset characteristics

 

 

 

Dataset Features Format Spatial Period of Projection
resolution Analysis
Land use Forest, Arc/Info 30m x 30m 1992, UTM,
and land deforestation, grid 1996, and Zones 19
cover GIS secondary 1999 through 23.
layers by regrowth, Projection,
Landsat cerrado, clouds, Datum, and
scene and shadows Spheroid
(path/row) WGSB4
Land use Forest, Arc/Info 30m x 30m 1992, Sinusoidal
and land deforestation, vector 1996, and projection,
cover secondary and grid 1999 Datum and
mosaic regrowth, Spheroid
cerrado, clouds, WGS84

and shadows

3.3.2.2. Selective logging dataset

Unsupervised classification has frequently been applied to assess land
use and land cover change in the Brazilian Amazon. This procedure does not
detect most of the selectively logged forests. As a result, additional remote
sensing techniques were required to accurately detect and map them.

Matricardi et al. (2007) conducted a multi-temporal analysis of selective
logging of the entire Brazilian Amazon. In that study, forests showing evidences
of selective logging were mapped and estimated for up to 38 Landsat scenes
throughout the Brazilian Amazon for 1992, 1996 and 1999. The authors used
semi-automated (texture algorithm) analysis of Landsat band 5 to detect log
landings (i.e. log storage patios) and, subsequently, applied variable buffer zones
(180 and 450 meters) around patios to estimate areas affected by selective
logging. In addition, Matricardi et al. (2007) mapped selectively logged forests
with visible canopy disturbance on Landsat images through visual interpretation,

and circumscribed them digitally on a computer screen. In that analysis, visual

78

interpretation and semi-automated method could detect 83.4% and 63.7% of the
total logged forests, respectively. Common areas between these two techniques
were around 50% but they were not double—counted. Areas that had been
previously missed using automatic analysis were added by visual interpretation
and vice-versa. Moreover, semi-automated and visual interpretation technique
combined showed 92.9% and 0.82 of overall accuracy and overall kappa

statistic, respectively. Concomitantly, a deforestation dataset was produced for
1992, 1996, and 1999 and composed by individual Landsat scenes and mosaics
for the entire region. Characteristics of the logging dataset are presented in Table
3.2.

Table 3.2 Selective Logging dataset characteristics

 

 

 

Dataset Classification Format Spatial Period of Projection
Features resolution Analysis
Selectively Binary classes Arc/Info 30m x 30m 1992, UTM,
logged forest (1 = logging, grid 1996, Zones 19,
GIS layers 0 = no data) and through 23.
(Landsat 1999 Datum and
path/row) Spheroid
WGS84

Selectively Binary classes Arc/Info 30m x 30m 1992, Sinusoidal,
logged forest (1 = logging, grid 1996, Datum and
GIS layers 0 = no data) and Spheroid
(mosaic) 1999 WGS84

 

The present study included the previously described selective logging
dataset, involving multi—annual measurements for1992, 1996, and 1999 of
selectively logged forests in ten'a-fimie (upland) of the Brazilian Amazon. In
addition, information derived from a multi-annual analysis of selective logging

from 1992 to 2002 for a case study (Landsat scene path226 and row 068) in the

79

State of Mato Grosso, conducted by Matricardi et al. (2005), was used to
estimate annual increase of selectively logged forests in the study region.
3.3.2.3. Landsat imagery

The Landsat imagery used in this analysis were drawn from the Tropical
Rain Forest Information Center (httpzllwwwtrficmsuedu) at Global Observatory
for Ecosystem Services, Michigan State University. For 1992 and 1996, 38
Landsat TM5 images were used for the analyses, and for 1999, 46 Landsat
ETM+ images were used. Three years (1992, 1996, and 1999) were used as
reference for measurements and analytical purposes. Some Landsat images,
however, were acquired on previous or following year of reference, according to
remotely sensed data availability and quality. Acquisition dates of each Landsat
scene used in this analysis are provided in Appendix 02.
3.3.2.3.1. Radiometric correction

Remotely sensed data is affected by changes in the instrument over time.
In this case, radiometric correction can be used to normalize the sensor
degradation to generate products suitable for multi-temporal analysis, especially
when the remotely sensed data is derived from different sensors and acquisition
dates (Chander and Markham, 2003). Therefore, radiometric correction was
conducted for all Landsat scenes used in this analysis. Further details on this
correction procedure were shown in Chapter 2.
3.3.2.3.2. Landsat imagery normalization

Multi-temporal analysis using satellite imagery may be affected by

seasonal variations of landscape components such as soil and vegetation

80

characteristics (Mas, 1999). To prevent this sort of distortion, image
normalization may be applied to reduce seasonality effects and atmospheric and
sensor variations (Elvidge et al., 1995).

Image normalization was applied for all Landsat scenes used in this
analysis. A good quality Landsat image, p226/r068 acquired in 1999, was
selected as reference. As suggested by Elvidge et al. (1995), the brightest and
darkest pixels or sample points were selected from the reference Landsat image
and from the corresponding scene acquired in 1996. Sample points consisted of
areas of undisturbed forests, soil exposure, and deep water bodies. Around 20
sample points of reflectance values from the reference image (dependent
variable) and from the corresponding image acquired in 1996 (independent
variable) were used to generate coefficients for a linear regression model.
Normalization coefficients (slope and intercept) for each band of interest were
produced. The acceptable R2 for each equation was greater than 0.95, and,
when needed, additional sample points were acquired to increase it. The

equation 1 describes the linear regression model used for image normalization.
y = ,50 + 31951 (1)

where y is the reference image reflectance (dependent variable), ,60 is the

intercept or constant coefficient, ,6, is the parameter associated with x,, and xl is

the reflectance of the normalizing image (independent variable).
The estimated coefficients and Landsat image acquired in 1996
(independent variable) were used as input to the equation 1, which generated a

spectrally adjusted (normalized) image to the reference image. The same

81

procedure was applied to normalize the Landsat image (p226/r068) acquired in
1 992.

Common areas (overlaps) observed between Landsat scenes were used
to obtain reflectance values of brightest and darkest pixels. In this case, each
previously normalized image served as a reference image for its neighboring
scene. This procedure was applied for all images acquired In 1999.
Consequently, the normalized Landsat imagery acquired in 1999 were used as
reference to normalize those acquired in 1992 and 1996.
3.3.2.3.3. Geometric correction

The Landsat imagery used in this analysis was previously system-
corrected at the TRFIC using sensor geometry and a polynomial distortion
model. The root-mean-square (RMS) positional errors resulting from this
geometric correction varied from 0.20 to 0.47 pixels. Additionally, ground control
points at known locations and intersections of road segments acquired during
fieldwork in 2004 were used to test geometric accuracy of three Landsat scenes
(p002/r067, p232/r067, and p226/068) located in the States of Acre, Rondbnia,
and Mato Grosso. The measurements between these three Landsat scenes and
the ground control locations averaged less than 36 meters :I: 31.4 meters and 32
meters 1: 37.1 meters for x and y directions, respectively.

3.3.3. Field study

The fieldwork observations and measurements were conducted during

June 2003, June and July 2004 in study sites in the States of Acre, Rondonia,

and Mato Grosso (Appendix 03). Various types of selective logging (ongoing

82

and previous logging), logged and burned forests, and undisturbed forests were
sampled across the study sites. The study site locations and characteristics are
described in Appendix 04. The field studies were designed to enhance the
accuracy assessment of remote sensing approaches and to support further
analysis of forest disturbances.

3.3.3.1. Field sampling and measurement

Field survey sampling was necessary to estimate the potential biological
composition changes of several forest use types (undisturbed, selective logging,
and forest fire). Fuel load such as dead trees, wood debris, and litter were also
quantified in each study site. Moreover, various forest degradation types and
intensities were identified and sampled (Appendix 04). Therefore, this survey
included simultaneous measurements of fuel load and forest composition and
regeneration assessments.

A sampling design scheme developed by Cochrane and Schulze (1998)
was used to guarantee accuracy and precision of the field survey. As a result, the
sample unit was the same for all sampled sites (Appendix C.5). Twenty transects
(approximately 9.4 hectares) were plotted and measured in Acre, Rond6nia, and
Mato Grosso. Each sampled site was geographically located using a GPS. The
sample unit on the ground consisted of one transect of 500 x 10 m in which the
number of trees, tree identification, tree condition (i.e. dead, alive, rotten, or
broken), and DBH measurement of trees greater than 10 cm DBH were
recorded. Within the main transect, a quadrat of 10 x 10 m was also plotted

every 25 m. Each 10 x 10 m quadrat contained a nested sub-quadrat of 5 x 5 m

83

for sampling tree regeneration and trees within two different diameter classes (<
5 cm and 5-10 cm). Secondary regrowth species and biomass (i.e. dry, rotten,
alive, broken, etc.) were identified within the sub-quadrats (DBH < 10 cm) and
main transects (DBH >10 cm). Additionally, along each main transect a 10 m fuel
line was plotted every 25 m. Each fuel line started at the center of every quadrat
(Q1, Q2, Q3, or Q4) towards a randomized direction (0° to 90°, 90° to 180°, 180°
to 270°, and 270° to 360° for Q1, Q2, Q3, and Q4, respectively). Subsequently,
litter depth was measured at the beginning, middle, and end of the line. In
addition, the remaining biomass under the last meter, under the last three
meters, and under the entire line for classes of diameter less than 2.5 cm,
between 2.5 and 7.6 cm, and greater than 7.6 cm were also counted.

Finally, hemispherical photos were acquired every 10 m along the main
transect line. These photos were used to measure canopy openness in each
location.

3.3.3.2. Identification of species

Identification of tree species was one of the most difficult tasks during the
field survey. In theory, plant components could be collected and properly
identified in a herbarium (e.g. Emilio Goeldi Museum in the State of Para).
However, in practical terms, a great amount of time, energy, and financial
resources would be required to fulfill this task. To avoid such logistical trap, I

used the help of a Mateiro6 in each study area to make an approximate

 

6 A Mateiro is a knowledgeable local person whom has been working in a given area long enough
to have the ability to identify most of the tree species using common or local names.

84

identification of most tree species. Finally, I used a description of trees produced
by the Brazilian Institute of Environment and Renewable Natural Resources-
IBAMA to provide the scientific name for each tree species previously identified
in the field. Scientific names helped to avoid double counting of the tree species

identified by local names.

3.3.3.3. Data processing for biodiversity

Forest diversity can be quantified in many different ways. Key
measurement concepts of the measurement of diversity include specie richness,
heterogeneity, and equitability. Alternative approaches to measure biological
diversity suggested by Peet (1974) are based on specie richness and
heterogeneity indices. Species richness is an indicator of species diversity based
on the number of species per sample. As a result, biological richness will
increase as the number of species observed in a sample increases. Therefore,
richness is dependent of the sample size. Following this reasoning, a larger
sample size with more species present will be understood as a 'richer' sample.
Alternatively, heterogeneity indices can measure diversity by taking into account
the number and equitability (relative diversity) of species and, because of that,
these indices are less sensitive to sample size (Peet, 1974).

To compare tree species diversity of different sample sizes conducted in
the State of Acre, Rond6nia, and Mato Grosso, I applied two heterogeneity
indices that are not sensitive to sample size, the Simpson’s Index of Diversity (1-
D) and the Simpson’s Reciprocal Index (1/D), both variants of Simpson’s Index

(D) (Peet, 1974). The Simpson’s Index of Diversity varies directly with

85

heterogeneity and ranges between 0 and 1, where higher values indicate greater

species diversity. This index can be defined as:

S
D =1— 2H p,-2 (2)

The Simpson’s Reciprocal Index starts with 1 as the lowest possible value,
representing a community containing only one species, and the maximum value

is the number of species in the sample. It can be defined as:

N S
D=1/Z,=,pi (3)

where D and D is the Simpson’s Index of Diversity and the Simpson’s
Reciprocal Index, respectively, p, is the proportion of individuals in species i,

and S is the number of species (Peet, 1974).
3.3.3.4. Field validation for remote sensing approaches

Canopy openness was measured for both disturbed and undisturbed
forests using hemispherical photos (further details are provided below). Similarly,
tree species diversity (including secondary tree species) was qualitatively and
quantitatively established for each site. Hemispherical photos acquired with “fish
eye” lens and digital camera (3.2 mega pixels resolution) beneath undisturbed
and disturbed forest canopies were used to estimate canopy openness. The
photo-shots follow 500 meter transect trails, for each 10-meter intervals. Canopy
openness measurement for each individual location was calculated using the
Gap Light Analyzer (GLA) software. Fraction canopy openness was estimated

using a semi-automated technique because each individual photo requires a

86

threshold empirically defined by the photo interpreter (Appendix 8.6).
Subsequently, the canopy fraction was obtained by calculating the difference
between 100% and forest canopy openness (%) calculated from the
hemispherical photos.

Finally, potential occurrences of forest fires were surveyed in 120
randomly distributed locations of forested lands within the Landsat path 226 and
row 068. The resulting information was used to perform an accuracy assessment
of the burned forest detection technique. Further details of the accuracy
assessment of burned forest detection are described in Chapter 2.

3.3.4. Remote sensing approaches
3.3.4.1. Burned forests detection

Burned forests were detected and mapped using a remote sensing
approach based on Spectral Mixture Analysis. This approach was successfully
applied by Cochrane and Sousa Jr (1998) and Souza et al. (2003) to detect
selectively logged and burned forests using Spot imagery in a case study in the
State of Para, Brazil. Cochrane and Sousa Jr (1998) used fraction images based
on particular endmembers (spectral pure signatures) that represented specific
image components such as green vegetation, soil, non-photosynthetic
vegetation, and shade. Based on this approach, a fully automated technique was
developed and tested to detect burned forests using non-photosynthetic
vegetation fraction image from Spectral Mixture Analysis (SMA), retrieved from
Landsat imagery. This technique was developed and tested for a 30000 km2

study site in the Amazon State of Mato Grosso. Accuracy assessment results

87

indicated 96.0% and 0.91 of overall accuracy and overall kappa statistic,
respectively. Additional details regarding this technique and its usage can be
found in Chapter 2.

The approach introduced in Chapter 2 was applied where signs of burned
forest were observed throughout the Brazilian Amazon. Burned forests were
detected and binary classified (1 = burned forest, 0 = non-burned forest) in raster
format for 1992, 1996, and 1999.
3.3.4.2. Fractional coverage

The theoretical foundation for the assessment of forest canopy impacts by
selective logging and forest fires using vegetation index derived from remotely
sensed data was previously described in Chapter 2 for a case study in the State
of Mato Grosso. In that study, Modified Soil-Adjusted Vegetation Index - MSAVI,
Global Environmental Monitoring Index - GEMI, modified GEMI“, Normalized
Difference Vegetation Index - NDVI, Aerosol Free Vegetation Index - AFRI,
Modified Soil-Adjusted Vegetation Index aerosol free — MSAVIaf, and Green
Vegetation derived from Spectral Mixture Analysis were retrieved from the
Landsat path 226 and row 068, and their performance was assessed under
different atmospheric conditions and forest characteristics.

For the basin-wide analysis purposes, two additional case studies in the
Amazon States of Rondonia and Acre (Appendix C.3) were conducted in the
present study. Vegetation indices’ performances were assessed again for
different Amazon sub-regions and forest sites by comparing fractional coverage

values derived from remotely sensed data to canopy cover measurements from

88

hemispherical photos acquired in the field. This assessment contributed to select
the optimum vegetation index to be applied in the basin-wide analysis.

Time-scale variation of forest canopy also was verified through multi-
temporal analysis of fractional coverage retrieved from Landsat imagery acquired
in 1992, 1996, and 1999. Finally, partial effects of each land use and land cover
change on fractional coverage were identified using multiple regression analysis.
Further details of this multiple regression analysis are described bellow.
3.3.5. Forest canopy impact assessment

Forest canopy degradation by selective logging, forest fire, logging and
forest fire combined, and forest regeneration following each of these forest
disturbances in the study area were assessed using multiple regression analysis.
Multiple regression analysis is useful to estimate partial effects of a given
independent variable by controlling other independent variables that
simultaneously affect the dependent variable (Wooldridge, 2000).

Based on Wooldridge (2000), the multiple linear regression model can be

defined as:

FC%: [30 +iBIxI +3239 +---+,314x14 +14: (4)
where FC% is the observed continuous variable (green fractional percentage
derived from remotely sensed data), and ,60 is the intercept or constant, ,3] is
the parameter associated with x1 (independent variable 1), ,62 is the parameter

associated with x2 (independent variable 2) and so on. The variable u is the error
term, which contains factors other than the independent variables included in this

model that affect FC%.

89

The vector of independent variables (x) tested is described as following:

x1= Deforestation

x2 = Secondary regrowth

x3 = Water body

x4 = Cerrado or Savannah

x5 = Newly logged forest

x6 = Newly logged and burned forest

x7 = Newly burned forest

x8 = Previously logged forest

x9 = Previously burned forest

x10 = Previously logged and burned forest
X11= Year 1996

X1; = Year 1999

x13 = Longitude or x location (kilometer)
x14 = Latitude or y location (kilometer)

OOOOOOOOOOOOOO

Note: Undisturbed forest and the year 1992 were used as omitted variables for
the land use classes and annual changes, respectively.

By using this multiple regression analysis, I was able to measure the effect
of each land use or year of analysis on forest canopy cover in the study area,
holding constant the other independent variables.
3.3.5.1. Testing Hypotheses

Based on field observations, I expected that impacts by anthropogenic
activities should produce measurable effects on forest canopy. Therefore, any
observed type of land use and land cover should affect (increase or decrease)
values of fractional coverage derived from Landsat imagery when compared to
undisturbed forests. I also expected that effects of non-controlled seasonal and
natural variations should have effects on values of fractional coverage, which
ultimately could be observed by using multi-temporal dataset. Finally, I expected
that values of fractional coverage could be affected by uncontrolled spatial

variability of natural forests and human activities in the study area. In this case,

90

location itself could be used to identify potential spatial variation of fractional
coverage values throughout the study region. Based on these assumptions, l

hypothesized about each value of the fourteen slope coefficients (,6, ”62,
...,,6,4 ), once the effect of other coefficients had been account for, using

statistical inferences to test the following hypothesis:

Ho: Null hypothesis:
Land use and land cover change variables (,6, to 6,0 ):
,6, = 0: Deforestation has no effect on fractional coverage.
,62 = 0: Secondary regrowth has no effect on fractional coverage.
,63 = 0: Cerrado has no effect on fractional coverage.
,64 = 0: Water body has no effect on fractional coverage.
,65 = 0: Newly logged forest has no effect on fractional coverage.
,66 = 0: Newly burned forest has no effect on fractional coverage.
,67 = 0: Newly logged and burned forest has no effect on fractional coverage.
,68 = 0: Previously logged forest has no effect on fractional coverage.
,69 = 0: Previously burned forest has no effect on fractional coverage.
6,0 = 0: Previously logged and burned forest has no effect on fractional

coverage.
Temporal variables (,6, , to 6,2 ):

,6,, = 0: Non-controlled seasonal or temporal variations for 1996 have no

effect on fractional coverage.
6,2 = 0: Non-controlled seasonal or temporal variations for 1999 have no

effect on fractional coverage.
Spatial variables (63 to ,624 ):
,6,3 = 0: Anthropogenic and natural spatial variations over longitude (x)

direction have no effect on fractional coverage.
6,4 = 0: Anthropogenic and natural spatial variations over latitude (y)

direction have no effect on fractional coverage.

H1: Alternative hypothesis:
Land use and land cover change variables (,6, to 6,0 ):

,6, ¢ 0: Deforestation has effect on fractional coverage.

,62 at 0: Secondary regrowth has effect on fractional coverage.
,63 -'/-' 0: Cerrado has effect on fractional coverage.

,64 it 0: Water body has effect on fractional coverage.

91

,65 it 0: Newly logged forest has effect on fractional coverage.

,66 it 0: Newly burned forest has effect on fractional coverage.

,67 75 0: Newly logged and burned forest has effect on fractional coverage.
,68 ¢ 0: Previously logged forest has effect on fractional coverage.

,69 it 0: Previously burned forest has effect on fractional coverage.

6,0 at 0: Previously logged and burned forest has effect on fractional

coverage.
Temporal variables ( ,6, , to 6,2 ):

,6, , ¢ 0: Non-controlled seasonal and temporal variations for 1996 have effect
on fractional coverage.

6,2 it 0: Non-controlled seasonal and temporal variations for 1999 have effect
on fractional coverage.

Spatial variables ( 6,3 to ,6, 4 ):

,6,3 76 0: Anthropogenic and natural spatial variations over longitude (x)

direction have effect on fractional coverage.
6,4 it 0: Anthropogenic and natural spatial variations over latitude (y)

direction have effect on fractional coverage.

3.3.5.2. Data Sampling

As previously discussed in Chapter 2, multiple linear regression models
are sensitive to spatial auto-correlation in the dataset. The auto-correlation
effects can be minimized by systematically sampling the dataset (Anselin 2002,
Arima et al. 2007). Therefore, I used a 5km x 5km point grid to sample all
variables of interest over the study region (Appendix C.1). Since pixel was used
as the unit of observation in this analysis, a unique value of the dataset was
acquired for each point of the sampling grid to represent each of the studied
variables. Additionally, each grid cell was located with x and y coordinates, which
also were used as independent variables. This procedure is expected to allow

the identification of spatial trends in the data and to reduce the error term in the

92

regression model due to spatial autocorrelation in the data population (Chomitz
and Gray 1996, Arima et al. 2007).
3.4. Results and Discussion
3.4.1 . Multi-temporal assessment of selective logging and forest fires
Matricardi et al. (2007) developed remote sensing approaches to detect
selectively logged forests in the entire Brazilian Amazon. Using such
approaches, a total of 5980 kmz, 10064 kmz, and 26085 km2 of selectively logged
forests were detected in 1992, 1996, and 1999, respectively. By using linear
mixture analysis, I could detect 8280 kmz, 7580 km2, and 11508 km2 of burned
forests in the Brazilian Amazon in 1992, 1996, and 1999, respectively. The total
area of tropical forests impacted by selective logging and forest fire and detected
using remotely sensed data were 11800 kmz, 16500 kmz, and 35600 km2 in
1992, 1996, and 1999, respectively. This magnitude of degraded forests had not
been reported in previous studies of deforestation (Skole et al., 2004). Table 3.3
shows further details of the detected areas of selectively logged and burned
forests, taking into account the overlap between both types of forest
disturbances.

Table 3.3 Detected selective logging, burned forests, and dataset characteristics

 

Year Burned only Burned & Logged Logged only Total degraded

 

(km2) 4km2) 4km2) (km )
1992 5,889.4 391.8 5,588.2 11,8893
1998 8,177.7 1.4033(3) 8,951.0 18,5320
1999 9,038.4 2,470.7(“) 24,188.1 35,897.3

 

a Total includes 290.2 km2 of burned forest detected in 1996, which was
previous logging detected in 1992 and not detected in 1996;

Total includes 573.4 km2 of burned forest detected in 1999, which was
previous logging detected by 1996 and not detected in 1999.

 

 

93

Landsat path and row analysis indicated that 1205 km2, 1719 kmz, and
4241 km2 of selective logging were detected in 1992, 1996, and 1999,
respectively, within the path 226 and row 068, located in Mato Grosso. This was
the greatest forest extension impacted by selective logging for a particular
Landsat scene (path/row) in the entire Brazilian Amazon. Meanwhile, the
greatest extension of burned forest detected was observed within the path 222
and row 063 in 1992 (around of 644 km2) and within the path 224/068 in 1996
and 1999 (around of 605 km2 and 1227 kmz, respectively). Appendix C.9 shows
basin-wide results of selectively logged and burned forest detected within each
Landsat path and row in 1992, 1996, and 1999.

Most of the detectable burned and selectively logged forests were
observed in Mato Grosso and Para. These two states accounted for more than
90% of total selective logging only, 89% of total burned forest only, and 87% of
total selectively logged and burned forest combined detected in 1992, 1996, and
1999. Mato Grosso was responsible for approximately 59% of burned forest in
1992 and Para was responsible for approximately 76% of selectively logged and
burned forests detected in 1996 (Appendix 0.10).

An additional element of concern in this analysis was related to the
occurrence of forest disturbances within the borders of land set aside from
economic development for different types of non-economic uses. Around 1.8%,
2.3%, and 4.7% of the total selectively logged forests were detected within the
limits of Amerindian reserves, protected lands, and military areas in 1992, 1996,

and 1999, respectively (Matricardi et al., 2007). Approximately 0.6%, 2.6%, and

94

5.1% of the total forests burned were detected within these areas in 1992, 1996,
and 1999, respectively. Amerindian reserves constituted the vast majority of the

detected selectively logged and burned forests within these areas (Table 3.4).

