i r .2 i
3.... v5.3.3}!-
Eight:
fig
“WW-«mm.
«.2 at"
. 43
x
. 3. .
a?» r. 2.. “a...
nu» ,
.afl in».
19:?
am...» v.2...
. r a»...
..
...R.»R .,
uidfifit
. um.»
.
r
2.5.... 3. 1..
3 5.1..
. s... .. 2d a...
I» H ; trs; !
1. ii
{ling
31:3: 25.0.1.3
1.2.44.2!!! .54..
12:33.»
l...3.:.:{.
. 151.13.. .
n . .1...
.31 .24...
t
5‘3 .
D‘Itvf ¢f
. 1.1.1.. 1
3.. .
9. n . 41.44.4233.” :2 . .. . 4):.2»....1.: V [12‘
ifs... sfiafinfi. Agriéwaégig
..,, 4......3EJ.
,WJEEV .u.
..
".3
/ 4/; .. '
/’ ”WI-b
0cm ,
41‘?
LIg‘ATE UiIIVERSITY
CHIGAN S
EXIST LANSING, MICH 48824-1048
This is to certify that the
dissertation entitled
SPATIAL STRUCTURE AND INVASION SUCCESS OF
POPULATIONS UNDERGOING RANGE EXPANSION: EFFECTS
OF DISPERSAL STRATEGY AND LANDSCAPE HETEROGENEITY
presented by
Genevieve Marie Nesslage
has been accepted towards fulfillment
of the requirements for the
Ph.D. degree in Department of Fisheries and
Wildlife and Program in Ecology,
Evolutionary Biology and Behavior
MAW/vim»
Major Professor’s Signature
WW Xyflaé
fl /
Date
MSU is an Affirmative Action/Equal Opportunity Institution
_ - - -._.
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
JAN 1 81009
010509
‘——_————
H ‘
2/05 chlRC/DatoDueJndd-sz
“
SPATIAL STRUCTURE AND INVASION SUCCESS OF POPULATIONS
UNDERGOING RANGE EXPANSION: EFFECTS OF DISPERSAL STRATEGY AND
LANDSCAPE HETEROGENEITY
By
Genevieve Marie Nesslage
A DISSERTATION
Submitted to
Michigan State University
in partial fulfillment of the requirements
for the degree of
DOCTOR OF PHILOSOPHY
Department of Fisheries and Wildlife and
Program in Ecology, Evolutionary Biology and Behavior
2005
ABSTRACT
SPATIAL STRUCTURE AND INVASION SUCCESS OF POPULATIONS
UNDERGOING RANGE EXPANSION: EFFECTS OF DISPERSAL STRATEGY AND
LANDSCAPE HETEROGENEITY
By
Genevieve Marie Nesslage
This dissertation explores the effects of dispersal strategy and landscape
heterogeneity on the spatial structure and invasion success of populations undergoing
range expansion. First, the spatial structure of an invasive population was evaluated
using semivariogram analysis (a characterization of the dependence of data on spatial
location). Twelve years (1985—1996) of historic gypsy moth (Lymantria dispar)
monitoring records from the Lower Peninsula of Michigan were analyzed. Seven models
were statistically fit to annual plots of semivariance against distance at two different
spatial extents. The model with the lowest Akaike’s Information Criterion value was
chosen to characterize overall semivariogram behavior for each year and spatial extent.
At large spatial extents, three distinct patterns of semivariogram behavior were observed
as the invasion progressed. This three-stage semivariance progression pattern likely
represents the invasion stages of establishment, expansion, and stabilization and may be
indicators of important mechanistic changes in invasion dynamics.
Second, the relationship between landscape heterogeneity and invasion success of
the gypsy moth in Michigan was analyzed by regressing invasion success on seven
measures of landscape fragmentation and by identifying potential threshold responses
(sudden, nonlinear changes) in invasion success with habitat loss. Invasion success
thresholds were then compared with that of seven landscape structure metrics. Both
patch- and gap-based landscape metrics exhibited threshold responses. However,
invasion success increased linearly with increasing proportion of habitat, indicating that
the negative effects of habitat fragmentation do not compound that of habitat loss. The
absence of detectable thresholds in invasion success suggests that aerial dispersers like
the invasive gypsy moth may not be as strongly affected by habitat fragmentation as
previously thought.
Lastly, semivariogram analysis was used to evaluate patterns of spatial
autocorrelation generated by simple and stratified dispersal models. Dispersal was
simulated across a homogeneous landscape composed of 100% habitat and a real,
heterogeneous landscape composed of 54% habitat. Population growth rate and stratified
dispersal distributions were varied as well. Semivariograms were calculated from model
simulations at every other time step and fit with four different models to characterize
changes in spatial structure of the population over time. Both simple and stratified
dispersal modeled across real, heterogeneous landscapes generated semivariogram
progression patterns similar to that of gypsy moths. Models of dispersal across 100%
habitat landscapes and models of stratified dispersal with a long-tailed dispersal
distribution across real landscapes generated semivariance progression patterns dissimilar
to that of gypsy moths. Thus, landscape heterogeneity (and not dispersal strategy) was a
major driving force behind semivariance patterns observed during the gypsy moth
invasion of Michigan.
ACKNOWLEDGEMENTS
I sincerely thank all those who have helped me in the pursuit of my doctorate. In
particular, I would like to thank my advisor Dr. Brian Maurer for his encouragement of
my research and career. I have been privileged to work with one of the most outstanding
ecologists of our time. In addition, I owe the success of my dissertation to Dr. Stuart
Gage who provided me with both the data and the knowledge of dispersal ecology
necessary to conduct this research project. I am deeply indebted to Dr. Michael Jones
and Dr. Richard Kobe for their thoughtful and constructive reviews of my research ideas
and manuscripts. I also thank Dr. Michael Wilberg for providing me with endless
statistical advice and personal support over the last five years. His exceptional
intelligence and kindness have inspired me to be a better scientist and a better person.
I would like to thank all those who offered me advice, support, and career-
advancing Opportunities over the last five years: Michael Donovan of the Michigan
Department of Natural Resources; Drs. William Taylor, Thomas Coon, Shawn Riley,
Henry Campa, and Robert Batie in the Department of Fisheries and Wildlife; my lab
mates Jennifer Skillen, Katherine Kahl, Anne Axel, Robert Lanfear, Hui-Yu Wang, and
Smruti Damania; and my parents Lee and Arthur Nesslage.
Funding for this project was provided in part by The Graduate School and the
Ecology, Evolutionary Biology, and Behavior Program at Michigan State University.
iv
TABLE OF CONTENTS
LIST OF TABLES ............................................................................................................ vii
LIST OF FIGURES ......................................................................................................... viii
INTRODUCTION .............................................................................................................. 1
CHAPTER 1
CHARACTERIZING THE INVASION PROCESS USING PATTERNS OF
SEMIVARIANCE: A CASE STUDY OF THE GYPYSY MOTH IN MICHIGAN ........ 5
Introduction .................................................................................................................. 5
Methods ..................................................................................................................... 11
Results ........................................................................................................................ 15
Variogram behavior in Michigan’s Lower Peninsula .................................... 15
Variogram behavior at the ecoregion scale ................................................... 16
Discussion .................................................................................................................. 20
Variogram behavior in Michigan’s Lower Peninsula .................................... 20
Variogram behavior at ecoregion scale ......................................................... 21
Timing of shifts in variogram behavior ......................................................... 23
Use of variograms in invasion analysis and interpretation ............................ 25
CHAPTER 2
THE RELATIONSHIP BETWEEN INVASION SUCCESS AND LANDSCAPE
HETEROGENEITY: A CASE STUDY OF THE GYPSY MOTH IN MICHIGAN ...... 35
Introduction ................................................................................................................ 35
Methods ..................................................................................................................... 39
Invasion data .................................................................................................. 39
Land cover data .............................................................................................. 39
Scale of analysis ............................................................................................ 4O
Characterizing the relationship between invasion success and landscape
structure ......................................................................................................... 41
Results ........................................................................................................................ 43
Relationship between invasion success and habitat loss ............................... 43
Relationship between landscape metrics and habitat loss ............................. 43
Relationship between invasion success and habitat fragmentation ............... 45
Discussion .................................................................................................................. 46
Effects of habitat loss and fragmentation on invasion success ...................... 46
Effect of scale of analysis and land cover map .............................................. 49
Management implications .............................................................................. 51
CHAPTER 3
PATTERNS OF SEMIVARIANCE GENERATED BY SIMPLE AND STRATIFIED
DISPERSAL ..................................................................................................................... 68
Introduction ................................................................................................................ 68
Methods ..................................................................................................................... 70
Model structure .............................................................................................. 7O
Characterizing changes in semivariance ........................................................ 72
Results ........................................................................................................................ 74
Discussion .................................................................................................................. 77
CONCLUSIONS ............................................................................................................... 96
Detection of emergent properties ............................................................................... 96
Semivariance analysis ................................................................................................ 98
Future directions ........................................................................................................ 98
LITERATURE CITED ................................................................................................... 100
vi
LIST OF TABLES
Table 1.1. Modeling results from variogram analysis of annual gypsy moth (Lymantria
dispar) trap catch in the Lower Peninsula (LP) and 10 ecoregions of Michigan,
1985-1996. The best model (model with lowest Akaike Information Criterion value)
of seven models fit to annual variograms of gypsy moth trap catch is listed as
“Type”. Note that the exponential (Expon.), Gaussian (Gaus.), and spherical
(Spher.) model names were abbreviated. Model shape (Shape) describes the overall
shape of the variogram being fit (B = bell or wave shaped, L = linear, and U =
untrended variogram with typical increase in semivariance to an asymptote). ........ 28
Table 2.1. Summary of critical thresholds in landscape metrics with habitat loss at three
scales of analysis (i.e. analysis window sizes of 75, 45, and 15 km on a side) using
two land cover maps. Dashes indicate no threshold response was detected. ........... 54
Table 3.1. Description of twelve model scenarios, including dispersal strategy modeled,
landscape type, population growth rate (r), and dispersal distributions used in model
simulations. Distribution 1 is a gamma distribution with rate parameter of 0.5 and
shape parameter of 1.2; Distribution 2 is a long-tailed gamma distribution with rate
parameter of 0.043 and Shape parameter of 0.5. ....................................................... 82
vii
LIST OF FIGURES
Figure 1.1. Map of Michigan’s Lower Peninsula showing its location in North America
and ecoregion boundaries used for variogram analysis of gypsy moth (Lymantria
dispar) catch between 1985 and 1996 ....................................................................... 30
Figure 1.2. Annual semivariograms of male gypsy moth (Lymantria dispar) trap catch in
the Lower Peninsula of Michigan, 1985—1996. Circles represent the original data
and Open squares represent predicted values generated by the best—fit models.
Predicted values were generated by either the cubic (1985—1988), wave (1989),
spherical (1990), power (1991—1995), or exponential (1996) models. Note the shift
in variogram shape from bell—shaped (1985—1989) to linear (1990—1995) to an
untrended pattern of increasing semivariance and leveling off at an asymptote. ..... 31
Figure 1.3. Total annual male gypsy moth (Lymantria dispar) catch in the Lower
Peninsula of Michigan between 1985 and 1996. Arrows indicate years in which
variogram patterns changed from bell—shaped to linear (right—facing arrow) and
from linear to typical increase to asymptote (left—facing arrow). Note that inspection
of the time series alone (without consideration of Spatial location of data points)
would indicate that the invasion had stabilized by 1990. ......................................... 33
Figure 1.4. Total annual male gypsy moth (Lymantria dispar) trap catch in each often
ecoregions of the Lower Peninsula of Michigan between 1985 and 1996. .............. 34
Figure 2.1. Range in years to colonization among traps located in the same analysis
window declined with increasing proportion of habitat. Results are reported using
Map 1 for each scale of analysis: a) 75 km, b) 45km, and c) 15km. ........................ 55
Figure 2.2. Thresholds in landscape metrics with increasing proportion of habitat
calculated using land cover Map 1 and a 75 km analysis window. .......................... 56
Figure 2.3. Thresholds in landscape metrics with increasing proportion of habitat
calculated using land cover Map 1 and a 45 km analysis window. .......................... 58
Figure 2.4. Thresholds in landscape metrics with increasing proportion of habitat
calculated using land cover Map 1 and a 15 km analysis window. .......................... 60
Figure 2.5. Relationship between range in years to colonization among traps located in
the same analysis window and landscape metrics calculated using land cover Map 1
and a 75 km analysis window. .................................................................................. 62
Figure 2.6. Relationship between range in years to colonization among traps located in
the same analysis window and landscape metrics calculated using land cover Map 1
and a 45 km analysis window. .................................................................................. 64
viii
Figure 2.7. Relationship between range in years to colonization among traps located in
the same analysis window and landscape metrics calculated using land cover Map 1
and a 15 km analysis window. .................................................................................. 66
Figure 3.1. Two dispersal distributions used to determine the proportion of dispersers
moving a given dispersal distance (number of cells in the landscape grid). The solid
line represents a gamma distribution with shape = 1.2, and rate = 0.5. The dashed
line represents a long-tailed gamma distribution with shape = 0.043, and rate = 0.5.
................................................................................................................................... 83
Figure 3.2. Change in variogram behavior for model scenario realsim1.7 characterized
by the total number of times each model was chosen as best-fit to the data in each
time step (a) and the mean probability that each model was the actual best model in
each time step (b). Mean probability estimates include 95% confidence intervals.
Variograms were fit with a quadratic equation (closed circles), a linear model (open
triangles), a model of the mean semivariance (closed squares), and an exponential
model (open diamonds). The solid line and values on the right ordinate represent
total population size averaged over all 50 scenarios. ................................................ 84
Figure 3.3. Change in variogram behavior for model scenario realsim2.1 characterized
by the total number of times each model was chosen as best-fit to the data in each
time step (a) and the mean probability that each model was the actual best model in
each time step (b). Mean probability estimates include 95% confidence intervals.
Variograms were fit with a quadratic equation (closed circles), a linear model (open
triangles), a model of the mean semivariance (closed squares), and an exponential
model (open diamonds). The solid line and values on the right ordinate represent
total population size averaged over all 50 scenarios. ................................................ 85
Figure 3.4. Change in variogram behavior for model scenario realstrat1.7Dl
characterized by the total number of times each model was chosen as best-fit to the
data in each time step (a) and the mean probability that each model was the actual
best model in each time step (b). Mean probability estimates include 95%
confidence intervals. Variograms were fit with a quadratic equation (closed circles),
a linear model (open triangles), a model of the mean semivariance (closed squares),
and an exponential model (open diamonds). The solid line and values on the right
ordinate represent total population size averaged over all 50 scenarios. .................. 86
Figure 3.5. Change in variogram behavior for model scenario realstrat2.lD1
characterized by the total number of times each model was chosen as best-fit to the
data in each time step (a) and the mean probability that each model was the actual
best model in each time Step (b). Mean probability estimates include 95%
confidence intervals. Variograms were fit with a quadratic equation (closed circles),
a linear model (open triangles), a model of the mean semivariance (closed squares),
and an exponential model (open diamonds). The solid line and values on the right
ordinate represent total population size averaged over all 50 scenarios. .................. 87
ix
Figure 3.6. Change in variogram behavior for model scenario realstratl .7D2
characterized by the total number of times each model was chosen as best-fit to the
data in each time step (a) and the mean probability that each model was the actual
best model in each time step (b). Mean probability estimates include 95%
confidence intervals. Variograms were fit with a quadratic equation (closed circles),
a linear model (open triangles), a model of the mean semivariance (closed squares),
and an exponential model (open diamonds). The solid line and values on the right
ordinate represent total population size averaged over all 50 scenarios. .................. 88
Figure 3.7. Change in variogram behavior for model scenario realstrat2.1D2
characterized by the total number of times each model was chosen as best-fit to the
data in each time step (a) and the mean probability that each model was the actual
best model in each time step (b). Mean probability estimates include 95%
confidence intervals. Variograms were fit with a quadratic equation (closed circles),
a linear model (Open triangles), a model of the mean semivariance (closed squares),
and an exponential model (open diamonds). The solid line and values on the right
ordinate represent total population size averaged over all 50 scenarios ................... 89
Figure 3.8. Change in variogram behavior for model scenario sim1.7 characterized by
the total number of times each model was chosen as best-fit to the data in each time
step (a) and the mean probability that each model was the actual best model in each
time step (b). Mean probability estimates include 95% confidence intervals.
Variograms were fit with a quadratic equation (closed circles), a linear model (open
triangles), a model of the mean semivariance (closed squares), and an exponential
model (open diamonds). The solid line and values on the right ordinate represent
total population size averaged over all 50 scenarios. ................................................ 90
Figure 3.9. Change in variogram behavior for model scenario sim2.1 characterized by
the total number of times each model was chosen as best-fit to the data in each time
step (a) and the mean probability that each model was the actual best model in each
time step (b). Mean probability estimates include 95% confidence intervals.
Variograms were fit with a quadratic equation (closed circles), a linear model (open
triangles), a model of the mean semivariance (closed squares), and an exponential
model (open diamonds). The solid line and values on the right ordinate represent
total population size averaged over all 50 scenarios. ................................................ 91
Figure 3.10. Change in variogram behavior for model scenario stratl .7D2 characterized
by the total number of times each model was chosen as best-fit to the data in each
time step (a) and the mean probability that each model was the actual best model in
each time step (b). Mean probability estimates include 95% confidence intervals.
Variograms were fit with a quadratic equation (closed circles), a linear model (open
triangles), a model of the mean semivariance (closed squares), and an exponential
model (open diamonds). The solid line and values on the right ordinate represent
total population size averaged over all 50 scenarios. ................................................ 92
Figure 3.11. Change in variogram behavior for model scenario strat2.lD1 characterized
by the total number of times each model was chosen as best-fit to the data in each
time step (a) and the mean probability that each model was the actual best model in
each time step (b). Mean probability estimates include 95% confidence intervals.
Variograms were fit with a quadratic equation (closed circles), a linear model (open
triangles), a model of the mean semivariance (closed squares), and an exponential
model (open diamonds). The solid line and values on the right ordinate represent
total population size averaged over all 50 scenarios ................................................. 93
Figure 3.12. Change in variogram behavior for model scenario stratl .7D1 characterized
by the total number of times each model was chosen as best-fit to the data in each
time step (a) and the mean probability that each model was the actual best model in
each time step (b). Mean probability estimates include 95% confidence intervals.
Variograms were fit with a quadratic equation (closed circles), a linear model (open
triangles), a model of the mean semivariance (closed squares), and an exponential
model (open diamonds). The solid line and values on the right ordinate represent
total population size averaged over all 50 scenarios. ................................................ 94
Figure 3.13. Change in variogram behavior for model scenario strat2.lD2 characterized
by the total number of times each model was chosen as best-fit to the data in each
time step (a) and the mean probability that each model was the actual best model in
each time step (b). Mean probability estimates include 95% confidence intervals.
Variograms were fit with a quadratic equation (closed circles), a linear model (open
triangles), a model of the mean semivariance (closed squares), and an exponential
model (open diamonds). The solid line and values on the right ordinate represent
total population size averaged over all 50 scenarios. ................................................ 95
xi
INTRODUCTION
Range expansion is a complex process that is affected by the interaction of many
factors, including short and long distance dispersal (Clark et a1. 2001 , Trakhtenbrot et a1.
2005), landscape heterogeneity (Hengeveld and Van den Bosch 1997, With 2002),
interspecific interactions (Richardson et al. 2000a, French and Travis 2001), and
evolution (Carroll and Dingle 1996, Travis and Dytham 2002). Recent work in the areas
of dispersal and invasion ecology indicates that environmental heterogeneity may be one
of the most influential factors affecting dispersal success of invasive species (With 2002,
Hastings et al. 2005). However, these studies do not synthesize all the spatial
information potentially available to ecologists. In particular, changes in the spatial
autocorrelation structure of populations expanding across heterogeneous landscapes have
not been thoroughly examined. Brodie et a1. (1995) first noted changes in spatial
autocorrelation patterns as a result of local dispersal in a study of the post—fire re—
invasion of a Populus balsamifera clone. Tracking such changes in the Spatial structure
may be a reliable indicator of important changes in population growth and dispersal
patterns during the course of an invasion. However, changes in spatial population
structure have not been identified in other environments or for species with difi‘erent
dispersal behaviors.
This dissertation contains three chapters that, respectively, address each of the
following questions:
1. Can important spatial changes in population dynamics over the course of
an invasion be characterized by geostatistics such as semivariance?
