d

.5' ‘1)
mg

_ a}. “than

V

. a}.
”Elma”; .n u
‘ Larry,

.J.

as?

“.1...

(3.4.: Vin. A .
gummy: quflinwy$
‘ ”we. .<
..u.)...\..i.\m¢z
a .. :3 2w nr J 1
x 1.»; I: :rZfiu.
.834. . .J u »v .51, Hutu”, .
. «‘1: 3s.» .. O. I .
am: 12.43:)! 5
35.2.1

.21 9.!
qt . v.

\ 1.0.134.

‘ J 3’;

:II. .
mm." .

«L9: 53...;
309.3

:huvm.e., Mr." .4

c n00l9.1tv' Iii!
. n . ‘ A n .1 O

fiafifiééwfifi, ‘ . __ . I . gammwfifi."

 

 

THES.
2
W5}
5 ’l .9 5} ll *7 r7

This is to certify that the
thesis entitled

MAPPING AND MODELING TROPICAL DEFOLIATION:
A CASE STUDY ON THE EFFECTS OF PLAN COLOMBIA

presented by

Paul Larry Delamater

has been accepted towards fulfillment
of the requirements for the

MA. degree in Geography

 

 

/ file/jor Professor’s Signature
// M/J y &cj/

 

Date

MSU is an Affirmative Action/Equal Opportunity Institution

 

 

LIBRARY

Michigan State

University

 

 

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

 

J"L I 6 2007

 

@1240?

 

 

 

 

 

 

 

 

 

 

 

 

6/01 c:/CIRC/DateDue.p65-p.15

 

MAPPING AND MODELING TROPICAL DEFOLIATION:
A CASE STUDY ON THE EFFECTS OF PLAN COLOMBIA

By

Paul Larry Delamater

A THESIS

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

MASTER OF ARTS
Department of Geography
2004

ABSTRACT

MAPPING AND MODELING TROPICAL DEFOLIATION:
A CASE STUDY ON THE EFFECTS OF PLAN COLOMBIA

By

Paul Larry Delamater

This research explores the effects of aerial fumigation as part of the
Drug War on Putumayo, Colombia. Green fractional coverage (fc) is
examined in a time series of Landsat ETM+ images to quantify and
describe the damage resulting from fumigation. No studies have
attempted to quantify these effects.

I used remote sensing technology and fieldwork to build a direct
parameterization of surface phenomena and a classification of landuse
and landcover (LULC) types. This was accomplished by using fc as a
biophysical variable that characterized the effects of defoliation and a
hybrid classification method of LULC. I validating the use of fc by
collecting ground truth data to show the linear relationship that exists
between image fc and ground fc. I proved that Plan Colombia spraying
during 2002 was not discriminate over the landscape by relating LULC

classes to changes in fc over space and time.

ACKNOWLEDGEMENTS

First, I would like to thank my advisor, Dr. Joseph Messina, and my
guidance committee, Dr. Jiaguo Qi and Dr. Mark Cochrane, for their expertise
and their valuable time donated towards this research. Without their excellent
help and direction, I would not have been able to conduct this study. I would like
to acknowledge that NASA funded this research under grant # NAG 5-16217.

I thank the University of North Carolina at Chapel Hill’s Carolina Population
Center for supplying me with very much needed data and knowledge.

I would also like to thank my fellow graduate students, professors, and staff
in Michigan State University’s Depaitment of Geography and the Center for
Global Change and Earth Environments for helping me to stay on the right path
throughout my study. i thank my fellow graduate student Stephen Aldrich for
helping me unlock the mysteries of word processing software and keeping me
abreast of many deadlines.

Last, but certainly not least, I would like to thank my family and friends. To
all of you... I love you and I thank you for the strength and confidence that you

have bestowed in me throughout my life.

iii

TABLE OF CONTENTS

LIST OF FIGURES ............................................................................................... vi
LIST OF EQUATIONS ......................................................................................... ix
LIST OF ABBREVIATIONS ................................................................................... x
Chapter 1 : INTRODUCTION ................................................................................ 1

PROBLEM STATEMENT & RESEARCH OBJECTIVES ............................ 9

RESEARCH HYPOTHESES .................................................................... 10
Chapter 2 : LITERATURE REVIEW .................................................................... 1 1

THE UNITED STATES AND THE DRUG WAR IN SOUTH AMERICA....11

COCA AND COLOMBIA ........................................................................... 12
PLAN COLOMBIA .................................................................................... 17
MONITORING VEGETATION USING REMOTE SENSING .................... 22
Chapter 3 : METHODS ....................................................................................... 25
IMAGE PRE-PROCESSING .................................................................... 25
SAMPLE POINTS .................................................................................... 34
FIELD EQUIPMENT AND MAPS ............................................................. 38
FIELD DATA COLLECTION ..................................................................... 39
FIELD DATA PROCESSING .................................................................... 47
fc VALIDATION ........................................................................................ 50
LULC CLASSIFICATION .......................................................................... 53
CHANGE DETECTION ............................................................................ 55
Chapter 4 : RESULTS AND DISCUSSION ......................................................... 57
fc VALIDATION RESULTS ...................................................................... 57

iv

fc AND LULC CHANGE DETECTION ...................................................... 63

CONCLUSIONS ....................................................................................... 68

FUTURE RESEARCH .............................................................................. 69
LITERATURE CITED .......................................................................................... 72
APPENDIX A ....................................................................................................... 79
APPENDIX B ....................................................................................................... 81
APPENDIX C ...................................................................................................... 82
APPENDIX D ...................................................................................................... 84
APPENDIX E ....................................................................................................... 86

LIST OF FIGURES

Figure 1.1 Study area maps ................................................................................. 2
Figure 1.2 Photos showing Oxisols and tropical rainforest vegetation ................. 3
Figure 1.3 Colombia Departments ....................................................................... 6

Figure 1.4 Coca, Ecuador; Lago Agrio, Ecuador; and coca growing region in

Putumayo, Colombia ...................................................................................... 8
Figure 2.1 UNODC estimates of coca cultivation in Colombia ........................... 16
Figure 2.2 UNODC estimates of coca fumigation in Colombia ........................... 20
Figure 2.3 Guamez Valley Map .......................................................................... 21
Figure 3.1 Landsat ETM+ Scenes (Path 9, Row 60) .......................................... 26
Figure 3.2 Green Fractional Coverage ............................................................... 31
Figure 3.3 Sampling area and sample points ..................................................... 37
Figure 3.4 Example of a field map sheet ............................................................ 40
Figure 3.5 Collection of homogenous (LULC or fc) polygon information ............ 42

Figure 3.6 Examples of possible non-homogenous vegetation covers at sample

point locations .............................................................................................. 44
Figure 3.7 Examples of canopy photos taken with the fisheye lens ................... 45
Figure 3.8 Field data collection form .................................................................. 47

Figure 3.9 Sample point buffer where (A) the sample point falls in the center of
four pixels and (B) falls near the center of one pixel .................................... 52
Figure 4.1 Regression results using (A) the point to point overlay and (B) the

point to mean of points within 21 .5m overlay ............................................... 58

vi

Figure 4.2 Regression results minus outliers in the data using (A) the point to
point overlay and (B) the point to mean of points within 21 .5m overlay ....... 61

Figure 4.3 Maps for defoliation Event 1 and Event 2 (notice extent box for Figure
4.4) ............................................................................................................... 65

Figure 4.4 Close up of defoliation near the Ecuador-Colombia border in Event 2

(notice where identified in Figure 4.3) .......................................................... 66

Images in this thesis are presented in color

vii

LIST OF TABLES

Table 3.1 SS atmospheric correction input parameters ...................................... 28
Table 3.2 NDVI values used to calculate fc for Landsat images ........................ 32
Table 3.3 Geometric rectification errors ............................................................. 33
Table 3.4 Criteria for a sample pixel ................................................................... 36
Table 3.5 fc classes for stratified random points ................................................ 36
Table 3.6 GLA threshold values for fisheye lens canopy photos ........................ 48
Table 3.7 LULC classification scheme ............................................................... 55
Table 4.1 Regression results for point to point overlay with the full dataset ....... 59

Table 4.2 Regression results for point to mean of points overlay with the full
dataset ......................................................................................................... 59
Table 4.3 Removed sample points ..................................................................... 60
Table 4.4 Regression results for point to point overlay without the outliers in the
dataset ......................................................................................................... 62
Table 4.5 Regression results for point to mean of points overlay without the
outliers in the dataset ................................................................................... 62

Table 4.6 Defoliation classes and LULC classes represented in each ............... 67

viii

LIST OF EQUATIONS

Equation 3.1 Conversion of DNs to Radiance values ......................................... 26
Equation 3.2 Conversion of Radiance to TOA reflectance ................................. 27
Equation 3.3 58 atmospheric correction ............................................................ 29
Equation 3.4 NDVI calculation ........................................................................... 30
Equation 3.5 fc definition ................................................................................... 30
Equation 3.6 fc calculation ................................................................................. 32

ix

LIST OF ABBREVIATIONS

_5§_ : Simulation of the Satellite Signal in the Solar Spectrum
5g : Area Of Interest (in ERDAS Imagine)

AUC : Autodefensas Unidas de Colombia (United Self-Defense Groups of
Colombia)

LE: United States Drug Enforcement Agency

EM : Digital Number or Digital Pixel Value

wt; : Direccion Nacional de Estupefacientes (National Narcotic Direction)
QQ_S_ : Dark Object Subtraction (atmospheric correction)

§_L_N_ : Ejercito de Liberacion Nacional (National Liberation Army)

[m : Enhanced Thematic Mapper +

FARC : Fuerzas Armadas Revolucionarias de Colombia (Revolutionary Armed
Forces of Colombia)

Lg : Green Fractional Coverage

_Gfi : Geographic Information System
GLA : Gap Light Analyzer

§_E’§ : Geographic Positioning System

HDF : Hierarchical Data Format

DL : Interactive Data Language

__l: Leaf Area Index
MSAVI : Modified Soil Adjusted Vegetation Index
NDVI : Normalized Difference Vegetation Index

ONDCP : Office of National Drug Control Policy

13

IF : Psuedo-lnvariant Feature (atmospheric correction)

:1

__|: Research Systems Inc.

§_A_V_I : Soil Adjusted Vegetation Index

T_O_A : Top of Atmosphere

M933 United Nations Drug Control Program
M: United Nations Office on Drugs and Crime
LJ_S : United States of America

yiDS: United States Department of State

USGAO : United States General Accounting Office

xi

Chapter 1 : INTRODUCTION

This research explores the effects of aerial fumigation as part of the Drug
War on Putumayo, Colombia. More specifically, green fractional coverage (fc)
change is calculated, validated, and examined in a time series of Landsat ETM+
images to explore the effects of aerial fumigation on a landscape. The location of
this study is Colombia, in the Putumayo Department, where fumigation of coca
fields has been and continues to occur as a part of Plan Colombia, a US funded
program to combat drug production in Colombia. Fumigation is one method used
by the Colombian government to eradicate the cultivation of coca in this region.
The coca plant is a perennial shrub of the genus Erythroxylum and grows very
well in poor, acidic soils that cannot support many other commercially cultivated
crops (Peterson 2002). A coca plant can produce harvestable leaves within one
year of being planted, can produce harvestable leaves up to 4 times per year,
and can remain productive for up to twenty five years (Clawson and Lee 1996;
Gardner 2001 ). The leaves of the coca plant are ground to a pulp that is one of
the primary ingredients in the illegal narcotic, cocaine.

The Putumayo Department is located in the southwestern Colombia and
shares borders with the Sucumbios canton of Ecuador and the Loreto
department of Peru. The intensive study area is roughly 40 km x 60 km located

in southwestern Putumayo, bordering Sucumbios (see study area box in Figure

1.1). Putumayo is just east of the Andes Mountains and is a part of the greater

Amazon River Basin.

 

   

 

 

 

Figure 1.1 Study area maps

Located very near the Equator, Putumayo’s geographic region has a
Kbppen-Geiger-Pohl Climate Classification of Af or Tropical rainforest climate.
The average temperature of every month is above 18°C and annual precipitation
generally is above 500cm and exceeds annual evaporation. This water surplus
and the warm soil temperatures produce soils rich in iron oxides with a deep red
color (see Figure 1.2); these soils belong to the Oxisols soil class (Strahler and
Strahler 1992). The landscape can be characterized as a mix of agricultural
(mostly subsistence) fields and lowland (Amazonian) tropical forest (see Figure
1.2). The province has a population of roughly 330,000 people who are mostly

small-scale farmers and colonized the area through in-migration (Dudley 2000).

 

 

 

 

 

 

Figure 1.2 Photos showing Oxisols and tropical rainforest vegetation

Putumayo and the proximal regions in Colombia are considered only
marginally controled by the Colombian government. The left-wing guerilla group,
the Fuerzas Armadas Revolucionan'as de Colombia (FARC) (Revolutionary
Armed Forces of Colombia) and various other right-wing paramilitary groups

such as the Autodefensas Unidas de Colombia (AUC) (United Self-Defense

Groups of Colombia) have a significant local presence, while the government has
very little control. Due to its remote location and marginal governmental control,
over the past 10-15 years, Putumayo (and especially southwestern Putumayo)
has become one of the world’s largest coca producing areas. Various reports
estimate that Colombia produces the leaves for 80% of the world’s cocaine
(Bibes 2001) and Putumayo produces about 50% of the leaves in Colombia
resulting in an astonishing 40% of the world’s coca leaf production originating in
Putumayo (Penhaul 2001). Because of these enormous levels of coca
cultivation, Putumayo has become an area that is heavily targeted by the coca
eradication efforts of Plan Colombia, the attempt by the Colombian and US
governments to reduce the supply of coca for cocaine production. Putumayo has
been referred to as “the ground zero” of Plan Colombia (LaFranchi 2001; lsacson
and Vaicius 2001).