Table 3.4.Total area of selectively logged and burned forests detected in 1992,
1996, and 1999 within protected areas (indigenous land, conservation unit, and

military area)

 

Indigenous land Conservation units Military area
Burned Logged Burned Logged Burned Logged
Year forest Forest (*) forest Forest (*) forest Forest (*)
(km2) (km2) (km2) (km2) (km2) (km2)

 

 

 

1992 33.3 106.5 2.8 3.6 0.0 0.0
1996 193.2 216.6 0.5 11.1 0.0 0.0
1999 574.3 1189.8 17.3 44.7 0.4 0.0
Total 800.8 1512.8 20.5 59.4 0.4 0.0

 

* Source: Matricardi et al. (2007)

Multi-temporal analysis between forest fire and deforestation indicated that
2968 km2 of burned forest only detected in 1992 were deforested by 1996 and an
additional 546 km2 by 1999. Around of 3225 km2 of burned forest only detected in
1996 were deforested by 1999. Comparisons between selective logging and
deforestation in this period of analysis indicated that 651 km2 of logged forest
only detected in 1992 was deforested by 1996 and an additional 708 km2 by
1999. Meanwhile, 984 km2 of logged forest detected in 1996 were deforested by
1999. Additionally, comparisons between deforestation, selective logging, and
forest fire indicated that around of 145 km2 of selectively logged and burned
forest combined in 1992 were cleared by 1996 and an additional 188 km2 were

cleared by 1999. Around of 720 km2 of selectively logged forest in 1992 and/or

95

1996 and burned forest in 1996, were deforested by 1999. These results also
showed that around of 457 km2 and 148 km2 of forest disturbances detected
could be considered recurrent fires and revisited logging, respectively. I defined
recurrent fire and revisited logging as burned and logged forests, respectively,
detected in 1992 that had formerly recovered sufficiently to become undetectable
in 1996 and again were detected in 1999 as burned or logged forests.
3.4.2. Optimum vegetation index selection

In Chapter 2, an optimum vegetation index was selected based on the
estimated relationships among various vegetation indices retrieved from a
Landsat image and canopy cover derived from hemispherical photos. The
statistical results showed that AFRI could perform slightly better than MSAVIa,
under clear sky and anomalous atmospheric conditions for the study region in
Mato Grosso. However, linear relationship between fractional coverage values
derived from each vegetation index and those calculated from hemispherical
photos acquired in the field indicated that MSAVI and its modified version
(MSAVIaf,) could better estimate forest canopy cover for the same region. Finally,
MSAVIa, was selected as the optimum vegetation index given its property of
acquiring more accurate ground information than the other indices in the
presence of smoke and aerosols.

Additional hemispherical photos were acquired in the field sites located in
Acre and Rondonia. MSAVI showed highest R-squared for Rond6nia and Acre
study sites (0.74 and 0.81, respectively). Immediately following MSAVI, MSAVIaf,

GEMI, GEMI“, and GV showed the same R-squared (0.69) for the Rondbnia

96

study site. MSAVIaf showed the second highest R-squared (0.80) for the Acre
study site. Based on these statistical results and on its performance under clear
and smoke sky condition, I selected MSAVIa, as the optimum vegetation index to
be used in this basin-wide analysis. The Appendix C.8 shows the statistical
results of the empirical linear relationship between various vegetation indices and
forest canopy coverage based on field measurements for the study sites in Mato
Grosso, Rond6nia, and Acre.

3.4.3. Forest canopy impact assessment based on remotely sensed data

Impacts on forest canopy by selective logging and forest fire were
estimated using a multiple regression model (Equation 4) and land use and
fractional coverage derived from remotely sensed data. The statistical results
indicated that this equation can significantly explain 52.8% of the dependent
variable (fractional coverage).

Analysis of the partial effects of different anthropogenic activities that
potentially could be affecting fractional coverage derived from vegetation index
domain in the study region from 1992 to 1999 indicated that deforestation alone
was responsible for a decrease of 49% (d: 0.38) in the fractional coverage values
when compared to fractional coverage values for undisturbed forests and holding
fixed all other factors. Although secondary regrowth is defined as a formerly
deforested area, this land use was much less responsible for the reduction of
fractional coverage values (around of 7%) than was deforestation itself.

In terms of forest uses, around of 28%, 19.8%, and 2% of estimated

forest canopy decreases in the studied period were due to new areas of selective

97

logging and forest fire combined, forest fire only, and selective logging only,
respectively. Previously burned and logged forests were responsible to decrease
approximately 10% of fractional coverage values. When considered previously
burned forest only, I could not reject the null hypothesis that fractional coverage
values for this forest use are not significantly different from fractional coverage
values for undisturbed forests (omitted variable). My analysis also indicated that
forest canopy is sufficiently recovering after selective logging activities
throughout the Legal Amazon and such recovery was responsible for an increase
of approximately 6% in fractional coverage values when compared to
undisturbed forests. Figure 3.1 shows partial effects of each anthropogenic

activitv on forest canopv change.

 

15.00
10.00
5.00
0.00
—5.00
-10.00
-15.00
—20.00
-25.00
-30.00
-35.00

Canopy change (%

 

t'iil'e’wiy loggiediforiest ’ ’ ’ T ENewIJSJmeIrcYest
l Newly logged/burned forest [I Previously logged forest
_ Previously burned forest :1 Previously logged/burned forest

 

 

 

Figure 3.1. Partial effects of various anthropogenic activities on forest canopy
cover in the Brazilian Amazon using linear multiple-regression model and multi-

temporal remotely sensed data

98

The null hypothesis that anthropogenic and natural spatial variations over
longitude and latitude (x and y) directions have no effect on fractional coverage
values could be rejected and the alternative hypothesis was consequently
accepted. Fractional coverage values significantly decrease 0.8% (3: 0.03
standard deviation) each 100 km towards east direction and significantly increase
1.2% (:I: 0.04) each 100 km towards north direction.

In terms of temporal change, the results indicated an increase of 1% of
fractional values in 1996 compared to year 1992 and by holding all other
independent variables fixed. The greatest annual loss of fractional coverage,
around 5% (:I: 0.3), was observed in 1999. A possible explanation is related to the
factors that vary annually, such as the rainfall reduction caused by the 1997-1998
“El Nino” observed by Siegert et al. (2001), absent in the list of independent
variables, may have contributed to decrease fractional coverage in 1999.
Appendix C.11 shows further results of the linear multi-regression analysis and
Appendices 018 to 0.20 show mosaics of fractional coverage derived from
Landsat imagery acquired in 1992, 1996, and 1999 for the study region.

3.4.4. Field measurements

In the field survey conducted in the States of Acre, Rondcnia, and Mato
Grosso, a total of 3269 individuals of tree species (DBH > 10 cm) were recorded
for an area of 9.38 hectares, an average greater than 348 tree individuals per
hectare. This estimate is well in the range of Whitmore’s (1984) estimate that
indicated 250 or more species of trees DBH greater than 10 cm for tropical rain

forests. Approximately 80% of the total observed tree species observed in the

99

present study could be identified by the Mateiros, which indicated that a minimum
of 119 different tree species were observed in those sampled sites (Appendix
C6).

The number of new tree species decreased as the sample size (number of
transects) increased. The number of tree species also increased between the
transects #3 and #4 and between transects #8 and #9, which represented the
transition sites between the State of Acre and Ronddnia and between the State
of Rond6nia and Mato Grosso, respectively. The observed increase of new tree
species in the first samples for each sampled region indicated potential
differences in tree species diversity among the study sites. However, the fact that
the same Mateiro was not available to assist the field work data gathering
process in three regions, could have affected the detected number of unknown
species.

Based on field observations, the sites sampled in Acre and Rondcnia
seemed to have greater number of tree species (DBH > 10 cm) than did those
sites located in Mato Grosso. Forest disturbances by selective logging and forest
fire were also visually more intensive in Mato Grosso. Based on these facts, I
expected that tree species diversity would be higher in Acre and Rondcnia.
Statistical results (t-test for independent groups, p1¢u2) confirmed that sampled
sites in Rond6nia and Acre combined had higher Simpson’s Diversity Index (M:
0.94, fl= 0.008) than did those sites sampled in Mato Grosso only (M = 0.81,
E =0.087). This difference was significant, 1(18) = 3.9937, g = .0004. Similarly,

sites in Rondbnia and Acre combined had significantly higher Simpson’s

100

Reciprocal Index (M: 16.3, S_D= 2.31) than did those sites in Mato Grosso only
(M = 6.1, S_D =1.88), 1(18) = 10.8658, 9 = .0000. The greatest decrease of tree
species diversity index was observed for a severely burned site (transect # 20) in
Mato Grosso. The Simpson’s Index of Diversity, Simpson’s Reciprocal Index, and
the number of tree species variation across the sampled sites are shown in

Figure 3.2.

 

140

 

120 .

1005

 

 

 

0 i i t t . 1 f i i l . 4 i t i i i l . 0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
TransectlD

 

 

+ Nrrberoftree species (DBH> 10cm) + Sirrpsorls Reciprocal hdex(1lD)
—o— Sirrpsorls Diversityhdex(1-D)

 

 

 

 

 

 

Figure 3.2. Number of tree species, Simpson’s index of diversity, and Simpson’s
reciprocal index against number of sampling (transect)

By grouping the sampled sites in undisturbed and disturbed (logged
and/or burned) forests, the difference of mean values of Simpson’s Index of
Diversity for these two groups was not significantly different from zero
(undisturbed M = 0.90, S_D = 0.06, disturbed M = 0.85, g; = 0.1, t(18) = 0.9199,

g = 0.1849). Similarly, the difference of mean values of the Simpson’s Reciprocal

101

Index was not significantly different for undisturbed (M = 12.03, E = 5.38) and
disturbed (M = 9.53, S_D = 5.59) forests, 1(18) = 0.8742, g = 0.1965). However,
the combination of undisturbed and logged forest sites (M = 0.892, S_D = 0.05)
produced significantly higher Simpson’s Index of Diversity than did those burned
forest sites (M = 0.796, g = 0.13), 1(18) = 2.401, g = 0.0135.

Undisturbed forest sites had significantly higher canopy cover (M = 87.7,
_D = 2.7) than did the forest sites disturbed by selective logging and forest fires
(M = 80.1, _S_D_ = 7.7), 1(18) = 2.138, p = 0.0232. Undisturbed forest sites also had
higher canopy cover (M = 88.6, E = 1.8) than did the forest sites disturbed by
forest fire only (M = 76.6, S_D = 11.1), 1(9) = 2.37, p = 0.021. Similarly,
undisturbed forests had higher canopy cover (M = 88.6, E = 1.8) than did the
forest sites disturbed by selective logging only (M = 81.9, S_D = 2.886), 1(12) =
4.627, p = 0.001.

Mean values of fuel load (metric tons per hectare) and litter depth (cm) did
not show substantial difference among various types of forest uses. For example,
the estimated difference of mean values of fuel load for undisturbed (M = 68.1,
§[_) = 42.1) and disturbed forests (M = 60.5, SQ = 36.5) was not significant, _t_(18)
= 0.380, p = 0.351. This result showed, however, that variance (standard
deviation) is substantial among those forest sites. The greatest fuel loads of
approximately 141 and 138 metric tons per hectare were observed in previously
logged forests in Mato Grosso and Rond6nia, respectively. As opposed to forest
uses, amount of fuel load seemed to vary significantly among the study regions.

Study sites located in Rondénia and Mato Grosso (M = 71.1, SQ = 35.8) had

102

higher fuel loads than did those located in Acre (M = 27.9, _S_D = 14.8), t(18) =
2.32, p = 0.016. Similarly, study sites located in Rondénia and Mato Grosso (M =
4.6, S_D = 2.3) had higher litter depth than did those located in Acre (M = 2.1, _S_D
= 1.1), 1(18) = 2.157, p = 0.023.

The difference of mean of mortality rates of tree species (DBH > 10 cm)
for undisturbed (M = 7.3, SD = 3.9) and disturbed forests (M = 15.9, S_D = 15.5)
was not significantly different from zero, 1(18) = -1.2163, g = 0.12. Undisturbed
and logged forest combined had significantly lower tree mortality rates (M = 7.6,
S_D = 4.5) than did those burned forest sites (M = 28.1, SD = 18.2), 1(18) = -
4.064, g = 0.000. Undisturbed and logged forest combined also had lower rates
of dead basal area (M = 8.47, _S_D = 5.2) than did those burned forest only (M =
22.7, S_D = 17.2), 1(18) = -2.897, g = 0.000. The difference of mean values of the
rate of dead basal area for undisturbed and disturbed forests was not
significantly different from zero, 1(18) = -0.9060, g = 0.1885.

3.4.5. Links between forest canopy and further ecological indicators
Pearson correlation values among various field measurements that
indicate different level of forest disturbances are presented in Table 3.5. Further

estimates of forest disturbance indicators based on field measurements are
presented in Appendix 07.

High negative correlation values were observed between forest canopy
cover, tree mortality rate, and dead basal area. These results indicate that tree
mortality rate and dead basal area increase when there is a decrease in forest

canopy cover.

103

Table 3.5 Correlation matrix among various forest parameters measured in the

study sites in the State of Acre, Rond6nia, and Mato Grosso.

 

 

FC TMR DBA REG REG5 REGI° Fuel Litter
load depth SID SRI

FC 1
TMR -0.77 1
DBA -0.78 0.93 1
REG 0.03 -0.19 -0.17 1
REG5 0.14 002 -0.01 -0.29 1
REG1o 0.31 -0.24 -0.18 —0.18 0.79 1
Fuel load -0.13 0.28 0.17 -0.27 -0.40 -0.41 1
Litter depth 0.40 -022 -0.17 -0.37 -0.04 -0.08 0.37 1
SID 0.70 -0.74 -0.73 0.37 0.02 0.22 -0.20 0.10 1
SRI 0.36 -0.47 -0.43 0.35 -0.02 0.28 -0.25 -0.19 0.82 1

 

Pearson correlation is significant at the 0.01 level (1-tailed) for bold values, which
represent special interest in this analysis. N= 20

FC = Canopy cover (%) derived from hemispherical pictures; TMR = Tree mortality

rate; DBA = Dead basal area; REG = Tree regeneration > 10 cm DBH; REG5 = Tree

regeneration < 5 cm DBH; REG10 = Tree regeneration between 5 and 10 cm DBH; SID

= Simpson’s Index of Diversity; RSI = Simpson’s Reciprocal Index

The Simpson’s Diversity Index showed a high positive Pearson correlation

value with forest canopy cover, indicating that tree diversity increases when
canopy cover increases. On the other hand, the Simpson’s Diversity Index had
high negative Pearson correlation values with tree mortality rate and dead basal
area. These results indicate that tree diversity decreases when tree mortality rate
and dead basal area increase. Tree mortality rate and dead basal area had high
positive correlation. Finally, two size groups (i.e. <5 cm DBH and 5-10 cm DBH)
of regeneration tree species had high positive correlation. These results indicate

that forest regeneration is not only a process with temporal fluctuations but

dependent on the existence of viable tree individuals.

104

3.5. Conclusions

The present results showed a substantial increase (~300%) of forest
degradation by selective logging and forest fires in the Brazilian Amazon
between 1992 and 1999. The total area of forest impacted by these activities was
around of 11800 kmz, 18500 kmz, and 35800 km2 in 1992, 1996, and 1999,
respectively. Total estimates indicated that at least 52859 km2 of natural forests
(1.1% of the entire Brazilian Amazon) were selectively logged and/or burned
between 1992 and 1999. From this total, approximately 10000 km2 were
deforested by 1999. Selective logging alone was responsible for more than 60%
of forest disturbances detected in the studied period and, approximately, 33%
and 7% of forest disturbances were due to forest fire alone and forest fire and
selective logging combined, respectively. The States of Mato Grosso and Para
were affected by approximately 90% of this forest degradation detected using
remotely sensed data acquired in 1992, 1996, and 1999 (see the spatial
distribution of selectively logged and burned forests detected on maps presented
in Appendices 0.12 to 0.17).

These amounts of forest disturbances detected in 1992, 1996, and 1999
do not represent, however, annual rates because they did not occur exclusively
on these given years. Most of selectively logged forests are detectable on
satellite imagery up to a year following logging activities, but none of the
disturbed forests should remain detectable more than 2-3 years due to rapid
forest regeneration in tropical regions (Stone and Lefebvre, 1998, Sousa Jr and

Barreto, 2000). Evidences of burned forests can be even more persistent on

105

satellite data. However, their accurate detection would require multi-annual
dataset in order to calculate annual rates of these types of forest disturbances
(Souza et al., 2005). Based on this assumption, I used the results of the multi-
annual analyses conducted by Matricardi et al. (2005) and the results of the
analysis presented in Chapter 2 to estimate the increase of new areas of
selectively logged and burned forests, which were then extrapolated to the entire
Brazilian Amazon.

In the case of selective logging, I used the average relations between total
detected and the increase of logging areas from 1992 to 1999 observed by
Matricardi et al. (2005) in the Mato Grosso and Rondonia study cases, 39.5%
and 91.4%, respectively. Furthermore, Matricardi et al. (2005) assumed that
logging activities may persist up to 3 years on satellite imagery and high intensity
logging occurred on those Landsat scenes that showed an average of total
detected logging areas greater than 100 km2 yr" in 1996 and 1999. Low intensity
logging where fewer marketable trees are harvested, lesser infra-structure built,
and, consequently, lower forest disturbances, was assumed to occur on Landsat
scenes that showed an average of total detected logging areas below 100 km2 yr’
1 in 1996 and 1999. Finally, the States of Mato Grosso and Rond6nia average
relations (39.5% and 91.4%) were assumed to represent Landsat scenes
showing high and low intensity logging activities, respectively.

For annual rates of burned forest, I used the average relations between
total detected and the increase of burned areas from 1992 to 1999 presented in

Chapter 2 for the Mato Grosso case study. The results in this case were 41%

106

and 50.4% observed for burned forests and selectively and burned forests
combined, respectively. I also assumed that forest fires may persist up to 5
years on satellite imagery and heavily burned forests occurred on those Landsat
scenes showing an average of total detected burned forests greater than 20 km2
yr" in 1996 and 1999. Lightly burned forests were assumed to occur on Landsat
scenes showing an average of total detected logging areas less than 20 km2 yr"
in 1996 and 1999. Based on the Rond6nla case study, an increase of
approximately 100% of burned forests was observed from 1992 to 1999 on
Landsat imagery. Additionally, field observations showed that lightly burned
areas encompassed small forest plots, sparsely distributed within the Landsat
path 232 row 067. Very unique site characteristics (i.e. moist microclimate and
dense vegetation) may have reduced fire intensity and prevented fire propagation
into neighbor forests. Finally, the average relations (41% and 50.4%) were
assumed to represent Landsat scenes showing heavily burned forest only and
selectively logged and burned forest combined, respectively. Landsat scenes
showing less than 20 km2 yr" in 1996 and 1999 were assumed as having an
average relation of 100%. Nevertheless, these lightly burned sites represented
less than 10% of the total burned forest detected in the studied period and should
produce minor impact on final results.

Assuming that the average relations previously presented can be
extrapolated to a basin-wide scale, the total increase of selectively logged forest
only, burned forest only, and selectively logged and burned forests combined in

1992 was at least 2391 kmz, 2883 kmz, and 193 kmz, respectively. These

107

amounts increased to 3951 kmz, 3018 kmz, and 849 km2 by 1996. Following this
increasing trend, estimates for 1999 indicated that areas of selectively logged
forest only, burned forest only, and selectively logged and burned forests
combined expanded to 11767 kmz, 4514 kmz, and 1156 kmz, respectively. These
new areas of selective logging and forest fires represent around of 46% of the
total disturbed forests detected in 1992 and 1996, and 49% in 1999. These
areas were considered to have been previously impacted by selective logging
and forest fire and, due to their more severe forest disturbances, were more
persistent on satellite imagery. The remaining detected disturbed forests
increased from 8401 km2 in 1992 to 8913 km2 in 1998 and 18258 km2 in 1999.
The total areas detected and annual rates of selective logging and forest fires are
presented in Tables 3.3 and 3.6, respectively.

New areas of natural forests impacted by selective logging and forest fires
increased 139% and 229% in the period between 1992 and 1996 and between
1996 and 1999, respectively. Matricardi et al. (2007) had estimated rather similar
increases of new areas of selective logging only in the same period for the entire
Brazilian Amazon. These results support previous reports of increasing trends in
degradation of natural forests in the Brazilian Amazon by selective logging
activities (Uhl et al., 1997, Nepstad et al., 1998, Stone and Lefebvre, 1998,
Souza and Barreto, 2000) and forest fires (Uhl and Buschbacher, 1985, Uhl and
Kauffman, 1990, Cochrane and Schulze, 1998, Cochrane, 2001, Cochrane et al.,

2004).

108

Table 3.6. Estimated increase of new selectively logged, forests burned, and

selectively logged and burned forests in the Brazilian Amazon.

 

Annual increase (km2)

 

 

 

Year Logged only (1) Burned only (2) Burned & Total
Logged (3)

1992 2,391.0 2,883.6 193.5 5,468.2

1996 3,951.6 3,018.6 649.4 7,619.6

1999 11,767.9 4,514.4 1,156.9 17,439.2

 

1 Logging average 1 (39. 5%) and average 2 (91.4%) of selective logging increases were
estimated for Mato Grosso and Rondonia study cases, respectively. (Matricardi et al. 2005)
2Logged and burned forest average 1 (50. 4%) and average 2 (100%) of burned and logged
forests combined were estimated for Mato Grosso and Rondonia study cases, respectively
3Burned forest average 1 (41%) and average 2 (100%) of burned forest“ Increases were
estimated for Mato Grosso and Rondonia study cases, respectively.
These averages were calculated as total area of forest Showing selectively logged and/or burned
forests for the first time divided by the total area of forest wherein selective logging and/or forest
fire was detected. Logging averages 1 and 2 were used to estimate new areas of forests
impacted by selective logging and forest fire for all Landsat scenes showing total area of
selective logging more or less than 100 km2 yr", respectively. The burned forest averages 1 and
2 were used to estimate new areas of forests burned for all Landsat scenes showing total areas
of this forest disturbance more or less than 20 km2 yr", respectively. This last procedure was
applied to estimate new areas of selective Ioggi_ng and forest fires using the averages 1 and 2.

 

 

The intensity of forest disturbances is also increasing in the Brazilian
Amazon. Formerly logged forests are increasingly been revisited by loggers
searching for the second tier economic species as raw material and undisturbed
forests became scarcer (Uhl et al., 1997, Matricardi et al., 2005). Matricardi et al.
(2005) reported that revisiting logging increased approximately to 7.3% of the
total logged forest in the period of 1993-2002. Similar increase (7.1%) of
recurrent fires was observed in the studied period in Mato Grosso (see Chapter
2). The study case results support the research conducted by Cochrane and

Schulze (1998) who observed that recurrent fire and its intensity are likely to

109

increase because of favorable local environmental created by the previous fire
events.

Forest disturbances within the limits of protected areas represented an
average of 2.9% and 2.8% of the total area of the detected selective logging and
forest fire, respectively. The results indicated that such areas have been very
effective in constraining these types of anthropogenic activities in Brazilian
Amazon. The results also indicated that between 1992 and 1999 the total area
detected of selectively logged and burned forests within protected areas
increased 2.9% and 4.5%, respectively. This trend of increasing forest
degradation within protected areas supports the analysis by Pedlowski et al.
(2005) who reported that many conservation units in the Brazilian Amazon have
been increasingly been exposed to logging and deforestation activities.

Finally, the results of the methods used in this analysis are considered
conservative, since, in different circumstances, degraded forests by selective
logging and forest fires cannot be detected. Indeed, Matricardi et al. (2005) and
Matricardi et al. (2007) observed that very low intensity selective logging cannot
be easily detected using remote sensing approaches, nor it may cause
substantial environmental impacts on natural forests and on fluxes of carbon
dioxide. However, the total areas of natural forests here classified as selectively
logged and burned forests represent additional land use types that constitute
major forest impacts and have not yet been properly accounted for by

conventional remote sensing approaches of deforestation assessment.