2. How is the movement of an invasion wavefront related to measures of
landscape heterogeneity?
3. How do dispersal strategy and landscape heterogeneity affect the
generation of semivariance progression patterns during range expansion?
In my first chapter, I identified new applications of geostatistics, specifically
semivariance, to the problem of detecting invasion stage transitions. The term “invasion
stages” refers to changes in population growth and dispersal patterns during the course of
an invasion that may be important indicators of mechanistic changes in the invasion
process. The goal of this chapter was to make and test predictions about the types of
semivariogram models that would best fit expected patterns of semivariogram behavior at
each stage of the invasion process. My objectives were to compare variogram behavior
at different spatial extents over the course of an invasion, and to identify potential
benefits and drawbacks of using semivariogram analysis as a tool in analyzing the
progress of biological invasions.
To meet my Objectives, I analyzed monitoring records from the gypsy moth
(Lymantria dispar) invasion of Michigan (Yang et al. 1998). The gypsy moth is an
exotic insect native to Europe that has caused extensive forest defoliation across much of
the eastern United States Since its introduction to Massachusetts in 1868 (Elkinton and
Liebhold 1990). The first breeding population of gypsy moths in Michigan was the result
of an independent introduction in 1954 (O‘Dell 1955, Hanna 1981). Gypsy moths
quickly spread across the state, reaching most areas of the Lower Peninsula by the early
19905 and the western Upper Peninsula by the late 19903 (Lele et al. 1998, Yang et al.
1998). The gypsy moth monitoring program in Michigan involved the placement of
pheromone-scented traps in sections eight and 26 of every township in Michigan (Yang
et al. 1998). At each trap, the total annual male moth catch and the trap location in UTM
coordinates were recorded. Traps were run between 1985 and 1996 in the Lower
Peninsula (LP) and between 1986 and 1996 in the Upper Peninsula. The Michigan gypsy
moth monitoring program records were ideal for analyzing changes in Spatial population
structure during range expansion because they contained spatially explicit data collected
at the same locations across a wide area for a long period of time.
My first chapter analyses characterized spatial changes in population structure of
the gypsy moth population over time, but those analyses did not consider the effects of
habitat on the range expansion process. Therefore, in my second chapter I investigated
the relationship between invasion success and measures Of landscape heterogeneity. To
date, landscape ecologists have used simple models of biological invasions across
computer-generated landscapes to make predictions about how the dispersal success is
affected by landscape heterogeneity (With and Crist 1995, With and King 1999a,
Collingham and Huntley 2000). Such models indicate that dispersal success declines
with habitat loss, but a precipitous decline in dispersal success may be observed if habitat
patches begin to break down and the negative effects of such habitat fragmentation are
additive (With and Crist 1995, With and King 1999a, Collingham and Huntley 2000). To
characterize the effects of landscape heterogeneity on invasion success, threshold
responses (sudden, nonlinear changes) in dispersal success with habitat loss are compared
to thresholds in landscape structure measurements. Landscape measurements that exhibit
thresholds most Similar to thresholds in dispersal success are thought to be the best
candidate measurements for predicting the effects of landscape structure on dispersal
success (With 2002). The goal Of this study was to characterize the relationship between
invasion success of a real species and the structure of a real landscape. My objectives
were to compare thresholds in gypsy moth invasion success and the structure of
Michigan’s landscape with thresholds produced by dispersal models run on computer-
generated landscapes. Also, I assessed the effect of scale on my results by repeating
invasion and landscape structure calculations at three scales of analysis using two
different classified land cover maps of Michigan.
The goal of my third chapter was to identify the underlying process or processes
responsible for the generation of semivariance progression patterns Observed in my first
chapter. I hypothesized that differences in either dispersal strategy (simple vs. stratified)
or landscape heterogeneity would affect the spatial structure of a population as it
underwent range expansion. To test this hypothesis, I assessed the ability of two
dispersal models to generate distinct changes in semivariance patterns over time and
compared model results with gypsy moths semivariance patterns (Chapter 1). My
Objectives were to explore the effects of landscape heterogeneity, population growth rate,
and the amount of long-distance dispersal on semivariance patterns.
CHAPTER 1
CHARACTERIZIN G THE INVASION PROCESS USING PATTERNS OF
SEMIVARIANCE: A CASE STUDY OF THE GYPYSY MOTH IN MICHIGAN
Introduction
Predicting the course of biological invasions has been an important question since
the emergence of invasion ecology almost 50 years ago (Reichard and White 2003).
However, a broad predictive framework for the invasion process has yet to be proposed
(Williamson 1999). One problem with accurately predicting the course of an invasion is
that data are not typically collected at a temporal and spatial extent large enough to
document and analyze the ecological processes of interest, namely dispersal and range
expansion. Even when the scale of data collection matches the scale of the invasion
process, prediction is complicated by the fact that overall patterns of population growth
and spread change as an invasion progresses (Townsend 1991).
Changes in growth and dispersal patterns during the course of an invasion are
often referred to as "invasion stages" (Hengeveld 1989, Brown 1993, Cousens and
Mortimer 1995, Shigesada et al. 1995, Hastings 1996b). Typically, three distinct stages
are recognized following “introduction”, or the initial arrival of an invasive species in a
new area. The first of these three stages is called “establishment” (or “colonization”) and
represents the formation of a self-sustaining population through successful survival and
reproduction (Anderson and May 1986, Cronk and Fuller 1995, Shigesada and Kawasaki
1997, Richardson et al. 2000b). The second stage is called “expansion” (or “spread”),
and represents the dispersal and potential formation of new self—sustaining populations
(Hastings 1996a). A final stage called “stabilization” follows expansion and represents
the stage at which an invader ceases to spread and becomes integrated into its new
environment (Cronk and Fuller 1995, Vermeij 1996). Reliably identifying invasion
stages is a critical first step in breaking down the invasion process and understanding how
the effects of different ecological factors (e.g. landscape heterogeneity) change
throughout the course of an invasion (With 2002).
The most common method for identifying the stages of invasion in empirical data
is analyzing temporal trends in population growth. The population trajectory of an
invader is theorized to undergo an exponential increase until environmental factors
become limiting (Ashton and Mitchell 1989). The stages of population growth are
usually characterized by an initial period of low density during which numbers increase
exponentially (Mack 1985, Cousens and Mortimer 1995). This stage is followed by a
period of density dependent population growth that ends when the upper density limit of
the population is reached. After these stages are complete, invaders may exhibit decline,
damped oscillations, stable oscillations, a stable equilibrium, multiple equilibria, or chaos
(May 1974, Brown 1993).
Mack (1985) suggested that the three stages of population growth correspond with
invasion stages (e. g. establishment, expansion, and stabilization). However, there are
several problems with using density dependent phase transitions to define invasion
stages. Population increase with and without spatial spread (establishment vs. expansion)
cannot be distinguished without corresponding spatial information. Also, population
growth during the expansion stage may not be density limited (Mack 1985). Kowarik
(1995) proposed that a time lag of slow, constant population growth signifies a clear
transition between the conceptual stages of establishment and expansion. However, low
initial population densities can make detection of early invasion stages difficult and
apparent time lags should be interpreted with care. Cousens and Mortimer (1995) caution
that it is extremely difficult to distinguish between a time lag and the early stages of
exponential growth with a constant, proportional rate of increase. At the beginning of an
invasion, low initial population size or low intrinsic growth rate may give the impression
of a time lag when there is none (Case 2000). Additional problems arise when
identifying the transition between expansion and stabilization. Curve—fitting an
incomplete time series may lead to false conclusions about the status of an invasion if
sampling error or year—to—year variation are significant (Case 2000). Given the
difficulties of identifying invasion stages in time series data, invasion dynamics may be
too variable among organisms for a common set of invasion stages to be reliably
identified.
An alternative explanation may be that stages of invasion could be more clearly
characterized when spatial dynamics are explicitly considered. If this is true, a more
spatially explicit approach is needed to identify the stages Of an invasion. Ecological data
are typically spatially autocorrelated (Rossi et al. 1992, Legendre 1993) so that data
collected at locations close to one another are more similar than data collected at
locations farther apart (Cliff and 0rd 1981, Griffith 1987). Analysis of spatially
autocorrelated data collected from permanent locations over a large area relative to the
area occupied by an invasive species may provide reliable indications of transitions
between invasion stages. Spatial autocorrelation often violates crucial assumptions of
many traditional statistical analyses because sample values are not independent
(Legendre and Legendre 1998). However, patterns of spatial autocorrelation result from
important ecological processes and should be considered a fundamental component of the
ecology of a Species (Legendre 1993). Careful inspection of these patterns may be
critical in identifying the stages of invasion. For example, detailed information on
patterns of spatial autocorrelation has been used to identify the stages of post-fire re—
invasion of a Populus balsamifera clone using correlograms, or plots of Moran’s I (an
index of spatial autocorrelation) against distance (Brodie et al. 1995). Three distinct
patterns of spatial autocorrelation were observed for each correlogram representing a
different age class, suggesting three separate stages of clonal development: post-fire
colonization, consolidation, and directional expansion (Brodie et al. 1995). These stages
may be roughly analogous to the invasion stages of introduction, establishment, and
expansion.
A related method for identifying changes in spatial autocorrelation patterns over
time is semivariance analysis. Semivariance is a spatial statistic used to characterize
dependence of sample values across space. The graph of semivariance against distance
class is called a semivariogram (or simply variogram) and depicts the average
dissimilarity between values based on separation distance between sampling points. The
standardized semivariance is equal to 1 minus the correlation when the population mean
and variance remain constant across the study area (Rossi et a1. 1992). Semivariogram
models are often used to interpolate between sample locations within a study area (e. g.
kriging, Burgess and Webster 1980), so variograms are most often treated as a means to
an end. Variograms by themselves, however, provide important information about the
invasion process that other methods such as time series analysis cannot (Rossi et al.
1992). For instance, semivariance is typically more sensitive to proportional effects
(correlation between local means and variances) than other measures of spatial
correlation or covariance; proportional effects cause “first—order” or linear trends in a
variogram and can result in overestimation of the range of distances at which data are
correlated (Rossi et al. 1992). Generally, this property of variograms is highly
undesirable for spatial modeling. For the purposes of identifying large—scale invasion
stage transitions, however, variogram analysis may be superior to other spatial analyses
because first—order trends in variograms may be indicative of important mechanistic
changes in invasion dynamics.
Given a general understanding of invasion dynamics, specific predictions can be
made regarding the pattern of semivariance expected at each invasion stage
(establishment, expansion, and stabilization). During the establishment stage, the
invasion is likely restricted to a small area relative to its potential range. Therefore, only
local, small-scale patterns of semivariance should be observed. If sampling occurs over
an area equal to or larger than the area occupied, a bell- or wave-shaped variogram may
result (Villard and Maurer 1996, Pebesma 1999). In this case, low semivariance (high
spatial autocorrelation) is observed not only between pairs of sample points located close
together, but also between pairs of sample points located far apart at opposite edges of
species’ range. Movement of an invasion wavefront across the study area during the
stage of expansion is a process that may result in first—order variogram trends. A first—
order trend indicates that a pattern is emerging in the data that is larger than the lag
classes modeled; therefore, local means and variances will differ in different locations
and directions (Rossi et al. 1992). During the expansion stage Of invasion, a gradient of
high to low population index values may occur that would appear in a variogram as a
straight line (Rossi et al. 1992, Bailey and Gatrell 1995, Legendre and Legendre 1998).
During stabilization, populations begin to settle and resume normal fluctuations with no
new large additions to the range unless a species is migratory or overcomes a large
physical barrier to expansion (Isard and Gage 2001). Therefore, at this stage one might
expect to observe an untrended variogram pattern in which semivariance increases until it
reaches the maximum distance at which data are spatially dependent and then levels off.
Obtaining the data necessary to conduct variogram analysis and test these
predictions requires careful, long—term monitoring over a large area. One such
monitoring program tracked the invasion of the gypsy moth (Lymantria dispar) across
Michigan (Gage et al. 1990, Yang et al. 1998). The gypsy moth is an exotic insect native
to Europe that has caused extensive defoliation across much of the eastern United States
since its introduction to Massachusetts in 1868 (Elkinton and Liebhold 1990). The first
breeding population of gypsy moths in Michigan was the result of an independent
introduction in 1954 and a second possible reintroduction occurred near Midland in the
19803 (O'Dell 1955, Hanna 1981). Gypsy moths quickly spread across the state, reaching
most areas of the Lower Peninsula by the early 19903 and the western Upper Peninsula
by the late 19903 (Lele et al. 1998, Yang et al. 1998).
In this chapter, my goal was to make and test predictions about the types of
models that would best fit expected patterns of variogram behavior at each stage of the
invasion process. My objectives were to (1) model and characterize variogram behavior
over the course of an invasion, (2) compare variogram behavior at different spatial
10
extents over the course of an invasion, and (3) present the potential uses of variogram
analysis as a tool in monitoring the progress of additional biological invasions.
Methods
Pheromone—baited traps were placed in sections eight and 26 of every township in
Michigan over 12 years from 1985 to 1996 (Gage et al. 1990, Yang et al. 1998). At each
trap, the total annual male moth catch and the trap location in UTM coordinates were
recorded. Traps were monitored between 1985 and 1996 in Michigan’s Lower Peninsula
(LP) and between 1986 and 1996 in Michigan’s Upper Peninsula. A subset of 1,090
traps was selected for this analysis from over 3,000 placed in a regular grid across the
state. Only traps Operated all 12 years at the same location were selected to avoid change
of support in the analysis (Isaaks and Srivastava 1989). Traps located in the Upper
Peninsula were excluded because the invasion was still in the early stages by 1996.
Variogram analysis was conducted at two different scales for this study. First,
semivariance was calculated for traps located across the entire LP. Semivariance was
calculated as:
NU?)
2N(h),= _—_le(xi_ xi+h)
7(h)— -
where y(h) is the semivariance for the distance interval h, N(h) is the total number of
sample pairs for the distance interval h, Xi is the measured sample value at location i, and
Xi +h is the measured sample value at point i + h (Isaaks and Srivastava 1989). Second,
data were grouped and analyzed by ecoregion using subsection delineations (Albert et al.
1986) that were modified to ensure continuous yet distinct groupings (Figure 1.1).
11
Ecoregions were chosen as units Of analysis because they represent areas of similar
habitat and climate without regard for political boundaries. Isotropic variograms were
calculated for each unit of analysis (entire LP or individual ecoregion) in each year of the
study using the geoR package in R (Ribiero and Diggle 2001, R Development Core Team
2004). Semivariance was calculated for 13 distance intervals or lag classes; the number
of pairs in each lag class ranged from 9,643 to 59,594 pairs for the LP analysis and 16 to
1,983 pairs for the ecoregion analysis. For the LP analysis, the maximum distance was
set at 300,000 111, roughly representing the north—south length of the LP; for ecoregion
analyses, the maximum distance was set at 100,000 m, roughly representing the length of
most ecoregions. Large maximum distances were used to observe variogram behavior
across the extent of the unit being analyzed.
To characterize the shape of each variogram, seven models were fit to each
variogram using maximum likelihood estimation (Bumham and Anderson 2002).
Although maximum likelihood may produce biased variogram parameter estimates in
cases of small to moderate sample size (Cressie 1993), this study included a sufficient
number of data points (>1,000) to avoid such problems. Each model included 3—4 of the
following parameters: range (a), the maximum distance at which data are spatially
dependent; sill (C1), the semivariance value at the which the range is reached; nugget
(Co), the component of the variance caused by local variability at scales smaller than the
sampling interval; or smoothness (K), a shape parameter signifying the order of the
modified Bessel functions of the third kind.
12
As noted above, a bell— or wave—shaped variogram would be expected during the
initial stage of establishment. Such variogram patterns were predicted to be best fit with
either the cubic model
2 3 5 7
)7(h) = C0 + C1[7(£) —8.75[£) +3.5[flj — 075(2) ] if h < a;
a a a a
otherwise y(h) = C O
or wave (also known as cardinal sine or hole effect) model
m) = C0 + C,[1- {5’— sin[11-D]
h a '
Linear, first—order trends are expected during the expansion stage that would be best fit
with the power model (exponent equal to l)
79(11)==C0 +C1h",00
or possibly the exponential model
a
}9(h) = C0 +C1I:1——exp [—3—]:H
Finally, a variogram was predicted to exhibit untrended behavior once the stage of
stabilization is reached. In this situation, semivariance would increase until an asymptote
is reached. Such a pattern would be best fit with either the power (exponent < 1),
exponential, spherical
3
)9(h) =CO +Cl[1.5fi—0.5[£] Jitha; 7(h) =CO ifh>a
a a
Gaussian
13
. h2
y(h) = C0 +C,|:l—exp [—3a—z—H
or generalized Cauchy model
2
}9(h)=CO+C1 1-— 1+6?) whereK>0,a>0.
These seven models were used to characterize variogram behavior because each
characterizes a slightly different pattern of spatial autocorrelation. The first five models
are bounded models for which the semivariance increases until the range is reached and
the variogram levels off (Goovaerts 1997, Chiles and Delfiner 1999). The spherical and
cubic models reach their sill at the range. The exponential, Gaussian, and generalized
Cauchy models approach their sill asymptotically. The spherical and exponential models
exhibit linear behavior near the origin, whereas the cubic, Gaussian, Cauchy, and wave
models exhibit parabolic behavior near the origin. The exact behavior of the generalized
Cauchy is further governed by the smoothness parameter (K). The power model is an
unbounded model that is linear when a = 1, curves upward when a > 1, and curves
downward when a < 1; the behavior of the power model at the origin also depends on the
value of the exponent. Note that the sill and range in the power model equation are not
equivalent to their counterparts in other models because this equation models an
unlimited dispersion process and is not intrinsically stationary (Chiles and Delfiner
1999).
The best—fit model for each year and unit of study was chosen by selecting the
model with the lowest Akaike Information Criterion (AIC) value (Bumham and
Anderson 2002). Comparing differences in magnitude of AIC values among models
14
allowed us to evaluate the amount of evidence in favor of each model. Finally, the timing
of any observed changes in variogram shape were compared with temporal patterns of
total trap catch in each unit of study.
Results
Variogram behavior in Michigan’s Lower Peninsula
Annual patterns of spatial autocorrelation, quantified as variograms of gypsy
moth trap catch, changed markedly between 1985 and 1996 in the LP (Figure 1.2, Table
1.1). In the first four years of the gypsy moth invasion (1985—1988), variograms
displayed a bell—shaped curve and the best—fit model was cubic, as predicted. In 1989,
the bell—shaped variogram was best fit with the wave model, but the cubic model
exhibited a very similar AIC value. From 1990 to 1992, variograms were linear in shape
and the best—fit model was either the power (1991—1992; a = 1), as predicted, or the
spherical (1990) model. By 1993, first order trends were no longer prevalent in the data
and semivariance increased steadily to an asymptote. The best—fit models for these
untrended variograms were the power (1993—1995; exponent < 1) or the exponential
(1996), as predicted. Changes in variogram behavior corresponded only partially with
temporal changes in trap catch (Figure 1.3). Variogram patterns changed from bell—
shaped to linear in 1990 when the temporal patterns in total catch began to level Off.
However, the switch from a linear to an untrended variogram in 1993 did not correspond
with a change in temporal trends.
15
Variogram behavior at the ecoregion scale
At the ecoregion scale, variogram behavior was more variable and predictions for
variogram patterns at each stage of the invasion were only supported weakly by the three
largest ecoregions (Table 1.1). Variograms for the High Plains ecoregion were linear for
the first five years of the study. The best fit for these variograms was provided by the
Gaussian (1985—1986) or wave (1987—1989) model due to a small downward curve in the
first few lag classes. By 1990, a small wave pattern emerged that was, not surprisingly,
best fit by the wave model. Between 1991 and 1996, an untrended variogram pattern was
Observed and the best-fit model was either the Gaussian (1991 and 1994), the cubic
(1992), the wave (1993 and 1996), or the power (1995, exponent < 1). The observed
switch in behavior of the variograms did not correspond with patterns observed in the
time series. Trap catch in the High Plains ecoregion grew exponentially until 1990 when
catch began to decline then level off by 1993 (Figure 1.4).
Between 1985 and 1993, variograms in the Ionia ecoregion displayed a slight
bell—shaped pattern that was best fit by the wave (1985, 1990—1991) cubic (1988—1989)
or Gaussian (1986—1987) model. By 1992, an untrended variogram emerged and the
Cauchy (1992), wave (1993), spherical (1994—1995) or the power (1996, exponent < l)
was the best—fit model. Trap catch in the Ionia ecoregion did not correspond with
changes in variogram behavior. Trap catch increased over the course of the study with a
slight dip reported in 1992 (Figure 1.4).