According to the United Nations Office on Drugs and Crime, over 400,000
ha of coca have been fumigated in Colombia since 1994, but, astonishingly, the
amount of area in coca cultivation throughout the country has grown from 3,871
ha in 1994 to 102,071 ha in 2002 (however, the area in cultivation has dropped
from a maximum of 163,289 ha in 2000) (UNODC 2003a). Many members of the
media have spoken out against the fumigation efforts of Plan Colombia over
these statistics, over reports of indiscriminate spraying of the landscape, over the
amount of military equipment being supplied to Colombia, and over other social,
political, and environmental issues (Alvarez 2003; Berger 2001; Brown 2000;

Driver 2001; lsacson and Vaicius 2001; Kratz 2002; Penhaul 2001; Peterson

2002; Vaicius and lsacson 2003; Wilson 2003). The US. Department of State
(2002) published Colombia's Environmental Auditor's parameters detailing
conditions that must be met for spraying to be conducted, including thresholds for
wind speed, temperature, and relative humidity. The USDS acknowledges that
although they try to minimize mistakes in fumigation due to human and
mechanical errors, occasional errors are unavoidable (USDS 2002).

Early fumigation efforts focused on large, “industrial” type coca fields,
often ignoring smaller fields. However, between July and October of 2002, the
push into Putumayo and Caqueta (a neighboring department, see Figure 1.3)
commenced in which no differentiation between large and small coca fields was
made and any field identified with coca became a target (Vaicius and lsacson
2003). Although Putumayo (and southwestern Colombia) was the focus of much
of the fumigation in 2002 (over 50% of the total area fumigated in Colombia
during 2002), the trend, more recently, has shifted away from these departments
and to the departments of Guaviare and Nariflo (see Figure 1.3) (UNODC
2003b).

Life for the people of southwestern Putumayo is not a life of luxury. Most
of the farmers are very poor and are accustomed to violence associated with the
presence of the FARC and other paramilitary groups and the results of cultivating
an illegal crop. One local farmer remarked to lsacson and Viacius (2001) that,
“life in Putumayo is not worth 1500 pesos ($0.75).” Unfortunately for the people

in the Putumayo, there are few other alternatives for cash crops (Cooper 2001).

 

   

 

 

 

Figure 1.3 Colombia Departments

Because of the unrest of the people living in Putumayo and the presence
of the FARC, conducting research in this region is far too dangerous for
outsiders. This danger has also spread south from Colombia into northern
Ecuador. Whereas, field teams from the University of North Carolina at Chapel
Hill had previously conducted research in and around Lago Agn'o, Ecuador
(Messina and Walsh 2001), the base for field studies has been moved more
south to Coca, Ecuador (see Figure 1.4). Lago Agrio has a population of around
24,000 people and the region was covered by primary forest 40 years ago. The
population of Lago Agrio has risen greatly since the discovery of oil in 1967 by
Texaco/Gulf and the agrarian reform laws that were passed in 1964 and 1973
(Perreault 2003). Although visiting Lago Agrio is safe, conducting research in the
countryside is considered dangerous, as many believe that FARC members

cross the border for rest and recovery in the Lago Agrio area. Famam (2002)

reported that from January to June of 2002, assassins of the FARC or the other
Colombian paramilitary groups working in Ecuador have killed more than 100
people in Lago Agrio. The situation in Lago Agrio had not gotten any safer in
2003 (when field work was scheduled for this research) and Ecuadorian
colleagues advised the field team to go elsewhere. Conducting research in
Coca, Ecuador allowed for a safe observation of a landscape with similar
biophysical characteristics and similar climatic conditions while still being located
in the same Landsat ETM+ scene as southwestern Putumayo (see Figure 1.4).
As mentioned previously, this region has a tropical rainforest climate. Although
the distance between these two areas is roughly 100 km, there are no significant
elevation changes between the coca growing areas of Putumayo and the areas
surrounding Coca, Ecuador. Also, no geomorphical features exist between the
areas that would significantly differentiate their respective biophysical or climatic

conditions.

 

OMA--L,.,
5 l

(ff/ COLOMBIA

Putumayo
00>—

EECUADOR/M we 9/12/02 Landsat ETM+ Scene
/

 

 

(Band combination 4,3,2)

 

 

 

Coca growing
region

2 25km

Lago Agrio

2 100km

 

 

 

 

Figure 1.4 Coca, Ecuador; Lago Agrio, Ecuador; and coca growing region in
Putumayo, Colombia

PROBLEM STATEMENT & RESEARCH OBJECTIVES

Descriptions of the visual damage from the spraying have been reported
at great length, however no follow-up studies have been attempted that quantify
the defoliation and subsequent regeneration of the coca fields, agricultural fields,
or other vegetation types affected by fumigation in Putumayo. The UNODC has
reported the amount of area in coca cultivation and the amount eradicated, but
the methods used are questionable, particularly as the results do not explain
exactly what is considered eradicated. There also have been reports of
indiscriminate spraying on the landscape. These, however, have only been
anecdotal reports from farmers and unsubstantiated stories by reporters. They
also stand in conflict with the strict guidelines laid out by the Colombian
Govemment’s Environmental Auditor. Because of the paramilitary presence
(FARC and AUC) and the unrest of the general population in Putumayo, such
assessments of the effects of Plan Colombia cannot be safely completed in the
field. Alternative methods of data collection and quantification must be employed.

The overall objective of this study is to quantitatively and qualitatively
measure the defoliation and regeneration of the vegetation due to the Plan
Colombia fumigation in Putumayo, Colombia. More specifically, the objectives of
this research are to:

1. Use remote sensing techniques to detect the biophysical attribute changes
that have occurred as a result of fumigation.
2. Verify and calibrate the detected biophysical attribute changes by

conducting field measurements in Ecuador.

3. Quantify the amount of defoliation of vegetation using the field-calibrated

4.

data.
Identify the primary land cover classes affected by fumigation and attempt

to identify whether the spraying is discriminate.

RESEARCH HYPOTHESES

| hypothesize that:

1.

Green fractional coverage derived from ETM+ data accurately quantifies
vegetation cover.

Hybrid classification techniques applied to ETM+ data identify LULC
classes.

Changes in is indicate changes due to defoliation.

Defoliation in Putumayo, Colombia occurs indiscriminately over the
landscape affecting cover types other than coca.

Landsat ETM+ data provides the synoptic coverage with adequate

resolution characteristics suitable for study.

10

Chapter 2 : LITERATURE REVIEW

THE UNITED STATES AND THE DRUG WAR IN SOUTH AMERICA

Reports of US government involvement with South American countries in
battling illegal drug related activities go back about 30 years. In 1973, the US
government began a four year, $6,000,000 training program for 600 law
enforcement officials to fight marijuana cultivation in Colombia (Sharpe 1988).
However, the US interdiction efforts in South American illegal drug trafficking
escalated greatly with President Reagan’s “war on drugs” of the early 1980’s
(Bagley 1988a). Many of the efforts resulting from this war have targeted the
supply end of the drug chain. In the Anti-Drug Act passed by Congress and
signed by President Reagan in 1986, approximately 75% of the $3.9 billion to be
spent in 1987 were dedicated to supply and programs such as expanded
enforcement, interdiction, and eradication/substitution versus the 25% dedicated
to demand end programs such as education, prevention, treatment, and
rehabilitation (Bagley 1988b). Over the last 20 years, narcotics control has
become a foreign policy priority as the US has carried out massive drug
interdiction operations in Colombia, Peru, and Bolivia (Vellinga 2000). The US
military has also been very active in fighting the production and trafficking of
drugs during this time period as spending increased from $0 in 1981 to $389

million in 1987. The Department of Defense (DOD) also loaned $303.5 million

11

dollars worth of equipment to various authorities enforcing drug laws from 1981
to 1987 (Mabry 1988).

Some of the early interdiction efforts centered around attempts to target
production at the source. During the early 1980's in Peru, local police carried out
manual eradication of coca fields with logistical support provided by the US Drug
Enforcement Agency (DEA), yet these programs showed only minor localized
reductions while total cultivation in Peru rose. The program expanded to include
spraying coca fields with herbicides in 1989, but demonstrations and a wave of
guerrilla attacks prompted a return to the previously used manual procedures of
eradication (Kay 1999).

No South American nation has received more US assistance in the drug
war than Colombia. From the years 1986 to 1996 the US State Department

estimated that over a half a billion dollars were spent in Colombia (Millett 1997).

COCA AND COLOMBIA

Colombia’s recent history has been littered with violent and bloody clashes
between people and the government and also between right-wing and left-wing
political groups. These clashes have often left many people dead such as the
period of La Violencia (from the late 1940’s to early 1960’s) which was
characterized by extreme violence, cruelty, and wanton killing, and claimed an
estimated 2-3% of the country’s total population (Guzman et al. 1962 and
Kalmanovitz 1988 as cited in Thoumi 1992). Colombia has also seen the rise to

power of left-wing guerilla groups, such as the FARC, the oldest guerrilla force in

12

the Americas (Browitt 2001), and the Ejercito de Liberacion Nacional (ELN)
(National Liberation Army), the rise of right-wing paramilitary groups, such as the
AUC, who attempt to combat the left-wing groups, and a very powerful
emergence (and a later disappearance) by major drug-trafficking groups such as
the Medellin and Cali cartels, all during the last 50 years. The FARC has
garnered so much political power in Colombia that the Colombian government
ceded a large portion of land in southern Colombia (roughly 40% of the total area
of Colombia) to the group in 2002 (Nagle 2002). The Putumayo department is a
stronghold of the FARC guerillas (Penhaul 2001) and is an underdeveloped area
of very marginal governmental control with a long history of illegal activity
(Marcella 2002).

Colombia also has a dubious and complex history of illegal narcotic
production. The de-Iegitimization of its government system over the past 45
years has allowed the country to gain an “advantage” in terms of illegal drug
production (T houmi 1992). The Colombian government continues to struggle for
legitimacy as it faces challenges such as questioning of the current peace
process, corruption, weak institutions, and inability to provide basic public goods
(Meltzer 2001).

During the 19705, marijuana cultivation came to Colombia as eradication
efforts were commencing in Mexico. Colombians then shifted to exporting
cocaine during the late 1980s as marijuana cultivation shifted back to Mexico.
During this time period, the raw materials (the coca leaves) were being imported

from Bolivia and Peru and the coca paste was produced in Colombia. This

13

changed again however in the early 19903 as Colombia began growing opium
poppies, began producing and exporting heroin, and began growing much more
coca (Vargas 2002). Poppy production has also been a target of Plan Colombia,
however the small amount of area in poppy production (3,828 ha in 2002 ) and
the sparse distribution throughout the mountainous areas of Colombia make it
very difficult to study (UNODC 2003). The 19803 and 1990s also saw the rise
and fall of the major drug cartels in Colombia. The Medellin and Cali cartels
grew in power and during this time period owned over one twelfth of Colombia’s
land or roughly 9.5 million ha (Ehrenfeld 1990). The Medellin cartel, during the
early 1980s, launched a brutal campaign of terrorism against Colombia's
government and citizens (Dishman 2001). The emergence of groups, as
powerful as the cartels, specifically the left-wing guerillas and paramilitaries, has
been a major factor in the destabilization of Colombia. The coca growth has
escalated since the early 1990s (as noted earlier) as Bolivia and Peru’s
production of coca has declined significantly. Much of the coca in Colombia is
grown in regions located outside of the realm of control of the Colombian
government (areas controlled by guerillas or paramilitaries). Although the
numbers vary significantly, estimates place the 68% of Colombia's cocaine
production in the departments of Putumayo, Guaviare, and Caqueta (DNE 2000)
as reported in (Moreno-Sanchez et al. 2003) and 50% in Putumayo alone
(Penhaul 2001). Coca came to Putumayo in the late 1970’s, but was not a major
factor until the mid-1990’s when fumigation efforts by the Colombian government

in Guaviare and Caqueta drove much of the coca production to the more remote

14

Putumayo department (lsacson and Vaicius 2001). The late 1990’s also saw the
demise of the Cali cartel, which had purchased their much of their coca leaves
from Peru. This, along with Peruvian efforts to battle coca growth locally, led
narcotics traffickers to promote the growth of coca in Colombia (Rojas 2003).
Between 1990 and 1995, the area in coca cultivation more than doubled to
80,000 ha and that number swelled to between 120,000 and 150,000 ha in
cultivation by 2000 (Thoumi 2002); (Uribe 1997), rose to 169,800 ha in 2002,
before finally falling to 144,450 ha in 2003 (United States General Accounting
Office (USGAO 2003) as reported by the Office of National Drug Control Policy
[ONDCP]). However, as is often the case with many of the reports, there are
discrepancies between one group’s figures compared to others. The UNODC
(2003a) estimates for area in coca cultivation are found in Figure 2.1.