110

CHAPTER IV
Spatially explicit probabilistic models of forest fire and
selective logging in the Brazilian Amazon
4.1. Abstract
In this chapter, multivariate probit models of forest fire and selective logging were
designed for the Brazilian Amazon. I developed these models using variables
that affect the occurrence of forest fire and selective logging throughout the
region. Therefore, the effects of each variable on likelihood of forest fire and
selective logging were estimated. In spite of the results of a forest fire probit
model indicated that distance to deforestation is the most correlated factor to
forest fires, a combination of factors (e.g. prolonged water deficit, selective
logging, vegetation type, and forest fragmentation) were apparently the more
realistic causes of forest fires in the Brazilian Amazon. In apparent opposition to
forest fire probability estimates, the selective logging probit model demonstrated
that logging activities were not significantly related to distance to deforestation
events. In fact, distance to access roads and timber centers were the most
significant statically and substantively factors affecting occurrence of selective
logging events. The average total timber volume per hectare was significant
statistically but not substantively, which indicated that abundance of timber
sources was not as important as its accessibility for loggers in the period of
analysis. Like forest fire probability, all factors combined (e.g. distance to roads
and timber centers, timber volume, vegetation types, land status, etc.) seemed to

define chances of selective logging occurrence. Although I could observe some

111

synergistic interactions between selective logging and forest fire, selective
logging activities were spatially located predominantly in new frontiers of
undisturbed forests, while most of forest fire events occurred in more fragmented
forests patches next to older deforestation areas. Finally, field information
indicated that the landowner’s desire of converting natural forests into agricultural
land may strongly affect forest fire probability. This factor, however, was not
included in the forest fire probit model because it is considered a criminal action
without previous authorization from the environmental agencies and, therefore, it
is very difficult to accurately survey it in the Brazilian Amazon.
4.2. Introduction

As already discussed in Chapters 1, 2, and 3, selective logging and forest
fire are becoming anthropogenic activities of special concern in tropical regions
due they potential impacts both on people and the environment. Various authors
have been studying these phenomena in the Brazilian Amazon and the resultant
synergistic interaction among them. Uhl and Buschbacher(1985) and Uhl et al.
(1991) observed that impacts by selective logging on undisturbed forests can be
severe, involving forest canopy damages, dead slash vegetation, and
degradation by logging infrastructure such access roads, skid trails, and patios.
These forest disturbances can be substantial even upon the occurrence of a very
selective timber harvesting event (Frumhoff, 1995). Consequently, forests
damaged by selective logging become more susceptible to fires, and fire-
maintained agricultural lands work as powerful ignition sources and fire can

rapidly spread into adjacent and more flammable selectively logged forests (Uhl

112

and Buschbacher, 1985). Furthermore, forest roads left behind by logging
activities provide access to occupation of new frontiers by landless peasants and
squatters (Uhl and Vieira, 1989).

In addition to forest disturbances by selective logging and deforestation,
several other factors might be contributing to increase forest fire in the Brazilian
Amazon. Studies conducted by Kuntz and Siegert (1999) and Siegert et al.
(2001) indicate that severe drought can be observed during years of climatic
events such as El Nino in the Amazon region. Such events substantially increase
forest susceptibility to fire (Cochrane, 2001, Laurance and Williamson, 2001,
Nelson, 2001). For example, Nelson (2001) reported that area and severity of
forest fires varied according to vegetation type in the state of Roraima in which
secondary and igapo (low land) forests were the most severely damaged
vegetation types during the 1997-98 dry season in that region.

Based on that the evidence present in the literature, I conducted a
spatiotemporal analysis using probabilistic models to assess probability of
selective logging and forest fire occurring in the Brazilian Amazon. To reach this
goal, I applied multivariate probit models and variables derived from remotely
sensed data. Each variable represented a specific factor potentially affecting the
spatial and temporal distributions of selective logging and forest fire throughout
the region. I also used multi-temporal datasets for 1992, 1996, and 1999,
including variables representing climatic, landscape, and land use characteristics

such as vegetation types, water deficits, deforestation, and forest fragmentation.

113

The results of the final probabilistic models substantially improved the
current understanding of spatiotemporal distribution of forest fires, selective
logging, and their interaction. These results also contribute to improve land use
and land cover change modeling, and the development of forest fire control and
selective logging monitoring programs by governmental agencies in Brazil.

4.3. Methods
4.3.1. The setting

This research focused on tropical rainforests located within the Brazilian
Amazon, which encompasses an area of approximately 5 million square
kilometers and includes all of the states of Acre, Amapa, Amazonas, Para,
Rond6nia, Roraima, and portions of the Mato Grosso, Maranhao, and Tocantins.
However, I centered my spatiotemporal analysis of selective logging and forest
fire occurrence in a study region (Appendix 0.1) within the Brazilian Amazon
where signs of selectively logged and forest fires were most evident in Landsat
imagery in 1992, 1996, and 1999 (Appendices C.12 to C17). Therefore, I
excluded those areas where these types of human activities were highly unlikely.
4.3.2. Datasets
4.3.2.1. Deforestation dataset

A deforestation dataset produced by the Tropical Rain Forest Information
Center (http://www.trfic.msu.edu) at the Michigan State University was used in
this study. This dataset involves a multi-annual classification of 229 Landsat
scenes covering the entire Brazilian Amazon. Landsat images were individually

classified into seven thematic classes: forest, deforestation, regrowth, cerrado,

114

cloud, cloud shadow, and water using unsupervised image classification
procedures. Three mosaics of the Landsat scenes available for the entire
Brazilian Amazon for 1992, 1996, and 1999 (Appendix D.2) were used to perform
the spatially explicit probabilistic models of fire and selective logging in the study
region, and details of the deforestation dataset are presented in Table 4.1.

Table 4.1 Deforestation dataset characteristics

 

 

Dataset Features Format Spatial Period of Projection
resolution Analysis
Land use Forest, Arc/Info 30m x 30m 1992, Sinusoidal
and land deforestation, vector 1 996, projection,
cover secondary and and Datum and
mosaic regrowth, grid 1 999 Spheroid
cerrado, WGS84
clouds,and
shadows

 

4.3.2.2. Selective logging dataset

As previously addressed in Chapter 2 and 3, selectively logged forests
were not included as a land use class in the deforestation dataset produced by
the TRFIC at MSU because more sensitive remote sensing techniques are
required to accurately detect and map their presence. Therefore, selectively
logged forests were detected and mapped using a semi-automated (texture
algorithm) technique and Landsat band 5 to detect log landings (i.e. log storage
patios) and, subsequently, variable buffer zones were applied (180 and 450
meters) around patios to estimate the size of areas affected by selective logging.

Following Matricardi et al.(2007), visual interpretation was used to map

selectively logged forests with visible canopy disturbance on Landsat images.

115

The combination of these remote sensing approaches showed 92.9% and 0.82 of
overall accuracy and overall kappa statistic, respectively.

The selective logging dataset available at Global Observatory for
Ecosystem Services (GOES), Michigan State University included three multi-
temporal mosaics of selective logging of the entire Brazilian Amazon (Table 4.2).

Table 4.2 Selective Logging dataset characteristics

 

 

Dataset Classification Format Spatial Period of Projection
Features resolution Analysis
Selectively Binary classes Arc/Info 30m x 30m 1992, Sinusoidal,
logged forest (1 if selectively grid 1996, WGS84
GIS layers logged, 0 and Datum and
(mosaic) otherwise) 1999 Spheroid

 

Such mosaics included a binary classification of natural forests (1 if
selectively logged, 0 if othenlvise) for up to 38 Landsat scenes throughout the
Brazilian Amazon for 1992, 1996 and 1999

The present study also included the selective logging mosaics for 1992,
1996, and 1999 produced by Matricardi et al. (2007) (Appendices C14, C15,
and 0.16)
4.3.2.3. Thematic maps

The present analytical effort included thematic maps of urban areas, road
network, protected areas, vegetation types, timber centers, and state borders in
the Brazilian Amazon. The urban area, road network, and state maps were
digitized by SIVAM (2004) from datasets produced by the Brazilian Institute of
Geography and Statistics (IBGE), derived from field surveillance and remotely
sensed data. The vegetation map was compiled and digitized by SIVAM (2004)

from dataset produced by the RADAMBRASIL Project in 1980 (Appendix 03).

116

The map of timber centers in the Brazilian Amazon was composed of timber
center locations reported by Lentini et al. (2003), and the map of urban areas
produced by SIVAM (2004). The timber center locations were defined based on
the spatial distribution of 1190 saw mills surveyed by the Amazon Institute of
Man and Environment — IMAZON throughout the Brazilian Amazon in 1998. The
identification of city location for each timber center was based on Lentini et al.
(2002) and their geographic location was plotted accordingly on the map of urban
areas (Table 4.3).

Table 4.3. Description of the thematic maps

 

 

Dataset Features Format Original Projection Source
scale

Map of Lines and table Arc/Info Derived Geographic, SIVAM
road of road types coverage from Datum IBGE
network (paved, and *.dbf Landsat SAD69

unpaved, imagery, 30

access, etc.) m spatial

resolution

Map of the Polygons, arcs, Arc/Info 1:250,000 Geographic, SIVAM
Amazon table of the coverage Datum IBGE
States State borders and *.dbf SAD69
Map of Polygons, arcs, Arc/Info 1:250,000 Geographic, SIVAM
urban table of city coverage Datum IBGE
areas locations and and *.dbf SADS9

urban areas
Map of Polygons, arcs Arc/Info 1:250,000 Geographic, ISA
protected and table coverage Datum IBAMA
areas and *.dbf SADGQ
Map of Polygons, arcs, Arc/Info 1:250,000 Geographic, SIVAM
vegetation table of coverage Datum IBGE

vegetation and *.dbf SAD69

types
Timber Points and Arc/Info 1:250,000 Geographic, IMAZON
centers tables coverage Datum

and *.dbf SA069

 

The road network map included different road types, which mostly

included official roads constructed by the federal, state and local governments.

117

The road classification includes paved road, unpaved, secondary roads, and
access roads. Logging roads and trails were not included. The state map
included borders of all states within the Legal Brazilian Amazon. The map of
urban areas included identification, location and administrative classification (i.e.
capital, city, and village) of all urban areas in the Amazon region for 1999.
Indigenous land, conservation unit classification (i.e. parks, ecological reserves,
extractive reserves, national and state forests, etc), and military areas were
included in the protected areas map. Appendix D.5 shows a composite map of
road network, protected areas, and timber center location.

Finally, the vegetation map included several vegetation types and sub-
types. The present analysis included 6 classes of forest types: dense, open,
semi-deciduous, deciduous, transitional, and pioneer forests. By 1999,
approximately 27%, 23%, 23%, 21% of the study area was covered by dense
forest, open forest, transitional forest, and cerrado (savannah), respectively.
Therefore, tropical rainforest was the predominant vegetation in the study region
covering 73% of its territorial area by 1999.

All these thematic maps were used to support the spatial analysis of forest
degradation by selective logging and forest fires in the Brazilian Amazon.
4.3.2.4. Timber volume in the Brazilian Amazon

In the 1970s the RADAMBRASIL Project conducted a forest inventory that
included tree species identification and measurements (i.e. basal area and
volume) within thousands of sample units throughout the Brazilian Amazon. In

this study, I used a dataset digitized by SIVAM (2004) containing the geographic

118

location and the total timber volume of tree species measured within 57,971
sample units generated by the RADAMBRASIL Project. In addition, a variogram
model was created by using these sample points, and ordinary kriging using the
values derived from this model was run in ArcGlSTM. The variogram model used
for ordinary kriging had an exponential curve, with a nugget of 2182.8, a range of
3,076,600, and a partial sill of 4460.1. As a result, a map of the timber volume
per hectare was created for the Brazilian Amazon and then converted into a grid
(Appendix 05).
4.3.2.5. Water balance

Water balance is a good measure of the moisture available in a terrestrial
ecosystem and, ultimately, can be a good indicator of forest fire flammability. The
water balance dataset used in this analysis was generated by (Willmott, 1999)
based in a modified procedure developed by Thornthwaite and Mather in 1955.
Basically, the calculation procedure for this model included multiple variables
such as: soil water-holding capacity, the gridded average-monthly precipitation,
evapotranspiration, and temperature. The average water balance datasets for
1992, 1996 and 1999 were then used to identify areas with water deficit during
the dry season7 in these given years. Appendix D.6 shows the water balance for

the Brazilian Amazon during the dry season (May to September) in 1999.

 

7 The dry season is defined as a time period from May to September, when a prolonged low
precipitation and high temperature can be observed in the Amazon region. The dry season for the

Northern Amazon, however, normally occurs from January to March.

119

4.3.3. Forest Fragmentation

Skole and Tucker (1993) introduced the concept of fragmented forests in
the context of tropical deforestation studies. Deforestation was previously
addressed through a binomial classification of forest and non-forest. These
authors observed that forest fragmentation could affect larger areas than
deforestation itself, especially in the context of the Brazilian Amazon. In their
ground-breaking research, Skole and Tucker (1993) defined fragmented forests
as forested areas smaller than 100 hectares surrounded by deforested lands,
and edge effects as a series of ecological responses affecting a buffer zone
distant up to 1 km from adjacent deforested areas.

Concepts drawn from Landscape Ecology have also been applied to
enhance the understanding of the ecological implications of anthropogenic
activities on forest ecosystems (Turner et al., 2001). Accordingly, I used two of
the patch-based measures (patch area and fractal dimension) of landscape
patterns proposed by Skole and Tucker (2001) to assess the presence of distinct
levels of forest fragmentation in the study area. Patch area can be used to
display the frequency distribution of number of patches and patch Size.
Therefore, a high frequency of small patch size is a good indicator of forest
fragmentation intensity, although it disregards the shaping of the patches.
Conversely, patch shape can be a good indicator of temporal changes in
landscape patterns. For example, small forest patches tend to have simpler

shapes than larger ones due to human activities and natural boundary effects,

120

respectively. Thus, fractal dimension is a useful indicator of shape complexity for
patches of different sizes (Turner et al., 2001).

_ In(A)
" In(P) + ln(k)

 

(1)

where d = fractal dimension, A = area of each patch, P = perimeter of each
patch, and k = 0.25 constant value if the actual patch perimeter is used to
estimate P.

By applying the equation 1, fractal dimension can depict the shape of each
forest patch that may range from 1.0 to 2.0 for a straight line and a circle or
square forest patch, respectively.

These two patched based measurements were applied in the study region
using land use and land cover classification maps from 1992, 1996 and 1999. I
used log transformed of path size (m2) in order to minimize the effects of very
large forest patches, and fractal dimension was applied using the equation 1. The
constant k = 0.25 was applied to estimate the fractal dimension of forest patches.
Forested areas are shown in Appendix 02.

4.3.4. Modeling forest fie and selective logging probabilities in the Amazon

Spatial pattern of selective logging and forest fire, their impacts and
interactions with other land use and socioeconomic variables were assessed
using two multivariate Probit models. The Probit is a binary response model,
which implies that the estimated probabilities will range between 0 and 1. This
model uses the standard normal cumulative distribution function and its use is
most indicated when the dependent variable is discrete and the partial effect of

any independent variable is potentially not constant (Wooldridge, 2000).

121

The probit models were used to identify the factors influencing the
probability of selective logging and forest fire occurrence throughout the study
region within the Brazilian Amazon based on their spatial distribution, local
biophysical characteristics, and socioeconomic factors.

According to (Wooldridge, 2000), the probit model can be defined as:

v*=ito+xis+s. v=1iy*>01. (2)
where y* is the unobserved or latent variable (in this study, a pixel representing
forest fire) that satisfies a normal, homoskedastic distribution with a linear

conditional mean; x is a vector of independent variables; 8 is a vector of

unknown coefficients; and § is an independently distributed error term assumed
to be normal with zero mean and constant variance 02. The indicator function
requires that observations take 1 as their value if the event (forest fire) is true,
and 0 if othenlvise. Therefore,

y=1ify*> 0, and

y= 0 if y* .<. 0.

Based on this approach, the probability of forest fire and selective logging in the

study region can be defined using equation 3:

P0! = 1 IX) = P8" > OIX) = P(8 > -XB|X) = $049 (3)
where, <I> (.) is the standard cumulative normal distribution function (cdf); x

is a vector of independent variables, and B is a vector of unknown coefficients.
4.3.4.1. Forest fire conceptual model
Previous studies have shown that forest fire is a growing problem in the

Brazilian Amazon due to different factors such as: the sprawl of ignition sources

122

from fire-managed agricultural lands Uhl and Buschbacher (1985) and the
increasing forest flammability (caused by greater fuel availability and drier
microclimates) as natural forests become more fragmented Cochrane(2001) and
Cochrane (2004) and disturbed by deforestation and selective logging activities
(Uhl and Buschbacher, 1985, Uhl and Kauffman, 1990, Holdsworth and Uhl,
1997, Cochrane et al., 2004). Environmental conditions may also substantively
increase forest flammability. For example, Uhl and Buschbacher (1985), Nepstad
et al. (1999), Houghton et al. (2000) , and Nelson (2001) reported that fires are
more likely during years of prolonged drought.

Many authors (Uhl and Buschbacher, 1985, Uhl and Kauffman, 1990,
Holdsworth and Uhl, 1997, Cochrane and Schulze, 1998, Cochrane, 2001,
Nelson, 2001, Cochrane et al., 2004) have observed that agricultural land is a
major ignition source for forest fires in the Brazilian Amazon. In addition to that, I
could observe during the fieldwork that fires frequently occur along roads and,
consequently, one may need to take into account the importance of
transportation networks as ignition sources. I could not find evidences, however,
of lightning strikes as important ignition sources of forest fire. This finding is
probably explained by the fact that most lightning strikes occur during the rainy
season when forests are too moist to allow fire ignition. Therefore, I hypothesized
that any forest fires observed in this study would be human-driven whether
accidentally or intentionally.

In addition to the ignition sources of fire, forests are becoming increasingly

fragmented and, hence, more susceptible to fires. Consequently, selective

123

logging which has been reported as a major source of forest disturbances in the
Brazilian Amazon (Nepstad et al., 1999, Skole et al., 2004, Asner et al., 2005,
Matricardi et al., 2007) also is deemed as a cause of increased forest fire (Uhl
and Buschbacher, 1985, Uhl and Kauffman, 1990, Holdsworth and Uhl, 1997,
Cochrane and Schulze, 1998, Nepstad et al., 1999, Cochrane, 2001, Nelson,
2001, Cochrane et al., 2004). Additionally, small forest patches and larger forest
edges left behind by the deforestation contribute substantially to forest
fragmentation (Skole and Tucker, 1993, Cochrane et al., 2004, Skole et al.,
2004). Once fragmented, tropical forests become more susceptible to fires
(Cochrane, 2001, Cochrane et al., 2004).

Water balance is an important environmental factor that may contribute to
increase forest fire susceptibility in tropical regions (Cochrane, 2001, Laurance
and Williamson, 2001, Nelson, 2001). Forest fires are more likely to occur in
tropical regions during the dry season, when precipitation is very low (Uhl and
Buschbacher, 1985, Houghton et al., 2000) and during years of extreme moisture
deficit and prolonged drought , where even undisturbed forests may become
flammable (Nepstad et al., 1999, Nelson, 2001). For example, a prolonged water
deficit such as the extreme drought observed during year of climatic events such
as the 1997-98 El Nir‘io (Kuntz and Siegert, 1999, Siegert et al., 2001)
substantially increased forest flammability in the Amazon region.

The forest fire conceptual model was based on the assumption that three
major factors may increase the probability of forest fire in the study region:

ignition sources, forest disturbances, and environmental conditions. I created a

124

vector of independent variables, which included the previously discussed
biophysical and socioeconomic factors that affect forest susceptibility to fire.
First, a dummy variable was created for the land status (1 if within protected
areas, 0 if othenlvise). It is expected that the land status of protected areas (e.g.
ecological reserves, indigenous land, extractive reserves, national parks) should
be a legal barrier for deforestation and selective logging activities (Pedlowski et
al., 2005) and, therefore, a grid cell (unit of observation) representing forest fire is
less likely to occur within their territorial borders. Dummy variables for each year
of selective logging (1 if logged, 0 if otherwise) also were created. It is expected
that selectively logged forests are more susceptible to fire than undisturbed
forests because of the canopy degradation caused by logging lead to a rapid
decrease moisture and humidity in logged forests (Uhl and Kauffman, 1990).
Subsequently, I created two continuous variables based on the Euclidean
distances to agricultural lands (i.e. deforestation) and roads, which estimated the
proximity of grid cells representing burned and non-burned forests to roads and
deforested areas. Another continuous variable was created to represent the
water deficit during the dry season of each year of analysis. As previously
discussed, it is expected that forests under prolonged water deficit would be
more vulnerable to fire than those under moisture conditions. Two other
continuous variables (logarithm of forest patch size and fractal dimension) were
created to represent landscape fragmentation, since smaller forest patch sizes
and lower values of fractal dimension are likely to be more fragmented, therefore,

more susceptible to fire.

125

I also created a set of dummy variables to represent vegetation factors,
assuming that forest fires are more likely to occur in drier microclimate observed
in deciduous, semi-deciduous, and transitional forests than in moisture
microclimate observed in dense and open forests.

Finally, dummy variables were created to represent each state within the
study region, which were intended to capture potential variations at state level
due to different environmental legislation and law enforcement capabilities. As
previously presented in Chapter 3, the bulk of selective logging and forest fire
activities were detected in the states of Mato Grosso, Paré, and Maranhéo in
1992, 1996, and 1999. These states are located within the so-called Arc of
Deforestation created by the Brazilian Institute of Environment and Renewable
Natural Resources- IBAMA as a priority region for forest fire control and
prevention in the Brazilian Amazon (IBAMA, 2007).

Based on that, I tested the vector of the following independent variables (x)
for the Probit model described in Equation 2:

Protected areas (1 if within protected areas, 0 if otherwise)
Proximity to agricultural lands (deforestation) (km)

Proximity to agricultural lands squared

Proximity to roads (km)

Proximity to roads squared

Logging in 1999 (1 if selectively logged in 1999, 0 if otherwise)
Logging in 1996 (1 if selectively logged in 1996, 0 if othenlvise)
Logging in 1992 (1 if selectively logged in 1992, 0if otherwise)
Water deficit (1 if water deficit persisted 2 5 months, 0 if otherwise)
Forest patch size (In transformed area)

Fractal dimension (1 to 2)

Vegetation type 1 (1 dense forest, 0 if otherwise)

Vegetation type 2 (1 open forest, 0 if othenlvise)

Vegetation type 3 (1 semi-deciduous forest, 0 if otherwise)
Vegetation type 4 (1 deciduous forest, 0 if otherwise)
Vegetation type 5 (1 transitional forest, 0 if otherwise)

OOOOOOOOOOOOOOOO

126

o Vegetation type 6 (1 pioneer forest0 if otherwise)
0 Temporal variable 1 (1 if forest was burned in 1999, 0 if othenrvise)
0 Temporal variable 2 ( 1 if forest was burned in 1996, 0 if otherwise)

The control variables were indexed to a time scale (temporal variables) in
the model. Thus, the interaction between time and other variables was assessed,
which contributed to explain spatiotemporal changes of forest. Additionally,
potential interactions between forest fires and very large distance from
agricultural lands and roads (ignition sources) were also considered. This
procedure required creating two new variables, distances to deforested areas
and roads squared, which represented spatial patterns of forest fires
geographically located at large distance from deforestation and roads.
4.3.4.2. Selective logging conceptual model

A multi-temporal assessment of selective logging in the Brazilian Amazon
conducted by (Matricardi et al., 2007) showed that selectively logged forests
increased more than 336% between 1992 and 1999. The vast majority of these
anthropogenic activities were observed within the states of Para and Mato
Grosso (Matricardi et al., 2007) and the logging industrial centers were clustered
along major Amazon transportation axis (Nepstad et al., 1999, Lentini et al.,
2003, Lentini, 2005). (Matricardi et al., 2007) also speculated that some
biophysical and socioeconomic factors such as raw material for saw mills,
environmental law changes occurred in 1996, and extreme droughts, might have
contributed to increase intensity and area of selective logging. Indeed, (Uhl et al.,
1997) reported that loggers were increasing both volume and number of tree

species harvested in more mature logging regions in which timber products are

127

becoming scarcer. As in the case of forest fires, given their legal status,
protected areas also are deemed as offering a barrier for logging activities
(Pedlowski et al., 2005). In this case, although the raw material may be abundant
and economically viable, it is not legally accessible.