Variograms for the Washtenaw ecoregion were largely linear in nature until the
last year of the study. Four of the first five years (1985, 1986, 1988, and 1989) were best
fit with the wave model. In 1987, and 1990—1992, the Gaussian model provided the best
16
fit. Between 1993 and 1995, the power model (exponent ~1) best fit the variogram. By
1996, however, semivariance increased then leveled off; in 1996, the best—fit model was
the generalized Cauchy. Inspection of the time series for the Washtenaw ecoregion
indicated the population was growing in an exponential fashion over the 12 years of the
study (Figure 1.4).
Variogram behavior over the study period for the Saginaw Bay ecoregion was
highly variable. In the first two years of the study, a bell—shaped curve was observed that
was best fit with the cubic (1985) and wave (1986) models. In 1987 and 1988, a linear
pattern emerged that was best fit with the wave model. By 1989, the bell—shaped curve
was observed again and best fit with the cubic model. Between 1990 and 1992, linear
variograms were observed and best fit with either the Gaussian (1990), spherical (1991),
or wave (1992) model. In 1993 and 1994, the bell shape emerged again and was best fit
using the wave (1993) and cubic (1994) models. In 1995, another linear variogram was
best fit with the Gaussian model. Finally, in 1996, an variogram emerged that was linear
with a negative slope; none of the seven models fit this pattern well with the possible
exception of the Cauchy model. Variogram patterns did not correspond with temporal
trends in trap catch. In general, trap catch in the Saginaw Bay ecoregion increased until
1989 when it dropped back to levels Observed in 1985 and then leveled off (Figure 1.4).
Variogram behavior for the Huron ecoregion oscillated between a bell—shaped
curve and a linear pattern. The bell—shaped curve was best—fit with the cubic (1985 and
1989) and the wave (1986—1987 and 1992) models. Linear variograms were best fit with
either the Gaussian (1988 and 1996), power (1990 and 1993), or the wave (1991, 1994—
1995) model. Trap catch in the Huron ecoregion did not correspond with changes in
17
variogram behavior. Trap catch in Ecoregion 2 increased exponentially until 1989 when
it began to level off (Figure 1.4).
Variograms for the Arenac ecoregion were mostly linear with the exception of 3
years in which an untrended variogram pattern was observed. Linear patterns were best
fit with either the wave (1985, 1987, 1989-1990, and 1994), power (1988 and 1991;
exponent ~ 1), or Gaussian (1993 and 1996) model. In 1986, 1992, and 1995, untrended
variogram patterns were observed and best fit with the wave (1985), Gaussian (1992),
and cubic (1995) models. Trap catch in the Arenac ecoregion did not correspond with
changes in variogram behavior. Trap catch increased exponentially until 1989 when it
dropped and leveled off by 1992 (Figure 1.4).
Variograms for the Presque Isle ecoregion exhibited a bell—shaped curve pattern
until 1996 when an untrended variogram pattern was observed. Bell—shaped variograms
were best fit with either the wave (1985, 1988—1994) or Gaussian (1986—1987 and 1995)
model. The Gaussian was also the best—fit model in 1996 when semivariance increased
and leveled off at the sixth lag class. Changes in trap catch did not correspond with
changes in variogram behavior. Trap catch in the Presque Isle ecoregion exhibited a
possible time lag between 1985 and 1988 then increased exponentially until 1992 when
catch dropped and leveled off (Figure 1.4).
Variogram patterns for the Kalamazoo ecoregion were dominated with a bell-—
shaped curve until 1996 when an untrended variogram pattern was observed. The bell—
shaped pattern was best fit with either the wave (1985 and 1994), cubic (1986, 1990—
1991, and 1993—1994), generalized Cauchy (1987), or Gaussian (1988—1989 and 1992).
In 1996, the untrended variogram shape was best fit with the spherical model. The
18
switch between bell—shaped and untrended variogram patterns was not reflected in the
graph of trap catch that increased exponentially throughout the course of the study
(Figure 1.4).
Variogram behavior for the North Allegan/Manistee ecoregion was mostly
characterized by a bell—shaped curve from 1985—1992; one exception was the largely flat
variogram in 1986. Variograms were best fit with either the cubic (1985, 1988, and
1991—1992), Gaussian (1987), or wave (1989—1990) model. Between 1993 and 1995, an
untrended variogram pattern emerged that was best fit with the power (1993,exponent
<1), Gaussian (1994), and cubic (1995) model. In 1996, a linear pattern was observed
and best fit with the wave model. Again, changes in total trap catch did not correspond
temporally with changes in variogram behavior. Trap catch in the North
Allegan/Manistee ecoregion exhibited a possible time lag from 1985 to 1988 then
increased exponentially until 1990 when it began to level Off (Figure 1.4).
The South Allegan ecoregion exhibited largely erratic variogram behavior. The
1985 variogram exhibited a bell-shaped curve that was best fit with the cubic model.
Over the next 11 years, however, variograms oscillated between untrended and linear
behavior. Untrended variograms were best fit with the exponential (1986), the wave
(1987, 1990, and 1995-1996), and spherical (1989 and 1991). More linear variograms
were best fit with the exponential (1992 and 1994), wave (1988), and power (1993,
exponent ~1). Time series data did not reflect the instability observed in variogram
analysis. The South Allegan ecoregion exhibited a possible time lag of nine years
followed by exponential growth and a leveling Off after the peak catch in 1993 (Figure
1.4).
19
Discussion
Variogram behavior in Michigan ’3 Lower Peninsula
The three, distinct variogram patterns displayed by gypsy moth trap catch data
across the LP from 1985—1996 likely represented expected patterns of spatial
autocorrelation for the invasion stages of establishment, expansion, and stabilization
because sampling occured across a large enough spatial scale. During the early stages of
invasion, bell—shaped variograms best fit by the cubic and wave models were observed
that may signify the establishment stage of invasion during which the population spreads
very little as it increases locally. Next, linear variograms were observed that were best fit
typically by the power model (exponent = 1); this linear behavior was likely caused by
expansion of the gypsy moth as it disperses across the state, resulting in a gradient of
high to low trap catch in the north—south direction (Yang et al. 1998). From a purely
spatial perspective, gypsy moths could be characterized as having reached the stage of
stabilization by 1993 in the LP of Michigan. From 1993—1996, local patterns of spatial
autocorrelation were no longer masked by first order trends and an untrended variogram
emerged. Gypsy moths were well established across most of the LP by 1993 and the
gradient in trap catch was not as strong (Yang et al. 1998).
A similar progression of variogram patterns was reported by Liebhold et al.
(1991) in their study of historical egg mass distributions over a large portion of
northeastern Massachusetts, southeastern New Hampshire, and southern Maine.
Liebhold et al. (1991) reported a slightly bell— or wave—shaped variogram in 1910,
largely linear variograms between 1911 and 1914, and an untrended variogram in 1915.
Given that gypsy moths were introduced to Massachusetts in the late 18803, it is not
20
unreasonable to assume that the gypsy moth invasion of New England had begun to
transition between stages over the next 20 years. The observed variogram shifts over
time potentially represent changes in invasion stages, but additional studies are needed on
new invasive Species using appropriate spatial sampling protocols to unravel this
complex process.
Variogram behavior at ecoregion scale
While variogram analysis for the entire LP showed three distinct behavioral
switches, analyses conducted at the ecoregion extent were not as straightforward to
interpret. Overall, predictions of variogram shape and model fit at each stage of the
invasion were only supported in part by the three largest ecoregions. A variogram
progression similar to that of the LP was displayed by the High Plains ecoregion data.
Also, the Ionia ecoregion exhibited linear variograms followed by untrended variograms
in the early 19903 in a fashion similar to the LP. These observed variogram patterns
make sense given the invasion history of the area; gypsy moths partially invaded this
ecoregion before the monitoring period began, then halted or slowly spread across it,
establishing a gradient in trap catch until the early 19903 when the invasion began to
spread more extensively across south—central Michigan (Yang et al. 1998). The
Washtenaw ecoregion was the last ecoregion in the LP to be completely invaded by
gypsy moths; this ecoregion contains the city of Detroit and the surrounding suburbs and,
in general, is a highly developed urban area with little suitable gypsy moth habitat. Until
the last year of the study, variograms were largely linear in nature, indicating that the
invasion wavefront was late in arriving in this region of the state. The remaining 7
21
ecoregions exhibited variogram behavior changes over time that were dissimilar to that of
the LP.
The Saginaw Bay, Huron, and Arenac ecoregions contained the original
establishment area for gypsy moths in Michigan (Hanna 1981), so trap catch in these
ecoregions reflect numbers Observed after the wavefront has passed. Not surprisingly,
variogram patterns in these ecoregions were different fiom patterns observed in
ecoregions colonized later in the study period. The alternation between bell—shaped and
linear (Saginaw Bay and Huron) or between linear and untrended (Arenac) variograms
may reflect changes in spatial autocorrelation that accompany natural cycling patterns of
gypsy moths after population stabilization.
Other patterns of variogram change over time were more difficult to explain.
Both the Presque Isle and Kalamazoo ecoregions exhibited bell—shaped variograms until
an untrended variogram pattern emerged in 1996. However, the invasion history of these
two ecoregions is strikingly different. The Presque Isle ecoregion was invaded within a
two—year period (1990—1991), whereas the invasion wavefront did not hit Kalamazoo
until 1993. It is unclear why both these ecoregions with dissimilar histories and shapes
would exhibit similar variogram patterns over time. Another mixed series of all three
variograms types were displayed in the North Allegan/Manistee ecoregion, perhaps
because two somewhat dissimilar, narrow ecoregion subsections were inappropriately
combined to create the area of analysis. In addition, the South Allegan ecoregion
alternated erratically between linear and untrended variograms over the study period.
This ecoregion is also long and narrow and may not be an appropriate shape for isotropic
variogram analysis.
22
In summary, only the three largest ecoregions (High Plains, Ionia, and
Washtenaw) exhibited variogram patterns over time that were consistent with their
invasion history, my initial predictions, and the patterns observed at the LP analysis
extent. This Observation is also supported in part by an egg mass study conducted by
Liebhold et al. (1995). Of the 4 large spatial datasets analyzed, only the dataset largest in
area (the tri—state dataset) exhibited the three—stage variogram patterns observed in the
Michigan LP analysis; an additional dataset collected over three years across the state of
Massachusetts was too short in duration for fair comparison. Therefore, the use of
variogram analysis to monitor invasion progress may need to be conducted across
relatively large areas (e.g. area of the High Plains ecoregion is 21,000 kmz). These
results reaffirm the fact that variograms are designed to statistically model structural
spatial dependence, or large—scale trends using many data locations simultaneously
(Rossi et al. 1992). However, the exact size of the monitoring region will likely depend
on the dispersal ability and other life history characteristics of the species in question.
Timing of shifts in variogram behavior
One important clue that variogram analysis may be a vital component in the
analysis of invasion progress is that the timing of variogram behavior transitions did not
correspond with changes in total trap catch, an index of population growth (Figure 1.3).
Without consideration Of spatial patterns, inspection of the time series alone would have
indicated that the invasion had stabilized by 1990. A similar discrepancy between
variogram patterns and trends in total catch over time was observed at the ecoregion
extent. For example, the time series for the Washtenaw ecoregion did not show any
23
stabilization in trap catch, whereas the 1996 variogram hinted at the possibility of a more
stable pattern of semivariance late in the study period. A similar disparity between the
timing of population growth and range expansion was observed by Maurer er al. (2001)
in their study of the European starling (Sturnus vulgaris) invasion of North America. In
this case, population size lagged behind geographic range size (kmz) such that marked
population grth did not occur until the invasion wavefront began to slow upon arrival
at the west coast of North America.
Because shifts in variogram behavior over time do not appear to coincide with
major changes in observed time series, one or more unidentified processes may be
driving the spatial dynamics of an invading species. For example, the final stage of the
gypsy moth invasion (stabilization) appeared to coincide with two major changes: 1) a
temporal leveling off of total trap catch evident in the time series, and 2) the switch from
linear to untrended variograms. While the concept that an invading population will
eventually reach its maximum is widely accepted, the concept that local spatial
autocorrelation is necessary for invasion stabilization is not as intuitive. One possible
explanation may be that stabilization of an invasion requires the establishment of
metapopulations (Moller 1996). A metapopulation is a set of local populations within
some larger area that experiences inter—population migration (Hanski and Gilpin 1997).
Moller (1996) suggested that metapopulation structure may be an important mechanism
for buffering the effects of local demographic and environmental fluctuations on the
overall invasion. However, barriers to interpopulation dispersal may prevent stabilization
of the invasion process (Anderson and May 1986, Moller 1996). Stabilization of
invasion dynamics is thought to occur in part by immigration between populations
24
subject to different environmental conditions. Brown and Kodric—Brown (1977)
described this phenomenon as the rescue effect because influx of individuals fi'om
neighboring populations may “rescue” a population from extinction. Stabilizing
metapopulation dynamics may be especially important for small invader populations or
populations located in less than optimal conditions (Roughgarden 1986, Stacey and Taper
1992). However, the extent to which metapopulation dynamics or other potential
mechanisms contribute to the invasion process is still unknown.
Use of variograms in invasion analysis and interpretation
If spatial analysis of other invasion data reveals similar spatial patterns that
correspond to invasion stages, the use of variogram analysis should prove insightful in
the improved monitoring and prediction of invasive spread. However, the use of
variograms in this manner will require a change in the way ecologists typically collect,
display, and analyze patterns in spatial data. To obtain enough data for proper variogram
analysis, extensive long—term, spatially explicit monitoring programs for multiple
invaders must be developed that include a static sampling design (i.e. sampling method or
location of traps does not change significantly over the study period). Any method of
identifying invasion stages will falsely identify stage transitions if there is a change in
monitoring effort or sampling methodology (Cousens and Mortimer 1995). Given the
ever—increasing number of new invasions and billions of dollars in damage that result
(Pimentel et al. 2000), applying more resources and effort into large—scale monitoring
programs will likely prove to be a good investment.
25
The second change in approach to invasion analysis involves plotting variograms
of data collected across or even beyond the scope of the spatial process (i.e. beyond the
invasion wavefront, Rossi et al. 1992, Legendre and Legendre 1998). Even when all zero
counts are excluded, a similar bell—shaped curve results that is best fit using either the
wave or cubic models. Although these variograms may be informative for exploring
invasion stage transitions, bell—shaped variograms should not be used to for interpolative
modeling unless localized kriging is employed with an appropriately small search radius
(Pebesma 1999).
The third change in approach to spatial data will require ecologists to tolerate the
presence of first order trends in a variogram. First order trends typically linearize a
variogram and are best fit using the power or exponential models. First order trends
imply that the attribute value measured at a given point depends on the location of that
point in the study area, a clear violation of the assumption of stationarity (no proportional
effects) required for many methods of interpolative modeling (Isaaks and Srivastava
1989). Despite the drawbacks of trended data, such patterns are indicative of a species
that is undergoing range expansion and should not be ignored. Examining variograms for
first order trends may be useful as an important invasion analysis tool. If interpolative
modeling or estimating spatial process parameters (e.g. range, nugget effect) are of
interest, first order trends should be modeled, if possible, and the residuals used to
calculate a variogram and krige (Cressie 1993, Chiles and Delfiner 1999).
Variogram analysis has great potential for detecting invasion stages. If an
invasion is monitored at an appropriately large spatial extent, changes in variogram
behavior over time may provide a quick identification of invasion stage. Invasions
26
contained in the establishment stage are more likely to result in successful eradication or
reasonable control; therefore, managers may decide to concentrate efforts on populations
identified by semivariance analysis to be in the early stages of invasion. Given that
timing of changes in semivariogram behavior did not coincide with major temporal
changes in total trap catch, semivariance may be the most reliable method available for
identifying invasion stages. Alternatively, a lag between arrival of the invasion
wavefront and changes in the number of adult male moths caught may have resulted in
the timing discrepancy between spatial and temporal patterns. Further examination of
when and why invasions transition between stages may improve prediction of invasion
dynamics if important mechanistic changes in population dynamics prove to be involved.
27
Table 1.1. Modeling results from variogram analysis Of annual gypsy moth (Lymantria
dispar) trap catch in the Lower Peninsula (LP) and 10 ecoregions Of Michigan, 1985-
1996. The best model (model with lowest Akaike Information Criterion value) of seven
models fit to annual variograms of gypsy moth trap catch is listed as “Type”. Note that
the exponential (Expon.), Gaussian (Gaus.), and spherical (Spher.) model names were
abbreviated. Model shape (Shape) describes the overall shape of the variogram being fit
(B = bell or wave shaped, L = linear, and U = untrended variogram with typical increase
in semivariance to an asymptote).
LP High Plains Ionia Washtenaw Saginaw Bay Huron
Year Type Sh. Type Sh. Type Sh. Type Sh. Type Sh. Type Sh.
1 985 Cubic B Wave L Wave B Wave L Cubic B Cubic B
1 986 Cubic B Gaus. L Gaus. B Wave L Wave B Wave B
1 987 Cubic B Wave L Gaus. B Gaus. L Wave L Wave B
l 988 Cubic B Wave L Cubic B Wave L Wave L Gaus. L
1 989 Wave B Wave L Cubic B Wave L Cubic B Cubic B
1990 Spher. L Wave B Wave B Gaus. L Gaus. L Powerl L
1991 Power1 L Gaus. U Wave B Gaus. L Spher. L Wave L
1992 Power‘ L Cubic U Cauchy U Gaus. L Wave L Wave B
1993 Power2 U Wave U Wave U Powerl L Wave B Powerl L
1994 Power2 U Gaus. U Spher U Power‘ L Cubic B Wave L
1995 Power2 U Power2 U Spher. U Power‘ L Gaus. L Wave L
1996 Expon. U Wave U Power2 U Cauchy U Cauchy L Gaus. L
1 exponent = 1
2 exponent < l
28
Table 1.1 (cont’d).
Arenac Presque Isle Kalamazoo North Allegan South Allegan
Year Type Sh. Type Sh. Type Sh. Type Sh. Type Sh.
1985 Wave L Wave B Wave B Cubic B Cubic B
1986 Wave U Gaus. B Cubic B Expon. B Expon. U
1987 Wave L Gaus. B Cauchy B Gaus. B Wave U
1988 Power L Wave B Gaus. B Cubic B Wave L
1989 Wave L Wave B Gaus. B Wave B Spher. U
1990 Wave L Wave B Cubic B Wave B Wave U
1991 Power' L Wave B Cubic B Cubic B Spher. U
1992 Gaus. U Wave B Gaus. B Cubic B Expon. L
1993 Gaus. L Wave B Cubic B Power2 U Powerl L
1994 Wave L Wave B Cubic B Gaus. U Expon. L
1995 Cubic U Gaus. B Wave B Cubic U Wave U
1996 Gaus. L Gaus. U Spher. U Wave L Wave U
' exponent = 1
2 exponent < 1
29
North Allegan!
Manistee
Allegan
Washtenaw .
Kalamazoo
Figure 1.1. Map of Michigan’s Lower Peninsula showing its location in North America
and ecoregion boundaries used for variogram analysis of gypsy moth (Lymantria dispar)
catch between 1985 and 1996.
30
200
56000
48000 ,
40000 -
32000 ,
24000 -
16000
8000
0.7,. ,7
0 100 200
1987
300
1989
300
1986
35000
30000
25000
20000
15000
10000
5000
1988
96000 1990
84000 ‘
72000 .
60000
48000
36000
24000 .
12000
O .
Lag distance (km)
Figure 1.2. Annual semivariograms of male gypsy moth (Lymantria dispar) trap catch in
the Lower Peninsula of Michigan, 1985—1996. Circles represent the original data and
Open squares represent predicted values generated by the best—fit models. Predicted
values were generated by either the cubic (1985—1988), wave (1989), spherical (1990),
power (1991—1995), or exponential (1996) models. Note the shift in variogram shape
from bell—shaped (1985—1989) to linear (1990—1995) to an untrended pattern of
increasing semivariance and leveling off at an asymptote.
31
56000 :
49000 -
42000 '
35000~"
28000 -
21000 .