In Colombia, much of the coca (estimates vary greatly) is grown on small
plots by poor subsistence farmers who have few alternatives for the production of
cash crops (Cooper 2001; Fratepietro 2001; Stauder 2001). The first step in the
processing of the coca leaves usually occurs very near the fields in rudimentary
laboratories (often nothing more than wooden shacks) as farmers use gasoline,
bleach, and sulfuric acid to create the coca paste from the coca leaves (Dudley
2000). A farmer can sell a kilogram of coca paste for $900, which is enough to

cover production costs, pay his workers, and feed his family (Stauder 2001).

15

 

 

180
160- '
140 q .__o.

120-

Ha x 1000 in coca cultivation

 

 

 

j I T I I I I I I
1994 1995 1996 1997 1998 1999 2000 2001 2002
Year

 

 

 

Figure 2.1 UNODC estimates of coca cultivation in Colombia

However, the paramilitary group or guerilla group who control the area
often tax the money earned by the farmer from growing coca. The paramilitaries
and guerilla groups use this tax to finance their wars against each other and the
Colombian government. Colombian authorities have reported that the FARC
makes nearly $500 million dollars per year in coca growing areas and the right-
wing paramilitaries, with unofficial help of the Colombian military and police,
make over $200 million per year via drug trafficking (Dudley 2000). These taxes,
combined with farming expenses, often leave the farmer making only slightly
more than the Colombian minimum wage of $150 per month ($1800 per year)

(Fratepietro 2001; lsacson and Vaicius 2001). Although the farmers are

16

surviving, they are not prospering. In some regions of Colombia, the guerilla
groups are so involved in the dmg production and trafficking that the distinction
between fighting the drug war and fighting the guerilla war is blurred as the same
equipment, intelligence, and personnel are used in both. US policy has been
tough to implement as the focus is solely to combat drug trafficking, not the

guerilla insurgency (Bibes 2001).

PLAN COLOMBIA

The US has a history of involvement in Colombian affairs that dates back
to the early 19603 (Petras 2001). “Plan Colombia” was developed during the
Clinton Administration as an attempt to reduce the impacts of drugs in the United
States and was signed into law in 2000 (Crandall 2002). The specific goals of
Plan Colombia are: a push into Southern Colombia; support for narcotics
interdiction efforts; support for the Colombian National Police; support for
developmental and particularly alternative development programs and
approaches; support for justice and other social sector reform; and support for
issues outside of Colombia, forward-operating locations, as well as several other
counties that are affected by what happens in Colombia (USDS 2001). By 2002,
Colombia had received over $2 billion for Plan Colombia with $1.7 billion coming
from the United States and the rest contributed by various European countries,
Canada, and Japan (Frechette 2003). Plan Colombia’s support for narcotics
interdiction efforts includes the use of aerial fumigation, which is employed in an

effort to eradicate the cultivation of coca plants in Colombia, and voluntary

l7

manual eradication pacts with aid to help farmers make a living with legal crops
(USDS 2002). Plan Colombia was renamed the Andean Regional Initiative by
President George W. Bush and continues to be sold as the key component in the
war on drugs (Driver 2001). (Although the program has been renamed, it is still
generally referred to as “Plan Colombia” and will be referred to as such
throughout the rest of this document)

The aerial fumigation program is carried out using crop dusters to spray a
herbicide called RoundUp Ultra (a combination of glyphosate, Cosmo Flux-411f,
and Cosmo-iN-D) onto the fields (lsacson and Vaicius 2001, many others have
also covered this topic). Glyphosate, the main ingredient used in RoundUp Ultra,
is the most widely used herbicide used in the United States and is generally not
harmful to humans and animals (USDS 2002). The other ingredients in
RoundUp Ultra are used to 1: more easily covert the liquid into a mist and 2:
allow the substance to stick to the vegetation better (as RoundUp in its native
form can be washed off with water, a concern in the humid tropics). The
fumigation is carried out using a mixture of methods and equipment. The first
stage is a reconnaissance flight where the aircraft is equipped with a digital
imaging system that records multispectral information and identifies crop types.
This device is connected to 3 Geographic Positioning System (GPS) receiver that
records position information to be used with the crop type data. Mission planning
for the actual spraying aircraft is carried out using this data, however the pilots
use a visual identification method before spraying (USDS 2002). In 2000, with

significant support from the US, the Colombian Government began this major

18

aerial eradication program in Putumayo and promised support to farmers who
willingly self-eradicated their own coca crops (Ernst 2002). Over the last few
years, two major fumigation efforts have taken place in Putumayo. The first
period of flights lasted from December 2000 to February 2001 covering 25,000
hectares and the second, from July to October of 2002, covered 60,500 hectares
(Vaicius and lsacson 2003). Because of the different methods used, reports of
the amount of area fumigated vary greatly (Ford 2003). The UNODC, using
remotely sensed imagery, report slightly higher numbers in Putumayo with
32,506 ha fumigated in 2001 and 71,891 ha in 2002 (UNODC 2003). In a
hearing before the US Senate, Colombian Vice President Santos-Calderon (US
Senate 2003) reported that overall in Colombia, 130,000 ha were sprayed in
2002, 65,000 ha were sprayed between January 2003 and June 2003, and the
goal was to spray 150,000 ha by the end of 2003. The UNODC estimates for the
amount of area fumigated in Colombia from 1994-2002 are found in Figure 2.2.
Although spraying has occurred over large, industrial-type coca fields,
much of the coca growing area in Putumayo is composed of small coca and
agricultural fields interwoven together on the landscape. The complex layout of
the Colombian landscape and the effects of wind drift and runoff have resulted in
many non-coca fields being adversely affected by the fumigation. Many reports
exist that detail how food and other subsistence crops have been sprayed and
destroyed (Berger 2001; Cooper 2001; Fratepietro 2001; lsacson and Vaicius
2001; Kratz 2002; Penhaul 2001; Peterson 2002; Stauder 2001; Vaicius and

lsacson 2003; Wilson 2003).

19

 

 

 

 

 

150
'o '.
.93 120 - '
m
2’ ,.
“" 90 ~
O "
O
O
‘— o
X _. .*
:‘E 60 2 g

o 0
$6? 30
O o .............. _.
0 I I I I I I I I I
1994 1995 1996 1997 1998 1999 2000 2001 2002
Year

 

 

 

Figure 2.2 UNODC estimates of coca fumigation in Colombia

Often, the farmers grow food crops in fields directly adjacent to their coca crops
or intermixed with coca crops and have seen these fields damaged or destroyed
by the fumigation efforts (Forero 2002). County officials in the Guamuez Valley
have tallied more than 800 cases where there was a claim of legal crops being
destroyed by spraying (Hodgson 2001) (see Figure 2.3).

Others argue that fumigation is also harmful to ecosystems and the
environment, threatening the biodiversity of the area (Peterson 2002; Vargas
2002), and fosters displacement of farmers, causing deforestation (Dudley 2000;

Ecuador 2001; Forero 2002; Vargas 2002; Wilson 2003).

20

 

 

   

COLOMBIA

0°..-7.i.----.7 7- i- _.._.i.._

ECUADOR

 

 

 

 

 

 

Figure 2.3 Guamez Valley Map

As of 2000, there were already reports of hundreds of Colombian refugees
fleeing across the border to Ecuador causing the US to increase related aid to
Ecuador (LaFranchi 2000) and during the first 10 months of 2001, an estimated
13,500 people fled Colombia with most of them going to Ecuador (Valenzuela
2002)

There also exists the possibility of moving coca production to bordering
countries. Reports have surfaced of Ecuadorian farmers living near the southern
Colombian border being offered money by drug traffickers to begin coca
production (Cooper 2001). The overseer of the United Nations Drug Control
Programs’s (UNDCP) office in Colombia, Klaus Nyholm, commented,
“Fumigation has an effect, but we would argue it‘s an effect of displacement”
(Forero 2002). Critics of Plan Colombia also point to statistics from the UN and

UNDCP (2002a) that showed a rise in the amount of area in coca cultivation

21

despite the eradication efforts. They use this as evidence that eradication is a
failed approach to drug control (Moreno-Sanchez et al. 2003).

More recently the US has attempted to distance themselves from further
promising more help to Colombia. President Bush’s administration has used
words such as “endgame” and “exit strategy" in the descriptions of future
strategies pertaining to Colombian aid, even as Colombian President Alvaro

Uribe seeks more help (Wilson 2003).

MONITORING VEGE TA TION USING REMOTE SENSING

The data collected by many earth observing satellites throughout the past
30+ years has given birth to a variety of methods for researchers to use in
monitoring vegetation change. The development of extensive imagery
databases has been the major factor contributing to the growth of remote sensing
technologies for studying change detection (Lunetta et al. 2002). Also, as time
progressed, visual interpretation of aerial photography and satellite imagery has
given way to quantitative and qualitative measures of the information collected by
the sensor. These measures attempt to remove the bias that can occur during
human interpretation. Included in these measures are numerous vegetation
indices that try to model the characteristics and the biophysical attributes of the
vegetation present on the ground. These indices are a primary source of
information for operational monitoring of vegetation cover (Gilabert et al. 2002).
Many of the indices have become very popular such as the Normalized

Difference Vegetation Index (NDVI) (Rouse et al. 1974), the Soil Adjusted

22

Vegetation Index (SAVI) (Huete 1988), the Modified SAVI (MSAVI) (Qi et al.
1994), the Leaf Area Index (LAI) (Price and Bausch 1995), and Green Fractional
Coverage (fc) (Qi et al. 2000) and research continues towards efforts to find the
most accurate relationships between the calculated biophysical variables and the
true state of the vegetation on the ground (Cohen et al. 2003; Xavier and
Vettorazzi 2004). Other research has also been conducted that models
statistical relationships between each of these indices (Carlson and Ripley 1997;
Leprieur et al. 2000; North 2002; Purevdorj et al. 1998; Zha et al. 2003). The
most widely used vegetation index is NDVI as it is a simple algorithm to compute
and most of the satellite imagery in use provide the necessary spectral bands for
computation. NDVI and the other vegetation indices can be extremely useful in
change detection studies using time-series data over the same area. Fung and
Siu (2000) used the differences between image NDVI values to monitor the
environmental changes occurring in Hong Kong from 1987 to 1995 due to such
factors as land reclamation (from the sea), new urban development, and forest
fires. Weiss et al. (2001) used the Coefficient of Variation observed between
NDVI values to evaluate the effects of grazing in Saudi Arabia’s rangelands.
Other studies have used observed changes in NDVI to create masks to aid in the
search for areas most likely to have undergone a LULC change (Lunetta et al.
2002)

An index that has been derived from NDVI in an attempt to model the
percent of a pixel covered with green vegetation is fc. Green vegetation cover is

an important factor in vegetation status and as an indicator of land degradation

23

(Purevdorj et al. 1998). Although multiple methods exist that attempt to derive fc.
the most popular is a process of linear spectral unmixing of the NDVI value. The
radiation reflected from heterogeneous materials and recorded in a satellite pixel
can be considered a mixture of a number of spectrally pure materials (Van Der
Meer 1999). Di et al. (2000) used this concept assuming that the pixel signal is
made up of two components, vegetation and non-vegetation. This study was a
preliminary investigation without a proven method of ground truth data collection,
however the results were encouraging. Previous studies have also been
attempted and showed success using both NDVI and LAI in the estimation of fc

(Carlson and Ripley 1997; Choudhury et al. 1994; Gillies and Carlson 1995).

24

Chapter 3 : METHODS

IMAGE PRE—PROCESSING

Three Landsat ETM+ scenes were ordered from the Center for Global
Change and Earth Observations at Michigan State University (through
www.landsat.org). The three co-located scenes (Path 9, Row 60 of the World
Reference System) and were acquired by the satellite on September 9, 2001;
September 12, 2002; and October 14, 2002 (see Figure 3.1). These scenes
represent the only available images with less than 30% cloud cover that have
been acquired by the satellite since the most recent Plan Colombia fumigation
commenced. Although these are the only images available, the September and
October images were collected during the second of the two major fumigation
efforts in Putumayo which (as previously mentioned) lasted from July to October
of 2002.

The original images on Compact Discs were imported from the
Hierarchical Data Format (HDF) directly into ERDAS Imagine as raw 8-bit data
representing Digital Pixel Values or Digital Numbers (DNs). Because the images
were going to be compared against each other quantitatively, a series of steps
were taken to normalize sensor collection errors and atmospheric effects present

in each image.

25

 

Landsat ETM+ Scenes (Band combination 4,3,2)

,7 a

         

h’. "..

 

09/09/2001 09/12/2002 10/14/2002

 

 

 

Figure 3.1 Landsat ETM+ Scenes (Path 9, Row 60)

The first step was to perform a Radiometric correction on each of the images.
The Radiometric correction involves converting the DN3 of the image to
Radiance values then converting the Radiance values to Top of Atmosphere
(TOA) reflectance values. The equation for the first step, conversion to Radiance

(Equation 3.1), is found below (NASA).

Lma _Lmin
L1 =( m )*( Qcal—Qcalmin )+Lmimt

where:

L ,1 = Spectral Radiance at the sensor’s aperture
Lmaxl = Spectral radiance that is scaled to Qcalmax
me = Spectral radiance that is scaled to Qcalmin
Qcal = Quantized calibrated pixel value or DN
Qcalmax = Maximum quantized calibrated pixel value
Qcalmin = Minimum quantized calibrated pixel value

Equation 3.1 Conversion of DNs to Radiance values

26

Next, the images were converted to TOA reflectance values using Equation 3.2.