Therefore, it was assumed for the selective logging conceptual model that
land legal status, raw material availability, transportation networks, and the
geographic locations of the timber centers were the major factors driving loggers
over natural forests. A dummy variable was created for the land status (1 if within
protected areas, 0 otherwise). A grid cell (unit of observation) representing
selective logging within protected lands is less likely to be observed. Dummy
variables for each year of selective logging (1 if logged, 0 otherwise) also were
created. (Matricardi et al., 2007) observed that selective logging substantially
increased in area and intensity in Brazilian Amazon, growing approximately 73%
and 345% in 1992-96 and 1992-99 periods, respectively.

Three continue variables were created to measure selective logging
proximity to logging centers, roads, and deforestation. I expected that loggers
have to operate at certain buffer distances from logging centers due to the
transportation costs and, therefore, selective logging areas would tend to
decrease as distance to logging centers increase. I also assumed that saw mills
can not be easily relocated closer to sources of raw material because they are
strongly attached to the consumption markets, major transportation axis, and
labor and electricity availability. Similarly to logging centers, it was expected that

undisturbed forests closer to the road networks and previous deforestation are

128

more accessible and, therefore, more likely to be selectively logged than those

located at larger distance from access roads and agricultural lands.

An additional continue variable was created to represent timber volume

(m3/ha. In this case, I assumed that loggers are also attracted by the amount of

timber volume (ma/ha) observed in undisturbed forest throughout a given area of

interest. Finally, a set of dummy variables was created to represent vegetation

factors, assuming that marketable tree species vary according to forest type and,

therefore, some forest types should be more attractive for loggers than others.

Based on these assumptions, I tested the vector of the following independent

variables (x) using the Probit model described in Equation 2 for selective logging:

OOOOOOOOOOOOOOOOOO

Protected areas (1 ifwithin protected areas, 0 if otherwise)
Proximity to agricultural lands (deforestation) (km)

Proximity to agricultural lands squared

Proximity to roads (km)

Proximity to roads squared

Proximity to timber centers (km)

Proximity to timber centers squared

Timber volume (m3 )

Forest patch size (In transformed area)

Fractal dimension (1 to 2)

Vegetation type 1 (1 if dense forest, 0 if otherwise)

Vegetation type 2 (1 if open forest, 0 if otherwise)

Vegetation type 3 (1 if semi-deciduous forest, 0 if otherwise)
Vegetation type 4 (1 if deciduous forest, 0 if otherwise)

Vegetation type 5 (1 if transitional forest, 0 if othenNise)

Vegetation type 6 (1 if pioneer forest, 0 if otherwise)

Temporal variable 1 ( 1 if selectively logged In 1999, 0 if otherwise)
Temporal variable 2 ( 1 if selectively logged In 1996, 0 if otherwise)

In the selective logging probit model the control variables were also

indexed to a time scale (temporal variables), to allow the assessment of selective

logging spatiotemporal changes. Finally, I created three variables (distances to

129

deforested areas, roads, and timber centers squared) to observe the spatial
distribution of selective logging at very large distances from logging centers,
deforested areas, and road network.

4.3.4.3. Data sampling

As previously indicated, the present analytical exercise focused on a study
region where evidences of selective logging and forest fire have been observed
using remotely sensed data, including 'portions of the states of Para, Mato
Grosso, Tocantins, Acre, Rondbnia, and Maranhéo. This option helped to
improve the sampling process and the statistical model performance. Analyses
involved datasets for 1992, 1996, and 1999, which is the most recent year of
mapping of forest fire and selective logging available for the Brazilian Amazon.

In this spatially explicit probabilistic analysis of selective logging and forest
fires, a stratified random sampling was applied to collect spatial data from point
locations. First, a vector of 5x5 km grid cell size was located evenly over the
study region. Subsequently, one point was randomly located within each 25 km2
grid cell (Figure 4.1).

The pixel was used as the unit of observation and, therefore, a unique
value of the dataset was collected for each point of the sampling grid to represent
forest fire (1 if burned, 0 if othenlvise) and another to represent selective logging
(1 if logged, 0 if otherwise) (Appendix D.1). Similarly, a pixel value was collected

for each independent variable and point location.

130

 

 

 

 

 

 

 

 

 

 

 

 

 

55:: T—T'l Rand”

. I sampling
5 0 5 10 15 20Km j _ , points

 

 

 

Figure 4.1 Stratified random sampling used to collect spatial data for the probit
model of forest fire and selective logging in the study region.

This stratified sampling procedure was designed to reduce the effects
autocorrelation in the dataset (Anselin 2002, Arima et al. 2007). The random
distribution of points within a regular sampling network can reduce effects of
potential coincidence between spatial patterns of landscape and sampling points
(Burrough, 1998). Stratified and random sampling combined can minimize
statistical biasness of as given model (Burrough, 1998, Anselin, 2002) and the
error term in the probit models due to spatial autocorrelation in the data

population (Chomitz and Gray 1996, Arima et al 2007).

131

4.4. Results and Discussion
4.4.1. Forest fire probabilistic model

The analysis of forest fire dataset showed that although forest fire
observation was relatively rare (unconditional probability of forest fire = 0.77%) in
the study region, encompassing more than 1.7 million km2 of tropical rainforests
by 1999, it has increasingly become a socioeconomic and environmental concern
in the last decades. The statistical significances of the previously discussed
factors that affect forest fire probability were estimated using a probit regression
model (Table 4.4).

The estimated model predicts that forest fires are inversely related to
distance to deforestation (agricultural lands). The probability of forest fires
proved to be significant both statistically and substantively. Forest fires increased
as distance to deforestation decreased. For example, by decreasing distance to
deforestation areas from 10 to 0 km would increase forest fire by 42%. The
predicted non-linear effect of distance to deforestation in the probability of forest

fire is presented in Figure 4.2.

132

Table 4.4. Probit regression analysis results of forest fire

 

 

Variables and Constant Coefficient Rob. St. P > |z| dF/DX
Error

Protected areas ( 1 if yes, 0 otherwise) -0.1955 0.0681 0.004 -0.00002
Distance to deforested areas (km) -0.1594 0.0144 0.000 -0.00002
Distance to deforested areas squared 0.0011 0.0001 0.000 0.0000001
Distance to roads (km) -0.0294 0.0043 0.000 -0.000003
Distance to roads squared 0.0002 0.00003 0.000 0.00000002
Selectively logged 1999 (1 if yes, 0
otherwise) 0.7075 0.0567 0.000 0.0004
Selectively logged prior to 1999 (1 if yes, 0
otherwise) 0.5716 0.0853 0.000 0.0002
Water deficit between May and September 0.0021 0.0008 0.011 0.0000002
Patch size (log area m2) -0.0113 0.0025 0.000 -0.000001
Fractal dimension ( 1 to 2) -0.2445 0.0444 0.000 -0.00003
Vegetation type: (omitted Pioneer)
Dense forest 0.2033 0.0970 0.036 0.00003
Open forest 0.4145 0.0969 0.000 0.00007
Semi-deciduous forest 0.4337 0.1591 0.006 0.0001
Deciduous forest 0.0210 0.1429 0.883 0.000003
Transitional forest (Forest/Savannah) 0.4555 0.0936 0.000 0.00009
State: (omitted Acre)
Para 0.9668 0.2224 0.000 0.0002
Maranhao 0.7751 0.2276 0.001 0.0005
Tocantins 0.3608 0.2517 0.152 0.0001
Mato Grosso 1.0190 0.2224 0.000 0.0005
Rondbnia 0.1016 0.2342 0.664 0.00001
Year: (omitted 1992)
1999 0.5476 0.0458 0.000 0.0001
1996 0.3067 0.0474 0.000 0.00005
Constant -2.8753 0.2439 0.000

 

Log likelihood = 418839944; Pseudo R2 = 0.2517; n= 144,793

 

I also estimated the effect of distance to deforestation in quadratic form.
As opposed to the estimated parameter of the level term, the squared term was
positive indicating that probability of forest fires eventually increases as distance
to deforestation increases at very large distances from deforestation. Indeed, I
could observe some isolated forest fires in the study region, geographically
located away from deforestation areas, which could be associated with selective

logging activities.

133

 

Probabiliy

 

0.50
0.45
0.40
0.35
0.30
0.25
0.20
0.15
0.10
0.05
0.00

 

 

 

 

 

 

 

 

—<>— Forest Fire

 

 

 

 

 

_;A§—__ -__- __ ____E
-___\.. WW 3%- _ ___ -2
\
Av.
\
_,_. \ _v a i
\
E,____ __ ___\\ _
\
..._. _ _ _u _ ___E.-\
\

_._- \ ,,

\

\
_.___ _ E _E \ \
0 2 4 6 8 10 12 14

Proximity to deforestation (km)

 

Figure 4.2. Probability of forest fire with respect to distance to deforestation

forest fire occurrence is inversely related to distance to roads. By decreasing
distance to roads from 50 to 0 km, forest fires would increase 7%. The predicted

non-linear effect of distance to roads in the probability of forest fire is presented

Similarly to distance to deforestation, the estimated model predicts that

in Figure 4.3.

The effect of distance to roads in quadratic form is positively related to

forest fires, indicating that probability of forest fires eventually increases as

distance to roads increases at very large distances from roads. It is likely that this

type of spatial pattern of forest fire occurs during logging operations in new

timber sources located at large distance from roads.

134

 

 

0.08
0.07 -
0.06 —
0.05 -
0.04
0.03 .
0.02 ~ ~—
0.01 1

0 4 8121620242832 3640444852 56 60

Proximity to roads (km)

 

 

 

 

 

—+—- Forest fire I

 

Probability

 

 

 

 

 

 

 

 

Figure 4.3. Probability of forest fire with respect to distance to roads
The water deficit observed in the study region significantly contributed to
increase the likelihood of forest fires at 95% probability. The direct effect is that
higher water deficit level increases forest flammability. For example, by
increasing the water deficit from 0 to 200 millimeters during the dry season (May
to September) forest fire likelihood would increase 2.5%. The predicted effect of

water deficit in the forest fire likelihood is presented in Figure 4.4.

135

0.045

0.040 ‘
0.035
l

:E‘ 0.030
5 0.025 W ,7,
g +ForestfiEI
9 0.020 ‘7 T ’
LL

0.015
0.010
0.005
0.000

 

0 20 40 60 80 100120140 160 180 200 I

Water Deficit (mm)

Figure 4.4. Probability of forest fire with respect to water deficit from May to

. *4

September

The model predicted that the effect of an increase in forest fragmentation
was significantly related to forest fires. The results showed that the applied
measurements of forest fragmentation (forest patch size and fractal dimension)
are significant explanatory variables, because the average probability of forest
fires increased by 1% and 2% as fractal dimension decreased from 2 to 1 and
forest patch size decreased from 533,000 km2 to 1 kmz, respectively. These
increases in absolute probabilities might be considered small in absolute terms,
but they are relatively high considering that the unconditional probability of forest
fire is 0.77% in the study region. The predicted effect of forest patch size and
shape (fractal dimension) in the probability of forest fire is presented in Figure

4.5.

136

 

Fractal dimension
1 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2
0.040 I I I I l l I I I I I I l 1' l I I I i l

 

0.035
0.030 -
0.025 -

Vi—a—iPiatch Size
0.020 -

Probability

1+ éhepe
0.015

 

0.010 -
0.005 -

 

 

 

0000 ALL]!ll'llllllllllllllllltgllgllllllglllll

0481216202428323640
Log of forest patch size

 

 

 

Figure 4.5. Probability of forest fire with respect to forest patch size and shape
Meanwhile, selective logging directly affects forest susceptibility to fires.
When compared to undisturbed forests, newly and previously detected selective
logging areas significantly increased the chance of fire 3.4% and 2.7%,
respectively. Selective logging contributes to increase canopy openness and
decrease moisture (Uhl and Kauffman, 1990), and hence, higher forest
flammability. In addition to selective logging effects, probability of forest fires was
1% greater outside of protected areas than inside of them. This result supports,
at least temporarily, the assumption that environmental laws and conservation
policies for protected areas has been successfully applied in the Brazilian
Amazon, contributing to impede or slow anthropogenic activities (e.g. forest fires)

Most of the vegetation dummy variables were statistically significant. Open, semi-

137

deciduous, and transitional forests with 2% of likelihood of forest fires were the
most fire-prone forest types by holding all other independent variables fixed at
mean value. Dense forest, with 1% of likelihood of forest fires was the least fire
prone forest type at 95% probability. Deciduous forest was not statistically
different from the omitted forest type (pioneer vegetation). These results show
that dense forest is the predominant and least fire prone vegetation type in the
Brazilian Amazon region. However, open and transitional forests become of
special concern because they were two-fold more likely to fire than dense forest,
and, combined, they cover approximately 46% of the study area. The predicted
effect of forest types, state political-administrative peculiarities, selective logging,
protected areas, and temporal variables on the probability of forest fire is
presented in Figure 4.6.

The state dummy variables for Rondonia and Tocantins were not
statistically different from the omitted state (Acre) in terms of forest fire
probability. The dummy variables for Para, Maranhao, and Mato Grosso were
both statistically and substantively significant. Mato Grosso has 5.7% greater
chance of experiencing fire than the omitted state, while Para and Tocantins
have 5.4% and 4.4% likelihood of fire, respectively. These results show that Mato
Grosso, Para, and Tocantins are the most fire-prone states in the Brazilian
Amazon by holding all other factors constant at mean value. However, the extent
of the “arc of forest fire” as defined by the present study does match perfectly the

arc of deforestation defined by (IBAMA, 2007). It can be hypothesized that

138

potential differences in state environmental policies and legislation might be

affecting forest fire likelihood in the Brazilian Amazon.

 

 

Probabiity

 

[El Para
Mato Grosso
‘ I Open forest
’El Transitional
‘ I Logged by 1999

’IYear1996

 

 

 

 

 

E] Maranhao Tocantins ‘
El Rondonia I Dense forest ‘
I Semi—Deciduous I Deciduous

Outside of protected areas I Inside of protected areas . .
I Logged by 1996 I Year 1999 ,

l
T T T 7 TTT l

Figure 4.6. Probability of forest fire with respect to forest types, states, selective

logging, protected areas, and year of analysis

Finally, the multi-temporal dummy variables proved to be statistically

significant. The year 1999 and 1996 had 3.4% and 2.6% greater chance of

experiencing fire than the omitted year (1992). These results contribute to explain

the temporal changes of forest fires which showed to be consistently and

significantly increasing along the period of analysis.

4.4.2. Selective logging model

Based on the selective logging dataset the unconditional probability of

selective logging is 1.15% for the study region, which encompassed more than

1.7 million km2 of tropical rain forests by 1999. Additionally, the statistical

139

significances of various factors that affect selective logging probability were
estimated using a probit regression model. The results of probit regression
analysis of selective logging are presented in Table 4.5.

Table 4.5. Probit regression analysis results of selective logging

 

 

Rob.
Variables and Constant Coefficient St. P > lz| dF/DX
Error
Protected areas ( 1 if yes, 0 otherwise) -0.3168 0.0486 0.000 00001
Distance to deforested areas (km) 0.0051 0.0132 0.697 0.000003
Distance to deforested areas squared -0.0017 0.0009 0.076 -0.0000009
Distance to roads (km) -0.0383 0.0031 0.000 000002
Distance to roads squared 0.0003 0.0000 0.000 0.0000001
Distance to timber center (km) -0.0058 0.0006 0.000 -0.000003
Distance to timber center squared -0.0000004 0.0000 0.873 00000000002
Timber volume (m3/ha) 0.0009 0.0004 0.031 0.0000005
Vegetation type: (omitted Pioneer)
Dense forest 0.4229 0.1054 0.000 0.0003
Open forest 0.3101 0.1106 0.005 0.0002
Semi-deciduous forest 0.2576 0.1754 0.142 0.0002
Deciduous forest -0.0714 0.1596 0.655 -0.00004
Transitional forest (Forest/Savan nah) 0.6567 0.1071 0.000 0.0008
State: (omitted Acre)
Para 0.9108 0.1948 0.000 0.0008
Maranhao 1.2005 0.1980 0.000 0.007
Tocantins -0.4176 0.3580 0.243 -0.0001
Mato Grosso 0.9233 0.1940 0.000 0.001
RondOnia -0.1841 0.2086 0.378 -0.00008
Year: (omitted 1992)
1999 0.6677 0.0297 0.000 0.0007
1996 0.2478 0.0323 0.000 0.0002
Constant -3.1936 0.2251 0.000

 

Log likelihood = -6947.22 ; Pseudo R2 = 0.236; n= 144,793

 

Based on the model results, I estimated the effect of distance to

deforestation in linear and quadratic forms. As opposed to the forest fire probit
model, the selective logging probit model estimated that neither the parameter of
the level term nor the squared term were statistically significant, indicating that
distance to deforestation has no effects in probability of selective logging. The

estimated model predicted that selective logging is inversely related to distance

140

to roads and, therefore, selective logging increases as distance to roads
decreases. Probability of distance to roads proved to be statistically significant.
For example, a decrease in distance to roads from 30 to 0 km would increase
selective logging by 2.8%. Similarly, selective logging is inversely related to
distance to timber centers. The probability of distance to timber centers was also
statistically significant. These results show that natural forests become less
attractive to loggers if they are obligated to move away from roads and logging
centers to harvest timber. The predicted non-linear effect of distance to roads
and logging centers in the probability of selective logging is presented in Figure

4.7.

 

Distance to roads (km)
0 3 6 9 121518212427 303336394245

0035 l ‘l llllAilli‘i l‘ l
. TIIII’ITIIIIIITIIIIIIIIIIIIII‘ITT

 

0.030 r
0.025 T
0.020 -- EQands

0.015 _ + Timbericentersg

Probability

0.010 -~
0.005 ~

  

 

 

   

IlllllllllillllllTIII
i , v

0 21 42 63 84 105126 147 168 189 210
Distance to timber centers (km)

 

 

 

Figure 4.7. Probability of selective logging with respect to distance to roads and
timber centers
Meanwhile, I also estimated the effect of distance to roads and logging

centers in quadratic form. As opposed to the estimated parameter of the level

141

term, the squared term of distance to roads was positive indicating that
probability of selective logging eventually increases as distance to roads
increases at very large distances from roads. This supports the assumption that
loggers, although reluctantly, could direct their attention toward undisturbed
forests geographically located farer away from existing road networks if timber
sources become too scarce in those forests close to the road network .The
estimated parameter of the squared term of distance to timber centers was not
statistically significant, which indicate no effect of distance to timber centers at
very large distances from saw mills in the probability of selective logging
occurrence.

The model predicted that the effect of an increase in timber volume was
positively related to selective logging. The results showed that timber volume is
significant at 95% probability, with the average probability of selective logging
increased by 0.1% as timber volume increased 100 m3. Again, although such
increases in absolute probabilities might be considered small in absolute terms,
they are considerably high given that unconditional probability of selective
logging is 1.15% for the entire study region. The predicted non-linear effect of

timber volume in the probability of selective logging is presented in Figure 4.8.

142

 

0.010
0.009 —~ — T T V
0.008
0.007
0.006
0.005
0.004
l 0.003
0.002 ,. _‘ g
0.001 Tam-m . ___T

" li‘llTTlliilll-TTTlllillll

0 60 120 180 240 300 360 420
Tlmber volume (m3/ha)

 

 

 

 

 

+ Selective logging

 

 

Probability

 

 

 

 

 

 

 

 

Figure 4.8. Probability of selective logging with respect to timber volume

Vegetation types

The vegetation dummy variables for dense, open, and transitional forests
were all statistically significant. Transitional forest, with 0.9% of likelihood of
selective logging was the most selective logging-prone forest type when all other
factors are held constant at mean value. Dense and open forests showed 0.7%
and 0.6% of likelihood of selective logging, respectively. Deciduous and semi-
deciduous forests were not statistically different from the omitted forest type
(pioneer vegetation). Although transitional forest is not the largest forest type and
does not have the highest timber volume per hectare, it showed to be the most
attractive ecotype to loggers. As previously presented, transitional forest also
have been object of more forest fires than any other forest types. The predicted
effect of forest types, states, protected areas, and temporal variables in the

probability of selective logging is presented in Figure 4.9.

143

 

0.020
0.018
0.016
0.014
0.012
0.010
0.008
0.006
0.004
0.002

Probability

 

 

 

El Para uTmnhao Tocantins , fl
Mato Grosso u Rondonia I Dense forest

I Open forest I Semi-deciduous I Deciduous

[j Transitional Outside of protected areas I Inside of protected areas

‘l-Year1999 lYear1996 ‘

 

—4 Tl

 

 

 

Figure 4.9. Probability of selective log’Wpemegetation
types, protected areas, and year of analysis

The probability of selective logging was three-fold greater outside than
inside of protected areas with 0.6% and 0.2% of likelihood of selective logging,
respectively. This result shows that protected areas have been reasonably
effective in impeding or even slowing the pace of selective logging in the
Brazilian Amazon

The state dummy variables for Rondénia and Tocantins were not
statistically different from the omitted state (Acre) in terms of selective logging
probability. The dummy variables for the states of Para, Maranhao, and Mato
Grosso were statistically significant. Natural forests located within the state of
Maranhao have 1.8% greater chances of selective logging than the omitted state

(Acre), immediately followed by forests located in Mato Grosso and Para, both

144

with 1.3% likelihood of selective logging and by holding all other factors constant
at the mean value. Similar of what was observed for forest fires, the differences
in state environmental policies and legislation may be affecting the likelihood of
selective logging within each state.

And, finally, the multi-temporal dummy variables also proved to be
statistically significant. The year 1999 and 1996 had 0.9% and 0.5% greater
chance of selective logging than the omitted year (1992). These results and the
results presented in chapter 2 and 3 show that selective logging activities are
increasing in area, intensity, and over time in Brazilian Amazon. (Matricardi et
al., 2007) speculated that major factors contributing to the increase of selective
logging during the period of analysis were the changes in the Brazil’s
environmental law, the scarcity of raw material for timber industries, and the
occurrence of extreme climatic events. Indeed, (Uhl et al., 1997) had observed
that loggers were going farther searching for new timber sources (undisturbed
forests) as well as revisiting previously logged forests for species that recently
became marketable.

4.5. Conclusions

The results derived from the two probit models could explain
important spatiotemporal dynamics of selective logging activities and forest fires
in the Brazilian Amazon. Based on these results, the occurrence of forest fires
were highly correlated with distance to deforestation. This result is supported by
Uhl and Buschbacher (1985) who observed that agricultural lands are important

ignition sources for forest fires in the Brazilian Amazon. The likelihood of forest

145

fire events is even greater if additional anthropogenic and environmental factors
that may increase forest flammability are considered. For example, selectively
logged forests were more susceptible to fire than undisturbed forests. This fire
susceptibility due to selective logging, however, tends to decrease as forests
recover from logging impacts and, four years following selective logging
activities, forest moisture can be similar to undisturbed forest (Holdsworth and
Uhl, 1997). It is very important to point out that although selective logging
seemed to be a factor fueling the increase of forest fires in the study region
during the period of analysis, a large amount of selective logging activities and
forest fires were spatially distinct events. While distance to deforestation is highly
related to forest fires, this factor was not statistically significant in selective
logging probability. For instance, the reality of timber scarcity in the old frontier
has forced loggers to move ahead of deforestation frontier searching for new
timber sources in undisturbed forests (Schmink and Wood, 1992, Pinedo-
Vasquez et al., 2001).

.Prolonged water deficit also showed to be an important factor that might
increase likelihood of forest fires. This result is also corroborated by different
authors (Uhl and Buschbacher, 1985, Nepstad et al., 1999, Houghton et al.,
2000, Nelson, 2001) who have observed that extreme droughts can increase
likelihood of forest fires in tropical regions. Complementarily, (Nelson, 2001)
reported that even undisturbed forests became flammable in the Brazilian
Amazon upon a severe drought as the one related to the El Nifio event of

1997/98. Similarly, forest fragmentation significantly contributed to increase the

146

chances of forest fires. Indeed, deforestation activities are responsible for
creating non-continuous and small forest patches and for increasing forest edges
with agricultural lands (Skole and Tucker, 1993, Cochrane et al., 2004, Skole et
al., 2004) that results in greater forest flammability (Uhl and Kauffman, 1990,
Cochrane, 2001, Cochrane et al., 2004). This anthropogenic phenomenon was
studied by (Cochrane, 2001), who concluded that forest fires were strongly
correlated with fragmented forest edges in a study site located in the state of
Para, Brazil.