14000 -
7000 a
0--,,-
0 100
1991
35000*
30000*
25000 ‘
20000 ~
15000
10000 -
5000 “
1993
Senfivafiance
0 100
21000
18000
15000 4
12000 «
9000 A
6000 J
3000 -
1995
0 100
Figure 1.2 (cont’d).
120000 ~
105000 .
90000 .
75000 ~
60000 .
45000 -
30000 ~
15000 -
0--- -- A ~-
200 300 0 100 200 300
1992
40000
35000 -
30000 -
25000 -
20000 ~
15000 -
10000 .
5000 -
O - ,,_, . .. if z.
200 300 0 100 200 300
1994
35000 ' 1996
30000 ~
25000 —
20000 ~
15000 .
10000 ~
5000 ~
200 300 0 100 200 300
Lag distance (km)
32
300000
250000
200000 ~
150000-
100000
50000 ~
0 no,
1985 1987 1989 1991 1993 1995
Ybar
Total trap catch
Figure 1.3. Total annual male gypsy moth (Lymantria dispar) catch in the Lower
Peninsula of Michigan between 1985 and 1996. Arrows indicate years in which
variogram patterns changed from bell—shaped to linear (right—facing arrow) and from
linear to typical increase to asymptote (left—facing arrow). Note that inspection of the
time series alone (without consideration of spatial location of data points) would indicate
that the invasion had stabilized by 1990.
33
120000 +HighPlains
100000 «
80000 4
60000 -
40000 ,
1985 1987 1989 1991 1993 1995
_= .
.3 35000 -- ° sag” Bay
§ 30000 - ”A Ammc
3 25000 a +Hm°n
a; 20000 .
Q A
35 15000 ‘
.§ 10000 ~ ‘j ‘ 2A ‘ 4, ‘
5 5mm— //
a - ,
+5 0 A A _, _ 7, z 7 “,7 a- - ”-2 ,, -
[-1
1985 1987 1989 1991 1993 1995
' —o—North Allegan
40000 I .g—t—SouthAlleJ:
30000 ‘ —x—Kalarmzoo
20000 -
10000 a
I
O -‘ ___. __" 7 A A A 7.,
1985 1987 1989 1991 1993 1995
Year
Figure 1.4. Total annual male gypsy moth (Lymantria dispar) trap catch in each often
ecoregions of the Lower Peninsula of Michigan between 1985 and 1996.
34
CHAPTER 2
THE RELATIONSHIP BETWEEN INVASION SUCCESS AND LANDSCAPE
HETEROGENEITY: A CASE STUDY OF THE GYPSY MOTH IN MICHIGAN
Introduction
Dispersal success is a measure of the ability of an organism to move away from
its site or origin and arrive in a patch of suitable habitat. The success of dispersal events
is thought to be affected in part by the structure and composition of the landscape across
which organisms move (Doak et al. 1992, King and With 2002). Individual-based
simulation models Of dispersal have shown that as much as 89% Of the variation in
dispersal success can be accounted for by differences in size and isolation of habitat
patches (Gustafson and Gardner 1996). The relationship between dispersal success and
landscape structure is of great conservation importance because habitat loss and
fragmentation may inhibit the dispersal or migration of organisms across the landscape
(Schwartz 1992, Malanson and Cairns 1997, Collingham and Huntley 2000, Mennechez
et a1. 2003).
Two different methods have been used to quantify the relationship between
dispersal success and landscape heterogeneity. The first method tests the linear
relationship between dispersal success and landscape structure metrics in order to identify
negative effects of habitat fragmentation on dispersal success. Either correlation
coefficients are calculated directly, or dispersal success is regressed on several landscape
structure metrics individually (Schumaker 1996, Li and Wu 2004). In general, patch-
based metrics of landscape structure such as connectivity and patch cohesion are
35
positively correlated with dispersal success (Taylor et al. 1993, Schumaker 1996),
whereas metrics that characterize the gap structure of the landscape such as lacunarity
(Plotnick et al. 1993, Plotnick et al. 1996, Dong 2000) are negatively correlated with
dispersal success.
The second method involves identifying potential thresholds (sudden, nonlinear
changes) in dispersal success and landscape structure with habitat loss; the levels of
habitat loss at which dispersal and landscape structure thresholds occur are then
compared (With and Crist 1995, With and King 1999a, Collingham and Huntley 2000).
Landscape metrics that exhibit thresholds most similar to thresholds in dispersal success
are thought to be the best measurements for predicting the effects of habitat
fiagmentation on dispersal success (With 2002). Dispersal success thresholds have been
observed in model simulations and empirical data (Wiens and Milne 1989, With and Crist
1995, Schumaker 1996). Landscape structure measurements have been calculated using
either real landscapes (classified land cover maps, Gardner et al. 1987) or computer-
generated neutral landscapes (With and King 1999a). Neutral landscapes are typically
binary (suitable or unsuitable) habitat maps generated using simple random or fractal
algorithms; these landscapes are called “neutral” because they do not contain the effects
of topography, disturbance history, climate, or ecological processes (Gardner et al. 1987,
With and King 1999a, With 2002).
Although dispersal success and many landscape structure metrics have been
shown to display threshold responses to habitat loss, the critical proportion of habitat at
which such thresholds occur (pcrit) is unknown (Andren 1994, With and Crist 1995,
With and King 1999b). Percolation models, based on the physics of liquid flow (Stauffer
36
1985), have been used to simulate random dispersal of organisms across a random
landscape. Percolation models predict that when proportion of habitat (p) is reduced
below pcrit = 0.59, a small loss of habitat will result in a disproportionately large decline
in dispersal success (With and Crist 1995). In other words, when < 59% of the landscape
is composed of suitable habitat, dispersal success drops dramatically because large
habitat patches that span most of the landscape and aid dispersal begin to break down into
many small habitat patches (O'Neill et al. 1988, Andren 1994, With and Crist 1995).
Thus, habitat fragmentation has a relatively small effect on dispersal success when p >
0.59; once pcrit is reached, habitat fragmentation compounds the negative effects of
habitat loss on dispersal success (Wiens et al. 1997, With and King 1999b). However, a
much lower value ofpcrit = ~0.05-0.1 has been predicted by models of dispersal across
more complex neutral landscapes (Gardner et al. 1987, With and King 1999a, With
2002).
Many landscape metrics also exhibit threshold responses to habitat loss that are
comparable to thresholds in dispersal success (With and King 1999a, With 2002). With
and King (1999a) used dispersal models across neutral landscapes to compare thresholds
in simulated dispersal success with thresholds in six landscape metrics: landscape
connectivity, average distance between patches, size of largest patch, total number of
patches, total length of edges, and lacunarity. The only landscape metric that displayed a
critical threshold value similar to that of dispersal success was lacunarity (landscape
“gappiness”) which increased suddenly below pcrit = ~0.05-0. 1. The remaining patch-
based metrics exhibited either a peaked response (total number of patches and total length
of edges) or critical threshold values at p > 0.3 (landscape connectivity, average distance
37
between patches, and size of larges patch). Thus, the authors concluded that lacunarity
(gap structure) may be more important than patch structure in determining dispersal
success (With and King 1999a).
To date, the relationship between real landscape structure and the actual dispersal
success of an invasive species has yet to be investigated. Although computer simulations
and experimental studies of individual dispersal movements are important to the
development of dispersal theory, such data are not easy to obtain across a large spatial
extent or over long time periods and, thus, are not practical for use in invasion
management. In order to study the effects of landscape structure on biological invasions,
measurements of invasion progress must be collected systematically across real
landscapes throughout the course of an invasion. One such monitoring program tracked
the invasion of gypsy moths (Lymantria dispar) across Michigan (Gage et al. 1990, Yang
et al. 1998). The gypsy moth is an exotic insect native to Europe that has caused
extensive defoliation across much of the eastern United States since its introduction to
Massachusetts in 1868 (Elkinton and Liebhold 1990). The first breeding population of
gypsy moths in Michigan was the result of an independent introduction in 1954 and a
second possible reintroduction occurred near Midland in the 19803 (O'Dell 1955, Hanna
1981). Gypsy moths quickly spread across the state, reaching most areas of the Lower
Peninsula by the early 19903 and the western Upper Peninsula by the late 19903 (Lele et
al. 1998, Yang et al. 1998).
The goal of this chapter was to quantify the relationship between progression of
the gypsy moth invasion wavefront and the structure of Michigan’s landscape. My
objectives were to (1) test the linear relationship between dispersal success and seven
38
landscape structure metrics, (2) identify critical thresholds in invasion success and
landscape structure with habitat loss, (3) compare empirical critical threshold values to
the predictions of percolation and neutral landscape theory (p = 0.59 and p = 0.05-0.1,
respectively), (4) explore the effect of scale on my results by repeating all analyses at
three spatial scales, and (5) compare analyses that used landscape metrics calculated from
two different classified land cover maps of Michigan.
Methods
Invasion data
Pheromone—baited traps were placed in sections eight and 26 of every township in
Michigan over 12 years from 1985 to 1996 (Gage et al. 1990, Yang et al. 1998). At each
trap, the total annual male moth catch and the trap location in UTM coordinates were
recorded. Traps were monitored between 1985 and 1996 in Michigan’s Lower Peninsula
and between 1986 and 1996 in Michigan’s Upper Peninsula. A subset of 1,090 traps was
selected for this analysis from over 3,000 placed in a regular grid across the state. Only
traps Operated all 12 years at the same location were selected to avoid change of support
in the analysis (Isaaks and Srivastava 1989). Traps located in the Upper Peninsula were
excluded because the invasion was still in the early stages by 1996.
Land cover data
Analyses of landscape structure were repeated using two 30 m resolution raster
images representing the land cover of Michigan. The first map (Map 1) was the
Michigan Resource Information System statewide land cover classification (Michigan
39
Department of Natural Resources. 1999). Map l was derived from 1978 color-infrared
aerial photographs and depicts 52 categories of urban, agricultural, wooded, wetland, and
other land cover types. The second map (Map 2), the 2001 Michigan Gap Analysis
Project land cover image created for the Michigan Department of Natural Resources
(Donovan et al. 2004), was derived from the classification of Landsat Thematic Mapper 5
and 7 imagery collected during spring, summer, and fall from 1999 to 2001. This map
depicts 32 categories of urban, agricultural, wooded, wetland, and other land cover types.
Gypsy moths are polyphagous herbivores that prefer oaks (Quercus spp.) and
aspens (Populus spp.) but will eat a wide variety of other deciduous tree species as well
(Elkinton and Liebhold 1990). Therefore, each land cover map was reclassified so that
all deciduous forest cover classes were combined to represent gypsy moth habitat. All
remaining types of land cover were considered unsuitable for gypsy moths. This simple
reclassification allowed me to explore broad patterns of structure in gypsy moth habitat
across the landscape without complicating the analysis with more detailed (and Often less
accurately classified, Donovan et al. 2004) distinctions among deciduous tree
communities or species (Li and Wu 2004).
Scale of analysis
I repeated analyses of landscape structure and invasion success at three different
spatial extents to assess the effect of scale (Gamder et al. 1989, Doak et al. 1992, Li and
Wu 2004). Because female gypsy moths are incapable of flight and male moths
generally do not disperse beyond 1 km/year (White et al. 2003), it was assumed that the
invasion wavefront was driven largely by wind-dispersed larvae that are passively
40
transported up to ~40 km away from their hatching site (Elkinton and Liebhold 1990).
Although a small percentage of larvae may actually survive such a long trip, occasional
long-distance dispersal events (in combination with human-aided dispersal of egg
masses) likely explain the high observed expansion rates of 20-40 km per year in some
areas of the country (Taylor and Reling 1986, Liebhold et al. 1992). Therefore, analysis
windows of 75, 45 and 15 km on each side were chosen, representing spatial extents that
were slightly larger, slightly smaller, and roughly equal to the spatial scale at which the
invasion process was likely occurring.
At each scale of analysis, the Lower Peninsula of Michigan was clipped into
multiple subsections using square-shaped analysis windows (the shape required for
lacunarity calculations). At each scale of analysis, the maximum number of windows
that could fit inside the Lower Peninsula of Michigan were created and aligned so as to
maximize the number of traps included in the analysis. Altogether, 12 boxes 75 km on
each side, 37 boxes 45 km on each side, and 207 boxes 15 km on each side were created
and used to clip each land cover map.
Characterizing the relationship between invasion success and landscape structure
Time series in total annual catch at each trap revealed a dramatic increase (from
tens to hundreds or thousands) in the number of moths caught at most traps once about 25
individuals had been caught in a given year. Therefore, year of colonization was defined
as the year in which total trap catch reached 25 or greater individuals. To obtain a rough
estimate of how close each trap was located to the initial site of gypsy moth introduction
in Michigan, the distance from each trap to a trap located near the Midland and Bay
41
County border was calculated. At each scale of analysis, traps were grouped by analysis
window and summary statistics were calculated, including average years to colonization,
invasion rate (distance from original introduction site/average years to colonization), and
variance and range in years to colonization. These summary statistics were considered
potential indices of invasion success.
Analysis windows were used to clip out subsections of both land cover maps for
landscape metric calculations. Each clipped land cover raster file was converted to an
ascii text file and imported into the software package APACK Version 2.22 (Mladenoff
and DeZonia 2002). Within APACK, proportion of gypsy moth habitat (p) and multiple
landscape structure metrics were calculated for each subsection of the landscape
including lacunarity, total number of habitat patches, size of largest habitat patch, total
length of edges, average area per patch, fractal box dimension, and centroid connectivity.
Lacunarity was calculated using moving window sizes of l, 5, 10, 50, 100, 150, 200, and
250 cells.
Once all indices and metrics were calculated, invasion success was regressed on
landscape structure metrics individually to test for linear relationships between invasion
wavefront movement and measurements of habitat fragmentation. Next, all indices and
landscape structure metrics were plotted individually against p to identify potential
threshold responses to decreasing proportion of habitat in the landscape. To identify the
presence of a threshold response, I looked for large declines or increases in the slope of
the plot. Subsets of 5, 11, or 31 successive points (at the 75, 45, and 15 km scales of
analysis, respectively) were created for each plot. I then Obtained a rough estimate of the
derivative at the midpoint of each subset by estimating the slope of a simple linear
42
regression line through the subset of points. For all decreasing monotonic curves,
thresholds were defined as the value of p corresponding with the subset midpoint at
which the regression slope declined below 10% of the maximum. For all increasing
monotonic curves, thresholds were defined as the value Of p corresponding with the
subset midpoint at which the regression slope rose above 10% of the maximum. For all
peaked curves, thresholds were defined as the value of p corresponding with the subset
midpoint at which the sign of the regression slope switched consistently from positive to
negative. Inverse predictions were made to calculate standard errors and 95% confidence
intervals for each threshold (N eter et al. 1985).
Results
Relationship between invasion success and habitat loss
Proportion of gypsy moth habitat (p) ranged from: 0.13-0.54 and 0.15-0.50 for
Maps 1 and 2 in 75 km landscapes; 0.09-0.55 and 0.11-0.58 for Maps 1 and 2 in 45 km
landscapes; 0004-071 and 0.02-0.70 for Maps 1 and 2 in 15 km landscapes. No indices
of invasion success displayed a threshold response to decreasing proportion of habitat.
Among the four indices of invasion success calculated, only range in years to
colonization declined significantly (a = 0.05) with increasing p in each landscape (Figure
2.1). Because the pattern of increase in range in years to colonization with decreasing p
was similar between land cover maps, only Map 1 results are presented.
Relationship between landscape metrics and habitat loss
Most landscape metrics exhibited a threshold response to changing p with both
land cover maps and at all three scales of analysis (Table 2.1 and Figures 2.2-2.4).
43
Because the overall behavior of landscape metrics in response to increasing p was similar
between land cover maps, only Map 1 results are presented. However, several large
differences in computed thresholds were observed between the two maps for size of
largest patch (at 75 and 45 km), average area per patch (at 75, 45, and 15 km), and
centroid connectivity (at 75 and 15 km). Two-fold increases in patch metric thresholds
were typically observed for the 75 and 45 km analyses, indicating that Map 2 displayed a
sudden increase in habitat fragmentation at a higher value of p than Map 1. In contrast,
thresholds occurred at a lower p for Map 2 than Map 1 at the 15 km scale of analysis.
Several landscape metrics exhibited threshold responses that were less sudden but
similar in location to thresholds predicted by neutral landscape models (pcrit = ~ 0.05-
0.1). Most analyses of size of largest patch and average area per patch displayed critical
thresholds at or near p = 0.1 (Table 2.1). Also, at the 15 km scale of analysis, lacunarity
exhibited critical thresholds of 0.1 and 0.153 for Maps 1 and 2, respectively; although
only thresholds for box size 1 are graphed, similar thresholds were observed for box sizes
5, 10, and 50. Only centroid connectivity calculated using Map 2 at the 75 km scale of
analysis exhibited a threshold (pcrit = 0.51) similar to that predicted by percolation
theory (pcrit = 0.59); however, a low pcrit = 0.09 was also observed at the 15 km scale of
analysis for Map2.
All other landscape metrics either displayed thresholds at p > 0.1 or no detectable
threshold behavior. Most lacunarity analyses either did not exhibit detectable threshold
responses to habitat loss or exhibited a threshold (pcrit = 0.44 at the 45 km scale of
analysis for Map 1) larger than that predicted by neutral landscape models. Total number
of habitat patches on the landscape peaked as expected for both land cover maps between
44
p = 0.22 and 0.30 for all maps and scales of analysis with the exception of Map 2 at 15
km which exhibited no observable threshold. Total length of edges also peaked at about
p = 0.26-0.37. Fractal box dimension began to level off at about p = 0.40-0.43 at the 15
km scale of analysis for both maps. However, this metric did not exhibit a critical
threshold for either land cover map at the 75 and 45 km scales of analysis.
Relationship between invasion success and habitat fragmentation
Although no thresholds were detected in measures of invasion success, range in
years to colonization is likely negatively affected by habitat fragmentation because it was
linearly related with several measures of landscape structure (Figures 2.5-2.7). Similar
results were observed for Map 2, so only Map 1 results are presented. As might be
expected, range in years to colonization declined significantly (a = 0.05) with size of
largest patch, average area per patch, fractal box dimension, and connectivity at the 75
km scale of analysis for Map 1; also, range in years to colonization increased
significantly with increasing lacunarity as expected (Figure 2.5). Total number of
patches and total length of edges did not exhibit a significant linear relationship with
range in years to colonization as might be expected given their peaked response to
declining habitat (Figure 2.2-2.4). Results for Map l at the 45 km scale of analysis
(Figure 2.6) were similar to those obtained at 75 km. However, at this scale, range in
years to colonization declined with increasing lacunarity; also, range in years to
colonization was found to increase slightly with increasing total number of patches.
Finally, results for Map 1 at the 15 km scale of analysis (Figure 2.7) were similar to those
obtained at the 75 km scale of analysis with the exception that range in years to
45
colonization declined slightly with increasing total length of edges. Also, range in years
to colonization was found to increase slightly with increasing total number of patches and
decrease slightly with total length of edges. Although many of the 45 and 15 km
regression results exhibited significant slopes, most linear relationships were weak (R2 <
0.2) and several of the 15 km analyses were likely influenced by extreme outliers.
Discussion
Efibcts of habitat loss and fragmentation on invasion success
Contrary to the predictions of percolation and neutral landscape theory, the gypsy
moth invasion wavefront exhibited a linear response to changes in landscape structure
across the Michigan landscape. Percolation and neutral landscape models predict a
sudden decline in dispersal success at p = 0.59 (O'Neill et al. 1988) and p = 0.05-0.1,
respectively (With and King 1999a). However, range in years to gypsy moth
colonization declined linearly with increasing proportion of habitat in the landscape
(Figure 2.1), indicating that habitat fragmentation did not compound the negative effects
of habitat loss.
Although thresholds in invasion success were not detected, the amount of time
necessary to complete the invasion of a given area was likely increased by habitat loss
and fragmentation. Range in years to colonization was significantly correlated with
several landscape metrics at all scales of analysis (Figures 2.1, 2.5-2.7). For example,
landscapes with p > 0.4 in Figure 2.1a represent areas of the northern Lower Peninsula
where the number of years to colonization were all small even though these traps were
located relatively far away from the site of original gypsy moth introduction. Landscapes
46
with p < 0.4 represent areas from across the peninsula that were not invaded as
uniformly; in other words, some traps in a given analysis window were invaded early in
the monitoring period while others were not colonized for up to 11 years later.