 

mL *dz
( i )

Pp: E

sun/1 * 60868

where:

p, = Planetary reflectance

L ,1 = Spectral Radiance at the sensor’s aperture

d = Earth-Sun distance in astronomical units
Em) = Mean solar exoatmospheric irradiances

c0361)s = Solar zenith angle

Equation 3.2 Conversion of Radiance to TOA reflectance

After the images were converted to TOA reflectance values, a Simulation
of the Satellite Signal in the Solar Spectrum (58) atmospheric correction (T anre
et al. 1990) was applied to remove the effects of the atmosphere on the signal
and to bring the pixel values to surface reflectance values. The code in the 5S
correction computes the solar radiation that is backscattered by the surface and
atmosphere as observed by the satellite (Mackay et al. 1998). The decision to
use an atmospheric correction was driven by studies showing the advantages of
correction for change detection studies (Song et al. 2001) and because NDVI
was going to calculated. Teillet et al. (1997) point out that the best
representation of NDVI derived from image data is the NDVI of the surface

reflectance values (p) as it is a function of the reflectance values of the

vegetation and will be skewed by the effects of the atmosphere found in pp.

27

Although, there are much more robust methods for this correction, they require
detailed information on the local atmospheric conditions on the date of image
collection. That information was not available for this area so these methods
could not be used. The input parameters used for the SS correction are found in
Table 3.1 and were easily obtainable for the study area except for visibility. l was
not able to locate the on-ground visibility data for this region at the specified time
period. I chose to use the default value of 17 miles and used this for all the
images. The parameters, site elevation, sensor elevation, sensor zenith, sensor
azimuth, model type, standard atmospheric model, aerosol model, and visibility
did not change between images as all images were collected by the same
satellite and the study area is in the same location in all images. The only
parameters that were different between images were solar zenith, solar azimuth,
and date. I obtained these parameters by viewing the header file of the original

image data.

 

Site Elevation
Sensor Elevation
Solar Zenith
Solar Azimuth
Sensor Zenith
Sensor Azimuth
Model Type
Standard Atmospheric Model
Aerosol Model
Visibility

Date

 

 

 

 

 

 

 

 

 

 

 

 

 

Table 3.1 58 atmospheric correction input parameters

28

Other methods of atmospheric correction such as Dark Object Subtraction (DOS)
and Psuedo-Invariant Feature (PIF) are easier to use, however they do not
perform as well (Song et al. 2001). The output parameters of the 5S correction
are a series of values modeling surface reflectance at each band. A model was
constructed by using a linear regression through the corresponding p and pp
values. The 58 atmospheric correction was implemented by using the equations
gathered from the 58 code and thelinear regression to create the new image in
ERDAS Model Maker. The detailed input and output parameters used for the
correction of each image and can be found in Appendix A and Appendix B.

Equation 3.3 is a simplified representation of the 58 correction of the image.

p = f ( p p , atmospheric conditions)
where:

p p = Planetary reflectance
,0 = Surface reflectance

Equation 3.3 58 atmospheric correction

Once the images were converted to surface reflectance, the Normalized
Difference Vegetation Index (NDVI), a measure of vegetation vigor, was
calculated (see Equation 3.4) using Band 4 (NIR) and Band 3 (Red) of each
image (Weiss et al. 2001). NDVI was calculated as a transition product for fc
calculation. A model was constructed in ERDAS Imagine to create the NDVI

layer for each image.

29

NDVI ____( PNIR - pRed )
PNIR + pRed

where:

pNIR = Near Infi'ared pixel reflectance value (Landsat ETM+ Band 4)
pm = Red pixel reflectance value (Landsat ETM+ Band 3)

Equation 3.4 NDVI calculation

The next step was to calculate the fc for each pixel. Green fractional
coverage (fc) is a measure of the areal amount of a pixel that is covered by
green vegetation (Qi et al. 2000). This is not biomass or Leaf Area Index (LAI)
which attempt to add a 3rd dimension (volume) to the calculation (in the case of
LAI, although it’s name implies the measurement of area, an attempt is made to
model and detect situations where more than one piece of vegetation is directly
below another and is more of a density measure). The calculation of fc is an
attempt to construct a 2D model of the amount of area within a pixel that has
green vegetation between the sensor and the ground. Typically, the vegetation
contained within pixel boundaries is not homogenous therefore fc is the measure
of the combination of all the vegetated areas and bare soil areas contained within

a pixel boundary (Equation 3.5).

II
fc = Zarean * % vegetation covern
o
where:

arean = Area represented by a homogenous % vegetation cover
% vegatation covern = The percent of green vegetation covering the ground

Equation 3.5 fc definition

30

Figure 3.2 shows how the fc of a pixel is determined by the amount of area
covered with green vegetation within the pixel boundaries. It also shows some of

the infinitely possible scenarios for an fc pixel value (Scanlon et al. 2002).

 

Theoretical pixel 1 Theoretical pixel 2
Bare soil

    

 

Forest with
75%

canopy
coverage

  
 
  
   

Forest with Bare $0“
100% canopy vevltgtgti/oon
coverage cog/erage

 

 

Theoretical pixel 3 Theoretical pixel 4

 

Pasture with
10% vegetation

coverage Pasture Forest
with 10% with 90%
vegetation canopy

Forest with coverage coverage

90% canopy
coverage

 

 

prier1 = (75%)*(100%) + (25%)*(0%) = 75%
prier 2 = (75%)*(0%) + (25%)*(75%) = 18.75%
prier 3 = (50%)*(90%) + (50%)*(10%) = 50%
prier 4 = (50%)*(90%) + (50%)*(10%) = 50%

 

 

 

Figure 3.2 Green Fractional Coverage

31

The fc layers were calculated from the existing NDVI layers using Equation 3.6
(Qi et al. 2000).

NDVI -— NDVIsoil )
NDVIveg — NDVIsoil

 

fc=(

where:

NDVI = NDVI value of pixel
NDVIso" = NDVI value of a 100% bare soil pixel
N DVIveg = NDVI value of a 100% vegetated pixel

Equation 3.6 fc calculation

Because the possibility for human error involved in calculating fc from NDVI, the
NDVlsoi. and NDVI...g values are very important. These values were selected by
first creating a small Area Of Interest (AOl) (25 x 25 pixels) over primary forests
and the African palm plantations for the 100% vegetated pixels (NDVlveg) with
downtown urban areas and cleared areas for the 100% bare pixels (NDVIsoil).
Because this step involved a visual interpretation of the images, an attempt was
made to keep the location of the AOls very close between all three images. The
descriptive statistics from inside the AOI were then used to select the end

member values (see Table 3.2).

 

 

 

 

Landsat Image NDVI,“ NDVlson
10/14/2002 0.97 0.10
9/12/2002 0.95 0.10
9/9/2001 0.95 0.10

 

 

 

 

 

Table 3.2 NDVI values used to calculate fc for Landsat images

The maximum NDVI value inside the AOls was assumed to be the best

representation of the NDVI value for a 100% vegetated pixel and minimum NDVI

32

value (taken from separate AOls) for the 100% bare pixel). Although a concerted
effort was taken to normalize the images, the variation in the NDVI.” value from
the10/14/2002 image is a result of the small collection differences that occur
between every Landsat image and cannot be fully removed.

The decision to use fc to measure defoliation was based on the theory
that fc actually represents a biophysical measure of the vegetation on the
ground. As noted earlier, the defoliation statistics produced by the UNODC and
other sources are based on the visual interpretation of satellite imagery. l
hypothesize that visual interpretation of the defoliation cannot accurately identify
the smaller amounts of defoliation occurring due to drift and/or runoff.

The last step in the image pre-processing step was to geometrically rectify
the images and the fc layers. This correction was completed using an image
previously rectified (supplied by the University of North Carolina at Chapel Hill,
Carolina Population Center) and finding features identifiable on all images to use
as control points. Once the control points were identified, each image was
geometrically rectified using a 2"d order polynomial transformation (using the
Nearest Neighbor interpolation method) and a model of the rectification

parameters for each image was generated. The three images were all corrected

to an acceptable level of accuracy (see Table 3.3).

 

 

 

 

 

 

 

 

 

 

 

 

9/9/2001 9/12/2002 10/14/2002
Error (pixels) Error (meters) Error (pixels) Error (meters) Error (pixels) Error (meters)
X 0.0179 0.537 0.0884 2.652 0.0234 0.702
Y 0.0227 0.681 0.0456 1.368 0.0262 0.786
Total 0.0289 0.867 0.0884 2.652 0.0351 1.053

 

Table 3.3 Geometric rectification errors

33

 

The fc layers and the original images were then rectified using the
correction models generated in the previous step. This correction was applied to
the fc layer after computation in an effort to reduce the affects that could occur
by calculating a biophysical variable (which are very dependent on pixel values)
from pixels that had been interpolated during the rectification process.

After the rectification process was completed, a cloud mask was created
by heads-up digitizing any and all areas covered by clouds or their shadows in
the study area. I attempted to create a classification method that would identify
clouds automatically, however I was not able to accurately classify the clouds
and shadows. Most of the problem in this was due to the edges of the clouds
were signal from the ground reaches the sensor, but is slightly degraded by the
cloud. After multiple attempts to solve this problem, I decided that the only
method that could produce the level of accuracy for cloud removal was to
manually trace the borders of the clouds and shadows, hence the heads-up
digitizing. Unless othenivise noted, this cloud mask was used to exclude these
areas in all of the following steps in image processing, sample point generation,

and change detection.

SAMPLE POINTS

The to layers were created for use in preliminary investigations and also
as the basis for a stratified random sampling scheme. The stratified random
scheme was implemented as an attempt to sample pixels that covered the range

(0% - 100%) of fc values that could be found both on the landscape and in the

34

image. The sample points were based on fc values from the 09/12/02 image as
it contained significantly less cloud cover throughout the study area. Although
the 10/14/02 image date is closer to the dates of ground truthing, the clouds
sprinkled throughout the study area would have voided too much of the image.
The first step in creating the stratified random sampling scheme was to
subset the pixels of the fc layer of the 09/12/02 image that were within 40 km of
Coca, Ecuador and within 250 meters of roads, dirt tracks, and foot paths (see
Figure 3.3). The distance of 40 km was chosen using knowledge gained from
people who had navigated this region. This distance that would allow the field
team adequate time to drive to a point, collect data, and return to the base of
operations in one day. The 250 meter buffer of roads, dirt tracks, and foot paths,
was based on similar knowledge as mentioned in the previous sentences. The
region offers many obstacles that hinder travel away from main roads such as
streams, thick forests, varying topography, wild animals, and heavy rainfalls. By
only including pixels within 250 meters of roads, this would reduce the time to
walk to each sample point. Additional masks were created as AOls (by heads-up
digitizing) to remove unreachable areas (African Palm plantations and river
islands). These masks and the aforementioned cloud mask were used to
exclude unreachable and unusable pixels from consideration for sample
locations. The combination of pixels not masked and within the buffered distance
of roads and the city of Coca left an fc layer that included all of the pixels that

were considered acceptable for sampling (see Table 3.4 and Figure 3.3).

35

 

A pixel to be considered for sampling must be:
Within 40 km of Coca, Ecuador

Within 250 m of road, dirt track, foot path
Not within a cloud or cloud shadow

Not within an African Palm plantation

Not within an island

 

 

 

 

 

 

 

 

Table 3.4 Criteria for a sample pixel

This layer was then separated by fc value into 10 separate classes (see Table
3.5) so an equal amount of sample points would be generated for each class.
Although most of the landscape is composed of pixels of 50% fc or above, the
less than 50% fc classes were collected as an attempt to observe as wide a

range of fc values as possible for the construction of a statistical model.

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Class fc values (%)
l 0 - 10
2 10 - 20
3 20 - 30
4 30 - 40
5 40 - 50
6 50 - 60
7 60 - 70
8 70 - 80
9 80 - 90

10 90 - 100

 

Table 3.5 fc classes for stratified random points

30 spatially random points were created inside each class giving 300 unique
sample locations. Another 300 points were generated as a backup to the original

300 in case of loss of data or inaccessibility of the original locations.

36

 

 

 

 

 

 

Figure 3.3 Sampling area and sample points

The decision to create 300 points was based on the amount of time that was
allotted for fieldwork. When collecting points for an accuracy assessment of a

LULC classification, the recommended amount of ground truth points per class is

37

50 (Congalton and Green 1999). However, the sample point generation was
based on collecting data for validating fc and not LULC. Therefore, the number
of ground truth points only had to be enough to build a statistically significant

relationship between the image fc values and the ground truth fc values.

FIELD EQUIPMENT AND MAPS

Leica LRF 800 Rangefinde — used to measure distances from observer to target

Pentax lQZoom 1045 WR 35mm Camera — used for general photos of the sample
point

Garmin GPS 12XL Unit - used to navigate to sample points

Nikon Coolpix 4500 Digital Camera — used for general photos and for fisheye
canopy photos

Trimble GeoEprorer lI Unit — used to collect a GPS position at each sample
point

Spherical Densiomete — carried as backup instrument for collecting canopy
coverage data

Gateway 3450 Laptop — used for storing data collected each day

Brunton Compass - used for recording bearing data at sample points

A series of maps were constructed for field navigation. These maps were
created from the 10/14/02 Landsat ETM+ image that had been sharpened using
the 15m resolution panchromatic band. I chose to create the maps with a 7,4,2
band combination, which results in vegetation appearing green on the image. I
was attempting to avoid the confusion that often results when showing the maps

produced with the popular 4,3,2 combination (resulting in vegetation appearing

38

red) to people not familiar with satellite imagery (especially drivers and property
owners who I assumed would be viewing the maps). The maps’ scale was
1:60388.57 which maximized the map area on a sheet of 8.5" x 11” paper and
each map contained the image as a backdrop with the sample points and the

road network overlayed.

FIELD DA TA COLLECTION

Field data collection took place in and around Coca, Ecuador from June 4
to June 21 of 2003. Coca is roughly 100 km from the center of the intensive
study area. However, as mentioned previously (Chapter 1), the physical
(vegetation and soils) and climatic (temperature and precipitation) conditions are
the same in both areas. The only differences in the area are due to
anthropogenic causes. Considering that both areas fall in the same Landsat
scene and are physically similar, data collected in and around Coca will be used
to validate fc over the entire study area.

We navigated by vehicle or on foot to each sample point location using the
Garmin GPS unit. When the point fell on land that was not privately owned, we
walked to the point and collected our information. If the point fell on private
property, such as in a field behind someone’s house, we explained our purpose
to the owner of the property and asked permission to collect information on their

land. In most situations, the owner complied and allowed us to collect our data,

39

 

300000 301“” 302000 303000 304000 305000 306000 307000
9907060 .

9953050
9952050
9901050
9960060 .~

9969060

9956060

   

9965060

Map datum: W65 84
Map projection: UTM 188
Map units: Meters

308000 3090‘!) 310000

 

300000 301000 302000 303000 304000 306000 300000 307000 309000 309000 310000

Sheet 17 10/14/2002 ETM+ Study Area

9967050

9986060

9964060

9903060

9962060

9961060

9900060

9969060

9967060

9960060

  

9966060

Scale

 

Paul Delamater / Joe Messina

Ecuador field 3

06/01/2003 - 06/26/2003

M' h' State University

Center for Global Change and Earth Observation:

 

gll

0
1 : 00368.67

 

Figure 3.4 Example of a field map sheet

40

 

however, in a few cases we were denied permission. Once we reached a point,
the Trimble GeoExplorerll GPS unit was used to collect the precise location of
the point (minimum of 200 readings with a PDOP blanket of 4). While collecting
the location with the Trimble unit, the LULC and fa attribute information was
recorded. In each location, a GPS point was collected. However, in areas of
homogenous cover (LULC or fc cover), polygon information was recorded. This
method included using the laser rangefinder to find the distance from the point to
a different cover and then using a compass to observe the bearing of which the
distance measurement was taken (see Figure 3.5). In most cases, this involved
collecting 4 readings to accurately describe the homogenous area of the
landscape. As many as 8 readings were collected in other situations where the
homogenous area was oddly shaped. Collecting the polygon information allowed
for more than one pixel of ground truth data to be collected while only taking the
time to collect one GPS point.

Whereas, there does not exist a proven method for collecting fc ground
truth data where a canopy is not present, a method of collection and a set of
decision rules were created in an effort to match ground truth data to the theory
behind fc. At each sample location, a theoretical “pixel” (matching the 30m x
30m Landsat pixel) was constmcted using the rangefinder and compass. In
areas where the ground cover is homogenous, l visually interpreted the fraction

coverage. This visual interpretation was only completed by one person (Paul

41

fl surveyed
homogenous

polygon

> distance (m)
bearing (°)

 

 

Figure 3.5 Collection of homogenous (LULC or fc) polygon information

Delamater) and, as a quality control, verified by others in the group (Dr. Joe
Messina or Francis Baquero).

o The visual estimation of the fractional coverage was based on the
vegetation in a polygon area (collected with the rangefinder and compass)
or the vegetation within the distance of one pixel (15-20 meters in every
direction from the GPS point). The following factors (in order of
importance) were used to determine the actual estimate at each collection
point.

1. Amount of vegetation — the amount of individual pieces of

vegetation

42

2. Greenness of the vegetation — the greenness of the vegetation (as
opposed to brownness or yellowness)

3. Proximity of the vegetation — the distance between the individual
pieces of vegetation

4. Thickness of the vegetation — the width of the individual pieces of
vegetation

5. Height of the vegetation — the distance from the ground to the top of
the vegetation

6. Other factors such as amount of water present went into the
estimation as often areas were slightly flooded because of heavy
rains that had just occurred. An attempt was made to distinguish
between permanently flooded areas and areas that were only
temporarily flooded.

Considering all of these factors, an fc percentage was estimated from 0%
to 100%. The units were estimated in 5% minimum unit increments (eg. 80 or
85, but not 83). The decision to estimate at this level was based on the
confidence of estimation and the lack of another method to gather fractional
coverage information of a non-canopy area. In a few cases (urban settings), the
fractional coverage was listed at 2% or 3%. This was a result of the amount of
vegetation in an area being present but confidence of at least 5% in the pixel
area being low. In non-homogenous areas, the fa estimation was completed by
using the previously stated method, then factoring in the amount of bare area

and/or the amount of different vegetation percentages in the pixel area (see

43

Figure 3.6). This was often accomplished easily in the field estimation. Very
complex and heterogeneous pixels often required a sketch that was later
interpreted manually with a calculator and a transparent grid. This is much like
former methods for estimating area from an aerial photo. (Lay the grid on top of
the sketch and count boxes at each fractional coverage percent. Take the

counts and percents and compute the fractional coverage for the whole area).

 

Example sample 1 Example sample 2

sample
point
0% 0% _
vegetation vegetation

coverage coverage

   

 

 

 

 

 

 

Figure 3.6 Examples of possible non-homogenous vegetation covers at sample
point locations

In forested areas we used the 4500 camera with the fisheye lens for our field
measurement of fc. To do this, we held the camera above head, and as level as
possible (as close to vertical) and shoot the picture of the canopy (See Figure 3.7
for examples of the fisheye canopy photos). This was made quite easy as the
the 4500 is designed to photograph vertically (it has a swiveling viewfinder that
allows the camera to be held comfortably while the lens is pointing straight up).

Then I recorded the fc information of the understory, which was defined as any

 

June 7, 2003

June 18, 2003

 

 

 

 

Figure 3.7 Examples of canopy photos taken with the fisheye lens

45

vegetation lower than the level of the camera at photo collection, using the
previously detailed visual collection method.

All information was collected on a field sheet (see Figure 3.8 and
Appendix C), which was created, from previous field sheets used in Ecuador by
Dr. Joe Messina during similar data collection trips. The field sheet was
designed to maximize the amount of area for the primary information to be
recorded while also allowing for ease of use in recording. Also considered were
areas on the sheet for ancillary information to be collected and an attempt was
made to keep the sheet at one half of a page (to fit 4 sample collections on one
piece of paper).

The field team was able to collect 211 points in the 18 days in Coca,
Ecuador. This number was less than the 300 point goal, however considering
the unpredictability of the weather (rain) and the sometimes rigorous challenges
associated with collecting points on the Amazonian landscape, we were pleased

with the amount of data collected.

46

 

Site#

 

Date Collection type

 

LULC type

Travel time

 

 

General description

 

Waypoint file name

Coordinate x Coordinate Y

 

 

# pts in file

Rover file name

 

Photo #

 

 

Photo type Canopy height

 

Polygon collection and Sketch area

 

 

 

 

Visual fractional coverage Information and notes

 

Figure 3.8 Field data collection form

Each night during field data collection, the points collected during the day
were downloaded from the Trimble GPS units to a laptop computer.
attribute data gathered were entered into a Microsoft Access database that

contained a relational field to link the database information to the GPS point

FIELD DA TA PROCESSING

47

 

location. After field data collection, the GPS points were differentially corrected
using base point data collected in Quito, Ecuador. Unfortunately, three days
worth of base point data were unavailable from the base station and therefore,
the points collected on these days were not useable. Once the GPS data points
were differentially corrected, they were converted to an ArclNFO coverage using
ArcTools.

The next step was to calculate the fa of the points taken under a canopy.
This was performed using the Gap Light Analyzer” (GLA) software created by
Simon Fraser University and the Institute of Ecosystem Studies. This software
was specifically built for use with fisheye lens photography to measure the
amount of canopy coverage by converting the photograph to a binary image and
comparing pixels of sky to canopy. The inputs for the GLA software included:
absolute center of the image, the dimensions of the photo, the distance from the
center of the image, and the threshold for sky. The threshold value is where
adjustments were made for the varied conditions of the sky in each image. After
viewing all of the images, sky conditions were categorized into the following table
and a threshold value (between 0-255) was chosen to best represent the

difference between sky and canopy for each category (see Table 3.6).

 

 

 

 

 

 

 

 

Sky Condition Threshold level
Sunny 180
Cloudy 140
Hidden Sunny 125

ery Cloudy 75

 

Table 3.6 GLA threshold values for fisheye lens canopy photos

48

After the threshold value was selected, the image was converted to binary data
type as a representation of sky and canopy pixels. The GLA software then
constructed a table with the photograph's file name and the canopy coverage
attributes. The percentage of canopy coverage at each point (generated by the
GLA software) and the understory fc values (recorded in the field) were then
used to construct a model of the true fc at each point by utilizing a custom
neutral model built in the Research Systems Inc. (RSI) Interactive Data
Language (IDLW). Neutral models are spatial models often used in landscape
ecology studies (Gardner et al. 1987). One specific use is as a means of
generating landscapes that share statistical properties with an observed
landscape (Keitt 2000). Although the canopy photos are not a traditional
landscape generally used in landscape ecology, they are a spatial representation
of the vegetation over an area. The fc value at each point was calculated by
modeling the canopy fc and the understory fc using neutral models, then
overlaying the layers to create a comprehensive model of the vegetation cover.
This method implies a random distribution of the vegetation throughout the
canopy and understory. Although this may not be the most robust method to
model the vegetation, it makes an attempt to include all of the vegetation found at
each point. First, a binary neutral model representing the canopy photo, with the
exact same dimensions and shape (from the inputs of the GLA software), was
randomly generated using the fc value calculated by the GLA software. Next,
another binary neutral model (with the same dimensions as the canopy model)

was created using the fc value of the understory. The two neutral models were

49

combined as any pixel with a value of 1 or 2 representing vegetation cover (see
Appendix D for IDL code). The vegetated pixels were then compared to the total
number of pixels to create the fc value at the point location. These values were
then entered into the attribute data contained in the Access database.

The final attribute data were imported as a table from the Access
database and joined to the point coverage using a table join in ArcMap. The
GPS points with polygon information were then subsetted using the “collection
type” field. The polygons were constructed according to a set of steps created by
the Carolina Population Center. The steps included creating a temporary line
coverage that held the distance and bearing information collected at the point.
Once the lines representing the distance and bearing information were
constructed around each point, a polygon coverage was created by heads up
digitizing each feature using the endpoints from the line coverage. This process
was greatly aided by the snapping options in ArcMap. The name of the point
with the polygon information was then added as a field in the attributes of the
coverage. The attribute data were joined to the polygon coverage using the

same method discussed earlier.

fc VALIDATION

The first step in validating fc was to remove the points from the ground
truth data where fc information was not collected (e.g. road control points). Next,
to compare the fc values calculated from the image to the ground truth data, the

image fc layer was converted to a point coverage using ArcToolbox. This

50

conversion creates a point feature at the center of each pixel in the image
containing the image value In the attributes of the corresponding point. A spatial
join was then performed between the ground truth points and the image points.
A point to point spatial join finds the point nearest the point being joined to
(ground truth) and assigns the attribute value (image fc) to that point. The
resulting point coverage contained the ground truth and image fc values in its
attributes. The attributes were then exported to an Excel spreadsheet and a
linear regression was modeled on the data using the image fc value as the
dependent variable and the ground truth fc as the independent variable (further
discussed in Chapter 4). However, after noticing that many of the ground truth
points did not fall directly in the center of a pixel, the decision was made to
include the image fc values of neighboring pixels in an attempt to model the
image is value away from the center of a pixel. Because the fc value recorded
on the ground was a function of the vegetation surrounding the point (in my
theoretical pixel), I needed to account for when my theoretical pixel bounds did
not match the true pixel bounds in the image. A buffer of 21.5 meters (just over
half the distance from the pixel center to the pixel comer, see Figure 3.9) was
constructed around each of the ground truth points. This distance was chosen
considering that if a ground truth point fell directly in the comer of four pixels,
then the fa value of all four pixels would be represented as the image fc value at
that point. By including the values of the adjacent pixels, l was able to construct
an image fc value that was a better representation of what was on the ground at

points away from the center of a pixel.

51

 

 

 

® pixel boundariesL,
//"'——‘\,\\\
fc - \
prer 1 ‘fcpixel 2

\\\\

/ \

\
/ l
, Asample point I,

 

fc pixel 3 y’éixel 4
\‘

21 .5m buffer
around sample point

/P_\\\

 

 

 

 

 

 

 

 

/ \
/ “‘\
\‘.
\
bsample pornl i
O O / 0
\ //
\\\\ //
\
\V//
0 O O

 

 

 

Figure 3.9 Sample point buffer where (A) the sample point falls in the center of
four pixels and (B) falls near the center of one pixel

52

However, if the sample point fell directly on or very near the center of a pixel, no
other pixel values would be represented. The points generated from the image
fc values were then spatially joined to the buffer polygon layer. A point to
polygon spatial join finds the points that fall within the polygon and assigns the
attributes of the points to the polygon. The attributes were then exported to an
Excel spreadsheet. For each of the polygons, the mean of the image fc values
contained within the polygon was calculated. A linear regression was modeled
using the mean image fc value as the dependent variable and the ground truth fc

value as the independent variable.

LULC CLASSIFICATION

The LULC classification, performed on the 9/9/2001 image, was a hybrid
method, combining an unsupervised classification, evaluation of classes, and a
supervised classification. A very similar LULC classification was performed in
this region by Messina and Walsh (2001). The image was chosen for
classification in an attempt to identify classes previous to the fumigation efforts of
2002. The ISODATA unsupervised classification was initialized from the
statistics of the image along the diagonal axis and had 255 potential classes.
The significance of 255 classes is to allow the maximum number of classes to be
generated while keeping the resulting dataset in an 8-bit data structure (which
keeps processing time and file size low). The convergence threshold of the
classification was set at 0.98 or 30 iterations before completion. Each image

reached the desired threshold level before the maximum number of iterations.

53

The signatures gathered from this classification were then compared
against each other in an effort to isolate the spectrally significant classes. This
comparison included the evaluation of the transformed divergence matrix
composed of the best average values between classes. The threshold for
keeping classes was a transformed divergence value of >1960. This threshold
value (1960) was chosen after considering the complexity of the landscape and
in an effort to keep the overlap between classes very limited. Each signature
was compared against all of the other signatures in the set. This method
included evaluating an array that contained the transformed divergence values
between signatures and eliminating those with that shared low values with many
other signatures. The end product of the evaluation was a set containing 90
spectrally significant signatures. The last step in the classification was a
maximum likelihood supervised classification performed using the spectrally
significant signatures. This created a classified layer with the value for each pixel
representing the signature (of the 90 spectrally significant signatures) that the
pixel was assigned to. This classified layer was then attributed to a very simple
LULC scheme (see Table 3.7) via a visual interpretation using colors, textures,
and location on the image and a knowledge of the research area (gained during
field work) (Messina and Walsh 2001). The Coca class was identified using the
visual interpretation and also by isolating agricultural areas that only appeared
north of the Rio San Miguel (dividing Colombia and Ecuador) and comprised a

small percentage of the overall image area.

54

 

 

 

 

 

 

 

 

 

 

 

 

LULC classes # of classes % of image area

lPrimary forest 3 38
Secondary forest 1 17
[Pasture with trees or Rastrojo 2 13
Agriculture or pasture with no trees 5 13
[P_ossible coca 5 6
IOther (includes urban, water, and other LULCs) 74 13

totals: 90 100

 

Table 3.7 LULC classification scheme

An accuracy assessment was not performed on the classification as 1) I did not
have ground truth data for the coca classes (see previously mentioned dangers
of conducting research in Colombia or near its border), 2) the confidence of the
classification was high considering the simple scheme and prior knowledge of the
area, and 3) the desired results did not call for an accuracy assessment (with

more emphasis placed on the number of classes affected).

CHANGE DETECTION

For change detection, a subset of the affected area was created via an
AOl in Imagine. The AOI was digitized (heads-up) and represents the area of the
image north of the Ecuador/Colombia border that appeared to be an agricultural
region. The change detection was accomplished by overlaying the fc layers,
subtracting the values of the earlier dated layer from the more recent layer,

recoding the data to only show the pixels that had lost 10% or more of their fc.

55

then subsetting the newly created layer using the AOI of the affected area. The
changes occurring between 9/9/2001 and 9/12/2002 were labeled as Event 1 and
the changes occurring between 9/12/2002 and 10/14/2002 were labeled as Event
2. The LULC classification layer was then overlayed with the results of Event 2

to identify the classes with an fc loss of more than 10%.

56

Chapter 4 : RESULTS AND DISCUSSION

fc VALIDATION RESULTS

The image fc values and the ground-truth fc values were compared using
a linear regression with the ground-truth fc values as the independent variable
and the image fc values the dependent variable. A linear regression was
chosen, as the expected results were to find a 1:1 relationship between the two
variables. The results of the original regression are found in Figure 4.1A and
Table 4.1. This regression was based on the point to point overtay where the
image to value was compared only against the ground-truth fc value of the pixel
that it fell in. The results of the regression of the mean of the image fc values
within 21.5 meters (see Figure 3.9) against the ground-truth value are found in
Figure 4.1B and Table 4.2. As the results show, these regressions both are
statistically significant. The regressions were also performed after taking out the
large outliers in the data. Given that the Northern Oriente is a very dynamic
landscape and that there was a 9-month lag between the date of image
acquisition and field data collection, I observed changes that had occurred
throughout the landscape. These changes included agricultural plots cleared for
less than 3 months and a farm, for example, where a large portion of secondary

forest had been cut down and cleared the day prior to our visit.

57

 

 

 

 

20 r e .f-‘x" . . Y = 0.690X + 15.586 .
(r2 = 0.643)

re...“ . ............ 11 line
.- l I I I I I I I

0 10 20 30 4o 50 60 7o 80 90 100
Ground truth fc value

 

 

 

 

Image fc value
01
O
i

 

 

4o-
30-
20 '

. — Y=0.691X+ 16.504
10_ I ‘E ' (r2=0.672)

 

 

03... i I I l

I I I I l I
o 10 20 30 4o 50 60 7o 80 90 100
Ground truth fc value

 

 

 

Figure 4.1 Regression results using (A) the point to point overlay and (B) the
point to mean of points within 21 .5m overlay

58

 

 

 

 

 

 

 

 

 

 

Observations 145

R Square 0.6426]
F 249.936
Si ificance F 7.4E-33
t Stat (intercept) 5.62976
t Stat (slope) 15.8094
P—value (intercept) 9.6E-08
P-value (slope) 7.4E-33

 

Table 4.1 Regression results for point to point overlay with the full dataset

 

 

 

 

 

 

 

 

 

 

Observations 145

R Square 0.67173
F 292.621
Significance F 2.1E-36
t Stat (intercept) 6.46025
t Stat (slope) 17.1062
P-value (intercept) 1.5E-09
P-value (slope) 2.1E-36

 

 

 

Table 4.2 Regression results for point to mean of points overlay with the full
dataset

At this point, the image to value was 70% and my field estimation, only
accounting for the remaining understory vegetation, was 15% (see Table 4.3 for
full list). I chose to remove 10% of the data points based on the LULC
information from my field forms and the evaluation of the most erroneous outliers
in the dataset. A total of 14 points were removed from each dataset and the
regressions were completed again. This number of points to remove was chosen
in an attempt to keep a high percentage of the data points (>90%), yet also

account for the changes that may have occurred to the landscape due to

anthropogenic or natural causes. The sample points that were removed were

59

chosen by selecting those with the largest difference between image fc value

and ground truth fc value and deleting them from the dataset.

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

ground fc image fc difference

1 15 55 40

2 15 70 55

3 20 52 32

4 30 69 39

5 40 79.5 39.5
6 45 93 48

7 50 82.333 32.333
8 65 36 29

9 80 35 45
10 85 52 33

1 1 85 49 36
12 90 44.5 45.5
13 90 32.5 57.5
14 95 49.667 45.333

 

Table 4.3 Removed sample points

 

The regression results can be found Figures 4.2A and 4.2B and Tables 4.4 and
4.5 and show very improved results. The linear regressions of the image to data
and the ground truth to data show a promising trend. Although none of the
relationships proved to be a perfect 1:1, there are factors that may have
contributed to this. The first factor may be the method of calculating the fc value
as a linear function of the NDVI value of the image. Other vegetation indices
such a SAVI and MSAVI are possibly better models of the biophysical properties
of the vegetation as they attempt to eliminate the effects of varying soil
conditions, however the elimination of soil effects should be inherent in the

calculation of fc.

60

 

 

 

 

 

 

40-
30 -
.3" _
20- -.- — Y2— 0.795x + 11.026
.3 . . (r =o.825)
1o — 8 1:1 line
0 j . I

I I I I I I I I
o 10 20 30 4o 50 60 7o 80 90 100
Ground truth fc value

 

O) \I
O O
I I

Image fc value
.6 01
O O
L l
O

 

20 _, — Y - 0.794x + 10.705
-. 8 g . (r2 = 0.849)
10 ‘ ‘13,»! O ............ 1:1 line

0 3.
o

 

 

 

l l I I I I l I I
10 20 30 4o 50 60 7o 80 90 100
Ground truth fc value

 

 

 

Figure 4.2 Regression results minus outliers in the data using (A) the point to
point overlay and (B) the point to mean of points within 21 .5m overlay

61

 

 

 

 

 

 

 

 

 

 

Observations 131

R Square 0.82472
F 606.969
Significance F 1.3E-50
t Stat (intercept) 5.46723
t Stat (slope) 24.6367
P-value (intercept) 2.3E-07
P-value (slope) 1.3E-50

 

Table 4.4 Regression results for point to point overlay without the outliers in the

 

 

 

 

 

 

 

 

 

dataset

Observations 131

R Square 0.8489

F 724.741
Si ificance F 8.8E-55
t Stat (intercept) 5.75114
t Stat (slope) 26.921

P-value (intercept) 6.1E-08
P-value (slope) 8.8E-55

 

 

 

 

Table 4.5 Regression results for point to mean of points overlay without the
outliers in the dataset

Subtracting the NDVIsoi. value in each pixel should eliminate the amount of
NDVI gained from bare soil in each pixel. Other studies have looked at the
possibility of estimating fc (referred to as VF [vegetation fraction] in the study)
from information in narrow visible bands instead of using NDVI values (Gitelson
et al. 2002). Although the results were promising, this study was conducted only
for wheat and corn over very small areas. Leprieur et al. (2000) and Purevdorj

(1998) have suggested that the relationship between NDVI and fa may be non-

linear when vegetation cover is sparse.

62

Another factor causing the non-perfect relationship may be the visual
estimation as a method of collection of ground-truth data. As Figures 4.1 and 4.2
show, the regression matches a variable with a continuous scale against a
variable that borders between continuous and categorical. Although the ground
truth data may appear categorical, it is really a continuous variable with a lesser
precision in collection. This problem was considered in the genesis of creating a
method for collecting ground-truth fc data, however no solution was apparent in
the literature. Zha et al. (2003) attempted to relate fc to NDVI in a semi-arid
grassland in China, but only collected ground fc measurements in 1m2 areas to
compare to the 30m2 Landsat pixel. Other studies used color photography taken
at each sample site to estimate the fa of an area (Purevdorj et al. 1998). This
“void” of a proven method of collection for fc data in non-canopied areas cannot
be filled without studies such as this, which attempt to discover a new method.
However, these results verify my hypothesis that a linear relationship exists

between image fc and ground to used in this study.

fc AND LULC CHANGE DETECTION

Considering the nature of the observed relationship between to calculated
from the image and the to observed on the ground and the likelihood for natural
changes on the landscape, the change detection was only considered for pixels
that lost more than 10% of their fractional coverage between images. 10%
change was chosen considering non-perfect (not 1:1) relationship between

image to and ground fc values and in an attempt to not include smaller

63

defoliation events that were not due to Plan Colombia fumigation. Zha et al.
(2003) showed that fc was 89% accurate in when comparing ground truth data
and image results grouped in 10% increments (the same increments used for the
sampling scheme in this study, see Table 3.5).

We found that 56,627 ha of 10 % or greater defoliation occurred during
Event 1 (9/9/2001 to 9/12/2002) and 49,551 ha during Event 2 (9/12/2002 to
10/14/2002) of the southwestern Putumayo coca eradication effort. The total area
experiencing 10% or greater defoliation during 2002 coca eradication effort in the
region was 106,178 ha. The results of the change detection in Events 1 and 2
are found in Figures 4.3 and 4.4. The distinguishing characteristic of the spatial
organization of defoliation in both events are the linear patterns that emerge.
Patterns such as these are not found in a naturally occurring environment and
the linearity of the patterns are believed to be the result of the flight lines taken by
the aircraft while spraying. Another observation that shows that the defoliation is
due to spraying is the scale at which it is occurring. As discussed earlier, the
farmers in this area live on very small plots. To have all of them clear their areas
at the same time would require a very large and very unrealistic organized effort.
The last distinguishing spatial characteristic is the temporal variability that exists
between the events. A very noticeable trend is that the defoliated areas of Event
2 are shifted away from the areas in Event 1. This shift in the defoliated areas of
Event 2 is noticeable as the linear patterns of defoliation still generally run north

to south, but are moved to the west of the linear patterns of Event 1.

64

 

COLOMBIA

ECUADOR

Defoliation Event 1 7157m‘

COLOMBIA

ECUADOR

 

I I ‘ I T I

10 15 20 25 30 35 40 50 100
% fc change (lost)

 

 

 

Figure 4.3 Maps for defoliation Event 1 and Event 2 (notice extent box for
Figure 4.4)

65

 

 

I I T I I I

10 15 20 25 30 35 40 50 100
% fc change (lost)

 

 

 

Figure 4.4 Close up of defoliation near the Ecuador-Colombia border in Event 2
(notice where identified in Figure 4.3)

This would signal that the defoliation occurring was planned in advance
and the result of fumigation, not due to natural causes or farmers clearing their
land.

According to the UNODC report, coca production over all of Putumayo
.was reduced by 71 ,891 ha versus our measured defoliation of 106,178 ha (which

does not cover all of Putumayo), an unexplained difference of 34,287 ha. I

believe that this unexplained difference is due to drift and runoff of the herbicide

66

and also (discussed in the next paragraph), due to non-coca LULCs being
excluded from the defoliation figures.

Discriminate spraying of coca would limit the total number of spectral
classes heavily impacted by spraying. Further, the areas of greatest magnitude
change in fractional cover would theoretically contain significantly fewer spectral
classes. The defoliation images for each event were broken into defoliation
classes to compare the number of classes represented as the intensity of
defoliation became greater. Surprisingly, in each defoliation class, the numbers

of classes represented are quite similar (see Table 4.6).

 

 

 

 

 

 

 

 

 

% fc lost Classes represented (Event 1) Classes represented (Event 2)
10-14 49 47
15-19 47 49
20-24 47 47
25-29 45 46
30-39 45 46
40-49 43 44
50-100 38 42

 

 

 

 

Table 4.6 Defoliation classes and LULC classes represented in each

Also for Events 1 and 2, the three dominant spectral classes represent 24% and
36% of the defoliated landscape, respectively, with the top 10 classes
representing 60% and 69%, respectively. Even assuming that the top 10 most
impacted classes capture most forms of coca production, this still leaves 40%
(Event 1) and 31% (Event 2) of the defoliation occurring in non-coca or coca
intercropped lands. For the 50% and above defoliation class, 57% of the total

spectral classes in event 1 and 63% in event 2, including crops and other non-

67

coca producing lands captured in both events, were represented (see Appendix
E for the expanded results).

While a trend exists with fewer classes being defoliated, it does not
represent the expected dramatic reduction one would expect. The complex
spatial organization of the Colombian coca producing landscape appears to

prevent discriminate spraying of defoliants.

CONCLUSIONS

This study used remote sensing technology and fieldwork to quantitatively
and qualitatively measure the defoliation of the vegetation due to the Plan
Colombia fumigation in Putumayo, Colombia. The objectives of the study were
mmbw

1. Identifying to as a biophysical variable that would sufficiently characterize
the effects of defoliation due to Plan Colombia.

2. Validating the use of fc in this study by collecting ground truth data and
using that data to show the linear relationship that exists between image
fc and ground fc.

3. Quantifying the amount of Plan Colombia defoliation using the field-
calibrated fc data.

4. Identifying the number of LULC classes affected by fumigation and
showing that the Plan Colombia spraying during 2002 was not

discriminate over the landscape.

68

I hypothesized that:
1. Green fractional coverage derived from ETM+ data accurately quantifies
vegetation cover.
2. Hybrid classification techniques applied to ETM+ data identify LULC
classes.
3. Changes in fc indicate changes due to defoliation.
4. Defoliation in Putumayo, Colombia occurs indiscriminately over the
landscape affecting cover types other than coca.
5. Landsat ETM+ data provides the synoptic coverage with adequate
resolution characteristics suitable for study.
My hypotheses were verified by 1) validating the linear relationship that exists
between image fc and ground truth fc. 2) performing a LULC classification using
the hybrid method resulting in 90 spectrally significant classes, 3) revealing the
linear spatial patterns of defoliation using an fc change detection, 4) using the
number of LULC classes represented in each defoliation class to show that a
dramatic reduction of classes does not exist and the most conservative estimates
place over 30% of defoliation in non-coca LULC classes, and 5) performing the

aforementioned analysis using Landsat ETM+ data.

FUTURE RESEARCH

Although this study identified effects of aerial fumigation on the landscape
in Putumayo, there exist many possibilities for improvement and also extension

of the research. One area that I would like to explore further is the algorithm

69

used to calculate fc. I believe that revisiting the statistical relationship between
NDVI (or any other of the previously mentioned vegetation indices) and fc could
lead to image to values that match ground fc values with a 1:1 relationship.
Another possibility is to refine the method used for ground tmth fc calculation.
The use of aerial photos and/or high spatial resolution satellite imagery to
calculate to added to the field estimation improvements could create a high level
of confidence in using fc derived from images of more coarse spatial resolution
such as Landsat ETM+.

I would also like to verify the LULC classes that were created using the
hybrid method. Although an accuracy assessment was not necessary for the
results of this study, the accuracy information would greatly enhance the
understanding of the LULC classes being fumigated and possibly help to explain
the effects of wind drift and runoff of the herbicide used in the fumigation.
Although the chances to find coca in Ecuador are limited, the possibility of
verifying the other LULC classes could illuminate, through a process of
elimination, the coca classes.

In future research, I would like to explore changes that have occurred on
Colombia’s landscape after fumigation. As was mentioned previously, many
believe that fumigation could lead to deforestation. I would like to conduct a
LULC change detection of the area in an attempt to illuminate possible rising
deforestation rates in Putumayo.

I also believe that I should extend this research to northern Ecuador.

Some of the possibilities are: showing whether Plan Colombia fumigation has

70

affected vegetation in Ecuador (near the border), calculating deforestation rates
near the border, locating coca cultivation in Ecuador, illuminating changes of the
spatial organization of the landscape in Ecuador due to the influx of Colombians,
and modeling the possible effects of Plan Colombia (the migration of Colombians

into Ecuador) using spatial simulation models.

71

LITERATURE CITED

Alvarez, M. D. 2003. Forests in the Time of Violence: Conservation Implications of the
Colombian War. Journal of Sustainable Forestry 16 (3/4):49-70.

Bagley, B. M. 19883. The New Hundred Years War? Us National Security and the War
on Drugs in Latin America. Journal of Interamerican Studies and World Aflairs
3O (1):l61-182.

Bagley, B. M. 1988b. US Foreign Policy and the War on Drugs: Analysis of a Policy
Failure. Journal of Interamerican Studies and World Aflairs 3O (2/3):189-212.

Berger, R. M. 2001. The time of coca: on the Colombian front of the drug war, it's hard to
tell who--or what--is the real enemy. Sojourners 30 (3):20-24.

Bibes, P. 2001. Transnational Organized Crime and Terrorism. Journal of Contemporary
Criminal Justice 17 (3):243-258.

Browitt, J. 2001. Capital punishment: the fragmentation of Colombia and the crisis of the
nation-state. Third World Quarterly 22 (6):1063-1078.

Brown, J. 2000. Doubt about US antidrug effort in Colombia. Christian Science Monitor,
July 3 2000, 3.

Carlson, T. N., and D. A. Ripley. 1997. On the Relation between NDVI, Fractional
Vegetation Cover, and Leaf Area Index. Remote Sensing of Environment 62:241-
252.

Choudhury, B. J ., N. U. Ahmed, S. B. Idso, R. J. Reginato, and C. S. T. Daughtry. 1994.
Relations between evaporation coefficients and vegetation indices studied by
model simulations. Remote Sensing of Environment 50: 1-17.

Clawson, P. L., and R. W. Lee. 1996. The Andean Cocaine Industry. New York: St.
Martin's.

Cohen, W. B., T. K. Maiersperger, S. T. Gower, and D. P. Turner. 2003. An improved
strategy for regression of biophysical variables and Landsat ETM+ data. Remote
Sensing of the Environment 84:561-571.

Congalton, R. G., and K. Green. 1999. Assessing the Accuracy of Remotely Sensed Data:
Principles and Practices. Boca Raton: Lewis Publishers.

Cooper, M. 2001. Plan Colombia: wrong issue, wrong enemy, wrong coutry. The Nation
272 (11):11-18.

72

Crandall, R. 2002. Driven by Drugs: U. S. Policy towards Colombia. Boulder: Lynne
Reinner Publishers.

Dishman, C. 2001. Terrorism, Crime, and Transformation. Studies in Conflict &
Terrorism (24):43-58.

DNE. 2000. Existencias historicas do 103 cultivos ilicitos y su erradicacion. Bogota:
Departamento Nacioal de Estupefacientes.

Driver, T. F. 2001. Colombia's war: drugs, oil and markets. The Christian Century,
November 7 2001, 14-19.

Dudley, S. 2000. The Colombia quagmire: the drug war goes south. American Prospect
11 (17):32-37.

Ecuador, E. o. 2001. Ecuador in Brief. Washington, DC: Embassy of Ecuador.
Ehrenfeld, R. 1990. Narco-terrorism. New York: Basic Books.
Ernst, J. 2002. Plan Colombia. SAIS Review 22 (1):97-102.

F amam, A. 2002. Colombia's civil war drifts south into Ecuador. Christian Science
Monitor, July 11 2002, 7.

Ford, J. T. 2003. Drug Contol: Coca Cultivation and Eradication Estimates in Colombia,
a letter to Charles H. Taylor, House of Representatives. Washington DC: US.
General Accounting Office.

F orero, J. 2002. US. is Stepping Up Drive to Destroy Coca in Colombia. The New York
Times, September 4 2002, A1, A6.

F ratepietro, S. 2001. Plan Colombia: the hidden float in the US. drug war. The Humanist
61 (4):]8-21.

Frechette, M. 2003. In Search of the Endgame: A Long-Tenn Multilateral Strategy for
Colombia. Coral Gables, FL: North-South Center and Strategic Studies Institute
(Joint Publication).

Fung, T., and W. Siu. 2000. Environmental quality and its changes, an analysis using
NDVI. International Journal of Remote Sensing 21 (5): 101 1-1024.

Gardner, D. 2001. The Failed War on Drugs. Chicago Sun-Times, January 21, 42.

Gardner, R. H., B. T. Milne, R. V. O'Neill, and M. G. Turner. 1987. Neutral models for
the analysis of broad-scale landscape patterns. Landscape Ecology 1:19-28.

Gilabert, M. A., J. Gonzalez-Piqueras, F. J. Garcia-Haro, and J. Melia. 2002. A
generalized soil-adjusted vegetation index. Remote Sensing of Environment
82:303-310.

73

Gillies, R. R., and T. N. Carlson. 1995. Thermal remote sensing of surface soil water
content with partial vegetation cover for incorporation into climate models.
Journal of Applied Meteorology 34:745-756.

Gitelson, A. A., Y. J. Kaufman, R. Stark, and D. Rundquist. 2002. Novel algorithms for
remote estimation of vegetation fi'action. Remote Sensing of Environment 80:76-
87.

Guzman, G., O. Fals-Borda, and E. Umana-Luna. 1962. La Violencia en Colombia:
estudio de un proceso social. 2nd ed. Bogota, Colombia: Ediciones Tercer
Mundo. '

Hodgson, M. 2001. Side effects hit Plan Colombia. Christian Science Monitor, February
27 2001, 6.

Huete, A. R. 1988. A soil-adjusted vegetation index (SAVI). Remote Sensing of
Environment 25:295-309.

lsacson, A., and I. Vaicius. 2001. Plan Colombia’s "Ground Zero". Washington DC:
Center for International Policy.

Kalmanovitz, S. 1988. Economia y nacion: una breve historia de Colombia. Bogota,
Colombia: Siglo XXI Editores.

Kay, B. H. 1999. Violent Opportunities: The Rise and Fall of "King Coca" and Shining
Path. Journal of Interamerican Studies and World Aflairs 41 (3):97-127.

Keitt, T. H. 2000. Spectral representation of neutral landscapes. Landscape Ecology
15:479-493.

Kratz, V. 2002. Lethal Dose. Sierra Magazine, J uly/August 2002.
LaFranchi, H. 2000. What's at stake? Christian Science Monitor, December 12 2000, 10.

LaFranchi, H. 2001. A drug war down on the farm. Christian Science Monitor, May 23
2001, 6.

Leprieur, C., Y. H. Kerr, S. Mastorchio, and J. C. Meunier. 2000. Monitoring vegetation
cover across semi-arid regions: comparison of remote observations from various
scales. International Journal of Remote Sensing 21 (2):281-300.

Lunetta, R. S., J. Ediriwickrema, D. M. Johnson, J. G. Lyon, and A. McKerrow. 2002.
Impacts of vegetation dynamics on the identification of land-cover change in a
biologically complex community in North Carolina, USA. Remote Sensing of
Environment 82:258-270.

Mabry, D. 1988. The US Military and the War on Drugs in Latin America. Journal of
Interamerican Studies and World Aflairs 3O (2/3):53-76.

74

Mackay, G., M. D. Steven, and J. A. Clark. 1998. An atmospheric correction procedure
for the ATSR-2 visible and near-infrared land surface data. International Journal
of Remote Sensing 19 (15):2949-2968.

Marcella, G. 2002. The US. Engagement with Colombia: Legitimate State Authority and
Human Rights. Coral Gables, FL: North-South Center.

Meltzer, J. 2001. The Enduring Colombian Conflict: A Canadian Perspective. Ottawa,
Ontario: Canadian Foundation for the Americas.

Messina, J. P., and S. J. Walsh. 2001. 2.5D Morphogenesis: modeling landuse and
landcover dynamics in the Ecuadorian Amazon. Plant Ecology 156:75-88.

Millett, R. L. 1997. The United States and Latin America's Armed Forces: A Troubled
Relationship. Journal of Interamerican Studies and World Aflairs 39 (1):121-136.

Moreno-Sanchez, R., D. S. Kraybill, and S. R. Thompson. 2003. An Econometric
Analysis of Coca Eradication Policy in Colombia. World Development 31
(2):375-383.

Nagle, L. E. 2002. Plan Colombia: Reality of the Colombian Crisis and Implication for
Hemispheric Security. Carlisle, PA: Strategic Studies Institue.

NASA. Landsat 7 Science Data Users Handbook.

North, P. R. J. 2002. Estimation of fapar, LAI, and vegetation fractional cover from
ATSR-2 imagery. Remote Sensing of the Environment 80:114-121.

Penhaul, K. 2001. Making a deal with the devil. US. News & World Report, April 16
2001, 28-29.

Perreault, T. 2003. Making Space: Community Organization, Agrarian Change, and the
Politics of Scale in the Ecuadorian Amazon. Latin American Perspectives 30
(128):96-121.

Peterson, S. 2002. People and Ecosystems in Colombia: Casualties of the Drug War. The
Independent Review 6 (3):427-440.

Petras, J. 2001. The geopolitics of Plan Colombia. Monthly Review 53 (1 ):30-48.

Price, J. C., and W. C. Bausch. 1995. Leaf Area Index Estimation fiom Visible and Near-
Infrared Reflectance Data. Remote Sensing of the Environment 52:55-65.

Purevdorj, T., R. Tateishi, T. Ishiyama, and Y. Honda. 1998. Relationships between
percent vegetation cover and vegetation indices. International Journal of Remote
Sensing 19 (18):3519-3535.

75

Qi, J ., A. Chehbouni, A. R. Huete, Y. H. Kerr, and S. Sorooshian. 1994. A modified soil
adjusted vegetation index (MSAVI). Remote Sensing of Environment 48:119-126.

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 and Forest
Meteorology 105:55-68.

Rojas, I. 2003. The push for zero coca: Democratic transition and countemarcotics policy
in Peru. Washington, DC: Washington Office on Latin America.

Rouse, J. W., R. H. Haas, J. A. Schell, D. W. Deering, and J. C. Harlan. 1974. Monitoring
the vernal advancement of retrogradation of natural vegetation. Greenbelt, MD:
NASA/GSFC (Type HI, Final Report).

Scanlon, T. M., J. D. Albertson, K. K. Caylor, and C. A. Williams. 2002. Determining
land surface fractional cover from NDVI and rainfall time series for a savanna
ecosystem. Remote Sensing of the Environment 82:376-3 88.

Sharpe, K. E. 1988. The Drug War: Going After Supply: A Commentary. Journal of
Interamerican Studies and World Affairs 30 (2/3):77-85.

Song, C., C. E. Woodcock, K. C. Seto, M. P. Lenney, and S. A. Macomber. 2001.
Classification and Change Detection Using Landsat TM Data: When and How to
Correct Atmospheric Effects? Remote Sensing of the Environment 75:230-244.

Stauder, M. 2001. Colombia's cocaine frontier. Mother Jones 26 (6):58-63.

Strahler, A. H., and A. N. Strahler. 1992. Modern Physical Geograpy. 4th ed. New York:
John Wiley & Sons, Inc.

Tanre, D., C. Deroo, P. Duhaut, M. Herman, J. J. Morcrette, J. Perbos, and P. Y.
Deschamps. 1990. Description of a computer code to simulate the satellite signal

in the solar spectrum: the SS code. International Journal of Remote Sensing 11
(4):659-668.

Teillet, P. M., K. Staenz, and D. J. Williams. 1997. Effects of Spectral, Spatial, and
Radiometric Characteristics on Remote Sensing Vegetation Indices of Ferested
Regions. Remote Sensing of the Environment 61 :139-149.

Thoumi, F. E. 1992. Why the Illegal Psychoactive Drugs Industry Grew in Colombia.
Journal of Interamerican Studies and World Aflairs 34 (3):37-63.

Thoumi, F. E. 2002. Illegal Drugs in Colombia: From Illegal Economic Boom to Social
Crisis. Annals of the American Academy of Political and Social Science 582: 102-
1 16.

76

UNDCP. 2002. Colombia: Annual Coca Cultivation Survey 2001: United Nations Drug
Control Program.

UNODC. 2003a. Colombia: Coca Survey for 2002: United Nations Office on Drugs and
Crime.

UNODC. 2003b. Colombia: Coca Survey for December 2002 and Semi-Annual Estimate
for July 2003: United Nations Office on Drugs and Crime.

Uribe, S. 1997. Evaluacion: Extension, tecnicas y tecnologias para la produccion y
rendimientos y magnitud de la industria. In Drogas ilicitas en Colombia: Su
impacto economico, politico y social, edited by F. E. Thoumi. Bogota, Colombia:
Editrial Planeta.

United States Senate. 2003. Caucus on Intenational Narcotics Control. US Policy
Regarding Narcotics Control In Colombia. lst. June 3, 2003.

USDS. 2001. International Narcotics Control Strategy Report. Washington DC: United
States Department of State.

USDS. 2002. Report on Issues Related to the Aerial Eradication of Illicit Coca in
Colombia. Washington DC: United States Department of State.

USGAO. 2003. Drug Control: Specific Performance Measures and Long-Term Costs for
US. Programs in Colombia Have Not Been Developed. Washington DC: United
States General Accounting Office.

Vaicius, I., and A. lsacson. 2003. The "War on Drugs" meets the "War on Terror".
Washington DC: Center for International Policy.

Valenzuela, P. 2002. Conflict Analysis- Colombia, Bolivia and the Andean Region.
Stockholm: Swedish International Development Cooperation Agency.

Van Der Meer, F. 1999. Iterative spectral urnnixing (ISU). International Journal of
Remote Sensing 20 (17):3431-3436.

Vargas, R. 2002. The anti-drug policy, aerial spraying of illicit crops and their social,
environmental and political impacts in Colombia. Journal of Drug Issues 32
( 1): 1 1 -60.

Vellinga, M. 2000. The War On Drugs and Why It Cannot Be Won. Journal of Third
World Studies 17 (2):113-128.

Weiss, E., S. E. Marsh, and E. S. Pfirman. 2001. Application of NOAA_AVHRR NDVI
time-series data to assess changes in Saudi Arabia's rangelands. International
Journal of Remote Sensing 22 (6):1005-1027.

77

Wilson, S. 2003. US. Seeks to Avoid Deeper Role in Colombia. Washington Post
Foreign Service, March 9, 2003, A16.

Xavier, A. C., and C. A. Vettorazzi. 2004. Mapping leaf area index through spectral
vegetation indices in a subtropical watershed. International Journal of Remote
Sensing 25 (9):1661-1672.

Zha, Y., J. Gao, S. Ni, Y. Liu, J. Jiang, and Y. Wei. 2003. A spectral reflectance-based
approach to quantification of grassland cover from Landsat TM imagery. Remote
Sensing of the Environment 87:371-375.

78

APPENDIX A

58 INPUT PARAMETERS

09/09/2001 Image ETM+

 

 

 

 

 

 

 

 

 

 

 

 

 

Site Elevation 500 meters
Sensor Elevation 705 kilometers
Solar Zenith 29.5321637
Solar Azimuth 79.5446957
Sensor Zenith 0
Sensor Azimuth 0
Model Type Predefined
Standard Atmos. Model Tropical
Aerosol Model Continental
Visibility 17 kilometers
Date 9/9/2001

 

 

09/12/2002 Image ETM+

 

 

 

 

 

 

 

 

 

 

 

 

 

Site Elevation 500 meters
Sensor Elevation 705 kilometers
Solar Zenith 29.3394215
Solar Azimuth 81.645768
Sensor Zenith 0
Sensor Azimuth 0
Model Type Predefined
Standard Atmos. Model Tropical
Aerosol Model Continental
Visibility 17 kilometers
Date 9/12/2002

 

 

79

10/14/2002 Image ETM+

 

 

 

 

 

 

 

 

 

 

 

 

 

Site Elevation 500 meters
Sensor Elevation 705 kilometers
Solar Zenith 27.727403
Solar Azimuth 107.930127
Sensor Zenith 0
Sensonr Azimuth 0
Model Type Predefined
Standard Atmos. Model Tropical
Aerosol Model Continental
Nisibility 17 kilometers
Date 10/14/2002

 

 

80

APPENDIX B

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

53 OUTPUT PARAMETERS
9/9/2001 Image
Reflectance Band 1 Band 2 Band 3 Band 4 Band 5 Band 7
0 0.0719 0.0395 0.025 0.0123 0.0026 0.0012
0.1 0.1492 0.1188 0.1089 0.0968 0.0893 0.0903
0.2 0.2292 0.1999 0.1944 0.1822 0.1762 0.1796
0.3 0.312 0.2831 0.2813 0.2684 0.2634 0.269
0.4 0.3977 0.3683 0.3697 0.3555 0.3508 0.3585
0.5 0.4866 0.4556 0.4597 0.4435 0.4384 0.4481
9/12/2002 Image
Reflectance Band 1 Band 2 Band 3 Band 4 Band 5 Band 7
0 0.0719 0.0395 0.025 0.0123 0.0026 0.0012
0.1 0.1492 0.1188 0.109 0.0968 0.0893 0.0903
0.2 0.2292 0.2 0.1944 0.1822 0.1762 0.1796
0.3 0.312 0.2831 0.2813 0.2684 0.2634 0.269
0.4 0.3978 0.3684 0.3698 0.3556 0.3508 0.3585
0.5 0.4867 0.4557 0.4598 0.4436 0.4385 0.4481
10/14/2002 Image
Reflectance Band 1 Band 2 Band 3 Band 4 Band 5 Band 7
0 0.0717 0.0394 0.0249 0.0123 0.0026 0.0012
0.1 0.1492 0.1 189 0.109 0.0969 0.0893 0.0904
0.2 0.2295 0.2003 0.1947 0.1824 0.1764 0.1797
0.3 0.3125 0.2836 0.2818 0.2688 0.2636 0.2692
0.4 0.3985 0.3691 0.3704 0.356 0.3511 0.3588
0.5 0.4877 0.4566 0.4606 0.4442 0.4388 0.4485

 

 

 

 

 

 

 

 

81

 

 

 

APPENDIX C

DESCRIPTION OF FIELD DA TA COLLECTION FORM

Site #2 The site # from the field maps
Date: The date that the information was recorded on

Travel time: The time traveled between the starting point and the current
collection point. Starting point should be noted

LULC type: The LULC at the collection point (see LULC sheet for guidelines)

Collection type: This is either a point, line, or polygon (using just GPS, or GPS
plus rangefinder)

General description: Any comments on general features at the collection point.
This can include vegetation, location, soils, man-made

11 pts in file: The number of points collected by the GPS unit at the collection
point

Rover file name: The name given to the rover file location by the GPS unit
Waypoint file name: The name given to each waypoint file by the GPS unit
Coordinate X: The X coordinate recorded by the GPS unit

Coordinate Y: The Y coordinate recorded by the GPS unit

Photo #: The photo number, if needed, at the collection point (may have
multiples)

Photo type: The name of camera that was used to take the photo (Pentax 105,
Pentax 4500)

Canopy height: The height of the canopy (where applicable) in meters
measured with the rangefinder

Polygon collection and Sketch area: The distance and azimuth of polygon
collection and sketches of the area

82

Visual fractional coverage information: The FC collection notes, what kind of

collection (photo, visual)

83

APPENDIX D

IDL CODE TO SIMULA TE CANOPYAND UNDERSTORY

;IDL program photo_fix.pro

;written by Paul Delamater for AAG ELWL and thesis
;creates neutral models to create one fc value when
;a fc photo and fc ground cover data are present

pro photo_fix
time = systime(1)

;define variables

fcphoto = [.83215, .808, .5681, .5612, .0973, .6587, $ .86415,
.1459, .7082, .7824, .7351, .87835]

fcground = [.75, .35, .3, .85, .71, .80, .40, .28, .80, $ .85,
.85, .80]

fcfinal = fltarr(12)

maxiter = 12

;open mask image
mask = read_tiff('c:\paul\research\thesis\mask.tif')

;get circle area
circlearea = n_e1ements(where(mask eq 1))

;start the loop
for i=0, maxiter-l do begin

;create neutral model 1
neutrall = randomu (seed, 1395, 1395)

;create photo model
photo_fc = neutrall 1e fcphoto(i)

;create neutral model 2
neutra12 = randomu (seed, 1395, 1395)

;create groung model
ground_fc = neutra12 1e fcground(i)

;create combined model
comb_fc = photo_fc + ground_fc
greencover = comb_fc gt 0

;mask out edges
greencover = greencover * mask

84

;calculate fractional coverage
greenpix = n_e1ements(where(greencover eq 1))
fcfinal(i) = (greenpix * 1.0) / (circlearea * 1.0)

endfor
print, fcfinal

;get the processing time... this code ran very fast
print, 'Total processing time =',(SYSTIME(1) - time)/60, '8

Minutes'

;all good things must come to an
end

85

APPENDIX E

EVENT 1.' % OF DEFOLIA TED AREA PER LULC CLASS

% of defoliated area

0.0027

0.0008

  

c‘D’

_w

0 ther (includes urban,

water, and other LULCs)

 
   

 

note: classes not represented have been removed

 

86

 

EVENT 2: % OF DEFOLIA TED AREA PER LULC CLASS

 

% of defoliated area

 

 

 

 

 

 

    

 

 

 

 

 

 

 

 

 

 

 

Class 7 0.0087

Class 13 0.0005

Class 16 0.0706

Classl7 0.0118 , ’ y . _ _ '
Class 18 0.0005 . -. " f .7 ,, ,1 .Q' .«
Class 19 0.0517 Other (includes urban, water, and other LULCs)
Class 20 0.0069

Class 21 0.3430 note: classes not represented have been removed
Class 22 0.0178

Class 23 0.2309

Class 24 0.0004

Class 25 0.6205

87

 

88

 

 

 

 

 

 

 

 

Class 83 2.2731
Class 84 1.6886
Class 85 0.5853
Class 86 1.6112
Class 87 0.4430
Class 88 0.5726
Class 89 0.2682

 

 

89

n. .31., ..................

l111111111311111111

 

 

m—A_