In additional to the previously discussed factors, I also observed that
probability of forest fires and selective logging varied according to forest types. It
is presumable that potential differences in soil and forest moisture present in
these ecosystems may affect forest fire susceptibility. In this analysis, open,
semi-deciduous, and transitional forests were more likely to get on fire than
dense and deciduous forests. This effect of vegetation type on forest fire
probability is supported by (Nelson, 2001) who observed that secondary and
igapo forests were more vulnerable to fire than any other vegetation types during
the El Nif‘lo event of 1997-98 in the state of Roraima. Similarly, selective logging
activities mostly occurred in transitional, dense, and open forests, which are the
predominant vegetation types in the Brazilian Amazon. Transitional forests,
however, showed slightly higher probability of selective logging than dense and
open forests. In this case, I hypothesized that economic factors (e.g. lower
harvesting cost and higher productivity) are driving loggers towards this particular

forest type.

147

Forest fire and selective logging likelihood also varied according to the
land ownership status. The legal status of protected areas seemed sufficient to,
at least, decrease the likelihood of the occurrence of fire and selective logging
within their territorial boundaries. However, given the observed increase of
selective logging and forest fire between 1992 and 1999, it seems reasonable to
expect that human pressure will increase in the future over the natural resources
existent within the protected lands established in the Brazilian Amazon.
Pedlowski et al. (2005) have already detected the occupation of areas and
exploitation of natural resources within the borders of protected lands in the state
of Rondénia.

Meanwhile, the greatest chances of forest fire and selective logging were
found in the states of Para, Mato Grosso, and Maranhao. These states fall within
the infamously famous “arc of deforestation” defined by IBAMA (2007).
Additionally, the states of Mato Grosso and Para were reported as responsible by
most of selective logging detected using remotely sensed data in 1992, 1996,
and 1999 in the Brazilian Amazon (Matricardi et al., 2007). Eventual differences
in state environmental legislation and law enforcement capabilities may be a
factor that promotes or hamper the occurrence of selective logging activities and
mmflfims

In terms of temporal changes, the results showed that selective logging
and forest fires significantly increased between 1992 and 1999. (Matricardi et al.,
2007) mentioned that scarcity of timber resources, changes in environmental law,

and extreme climatic events are perhaps the most important factors that

148

contribute to increase selective logging in the period of analysis. I would
hypothesize that the process of increasing socioeconomic pressure over
undisturbed forests will continue, unless intensive environmental monitoring and
law enforcement combined with governmental incentives for alternative uses of
land forest resources are implemented in Brazilian Amazon.

I also hypothesize that at least half of forest fires detected in 1992 and
1996 were ‘intentional’ rather than ‘accidental’ fires, which indicates that
landowners intended to burn those forests in order to achieve rapid conversion to
agricultural lands. This inference is supported by the results shown in Chapter 3
that indicated that more than 50% of burned forests detected in 1992 and 1996
had been deforested by 1999. This factor, however, was not included in the
presented probabilistic models.

Finally, it is my hope that the bulk of the results of this research can be
used in one hand the expand the modeling of land cover change in the Brazilian
Amazon, and on the other hand, also be used to improve governmental
programs aimed at enhancing the control of forest fires and the environmental

monitoring in the Brazilian Amazon region.

149

CHAPTER V
Concluding Remarks
5.1. Seating the current study in the context of global change research

In 2000, the area of the Earth covered with forests was estimated as being
around 39 million square kilometers. Most of those forest areas were located in
tropical areas (47%) and boreal zones (33%). The rest of the forested areas
(18%) were located in temperate and subtropics regions. Most of the forests
(95%) were then classified as natural forests and a smaller fraction (5%) as
forest plantations (VVRI, 2000). Meanwhile, tropical forests are estimated to store
460 to 575 billion metric tons of carbon, an average of 180 metric tons of carbon
per acre (NASA, 1998). Therefore, the highest carbon and biodiversity losses
occur when a high biomass forest area is converted into low biomass systems
such as pasture or agriculture; releasing carbon and other trace gases to the
atmosphere while natural habitats are gravely disturbed. The release of trace
gases into the atmosphere enhances the Greenhouse effect which contributes to
higher temperatures and, consequently, to dramatic global climatic changes
(Skole and Tucker, 1993, NASA, 1998, Skole et al., 2004).

Deforestation (i.e. conversion of natural forests to non-forest land covers)
is annually destroying a significant portion of the world forests. The gross and net
rates of global forest loss averaged 17 and 13 million hectares annually, in the
19808. Moreover, the gross and net rates of global deforestation grew to an
average of 14.6 and 9.4 million hectares annually in the 1990s. The 19903 rates

of annual forest loss correspond to 0.4% globally and 0.8% in the tropics.

150

Consequently, the highest deforestation rates have been observed in developing
countries in the tropics (F RA, 2005).

Deforestation in the tropics is a critical factor that has several implications
for biological diversity, local habitats, and carbon cycle at global scale (Skole and
Tucker, 1993, Skole et al., 2004). Approximately 15.4 and 16 million hectares of
forest loss was annually observed in tropical countries during 1980-90 and 1990-
00 periods, respectively (Matthews, 2001). More specifically, multi-temporal
measurements of deforestation in the Brazilian Amazon provided by (INPE,
2007) demonstrate that more than 2 million hectares of natural forests were
slashed annually from 1978 to 2005 in the region.

In addition to deforestation, several other processes of land use and land
cover change may increase forest degradation in tropical regions such as forest
disturbances by selective logging, forest fires, and forest fragmentation (Skole et
al., 2004). Skole and Tucker (1993) estimated that areas of disturbed forests due
to forest patch isolation and its edge effects in the 19803 were 150% larger than
deforested area itself in the Brazilian Amazon. Because of fragmentation, forests
will become more susceptible to fire due to an increase of forest disturbances
associated to edge effects including the exposure to fires (Cochrane et al., 2004).
Based on sawmill survey in the Brazilian Amazon, Nepstad et al.(1999)
estimated that the current carbon flux estimates for the Amazon region may be
underestimated because of logging areas that have not been accounted for.

Nevertheless, in spite of the important contribution of these land use

processes to global climate change, only a few attempts have been made to

151

detect and measure forest disturbances, and its extent and impacts using
remotely sensed data gathered for tropical regions. As a result, the total area of
forest damaged by human activities may have been underestimated. Therefore,
relevant questions have yet to be addressed before a better understand of
selective logging and forest fire dynamics, their impacts on tropical forests, and
their interactions with other land use types is achieved.

In this research, based on data obtained remotely and in field studies, I
presented new estimates of the extent of selectively logged and burned forests
for the Brazilian Amazon basin for 1992, 1996, and 1999. I also estimated the
impacts caused by selective logging and forest fire in the region. The results
show that the total combined areas of selectively logged and burned forest
increased from approximately 11800 km2 to 35600 km2 from 1992 to1999.
Selective logging alone was responsible for 60.4% of forest disturbances in the
studied period. My estimates indicate that the total active areas of selective
logging and forest fire in the Brazilian Amazon have substantially increased from
5468 km2 to 7620 km2 and from 7620 km2 to 17439 km2 in 1992-1996 and 1996-
1999 periods , respectively.

These amount of disturbed forests by selective logging and forest fires
represent additional land use types that have yet to be properly accounted for.
Skole et al. (2004) affirmed that these processes of forest degradation can
substantially degrade tropical forests and increase loss of biomass. F urtherrnore,
the impact of selective logging on undisturbed forests tropical forests creates a

net carbon flux equal to approximately 4-7% of the annual carbon release

152

generated by deforestation (Fearnside, 1997), which can be severely aggravated
by forest fires (Uhl and Kauffman, 1990).

The intensity of forest disturbances is also increasing in the Brazilian
Amazon. As previously mentioned, selective logging was responsible for the
largest extent of forest disturbance Amazon-wide. Forest fires, however, were
responsible for the greatest impact on canopy cover, causing an average of
18.8% of canopy loss and even greater average of 27.5% of canopy loss if forest
fire and selective logging results were combined. This result is supported by (Uhl
and Kauffman, 1990) who observed that forest fires in tropical forest can be
devastating because of their natural low resistance to fire. Meanwhile, selective
logging was responsible for around of ~2% of forest canopy losses. These results
indicate that although selective logging may destroy 40-50% of canopy cover in
particular study sites (Uhl and Vieira, 1989, Verissimo et al., 1992), it has lower
impacts on forest canopy cover at the regional scale.

Finally, the results of this analysis must be considered conservative, since
under different circumstances, the most cryptic forest degradation by selective
logging and forest fires cannot be properly detected. Indeed, Matricardi et al.
(2005) and Matricardi et al. (2007) observed that very low intensity selective
logging cannot be easily detected using remote sensing approaches. This type of
forest disturbance, however, does not cause substantial environmental impacts

on natural forests and on fluxes of carbon dioxide.

153

5.2. Research questions revisited

o Is selective logging and forest fire causing more forest degradation than

othem/ise would occur?

According to INPE (2007) estimates, a total area of 440186 kmz, 517065
kmz, and 569269 km2 of tropical forests in the Brazilian Amazon had been
deforested by 1992, 1996, and 1999 respectively. Those estimates from INPE
also indicate an average increase of deforestation of 18331 km2 year“1 between
1992 and 1999. The results of the present research indicate that the total area
affected by selective logging and forest fire correspond to 2.7%, 3.2%, and 6.3%
of total deforestation by 1992, 1996, and 1999, respectively. In addition, results
showed that active selective logging and forest fire represented an additional of
27.4%, 38.1%, and 87.3% to the outright increase of deforestation in 1992, 1996,
and 1999, respectively. Therefore, selective logging activities and forest fires are
substantially affecting large extents of forested areas in the Brazilian Amazon.
Further results from the case study in Mato Grosso indicate that this trend in
forest degradation by selective logging and forest fires may persist, while natural ,
resources are abundant and accessible in a given area. However, forest fires
appeared to be the most aggressive type of forest disturbance, causing an
average decrease of 18.8% in forest canopy cover. When forest fires and
selective logging were combined, the effect is even greater, leading to an
average decrease of 27.5% in forest canopy cover, whereas selective logging

alone contributed to a decrease of 2% of forest canopies in the study region.

154

Despite the fact that selective logging only contributed to a decrease of 2%, this
process impacted the largest extent of natural forests in the Brazilian Amazon
during the period of study. Nevertheless, this study demonstrated that the area
covered with natural forests under the impacts of fires and selective logging is
rapidly increasing, making these land uses of greater concern. Finally, forests
impacted by fires and selective logging showed the ability to rapidly recover from
these disturbances if left to fallow, for periods ranging from 5 to 10 years.
Therefore, forest regeneration is another very important phenomenon that must
be further investigated because it may indicate forest uses and management
alternatives aimed at mitigating impacts brought about by anthropogenic

activities in tropical regions.

o Can we accurately measure these anthropogenic activities in the Brazilian

Amazon basin?

The automated remote sensing technique based on Spectral Mixture
Analysis was an efficient methodological approach for estimating the area
affected by forest fire in the Amazon region. This approach was previously
employed by Cochrane and Sousa Jr (1998) and Souza et al. (2003) to detect
selectively logged and burned forests using Spot imagery in a case study in
Para. I further improved, tested, and validated this remote sensing technique for
a 30000 km2 study site in the state of Mato Grosso. The accuracy assessment
results of this forest fire detection approach indicated 96.0% and 0.91 of overall

accuracy and overall kappa statistic, respectively. This technique showed high

155

sensitivity to changes in Non-Photosynthetic Vegetation (NPV) within tropical
forests. NPV grows substantially when forests are affected by fires, and its
effects are prolonged throughout contiguous forested areas. Indeed, (Nelson,
2001) reported that even ground fires could be detected using Landsat TM
imagery. Therefore, based on my remote sensing analyzes and findings of the
field studies, I believe that the most fire-impacted forests were accurately

detected and mapped.

The semi-automated texture analysis and visual interpretation combined
also showed to be an efficient remote sensing approach to detect and estimate
areas affected by selective logging. The results of the research conducted by
Matricardi et al. (2007) showed that most of selectively logged forests were
detected and mapped using semi-automated and visual interpretation
approaches. These remote sensing approaches showed 92.9% and 0.82 of

overall accuracy and overall kappa statistic, respectively.

However, I must advise that the results of these methods must be considered
as conservative, since, in different circumstances, some cases of selectively
logged and burned forests could not be detected. | hypothesize that very low
intensity ground fires that did not affect canopy cover were difficult to detect
using remote sensing approaches. Similarly, selective logging activities may not
be easily detectable using Landsat imagery because of lack of soil exposure by

logging features (patios, access roads, etc.) on the ground.

156

o Where are the most severe selective logging and forest fire occurring in that
region? Why selective logging and forest fire are occurring there and not

somewhere else?

Based on the findings of the basin-wide results, most of the impacted
forests by selective logging and fires were detected in Mato Grosso, Para, and
Maranhao, which are geographically located within the so-called “Arc of
Deforestation” defined by the Brazilian Institute of Environment and Renewable
Natural Resources as a distinguished zone highly affected by anthropogenic
activities in the Brazilian Amazon. While spatial distribution of forest fires were
highly correlated with proximity to deforestation areas, selective logging activities
were more related to proximity to logging centers and road networks. These
findings are supported by Uhl and Buschbacher (1985), Uhl and Kauffman
(1990), and Cochrane (2001), who observed that fires from fire-maintained
agricultural lands are the main ignition sources for forest fires in the Brazilian
Amazon. Additional factors such as prolonged water deficit, selective logging,
and forest fragmentation also contributed to increase forest fires, though a
combination of all these factors seemed to be a more probable reason for forest
fire events occurring in the region. Meanwhile, I also observed that the spatial
distribution of selective logging activities is strongly influenced by the distance to
logging centers and road networks, which may be associated to transportation
costs. As a consequence, loggers have to improve their transportation efficiency
continuously to better exploit timber sources in new frontiers of undisturbed

forests as the forests next to logging centers become overexploited. During the

157

field work in Mato Grosso I could observe that loggers are commonly using
heavy-load (30 to 40 tons) trailer trucks to transport logs from forest sites to
sawmill. Finally, the status of land ownership seemed to play an important role in
the spatial distribution of both selectively logged and burned forests. For
example, only 3.7% of the total impacted forests by selective logging and fires
were geographically located within protected areas in the period of analysis.
There is growing evidence, however, of an increasing trend for the occurrence of

anthropogenic activities within the limits of protected areas.

o Is there any spatial pattern of these activities in that region? Are those

spatial patterns changing over time?

The results of this research show that forest fires and selective logging activities
were occurring throughout the Brazilian Amazon at an increasing intensity during
the period of 1992 to 1999 (see appendices 0.12 to 0.17). Forest fire events that
could rarely be observed in Rondonia and Acre in 1992 were evident in 1996 and
1999 in both states. The most meaningful forest fire and logging clusters were
located in the Eastern and Southern Para, Western Maranhao, and Northeastern
and Mid-Northern Mato Grosso in 1992, 1996, and 1999. The analysis for 1999
logging and fire mosaics (appendices C.12 to 0.17) revealed the surfacing of
new logging centers and fire events in Northwestern and Southwestern Para, and
Northwestern Mato Grosso. These mosaics also show that selective logging
activities and forest fires have encroached on protected areas in the study

region. This spatial feature is especially clear in the Northeastern Mato Grosso.

158

o Is there any interaction between selective logging and other land use or land
cover type that may be synergistically causing even more forest

disturbances?

As previously presented in Chapter 4, selective logging activities significantly
increased forest fire likelihood in the Brazilian Amazon. This fact is supported by
Uhl and Buschbacher (1985), Uhl and Kauffman (1990), and Cochrane (2001),
who observed that selective logging increases forest fire susceptibility in tropical
regions. The proximity to agricultural land, however, was the factor that most
substantively and significantly affected probability of forest fires in the study
region. On the other hand, this factor did not significantly affect selective logging
likelihood in the study region. Further juxtaposition between selective logging and
forest fire maps indicate less than 7% of overlap between these types of forest
disturbances in the period of analysis. These facts indicate that, at the regional

scale, most of selective logging and forest fire were spatially distinct events.

The results from the basin-wide analysis of forest fires also indicate that only
around of 457 km2 could be considered recurrent fires detected over the study
period. Moreover, field measurements showed that fuel load averages for study
sites in Acre, Rondonia, and Mato Grosso were not significantly different for
various forest uses (i.e. logged, burned, and undisturbed forests). Altogether,
these facts indicate that the positive association between selective logging and
forest fires at the local scale observed by many authors (Holdsworth and Uhl et

al. 1997, Cochrane 1999, and Cochrane and Schulze 1999) may not be

159

supported by a basin-wide analysis, since most of the selective logging and
forest fires detected did not occur at the same place and time. The results and
field measurements also indicated that some forested areas are naturally more
fire prone than others, whether logged or not, which indicates that forest fire
events at a particular site require very unique local conditions. These may
include, but are not limited to, water deficit, forest type, forest disturbance, and

the presence of ignition sources.

5.3. Hypotheses revisited

Hypothesis 1: Forest impactiby selective Ioggjmctivities increase over ting
Ml severely disturb nettle] forests in the B__razilian Amam

This hypothesis may be considered true since my results showed that areas of
selectively logged forests detected using remotely sensed data increased 500%
between 1992 and 1999. These results also showed that selective logging alone
contribute to degrade around of 2% of forest canopy compared with undisturbed
forests. Selective logging and forest fire combined substantively increases
canopy degradation up to 27.5%. Moreover, selective logging is increasing in
area and intensity throughout the Brazilian Amazon. It is also important to point
out that forest canopy regeneration was very strong (~6% greater than
undisturbed forests) in previously logged forest detected in the period of analysis.
At the local scale, field measurements indicated that newly selectively logged
forests may significantly decrease around 6.7% of canopy cover. lf selective

logging and fire combined, canopy cover is decreased by 7.7%. Finally, diversity

160

index, tree mortality and fuel load rates for logged sites were not significantly

different from undisturbed forests.

Hypothesis 2: Forest fires mostly occur on fire-prone selectively logged forests
_a_nd_further increase forest degidation in the §r_azilian Amazon.

This hypothesis cannot be accepted as it is stated. My results showed that
selective logging significantly increase forest fire probability. However, the
juxtaposition of forest fire and selective logging maps showed only approximately
7% of common area between logging and fire. Additional statistical analysis
indicated that proximity to deforestation was a factor that substantively and
significantly affected forest fire likelihood. Therefore, fire-prone selectively logged
forests may increase forest fire probability, but it did not seem to be the major
factor that increases forest fire events at the regional scale. The second half of
this hypothesis can be accepted as stated. Indeed, the results showed that, when
taken as a singular phenomenon, forest fires severely degraded canopy cover
(up to ~19%) in the study region. At the local scale, field measurements also
indicated that newly burned forests significantly decreased around 12% of
canopy cover when compared to undisturbed forests. Tree mortality rate was
also significantly higher and diversity index was significantly lower compared to
undisturbed and selectively logged forests.

Hypothesis 3: yflables that reflect landscape and land use clfirgcten'stics (e. g,

vegetation Me, soil, precipitation, deforestation, forest fragmentation. proximity

161

to road network, etc.) explain spatialeistripution and Mability of selective
logging and forest fire to occur in the Lrazilian Amazon.
This hypothesis can be accepted. In this analysis, variables derived from
remotely sensed data representing landscape and land use characteristics were
used as input for forest fires and selective logging probit models. These
multivariate models permitted to select statistically significant variables that
represented factors influencing selective logging and forest fire occurrences in
the study area. The final statistical models supported inferences of both the
selective logging and forest fire probabilities throughout the study region. The
effect of each variable on forest fire and selective logging probabilities was
previously presented and discussed.
5.4. Opportunities for further studies

lVly study showed a 10 year trend of forest fire and selective logging in the
Amazon, but the long-term impacts of these activities are still unknown. I also
estimated the extent and impact of selective logging and forest fires on pristine
forests in the Brazilian Amazon. In fact, synergistic interactions between these
two land use types are poorly understood at regional scales. Therefore, new
approaches using finer temporal, spatial, and spectral resolutions satellite
imagery should be developed to improve the combined assessment of forest fires
and selective logging in tropical regions. Field studies also should be conducted
in different sub-regions in the Brazilian Amazon to better understand the process
of forest degradation and its synergisms in the Brazilian Amazon. Finally, the

development of simulations and scenarios of forest fires and selective logging

162

could be useful to support public policies to achieve a sustainable development

in the Amazon region.

163

APPENDICES

164

Appendix A.1. Total deforestation in the Brazilian Amazon from 1978 to 2004

165

 

800,000
700,000
600,000 ,
500,000 -
400,000
300,000
200,000 ,
100,000 "

Area (km2)

 

1978 1988 1998 2001 2002 2003 2004 2005
Year

 

 

I: Cumulative deforestation

 

 

Source: INPE (2007)

Appendix A.1

166

 

Appendix 3.1. Study site location (Landsat path 226 row 068) in the Amazon

State of Mato Grosso, Brazil.

167

—..m XECOQQ<

“N“
e 8e

222:2; 82. com

  

 

 

E
cow

9.82.6.2 8v com o
I .-
V §l\r//\ \QIJ rrrrrrr
\ I
.\ T
\\ ll lT. I.-- l :o~aE< .33 9:35 B
\ .., irrr
\l.. T. -
T. W
.\
\ .5 \.
A 32.5w .\
T T, T
N 8° .
T Boa Ff
.T. J
\ mum T
..IJrK/ .3
.T ..l|tll1lll|l|
/ . .........................
2. ...........

zonaE<
:53 2: 2556 :55 I

    

 

..=N<mm

 

0320 055. To 39w

 

168

Appendix 82. Satellite imagery used in the case study in the Amazon State of

Mato Grosso, Brazil. Landsat Path 226 and Row 068

169

 

Acquisition Landsat Band Source

 

 

Date Sensor
19-May-92 TM 1,2,3,4,5,7 INPE
26-Aug-93 TM 1,2,3,4,5,7 INPE
12-Jul-94 TM 1,2,3,4,5,7 INPE
29-Jun-95 TM 1,2,3,4,5,7 INPE
1-Jul-96 TM 1,2,3,4,5,7 INPE
5-Aug-97 TM 1,2,3,4,5,7 INPE
5-Jun-98 TM 1,2,3,4,5,7 INPE
19-Aug-99 ETM+ 1,2,3,4,5,7 TRFIC
18-Jun-00 ETM+ 1,2,3,4,5,7 TRFIC
8-Aug-01 ETM+ 1,2,3,4,5,7 TRFIC
4-Mar—02 ETM+ 1,2,3,4,5,7 TRFIC
26-Jul-02 ETM+ 1,2,3,4,5,7 TRFIC
6-Aug-03 TM 1,2,3,4,5,7 INPE
21-Jun-04 TM 1,2,3,4,5,7 INPE
Appendix 8.2

170

Appendix B.3. Slope and intercept for atmospheric correction of the Landsat

imagery using as 58 model.