Therefore, time to complete colonization of a given area (i.e. uniformity of the invasion)
was likely slowed by habitat loss. Also, invasion success decreased with increasing
habitat fragmentation. Specifically, range in years to colonization increased as size of
largest patch, average area per patch, connectivity, and fractal box dimension decreased;
also, range in years to colonization typically increased with increasing lacunarity. The
only evidence for a non-linear threshold response of invasion success to changes in
habitat fragmentation may be the curvilinear decline of range in years to colonization
with increasing landscape connectivity, average area per patch, and possibly size of
largest patch at the 15 and 45 km scales of analysis. Overall, most relationships between
invasion success and landscape metrics were linear. Thus, for passive dispersers like the
gypsy moth, habitat fragmentation may slow the invasion wavefront but not cause a
sudden, nonlinear decline below a critical level of habitat loss.
One potential reason for the apparent discrepancy between these results and many
published dispersal model predictions may be that this study measured large-scale
movement of the invasion wavefront instead of relatively small-scale movements of
individual dispersers. Most research conducted on the relationship between dispersal and
landscape structure has involved either individual-based simulation models (Schwartz
1992, With and Crist 1995, Pitelka 1997, With and King 1999a, Matlack and Monde
2004) or experiments documenting short distance, terrestrial movements of beetles
(Wiens and Milne 1989, Wiens et al. 1997). My study differs from all others conducted
47
to date in that it tracks the movement of the invasion wavefront and not the movement of
individual dispersers; sudden increases in gypsy moth trap catch data indicate that the
invasion wavefront (including larvae and non-vagile adult females) recently arrived into
the area near that trap, not that a large number of adult male moths have recently flown
into the region. Therefore, the index of invasion success used in this study may be
displaying a macroecological phenomenon, an emergent property of the combined
individual dispersal movements of all individuals in the population (Brown 1995, Maurer
1999). Emergent properties are common features of large, complex systems, but they are
not observable in data collected at smaller scales. Therefore, data collected at the scale of
individual dispersal movements may not be sufficient to predict the relationship between
overall invasion success and landscape structure.
Alternatively, the relationship between invasion success and landscape structure
may be linear because it is strongly affected by non-random movement and
environmental factors not accounted for in most simulation models. Percolation models
assume organisms move randomly across a landscape composed of randomly distributed
patches of suitable habitat (Wiens et al. 1997); similarly, neutral landscapes do not
represent the true complexity of real landscapes and their associated dispersal models
assume relatively simple movements (With and King 1999a). However, gypsy moths are
carried passively by directed winds that do not allow for purely random dispersal
movement (Elkinton and Liebhold 1990). Increasing habitat fragmentation will, of
course, have the effect of increasing the probability that gypsy moth larvae will be
deposited by atmospheric motion systems or rainfall events into unsuitable areas and fail
to survive (Isard and Gage 2001). Anemochorous species like the gypsy moth (and,
48
perhaps, species that actively fly) may not be as negatively affected by the same loss of
habitat and patch connectivity as active terrestrial dispersers. Also, percolation and
neutral landscape models do not consider the effects of environmental conditions such as
topography, disturbance history, climate, or ecological processes such as competition and
predation (Gardner et al. 1987, With and King 1999a, With 2002). Therefore, real
invasion data may not exhibit threshold responses because environmental factors may
mediate the effects of habitat fragmentation. The exact response of a species to the
landscape likely depends on a number of additional factors such as dispersal ability,
degree of habitat specialization, and rate of habitat turnover in dynamic landscapes (With
and Crist 1995, Isard and Gage 2001, King and With 2002, Matlack and Monde 2004).
Effect of scale of analysis and [and cover map
Increased variation in range of years to colonization (Figure 2.1) and weak
correlations between invasion success and landscape structure (Figures 2.5-2.7) at the
two smaller analysis scales (45 and 15 km) suggest that movement of the invasion
wavefront is likely occurring at a large spatial extent compared to the majority of
individual moth movements. By comparing results from different scales of analysis, a
breakdown in ability to characterize invasion success was detected when analysis
windows of < 75 km on a side were used. Even though most gypsy moth larvae do not
disperse 40 km (Elkinton and Liebhold 1990), a small number may survive long distance
dispersal events. Those individuals may then begin forming their own new colonies, or
“nascent foci” (Moody and Mack 1988), ahead of the invasion wavefront that eventually
merge with the wavefront. The formation of nascent foci typically results in much higher
49
rates of expansion (Yang et al. (1998) reported 6 km/year in Michigan) than would be
expected given the average individual dispersal distance (Moody and Mack 1988,
Shigesada et al. 1995). With (2004) has suggested that landscape pattern is less crucial
for predicting colonization success if even a few long-distance dispersal events are
successful. However, these results indicate gypsy moth invasion success responds to
landscape structure at large spatial extents; thus, occasional long-distance dispersal
events and their interplay with landscape structure may be vital to our understanding of
the invasion process. If the effects of landscape structure on invasion progress are to be
quantified, invasions may need to be monitored, and landscape metrics calculated, at the
spatial extent of long-distance dispersal events.
Observed thresholds in landscape metrics were remarkably similar between the
two land cover maps of Michigan for several landscape metrics despite the fact that these
maps were generated almost 20 years apart using different types of imagery and
classification methodology; however, three patch-based metrics (size of largest patch,
average area per patch, and centroid connectivity) differed greatly between maps. One
reason these metrics are sensitive to choice of land cover maps is that they all measure
the exact size of patches in the landscape and the two maps used in this study differed in
their characterization of patch size. In general, Map 2 paints a much patchier picture of
the Michigan landscape than Map 1. For example, the total number of patches/1 ,000
ranged from 6.34-9.45 for Map 1 and 33.8-102.2 for Map 2 on 75 km landscapes. This is
likely due to the fact that Map 2 was generated from satellite imagery using a complex
classification methodology and because habitat fragmentation likely increased in
Michigan between 1978 and 2001. Threshold detection appeared to be reasonably
50
resilient to differences in the magnitude of other landscape metrics that are less
dependent on patch size. However, my choice of land cover map would have had a more
obvious effect had the landscape not been reclassified to reflect either habitat or non-
habitat. Landscape metrics would have been quite different and more complicated to
interpret had even a few of the original 52 land cover classes in Map 1 or 32 classes of
Map 2 been retained in the analysis.
Management implications
Gypsy moths did not exhibit a threshold response to declining proportion of
habitat, indicating that habitat fragmentation did not compound the negative effects of
habitat loss for this species. Thus, gypsy moths may be more resilient to habitat
fragmentation than expected. Although habitat fragmentation likely slowed the invasion,
fragmentation did not prevent the invasion from reaching all areas of the Lower
Peninsula in a relatively short period of time (<10 years in most cases). Therefore, using
habitat fragmentation as a “fire-break” to limit invasive spread as suggested by Sharov
and Leibhold (1998) and With (2004) will probably not be a successful long-term
strategy for generalists like the gypsy moth that show great dispersal capacity.
This study also indicates that lacunarity alone may not be sufficient for predicting
the response of invasive species to landscape heterogeneity as has been suggested by
With and King (1999a, With 2002). First, invasion success did not exhibit threshold
behavior similar to that of lacunarity (p = 0.05-0.1), suggesting that invasion success is
not affected by habitat loss in the same fashion as lacunarity. Second, my results show
that lacunarity was highly sensitive to scale of analysis, indicating that the effects of
51
habitat gaps on invasion success may be different at different scales. Lacunarity only
displayed thresholds at p = 0.05-0.1 at the 15 km scale of analysis; at larger spatial
extents, thresholds in lacunarity were either not observed, or were detected at p = 0.44.
The close correspondence between dispersal success and lacunarity thresholds observed
by With and King (1999a) may only be generated by processes occurring at spatial
extents similar to individual movements; this relationship may not “scale up” to the
movement of an invasion wavefront (Li and Wu 2004). Another important result
regarding lacunarity is that it was not the only metric that exhibited thresholds near the
dispersal success threshold of p = 0.1 reported by With and King (With and King 1999a).
Several patch-based metrics such as size of largest patch, average area per patch, and
connectivity also exhibited thresholds at p = ~0.1. Regression analyses also indicated
that indices of patch size and distribution may be just as good at describing the negative
effects of habitat fragmentation on invasion success as lacunarity. Lacunarity should be
used with caution given that the correlation between lacunarity and range in years to
colonization vacillated from positive to negative as analysis window size decreased
(Figures 2.5-2.7). In contrast, patch—based metrics such as size of largest patch, average
area per patch, and connectivity were negatively correlated with range in years to
colonization at all scales of analysis. Given the problems associated with the lacunarity
metric (especially at larger scales of analysis), I suggest that lacunarity not be used as the
primary or sole predictor of an invasive species’ response to landscape structure.
This study also suggests that the scale of data collection and analysis must be
carefully matched to the scale of the process of interest if invasion management is to be
effective. Although studies of individual dispersal movements are important to the
52
development of dispersal theory, such data are not easy to obtain across a large spatial
extent or over long time periods and, thus, are not practical for use in invasion
management. Therefore, studies conducted on individual dispersal movements may not
be applicable to processes occurring at much larger spatial extents and should be used
with caution when planning management actions.
53
Table 2.1. Summary of critical thresholds in landscape metrics with habitat loss at three
scales of analysis (i.e. analysis window sizes of 75, 45, and 15 km on a side) using two
land cover maps. Dashes indicate no threshold response was detected.
75 km 45 km 15 km
Map l Map2 Map 1 Map 2 Map 1 Map 2
. 0.44 0.1 0.15
Lacunamy ' ' i 0.02 i 0.001 1» 0.08
Total number 0.28 0.27 0.23 0.30 0.22 _
ofpatches a: 0.36 i 0.63 i 0.76 i 0.81 : 15.12
Size of largest 0.15 0.27 0.14 0.27 0.12 0.18
patch i 0.10 i 0.08 i 0.11 i" 0.13 i 0.002 :t 0.98
Total length of 0.28 0.27 0.32 0.37 0.315 0.26
edges 3: 1.4 i 2.2 i 1.84 x 1.68 i 0.36 i 5.12
Average area 0.15 0.22 0.14 0.28 0.22 0.09
per patch : 0.11 i 0.11 i 0.09 i 0.13 i 0.004 :t 0.08
Centroid 0.28 0.51 0.42 0.42 0.31 0.09
connectivity i 0.15 i 1.2 i 0.1 i- 0.12 :t 0.005 :r 0.11
Fractal box _ _ _ _ 0.43 0.40
dimension i 0.000] t 0.06
54
a) 75 km landscapes
. y=-19.71x+ 12.18
R2 = O.7l,p = 0.001
0
0 - b) 45 km landscapes
o o y= -6.42.x+ 5.9
. . . R2=0.19,p=0.004
Range in years to colonization (years)
0.
8-
A
v I
0 0.1 0.2 0.3 0.4 0. 0.6
c) 15 km landscapes
0 y=-3.13x+2.61
. . . . R2=0.14,p<0.0001
0
.0 0.2 0.4 0.6 0.8
Proportion of habitat
Figure 2.1. Range in years to colonization among traps located in the same analysis
window declined with increasing proportion of habitat. Results are reported using Map l
for each scale of analysis: a) 75 km, b) 45km, and c) 15km.
55
Lacunarity
H N b) «P U! ON \J 00
Total number of patches/1000
«BMQQOOOO
o
o
0.6 0.1 0.2 0.3 0.4 0.5 0.6
O
o
N
.o .
0)
O
4:.
O
u:
O
O
2 2500 1
$2
‘55 2000 i 0
CL
§ . .
L5 1500
.2
1;; 1000 1 ° .
Q)
9
E 5001 . o
g..— 0
O
a
0.1 0.2 0.3 0.4 0.5 0.6
800 ~ , 1.4 1
750 l 1.21 '
‘5 600 - ’ 5 '8‘
its 3 0.6
E 550 ~ g ..
500 4 U 0.4
450 ‘ 0.. 0'2 .~ . o o .
400 . . . . 0.0 e . . . .
0,1 0,2 0,3 0,4 0,5 0.1 0.2 0.3 0.4 0.5 0.6
Proportion of habitat
Figure 2.2. Thresholds in landscape metrics with increasing proportion of habitat
calculated using land cover Map 1 and a 75 km analysis window.
56
1600 ~
1400 * o
1200 ‘
1000 1
600 1
400 1 ‘
200 1 .. . ° '
Average area per patch
00
o
o
.
0.1 0.2 0.3 0.4 0.5 0.6
1.95
1.90 l 0‘
1.85 ‘
1.80 <
1.75 * .'
Fractal box dimension
0
1.70 . , . . a
0.1 0.2 0.3 0.4 0.5 0.6
Proportion of habitat
Figure 2.2 (cont’d).
57
12 - g 3.5 .
10 ‘ . E 3.0 . o
' a
.Q 8i \ 132.51 0" '. o
a . LE.- 0 . o .
g 6 1 ‘ O 2.0 ‘ '0 ‘
o .. h o . .
(U 4 . B 15 4 o .. O
.4 4 s. . E . . fl"
2 '-~.._§1.0‘ a...
S o
0 1 . . . . ES 0.5 - . . r 4 .
0.1 0,2 0,3 0,4 0,5 01 0.2 0.3 0.4 0.5 0.6
O
O
S 1200 ~
g .
.5 1000 ~
Q. o
2 800 < . .
E 600*
E20 4001 . ° 3 . -
:5 200 l :' . . .
0&1) 0 -L~—‘le ', . '. . . . .
E7: 0.1 0.2 0.3 0.4 0.5 0.6
350 ' 3.01
300 ‘ o 0 . 2.5 . . °
C . O . >.. .0
8 250 * 0 ' ° .0 o E 2'0 o 0
3.1.200 5:. ’ 8"5‘ . .
'0 c
m o. 8 1.0 i . '
150 ~ . . . r
‘ ' 0.5 . g .- .
100* ' 1 , , . 00 -',"'"';° . . . .
0.1 0.2 0.3 0.4 0.5 0.1 0.2 0.3 0.4 0.5 0.6
Proportion of habitat
Figure 2.3. Thresholds in landscape metrics with increasing proportion of habitat
calculated using land cover Map 1 and a 45 km analysis window.
58
1800 .
1600
1400
1200
1000
800 < . .. ' .
600 .
400
200
0
Average area per patch
0
O
_4_42
I
O
1.95 «
1.90 4 4,...-
l.85 - 0.
1.80 - '-
1.75 - . '
1.70 « . "
l.65«
1.60 a . . . . .
0.1 0.2 0.3 0.4 0.5 0.6
Fractal box dimension
Proportion of habitat
Figure 2.3 (cont’d).
59
250 g 0.6
200 1 230.5 J .
3 0.4 ...o 0
g 150 o. '. . '. .' '
a “03‘ *4? "° "
o 100 802 .p g .l. .u
:3 .8 . ‘3‘ ..:§ ’0'.” '
50 , 3 50.1 1 ‘0'. o . 0 s":
, . v . a .
0.0 02 04 06 00 0.2 04 06 08
8 200 '
2123‘ =
E 140 - .354.
‘3 120 ‘ 3'
Q 00. o
1;; 100 ‘ o . . ‘
15° 28‘ -'§"=--
c6 . o. 9
"" o. . 0.8'
95 40 0. .‘.
Q)
E
CD
400 -
Connectivity
p—s
O
O
300 .
200 .
0.0 0.2 0.4 0.6 0.8
. MW”.
0.0 0.2 0.4 0.6 0.8
Proportion of habitat
Figure 2.4. Thresholds in landscape metrics with increasing proportion of habitat
calculated using land cover Map 1 and a 15 km analysis window.
60
50-
0000
4321
83835.
0.6 0.8
0.4
0.2
0.0
10000 1
8000 .
6000 *
4000 ~
2000 ~
:83 Ba 83 owmco><
0.4 0.6 0.8
0.2
0.0
l
0.8.6.
211
865::
4
1
cxo
1.2 1
n
F—d
0.
1
308m
8.
O
0.2 0.4 0.6 0.8
Proportion of habitat
0.0
Figure 2.4 (cont’d).
61
12~
10*
Range in years to colonization (years)
12 « y = 0.74x + 0.20
010-50500
° R2=0.06, =0.49 -
10 . p .
8 « o o
6 /
4 q o o
y= l.l9x+ 1.12 2‘ . .
R2=0.53,p=0.01 ,
. . . a a . 0 . . .
3 4 5 6 7 8 6 7 8 9 10
Lacunarity Total number of patches/1 000
12 1 )1); -0.004x+ 8.44
' = .7, = .008
10 . . 05 p 0
8 .
6 .
4
2 .
0 . , . . .
0 500 1000 1500 2000 250(
Size of largest patch / 1000 (m2)
12
y=-0.015x+ 15.34
1 R2 = O.27,p= 0.10
V
400 450 500 550 600 650 700 750 800
Total length of edges/1000
Figure 2.5. Relationship between range in years to colonization among traps located in
the same analysis window and landscape metrics calculated using land cover Map 1 and a
75 km analysis window.
62
. y = -0.01x + 9.49
10 1 . R2=0.74,p=0.001
8 1
6 .
4 .
2 .
0 . . . .
0 400 800 1200 160(
A Average area per patch (m2)
g 12 ' )1); -6.78x + 8.09
.5 10 q . = 0.50, p = 0.02
E; o
'E 8
.9.
8 6~
8 4
8
>\ 2 ~ 0
.E
«in 0 , . . . . . —.
a? 0.0 0.2 0.4 0.6 0.8 1.0 1.2 1.4
Connectivity
12 ~
10 1
8 .
6 .
4 4
y = -37.55x + 74.8
2 . R2 = 0.63, p = 0.004
0 fl
1.70 1.75 1.80 1.85 1.90 1.95
Fractal box dimension
Figure 2.5 (cont’d).
63
10 - 10 * y=2.l9x—O.15
8. , ' .y=0.266x+2,74 84 R2=0.45,p<0.000'l .-
, . R2=O.08,p=0.05
6.
4.
2.
O
p
O
p
fl
0.0 0.1 0.2 0.3 0.4 0.5 0.
r
0.5 1.0 1.5 2.0 2.5 3.0 3.5
Lacunarity Total number of patches/ 1000
10
, , y = -0.003x + 4.87
8 R2 = 0.20, p = 0.004
o . . .r , . . .
0 200 400 600 800 1000120(
Range in years to colonization (years)
Size of largest patch/1000 (m2)
10 .y=-0.002x+4.38 ..
8 fl R2=0.003,p=0.25
6 . . . .
4 .W
2 . . .. . . .
0 e
100 150 200 250 300 350
Total length of edges/1000
Figure 2.6. Relationship between range in years to colonization among traps located in
the same analysis window and landscape metrics calculated using land cover Map l and a
45 km analysis window.
64
y = -0.003x + 5.30
R2 = 0.27, p = 0.0006
0 . .. - . fl
0 400 800 1200 1600 200(
Average area per patch (m2)
y=-l.26x+5.00
8 . , 122:0.22, p= 0.002
Range in years to colonization (years)
O . . = f . . .
0.0 0.5 1.0 1.5 2.0 2.5 3.0
Connectivity
10 1
8 w . y=-9.26x+20.71
. . R2=0.12,p= 0.02
6 « o o o o
4 .
2 .
0 c
1 .6 l .7 1.8 1.9 2.0
Fractal box dimension
Figure 2.6 (cont’d).
65
_ qy=6.85x+0.14 . .
R2 = 0.18,p < 0.0001
6 1' 6 . o
4 4 +
2 y = 0.02): + 1.55
R2 = 0.03, p = 0.004 2 l
0 . ' ' . ' 0 ~
0 50 100 150 200 250 .0 0.1 0.2 0.3 0.4 0.5 0.6
Lacunarity Total number of patches/1000
8 y=-0.le+2.l
- R2=0.12,p< 0.0001
0 40 80 120 160 200
Size of largest patch/1000 (m2)
Range in years to colonization (years)
8 -
y=-0.03x+2.37 o o o
6 ‘ R2=0.02,p= 0.02
o o no
4 ~ 0 a a... o. o
0 10 20 30 40 50
Total length of edges/1000
Figure 2.7. Relationship between range in years to colonization among traps located in
the same analysis window and landscape metrics calculated using land cover Map l and a
15 km analysis window.
66
y = -0.0006x + 1.97
R2 = 0.086, p < 0.0001
0 2000 4000 6000 8000 1000'
Average area per patch (m2)
y = -0.008x + 1.74
R2 = 0.06, p = 0.03
‘-
"
fl
0 100 200 300 400
Range in years to colonization (years)
Connectivity
8 -
y = -3.27x + 7.41
6 R2 = 0.09, p < 0.0001
0.9 1 .2 1.5 1 .8 2.1
Fractal box dimension
Figure 2.7 (cont’d).
67
CHAPTER 3
PATTERNS OF SEMIVARIANCE GENERATED BY SIMPLE AND STRATIFIED
DISPERSAL
Introduction
Ecological variables may be correlated across space as a result of both population
processes and environmental conditions (Legendre 1993, Koenig 1999). However,
patterns of spatial autocorrelation are not static and may change over time as populations
fluctuate and respond to a changing environment (V illard and Maurer 1996, Koenig
1999). Brodie et al. (1995) first noted changes in spatial autocorrelation patterns as a
result of local dispersal in a study of the post—fire re—invasion of a Populus balsamifera
clone. Correlograms, plots of the spatial autocorrelation statistic Moran’s I against
distance (Legendre and Legendre 1998), revealed three distinct patterns of spatial
autocorrelation representing three stages of clonal development: post—fire colonization,
consolidation, and directional expansion (Brodie et al. 1995).
Likewise, three general patterns of semivariance were observed during the gypsy
moth (Lymantria dispar) invasion of Michigan (Chapter 1). Semivariance is a spatial
statistic used to quantify the variance of the differences between all possible points
located a given distance apart (Isaaks and Srivastava 1989); the standardized
semivariance is equal to 1 minus the correlation when the population mean and variance
remain constant across the study area (Rossi et al. 1992). In the gypsy moth invasion
study, graphs of semivariance against distance called semivariograms (or simply
variograms) were used to depict the average similarity between values based on
68
separation distance between sampling points (Chapter 1). Initially, bell— or wave—shaped
variograms were obtained as the initial population became established. As the invasion
wavefront began to move across the study area, a gradient of high to low gypsy moth trap
catch occurred that produced linear variograms. Finally, as the invasion neared
completion in the Lower Peninsula of Michigan, untrended (asymptotic) variogram
patterns were observed in which semivariance increased until it reached the maximum
distance at which data are spatially dependent and then leveled off. A similar progression
of semivariance patterns was reported by Liebhold et a1. (1991) in their study of
historical gypsy moth egg mass distributions over a large portion of northeastern
Massachusetts, southeastern New Hampshire, and southern Maine.
The fact that two species like P. balsamifera and the gypsy moth both exhibit
substantial changes in spatial autocorrelation patterning is intriguing because each species
has quite different dispersal strategies. The P. balsamifera clone studied by Brodie et al.
(1995) expanded primarily via root suckering, an example of simple dispersal. Simple
dispersal (also called neighborhood diffusion) involves the movement of dispersing
individuals to areas adjacent to their point of origin (Hengeveld 1989). In contrast, gypsy
moth larvae disperse primarily through the air via stratified dispersal, a combination of
both short and rare, long-distance movements to areas far away from the point of origin
(Hengeveld 1989). Stratified dispersal differs from simple dispersal in that occasional
long-distance (or “jump”) dispersal events sometimes result in the establishment of new
populations independent of the original population (Sharov and Leibhold 1998). To date,
the way in which dispersal strategy (simple vs. stratified) affects the spatial structure of
populations as they undergo range expansion has not been explored.
69
In this study, my goal was to determine if computer models of dispersal generate
distinct patterns of semivariance that change over time as a population grows and spreads
across the landscape. My objectives were to (l) quantify and compare changes in
patterns of semivariance over time generated by models of both simple and stratified
dispersal, (2) characterize the effects of different population growth rates and dispersal
distributions on patterns of semivariance, and (3) compare model results to empirical
patterns of semivariance observed during the gypsy moth invasion of Michigan (Chapter
1). Stratified dispersal models with long-tailed dispersal distributions typically generate
the most realistic estimates of range expansion rate (Clark 1998, Higgins and Richardson
1999, Clark et al. 2001). Thus, I predicted such stratified dispersal models would also
generate variogram progression patterns most similar to empirical patterns displayed by
gypsy moths (bell-shaped, linear, and asymptotic).
Methods
Model structure
Two models were created in R (R Development Core Team 2004) to simulate
population dynamics with either simple or stratified dispersal movements. Population
change in both models was simulated using a discrete, stochastic Ricker model with
density dependent dynamics and non-overlapping generations
Ni+l : Ni exp(r —(uNi ))exp(sz)
where is N; is the population size at time step i, r is the maximum per capita rate of
growth, u is the effect of intraspecific competition on population growth, .5 is a scale
factor that represents the size of random fluctuations in r, and z is a standard normal
70
random variable with a mean of zero and a standard deviation of 1. For this study, u and
s remained constant at respective values of 0.005 and 0.1. Each model was run twice
with a different value for r. The first value (r = 1.7) represented a rate similar to that
exhibited by gypsy moths (Sharov and Leibhold 1998); a second, higher value (r = 2.1)
was also selected to generate population dynamics with a two-point limit cycle for
comparison (Case 2000). Reproduction occurred at discrete time intervals, and each
reproduction pulse was followed by a dispersal pulse. Models were initiated with N, at
carrying capacity (K = r/u) in a suitable habitat cell at the center of the landscape.
In both simple and stratified dispersal models, the number of individuals that
would disperse a given distance away from each cell in each time step was determined,
and then the direction of dispersal was selected. To simulate density dependent dispersal,
dispersal began in each cell when population abundance reached 75% of K, allowing each
population to build up substantially before emigration occurred. For the simple dispersal
model, proportion of individuals dispersing out of a given cell was selected from a
random uniform distribution between 0 and 0.1. Dispersing individuals were then moved
one cell away in a random direction. For the stratified dispersal model, the proportion
and number of individuals dispersing a given distance was determined by one of two
gamma distributions (Figure 3.1). Distribution 1 used a rate parameter of 0.5 and a shape
parameter of 1.2; this distribution was similar to half of a high-peaked normal distribution
in that most dispersers moved short distances and few to none moved moderate to long
distances (Clark 1998). Distribution 2 was a long-tailed gamma distribution with rate
parameter (a) of 0.043 and a shape parameter (6') of 0.5; use of this dispersal distribution
resulted in most dispersers moving short distances, but relatively more individuals made
71
moderate to long dispersal jumps across the landscape compared to Distribution 1 (Clark
1998). For both simple and stratified dispersal models, movement occurred in a random
direction once dispersal distance was determined for each dispersing individual.
Reflecting boundaries were employed such that no individuals were allowed to leave the
landscape.
Both models were run on two different landscapes consisting of 400 cells. The
homogeneous landscape consisted of 100% suitable habitat and was chosen to examine
the movement of organisms that were uninhibited by habitat loss or fragmentation. The
heterogeneous landscape was a selected section (360 kmz) of the northern part of
Michigan’s Lower Peninsula consisting of 54% suitable habitat; this landscape was taken
from the 2001 GAP/IF MAP Michigan land cover image (Donovan et al. 2004) in which
all deciduous hardwood types were reclassified with a value of “1 ” and all other land
cover classes were reclassified as “0” or unsuitable habitat (Chapter 1). In summary, 12
scenarios were created by varying growth rates, landscapes, and dispersal distributions
(for the stratified dispersal model) for each of the two dispersal models (Table 1). Each
model scenario was run for 100 time steps and repeated 50 times.
Characterizing changes in semivariance
Semivariance was calculated as:
1 N(h) 2
707) - m #2109 _xi+h)
where y(h) is the semivariance for the distance interval h, N(h) is the total number of
sample pairs for the distance interval h, Xi is the measured sample value at location i, and
72
Xi +h is the measured sample value at point i+ h (Isaaks and Srivastava 1989).
Variograms were created using all data from all cells in every other time step (total of
50/simulation) with 13 lag distance bins and a maximum distance equal to the length of
one side of the landscape.
To characterize the shape of each variogram, I fit four models to the data
generated by each scenario (Table 1) using maximum likelihood estimation (Bumham
and Anderson 2002). First, a quadratic model defined as
flh) = ah2 + bh + c
was fit to the data to characterize bell-shaped variograms; h represents the distance lag
class, c the intercept, and a and b the slopes for the second- and first-order terms,
respectively. To characterize linear variograms, I fit a linear model
flh) = bh + c
to the data. I also fit a horizontal line to the data in which the starting value for the
constant c was set to the mean semivariance
701) = c
in order to characterize variograms with no spatial autocorrelation. To characterize
asymptotic (untrended) variograms, I fit the data with the exponential variogram model
;9(h) = C0 + Cl [1 —exp [—3 3]]
a
which included the following parameters: range (a), the maximum distance at which data
are spatially dependent; sill (C 1), the semivariance value at the which the range is
reached; and nugget (C 0), the component of the variance caused by local variability at
73
scales smaller than the sampling interval. Models were fit in AD Model Builder (Limited
2000) and the best fit model was chosen by selecting the model with the lowest Akaike
Information Criterion (AIC) value (Bumham and Anderson 2002). Weight of evidence
in favor of each model was assessed by computing model probabilities using Akaike
weights (Bumham and Anderson 2002).
Results
As range expansion progressed, changes in variogram behavior were displayed by
all 12 model scenarios. Four general types of changes were exhibited as indicated by
changes in the model selected as best fit most ofien in each time step (Figures 3.2-3.13).
The first group of model scenarios (realsiml .7, realsim2.l, realstrat1.7Dl, realstrat2.1Dl)
displayed a clear switch from quadratic to exponential in the model most often chosen as
best fit, implying that variogram shapes progressed from bell-shaped to asymptotic
(Figures 3.2-3.5). One commonality among these four scenarios is that they all modeled
dispersal across real landscapes composed of 54% suitable habitat (Table 1). Simple
dispersal models (Figures 3.2 and 3.3) took longer to progress from bell-shaped to
asymptotic than stratified models (Figures 3.4 and 3.5). The linear model was chosen as
best-fit more often in model scenarios with r = 2.1 (Figures 3.3 and 3.5) than in models
with r = 1.7 (Figures 3.2 and 3.4); otherwise, no major differences were observed
between model scenarios with different population growth rates. Both stratified dispersal
model scenarios in this group (realstratl .7D1 and realstrat2.]Dl) generated dispersal
patterns using dispersal Distribution 1. In general, confidence intervals around
probability estimates were narrow relative to the second and third groups of model
74
scenarios. Also, models chosen as best-fit were clearly the best model most of the time
(i.e. had the highest probability of being the actual best model) except when variogram
behavior progressed from bell-shaped to asymptotic.
The second group of model scenarios (realstratl .7D2 and realstrat2.1D2 in
Figures 3.6 and 3.7) displayed a clear switch in the best-fit model from quadratic to a
mixture of exponential and (less ofien) linear models early in the time series, implying
that variogram shapes progressed from bell-shaped to either asymptotic or linear. One
commonality among these two scenarios is that they both modeled stratified dispersal
with the long-tailed dispersal Distribution 2 (Figure 3.1). Also, both model scenarios
were run on a real landscape composed of 54% suitable habitat. No major differences
were observed between scenarios with different population growth rates. Although the
exponential had a higher probability of being the actual best model than the linear or
quadratic (at time steps > ‘~10), confidence intervals around probability estimates
broadened once variogram behavior progressed from bell-shaped to asymptotic or linear.
Such broad confidence intervals indicate that evidence in favor of the exponential model
over the linear or quadratic models was not strong.
The third group of model scenarios (siml .7 and sim2.1 in Figures 3.8 and 3.9)
was best fit with the quadratic model early in the time series. However, later in the time
series, the number of times the quadratic model was chosen as best-fit declined and the
linear (and occasionally mean) model became more prevalent. Later in the time series,
the mean (and exponential for sim2.1) model was occasionally chosen as best-fit as well
but not as often as the quadratic or linear. These two scenarios both modeled simple
dispersal on landscapes composed of 100% suitable habitat. Interestingly, the scenario
75
with a higher population growth rate (sim2.l) took longer for convergence between the
top three models; also, confidence intervals around probability estimates for scenario
sim2.1 were narrower than that of the model with a lower growth rate (siml .7).
The fourth group of model scenarios (strat2.lD2, strat2.lDl , stratl .7D2, and
stratl .7D1 in Figures 3.10 to 3.13) produced variogram patterns that were best-fit by the
quadratic model a majority of the time. Occasionally, the mean or linear models were
chosen as well for model scenarios strat1.7D1 and strat2.lD2 (Figures 3.12 and 3.13),
respectively. However, the probabilities that either the mean or linear models were
actually the best-fit model were quite low. All four of these model scenarios were
stratified dispersal models run on landscapes composed of 100% suitable habitat.
Contrary to my prediction, both simple and stratified dispersal models exhibited
variogram progression patterns similar to that of gypsy moths in Michigan (Chapter 1)
when dispersal was modeled across a real section of the Michigan landscape. Among all
12 model scenarios, scenarios in the first group exhibited variogram progression patterns
most similar to those exhibited by gypsy moths (bell-shaped to linear to asymptotic).
However, the linear model was not typically chosen as an intermediate best-fit model
with the possible exception of scenarios realsim2.l (Figure 3.3) and realstrat2.1Dl
(Figure 3.5) in which the linear model sometimes fit the data well.
Another discrepancy between these results and the gypsy moth study is that
changes in variogram behavior generated by dispersal models showed close
correspondence with changes in population abundance over time. Progression in
variogram behavior typically occurred when population abundance began to reach an
asymptote. No apparent relationship was observed between the presence of limit cycles
76
(generated when r = 2.1) and higher variability in the probability of a given model
actually being best-fit.
Discussion
The results of this study indicate that landscape heterogeneity is the major driving
force behind semivariance patterns observed during the gypsy moth invasion of Michigan
(Chapter 1). In other words, spatially autocorrelated habitat (not dispersal strategy) may
be determining the way in which abundance is distributed across the landscape. Only
models of dispersal across sections of a real landscape produced variogram progression
patterns similar to that of the gypsy moths. A strong effect of landscape heterogeneity on
the range expansion process has been observed in other studies of dispersal and
biological invasions as well. Environmentally dependent rates of range expansion have
been observed repeatedly in both empirical and simulation studies (Bergelson et al. 1993,
Hengeveld and Van den Bosch 1997, Collingham and Huntley 2000, Hastings et al.
2005). Also, landscape heterogeneity may affect dispersal success (With and Crist 1995,
Schumaker 1996, King and With 2002, Matlack and Monde 2004), although not
necessarily to the extent suggested by many simulation studies (Chapter 2). It is unlikely
that unique characteristics of the Michigan landscape caused such variogram progression
patterns because similar patterns were observed during the gypsy moth invasion of the
east coast of North America (Liebhold et al. 1991).
Contrary to my predictions, variogram progression patterns similar to those of the
gypsy moth do not appear to be generated by differences in dispersal strategy.
Variogram progression patterns were remarkably similar between simple and stratified
77
dispersal models in the first group of model scenarios (realsim1.7, realsim2. 1 ,
realstrat1.7D1, realstrat2.1D1); the major observable difference was that simple dispersal
models took longer to progress from quadratic to exponential. Thus, stratified dispersers
like the gypsy moth are likely not the only organisms that produce a progression of
variogram patterns from bell-shaped to asymptotic as they undergo range expansion.
Although variograms were not calculated by Brodie et al. (1995), they did observe a
distinct progression of correlogram change over time; this suggests that variogram
analysis of their data might reveal that organisms with dispersal strategies quite different
from gypsy moths also exhibit distinct variogram progression patterns.
An alternative explanation for why I did not observe typical variogram
progression patterns in many of the stratified dispersal model scenarios (group 4) may be
that the size of my landscape was not large enough to allow for the process of long-
distance dispersal to be modeled properly. The second group of long-distance stratified
dispersal scenarios (realstrat2.1D2 and realstratl .7D2) exhibited a switch in variogram
behavior from bell-shaped to either linear or exponential; if reflecting boundaries and a
larger landscape had been used, this group of model scenarios may have exhibited
variogram progression patterns most similar to that of gypsy moths. Schneider (2003)
observed a linear relationship between mean movement distances of five butterfly species
and the size of the study area across which mark-recapture studies were conducted.
Therefore, the study area size may have a major impact on the movement distances
observed for volant organisms that exhibit long-distance dispersal.
I did not observe many occurrences when the linear model was best fit to
variograms in the first group of model scenarios (realstrat1.7D1, realsim2. 1 , realsim2. 1 ,
78
realstrat2.1Dl) possibly because of differences between the linear model used in this
study and the power model used in Chapter 1. In Chapter 1, I did not restrict the
exponent of the power model
flh) =C0 +Clha,00
to equal 1; therefore, models that were primarily linear with a slight curve downward at
the end of the time series were still best-fit by the power model with exponent < 1. In
this chapter, I used a strict linear model, so only variograms with a straight line were best
fit with the linear model. If the variograms in Chapter 1 were refit with the models used
in this chapter (quadratic, linear, mean, and exponential), the gypsy moth variograms that
were best fit by the power model with an exponent < 1 (Table 1.1) in years 1993 to 1995
(Figure 1.2) would likely be best fit by the exponential model and only two of the 12
variograms would have been characterized as linear.
The close correspondence between timing of variogram progression patterns and
changes in population dynamics in my simulation results indicates that a time lag may be
affecting the timing of variogram progression in the gypsy moth monitoring data.
Because female gypsy moths are incapable of flight and male moths generally do not
disperse beyond 1 km/year (White et al. 2003), the gypsy moth invasion wavefront is
driven largely by passively transported, wind-dispersed larvae (Elkinton and Liebhold
1990). However, the gypsy moth monitoring program in Michigan sampled the adult
male moth population in a given area (Yang et al. 1998). Therefore, estimates of the
timing of spatial changes in population structure may need to be adjusted to account for a
time lag if larval movements drive those changes. Although inspection of spatial patterns
in variogram behavior has potential implications for detecting stages in the invasion
79
process (Chapter 1), the biology of the invading organism should be taken into account
when interpreting the timing of changes in both temporal and spatial population
dynamics. Alternatively, the dispersal models used in these simulations may have
tracked closely changes in spatial structure because they simulated density dependent
dispersal. If gypsy moths do not exhibit density dependent dispersal, a time lag between
leveling off of total population abundance and the switch from linear to asymptotic
variograms would be expected in the gypsy moth variogram analysis (Chapter 1).
Another aspect of my prediction that was not supported by these simulation
studies was that gypsy moth-like variogram patterns would only be observed when
dispersal was modeled using a long-tailed dispersal distribution. In fact, models that
utilized a relatively high amount of long-distance dispersal generated less stable spatial
structure (i.e. fluctuations between exponential and linear models) than simple dispersal
models. Again, the second group of long-distance stratified dispersal scenarios
(realstrat2. 1 D2 and realstratl .7D2) may have exhibited variogram progression patterns
most similar to that of gypsy moths if reflecting boundaries and a larger landscape had
been used. Although long-tailed dispersal distributions were not required to generate
variogram progression patterns, they are often required to accurately predict the rate of
range expansion for species with stratified dispersal strategies such as trees (Clark 1998,
Higgins and Richardson 1999, Clark et al. 2001), insects (Kot et al. 1996, Lewis 1997,
Bailey et al. 2003), and birds and mammals (Van den Bosch et al. 1992, Shigesada et al.
1995, Veit and Lewis 1996). Therefore, different aspects of the range expansion process
may depend on different aspects of an organism’s dispersal strategy (Kinlan et al. 2005).
Range expansion rate (Moody and Mack 1988, Clark 1998) and genetic connectivity
80
(Trakhtenbrot et al. 2005) may rely heavily on rare, long-distance events. However,
maintenance of spatial structure within a species range may be driven by patterns of local
dispersal (this paper) across a spatially autocorrelated environment (Brown 1984). Clark
has shown that the use of mixed dispersal kernels (dispersal distributions split into local
and long-distance components) in range expansion models produce the best estimates of
range expansion rate (Clark 1998). Such an approach may be needed to tease apart the
relative importance of each dispersal component to the overall range expansion process,
and to identify the effect of landscape heterogeneity on each component.
81
Table 3.1. Description of twelve model scenarios, including dispersal strategy modeled,
landscape type, population grth rate (r), and dispersal distributions used in model
simulations. Distribution 1 is a gamma distribution with rate parameter of 0.5 and shape
parameter of 1.2; Distribution 2 is a long-tailed gamma distribution with rate parameter
of 0.043 and shape parameter of 0.5.