171

Slope and intercept for atmospheric correction (58)

 

 

 

 

 

 

 

 

Year Band Slope intercept R2 Sig. F
1992 B1 1.2424 -0.0900 0.9993 0.000
32 1.2346 -0.0483 0.9996 0.000
BB 1.1749 -0.0288 0.9998 0.000
B4 1.1846 -0.0141 0.9999 0.000
BS 1.1650 -0.0026 1 0.000
B7 1.1349 -0.0010 1 0.000
1993 B1 1.2455 -0.0896 0.9993 0.000
32 1 .2376 -0.0480 0.9996 0.000
B3 1.1764 -0.0284 0.9998 0.000
B4 1.1853 -0.0138 0.9999 0.000
B5 1.1656 -0.0025 1 0.000
B7 1.1378 -0.0010 1 0.000
1994 B1 1.2648 -0.0944 0.9992 0.000
32 1 .2547 -0.0507 0.9996 0.000
B3 1.1893 -0.0302 0.9998 0.000
B4 1.1958 -0.0149 0.9999 0.000
B5 0.1715 -0.0027 1 0.000
B7 1.1403 -0.0012 1 0.000
1995 B1 1.2682 -0.0953 0.9992 0.000
32 1.2576 -0.0512 0.9996 0.000
B3 1.1915 -0.0306 0.9998 0.000
B4 1.1976 -0.0151 0.9999 0.000
B5 1.1726 -0.0028 1 0.000
B7 1.1410 0.0012 1 0.000
1996 B1 1.2728 -0.0965 0.9993 0.000
B2 1.2582 -0.0510 0.9996 0.000
B3 1.1915 -0.0303 0.9998 0.000
B4 1.1972 -0.0148 0.9999 0.000
B5 1.1723 -0.0027 1 0.000
B7 1.138 -0.0013 1 0.000
1997 B1 1.2837 -0.0990 0.9992 0.000
32 1.2712 -0.0533 0.9996 0.000
B3 1.2017 -0.0319 0.9998 0.000
B4 1.2055 -0.0158 0.9999 0.000
B5 1.1773 -0.0030 1 0.000
B7 1.1443 -0.0012 1 0.000
1998 B1 1.2774 -0.0976 0.9993 0.000
82 1 .2663 -0.0523 0.9996 0.000
B3 1.1972 -0.0311 0.9998 0.000
B4 1.2015 -0.0152 0.9999 0.000
BS 1.1751 -0.0028 1 0.000
B7 1.1447 -0.0012 1 0.000

 

Appendix 3.3. To be continued

172

Slope and intercept for atmospheric correction (5S)

 

 

 

 

 

 

 

Year Band Slope intercept R2 Sig. F
1999 B1 1.2837 -0.0099 0.9992 0.000
32 1.2491 -0.0502 0.9996 0.000
B3 1.1859 -0.0300 0.9998 0.000
B4 1.1937 -0.0149 0.9999 0.000
B5 1.1701 -0.0028 1 0.000
B7 1.1383 -0.0012 1 0.000
2000 B1 1.2705 -0.0959 0.9993 0.000
82 1.2604 -0.0513 0.9996 0.000
B3 1.1927 -0.0304 0.9998 0.000
B4 1.1978 -0.0149 0.9999 0.000
B5 1.1729 —0.0027 1 0.000
B7 0.8754 0.001 1 0.000
2001 B1 1.2492 -0.0906 0.9993 0.000
BZ 1.2414 -0.0483 0.9996 0.000
B3 1.1786 -0.0285 0.9998 0.000
B4 1.1867 -0.0138 0.9999 0.000
B5 1.1665 -0.0024 1 0.000
B7 1.1373 -0.0011 1 0.000
2002 B1 1.2418 -0.0025 0.9993 0.000
32 1.2343 -0.0137 0.9996 0.000
B3 1.1743 -0.0282 0.9998 0.000
B4 1.1839 -0.0476 0.9999 0.000
B5 1.1647 -0.0029 1 0.000
B7 1.1349 -0.0010 1 0.000
2003 B1 1.2728 -0.0965 0.9993 0.000
B2 1.2624 -0.0516 0.9996 0.000
B3 1.1940 -0.0306 0.9998 0.000
B4 1.1988 -0.0149 0.9999 0.000
BS 1.1734 -0.0027 1 0.000
B7 1.1382 -0.0009 1 0.000
2004 B1 1.2173 -0.0849 0.9992 0.000
BZ 1.2118 -0.0457 0.9996 0.000
B3 1.1588 -0.0273 0.9998 0.000
B4 1.1722 -0.0134 0.9999 0.000
B5 1.1578 -0.0026 1 0.000
B7 1.1381 -0.0012 1 0.000

 

Appendix 8.3. Continued.

173

Appendix B.4. Slope and intercept for normalization of the Landsat imagery
using as a reference the path 226 row 068, acquired in August 8,

2001.

174

Slope and intercept for Landsat imagery normalization

 

 

 

 

 

 

 

 

Year Band Slope intercept R2 Sig. F
1992 31 1.9458 -17.023 0.9877 0.000
32 1.7109 -10.506 0.9849 0.000
B3 2.2996 -15.615 0.9884 0.000
B4 1.0635 9.8466 0.9884 0.000
B5 1.6272 -14.439 0.9761 0.000
B7 1.5619 2.6144 0.9814 0.000
1993 31 1.1355 -18.099 0.9053 0.000
32 1.0387 -9.2681 0.9202 0.000
B3 1.655 -11.368 0.9608 0.000
B4 0.3689 41.513 0.9771 0.000
B5 1.1254 -5.123 0.9965 0.000
B7 1.1151 -2.8036 0.9994 0.000
1994 31 1.3403 -7.4632 0.9967 0.000
32 1 .2786 -2.8302 0.9982 0.000
B3 1 .3281 -3.6665 0.9988 0.000
B4 1 .0825 -0.5237 0.9942 0.000
B5 1.2995 -3.0316 0.9983 0.000
B7 1.3192 0.9798 0.9963 0.000
1995 31 1.7059 -1S.641 0.9622 0.000
32 1 .4951 -6.068 0.9703 0.000
B3 1 .7877 -8.0787 0.9934 0.000
B4 1.2372 1.1525 0.994 0.000
35 1.5292 -7.2734 0.9882 0.000
B7 1 .8949 -3.7354 0.9874 0.000
1996 31 1.5642 -12.177 0.9913 0.000
32 1 .4677 -5.3S67 0.9963 0.000
B3 1 .4836 -4.5231 0.9876 0.000
B4 1.1978 2.3117 0.9736 0.000
BS 1.3547 -2.1899 0.9756 0.000
B7 1.6588 -3.297 0.9928 0.000
1997 31 1.513 -15.815 0.9588 0.000
32 1.4038 -8.9126 0.9717 0.000
B3 1 .4471 -9.2322 0.9785 0.000
B4 1.1821 -5.186 0.9756 0.000
BS 1 .3085 -3.382 0.997 0.000
B7 1.2852 -0.4181 0.9959 0.000
1998 31 1.0959 -0.912 0.9833 0.000
32 1 .2526 -2.0085 0.9944 0.000
33 1.7138 -5.9986 0.9972 0.000
34 1.2153 1.1125 0.985 0.000
35 1 .2027 0.742 0.9945 0.000
37 1.7216 -3.1168 0.9762 0.000

 

Appendix 3.4. To be continued

175

Slope and intercept for Landsat imagery normalization

 

 

 

 

 

 

Year Band Slope intercept R2 Sig. F
1999 31 1.0725 -4.698 0.972 0.000
32 1.1402 -4.7609 0.98 0.000
33 1.094 -2.9142 0.9919 0.000
34 1.0274 -1.7792 0.9919 0.000
35 1.0021 -0.498 0.9994 0.000
37 1.0361 0.00832 0.9667 0.000
2000 31 1.136 -4.2724 0.9763 0.000
32 1.1361 -3.3419 0.9701 0.000
33 1.4985 -6.1871 0.9888 0.000
34 1.077 0.9588 0.9842 0.000
35 1.2987 -7.6197 0.9921 0.000
37 1 .4035 -2.0002 0.9626 0.000
2002 B1 1 .4999 -8.6342 0.9874 0.000
32 1.3816 -4.4249 0.9881 0.000
33 1.2784 -1.1385 0.9973 0.000
34 1.0682 -0.8031 0.9651 0.000
35 1.505 -15.125 0.9601 0.000
37 1.7514 -S.9245 0.9594 0.000
2003 31 0.9417 -1.3961 0.9688 0.000
32 0.8743 0.0857 0.9875 0.000
33 1.0109 -4.7539 0.9653 0.000
34 0.9747 0.3068 0.9838 0.000
35 1.0121 -0.3465 0.9948 0.000
37 0.8776 1 .6558 0.9995 0.000
2004 31 0.9762 1.6281 0.9522 0.000
32 1.0554 0.051 0.9637 0.000
33 1.2116 -2.6322 0.98 0.000
34 1.2849 1 .5652 0.9872 0.000
35 1.5952 -4.9404 0.9914 0.000
37 0.9849 2.0037 0.9986 0.000

 

Appendix 8.4. Continued.

176

Appendix B.5. Fieldwork locations within the study area

177

Fieldwork locations within the study area

 

 

Terra Nova

 

Paranatinga

 

 

Site Latitude
ID (degree)
MT01 -1 1 .5569
MT02 -1 1 .2263
MT03 -1 1 .7591
MT04 -1 1 .7438

Appendix 35

Longfiude
(degree)
-55.4017
-55.3985
-55. 1 86
-55.21 19

UTM X
(meters)
6742876
6748368
6976763
6948595

178

UTM Y Altitude
(meters) (meters)
8721959 358
8758522 287
8699443 346
8701 157 352

 

Appendix 36. Canopy openness estimation using gap light analyzer and

hemispherical photos from forest canopy cover

179

 

 

 

 

Appendix 3.6

180

Appendix 87 Visible and SWIR (at 2.1 pm) reflectance, and MSAVI and

MSAVlaf retrieved from Landsat image acquired in 2003.

181

 

 

 

1
/
Q.
3
B
G
R

0 4|

40km

20

10

 

 

Appendix 3.?

182

Appendix B.8. Burned forests detection based on non-photosynthetic vegetation

fraction image

183

 

    

RGB 5/4/3 Red dots = NPV threshold of 0.12 Burned forest map

 

 

 

E
o a 6 12km
Appendix 8.8

184

Appendix 39. Cumulative deforestation and persistence of selective logging and

forest fire on remotely sensed data

185

em 5.2.2.2

.8365 .o: 963 858 ..m :o 8826:

>__m:::m 295 6:. .860: L063 Lo NEv. m. cm >_o~mE_xoLaa< .uoLEQm mm; @5085 :5: 3:20 o>=m_:E:o 05 To eEv. «3 Lo xmmE < .902
Low: :o>_m 55 L8 :ozmummLoLo: o>=m_:E:o Lo _Soh,L Loo: :62: 65 L8 3898:: cc: Lang m:o_>oLa :_ 99096:

528 8:5: Lo .02. e .me; :62: ES :_ 8826: $90: 8:5: Lo _EokeL Low: :95 65 L2 3898:: En Loo: w:o_>oL: :_ c.2090:
393 8:5: :8 830. .6 .30.: o Lao.» 52¢ 65 :_ 62096: :88: 8.5:: ::m comma. :0 _the Loo: :96 65 L8 88900:: :5
Low: 9039: :_ 8896: 6.28 830. Lo .90... 6 :8: $36 65 :_ 3826: 7.92 :33. Lo .90... .umoLoL 39:86:: o>=m_:E:o .30.: o

 

186

 

 

momma 0mm: 03 Qmm : Nmom fimmog Lu Rad :Vmww woom
0. e39 348 Doom mwmoé when: v.89: mmmmm 4.80.9 meow
D: E @va vmmm 0mm: 30me Hm teem tweed Named? meow
0.336 v.8: mdmm mam: Memo; 0. e :3 Odom: 5mm e. 3 room
NEW: Qmmm mwmw N ewe Comm; 03“: N Ema 0m 5.? ooow
.4. EN: WEN Owov 0mm v.8: 0:3: :8: ode El 39
4.936 33m ow? 0mm 0N9 Noam o.wmw.m @309 mm?
mowed wam :98 mm: N8 0305 Named @085. .83
@086 0.8m who who How Qwood momma 343.2 89
Q89: 3w em «.03 ad: Own: mowed comm; 33.8 mm?
seem: comm mdm 0mm 2.: Wmooé 36.3.? No.3. 5 ”some
@086 w. Sum mmwe mm? 0.34 ohm: mwewé 33.3 mm?
chow: Nmmm - rm: - mow: mammvm no?
89:98:: 506960 03898:: 3:28:80 08633:: 3:28qu
Tees: Tees: Tog: 8:5: ES: Tare:
,_ 8.58580 2:0 $98 855 .w 830. 262698 2:0 832 262898 . Emma“. Lmo>
0:33:50 39:56::

 

398 89320

 

Appendix 8.10. Forest canopy losses estimated using multi-regression model for

the study area

187

Forest canopy losses estimated using multi-regression model for the study area

 

N = 49583 observations (3000m x 3000m systematic sampling grid)

R2 = 0.6307; Adj. R2 = 0.6305; Prob > F = 0.000; F(23, 49559) = 3526.17

 

 

Constant and Coefficient Std. Error t P (2- [95% Conf. Interval]
Variables tails)

Constant -235.0 12.96792 -18. 12 0.000 -260.46 -209.626
Deforestation -53.9 0.2091 1 -257.7 0.000 -54.294 -53.474
Newly logged -5.0 0.40512 -12.34 0.000 -5.7937 -4.2057
Old logged forest
detectable(*) -1 .6 0.3423 -4.62 0.000 -2.2521 -0.9103
Old logged forest
non-detectable(“) 3.1 0.2428 12.58 0.000 2.57897 3.53075
Newly logged &
burned forest -35.6 1.05498 -33.78 0.000 -37.709 -33.573
Newly burned ' -28.9 0.99566 -29.05 0.000 -30.877 -26.974
Old burned forest
detectable (*) -12.6 0.76258 -16.49 0.000 -14.072 -11.082
Old burned forest
non-detectable(**) -1 .0 0.71549 -1 .45 0.147 -2.4412 0.36352
Old logged &
burned forest
detectable(*) -25.6 1 .03242 -24.75 0.000 -27.58 -23.533
Old logged &
burned forest non-
detectab|e(**) -4.7 0.54397 -8.57 0.000 -5.7269 -3.5945
Water body -61.3 1.81431 -33.8 0.000 -64.871 -57.759
Year of 1993 -8.8 0.36602 -24.05 0.000 -9.5186 -8.0838
Year of 1994 -4.5 0.36661 -12.23 0.000 -5.2039 -3.7668
Year of 1995 -1.2 0.3669 -3.33 0.001 -1.9414 -0.5031
Year of 1996 -3.1 0.36738 -8.45 0.000 -3.8251 -2.385
Year of 1997 -5.6 0.36799 -15.11 0.000 -6.2802 -4.8377
Year of 1998 -3.5 0.36895 -9.41 0.000 -4.1957 -2.7494
Year of 1999 84 0.36992 -22.61 0.000 -9.0905 -7.6404
Year of 2000 -4.4 0.37366 -11.83 0.000 -5. 1537 -3.6889
Year of 2001 -7.1 0.37501 -18.91 0.000 -7.8252 -6.3551
Year of 2002 -6.5 0.37629 -17.31 0.000 -7.2502 -5.7751
Year of 2003 -5.0 0.37739 -13.23 0.000 -5.7316 -4.2522
Year of 2004 0.1 0.37891 0.25 0.806 -0.6497 0.83561
X coordinate (km) -0.001 0.0015 -0.43 0.669 -0.0036 0.0023
Y coordinate (km) 0.037 0.001497 24.7 0.000 0.03403 0.0399

 

* The term detectable refers to previously logged and/or burned forests that persist detectable on
satellite imagery. ** The term non-detectable refers to previously logged and/or burned forests
that do not persist detectable on satellite imagery.

Note: In this multiple-regression analysis, undisturbed forest and the year of 1992 were used as
omitted variables for land use and annual changes, respectively.

Appendix 8.10

188

Appendix C.1. Study region location and Landsat scenes used in the
spatiotemporal assessment of selective logging and forest fire in

the Legal Amazon

189

to 5.95:2

 

“Hun
EM :3 2:. a 2:.

one I.

 

 

 

>2... 3% .
32 :5 .39 .33 .
3:1 .553:an 3:3: nan—Esau

 

 

DNN QNNBN NnN N8 An—Haa

L \

 

33%: «F: - a E
20N<E< 4<0m._ 9ng

 

H
Ev. 92:. com a com

      

cones _aaoJ 2... 038:0 =NEn I
conga Ema.— uhnsm I

..N<¢m

 

 

:onE< _mmo._ 9: :_

9: $90.. ocm @580. 6266.8 :0 EoEwmmmmm _wLooE9ozmgw 9t :_ now: 3:08 55:3 _ocm cosmoo. :29: >35

190

Appendix C.2. Acquisition dates of the Landsat imagery for 1992, 1996, and

1999

191

Acquisition dates of the Landsat imagery for 1992, 1996, and 1999

 

 

 

Path/Row Acquisition Date
1992 1996 1999
221/063 July 11, 1992 May 30, 1997 May 14, 2000
222/062 July 24, 1991 July 5, 1996 July 14, 1999
222/063 July 24, 1991 June 3, 1996 August 23, 1999
222/064(1) August 25, 1991 June 19, 1996 August 23, 1999
223/062 August 16, 1991 May 28, 1997 July 13, 1999
223/063 May 28, 1991 July 12, 1996 July 13, 1999
223/064 August 10, 1992 June 10, 1996 July 13, 1999
223/065 June 02, 1993 June 10, 1996 July 13, 1999
223/066 July 25, 1992 June 10, 1996 July 13, 1999
223/0670) - - August 30, 1999
224/062(2) - - October 24, 1999
224/063 June 22, 1992 July 19, 1996 October 8, 1999
224/064(2) - - October 8, 1999
224/065 July 16, 1992 June 19, 1996 October 8, 1999
224/066 July 16, 1992 July 3, 1996 October 8,1999
224/067 July 16, 1992 July 3, 1996 October 8,1999
224/O68 July 23, 1992 July 3, 1996 August 21, 1999
224/069 July 23, 1992 July 3, 1996 August 21, 1999
225/0670) July 23, 1992 June 8, 1996 August 12, 1999
225/0680) July 23, 1992 August 11, 1996 August 12, 1999
225/069 June 21, 1992 August 11, 1996 August 12, 1999
225/0700) - - August 12, 1999
226/063 July 20, 1991 June 18, 1997 August 3, 1999
226/066(2) - - August 3, 1999
226/067 May 19, 1992 July 1, 1996 August 19, 1999
226/068 May 19, 1992 July 1, 1996 August 19, 1999
226/069 May 19, 1992 July 1, 1996 August 19, 1999
226/070(2) - - August 3, 1999
227/0620’) - - August 10, 1999
227/0650) August 06, 1992 August 09, 1996 August 10, 1999
227/067 August 6, 1992 July 24, 1996 August 10, 1999
227/068 August6,1992 June 6, 1996 August 10, 1999
227/069 July 5, 1992 June 6, 1996 August 10, 1999
227/070 July 5, 1992 June 6, 1996 August 10, 1999
228/067(2) - - August 17, 1999
228/068 June 18, 1992 June 13, 1996 August 17, 1999
228/069 June 18, 1992 June 13, 1996 August 17, 1999
229/067 my 3, 1992 July 22, 1996 August 8, 1999

 

1 Selectively logged forest not mapped in 1992 and 1996.
2 Selectively logged and burned forest not mapped in 1992 and 1996.

Appendix C.2. To be continued

192

Acquisition dates of the Landsat imagery for 1992, 1996, and 1999

 

 

 

Path/Row Acquisition Date
1992 1996 1999

229/068 July 11, 1992 July 22, 1996 October 11, 1999
229/069 June 25, 1992 July 6, 1996 August 8, 1999
229/070 July 11, 1992 June 23, 1997 August 8, 1999
230/068 July 10, 1992 October 17, 1996 August 15, 1999
230/069 May 15, 1992 October 17, 1996 July 30, 1999
232/067 June 22, 1992 June 25, 1996 June 28, 2000

002/067 August 30, 1992 August 1, 1996

Appendix C.2. Continued.

193

August 2, 1999

Appendix C.3. Case study (fieldwork) locations in the Brazilian Amazon

194

no 565:?

 

Ex coo oov

com o

 

moez
hoe:
00:51.

nee:
eoez'

o
NOPE

 

 

 

5:86.36 e
533233335 0

Dzwow4

 

 

        

6T. :25 :o 35
03.52

\

    
    
    

\Wtf

Ex 000..

23:30...

.\ H
: ..\

\ 2am

 

x,

\t t
/
3:552 /, .L ,_, \x.)
:auE LL \p

2583:
ZON<_2< gem... o.r=N<¢m

com 0

2:22.31

”NCONNE
t2}.

 

                       

ooow com

o

 

         
      

    

 
    

Sow: .82
our. El 855 21
room oEHHMW. co< :0 >20

       

 

 

:o~mE< :mENSm 05 :_ 30380. 0.83205 >25 3:0

195

Appendix 0.4. Case study and transect locations and characteristics

196

to 5.2.8.2

 

 

:26 988 8 8:2 858 88.8 8 x 8m M38- 832- oms 88:0 682 8:2
6.6 88 a 8.5.. 858 2.58: S x 8m 83.8.. 88.2- are 886 682 8:2
:8 88 a 8:2 883 S x com 838. 8an w; 885 292 8:2
:8 83 3 8:2 883 S x 8.8 8.8.8- 8.0?- Ee 6880 682 8:2
882 893283 2 x 08 83.8- «8.2- .88 88.6 082 8:2
882 883283 3 x 8.8 838- mmmd- was 885 282 8:2
6.6 88 2 883 S x 8m NE?- 857 E 8 88.6 682 8:2
S_o 88 8 8539883 8 x 8m 8:8- 8:. e T 2 8 88.6 682 85
S_o .883 3 858 88. 88 co 3 x 08 808- mmm. e e- S 3 885 682
6.0 88> i 858 238: 8 x 08 80.8.- 8m. e T : 8 885 682 8:2
:8 88 a 8:2 883 S x 8m 83.8- 80 e T 9 3 885 682
66 88 3 858 3:8: S x 08 88.8- 837 8 e 885 682 5:2
882 88328: S x com. 888- 8:9- 8 e. 98 80m
85: 88328: S x 08 $8.8- 82:- B 8 9888: 80m
6.6 88 a 883 S x 8m «88. 88.8- 8 e 9888.
S_o 88 2 883 S x 8.8. N88- 888- 8 8 888$ 80m
6.0 88 a 883 S x 8: :08- @89- 8 we 83
88 : 858.883 8 x 08 98.8- 039- 8 s 83 89a
882 8898 88:0 8 x 08 88.5. 83- 8 8 82
8.82 88328: S x 08 88.5- Ema- 8 s 89a 8%
E: :o_m:®E_o $983 $833 0. Q_
mozmtofiogmzo Sew Hoomcwb- 00:56:04 $03:le— uoomcm-c. 39w 31w

 

197

Appendix 05. Transect design for fieldwork measurements

198

no 5883

 

 

 

€3.23on £00: :2... ucm .Eo On A ucm .Eo 0.5-Om
.Eo mN v 80820 00:5 55.3 32:05 u:_:_m:..0.. :0 E300 "0:2 .03“.

300:9: E can 05
320 So or A .30 $2: ”EoEqumoE ran 05 5:85.50. 00....-

.:.0 905 m 5230: .8085 0:... E0 m v 3.08.20 ”momma...
.2089: EEoEu 02: 55.3 moo: 05 5:80:30: 36 .58 30.6030

:8 0.. A Imo 53> moo:

 

2:080:8«0E 0mg _0:u_

:88. .828. 52... ll

€825: 800:0:
.8 .88. so:

 

$58. 88.6

 

 

 

:0 EoEoSmmoE Imo 0:0 €00.05 :0 .850: .950 .0003 520:8 ozm0m4
ooh 60:35.52 00... 600:: :0 30:5: 05 :0 ano 3000:“: £02
mhzgmm—Dm4g Dam—n.
a . NO Em

 

. \.\I II E...
8 _ :5 w 5 Eu.

 

 

 

 

 

 

 

ES
_ \8 ‘. _ _ _ _
Em_ _Em_Em_ Em._sm
\\
\\
=58

 

 

 

wecmEoSmmoE 0:02.20: :0: :9000 88:0:-

199

Appendix C.6. List of tree species identified and recorded during the field work in

the States of Acre, Rondbnia, and Mato Grosso.

 

List of tree species identified and recorded during the field work in the States of
Acre, Rondonia, and Mato Grosso

 

 

Species Scientific Name
Abacaterana Aniba burchellii Kostern
Abiorana Pouten'a macrophylla (Lam.) Eyma
Abiu Pouten'a sagotiana (Baill) Eyma
Acai Euterpe oleraceae Mart.
Acapu Minquartia guianensis Aubl.
Algodao Bombax sp
Amarelao Qua/ea dinizii Ducke
Amendoim-Bravo Pterogyne nitens Tul.
AmesclalBreu Trattinickia burseraefolia Mart. Willd.
Angelim Hymenolobium sen'ceum Ducke
Angelim Pedra Dinizia excelsa Ducke
Angico Cassia fastuosa Willd.
Apui Ficus tn'gona L.f.
Babacu (*) Attalea sp
Bacuri Symphonia globulifera L.f.
Bagaceira Bagassa guianensis Aubl.
Banana-do-mato (*) Musa spp

Bandarra (*)
Barriguda
Bauchorana
Bolateira
Branquilho
Branquinho
Breu grande
Cafezinho
Caja
CambaralCedrinho
Canafistula
Canela
Capirana
Caraipe
Caroba
Castanheira
Catanudo

Catuaba / Breu Leite

Caucho

Caxeta / Mandioqueira

Cedro

Champagne/Cumaru

Chicha/ Capote
Cinzeiro

Schizolobium amazonicum (H ubber) Ducke
Chon'sia integrifolia Ulbr.

Ecclinusa guianensis Eyma
Manilkara huben' (Ducke) Chevalier
Albizia hasslen'i (Chodat) Burkart
Sebastiana sp

Protium apiculatum Swartz

Caesan'a sp

Spondias dulcis F orst.

En'sma uncinatum Warm.

Cassia Ieiandra Benth.

Nectandra cuspidata Nees
Calycophyllum spruceanum Benth
Licania pruinosa R. Benoist
Jacaranda copaia (Aubl.) D. Don
Bertholletia excelsa Humb. Bonpl
Chrysophyllum sp

Th yrsodium schomburgkianum Benth.
Castilla uIei Warb.

Qua/ea albiflora Warm.

Cedrela odorata L.

Dipten'x odorata (Aubl.) Willd.
Sterculia speciosa K. Sch um
Terminalia amazonica QF. Gmell Excell.

 

Appendix C.6. To be continued...

201

List of tree species identified in the field work in the States of Acre, Rondénia,

and Mato Grosso

 

 

Species Scientific Name
Cipouba Parkia discolor Spruce ex. Benth.
Copaiba Copaifera guianensis Desf.
Copiuba Goupia graba Ambl.

Cumaru Ferro Dipten‘x trifoliata Ducke
Cupuacu Theobroma bicolor Humb & Bonpl.
Envireira Rollinia sp

Escorrega Macaco
Escova de Macaco
Espeteiro
Farinha Seca
Favefia

Figueira

Freijo
Garapa/Garapeira
Garrote
Goiabinha (*)
Guanandi
Guaranta
Guariuba
lmbauba (*)
Imburana

Inga

lpe

Mauba

Jambo

Jatoba (Jatai)
Jenipapo

Jito

Joao Mole

Lacre

Landim

Leiteiro

Louro
Macaranduba
Macuco

Mamica de Porca
Mamui
Mandiocao
Marapaju
Marmelada
Matamata

Capirona huben'ana Ducke

Apeiba sp

Casean'a sylvestris Sw.

Polygonanthus amazonicus Ducke
Parkia nitida Miq.

Ficus malacocarpa Standl.

Cordia sagoti L. M. Johnston.

Apuleia Ieiocarpa (Vogel) J .F. Macbr.
Brosimum utile (H.B.K.) Pittier

Bellucia grossulan'oides (L.) Triana
Calophyllum brasiliense Cambess.
Esenbeckia Ieiocarpa Engl.

Clan'sia racemosa Ruiz & Pav.

Cecropia uIei Snethl.

Amburana acreana (Ducke) A.C. Sm.
Inga macrophylla H.B.K.

Tabebuia senatifolia (Vahl) Nichols.
Mezilaurus itauba (Meissn.) Taubert ex Mez.
Eugenia malaccensis L.

Hymenea courban'l L.

Genipa caruto H.B.K.

Quarea silvatica CDC.

Neea opositifolia Ruiz & Pav.

Vismia guianensis Pers.

Calophyllum brasiliense Cambess.
Himatanthus sucuuba (Spruce ex. Mull. Arg.) Woodson
Ocotea rubra Mez.

Manilkara excelsa (Ducke) Standl.
Hirtella racemosa Lam.

Zanthoxylum rhoidolium Lam

Jaracatia spinosa (Aubl.) A. DC.
Schefflera morototoni (Aubl.) Decne. & Planch.
Manilkara Iongifolia (A.DC.) Dubard
Amaioua guianensis Aubl.
Eschweileragrandifolia (AublLSandwith

 

Appendix C.6. To be continued...

202

 

 

Species Scientific Name
Mirindiba Terminalia amazonica (J.F. Gmel) Excell.
Monjolo Enterolobium maximum Ducke
Mororo Bauhinia macrostachya (Raddi) J.F. Macbr.
Muiracatiara Astronium gracile Engler
Mulateiro Pentaclethra macro/oba (Willd.) Kuntze
Mulungu Malouetia tamaquan'na (Aubl.) A.DC.
Murici Byrsonima chrysophylla H.B.K.

Naja Maximiliana man'pa (Aubl.) Drude

Nao identificada Unknown

Orelha de Negro Enterolobium contortisiliquum (Vel|.) Morong
Orucuri Siagus sp

Paineira Ceiba Pentandra (L) Gaertn.

Palheira (*) Geonoma sp

Palmeira (*) Livistoma chinensis (Jacq.) R.Br.

Palmeira Norjesu (*)
Parna
Pata de Vaca

Pau cebola Clusia rosea Jacq.

Pau d'alho Seguien'a Iangsdorffii Mog

Paxiuba (*) PALMAE

Pente de Macaco Bagassa guianensis Aubl.

Pequia Caryocar gracile Wittm.

Peroba Mico Aspidospenna pyn'folium Mart.
Pindaiba Xylobia frutescens Aubl.

Pororoca Curatella americana L.

Quina Coutarea hexandra (Jack.) K. Schum.
Roxinho Peltogyne paradoxa Ducke

Sangue lryanthera elliptica Ducke

Seringueira Hevea brasiliensis Mull. Arg

Sucupira Bowdichia virgilioides H.B.K.
Sumauma Ceiba Pentandra Gaertn.

Tachi Tachigalia paniculata Aublet
Tamarindo Man‘iodendron elatum (Ducke) Gleason
Tauari Couratan' guianensis Aubl.

Tucuma Astrocaryum aculeatum G.F. W. Meyer
Uchi Endopleura uchi (Huber) Cuatr.

Unha de gato (*) Mimosa sp

Urtiga (*) Cnidoscolus pubescens (Fax) Fax & k. Hoffm.
Velame Sclelorobium paniculata Vogel

Violeta Peltogyne paradoxa Ducke

Virola Virola sun'namensis (Ro|.) Warb.

Livistoma sp
Brosimum sp
Bauhinia sp

 

* Regeneration (secondary) tree species

Appendix C.6. Continued.

203

Appendix C]. Estimates of indicators of forest disturbances by selective logging
and forest fires based on field measurements in the States of

Acre, Rondonia, and Mato Grosso

204

...00::_::00 0: 0 ._. ho 50:00}:

 

 

 

0: 00.0 EN: mm mm: 0mm 800:0 0:02 0Q:
om 3.0 00.3 00 mm 000 00080 0:02 or:
om Rm 00.0: mm 3» 0mm 800:0 0:02 0::
mm mm: 00.0: 0.0 «0 0mm 000:0 0:02 .20.:
3 0.0.0 omen 0: mm 000 800:0 0:02 0::
0: m: 00.8 00 m: 30 080:0 0:02 mt:
00 00.: omNm 00 mm mm: 00080 0:02 .3 u:
om 8.0 00.: om ow 05 800:0 0:02 m: u:
m: Bum 00.0: 0: mm 00m 800:0 0:02 m: ::
0m 0: 000: mm mm mom 800:0 0:02 S u:
0: 80 5mm om 00 «mm 00080 802 o: 0
0:. 3.: 00.0: m0 mm 3.0 800:0 0:02 mo ::
m0 000 mmem 00 0: 00m 9:00:00. mo 0.:
00 8.0 9.0m mm mm Sm 0_:00:0m B 0:
0:. NWN 00.0w mm: mm NR 0_:00:0m mo 0::
00 E: omwm 00: ON 0.00 0_:00:0m mo ::
00: mod 3.0m 00m mm 030 0:90. 00 u:
m: 00.0 mm: 00: 00 00.0 0:0< mo 0:
mm 9.: 00m: 0: 0 0mm 0:04 mo:
mm 00.0 3.9: om: 0: 0:0 0:0< :00

0:: 0&9: 00:0 :0::m:: 0:: 0:: 0::

:0:0:0:000m _000: 0000 00:0 _000m 0.00% 0000 20303.3. :0:000_ 05 o_
2.50 E0 0: A: 000:: 800:0:

 

205

00:00:00 0.0 50:00:;

 

 

 

 

0.: 00.0: 0:: 000: 00.0 00.0 :0. 00.00 0.00 00::
0.0 00.00 00: 00:0 00.0 00.0 0.0 00.00 «.00 0:0
0.0 00.00 00: 0000 00.0 00.0 0.:: 00:: 0.0: 0::
0:.0 :.:0: 3.0 0000 00.0 00.0 0.0 00.0 0.0: :.:00
0.0: 5.00 000 0000 00.0 00.0 0.0 0:.:: 0.0 0::"
0.0 :.00 000 $0.00 00.0 00.0 0.0 00.0 0.0 0 :0
0.0 0.00 000 00:0 00.0 00.0 0.0: 0.: 0.0 w: u
0.0 0.00 000 0000 00.0 00.0 0.0: 0.00 0.00 0: 0
0.0 :.: :: E: 000: 3.0 :00 0.0 0.0: 0.0: 0: n
0.0 0.00 000 0000 00.0 00.0 0.0 0.0 0.0: :2:
0.0 0.00 000 0000 00.0 :00 0.00 0.0: 0.0 0::"
0.0 0.:0 0:0 0:0: 00.0 00.0 0.00 0.0 0.0: 00:"
0.0 0.00 E 000 00.0: 00.0 0.0: 0.: 0.0 00:w
0.0 0.00 00 000 00.0: 00.0 0.0: 0.0: 0.0 00::
0.0 000: 00: 000 00.0: 00.0 0.00 :.0: 0.:. 000
:.0 0.00 00 30: 00.0: 00.0 0.00 :.0 0.0 00::
0.0 0.0: 000 0000 00.0: 00.0 0.00 0.0 0.0 00::
0.0 :.:0 000 0000 00.0: 00.0 0.00 0.0: :.0: 00::
0.0 0.00 000 0000 00.00 00.0 0.0 0.0 0.: 0
00:

:.0 :00 000 0000 00.0: «0.0 0.0 0.: 0.0 :0:W
0:3 020:9 :00 :00 2:82: 0:084 .080 2.08::

5000 0:002: ::0 0 :-0 Eu 0 v «:00 : _Qm :0:0:0:000m 30: 0:0: 320:0:

:05: 002 _000 :0::0_0:0_>_0:: D_

000:020 000:0>< 5000:0000. 0005 E0 0: A: 006000 00:: 600:0;

 

206

Appendix C.8. Empirical linear relationship between various vegetation indices
and forest canopy coverage based on field measurements for the

study sites in Mato Grosso, Rondonia, and Acre.

207

Mato Grosso site, N= 47

 

 

 

 

 

 

 

 

 

208

Vegetation Index R2 df F(1,45) Sigf.
MSAVI 0.840 45 236.83 0.000
GEMI 0.802 45 182.21 0.000
NDVI 0.672 45 92.26 0.000
MSAVIaf 0.810 45 191.9 0.000
GEM|2,1 0.776 45 155.79 0.000
AFRI 0.603 45 68.25 0.000
GV 0.762 45 144.24 0.000
Rondonia study site, N= 32

Vegetation Index R2 df F(1,30) Sigf.
MSAVI 0.7374 30 84.26 0.000
GEMI 0.6991 30 69.7 0.000
NDVI 0.6018 30 45.35 0.000
MSAVlaf 0.6983 30 69.44 0.000
GEMI2.1 0.6972 30 69.07 0.000
AFRI 0.5772 30 40.96 0.000
GV 0.6897 30 66.67 0.000
Acre study site, N = 30

Vegtation Index R2 df F(1,28) Sigf.
MSAVI 0.8125 28 121.3 0.000
GEMI 0.7304 28 75.85 0.000
NDVI 0.7181 28 71.34 0.000
MSAVIaf 0.8035 28 114.49 0.000
GEM|2.1 0.705 28 66.9 0.000
AFRI 0.7147 28 70.14 0.000
GV 0.7416 28 80.34 0.000

Appendix C.8

Appendix C.9. Basin-wide results of selective logging and burned forest detection

for 1992, 1996, and 1999

209

000:0:00 0: o... 0.0 50:000.:

 

 

 

0.000 0. :: 0.00: 0.00 0.0 0.000 0.0: 0. : 00:0 009000
0.00 - 0.0 - - - - - - 6000.000
0.0 - 0.0 0.00 :.0 0.0 0.0 - 0.: 009000
- - - - - - - €080.00
0.00 0.0 0.000 0.0: 0. : 0.000 0.: 0.0 0.000 009000
0.00 0.0 0.00: - - 0.000 - - 0.00: 300900
0.0: - 0.000 - - 0.000 - - :.:00 6009000
- 0.3.. - - 0.000 - - 0.:: €082.00
m: - 0:00.: :.o - 0.000 - - 0.00: 9008500
00 :0 0.0 0. :0: 0.0: 0. : 0.00 :.00 0.0 00:0 0900
0.000 :00 00.00.: 0. :0: 0.00 0.00: 0.000 0.00 0.000 009.00
0.00 :.0: :.0: 0.00 0.0 0. K 0.00 - 0.00 009.000
0.0.0 0.0 0.00 :.0: :.0 0.3. 0.... - ...00 009.000
0. :00 0.0: 0.00: 0. :: 0.0 0.00 - - - 9009000
- 0.0 - - 00 - - - 5:00.000
.10 0.0 :00 0.0 0.0 :.00 0.00 0.: 0. :0 009000
0.00 - 0.00 0.0: 0.0 09.00 0.00 :.: 0. : :: 009000
0.000 0.0 0.00 09.0 0.0 0.00 .000 - 0.0: 009000
0.03.0 0000 0.000 00:0: :.:0: 0.000 0.000 0.00 :.00 009000
:.000.0 0.000 000.0. 0.000.: 0.000 0. : :0 0.000 :00 0.00: 009000
0.:: 0.0 0. :0 0.: :.0 0.00 - 0.0 0.:: 6009000
0. :00.0 0.000 0.00:. 0.000 0.00: 0.00:. 0.0.00 0.00 0.0.0 009000
0. :00 :.000 0.0.00 0.000 0.00 0.3.0 0.000 0.00 :.000 009000
:00 :.0: 0.00 0.: :.0: 0.:: 0.00 0.00 0.000 009:00
0.00 - ...00 0.00 - 0. :: 0.: - :.0 009000
.05.. 0:3 .05.. 6%.. 0:2. 0:2. .05.. We: .05..
2:0 0.: 00000.. 2:0 2:0 C 00000.. 2:0 2:0 00000.. 2:0 500
00000.. .0 00:50 00:50 00000.. .0 00:50 0055 00000.. .0 00:50. 00:50 .0
000: 000: 000: :.0n.

 

210

. 0005.50 .06 50:00.2

 

 

 

 

:.00:.00 0000.0 0000.0 0. :00.0 0.000.: 0.00 :.0 :.000.0 0. :00 0000.0 .0000
0.00: - 0.0 - - 0.0 0.0: - :.0 009000
0.00 0.0 0.00 0.00 0.0 0.0: :.0: 0. : 0.00 009000
0.00: 0.0 0.00: 0. :0 0.0 0.00 0.0: - 0.00 009000
0.00 0. : 0.00 0.00 - 0.0 0.00 0.0 0.00 009000
0.0: 0.0 0.00 0.0 0.0 0.00 0.0 - 0.0 009000
0. : :0 0.00 0.00: 0.00 0.0 0. :: 0.0 :.0 0.00 009000
0.000 0.0 0.0: 0.00 - 0.0: 0.0 - 0.0: 009000 1
0.000 0.0 0.00 0.00 0.0 0. :: 0.0: 0.0 :. :: 009000 «A
0.000 0. :0 0.00: 0.00: 0.00 0.00: 0.00: 0.0 :.00: 009000
:. :: - :.00 :.0: - 0.0 - - - $009000
0.0: :.0 0.00 0.0 - - 0.0 - - 009000
0.000 0. :0 0.00: 0.000 0.0: 0.00 :.000 0.0 0.00 009000
0.00:. : 0. :0: 0.000 0.000 0.00 0.000 0.000 0. 00 0.0 :0 009000
0.00: 0.0 :.00: 0.0: 0.0 0.00: 0.0 0.0 0.000 009000
0.0:: 0. : :.00 - - 0.0 - - 0.0 $009000
0.000 0.0 0.00 0.0: :.0 :.0 - - - 7.009000
0. : - 0. :0 - - - - - - $009000
0000.0 0.000 0.000 :.000.: 0.00 0.0 :0 0.000 0.00 0.00: 009000
0. :00.0 0.000 0.000 0.0:0.: 0.00 0. :0: 0.000.: 0.00 0.000 009000
0:2. 0%.. 05... 0E... 0%.. 050 :5... 0:2. 0:2.
2:0 A... 00000.. 2:0 2:0 C 00000.. 2:0 2:0 00000.. 2:0 .500
00000.. .0 00:50 00:50 00000.. 0 00:50 00:50 00000.. .0 00:50 00:50 .0
000: 000: 000: £00

 

 

Appendix C.10. Selectively logged and burned forests detected by State in the

Brazilian Amazon

212

2.0 50:32

 

 

 

 

:.00:_00 0000.0 0000.0 0.:00.0 0.000.: 0.00:.0 :.000.0 0.:00 0000.0 .060
0.0: - 0.00 0.:: 0.0 0.00: 0.0 0.0 :.00 0:0:0000
- - - - - - - - - 08.050
0.000 0.0 0.00: 0.00: 0.: 0.00 0.00 0.0 0.:0 9:00:00
0000.0 0.0:0.: 0000.0 0.00:.0 0.000.: :.000.0 0.00:.0 0.000 0.000.: 0:0n.
0.000.:: 0.000 0.:00.0 00:00 0.000 0000.0 0000.0 :.00: 0.:00.0 00080 0.0.2
0000.0 0.000 0.000 0.000 0.0:: 0.000 0.000 0.00 0.000 005:0:02
- - - - - - - - - 00:80:00
- - - - - - - - - 000:2.
0.00 - 0.00 0.00 - 0.:: 0.0 0.0 0.0 0:00.
.05.. 005.. .05.. .05. .05.. .05.. .05.. .05.. .05..
2:0 F. 00000.. 2:0 2:0 A“. 00000.. 2:0 2:0 00000.. 2:0 0.0.0
00000.. .0 00:50 00:50 00000.. .0 00:50 00:50 00000.. .0 00:50 00:50
000: 000: 000:

 

213

Appendix C.11. Forest canopy losses estimated using multi-regression model for

the study region

214

 

 

Variables and Coef. Std. Err. t P (2 tails) [95% Conf.
Constant Interval]

Constant 95.5 0.2461 387.83 0.000 95.0 95.9
Deforestation -49.0 0.1724 -284.30 0.000 -49.4 -48.7
Secondary Regrowth -6.8 0.2297 -29.74 0.000 -7.3 —6.4
Cerrado (Savannah) -39.2 0.1573 -249.37 0.000 -39.5 -38.9
Water body -66.8 0.5647 -118.36 0.000 -67.9 -65.7
Newly logged (1) -1.6 0.4895 -3.28 0.001 -2.6 -0.6
Newly burned (1) -18.8 0.6250 -30.12 0.000 -20.1 -17.6
Newly logged &
burned ( ) -27.5 1.3681 -20.07 0.000 -30.1 —24.8
Previously logged (2) 5.8 1.4327 4.05 0.000 3.0 8.6
Previously burned (2) -2.5 1.1203 -2.25 0.024 -4.7 -O.3
Previously logged &
burned (2) -9.7 1.4907 -6.48 0.000 -12.6 -6.7
Year 1996 1.0 0.1381 7.19 0.000 0.7 1.3
Year 1999 -5.1 0.1324 -38.29 0.000 -5.3 -4.8
Longitude (km) —0.008 0.0001 -59.48 0.000 -0.008 -0.007
Latitude (km) 0.012 0.0002 58.34 0.000 0.011 0.012

 

‘ The term newly refers to logging and/or burning activities occurred during 1993-96
and 1997-99 intervals, detected by 1996 and 1999, respectively. 2 The term previously
refers to logging and/or burning activities detected by 1992 and 1996 that persisted
detectable on Landsat imagery by 1996 and 1999, respectively.

 

N = 154635 observations (5000m x 5000m systematic sampling grid)

R2 = 0.5285; Adj. R2 = 0.5284; Prob > F = 0.000; F(14, 154620) = 12377.16

 

Appendix C.11

215

Appendix C.12. Spatial distribution and intensity (logged area I625 km2) of

selectively logged forests detected in 1992

216

Spatial distribution and intensity (logged area / 625 km2) of selectively logged

forests detected in 1992

  
  

 

 

 

 

 

   

 

 

 

 

 

701w sogw
1'10 1 , «MLI
IU I 1 IV I
0° 0.
10°S -10°S
Logged area (1992)
‘ 0%
l :
l :
‘ = 35%
l l =
1 ' = 70%
l
20°s ‘ * 20°s
l l i
70"W 60°W 50°s

Appendix C.12

217

Appendix C.13. Spatial distribution and intensity (logged area I625 km2) of

selectively logged forests detected in 1996

218

Spatial distribution and intensity (logged area / 625 km2) of selectively logged

forests detected in 1996

 

70gw 60W
1

 

 

 

 

    

 

 

 

 

10°s- ‘ , -10°s
1 Rondonia” '
1| I I. '
l flab m Logged arm (1996)
l j _ 0%
‘ C
l ’ =
l “ g 35%
. l =
\ —=— i : 70%
0 300 600 900 Km
20°s 1 1 20‘s
7o'°w 605w 503w
Appendix C.13

219

Appendix C.14. Spatial distribution and intensity (logged area / 625 km2) of

selectively logged forests detected in 1999

220

Spatial distribution and intensity (logged area / 625 km2) of selectively logged

forests detected in 1999

 

 

 

   
 

 

707w 60’W
10°11 1 10°N
N 1
l
W<e>>E ]
S ,
Area of Study

 

 

0° ' ' ff . ' [I .. r f ...‘1 — L“

 

Logged area (1999)
%

w
u:
a?

IIIIIIIIII
O

N
o
33

WMC
Lu 0

 

 

 

l
20°s l
1

 

 

Appendix C.14

Appendix C.15. Spatial distribution and intensity (burned area / 625 km2) of

burned forests detected in 1992

222

Spatial distribution and intensity (burned area / 625 km2) of burned forests

detected in 1992

 

 

 

    

Burned forest area
(1 1

I
o
a?

or
a!

‘3

as
(l)

 

 

anon
£U O

 

 

 

Appendix C.15

223

Appendix C.16. Spatial distribution and intensity (burned area I 625 km2) of

burned forests detected in 1996

224

Spatial distribution and intensity (burned areal 625 km2) of burned forests

detected in 1996

701W

50M

 

10° I

1

 

    

i
.

law-mg“ 10%.-

f
.1; 1]

Area of Study

 

 

20°s-

 

 

 

    

 

Burned foreSt area
(1996)

o
33

 

 

Appendix C. 16

 

I
soow

Appendix C.17. Spatial distribution and intensity (burned area I625 km2) of

burned forests detected in 1999

226

Spatial distribution and intensity (burned area / 625 km2) of burned forests

detected in 1999

 

 

10°N * 777777 ,, 77 7 a «LV v» e7 M 10°N

 

 

 

 

 

or
a?

8
at

 

 

 

 

Appendix C. 17

227

Appendix C.18. Fractional coverage derived from Landsat imagery acquired in

1992 for the study region

228

Fractional coverage derived from Landsat imagery acquired in 1992 for the study

region

 

      

5‘5: Tocantins

 

 

Mato Grosso
:.:—Z:

0 250 500 750 1,000 Km

 

 

 

Appendix C.18

229

 

Appendix C.19. Fractional coverage derived from Landsat imagery acquired in

1996 for the study region

230

Fractional coverage derived from Landsat imagery acquired in 1996 for the study

 

I’Jf

region

A.10apa

   

1R“\ (“J E/\
/

, 11181201338

 
 
 
    

   

 

FC%
E10
E
=
II50
\ -
inlamGrosso/ =
I— -
o 250 500 750 1,000 Km Ax -100

‘x / \vslji

__/

 

 

 

Appendix C.19

231

Appendix C.20. Fractional coverage derived from Landsat imagery acquired in

1999 for the study region

232

Fractional coverage derived from Landsat imagery acquired in 1999 for the study
region

 

Jv/

Roraima

    

  

Amazonas

TI
0
o\°

IIIIIIIIIflfi
01 o
O

_I
O
0

III—:—
0 250 500 750

 

 

Appendix C20

233

 

Appendix D1 The study region location within the Brazilian Amazon

234

The study region location within the Brazilian Amazon

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

7o:w 60:W 50°w
5° N
0°
5°s
10°s-
15°s-

70°w 60°W 50°w
Appendix D.1.

235

Appendix D.2. Deforestation map for the Brazilian Amazon in 1992, 1996, and
1999

236

Deforestation map for the Brazilian Amazon in 1992, 1996, and 1999

 

Deforestation map - 1992

 

 

 

 

 

 

 

 

 

 

Deforestation map - 1999

 

 

 

 

 

Appendix D.2

237

Appendix 03. Vegetation map of the Brazilian Amazon in 1999.

238

Vegetation map of the Brazilian Amazon in 1999.

70°W 60" W 50" W

forest

.-:-
0 300 600 9001km

70° W

Appendix D.3

239

 

10°N

Appendix D.4. Map of road network, protected areas, and timber centers in the

Amazon

240

Map of road network, protected areas, and timber centers in the Amazon

70° W 60°W 500 W
P 1

l
l
l

 

  

.,_¥_—.V _50 N

 

 

 

   

 

 

 

 

 

 

 

1 5° 3 I! I.151: S
l I A Tlmber centers
1 - Protected areas
1 ll 0 250 500 750m 1 S Road network
. =—=l ‘ l
20081 #7; A VT in? if T —" " A A. v fi ,_ —.._ A. 2 fihzoos
r 1
70° W 60° W 500 w
Appendix 0.4

241

Appendix 05. Map of timber volume generated using ordinary kriging
interpolation and the dataset derived from forest inventory

conducted by RADAMBRASIL project

242

Map of timber volume generated using ordinary kriging interpolation and the

dataset derived from forest inventory conducted by RADAMBRASIL project

 

 

5°8-1

10°3-

   

 

. Timber center

 

 

 

 

1
15° 3'” 7 A ”‘17' v A ’ 1‘ Tlmber Volume h15°S
l < 50 m3
' I - 50-100 m3
1‘ 0 250 500 750 km ‘1 - 1m_150 m3
2008 _777 *‘l __ fifi *7 ‘ ik A. _{_ ‘_ - 150.2m "121.20,,
E 1 r - > 200 m S
I
70°W 60°W 50'W
Appendix D.5.

243

Appendix D.6. Water deficit (mm) for the Brazilian Amazon during the dry

season (May to September) in 1999.

244

Water deficit (mm) for the Brazilian Amazon during the dry season (May to

September) in 1999.

70°W 60°W 50°w
10° N - .

 

10°S-W —*'10°S

     
   

Legend
—Paved roads
Water deficit

'0 300 600 9001 km
20°8-
-60.1-80 mm

 

 

 

701w 601w 50°w
Appendix 0.6

245

6. References

ABlMCl (2002). Mechanically processed timber - Sectorial study 2001, Brazilian
Association for Mechanically Processed Timber. Available on lntemet site
hmazl/wwwsyntekocombr. Web site last accessed on July 2004.

Adams, J. B., D. E. Sabol, V. Kapos, R. Almeida Filho, D. A. Roberts, M. O.
Smith and A. R. Gillespie (1995). "Classification of multispectral images
based on fractions of endmembers: Application to land-cover change in
the Brazilian Amazon." Remote Sensing (finvironment 52(2): 137-154.

Alvarado, E. and D. V. Sandberg (2001). Logging in Tropical Forest: Literature
review on Ecological Impacts, University of Washington and USDA Forest
Service.

Anselin, L. (2002). "Under the hood : Issues in the specification and
interpretation of spatial regression models." Agricultural Economics 27(3):
247-267.

Arima, E., Simmons, C., Walker, R., and Cochrane, M. 2007. Fire in the Brazilian
Amazon: A spatially explicit model for policy impact analysis. “Journal of
[Legiongficience’i 47(3). 541-567.

Asner, G., M. Keller and J. N. M. Silva (2004). "Spatial and temporal dynamics of
forest canopy gaps following selective logging in the eastern Amazon."
Global Chang Biology 10: 765-783.

Asner, G. P., M. Keller, J. Pereira, Rodrigo and J. C. Zweede (2002). "Remote
sensing of selective logging in Amazonia: Assessing limitations based on
detailed field observations, Landsat ETM+, and textural analysis." Remote
Sensing of Environment 80(3): 483-496.

Asner, G. P., D. E. Knapp, E. N. Broadbent, P. J. C. Oliveira, M. Keller and J. N.
Silva (2005). "Selective Logging in the Brazilian Amazon." Science
310(5747): 480-482.

Bateson, A. and B. Curtiss (1996). "A method for manual endmember selection
and spectral unmixing." Remote Sensing oflinvironment 55(3): 229-243.

Boardman, J. W., F. A. Kruse and R. 0. Green (1995). Mapping target signatures
via partial unmixing of AVIRlS data. Boulder, CO, University of Colorado,
Geological Sciences Department: 26p.

Burrough, P. A., McDonnell, RA. (1998). Principles of geographical information
systems. Oxford, Oxford University Press.

Chander, G. and B. Markham (2003). "Revised Landsat-5 TM radiometric
calibration procedures and postcalibration dynamic ranges." Geoscience
§n_d Remote Sensing, lEEE Transactions on 41(11): 2674-2677.

Chomitz, K. and D. Gray (1996). "Roads, land use, and deforestation: a spatial
model applied to Belize." The World Bank Economic Review 10(3): 487-
512.

Cochrane, M. A. (2001). "Synergistic Interactions between Habitat Fragmentation
and Fire in Evergreen Tropical Forests." _Consegtion Biology 15(6):
1515-1521 .

246

Cochrane, M. A., A. Alencar, M. D. Schulze, C. M. Souza, Jr., D. C. Nepstad, P.
Lefebvre and E. A. Davidson (1999). "Positive Feedbacks in the Fire
Dynamic of Closed Canopy Tropical Forests." Science 284(5421): 1832-
1835.

Cochrane, M. A. and M. D. Schulze (1998). "Forest Fires in the Brazilian
Amazon." ConservatioLBiology 12(5): 948-950.

Cochrane, M. A., D. Skole, E. A. T. Matricardi, C. Barber and W. Chomentowski
(2004). Selective Logging, Forest Fragmentation, and Fire Disturbance. l_n
Working Forests in the Tropics: Conservation through Sustainable
Management? D. J. ZARIN. Columbia, USA, Columbia University Press.
Chapter 17: 310-324.

Cochrane, M. A. and C. Sousa Jr (1998). "Linear mixture model classification of
burned forests in the Eastern Amazon." International Joumil of Ramota
Sensing 19(17): 3433-3440.

De Moura and Galvao (2003). "Smoke effects on NDVI determination of
savannah vegetation types." lnterrfiional JOlflgj of Ramote Sensing
24(21): 4225-4231 .

Elvidge, C. D., D. Yuan, R. D. Weerackoon and R. S. Lunneta (1995). "Relative
radiometric normalization of Landsat Multispectral Scanner (MSS) data
using an automatic scattergram-controlled regression." PE&RS -
Photogrammetric Engineering & Remote Sensing 61(10): 1255-1260.

Fearnside, P. M. (1997). "Greenhouse gases from deforestation in Brazilian
Amazonia: Net commited emissions." Slimatic Changa 35: 321-360.

Fearnside, P. M. (1999). "Biodiversity as an environmental service in Brazil's
Amazonian forests: risks, value, and conservation." Environmental
ConserLation 26(4): 305-321.

Fearnside, P. M. (2002). "Avanca Brasil: Environmental and social consequences
of Brazil's planned infrastructure in Amazonia." Environmental
Management 30(6): 735-747.

Fearnside, P. M., A. T. Tardin and L. G. M. FlLHO (1990). Deforestation in the
Brazilian Amazonia. Manaus, AM, Brazil, Amazon Research Institute and
National Institute for Space Research.

FRA (2005). The Global Forest Resources Assessment, 2005 - Main Report.
Rome, Italy, FAO - Food and Agriculture Organization of the United
Nations: 320.

Fredericksen, T. S. and W. Pariona (2002). "Effect of skidder disturbance on
commercial tree regeneration in logging gaps in a Bolivian tropical forest."
Forest Eiology ang Management 171(3): 223-230.

Frumhoff, P. C. (1995). "Conserving Wildlife in Tropical Forests Managed for
Timber." BioScience 45(7): 456-464.

Gentry, A. H. (1988). "Tree Species Richness of Upper Amazonian Forests."
PNAS 85(1): 156-159.

Grainger, A. (1993). "Rates of deforestation in humid tropics: estimates and
measurements." The Geographical Journal 159: 33-44.

247

 

Grainger, A. (1999). "Constraints on modeling the deforestation and degradation
of tropical open woodlands." Glotfil Ecology and Biggeograghy 8: 179-
190.

Holdsworth, A. R. and C. Uhl (1997). "Fire in Amazonian selectively logged rain
forest and the potential for fire reduction." Ecological Aplications 7: 713-
725.

Houghton, R. A. (1997). "Terrestrial Carbon storage: global lessons for
Amazonian research." Ciéncia e Cultura 49: 58-72.

Houghton, R. A., D. L. Skole, C. A. Nobre, J. L. Hackler, K. T. Lawrence and W.
H. Chomentowski (2000). "Annual fluxes of carbon from deforestation and
regrowth in the Brazilian Amazon." Nature 403(6767): 301-304.

Huete, A. R., T. Miura and X. Gao (2003). "Land cover conversion and
degradation analyses through coupled soil-plant biophysical parameters
derived from hyperspectral EO-1 Hyperion." Geoscience and Remote
Sensing, IEEE Transactions on 41(6): 1268-1276.

Huth, A. and T. Ditzer (2001). "Long-term impacts of logging in a tropical rain
forest - a simulation study." Forest Ecology and Management 142(1-3):
33-51.

IBAMA (2007). PROARCO - Programa de Prevencao e Controle as Queimadas
e Incéndios Florestais no Arco do Desmatamento, The Brazilian Institute
of Environment and Renewable Natural Resources. Available on Internet
site http://www.ibama.gov.br/proarco/ web page last accessed on
February 2007.

IBGE (2006). Extracao vegetal, National Institute for Geography and Statistic.
Available on Internet site http://www.sidra.ibge.gov.br.Web page last
accessed on December 2006.

INPE (2007). Monitoramento de queimadas, National Institute for Space
Research. Available on the internet site
httg/lwwwcptec.inpe.br/queimadas/. Web site last accessed on January
2007.

INPE (2007). Monitoring of the Brazilian Amazon by satellite, National Institute
for Space Research. Available on lntemet site hmzllwwwirmebr/prodesl.
Web site last accessed on January 2007.

Johns, J. S., P. Barreto and C. Uhl (1996). "Logging damage during planned and
unplanned logging operations in the eastern Amazon." Forest Ecology and
Management 89(1-3): 59-77.

Karnieli, A., Y. J. Kaufman, L. Remer and A. Wald (2001). "AFRI -- aerosol free
vegetation index." Remote Sensing of Environment 77(1): 10-21.

Kuntz, S. and F. Siegert (1999). "Monitoring of deforestation and land use in
Indonesia with multitemporal ERS data." International Journal of Remote
Sensing(14): 2835-2853.

Lambin, E. F. (1999). "Monitoring forest degradation in tropical regions by remote
sensing: some methodological issues." Acta Amazonica 8: 191-198.

Lambin, E. F. and D. Ehrlich (1997). "The identification of tropical deforestation
fronts at broad spatial scales." lntemational Jotm of Lemote S_ensing
18(17): 3551-3568.

 

248

Laurance, W. F., M. A. Cochrane, S. Bergen, P. M. Fearnside, P. Delamonica, C.
Barber, S. D'Angelo and T. Fernandes (2001). "The Future of the Brazilian
Amazon." Science 291(5503): 438-439.

Laurance, W. F., S. G. laurance, L. V. Ferreira, J. Rankin-de Merona, C. Gascon
and T. E. Lovejoy; (1997). "Biomass collapse in the Amazonian forest
fragments." Science 278: 11 17-1118.

Laurance, W. F. and G. B. Williamson (2001). "Positive Feedbacks among Forest
Fragmentation, Drought, and Climate Change in the Amazon."
Conservation Biology 15(6): 1529-1535.

Lentini, M., A. Verissimo and L. Sobral (2003). Fatos florestais da Amazonia.
Belem do Para, Brazil, Instituto do Homem e Meio Ambiente da
Amazonia.

Lentini, M., Verissimo, A., Pereira, D. (2005). A Expansao Madeireira na
Amazonia. O Estado da Amazonia. Belém, PA, Brazil. Instituto do Homem
e Meio Ambiente da Amazonia: 4p.

Maas, S. J. (2000). "Linear Mixture Modeling Approach for Estimating Cotton
Canopy Ground Cover using Satellite Multispectral Imagery." Remote
Sensing of Environment 72(3): 304-308.

Macedo, D. S. and A. B. Anderson (1993). "Early Ecological Changes Associated
with Logging in an Amazon Floodplain." Biotropica 25(2): 151-163.

Martini, A. M., Z., N. A. Rosa and C. Uhl (1994). "An attempt to predict which
Amazonian tree species may be threatened by logging activities."
mironmentwnservation 21: 152-161.

Mas, J.-F. (1999). "Monitoring land-cover changes: a comparison of change
detection techniques." International Journal of Remote Sensing 20(1):

1 39-152.

Matricardi, E., D. Skole, M. Cochrane, M. Pedlowski and W. H. Chomentowski
(2006). "Multi-temporal assessment of selective logging in the Brazilian
Amazon using Landsat data." International Jogrnal of Remote Sensing.

Matricardi, E. A. T. (2003). Multi-temporal assessment of selective logging using
remotely sensed data in the Brazilian Amazon. Degartment of Geograghy.
East Lansing, MI, Michigan State University: 151p.

Matricardi, E. A. T., D. L. Skole, M. A. Cochrane, M. Pedlowski and W.
Chomentowski (2007). "Multi-temporal assessment of selective logging in
the Brazilian Amazon using Landsat data." International Journal of
Remote Sensing 28(1): 63-82.

Matricardi, E. A. T., D. L. Skole, M. A. Cochrane, J. Qi and W. Chomentowski
(2005). "Monitoring Selective Logging in Tropical Evergreen Forests Using
Landsat: Multitemporal Regional Analyses in Mato Grosso, Brazil." Earth
Interactions 9(24): 1-24.

Matthews, E. (2001). Understanding the F RA 2000. Washington, DC., The World
Resource Institute. Available on internet site http:/www.wri.cyq. Last
accessed in December 2006.

Moura, M. L. and L. S. Galvao (2003). "Smoke effects on NDVI determination of
savannah vegetation types." International Jogmal of Remote Sgising
24(21): 4225-4231 .

249

NASA (1998). The Tropical Deforestation. Greenbelt, Maryland, USA, National
Aeronautics and Space Administration: FS-1998-11-120-GSFC.

Nascimento, H. E. M. and W. F. Laurance (2002). "Total aboveground biomass in
central Amazonian rainforests: a landscape-scale study." Forest Ecology
m Management 168(1-3): 311-321.

Nelson, B. (2001). Epgo em florestas da Amazonia Central em 1997. X Simpésio
de Sensoriamento Remoto, Foz do Iguacu, PR, Brazil.

Nepstad, D., D. McGrath, A. Alencar, A. C. Barros, G. Carvalho, M. Santilli and
M. C. V. Diaz (2002). "Frontier governance in Amazonia." Science's
Compass. 295: 629-631.

Nepstad, D., A. Moreira and A. Alencar (1999). Flames in the rain forest: origins,
impacts, and alternatives to Amazonian fire. Brasilia, DF, Brazil. Programa
Piloto para Conservacao das Florestas Tropicais.

Nepstad, D., A. Moreira, A. Verissimo, P. Lefebvre, P. Schlesinger, C. Potter, C.
Nobre, A. Setzer, T. Krug, A. C. Barros, A. Alencar and J. R. Pereira
(1998). "Forest Fire Prediction and Prevention in the Brazilian Amazon."
_Consarvation Biology 12(5): 951-953.

Nepstad, D. C., A. Verssimo, A. Alencar, C. Nobre, E. Lima, P. Lefebvre, P.
Schlesinger, C. Potter, P. Moutinho, E. Mendoza, M. Cochrane and V.
Brooks (1999). "Large-scale impoverishment of Amazonian forests by
logging and fire." 398(6727): 505-508.

Pedlowski, M., E. Matricardi, D. Skole, S. R. Cameron, W. H. Chomentowski, C.
Fernandes and A. Lisboa (2005). "Conservation units: a new deforestation
frontier in the Amazonian state of Rondonia, Brazil." Environmental
Conservation 32(2): 149-155.

Peet, R. K. (1974). "The Measurement of Species Diversity." Annual Review of
Ecology and Svatematics 5: 285-307.

Pinard, M. A. and F. E. Putz (1996). "Retaining Forest Biomass by Reducing
Logging Damage." Biotropica 28(3): 278-295.

Pinedo-Vasquez, M., D. J. Zarin, K. Coffey, C. Padoch and F. Rabelo (2001).
"Post-boom logging in Amazonia." Human Ecolggy. 29: 219-239.

Pinty, B. and M. M. Verstraete (1992). "GEMI - A non-linear index to monitor
global vegetation from satellites." Vegetation 101: 15-20.

Qi, J. (2001). Interpretation of spectral vegetation indices and their relationship
with biophysical variables. Michigan State Universiy. East Lansing, MI,
USA. non-published manuscripted.

Qi, J., F. Cabot, M. S. Moran and G. Dedieu (1995). "Biophysical parameter
estimations using multidirectional spectral measurements." Remote
Sensing of Environment 54(1): 71-83.

Qi, J., A. Chehbouni, A. R. Huete, Y. H. Kerr and S. Sorooshian (1994). "A
modified soil adjusted vegetation index." _R_amota_Sensing of Environmaflt
48(2): 119-126.

Qi, J., R. C. Marsett, M. S. Moran, D. C. Goodrich, P. Heilman, Y. H. Kerr, G.
Dedieu, A. Chehbouni and X. X. Zhang (2000). "Spatial and temporal
dynamics of vegetation in the San Pedro River basin area." Agricultural
aag Forest Meteorolgqy 105(1-3): 55-68.

250

Qi, J., C. Wang, E. Matricardi and D. Skole (2002). Improveg_selective loggi_ng
dLetection with Laijsat images in tQpical regions. Geoscience and
Remote Sensing Symposium, 2002. IGARSS '02. 2002 IEEE International.

RADAMBRASIL (1980). Levantamento dos recursos naturais. Mining and Energy
Ministry. Folha SC.21. Juruena. Volume 20. Rio de Janeiro, RJ.

Rencz, A. N. (1999). Remote Sensing of the Earth Sciences. Manual of Remote
Sensing. American Soiegr for Photogrammetry a& Remote Sensing:
251-306.

Schmink, M. and C. Wood (1992). From rivers to road. In: Contesteg frontiers in
Amazonia, Columbia University Press: 387p.

Schroeder, W., J. T. Morisette, I. Csiszar, L. Giglio, D. Morton and C. O. Justice
(2005). "Characterizing Vegetation Fire Dynamics in Brazil through
Multisatellite Data: Common Trends and Practical Issues." Earth
Interactions: 1-26.

Siegert, F., G. Ruecker, A. Hinrichs and A. Hoffmann (2001). "Increased damage
from fires in logged forests during droughts caused by El Nino."
Nature(414): 437-440.

SIVAM (2004). Projeto do Sistema de \flgilancia da Amazonia. Povoamento das
bases de dados. Sistema de Vigilancia da Amazonia, Brasilia, DF, Brazil.

Skole, D., M. A. Cochrane, E. Matricardi, W. H. Chomentowski, M. Pedlowski
and D. Kimble, Eds. (2004). Eagern tgprowss in the Ama_zon Region:
measuring forest conversion. regeneration and degradation. In the Land
Change Science: Observing, Monitoring, and Understanding Trajectories
of Change on the Earth's Surface, Kluwer Academic Publishers.

Skole, D. and C. Tucker (1993). "Tropical Deforestation and Habitat
Fragmentation in the Amazon: Satellite Data from 1978 to 1988." Science
260(5116): 1905-1910.

Sousa Jr, C. and P. Barreto (2000). "An alternative approach for detecting and
monitoring selectively logged forests in the Amazon." International Journal
of Remote Sensing 21(1): 173-179.

Souza, C. and P. Barreto (2000). "An alternative approach for detecting and
monitoring selectively logged forests in the Amazon." International Journal
of Remote Sensing 21(1): 173-179.

Souza, C. M., D. A. Roberts and A. L. Monteiro (2005). "Multitemporal Analysis of
Degraded Forests in the Southern Brazilian Amazon." Earth Interactions:
1-25.

Souza, J., Carlos, L. Firestone, L. M. Silva and D. Roberts (2003). "Mapping
forest degradation in the Eastern Amazon from SPOT 4 through spectral
mixture models." Remote Sensing of Environment 87(4): 494-506.

Stone, T. A. and P. Lefebvre (1998). "Using multi-temporal satellite data to
evaluate selective logging in Para, Brazil." lntemational Journal of Remote
Sensing 19(13): 2517-2526.

Tucker, C. J. (1979). "Red and photographic infrared linear combinations for
monitoring vegetation." Remote Sansing ofEnvironment 8(2): 127-150.

251

 

Turner, M. G., R. H. Gardner and R. V. O'Neill (2001). Landscape Ecology in
Theom and Practice: Patterns and Process. New York, Springer-Verlag
New York, Inc. '

Uhl, C., P. Barreto, A. Verissimo, E. Vidal, P. Amaral, A. C. Barros, C. Souza, Jr.,
J. Johns and J. Gerwing (1997). "Natural Resource Management in the
Brazilian Amazon." BioScience 47(3): 160-168.

Uhl, C. and R. Buschbacher (1985). "A Disturbing Synergism Between Cattle
Ranch Burning Practices and Selective Tree Harvesting in the Eastern
Amazon." 17: 265-268.

Uhl, C. and J. B. Kauffman (1990). "Deforestation, Fire Susceptibility, and
Potential Tree Responses to Fire in the Eastern Amazon." Ecology 71(2):
437-449.

Uhl, C., A. Verissimo, M. M. Mattos, Z. Brandino and I. C. Guimaraes Vieira
(1991). "Social, economic, and ecological consequences of selective
logging in an Amazon frontier: the case of Tailandia." Forest Ecology and
Management 46(3-4): 243-273.

Uhl, C. and I. C. G. Vieira (1989). "Ecological impacts of selective logging in the
Brazilian Amazon: A case study from the Paragominas region of the state
of Para." Biotropica 21: 98-106.

Verissimo, A., P. Barreto, M. Mattos, R. Tarifa and C. Uhl (1992). "Logging
impacts and prospects for sustainable forest management in an old
Amazonian frontier: The case of Paragominas." Forest Ecology and
Management 55(1-4): 169-199.

Verissimo, A., P. Barreto, R. Tarifa and C. Uhl (1995). "Extraction of a high-value
natural resource in Amazonia: the case of mahogany." Forest Ecology and
Management 72(1): 39-60.

Vennote, E., D. Tanre, J. L. Deuze, M. Herman and J. J. Morcrette (1997).
Second simulation of the satellite signal in the solar spectrum - 6S. Uir
Guide Version 2.

Wang, C., J. Qi and M. Cochrane (2005). "Assessment of Tropical Forest
Degradation with Canopy Fractional Cover from Landsat ETM{plus} and
IKONOS Imagery." firth Interactions 9(22): 1-18.

Wang, C., J. Qi, M. S. Moran and R. C. Marsett (2004). "Soil moisture estimation
in a semiarid rangeland using ERS-2 and TM imagery." Remote Sensing
p_f_Environment 90: 178-189.

Whitmore, T. C. (1984). Tropical rain forests of the Far East. 2d ed. Clarendon,
Oxford.

Wiegand, C. L., A. J. Richardson, D. E. Escobar and A. H. Gerberrnann (1991).
"Vegetation indices in crop assessments." _Ramotefinsinggf
mironment 35(2-3): 105-1 19.

Willmott, C. J. a. M., K. (1999). Terrestrial water balance data archive: regridded
monthly climatologies, Cort. J. Willmott and Kenji Matsuura with support
from NASA's Seasonal to Interannual ESIP.AvaiIable on lntemet site
http://climate.qeog.udel.ed u/~climate/ page last accessed on February
2007.

252

 

Wooldridge, J. M. (2000). Introductory Econometrics: A modern approach,
South-Western College Publishing.

WRI (2000). "The World Resources: The Human Population." Unitad Nations
_Develgpment Prgqramme. Available on internet site http://www.wri.org/.
Last accessed in December 2006.

253

 

   

I"IIIIQIIIIIZIIIIIIl