Scenario Dispersal Landscape r Dispersal
name strategy drstrrbutron
siml .7 Simple 100% habitat 1.7 -
sim2.1 Simple 100% habitat 2.1 -
strat1.7Dl Stratified 100% habitat 1.7 Distribution 1
strat1.7D2 Stratified 100% habitat 1.7 Distribution 2
strat2.lDl Stratified 100% habitat 2.1 Distribution 1
strat2. 1 D2 Stratified 100% habitat 2.1 Distribution 2
realsim1.7 Simple 54% habitat (real) 1.7 -
realsim2.] Simple 54% habitat (real) 2.1 -
realstrat1.7Dl Stratified 54% habitat (real) 1.7 Distribution 1
realstratl .7D2 Stratified 54% habitat (real) 1.7 Distribution 2
realstrat2.1Dl Stratified 54% habitat (real) 2.1 Distribution 1
realstrat2.1D2 Stratified 54% habitat (real) 2.1 Distribution 2
82
0.4 -
Probability density
o .o
N t»
.O
y—d
0.0 . . . .
5 10 15 20
Dispersal distance (number of cells)
Figure 3.1. Two dispersal distributions used to determine the proportion of dispersers
moving a given dispersal distance (number of cells in the landscape grid). The solid line
represents a gamma distribution with shape = 1.2, and rate = 0.5. The dashed line
represents a long-tailed gamma distribution with shape = 0.043, and rate = 0.5.
83
50 8e+4
g
-o
0
2‘3. 40 .
£- 6e+4
‘fi
§ 30.
8
O 4e+4
E
8 20 -
.§
gES 2e+4
5 10 -
JD '1:
g o
'8
Z 0 ~ 0 g
20 40 60 80 100 8
2‘
1.2 - — 8 +4
(b) e a
E
1.0 i (D
:2 - 6e+4
B 0.8 ~
.2
33
a 0'6 ‘ ~ 4e+4
A?
:1 0.4 4
'8
'8 - 2e+4
a: 0.2 -
0.0 3
I I I I I 0
0 20 4O 60 80 100
Time step
Figure 3.2. Change in variogram behavior for model scenario realsim1.7 characterized
by the total number of times each model was chosen as best-fit to the data in each time
step (a) and the mean probability that each model was the actual best model in each time
step (b). Mean probability estimates include 95% confidence intervals. Variograms were
fit with a quadratic equation (closed circles), a linear model (open triangles), a model of
the mean semivariance (closed squares), and an exponential model (open diamonds). The
solid line and values on the right ordinate represent total population size averaged over all
50 scenarios.
84
- le+5
g
"C
d)
a - 8e+4
E
8
f - 6e+4
O
“O
O
a
m
,“g’ 4e+4
«I:
O
312' r 2e+4
E a
z '2
0 a?
100 g'
1.2 ~ - le+5 g‘
(b) 5‘
.3 1.0 '1 _ 8
1: 8e+4
E
”,3 0.8 __
a - 6e+4
3 1
a 0.6
Lg 0.4 .
.D
2
°‘ 0.2 1 - 2e+4
0.0 l 0
0 20 40 60 80 100
Time step
Figure 3.3. Change in variogram behavior for model scenario realsim2.] characterized
by the total number of times each model was chosen as best-fit to the data in each time
step (a) and the mean probability that each model was the actual best model in each time
step (b). Mean probability estimates include 95% confidence intervals. Variograms were
fit with a quadratic equation (closed circles), a linear model (open triangles), a model of
the mean semivariance (closed squares), and an exponential model (open diamonds). The
solid line and values on the right ordinate represent total population size averaged over all
50 scenarios.
85
50 0009 00990 00900 90 0 000900 000000 90000 00 — 8e+4
S . . . . o 00 e 0 0
C3
1:
.2
3 40 - l
,5 (a) 6e+4
3
.o
T) 30 ‘
'8
E ~ 4e+4
3’3 20 -
.22
C
2 ~ 2e+4
g 10 ~
g '0
z . '8
0 ” . A . . o g—
20 40 6O 80 100 §
1.2 e 8‘
E
o.
10 v I‘ " 'yg ' ! "U'. 'U'." _ 6e+4 g
. 1 A " . . .' ’. . . . .‘ '. . .‘.~',‘. 0 .'.'. .'. O o . O .'.'.'o.0.o.o . O o . 8
Probability model is best fit
0 20 40 60 80 l 00
Time step
Figure 3.4. Change in variogram behavior for model scenario realstratl .7D1
characterized by the total number of times each model was chosen as best-fit to the data
in each time step (a) and the mean probability that each model was the actual best model
in each time step (b). Mean probability estimates include 95% confidence intervals.
Variograms were fit with a quadratic equation (closed circles), a linear model (open
triangles), a model of the mean semivariance (closed squares), and an exponential model
(open diamonds). The solid line and values on the right ordinate represent total
population size averaged over all 50 scenarios.
86
50 - 9009... ..o 09 o .9999 0.0 o o 090 90 — le+5
O . O. O . .0099 O O.
s O O O .
8 .
g 40 - 8e+4
E (a)
‘67:
B 30 .1 ; ’ 6e+4
g l
E
g 20 ' ' - 4e+4
.._. l
(I; O
33 10 0/ - 2e+4
E H
Z 0}_ .‘1.AQ.AAA. OAAAIL. .. .l 0 g
20 4O 60 80 100 g
1.2 - 8
8-
- 8e+4 E
E it. .:.'.'9 .'.'9'.O. .:.|O.9..:.'......’.;59.009.'.OO.:. .:.O.9. g
8 w o . ('0
:2 0.8 ‘ ~ 6e+4
:g l (b)
g 0.6 ~ ‘
.1? l " 4e+4
'2'; 0.4 -
.o
e ..
D.- 0.2 7 A“: L ze+4
:12:
o o ifi'“ 7777777 . 7‘4“" 375% 77 . 777‘; 1"? $3533? . . 7777777
7 t I l j 1 0
0 20 40 6O 80 100
Time step
Figure 3.5. Change in variogram behavior for model scenario realstrat2.1Dl
characterized by the total number of times each model was chosen as best-fit to the data
in each time step (a) and the mean probability that each model was the actual best model
in each time step (b). Mean probability estimates include 95% confidence intervals.
Variograms were fit with a quadratic equation (closed circles), a linear model (open
triangles), a model of the mean semivariance (closed squares), and an exponential model
(open diamonds). The solid line and values on the right ordinate represent total
population size averaged over all 50 scenarios.
87
50 1 - 8e+4
a
g ( )
13
Q)
'5 40 ~ 7
E r 6e+4
E
g 30 '
g - 4e+4
V)
t... .
S - 2e+4
fig 10 l
l V
z A/" . g
0 , r , . M— 0 E.
20 40 60 80 100 g
S
1.2 1 (b) r 86+4 g.
E
E. 1.0 a E
g ~ 6e+4 n
.o
”‘2
E
8
E - 4e+4
.é‘
.5
8
g ‘ — 2e+4
0
100
Time step
Figure 3.6. Change in variogram behavior for model scenario realstrat1.7D2
characterized by the total number of times each model was chosen as best-fit to the data
in each time step (a) and the mean probability that each model was the actual best model
in each time step (b). Mean probability estimates include 95% confidence intervals.
Variograms were fit with a quadratic equation (closed circles), a linear model (open
triangles), a model of the mean semivariance (closed squares), and an exponential model
(open diamonds). The solid line and values on the right ordinate represent total
population size averaged over all 50 scenarios.
88
50 ~ le+5
8 a
g ( )
d.)
4:
g 40 * 86+4
E
E) 30 * 6e+4
o
E
3
.§ 20 * 4e+4
c:
o
0) .
"g 10 r 28+4
Z
5
0 n‘ . A A . K IA . A 0 ST
20 40 60 80 100 g'
1.2 - g
l se+4
: 10 - (b) g
l...
,3; 0.8 - " 6e+4
E
g 0.6 -
__ ~ 4e+4
g .l- a: 555+“ :_ _ __" 1+: _
g 0.4 ~ " l "‘” ‘"“ " " ___ ‘
.0 ppm m ’
8 RAM 5) 1
‘7“ 0,2 - _ q_ ___.~,:-=‘~— —: ,hfrfiztvw E5377: ’ 2e+4
10000. F011); 1110001)"! 1 ~-‘
00 ‘ ‘ minimum: 0
0 20 40 60 80 100
Time step
Figure 3.7. Change in variogram behavior for model scenario realstrat2.1D2
characterized by the total number of times each model was chosen as best-fit to the data
in each time step (a) and the mean probability that each model was the actual best model
in each time step (b). Mean probability estimates include 95% confidence intervals.
Variograms were fit with a quadratic equation (closed circles), a linear model (open
triangles), a model of the mean semivariance (closed squares), and an exponential model
(open diamonds). The solid line and values on the right ordinate represent total
population size averaged over all 50 scenarios.
89
5° - l.6e+5
s (a)
8 - l.4e+5
D
40 -
g * l.2e+5
‘63
3 30 ~ l l.0e+5
a . . . ‘
o . . ”‘5 . 'of ~ 8.0e+4
a ”m 1W
GE) 20 7 i . 7‘ ‘ r 6.0e+4
E 10 ” 4.0C+4
3 -
g - 2.0e+4 é“
Z O oowooooooooooooo- w‘woov‘ooMoWooao‘. 00 9E;
. 8'
20 4O 6O 80 100 3
cr
1'2 ' — l.6e+5 E
(b) g
0
1.0 . 'rv~- ‘ r 1.4e+5 a
t: . 1
7% l ’ - l.2e+5
B 0.8 ‘l
'9‘ - 1.0e+5
g 0 6
o . -
g r l‘ ‘t " ‘} "bn 8.0e+4
, . r .s‘ ‘
é 0'4 d a: ‘ AZA‘I‘ ‘} a.) ‘ nai‘ "' 6. 064-4
«s A .
IE ‘ I I - 4.0e+4
a. 0.2 - I I I
‘1 k' I ~ 2.0e+4
OO . #9ch ‘o' o o o o o o o o o I a .6 o o o 9 onto 9 o o o'o‘oio..‘.l..,..'.','.'!.L.$.1”.” O O
0 20 4O 6O 80 100
Time step
Figure 3.8. Change in variogram behavior for model scenario sim1.7 characterized by
the total number of times each model was chosen as best-fit to the data in each time step
(a) and the mean probability that each model was the actual best model in each time step
(b). Mean probability estimates include 95% confidence intervals. Variograms were fit
with a quadratic equation (closed circles), a linear model (open triangles), a model of the
mean semivariance (closed squares), and an exponential model (open diamonds). The
solid line and values on the right ordinate represent total population size averaged over all
50 scenarios.
9O
5 0 ~ 2e+5
(6
g (a) t 2e+5
\
{‘9’ 4o - V ‘ * 26+5
7; r le+5
o
f 30 _‘ " le+5
g L 1 +5
E e
g 20 . ' 86+4
5;: . . )- 6e+4
g 10 - ’ 4C+4
. . “U
0 ~ 2e+4 o
z ‘ 7 7 7 ‘ ‘ ‘3 ‘ 70.. £759.72 :70 7 o o IO
0 00300990999094“.oooo""‘. . .. O 93". l' O E
20 4O 6O 80 100 g
g.
1.2 - (b) _ l.8e+5 g
10 * l.6e+5 g
E ' - l.4e+5
i”;
.23 0.8 ‘ ~ l.26+5
T) __ ,_
g 0.6 - . l.0e+5
é‘ 8.0e+4
= 0.4 i“ l
E l H- e 6.0e+4
o
5: 0.2 ‘ 4.0e+4
‘ 2.0e+4
O 20 40 60 80 100
Time step
Figure 3.9. Change in variogram behavior for model scenario sim2.1 characterized by
the total number of times each model was chosen as best-fit to the data in each time step
(a) and the mean probability that each model was the actual best model in each time step
(b). Mean probability estimates include 95% confidence intervals. Variograms were fit
with a quadratic equation (closed circles), a linear model (open triangles), a model of the
mean semivariance (closed squares), and an exponential model (open diamonds). The
solid line and values on the right ordinate represent total population size averaged over all
50 scenarios.
91
50 1.6e+5
S
(U
73 - 1.4e+5
5 40 . W
E (a) ~ 1.2e+5
‘53
g . 1.0e+5
O)
'8
E L 8.0e+4
8
.g — 6.0e+4
“5
a; - 4.0e+4
E
Z r 2.0e+4 BU
.. '8
. 0.0 g
100 g'
g.
- l.4e+5 é
7:? 1'0 ‘ W v mm I 2’: I I t t I: g
cm) a ‘.‘ u n u a t a a I,— l.2€+5 (D
,9 l I I
U)
-— 0.8 l _
T) i (b) l.Oe+5
‘3
E 0.6 - ~ 8.0e+4
.Q‘
= II ~ 6.0e+4
g 0.4 - u
,5: 9' . 4.0e+4
N"
0.2 - "LL
“ill . - 2.0e+4
0 0 M}; M . A in: .‘.‘!‘_‘J‘s'!a‘s"a‘u‘s‘ '4‘". , .v. .v. .v. in ‘1'".
. 0.. QOOQOOOOOOOQOQO 906066000OO0,0.0,0000,0.00,000(60.0},0 00
O 20 40 60 80 100
Time step
Figure 3.10. Change in variogram behavior for model scenario stratl .7D2 characterized
by the total number of times each model was chosen as best-fit to the data in each time
step (a) and the mean probability that each model was the actual best model in each time
step (b). Mean probability estimates include 95% confidence intervals. Variograms were
fit with a quadratic equation (closed circles), a linear model (open triangles), a model of
the mean semivariance (closed squares), and an exponential model (open diamonds). The
solid line and values on the right ordinate represent total population size averaged over all
50 scenarios.
92
Number of times model best fit the data
aoucpunqe uouelndod
Probability model is best fit
Time step
Figure 3.11. Change in variogram behavior for model scenario strat2.lD1 characterized
by the total number of times each model was chosen as best-fit to the data in each time
step (a) and the mean probability that each model was the actual best model in each time
step (b). Mean probability estimates include 95% confidence intervals. Variograms were
fit with a quadratic equation (closed circles), a linear model (open triangles), a model of
the mean semivariance (closed squares), and an exponential model (open diamonds). The
solid line and values on the right ordinate represent total population size averaged over all
50 scenarios.
93
50 — l.6e+5
S
.8 - l.4e+5
0)
5 40 -
3'3 (a) - 1.2e+5
‘53
'8 30 - - l.0e+5
E
o - . +
E 8 0e 4
U)
“8’ 20 7 ~ 6.0e+4
<3:
3 ~ 4.0e+4
3 10 «
g ~ 2.0e+4
z 4 . a
O .v‘A......................-;~Q96‘2‘9'.'OOOOO""‘999909.9....7 0.0 :8...
2° 40 60 80 100 g.
:1
1.2 - g—
- l.4e+5 g
1.0 ‘ ‘ ‘. 1 _. Z. r. I‘- E‘
E ’1 I ‘ * l.2e+5 8
273
73) 0'8 d l' (b) - l.Oe+5
T)
“g": 0.6 + - 8.0e+4
g - 6.0e+4
35 0.4 -
(U
E r 4.0e+4
9‘ 0.2 - l
‘ i — 2.0e+4
.‘ "loll ,
0.0 70.“..o’ooetoooo9099000060'...0’09ooo”"¢o€'¢o.oo0’°7 00
0 2° 40 60 80 100
Time step
Figure 3.12. Change in variogram behavior for model scenario strat1.7Dl characterized
by the total number of times each model was chosen as best-fit to the data in each time
step (a) and the mean probability that each model was the actual best model in each time
step (b). Mean probability estimates include 95% confidence intervals. Variograms were
fit with a quadratic equation (closed circles), a linear model (open triangles), a model of
the mean semivariance (closed squares), and an exponential model (open diamonds). The
solid line and values on the right ordinate represent total population size averaged over all
50 scenarios.
94
50 ’ r 2e+5
:3
3 ~ ‘ H '. 2e+5
d.) A a . J; t
g 40 4 WV ' W “w _ 264.5
3::
*" ~ le+5
§ (a)
T. 30 ‘ l le+5
B
E \ - le+5
(I)
g 20 ' I - 8e+4
“E ‘ an
“E 10 ‘ . r 4e+4
Z J i - 2e+4 3"
“U
0 ‘0'09009609699099000969000606099090000.99906006090 0 g
20 40 6O 80 100 g'
1.2 - - 1.8e+5 “g
D.
- 1.6e+5 g
1.0 " 8
as ’1 | l u ’ ~1.4e+5
‘5 5 l t
3; 0'“ I, ll I VI II : : - 1.2e+5
6 ’ ,.
”8 0.64 t 1.0e+5
5. ~ 8.0e+4
E; 0.4 - i - 6.0e+4
'8
d: 0.2 - . - 4.0e+4
‘ la“ 2.0e+4
0.0 - ,
. . . . 0.0
o 20 40 6O 80 100
Time step
Figure 3.13. Change in variogram behavior for model scenario strat2.lD2 characterized
by the total number of times each model was chosen as best-fit to the data in each time
step (a) and the mean probability that each model was the actual best model in each time
step (b). Mean probability estimates include 95% confidence intervals. Variograms were
fit with a quadratic equation (closed circles), a linear model (open triangles), a model of
the mean semivariance (closed squares), and an exponential model (open diamonds). The
solid line and values on the right ordinate represent total population size averaged over all
50 scenarios.
95
CONCLUSIONS
The results of this dissertation illustrate two ways in which landscape
heterogeneity affects the range expansion process. First, the spatial distribution of
abundance across the landscape appears to be a population response to the composition
and configuration of habitat. For simulation models (Chapter 3) to produce semivariance
progression patterns similar to those observed empirically (Chapter 1), dispersal had to be
modeled across a real landscape that contained a mosaic of habitat and non-habitat
patches. The second way in which landscape heterogeneity may affect the range
expansion process is by reducing the uniformity and speed with which a species expands
across a landscape. In my second chapter, gypsy moth invasion success was shown to
decline linearly with habitat loss and fragmentation.
Detection of emergent properties
Spatial patterns observed in this analysis of gypsy moth range expansion across
Michigan reflect the combined individual responses of numerous insects. As shown in
Chapter 2, the population response to habitat loss and fragmentation was markedly
different from simulation models of individual dispersal events. Therefore, the linear
response of the gypsy moth invasion wavefront to landscape heterogeneity may be an
emergent property of a complex system. In other words, individual dispersal success
may exhibit a threshold (sudden, nonlinear) response to habitat loss and fragmentation
that is not exhibited by the entire population when invasion success of the wavefront is
measured.
96
In order to detect such emergent properties, large-scale monitoring programs are
essential. The spatial sampling protocol used in the Michigan gypsy moth monitoring
program was designed to identify large-scale processes such as movement of the invasion
wavefront. Most management programs cannot invest resources in tracking individual
dispersal events. Therefore, large-scale, long-term monitoring programs like the one
used to track movement of the gypsy moth invasion wavefront in Michigan are vital if we
are to improve the management of invasive species spread.
The best method for spatially sampling a population so that large-scale spatial
patterns (e. g. movement of an invasion wavefront or changes in spatial structure of a
population over time) can be detected has yet to be determined. The gypsy moth
monitoring program in Michigan was ideal for spatial analyses because most traps
remained in the same location across the entire 12 years of data collection. Without
permanent sampling locations, many spatial statistics cannot be compared among years
because of change in support in the data (Isaaks and Srivastava. 1989). Another aspect of
the gypsy moth monitoring program that was ideal for spatial analysis is that traps were
placed across the entire state. Good spatial coverage across all areas in which an invasive
species might spread is necessary if semivariance analysis is to be used to identify
invasion stages. Without sampling points located along the edge or outside the area that a
species occupies, the bell-shaped variograms that signal the establishment stage (Chapter
1 and 3) may not be detectable. In the future, I plan to explore the effect of different
sampling schemes on the detection of changes in spatial structure of expanding
populations.
97
Semivariance analysis
By comparing methodologies from Chapter 1 with Chapter 3, I learned that non-
traditional models of semivariance (e. g. quadratic and mean used from Chapter 3) should
be used in place of traditional variogram models (e. g. exponential, spherical, Gaussian
from Chapter 1) when semivariograms are being used as a tool in monitoring range
expansion. Traditional semivariogram models are not adequate for describing trended
data because they do not characterize well the wide range of semivariogram behavior
exhibited by trended data (e. g. bell-shaped variograms). Although such trended data
should never be used as the basis for interpolative modeling (i.e. kriging) because they do
not provide reliable estimates of the range and sill of a variogram, trended
semivariograms appear to be useful tools in identifying significant changes in spatial
structure of a population undergoing range expansion
Future directions
This dissertation suggests that landscape heterogeneity is a driving force behind
spatial patterns in abundance for populations undergoing range expansion. Therefore,
future research in dispersal and invasion ecology should concentrate more closely on the
relationship between landscape structure and dispersal. Dennis et al. (1998) and Lele et
al. (1998) have found marked improvements in population modeling when spatial
variation in environmental disturbance, population growth rates, and dispersal are
included. In the future, I will build upon their results by explicitly modeling stochastic,
density dependent population grth across an array of different heterogeneous
landscapes. I will then compare semivariogram progression patterns generated by
98
different dispersal scenarios across landscapes with substantially different levels of
habitat composition and fragmentation (a combination of approaches from my second
and third chapters). This approach will allow me to identify what types of structured
landscapes generate specific types of semivariogram patterns and also the level of habitat
loss at which semivariance progression patterns emerge (Chapter 3). I also plan to test
for improvements in range expansion model predictions when realistic population
responses to landscape structure are explicitly modeled. I expect that such an approach
will greatly improve predictions of the movement of harmful invaders (Moody and Mack
1988, With 2002), the range expansion of native species (Dark et al. 1998, Trakhtenbrot
et al. 2005), and the effects of climate change on the maintenance of existing species and
ecological communities (Dyer 1995, Higgins and Richardson 1999, Iverson et al. 2004).
Finally, I am interested in performing semivariance analysis on species undergoing range
contraction to see if semivariance progression patterns exhibited by such populations are
reversed in comparison to patterns exhibited by the gypsy moth.
99
LITERATURE CITED
Albert, D., S. Denton, and B. Barnes. 1986. Regional landscape ecosystems of Michigan.
University of Michigan, Ann Arbor.
Anderson, R. M., and R. M. May. 1986. The invasion, persistence and spread of
infectious diseases within animal and plant communities. Philosophical
Transactions of the Royal Society of London B 314:533-570.
Andren, H. 1994. Effects of habitat fragmentation on birds and mammals in landscapes
with different proportions of suitable habitat: a review. Oikos 71:355-366.
Ashton, P. J ., and D. S. Mitchell. 1989. Aquatic plants: patterns and modes of invasion,
attributes of invading species and assessment of control programmes. Pages 525
in J. A. Drake and H. A. Mooney, editors. Biological invasions: a global
perspective. John Wiley & Sons, New York, New York, USA.
Bailey, R. I., M. E. Lineham, C. D. Thomas, and R. K. Butlin. 2003. Measuring dispersal
and detecting departures from a random walk model in a grasshopper hybrid zone.
Ecological Entomology 28: 129-13 8.
Bailey, T. C., and A. C. Gatrell. 1995. Interactive spatial data analysis. Prentice Hall,
New York, New York, USA.
Bergelson, J ., J. A. Newman, and E. M. F loresroux. 1993. Rates of weed spread in
spatially heterogeneous environments. Ecology 74:999-1011.
Brodie, C., G. Houle, and M. Fortin. 1995. Development of a Populus balsamifera clone
in subarctic Quebec reconstructed from spatial analyses. Journal of Ecology
83:309-320.
Brown, J. H. 1984. On the relationship between abundance and distribution of species.
American Naturalist 124:255-279.
Brown, J. H. 1995. Macroecology. The University of Chicago Press, Chicago, Illinois,
USA.
Brown, J. H., and A. Kodric-Brown. 1977. Turnover rates in insular biogeography: effect
of immigration on extinction. Ecology 58:445-449.
Brown, M. W. 1993. Population dynamics of invading pests: factors governing success.
Pages 203-218 in K. C. Kim and B. A. McPheron, editors. Evolution of insect
pests: patterns of variation. Wiley, New York.
100
Burgess, T. M., and R. Webster. 1980. Optimal interpolation and isarithmic mapping of
soil properties I. The semi-variogram and punctual kriging. Journal of Soil
Science 31:315-331.
Bumham, K. P., and D. R. Anderson. 2002. Model selection and multimodel inference: a
practical information-theoretic approach, Second edition. Springer-Verlag, New
York, New York.
Carroll, S. P., and H. Dingle. 1996. The biology of post-invasion events. Biological
Conservation 78:207-2 14.
Case, T. J. 2000. An illustrated guide to theoretical ecology. Oxford University Press,
New York.
Chiles, J ., and P. Delfiner. 1999. Geostatistics: modeling spatial uncertainty. John Wiley
& Sons, Inc., New York.
Clark, J. S. 1998. Why trees migrate so fast: confronting theory with dispersal biology
and the paleorecord. American Naturalist 152:204-224.
Clark, J. S., M. Lewis, and L. Horvath. 2001. Invasion by extremes: variation in dispersal
and reproduction retards population spread. American Naturalist 157:537-554.
Cliff, A. D., and J. K. Ord. 1981. Spatial processes: models & applications. Pion,
London.
Collingham, Y. C., and B. Huntley. 2000. Impacts of habitat fragmentation and patch size
upon migration rates. Ecological Applications 10:131-144.
Cousens, R., and M. Mortimer. 1995. Dynamics of weed populations. Cambridge
University Press, Cambridge.
Cressie, N. A. C. 1993. Statistics for spatial data. John Wiley & Sons, Inc., New York.
Cronk, Q. C. B., and J. L. Fuller. 1995. Plant invaders: the threat to natural ecosystems.
Chapman & Hall, London.
Dark, S. J ., R. J. Gutierrez, and G. I. Gould. 1998. The barred owl (Strix varia) invasion
in California. Auk 115:50-56.
Dennis, B., W. P. Kemp, and M. L. Taper. 1998. Joint density dependence. Ecology
79:426-441.
Doak, D. F., P. C. Marino, and P. M. Kareiva. 1992. Spatial scale mediates the influence
of habitat fragmentation on dispersal success: implications for conservation.
Theoretical Population Biology 41:315-336.
10]
Dong, P. 2000. Lacunarity for spatial heterogeneity measurement in GIS. Geographic
Information Sciences 6:20-26.
Donovan, M. L., G. M. Nesslage, J. J. Skillen, and B. A. Maurer. 2004. The Michigan
Gap Analysis Project Final Report. Wildlife Division, Michigan Department of
Natural Resources, Lansing.
Dyer, J. M. 1995. Assessment of climatic warming using a model of forest species
migration. Ecological Modelling 79: 199-219.
Elkinton, J. S., and A. M. Liebhold. 1990. Population dynamics of gypsy moth in North
America. Annual Review of Entomology 35.
French, D. R., and J. Travis. 2001. Density-dependent dispersal in host-parasitoid
assemblages. Oikos 95.
Gage, S. H., T. M. Wirth, and G. A. Simmons. 1990. Predicting regional gypsy moth
(Lymantriidae) population trends in an expanding population using pheromone
trap catch and spatial analysis. Environmental Entomology 19:370-3 77.
Gardner, R. H., B. T. Milne, M. G. Turner, and R. V. O'Neill. 1987. Neutral models for
the analysis of broad-scale landscape pattern. Landscape Ecology 1:19-28.
Garnder, R. H., R. V. O'Neill, M. G. Turner, and V. H. Dale. 1989. Quantifying scale-
dependent effects of animal movement with simple percolation models.
Landscape Ecology 3:217-227.
Goovaerts, P. 1997. Geostatistics for natural resource evaluation. Oxford University
Press, New York.
Griffith, D. A. 1987. Spatial autocorrelation: a primer. Association of American
Geographers, Washington, DC.
Gustafson, E. J ., and R. H. Gardner. 1996. The effect of landscape heterogeneity on the
probability of patch colonization. Ecology 77:94-107.
Hanna, M. 1981. Gypsy moth (Lepidoptera: Lymantriidae) survey in Michigan. Great
Lakes Entomologist 14:103-108.
Hanski, I., and M. E. Gilpin. 1997. Metapopulation biology: ecology, genetics, and
evolution. Academic Press, New York, New York, USA.
Hastings, A. 19963. Models of spatial spread: a synthesis. Biological Conservation
78: 143-148.
102
Hastings, A. 1996b. Models of spatial spread: is the theory complete? Ecology 77 :1675-
1679.
Hastings, A., K. Cuddington, K. F. Davies, C. J. Dugaw, S. Elmendorf, A. Freestone, S.
Harrison, M. Holland, J. Lambrinos, U. Malvadkar, B. A. Melbourne, K. Moore,
C. Taylor, and D. Thomson. 2005. The spatial spread of invasions: new
developments in theory and evidence. Ecology Letters 8:91 -1 01.
Hengeveld, R. 1989. Dynamics of biological invasions. Chapman and Hall, New York,
New York, USA.
Hengeveld, R., and F. Van den Bosch. 1997. Invading into an ecologically non-uniform
area. Pages 217-225 in B. Huntley, W. Cramer, A. V. Morgan, H. C. Prentice, and
J. R. M. Allen, editors. Past and future rapid environmental changes: the spatial
and evolutionary responses of terrestrial biota. Springer-Verlag, Berlin, Germany.
Higgins, S. 1., and D. M. Richardson. 1999. Predicting plant migration rates in a changing
world: the role of long-distance dispersal. The American Naturalist 153:464-475.
Isaaks, E. H., and R. M. Srivastava. 1989. An introduction to applied geostatistics.
Oxford University Press, New York, New York.
Isard, S. A., and S. H. Gage. 2001. Flow of life in the atmosphere. Michigan State
University Press, East Lansing, Michigan, USA.
Iverson, L. R., M. W. Schwartz, and A. M. Prasad. 2004. How fast and far might tree
species migrate in the eastern United States due to climate change? Global
Ecology and Biogeography 13:209-219.
King, A. W., and K. A. With. 2002. Dispersal success on spatially structured landscapes:
when do spatial pattern and dispersal behavior really matter? Ecological
Modelling 147:23-39.
Kinlan, B. P., S. D. Gaines, and S. E. Lester. 2005. Propagule dispersal and the scales of
marine community process. Diversity and Distributions 11:139-148.
Koenig, W. D. 1999. Spatial autocorrelation of ecological phenomena. Trends in Ecology
& Evolution 14:22-26.
Kot, M., M. A. Lewis, and P. van den Driessche. 1996. Dispersal data and the spread of
invading organisms. Ecology 77 :2027-2042.
Kowarik, I. 1995. Time lags in biological invasions with regard to the success and failure
of alien species. in P. Pysek, K. Prach, M. Rejmanek, and M. Wade, editors. Plant
invasions: general aspects and special problems. SPB Academic Publishing,
Amsterdam.
103
Legendre, P. 1993. Spatial autocorrelation - trouble or new paradigm. Ecology 74:1659-
1673.
Legendre, P., and L. Legendre. 1998. Numerical Ecology, Second edition. Elsevier, New
York.
Lele, S., M. L. Taper, and S. H. Gage. 1998. Statistical analysis of population dynamics
in space and time using estimating functions. Ecology 79: 1489-1502.
Lewis, M. A. 1997. Variability, patchiness, and jump dispersal in the spread of an
invading population. Pages 46-69 in D. Tilman and P. Karieva, editors. Spatial
ecology: the role of space in population dynamics and interspecific interactions.
Princeton University Press, Princeton, New Jersey, USA.
Li, H., and J. Wu. 2004. Use and misuse of landscape indices. Landscape Ecology
19:389-399.
Liebhold, A. M., J. S. Elkinton, G. Zhou, M. E. Hohn, R. E. Rossi, G. H. Boettner, C. W.
Boettner, C. Bumham, and M. L. McManus. 1995. Regional correlation of gypsy
moth (Lepidoptera: Lymantriidae) defoliation with counts of egg masses, pupae,
and male moths. Environmental Entomology 24:196-203.
Liebhold, A. M., G. A. Halverson, and G. A. Elmes. 1992. Gypsy moth invasion in North
America: a quantitative analysis. Journal of Biogeography 19:513-520.
Liebhold, A. M., Z. Xu, M. E. Hohn, J. S. Elkinton, M. Ticehurst, G. L. Benzon, and R.
W. Campbell. 1991. Geostatistical analysis of gypsy-moth (Lepidoptera,
Lymantriidae) egg mass populations. Environmental Entomology 20:1407-1417.
Mack, R. N. 1985. Invading plants: their potential contribution to population biology. in
J. White, editor. Studies on plant demography: a festschrift for John. L. Harper.
Academic Press, London.
Malanson, G. P., and D. M. Cairns. 1997. Effects of dispersal, population delays, and
forest fragmentation on tree migration rates. Plant Ecology 131:67-79.
Matlack, G. R., and J. Monde. 2004. Consequences of low mobility in spatially and
temporally heterogeneous ecosystems. Journal of Animal Ecology 92:1025-1035.
Maurer, B. A. 1999. Untangling ecological complexity: the macroscopic perspective.
University of Chicago Press, Chicago.
Maurer, B. A., E. T. Linder, and D. Gammon. 2001. A geographical perspective on the
biotic homogenization process: implications from macroecology of North
American birds. in J. Lockwood and M. McKinney, editors. Biotic
homogenization. Kluwer Academic Publishers, London.
104
May, R. M. 1974. Biological populations with non-overlapping generations: stable
points, stable cycles, and chaos. Science 186:645-647.
Mennechez, G., N. Schtickzelle, and M. Baguette. 2003. Metapopulation dynamics of the
bog frittilary butterfly: comparison of demographic parameters and dispersal
between a continuous and a highly fragmented landscape. Landscape Ecology
18:279-291.
Michigan Department of Natural Resources. 1999. Michigan Resource Information
System (MIRIS). MDNR 1978 Landuse/Cover. Land Cover / Use 1. Lansing,
Michigan: MDNR. Land cover interpreted from aerial photography.
Mladenoff, D. J ., and B. DeZonia. 2002. APACK 2.22 User's guide.
Moller, H. 1996. Lessons for invasion theory from social insects. Biological
Conservation 78: 1 25- 142.
Moody, M. E., and R. N. Mack. 1988. Controlling the spread of plant invasions: the
importance of nascent foci. Journal of Applied Ecology 25:1009-1021.
Neter, J., W. Wasserman, and M. H. Kutner. 1985. Applied linear statistical models.
Irwin, Homewood, Illinois.
O'Dell, W. V. 1955. The gypsy moth outbreak in Michigan. Journal of Economic
Entomology 48: 1 70-1 72.
O'Neill, R. V., B. T. Milne, M. G. Turner, and R. H. Garnder. 1988. Resource utilization
scales and landscape pattern. Landscape Ecology 2:63-69.
Otter Research Limited. 2000. An introduction to AD Model Builder Version 4, Sydney,
British Columbia.
Pebesma, E. 1999. GSTAT User's Manual. Mathematics Department, Macquarie
University, Sydney, Australia.
Pimentel, D., L. Larch, R. Zuniga, and D. Morrison. 2000. Environmental and economic
costs of non-indigenous species in the United States. BioScience 50:53-65.
Pitelka, L. F. 1997. Plant migration and climate change. American Scientist 85:464-473.
Plotnick, R. E., R. H. Gardner, W. W. Hargrove, K. Prestegaard, and M. Perlmutter.
1996. Lacunarity analysis: a general technique for the analysis of spatial patterns.
Physical Review E 53:5461-5468.
105
Plotnick, R. E., R. H. Gardner, and R. V. O'Neill. 1993. Lacunarity indices as measures
of landscape texture. Landscape Ecology 8:201-211.
R Development Core Team. 2004. R: A language and environment for statistical
computing. in R Foundation for Statistical Computing, Vienna, Austria.
Reichard, S. H., and P. S. White. 2003. Invasion biology: an emerging field of study.
Annals of the Missouri Botanical Garden 90:64-66.
Ribiero, P. J ., Jr., and P. J. Diggle. 2001. geoR: A package for geostatistical analysis. R-
NEWS 1:15-18.
Richardson, D. M., N. Allsopp, C. M. D'Antonio, S. J. Milton, and M. Rejmanek. 2000a.
Plant invasions - the role of mutualisms. Biological Reviews 75:65-93.
Richardson, D. M., P. Pysek, M. Rejmanek, M. G. Barbour, F. D. Panetta, and C. J. West.
2000b. Naturalization and invasion of alien plants: concepts and definitions.
Diversity and Distributions 6:93-107.
Rossi, R. E., D. J. Mulla, A. G. Joumel, and E. H. Franz. 1992. Geostatistical tools for
modeling and interpreting ecological spatial dependence. Ecological Monographs
62:277-314.
Roughgarden, J. 1986. Predicting invasions and rates of spread. in H. A. Mooney and J.
A. Drake, editors. Ecology of biological invasions of North America and Hawaii.
Springer-Verlag, New York.
Schneider, C. 2003. The influence of spatial scale on quantifying insect dispersal: an
analysis of butterfly data. Ecological Entomology 28:252-256.
Schumaker, N. H. 1996. Using landscape indices to predict habitat connectivity. Ecology
77:1210-1225.
Schwartz, M. W. 1992. Modelling effects of habitat fragmentation on the ability of trees
to respond to climatic warning. Biodiversity and Conservation 2:51-60.
Sharov, A. A., and A. M. Leibhold. 1998. Model of slowing the spread of gypsy moth
(Lepidoptera: Lymantriidae) with a barrier zone. Ecological Applications 8:1170-
1179.
Shigesada, N., and K. Kawasaki. 1997. Biological invasions: theory and practice. Oxford
University Press, Oxford.
Shigesada, N., K. Kawasaki, and Y. Takeda. 1995. Modeling stratified diffusion in
biological invasions. American Naturalist 146:229-251.
106
Stacey, P. B., and M. L. Taper. 1992. Environmental variation and the persistence of
small populations. Ecological Applications 2:18-29.
Stauffer, D. 1985. Introduction to percolation theory. Taylor and Francis, London.
Taylor, P. D., L. Fahrig, K. Henein, and G. Merriam. 1993. Connectivity is a vital
element of landscape structure. Oikos 68:571-573.
Taylor, R. A. J ., and D. Reling. 1986. Density/height profile and long-range dispersal of
first-instar gypsy moth (Lepidoptera: Lymantriidae). Environmental Entomology
15:431-435.
Townsend, C. R. 1991. Exotic species management and the need for a theory of invasion
ecology. New Zealand Journal of Ecology 15:1-3.
Trakhtenbrot, A., R. Nathan, G. Perry, and D. M. Richardson. 2005. The importance of
long-distance dispersal in biodiversity conservation. Diversity and Distributions
11:173-181.
Travis, J ., and C. Dytham. 2002. Dispersal evolution. during invasions. Evolutionary
Ecology Research 4:1 1 19—1 129.
Van den Bosch, F., R. Hengeveld, and J. A. J. Metz. 1992. Analyzing the velocity of
animal range expansion. Journal of Biogeography 19:135-150.
Veit, R. R., and M. A. Lewis. 1996. Dispersal, population growth, and the Allee effect:
dynamics of the house finch invasion of eastern North America. American
Naturalist 148:255-274.
Vermeij, G. J. 1996. An agenda for invasion biology. Biological Conservation 78:3-9.
Villard, M. A., and B. A. Maurer. 1996. Geostatistics as a tool for examining
hypothesized declines in migratory songbirds. Ecology 77:59-68.
Wiens, J. A., and B. T. Milne. 1989. Scaling of 'landscapes' in landscape ecology, or,
landscape ecology from a beetle's perspective. Landscape Ecology 3:87-96.
Wiens, J. A., R. L. Schooley, and R. D. Weeks, Jr. 1997. Patchy landscapes and animal
movements: do beetles percolate? Oikos 78:257-264.
Williamson, M. 1999. Invasions. Ecography 2225-12.
With, K. A. 2002. The landscape ecology of invasive spread. Conservation Biology
16:1192-1203.
107
With, K. A. 2004. Assessing the risk of invasive spread in fragmented landscapes. Risk
Analysis 24:803-815.
With, K. A., and T. O. Crist. 1995. Critical thresholds in species' response to landscape
structure. Ecology 76:2446-2459.
With, K. A., and A. W. King. 1999a. Dispersal success on fractal landscapes: a
consequence of lacunarity thresholds. Landscape Ecology 14:73-82.
With, K. A., and A. W. King. 1999b. Extinction thresholds for species in fractal
landscapes. Conservation Biology 13:314-326.
Yang, D., B. C. Pijanowski, and S. H. Gage. 1998. Analysis of gypsy moth (Lepidoptera:
Lymantriidae) population dynamics in Michigan using geographic information
systems. Environmental Entomology 27:842-852.
108
11111111111111:will: