University of Southern California Dissertations and Theses

Assessing the transferability of a species distribution model for predicting the distribution of invasive cogongrass in Alabama

(USC Thesis Other)

Assessing the transferability of a species distribution model for predicting the distribution of invasive cogongrass in Alabama

PDF

Download a page range

Download transcript

Copy asset link

Request this asset

Transcript (if available)

Content Assessing the Transferability of a Species Distribution Model for Predicting the Distribution of Invasive
Cogongrass in Alabama
by
Rachel Eagle Shanks
A Thesis Presented to the
Faculty of the USC Graduate School
University of Southern California
In Partial Fulfillment of the
Requirements for the Degree
Master of Science
(Geographic Information Science and Technology)
December 2019
ii

Copyright © 2019 by Rachel Eagle Shanks

iii

To my mother, Angela Faye Eagle.

iv

Table of Contents
List of Figures ................................................................................................................................ vi
List of Tables ............................................................................................................................... viii
Acknowledgements ........................................................................................................................ ix
List of Abbreviations ...................................................................................................................... x
Abstract .......................................................................................................................................... xi
Chapter 1 Introduction .................................................................................................................... 1
1.1. Cogongrass ..........................................................................................................................4
1.2. Research Goals....................................................................................................................5
1.3. Study Organization and Structure .......................................................................................7
Chapter 2 Background .................................................................................................................. 10
2.1. Description of the Species ................................................................................................10
2.2. Modeling with Maxent ......................................................................................................16
2.2.1. Model Tuning...........................................................................................................17
2.2.2. Testing Maxent Results............................................................................................18
2.2.3. Transferability of Maxent Models ...........................................................................20
2.3. Related Research ...............................................................................................................20
Chapter 3 Data and Methods......................................................................................................... 26
3.1. Study Area ........................................................................................................................27
3.2. Scale of Study ...................................................................................................................32
3.3. Data Description ...............................................................................................................33
3.3.1. Species Presence Data..............................................................................................36
3.3.2. Soils Data Overview ................................................................................................39
3.3.3. Landcover Data Overview .......................................................................................48
3.3.4. Roads Data Overview ..............................................................................................51
v

3.4. Methods.............................................................................................................................54
3.4.1. Defining the Model ..................................................................................................54
3.4.2. Tuning the Model .....................................................................................................55
3.4.3. Gauging Fitness of the Model ..................................................................................57
3.4.4. Testing Transferability of the Model .......................................................................58
Chapter 4 Results .......................................................................................................................... 63
4.1. Area Under the Receiver Operating Characteristic Curve (AUC) ....................................64
4.2. Sensitivity (Omission) ......................................................................................................64
4.3. Variable Contributions and Gain ......................................................................................65
4.4. True Skill Statistic (TSS) ..................................................................................................77
Chapter 5 Conclusions .................................................................................................................. 80
5.1. Uncertainty in the Model ..................................................................................................81
5.2. Proposed Future Work ......................................................................................................82
5.3. Findings.............................................................................................................................83
References ..................................................................................................................................... 85
Appendix A: Soils Related Environmental Covariate Maps ........................................................ 90
Appendix B: Other Environmental Covariate Maps ................................................................... 111
Appendix C: Data Layer Conversion Steps ................................................................................ 120
Appendix D: Maxent Model Settings Screen Captures .............................................................. 122
Appendix E: Response Curves.................................................................................................... 125
Appendix F: Ecological Systems with Category Groupings ...................................................... 126

vi

List of Figures
Figure 1: Image of Imperata cylindrica (L.) Beauv. in Bloom in a Pine Plantation.. ..................... 1
Figure 2: Map of Cogongrass Infestation Presence Point Locations and Study Areas .................. 3
Figure 3: County Level Distribution with Density of Infestation Points of Cogongrass ................ 5
Figure 4: General Structure and Organization of the Project. ........................................................ 8
Figure 5: Morphology of Cogongrass. .......................................................................................... 12
Figure 6: Aerial View of a Cogongrass Infestation in a Young Pine Plantation. ......................... 15
Figure 7: High Level Overview of Maxent Steps ......................................................................... 26
Figure 8: Model Study Area with Cogongrass Infestation Presence Point Locations .................. 28
Figure 9: Test Area 1 with Cogongrass Infestation Presence Point Locations ............................. 29
Figure 10: Test Area 2 with Cogongrass Infestation Presence Point Locations ........................... 30
Figure 11: AFC Field Verified Cogongrass Infestation Lcations ................................................. 37
Figure 12: Depth to Restrictive Layer Thumbnail Images for each of the Study Areas. ............. 40
Figure 13: Drainage Class Thumbnail Images for each of the Study Areas. ................................ 41
Figure 14: The Soil Texture Triangle ........................................................................................... 42
Figure 15: Particle Size Thumbnail Images for each of the Study Areas.. ................................... 43
Figure 16: Percent Clay Content Thumbnail Images for each of the Study Areas.. ..................... 43
Figure 17: Percent Sand Content Thumbnail Images for each of the Study Areas. ..................... 44
Figure 18: Percent Silt Content Thumbnail Images for each of the Study Areas. ........................ 45
Figure 19: Soil pH Thumbnail Images for each of the Study Areas. ............................................ 46
Figure 20: Data Layer Creation Workflow for Soils Data ............................................................ 47
Figure 21: Data Preparation Workflow for Percent Canopy Layer. ............................................. 49
Figure 22: Percent Canopy Thumbnail Images for each Study Area. .......................................... 50
Figure 23: Ecological System Tumbnail Images for each Study Area. ........................................ 51
Figure 24: Distance to Nearest Road Thumbnail Images for each Study Area. ........................... 53
vii

Figure 25: Response Curves for Percent Canopy. ........................................................................ 67
Figure 26: Response Curves for Ecological System ..................................................................... 69
Figure 27: Consolidated Graph of Cogongrass' Response to Ecological System. ........................ 69
Figure 28: Response Curves for pH .............................................................................................. 71
Figure 29: Response Curves for Distance to Nearest Road. ......................................................... 71
Figure 30: Jackknife for the Model Study Area. ........................................................................... 74
Figure 31: Jackknife for Test Area 1. ........................................................................................... 75
Figure 32: Jackknife for Test Area 2. ........................................................................................... 76

viii

List of Tables
Table 1: General Biological Characteristics of Cogongrass ..................................................................... 11
Table 2: General Habitat Description of Cogongrass ............................................................................... 13
Table 3: Verified Point Locations, Total km
2
, and Average Points per km
2
. ........................................... 31
Table 4: Ecological System Categories. ................................................................................................... 32
Table 5: Datasets Used in the Study ......................................................................................................... 35
Table 6: Environmental Variables Used in the Study ............................................................................... 36
Table 7: Soil Variables Used in the Study ................................................................................................ 39
Table 8: Regularization Multiplier's Effect on AUC ................................................................................ 56
Table 9: Model Suitability Indicator Results ............................................................................................ 63
Table 10: Percent Contribution of Environmental Variables ................................................................... 66
Table 11: Permutation Importance of Environmental Variables .............................................................. 66
Table 12: Percent Geographic Area (km
2
) Occupied by Each Ecological System ................................... 70
Table 13: TSS for Each Replicate Run ..................................................................................................... 78

ix

Acknowledgements
I am grateful to the faculty and staff of the Spatial Sciences Institute at the University of
Southern California for the excellent program and exceptional professors and cohorts. I am
especially grateful to my advisor, Dr. Karen Kemp for her guidance and encouragement through
this process and ever-present patience as I worked through the challenges of life + graduate
school. I am grateful for the time and experience offered to me by my committee members, Dr.
Su Jin Lee and Dr. Katsuhiko Oda. This journey has been a long one, but one that has shaped my
life and career for the better.
I would also like to thank the Alabama Forestry Commission for providing the full
cogongrass presence point location dataset for use in this analysis as well as for guiding me
toward valuable contacts, resources, and literature used in this study. I would like to thank my
employer, Silvics Solutions, for providing the roads datasets used in this analysis and for being
my sounding board when ArcGIS baffled my sometimes less than analytical brain. I would also
like to send out special thanks to my manager and mentor, Greg Triplett, for allowing me the
time to focus effort on this project when timelines required, for his patience and understanding
when my priorities shifted, and for picking up my slack when I had to prioritize school over
work.
Finally, I would like to thank my family without whom I would not have been able to
juggle work, parenting, school, and life in general. Through all of the missed visits, distracted
phone calls, and general negligence of my responsibilities as a mother, daughter, sister, and
friend, we somehow came out of this still speaking to one another! I love you all and could not
have completed this journey without your understanding and support.

x

List of Abbreviations
ASCII American Standard Code for Information Interchange
AUC Area Under the Curve
CSV Comma-Separated Values
DEM Digital Elevation Model
GAP Gap Analysis Project
GIS Geographic information system
GISci Geographic information science
GIST Geographical Information Science and Technology
MSA Modeled Study Area
PRISM Parameter-elevation Regressions on Independent Slopes Model
ROC Receiver Operating Characteristic
SDM Species Distribution Model
SSI Spatial Sciences Institute
TSS True Skill Statistic
USC University of Southern California
USFS United States Forest Service
USGS United States Geological Survey

xi

Abstract
As of April 19th, 2018, there were 34,771 verified locations of cogongrass (Imperata cylindrica
(L.) Beauv.) infestations within the state of Alabama. Cogongrass is a highly invasive non-native
species of rhizomatous grass that is considered one of the ten worst weeds worldwide. This
highly invasive and environmentally destructive species has caused significant damage
throughout its current distribution and efforts to control and eradicate the threat have been
underway for almost a decade. This study utilized the Maximum Entropy (Maxent) model to
predict the location of invasive cogongrass within the state of Alabama. The model developed
using the presence locations and environmental data for the Model Study Area, one Alabama
Forest Commission (AFC) Work Unit, was applied to two additional AFC Work Units to test
transferability of the model to areas of similar and dissimilar ecological and geographic makeup.
The Model Study Area’s Maxent model resulted in an acceptable AUC (0.725 with sd = 0.0010)
and fair TSS score (0.4087) with a test omission rate of 0.0832. Transferability test results
differed between the two test areas. Using the Model Study Area’s model on Test Area 1, an area
similar in most aspects to the Model Study Area, resulted in an AUC of 0.746 with a standard
deviation of 0.002, a TSS score of 0.3944 and a test omission rate of 0.0807. These results
indicated that the original model was sufficiently transferable to the similar Test Area 1. Test
Area 2 was dissimilar from the Model Study Area in most environmental covariates as well as
number of verified presence point locations. Applying the model to Test Area 2 resulted in an
AUC of 0.846 with a standard deviation of 0.017, a TSS score of 0.2377 and a test omission rate
of 0.2941. These results suggest the need for some concern about the suitability of the transferred
model to Test Area 2.
1

Chapter 1 Introduction
Imperata cylindrica (L.) Beauv., commonly known as cogongrass (Figure 1) is a highly invasive
and environmentally destructive non-native species with serious biological, environmental, and
economic impacts to the Southeastern United States. In fact, as of October 22
nd
, 2018, the U.S.
Forest Service website lists cogongrass as “one of the 10 worst weeds worldwide and a pest in 73
countries.” Cogongrass, like most non-native
invasive species, can become an agent of
change in the ecosystem within which it
becomes established. As an agent of change, the
species can have a deleterious effect on native
biodiversity (McNeely 2001).
In an effort to better understand the
distribution and potential infestation threat of
invasive species, ecologist use tools such as
species distribution models (SDM) to assist in
their understanding of the potential species
spread and to plan for appropriate management
actions related to the species being studied. The
Maximum Entropy Model (Maxent) is a SDM which is commonly used by ecologists to study
the current and predicted future distribution of a species given presence-only datasets. The use of
this model helps researchers better predict infestation points based on environmental factors and
assist in their efforts to eliminate this blight by guiding eradication funds to appropriate areas of
high risks for infestation. Limited funding necessitates that eradication efforts must be focused
Figure 1: Image of Imperata cylindrica (L.)
Beauv. in bloom in a pine plantation. Image
curtesy of Chris Evans, University of Illinois
with permission via bugwood.org.
2

on areas that respond best to treatment to ensure maximum benefit to the environment,
community, and rural economy. In this analysis, Maxent was used to model the predicted
potential distribution of cogongrass infestation given suitable conditions within Alabama
Forestry Commission’s (AFC) Work Unit 11, which in this document is referred to as the Model
Study Area. The resultant model was then transferred to two other study areas to test model
transferability for the species and to theorize the potential for model transferability across the
state. The two transferability test areas were selected so that Test Area 1 was highly similar in
ecological niche and number of verified infestation point locations to the Model Study Area and
Test Area 2 is dissimilar. These study locations are shown in Figure 2.
3

Figure 2: Map of cogongrass infestation presence point locations with the Model Study Area and
two transferability test areas defined. The Model Study Area is outlined in cyan and the two Test
Areas are outlined in Magenta.
4

1.1. Cogongrass
Imperata cylindrica (L.) Beauv. (cogongrass), is a highly invasive non-native species of
rhizomatous grass that was originally introduced in the southeastern United States accidentally in
1912 as packing material in shipping crates from Japan for imported goods at the Port of Mobile
in Grand Bay Alabama (Tabor 1949; Tabor 1952; Dickens 1974; Dozier 1998; MacDonald 2004;
Damghani 2013). The species was later intentionally introduced from the Philippines in
Mississippi (Tabor 1949; Tabor 1952; Dickens 1974; Dozier 1998; Ervin and Holly 2011) and
Florida in the 1920s and 1930s by the USDA as forage and for erosion control (USDA NRCS
Plants Database). The var. rubra variety (a non-invasive ornamental cultivar) of Imperata
cylindrica is still sold by the nursery industry in some states as an ornamental grass under the
name Japanese Blood Grass, or Red Baron, (Dozier 1998; Missouribotanicalgarden.org, last
accessed 11/4/2018), however all other varietals are listed as a Federal Noxious Weed under the
Plant Protection Act, which limits its transport between states without an appropriate permit.
Currently the range of verified infestations of cogongrass within the continental United
States spans from East Texas, Southeast to South Florida, and as far north as North Carolina,
according to the Early Detection and Distribution Mapping System website developed by The
University of Georgia – Center for Invasive Species and Ecosystem Health (EDDMapS 2019).
Figure 3 shows a map of this distribution.
5

1.2. Research Goals
The research objectives of this study were two-fold. The first objective was to evaluate
the fitness for use of Maxent in modeling the predicted potential distribution of cogongrass
infestation given suitable conditions using the selected environmental covariates within the
Model Study Area. The second objective was to test the transferability of that model to other
study areas within the state of Alabama.
Alabama Forestry Commission Work Units were used to delineate the boundaries of
study areas within this project. AFC Work Unit 11 was selected as the Model Study Area
because the area contains a large verified point location dataset to use in the model (9242 points)
and this AFC Work Unit contains the transferability study area from the Ervin and Holly (2011)
study (Clarke County, Alabama) that initially sparked my interest in model transferability.
Percent canopy cover was shown to be the most influential variable on the Ervin and Holly
Figure 3: County level distribution with density of infestation points of cogongrass
verified sites. Source: EDDMapS 2019
6

Mississippi model, and therefore we hypothesize that percent canopy will have significant
influence on the models produced in this study as well.
The two test study areas were selected based on their similarity and dissimilarity to the
Model Study Area. It was hypothesized that Test Area 1, which is relatively similar in
environmental covariate values to the Model Study Area, will have a similar model result to the
Model Study Area. Further, it is hypothesized that Test Area 2, which is relatively dissimilar in
environmental covariate values to the Model Study Area, will have dissimilar model results to
the Model Study Area but will still produce an acceptable model.
The guiding motivation, beyond the desire to generate an appropriate model that is
transferable across various areas of the state, is the hope that the resulting model and
transferability tests will be useful in directing future survey efforts and funding decisions for
implementing control and eradication measures against invasive cogongrass in the state of
Alabama. Evaluating the model will help researchers better predict infestation points based on
environmental factors used and assist in their efforts to eliminate this threat by guiding survey
and eradication funds to appropriate areas of high risk for infestation. Invasive species
management has been shown to be more effective when management activities occur in the early
stages of infestation as attempted management of large, well-established colonies of invasives is
difficult and cost prohibitive (Ervin and Holly 2011). Limited funding necessitates that
eradication efforts must be focused on areas that respond best to treatment to ensure maximum
benefit to the environment, community, and rural economy. The use of Maxent to facilitate
targeted survey and eradication efforts is possible only if this type of SDM can be shown to be a
useful tool in predicting the distribution and spread of this species and is transferable across the
7

affected area. In addition, the results of this study can be used to further prompt research into this
species as well as the use of Maxent in predicting species distribution.
1.3. Study Organization and Structure
This study was structured to first define an appropriate Maxent model for cogongrass in a
specific area in the state of Alabama and then test the transferability of that model to other areas
within the state. Figure 4 depicts the overall study workflow. First the project goals and species
were defined. Then the model study area and study related questions were reviewed. These
questions, and the answers to them, as gleaned from research, guided the definition of the species
and environmentally appropriate datasets needed to complete the study. Once the required
datasets were identified, the data was prepared for use by Maxent using Esri’s ArcGIS 10.6
Desktop. The prepared data was then used within Maxent 3.4.1. A baseline model was trained
utilizing all gathered datasets and all Maxent default values and then the model for the Model
Study Area was tuned through iterative runs where maxent settings were modified and
environmental covariates that were deemed to add little to no added value to the model were
removed. A final model was created for the model study area and results were analyzed to verify
fitness for use given the species and environmental extent of the study. The model produced for
the Model Study Area was then used to test transferability to the two test areas.
8

Figure 4: General structure and organization of the project.
The remainder of this document is broken into four additional chapters. The next chapter
provides background context and additional information pertinent to the species being studied,
the modeling method used, and research that guided the decisions made as this project
progressed. Chapter 3 describes the data included in this study in detail, as well as the methods
used to build the models generated by this project. Chapter 4 discusses the results of the models
produced and specifically focuses on the key statistics for judging model fitness. Finally, Chapter
9

5 includes a discussion of the conclusions gleaned from this project and the models generated in
the process of this study.

10

Chapter 2 Background
This chapter provides background context and additional information pertinent to the species
being studied, the modeling method used, and research that guided the decisions made as this
project progressed. This chapter begins by describing, in detail, the morphology and biological
characteristics of the species and the habitat range in which it can grow. The chapter then
continues by discussing Maxent as a tool for modeling and the tuning and testing of the output
model. Finally studies pertinent to the decisions made in this study are reviewed.
2.1. Description of the Species
Cogongrass has many alternate common names throughout the world. It is also known as
kunai grass, blady grass, japgrass, alang-alang, lalang grass, as well as many others and is often
confused with Brazilian satintail (Imperata brasiliensis) which is a closely related species in the
genus. Cogongrass is fast-growing and can spread by rhizomatous shoot up to 4m
2
in as little as
11 weeks on productive sites (Dozier 1998; Wilcut et al. 1988a). The general biological
characteristics of the species are defined in Table 1 below. The species is stemless, forming rigid
leaf tufts developing directly from the rhizomes. Leaves can grow to 150cm in height and 4 to
10mm in width and have very sharp pointed apex. They have an off-center midrib that is white in
appearance and has finely serrated margins (Estrada and Flory 2014; Dozier 1998). Cogongrass
has a high rhizome to shoot ratio which increases its regenerative ability and exhibits allelopathic
tendencies which inhibit growth of competing native species (MacDonald 2004).
11

Table 1: General biological characteristics of cogongrass
Biological
Characteristics
Description
Reproduction Vegetative and seed
Flower
Branched panicle with dense white fluffy spikelets growing
10-20cm long
Growth Structure
Stemless, grows in loose tufts with leaves emanating from
rhizomes
Leaf Blade
Long and slender; 15-150cm tall; 4-10mm wide with off center
white midrib
Root Structure Rhizomes with attached dense fibrous root system

Rhizomes are the primary mechanism for local spread of the species once invasion has
occurred. Rhizomes are aggressive, hardy, branched, and grow in dense clumps. These clumps
form dense mono-species mats that impede the growth of other species that would otherwise
utilize that environment. In a 1977 study by Lee et al., rhizome density was measured to be 89m
(linear) per square meter of soil. Rhizome clumps restrict access to nutrients needed for native or
commercial species to thrive, further harming both the biological and economic environments in
which it is found. Rhizomes are whiteish in color with short, scaly nodes and sharp barbed tips
that can penetrate the roots of other species (Dozier 1998). The morphology of the species is
represented in Figure 5. It is important to note that buds do not form until the third or fourth leaf
stage of the plant’s life cycle. This is also when dense root development begins (Dozier 1998).
The importance of this is due to planned timing of infestation eradication. For cogongrass,
invasive species management has been shown to be more effective when management activities
occur in the early stages of infestation as attempted management of large, well-established
colonies of invasives is difficult and cost prohibitive (Ervin and Holly 2011). Therefore,
eradication efforts will be less costly if management activities can be conducted in young plants.
12

Figure 5: Example morphology of cogongrass. Images reproduced by permission from
bugwood.org.
Panicle flower heads are 5-20cm long and silvery-white. The panicle is fuzzy giving the
flower a soft cottony look (EDDMapS 2019). Some studies suggest that flowering occurs
generally after disturbance or stress but recent studies counter that thought and show that
cogongrass produces an abundance of seed even without disturbance or stress and the seeds are
easily distributed by wind. Each plant can produce up to 3000 seeds annually (MacDonald 2004;
Dozier 1998; Wilcut et al. 1988a; Holm 1977). Cogongrass spreads locally via rhizome growth
and long-distance via seed dispersal. Wilcut et al. (1988a) state the average flight of a one-
seeded spikelet was 15m, and a 2011 study by Yager, Miller, and Jones measured the maximum
flight distance for a spikelet, with seed removed, to be 37m in a pine-tallgrass environment.
Studies have suggested that the West to East wind patterns along major roads and Interstate
highways has created a dispersal route for cogongrass infestation by seed (Yager, Miller, and
Jones 2011; MacDonald 2004; Wilcut et al. 1988a; Hubbard et al. 1944).
Cogongrass is a very hardy species and is tolerant of shade, high salinity, moisture and
drought. The general habitat description of the species is defined in Table 2. Cogongrass grows
13

in tropical and subtropical climates ranging in latitude from 45°N to 45°S. occurring in a wide
range of ecological conditions (MacDonald 2004; Holm et al. 1977). Cogongrass thrives in
minorly disturbed sites and non-disturbed rural sites but not heavily disturbed sites. The species
has been shown to thrive along roadways, in pastures and mining sites, pine forests and other
open areas, but does not thrive in areas of heavy cultivation and repeated tillage (Dozier 1998;
Willard et al. 1990) Therefore one mechanism for control is repeat tillage treatments when
infestation occurs on sites where tillage is possible, and it would be expected that cogongrass
would not infest agricultural row crop sites where repeat heavy tillage occurs.
Table 2: General habitat description of cogongrass
Habitat Description
Range Tropical and subtropical climates (Latitudes 45ºN to 45ºS)
Site Highly adaptable (occurs in a wide range of ecological conditions from
xeric uplands to shaded mesic sites)
• Degraded forests, roadsides, arable land, young plantations, sandhills,
flatwoods, hardwood hammocks, grasslands, river margins, swamps,
scrub, and wet pine savanna communities
• Thrives in areas of minimal tillage and frequent burning
• Tolerant of varied soil conditions including variations in fertility,
organic matter and moisture
• Grows best in relatively acidic soils (pH 4.7)
• Relatively intolerant of shade
Rainfall 75 to 500cm
Elevation Sea level to 2000m
Temperature -4.5ºC or lower for more than 24 hours is lethal to rhizomes (however
dense thickets can insulate themselves and may survive temperatures as
low as -14ºC.

In 2009, the Alabama Forestry Commission received a three-year, 6.3-million-dollar
grant from the American Recovery and Reinvestment Act to initiate a proactive, coordinated
campaign to eradicate cogongrass in the state of Alabama. The Commission for the Campaign
Against Cogongrass was formed to detect, map, and plan an effective program for the eventual
14

eradication of cogongrass from the state. This grant was sufficient to get the initiative started;
however more funding is necessary to win the war on cogongrass (Bargeron 2009).
According to the U.S. Fish and Wildlife Service Invasive Species website, invasive non-
native species do not have the natural checks and balances that native species would have in an
ecosystem. Therefore, when a non-native species is introduced to a new environment, it can
become invasive if there are no natural elements to restrict its propagation. This invasion of the
ecosystem by a non-native species can have deleterious effects on the ecosystem, the economy,
and human health (US Fish and Wildlife Service last accessed 11/4/2018). Cogongrass has been
an invasive non-native species in the southeastern United States since its introduction in 1912
and has been shown to have significant impacts both economically and environmentally in
heavily infested areas. Figure 6 shows an aerial view of the impact of cogongrass infestation in a
young pine plantation. Cogongrass is a highly adaptive invasive species with a broad tolerance to
environmental and ecological conditions. Therefore, this species has the potential to adversely
change the structure and diversity of environments in which it invades. Cogongrass has been
linked to the reduction of native diversity and alteration of ecological processes within infested
ecosystems, especially in fire-dependent communities (Lippincott 2000). This highly invasive
and environmentally destructive species has caused significant damage throughout its current
distribution and efforts to control and eradicate the threat have been underway for almost a
decade.
15

Figure 6: Cogongrass infestation in a young pine plantation exhibiting its distinctive circular
infestation pattern and severity of infestation at occurrence locations. Image courtesy of Greg
Leach, International Paper via Bugwood.org.
Economic stressors resulting from the establishment of cogongrass include cost of
eradication, impact of eradication efforts on native species and agricultural crops, financial loss
due to disturbance, etc. (Hubbard et al. 1944; Soerjani 1970; Eussen et al. 1976; Daneshgar et al.
2008). Studies have also shown that cogongrass is particularly problematic in agricultural
systems where the species directly competes with agricultural crops for both space and nutrients.
This competition reduces crop yields and increases weed control costs (Ervin and Holly 2011;
Akobundu and Ekeleme 2000; Terry et al. 1997). Cogongrass has been the subject of numerous
and diverse studies throughout the Southeastern United States and therefore is a good candidate
for studying the effectiveness of a Maximum Entropy model (Phillips et al. 2017) in predicting
its distribution and testing transferability of the same.
16

2.2. Modeling with Maxent
SDMs are routinely used to predict the potential distribution of a species based on known
point locations. The species distribution modeling used in this study is the maximum entropy
method using Maximum Entropy Species Distribution Modeling (Maxent) Version 3.4.1
(Phillips et al. 2017). Maxent uses presence only species location points and environmental
variables to develop probability models for the distribution of the species being modeled
(Phillips et al. 2006; Ervin and Holly 2011; Elith et al. 2011). Presence locations are compared to
the environment through the use of background points (Crall et al. 2013).
Maxent has been gaining in popularity and use in the fields of ecology and environmental
sciences and has been shown to outperform other species distribution modeling methods in
predictive accuracy (Merrow, Smith, and Silander 2013). Machine Learning models such as the
Maxent model allow the model to “learn” from iterative model runs given a sample known
dataset to train the model and a sample known dataset to test the model’s understanding of the
data and associated environmental layers.
Maxent is a machine learning model that is well suited for species distribution modeling
(Phillips et al. 2017). Maxent models are good predictors of species distribution with limited
datasets and work especially well with presence only data (Phillips et al. 2017; Ervin and Holly
2011; Elith et al. 2011). Before running the model, the biology and ecological niche of the
species being studied must be closely examined to ensure that selected environmental variables
align with the biological and environmental factors that influence the species distribution within
the intended study area (Manzoor, Griffiths, and Lukac 2018). This species review and
conscientious environmental data selection can be time consuming but is vital to the production
of an informed SDM. The Maxent model allows for an understanding of which environmental
17

variables are most important to the distribution of the species being modeled and gives a
relatively unbiased prediction based on the constraints provided. The model requires a training
dataset, a testing dataset, and environmental layers to act against the model as predictors
(Merrow, Smith, and Silander 2013). The training and testing datasets are presence only data and
can be subsets of the same original larger dataset.
As Maxent is a presence only modeling application, the pseudo-absence (background)
points are generated by the model. It is important to note that some sampling bias may be
introduced into the model if your species presence point distribution does not cover the entire
range of your study’s geographic extent. If this is the case, you can create a background file in
ArcGIS to use within Maxent to limit where the model predicts background points so that it does
not create background points outside of the extent of the presence data points. This can be done
using the Create Minimum Bounding Geometry tool in ArcGIS then converting the output
polygon to raster (.asc) format for use in Maxent. In this study, no minimum bounding geometry
was needed as the presence points spanned the entire extent of the Model Study Area.
2.2.1. Model Tuning
In Maxent, model tuning is performed to optimize model complexity and fit. Tuning the
model smooths the response curves to the specific environmental variables included in the model
to reduce overfitting (Elith et al. 2011). Maxent provides default settings for parameters that
were determined to be the average optimal values (Phillips and D udı k 2008), however, it is
recommended that these settings be tuned for the specific species and region of study
(Radosavljevic and Anderson 2014).
To assist with potential issues related to spatial autocorrelation, a baseline (neutral)
model run can be performed with all parameters set to default. Based on the results of the
18

baseline model run, the model settings can be tuned until an appropriate model output from the
sample data is derived. Modifications to parameters, constraints, and environmental layers
included should be based on research of similar studies (Merrow, Smith, and Silander 2013).
Regularization is an available parameter in Maxent that relaxes the environmental
constraints so that the predictions do not have to fit the constraints exactly. This allows the
model to ignore variables that don’t impact the model and to determine the most impactful
variables on the model output. Regularization protects against overfitting by affecting how
closely the output distribution is fit to the provided presence data. To get a closer fit (more
localized output distribution) the regularization multiplier can be reduced (less than 1). To get a
more spread out distribution, increase the regularization multiplier (greater than 1). Care should
be taken if the regularization multiplier is modified to avoid overfitting or underfitting of the
model (Phillips et al. 2017).
2.2.2. Testing Maxent Results
The Maxent model provides a robust testing set for measuring uncertainty. Model-based
uncertainty methods in Maxent models are found in the form of sensitivity and uncertainty
analysis. Confrontational methods include visual tests and statistics-based tests. The Maxent
model utilizes both tests in evaluating the outcome of the model. Visual tests can be performed
from the layers that are generated from the model that can be rendered as graphics for
understanding the model outcomes. Statistics based tests including sensitivity, specificity,
threshold dependence plus standard deviation, and regularization help to describe and reduce
uncertainty. And finally, Maxent can be run iteratively with different parameters and or
constraints to heuristically observe patterns that arise from the model runs.
19

Indicators of model fitness include the area under the receiver operating characteristics
curve (AUC), Omission rate, and True Skill Statistic (TSS). The AUC measures the accuracy of
the model in predicting distribution based on sample data. The closer the AUC is to 1, the better
the model is at predicting the distribution. In the graphic output provided by Maxent, the mean
AUC is shown as the area under the red line and the steeper the angle of increase the closer the
AUC value is to 1. An AUC approaching 0.5 means the model cannot predict class separation
and therefore cannot predict the distribution at all given the input parameters (the random model
is the 1:1 random prediction line depicted in black on the graph in Maxent’s output .html file).
An AUC approaching 0 indicates reciprocity in the prediction (Narkhede 2018). The receiver
operating characteristics (ROC) curve itself is the probability curve measuring the probability the
model is a good fit for the data and question being answered. In general, AUC above 70% is
considered “sufficiently accurate to be used in conservation planning” (Elith et al. 2006, 141).
Since population size is generally not known but estimated, Maxent cannot produce true
occurrence rate per grid cell in the analysis. Sensitivity is a rating of how well the model predicts
positive outcomes (or presence). This is the omission rate of the model. Specificity is the
measurement of how well the model predicts negative outcomes (or absence). In other words,
specificity measures the percentage of absence points that are reported as presence (false
negatives) based on modeled probability. This is the commission rate of the model (Phillips
2017; Phillips et al. 2006, Elith et al. 2011; Anderson 2012). However, true commission cannot
be measured with presence only data. Sensitivity, however, can be used as a measure of fit along
with AUC. Sensitivity and specificity are inversely related. If we decrease the threshold (more
positive values) we increase the sensitivity (fewer false negatives) and decrease the specificity
(more false positives) (Narkhede 2018). When using AUC as a measure of model performance, it
20

is recommended that omission and commission rates be included in the evaluation where
possible (Lobo et al. 2008). Finally, The True Skill Statistic (TSS) is a variation on Kappa that
mitigates issues of prevalence that limit the use of Kappa in presence only models like Maxent
(Allouche, Tsoar, and Kadmon 2006).
2.2.3. Transferability of Maxent Models
Transferability (also called projection) of SDMs has been an issue of concern in research
studies as models built in one geographic space do not always project well to different
geographic space and/or time. To maximize transferability, the environmental layers used within
the model must align with the requirements of the species but should be broad enough to
encompass the entire extent of the originally modeled area and the intended projection area. This
alignment is necessary to allow the model to be transferred across space or time for the specific
species under review (Anderson 2012; Peterson et al. 2011). Model tuning is recommended to
maximize suitability of the model for the species and location being modeled and is especially
important when the ability to transfer the resultant model is a desired outcome of the study
(Radosavljevic and Anderson 2014).
2.3. Related Research
Background research for this thesis included review of research in the following areas:
Studies which similarly used the Maxent model to predict invasive species distribution, research
on model transferability, research on modeling distribution of cogongrass specifically, and
research on Maxent model parameters. There have been several studies of invasive species,
21

including cogongrass, utilizing the Maxent model that are useful background references for this
study.
Amanda West et al. (2016) endeavored to predict invasive species distribution of
cheatgrass (Bromus tectorum L.) utilizing Maxent. West states that presence-only models are
rarely evaluated against real field data, therefore, the authors determined to test field data
collected over a period of time against the Maxent presence only model. Presence data collected
between 2007 and 2013 were used as the sample data for the study. West et al. ran a Maxent
model across the area in 2007 using limited real data. Then, the Maxent model was rerun using
the new data from 2008 to 2013 using same parameters as used in 2007 to test the accuracy of
the previous results. A new model with updated parameters was also run and was tested with the
same sample dataset collected between 2008 and 2013. West used area under the curve (AUC),
percent correctly classified (PCC), sensitivity, specificity, and true skill statistic (TSS) to
evaluate and validate the models. The West et al. study concludes that the Maxent model is a
good fit for measuring the distribution of invasive cheatgrass in the Rocky Mountain National
Park .
Ervin and Holly (2011) performed a similar study at Mississippi State University on
cogongrass in southeastern Mississippi in an attempt to determine if the Maxent model they
designed for the Mississippi varietal would transfer to appropriately predict the distribution of
the Alabama varietal testing their Mississippi model against three subsets of Alabama
cogongrass data from the same geographic area but collected in three different years. They
determined that there was low transferability of the Mississippi model from Mississippi to
Alabama but noted several potential reasons for this low transferability including the landscapes
that were focused on (Mississippi focused on roadways while Alabama focused on managed
22

timberland) and discrepancies in soils. These concerns related to environmental factors affecting
transferability support the use of landcover data, distance to roads layer data and soils related
variables in the Maxent model generated in the current study. Another important point here is
that the lineage of cogongrass in Mississippi is from the Philippines while the lineage of
Alabama cogongrass is from Japan. That genetic difference may also be a factor in how the
species responds to environmental factors within the study (Lucardi, Wallace, and Ervin 2011).
A study of genetic impact is out of scope for this analysis; however, it is worth noting as lineage
can play a role in species response to environmental stimuli. Importantly, Ervin and Holly’s
Mississippi dataset was collected in a different fashion and for a different purpose (different
landscape focus as mentioned above) than the Alabama dataset used in that study. This could
have had an impact on the poor transferability of the model.
A 2005 case study on cogongrass published by the US. Forest Service indicated that
cogongrass may outcompete native species in poor soils due to its dense rhizome mat (Howard
2005) allowing the invasive to restrict access to soil nutrients and water for native grasses
(Howard 2005; Lippincott 1997). Cogongrass rhizomes have been shown to be present in the top
15cm of fine textured soils or top 40cm of coarse textured soils (MacDonald 2004). Howard also
noted that native species that outcompete cogongrass successfully generally have deeper root
systems or taller crowns, although MacDonald noted that cogongrass rhizomes formation may be
present to depths of 120cm (MacDonald 2004, Holm 1977). This study supports the use of soils
related variables such as depth to soil restrictive layer, soil texture layers, and drainage class as
environmental covariates for use with Maxent.
A 2000 study by King and Grace examined soil moisture content’s effect on cogongrass
seedling germination and growth, testing soil saturation ranging from dry to inundated.
23

Measurements of plant height and number of shoots were used to define seedling growth rate and
germination success respectively. This study found that cogongrass seedling germination was
weakest (reduced by 74%) when soils were inundated and that growth became increasingly
restricted, especially for smaller seedlings, as soil saturation increased. The authors suggested,
based on the results of their study that soil inundation in the early stages of cogongrass
establishment could restrict invasion by seed.
Roads as a pathway for seed dispersal was reviewed in a 2017 study by Rauschert,
Mortensen and Bloser. In this study, the authors followed physical seed dispersal of Carthamus
tinctorius L.(safflower) seeds, by routine rural road maintenance equipment, specifically by road
graders, on rural dirt roads. Safflower seeds were used in this study as the use of invasive seed
was restricted. The authors placed four patches of 5000 painted seeds in a grid across a rural road
that was planned to be graded using a typical three pass approach. Then, immediately following
the road grading event, Rauschert, Mortensen and Bloser measured the distance that seeds
traveled based on seed starting location and ending location. The study found that 41.8% of
seeds moved between 10 and 50 meters and only 1.6% traveled greater than 50 meters with a
maximum movement of 273m. This study focused on the physical movement of seed by road
maintenance equipment, however hitchhiking of seeds on vehicles and wind dispersal along
roadways was not included in this study but has been identified as additional key pathway for
dispersal related to transportation corridors. This study as well as mention of roads ad vectors for
dispersal in other studies provides incentive to include distance to road as an environmental
variable within the current study on cogongrass (Rauschert, Mortensen and Bloser 2017).
A study on woody shrubs as a barrier to wind dispersal of cogongrass seed (Yager,
Miller, and Jones 2011) was performed at Camp Shelby Training Site in Mississippi. This study
24

tested the travel distance of cogongrass spikelets (seed removed) released along three sites which
each contained blocks of pine-shrub forest and pine-tallgrass forest. The goal of the study was to
determine if forests with a woody shrub mid-story reduced the dispersal of cogongrass spikelets
and therefore reduced invasive introduction to forested areas along roadways. The study found
that although mean dispersal distance of cogongrass spikelets was not significantly different
between the two forest types, that more spikelets traveled further in the pine-tallgrass forest
(25% dispersed further than 5m) than did in the pine-shrub forest (8% dispersed further than 5m)
and the mean maximum dispersal distance was greater in the pine-tallgrass forest (37m) than in
the pine-shrub forest (23m). The study concluded that cogongrass dense woody shrub vegetation
along forest edges may impede the wind dispersal of cogongrass spikelets and subsequent
invasion of the species to the forest interior however it does point out that, in areas where
cogongrass is already present, infestation growth may still occur via vegetative spread as the
species some shows tolerance to shade.
A 2018 study reviewed the impact of grain size of predictor variables on the accuracy and
transferability of SDM models specifically using Maxent to test transferability for an invasive
plant species (Rhododendron ponticum (L.)) in Wales, U.K. (Manzoor, Griffiths, and Lukac
2018). The authors noted that the selection of grain size in SDMs is often dependent on the
availability of appropriate predictor variable (environmental covariate) data and the available
resolution of that data. As noted by the authors, finer grain size allows for more detailed and
potentially more accurate prediction of suitable environmental habitat and courser grain size
inhibits habitat delineation. Maxent requires all environmental covariates to utilize the same
grain size and therefore, The authors focused this study on comparing Maxent outputs of three
models. The modeled grain sizes were 1km, 300m, and 50m. For the 50m model, biophysical
25

variables of Altitude, Aspect, Slope, Land Cover, and Distance from water channels were used in
the model as environmental covariates. These same datasets were resampled to 300m to be used
in the 300m model and also resampled to 1km for the 1km model (see Manzoor, Griffiths, and
Lukac 2018 for details on methods used). For the 1km model, bioclimate variables were also
included as is common in SDM studies (Manzoor, Griffiths, and Lukac 2018). Model
transferability at all three grain sizes was also tested to determine if grain size has an impact on
the transferability of the model.
This study used the Continuous Boyce Index (CBI) to test transferability. The study
results show that CBI improves as grainsize is reduced in both the training model area as well as
the transfer test area. In the training model area, the CBI improved from 0.825 for the 1km model
to 0.895 for the 300m model to 0.964 for the 50m model. In the transfer test area, the CBI
improved from 0.65 for the 1km model to 0.90 for the 300m model but dropped to 0.77 for the
50m model. The reduction in CBI between the 300m and 50m models in the transfer test area
was attributed to differences in range and topography of the two geographic areas of study. The
authors concluded that, although the use of climate data is widely used in SDMs and in many
cases this is justified, biophysical variables based on the biology and ecology of the species
being studied as well as the spatial extent of the study area may be more important for localized
studies. Therefore, the authors suggest that the use of course grained climate datasets should be
considered in reference to their overall importance to the specific species and geographic extent
of the study, and that the inclusion of these climate datasets can produce less accurate SDMs due
to the required coarser grain size of the model.

26

Chapter 3 Data and Methods
As discussed in Chapter 1, this study focuses on Imperata cylindrica (L.) Beauv., more
commonly known as cogongrass, which is a highly invasive species with high tolerance to a
broad range of environmental conditions. Maxent was utilized to model the predicted potential
distribution of cogongrass infestation given suitable conditions for the AFC’s Work Unit 11
(Model Study Area) and then transferred to Work Units 12 (Test Area 1) and 8 (Test Area 2) to
test model transferability across the state.
As described in Figure 7,
the use of Maxent for species
distribution modeling has four
key steps. First, the species was
researched to ensure that
environmental variables selected
for the model are relevant to the
species and study location. This
is discussed in more detail in the
Data Description section below. Second, the datasets were cleansed, and data layers were
prepared. This included both the species presence data as well as the environmental variables.
Third, the Maxent model was run at default and then tuning occurred to ensure parameter
settings were appropriate for the study. Finally, the Maxent results were evaluated and model
performance was determined.
The ultimate goal of this study was to determine if Maxent is an appropriate tool to
predict cogongrass distribution and, if so, to determine if a locally constructed model could be
Figure 7: High level overview of Maxent steps
27

transferred to other areas within the state successfully. This study endeavors to create a model
that can be transferred to each work unit and reliably predict cogongrass infestation locations to
help guide survey and eradication efforts by the AFC.
3.1. Study Area
The primary Model Study Area (Figure 8) is the Alabama Forestry Commission’s (AFC)
Work Unit 11 encompassing 10,902 km
2
with an average point per km
2
of 0.85. This study area
includes Choctaw, Marengo, Clarke, and Washington counties. The Model Study Area was
chosen because the area contains a large verified point location dataset to use in the model (9,242
points) and this AFC Work Unit contains the transferability study area from the Ervin and Holly
2011 study (Clarke County, Alabama) that initially sparked my interest in model transferability.
Transferability of the resultant model was then tested against similar as well as dissimilar Work
Units within the state. Test Area 1 (Figure 9) consists of AFC Work Unit 12 which encompasses
7,306 km
2
with 6826 presence points that fall within the boundary of the study area and an
average point per km
2
of 0.93. Test Area 2 (Figure 10) consists of AFC Work Unit 8 which
encompasses 8,088 km
2
with only 78 presence points that fall within the boundary of the study
area and an average point per km
2
of 0.01. Table 3 shows the comparison of area and number of
points in each of the Alabama Work Units.
28

Figure 8: Model Study Area with Cogongrass Infestation Presence Point Locations Identified
29

Figure 9: Test Area 1 with Cogongrass Infestation Presence Point Locations Identified
30

Figure 10: Test Area 2 with Cogongrass Infestation Presence Point Locations Identified
31

Table 3: Count of verified point locations, total km
2
, and average points per km
2
in each AFC
Work Unit.
Work Unit Count of
Points
Square
Kilometers
Average
Points per
Square
Kilometer
1 5 6,468 0.00
2 47 6,688 0.01
3 2,355 10,378 0.23
4 4,253 7,867 0.54
5 9 7,009 0.00
6 22 5,840 0.00
7 6 8,347 0.00
8 78 8,088 0.01
9 13 6,689 0.00
10 1,898 6,700 0.28
11 9,242 10,902 0.85
12 6,826 7,306 0.93
13 7,389 7,302 1.01
14 256 7,219 0.04
15 435 5,330 0.08
16 88 6,425 0.01
17 1,557 5,916 0.26
18 262 6,702 0.04
Total 34,741 131,176
Average 1,930 7,288 0.24

As can be seen in Table 4, Work Unit 11 is mostly rural with 50% upland
forest/woodlands and 27% floodplain forest. Eight percent of this area is in agricultural use and
only three percent is developed. Like the Model Study Area, Test Area 1 is very rural in nature.
This area includes 47% upland forest/woodlands, 22% floodplain forest, 15% agricultural use,
and is three percent developed. Test Area 1 has the closest number of presence points per km
2
to
the Model Study Area. Based on visual inspection of the environmental layers using ArcGIS
10.6, it was determined that Test Area 1 has similar distribution of land use, PctClay, PctSilt,
PctSand, drainage class and PctCanopy with higher pH and slightly more agricultural use. (See
Appendix A and Appendix B for maps of each environmental layer used in the analysis). It was
32

therefore expected that Test Area 1 would display a similar predicted distribution to the Model
Study Area.
Table 4: Ecological System categories for comparison of study areas’ land use differences.
Category Model Study Area
% of area
Test Area 1
% of area
Test Area 2
% of area
Forest/Woodlands 49.99 46.88 52.34
Floodplain Forest 26.54 22.24 0.14
Agriculture 8.07 14.64 20.84
Developed 3.18 3.09 15.05
Disturbed 10.99 12.39 8.19
Water 1.18 0.68 2.80
Other 0.06 0.07 0.64

Test Area 2 includes the city of Birmingham, the largest city in the state, and is the most
urban area within the state of Alabama. This area’s land use includes: 52% upland
forest/woodlands, <1% floodplain forest, 21% in agriculture and is 15% developed (Table 4).
Again, based on visual inspection of the environmental layers, it was determined that Test Area 2
has significantly greater variability in depth to restrictive layer, lower PctCanopy, more PctClay
and PctSilt and less PctSand than the Model Study Area. Test Area 2 was included to test
transferability of the model defined for the Model Study Area to an area of dissimilar
environmental makeup.
3.2. Scale of Study
A key component of gridded data, such as the ASCII files used by Maxent, is the grain
size which represents the spatial resolution of the layers to be included in the analysis (Manzoor,
Griffiths, and Lukac 2018). For a model such as Maxent to function properly, all layers must be
set to the same spatial resolution. The resolution selected for this study was 30m resolution
33

which is the resolution of the primary datasets used to build and test the model. The Soils and
Land Use datasets are both native 30m resolution.
I considered testing at 200m and 800m to allow for the inclusion of Parameter-elevation
Regressions on Independent Slopes Model (PRISM) climate datasets. According to the PRISM
Climate group website at Oregon State University, PRISM data is provided by the PRISM
Climate Group which gathers climate data from monitoring networks and develops spatial
climate datasets to be used to show short- and long-term climate patterns. Although the use of
climate data in species distribution modeling is common (Manzoor, Griffiths, and Lukac 2018),
it was determined to be unnecessary in this study as the scale of data available was too coarse to
provide adequate detail to inform the model. Also, given that cogongrass is highly tolerant to a
wide range of climatic conditions, and the climate of the state meets this range for all pertinent
climate data in all but the extreme northeastern portion of the state, inclusion of climate data was
determined to be superfluous.
3.3. Data Description
A wide range of environmental variables are available both publicly and privately for use
in SDMs such as Maxent. To minimize potential overfitting of the model, care was taken to
select only variables that were relevant to the species and study location and to reduce
redundancy in variables where possible. For greater model relevance and to minimize correlation
between variables used, it is advised to select environmental variables that are relevant to the
species being studied (Manzoor, Griffiths, and Lukac 2018; Radosavljevic and Anderson 2014).
When employing data from multiple sources, however, ensuring proper alignment of the
data can be difficult. Maxent requires that all environmental layers used in a given model match
in geographic extent, grid cell size, and projection (Elith et al. 2011; Phillips 2017; Ervin and
34

Holly 2011). To this end, all environmental variables used within this study were sampled at a
30m by 30m grid cell size in the North American Datum (NAD) 1983 UTM Zone 16N
projection. The species point location dataset was also projected to NAD 1983 UTM Zone 16N
to match the projection of the environmental variables.
Datasets to be used within this analysis include the verified point location dataset of
cogongrass from the Alabama Forestry Commission, USGS GAP Land Cover data set, USDA
Soils data, and four roads datasets provided by Silvics Solutions LLC comprising Local, State,
US Highways, and Interstate features (see Table 5). For reference and visualization purposes
State, County, and Bing Maps base maps were also used in the study.
35

Table 5: Datasets used in this study
Dataset Source Description
Cogongrass
verified
infestation point
location dataset
Alabama Forestry
Commission
Point location dataset for all verified infestation
locations in the state of Alabama as reviewed and
verified by the Alabama Forestry Commission. The
publication of this dataset is 4/19/2018. This dataset
can be acquired through direct request from the
Alabama Forestry Commission.
Distance to
Nearest Road
Feature
Silvics Solutions
LLC
Calculated using the Euclidean Distance tool in
ArcGIS 10.6 from four road layers provided by
Silvics Solutions LLC.
USGS GAP Land
Cover Data Set
Databasin.org
(https://databasin.org/
datasets/e6c2c82715b
e44bba3579fa6010ac
fd5)
“The USGS GAP Land Cover Data Set includes
detailed vegetation and land use patterns for the
continental United States. The data set incorporates
the Ecological System classification system
developed by NatureServe to represent natural and
semi-natural land cover.” (USGS website).
Projection = NAD_1983_Albers.
USDA Soils data

United States
Department of
Agriculture Soil
Survey Geographic
Database (SSURGO)
Soils data as collected by the National Cooperative
Soil Survey. The survey is broken down into map
units (polygons) describing the soils and other
components of the soils such as productivity, and
soil horizons. The information was collected at
scales ranging from 1:12,000 to 1:63,360.
Projection = World Geodetic System 1984 in units
of decimal degrees.

All datasets were projected to NAD 83 UTM Zone 16 North and clipped to the boundary
of the state of Alabama prior to the outset of the study. Any data falling outside of that boundary
was removed from this analysis. All environmental variable (covariate) datasets are publicly
available for download with the exception of the specific road layers used for the Euclidean
Distance calculation, however similar roads datasets are available publicly. See Table 6 for the
specific source of each environmental layer used in the study. Maps showing each of these layers
are included in Appendix A and B.
36

Table 6: Environmental variables used in the study along with their layer name abbreviation and
specific source and tool used for creation where applicable
Variable Abbreviation Source
Percent Canopy Cover PctCanopy nlcd_2011_USFS_tree_canopy_2011_edition_201
6_02_08_cartographic
Ecological System EcolSys GAP Land Cover Data for Alabama, USA
(gap_30m_al)
Distance to Roads Distance Roads layer provided by Silvics Solutions, LLC.
Distance calculated using the Euclidean Distance
tool in the Spatial Analyst toolbox in ArcGIS 10.6
Soil pH pH gSSURGO Soils Data Development Tools toolbox.
gSSURGO Mapping Toolset. Create Soil Map tool
for ArcGIS
Soil Particle Size PartSize MUPOLYGON layer from gSSURGO_g_al
database
Drainage Class DC gSSURGO Soils Data Development Tools toolbox.
gSSURGO Mapping Toolset. Create Soil Map tool
for ArcGIS
Depth to Restrictive
Layer
Bed gSSURGO Soils Data Development Tools toolbox.
gSSURGO Mapping Toolset. Create Soil Map tool
for ArcGIS
Percent Clay Content PctClay gSSURGO Soils Data Development Tools toolbox.
gSSURGO Mapping Toolset. Create Soil Map tool
for ArcGIS
Percent Silt Content PctSilt gSSURGO Soils Data Development Tools toolbox.
gSSURGO Mapping Toolset. Create Soil Map tool
for ArcGIS
Percent Sand Content PctSand gSSURGO Soils Data Development Tools toolbox.
gSSURGO Mapping Toolset. Create Soil Map tool
for ArcGIS

3.3.1. Species Presence Data
The species presence data used in this study consisted of the verified point location
dataset for Cogongrass (Imperata cylindrica (L.) Beauv.) as provided by the Alabama Forestry
Commission. Specifically, this data was provided by Dana Stone, Forest Health Coordinator,
Alabama Forestry Commission, Montgomery, AL and was provided in shapefile format. Figure
11 shows the entire set of points. This is a very robust dataset including over fifty-four thousand
reported points and 34,771 field verified points.
37

Figure 11: Alabama Forestry Commission field verified cogongrass infestation locations in the
state of Alabama
38

The presence point dataset was cleansed prior to use in this study. First the dataset was
reduced to only those points that have been field verified, the dataset was then projected to NAD
1983 UTM Zone 16N and the Extract by Mask tool was used to remove all presence points that
fell outside of the State of Alabama. This tool was used extensively in the data preparation phase
of this study and therefore warrants a brief explanation of its function.
The Extract by Mask tool in the Spatial Analyst Toolset is used to extract cells from an
input raster that correspond to the area defined by a mask layer which can be vector or raster.
This tool is used to ensure that the output raster has the exact same cell count in number of
columns and rows as the mask layer. It is a requirement of Maxent that all environmental layers
have the same header values in the ASCII files used in the model run. In Maxent, if layers do not
have the same values in the header, the model will error out and cannot be run until the extent of
each ACSII file matches exactly. Since the mask feature specified was a vector layer rather than
a raster layer, the tool internally converts the vector to a raster and marks any point cell whose
cell center point falls outside of the original vector boundary as No Data (Esri 2019).
The points were then further extracted to create separate environmental data layers for the
Model Study Area and the transferability test areas. Once the final point location datasets needed
for the model were generated, new X and Y coordinate columns were added to the attribute table
and the X and Y values were calculated in meters. This table was then exported to Microsoft
Excel using the Table to Excel tool and the resultant Excel file was converted to a comma-
separated values (CSV) file format for use in Maxent. Steps for preparation of the presence point
location dataset for use in Maxent are outlined in Appendix C.
39

3.3.2. Soils Data Overview
Cogongrass’ tolerance to a wide range of environmental conditions (Terry et al. 1997;
Howard 2005) makes selection of appropriate environmental layers for analysis tricky. The
ultimate goal of the study is to provide a model that is transferable to areas across the state of
Alabama and therefore environmental variables considered for the study need to be both granular
enough to contribute usable outcomes and broad enough to be applicable across the entire state.
Descriptions of the soils related environmental variables included in this study can be found in
Table 7 and details on each one are given in the following paragraphs.
Table 7: Soil variables used in the study
Soils Variable Description
Bed Depth in centimeters to the layer that impedes water and air movement
or restricts root growth within the soil (depth to restrictive layer)
DC Drainage class is an indication of the soil’s wetness and/or saturation
PartSize Particle size is the general classification of the soil texture as determined
by grain size for the topmost horizon of soil (standards used by the U.S.
Department of Agriculture). Terms defined according to % of sand, silt
and clay.
PctClay Percent clay is the percentage by weight of soil with mineral particles
less than 0.002 mm in diameter.
PctSand Percent sand is the percentage by weight of soil with mineral particles
ranging from 0.05mm to 2mm in diameter
PctSilt Percent silt is the percentage by weight of soil with mineral that range
from 0.002mm to 0.05mm in diameter
pH Soil pH is a measure of the acidity or alkalinity of the soil using the 1:1
water method of measurement.

Depth to Restrictive Layer (Bed) quantifies the depth in centimeters to the layer that
impedes water and air movement or restricts root growth within the soil. According to the
metadata layer properties associated with the gSSURGO_CreateSoilMap.py script used to
generate this layer, the restrictive layer is a continuous layer and can be a physical, chemical, or
thermal barrier. The fire case study discussed in Section 2.3 indicated that cogongrass may
40

outcompete native species in poor soils due to its dense rhizome mat (Howard 2005) allowing
the invasive to restrict access to soil nutrients and water for native grasses (Howard 2005;
Lippincott 1997). Depth to soil restrictive layer as well as drainage class were selected as
environmental variables within this study as proxy for these considerations. Figure 12 shows
thumbnail images of this layer. The values range from greater than 201 cm (bright green) to 0 cm
(red). Note how Test Area 2 is very different from the other two areas as it has much shallower
depth to restrictive layer in much of its geographic area. For this and all subsequent thumbnail
images in this section, large images of these layers are included in Appendix A: Soils Related
Environmental Covariate Maps.

Figure 12: Depth to restrictive layer thumbnail images for each of the study areas. See Appendix
A for larger images.
Drainage Class (DC) is a representation of moisture content in the soil in its natural
condition. There are seven subclasses which range from excessively drained to very poorly
drained within the drainage class variable (www.epa.gov/enviroatlas). A study by King and
Grace showed that high water levels restricted cogongrass seedling growth and the seedlings
germination was reduced by 74% when soils were flooded (King and Grace 2000). The Model
Study Area had poorly drained soils over 26% of its total area, whereas Test Areas 1 and 2 had
41

15% and 4% of their areas consisting of poorly drained soils respectively. In contrast, the Model
Study Area had 68% of the total area covered with well drained soils, whereas Test Areas 1 and
2 had 73% and 86% of total area consisting of well drained soils respectively. Figure 13 shows
thumbnail images of this layer. Well-drained soils are shown in mossy greens and poorly drained
soils are shown in blues. Large images of these layers are included in Appendix A: Soils Related
Environmental Covariate Maps.

Figure 13: Drainage class thumbnail images for each of the three study areas. See Appendix A
for larger images
Particle size (PartSize) represents a general classification (grouping) of the soil texture as
determined by grain size for the topmost horizon of soil using the standards used by the U.S.
Department of Agriculture. This grouping places soils with somewhat similar properties in the
same particle size class and is helpful when a general view of soil texture is needed. PartSize
classification is defined according to percent of sand, silt, and clay as shown on Figure 14,
therefore there is some correlation between this variable and the individual percentages for sand,
silt and clay used within the model. Soil particle size, both in broad categorical terms as well as
percent clay, silt, and sand, were included in this analysis.
42

Figure 14: The soil texture triangle is used to convert the relative amounts of clay, silt, and sand
in the soil into texture classes. For example, a soil that is 20% clay, 30% silt, and 50% sand is a
Silty Loam. Image courtesy of Grow it Organically website, https://www.grow-it-
organically.com/facts-about-soil.html

Figure 15 shows thumbnail images of soil particle size . Particle size coloration in the
maps indicate PartSize categories as clayey soils (oranges), Silty soils (blues), Loams (greens)
and sandy soils (browns). Note that Test Area 2 has much more fine-loamy soil (light green) than
the Model Study Area or Test Area 1. Large images of these layers are included in Appendix A:
Soils Related Environmental Covariate Maps.
43

Figure 15: Particle size thumbnail images for each of the study areas. See Appendix A for larger
images.
The Percent Clay Content (PctClay) variable represents the percentage by weight of soil
with mineral particles less than 0.002 mm in diameter. The percentage and kind of clay found in
soil has significant impact on land use, drainage, fertility, etc. Figure 16 shows thumbnail images
of this layer. PctClay ranges across the study areas from 0% to 71.6% clay content where darker
color indicates higher percentages. Large images of these layers are included in Appendix A:
Soils Related Environmental Covariate Maps.

Figure 16: Percent clay content thumbnail images for each of the study areas. See Appendix A
for larger images.
44

The Percent Sand Content (PctSand) variable represents the percentage by weight of soil
with mineral particles ranging from 0.05mm to 2mm in diameter. Figure 17 shows thumbnail
images of this layer. PctSand ranges across the study areas from 0% to 94.1% sand content
where darker color indicates higher percentage. Large images of these layers are included in
Appendix A: Soils Related Environmental Covariate Maps.

Figure 17: Percent sand content thumbnail images for each of the study areas. Appendix A for
larger images

The Percent Silt Content (PctSilt) variable represents the percent of mineral soil particles
that range from 0.002mm to 0.05mm in diameter. Figure 18 shows thumbnail images of this
layer. PctSilt ranges across the study areas from 0% to 66% silt content where darker color
indicates higher percentage. Large images of these layers are included in Appendix A: Soils
Related Environmental Covariate Maps.
45

Figure 18: Percent silt content thumbnail images for each of the study areas. Appendix A for
larger images.
The soil pH (pH) variable represents a measure of the acidity or alkalinity of the soil
using the 1:1 water method of measurement. Cogongrass has been shown to grow best in
relatively acidic soils (pH of 4.7) and a study by Wilcut et al. (1988a) states that cogongrass
grew better in soils of pH 4.7 than at pH 6.7. In that study, soil pH of 6.7 was chosen to represent
typical soil pH of cultivated fields. Seed germination has also been shown to increase at pH less
than 5.0 (Sajise 1976). Although cogongrass has stronger growth rates in more acidic soils, the
species can grow in a broad range of pH values at sub-optimal growth rates.
The soil layers included in this study were selected for their relevance to the biological
and ecological niche of cogongrass. Several studies have raised the importance of soil pH, not
necessarily on the presence of cogongrass, but on the health of the species in its environment
(Ervin and Holly 2011; MacDonald 2004; Eussen and Wirjahardja 1973). Ervin and Holly
suggested that soil pH may play a larger role in cogongrass infestation in their transferability test
site in Clarke county Alabama (Ervin and Holly 2011), therefore it was determined that soil pH
would be a useful environmental variable to include in this study.
46

Figure 19 shows thumbnail images of this layer. Soil pH values range across the study
areas from 0 to 8.3 where red is the most acidic and blue is the most alkaline. The full pH scale
ranges from 0 to 10. Large images of these layers are included in Appendix A: Soils Related
Environmental Covariate Maps.

Figure 19: Soil pH thumbnail images for each of the study areas. See Appendix A for larger
images.
Although the soils dataset from gSSURGO is provided in 30m grid cell size, the original
soil mapping units were in vector format and were based on polygons with a minimum polygon
map unit size ranging from one to ten acres (Ervin and Holly 2011; Soil Survey Staff 2011).
Therefore, some reduction in granularity may occur when the Mapunit vector data was converted
to raster format using the Polygon to Raster tool in the Conversion toolbox in ArcGIS 10.6
(Ervin and Holly 2011). This tool was used to produce the Soil Particle Size raster layer. All
other soils related data layers were produced from the gSSURGO Soils Data Development Tools
toolbox, gSSURGO Mapping Toolset, Create Soil Map tool for ArcGIS. Figure 20 shows the
workflow used.
47

Figure 20: Data layer creation workflow for Soils data
48

3.3.3. Landcover Data Overview
Two data attributes from the Landcover dataset were included in the study. They include
percent canopy cover and ecological system. Percent canopy cover for this study was pulled
from the 2011 edition of the National Land Cover Dataset (NLCD) Tree Canopy cartographic
layer produced by the Multi-Resolution Land Characteristics Consortium (MRLC). The 2011
edition was chosen as it most closely matched the timeframe that the initial Cogongrass
infestation study was implemented and therefore would represent the percent canopy at the time
of that study. The NLCD data is downloadable in 30m raster format and was generated by the
United States Forest Service (USFS). Details related to the layer and its original creation can be
found on the MRLC website (MRLC accessed 04/13/2019). This layer followed the preparation
process as depicted in Figure 21.
49

Figure 21: Data preparation workflow for Percent Canopy layer.
50

The percent canopy cover (PctCanopy) dataset was chosen for this study as several
studies have suggested that canopy cover is a limiting factor in cogongrass growth as the species
is somewhat shade intolerant. Percent canopy cover had a 77% contribution in the Mississippi
portion of the Ervin and Holly study (Ervin and Holly 2011) and ability to survive as an
understory species (Gaffney 1996) and tolerance up to 50% reduction in sunlight (Patterson
1980) has been reported. Figure 22 shows thumbnail images of this layer. PctCanopy ranges
across the study areas from 0% to 100% where darker color indicates higher percentage. Large
images of these layers are included in Appendix B: Other Environmental Covariate Maps.

Figure 22: Percent canopy thumbnail images for each study area. See Appendix B for larger
images.
Cogongrass can survive in a broad range of environmental ecological habitats as
discussed in Chapter 1 and Chapter 2. Studies have also shown that cogongrass is particularly
destructive in agricultural systems where the species can compete directly with agricultural crops
thus creating not only an ecological impact but an economic one as well (Ervin and Holly 2011;
Akobundu and Ekeleme 2000; Terry et al. 1997; Hubbard et al. 1944). To test the importance of
ecological system on cogongrass infestation, the ecological system layer was generated from the
GAP Land Cover Data for Alabama, USA (gap_30m_al). The method for creating the raster
51

layer used in Maxent followed the same process as that used for the percent canopy layer
described in Figure 21. Figure 23 shows thumbnail images of this layer. Ecological system is
depcted in these images in grouped categories of Frest/Woodlands (green), Floodplain Forest
(blue-green), Agriculture (tan), Developed (dark orange), Disturbed (brown), water (blue) and
undefined/other (grey). Note the larger proportion of Floodplain Forest in the Model Study Area
and the larger propotion of Developed land in Test Area 2. Large images of these layers are
included in Appendix B: Other Environmental Covariate Maps.

Figure 23: Ecological System maps for each study area. See Appendix B for larger images.
3.3.4. Roads Data Overview
Cogongrass spread occurs via two mechanisms, rhizome growth for local spread, and
seed dispersal for long range spread as described in Chapter 2. Studies have indicated that spread
along roads occurs due to wind dispersal as well as seed dispersal via hitch hiking on road
maintenance equipment (Rauschert, Mortensen and Bloser 2017; Wilcut et al. 1988a; Wilcut et
al. 1988b; Willard 1990). Although studies have quoted wind as the primary long-distance
dispersal method (Yager, Miller, and Jones 2011), Willard suggests in his 1990 study that long
range spread in Florida was primarily due to rhizome pieces being transported in fill dirt (Willard
et al. 1990). In either case, roads play a part in infestation spread. To test for the impact of roads
52

on Cogongrass, roads datasets were procured and the distance from each grid cell within a
distance raster to the nearest road feature was calculated using the Euclidean distance tool in the
Spatial Analyst toolset in ArcGIS 10.6. Roads data was provided for use in this study by Silvics
Solutions, LLC in the form of four distinct road vector polyline layers. These layers were
provided in North American Albers Equal Area Conic projection and were projected to NAD
1983 UTM Zone 16N using the Project tool in ArcGIS 10.6. The original roads layers provided
consisted of a Local Roads layer, which contained both city and county roads, a State Highway
layer, a US Highway layer, and an Interstate layer. Some feature overlap occurred between
layers as some road features are captured in more than one dataset. This was mitigated when all
road layer features were combined into a single layer and duplicates were removed.
The Euclidean Distance tool creates a raster dataset where each cell within the layer
contains a value equal to the distance from the cell center to the nearest road feature. The use of
the Near tool in the Proximity toolset was also investigated, but it was determined that the
Euclidean Distance tool provided an output that best meet the needs of the Maxent model. The
Near tool was utilized in data review, however. The input data layer for the Euclidean Distance
tool was set to the consolidated roads layer. All four of the original roads datasets were
combined into one consolidated road data layer in order to run the Euclidean distance tool on a
raster layer depicting all roads at once. Figure 24 shows thumbnail images of this layer. The
distance from nearest road ranges from 0 to 5359 meters across the three study areas. Note that
in Test Area 2 there are substantially more local road features and therefore fewer cogongrass
presence points that fall at great distance from roads. Large images of these layers are included
in Appendix B: Other Environmental Covariate Maps
53

Figure 24: Distance to nearest road data maps for each study area. See Appendix B for larger
images.

There are two important notes regarding the Euclidean distance raster. First, road width
was not considered when this road layer was created as it was created from vector polyline layers
that was then transformed into a 30m raster layer using the Polygon to Raster tool in the
Conversion toolbox in ArcGIS 10.6. As previously mentioned, according to Ervin and Holly, the
Polygon to Raster tool can result in some reduction in granularity (Ervin and Holly 2011). The
output distance is to the center point of the road not the road edge. Second, when species
presence points of larger distances from roads were visually investigated (using the Near tool and
measure tool), many of these cells were within closer proximity to unmapped roads, such as
interior woods roads, than is indicated in the Euclidean Distance raster layer which depends on
mapped features being present in the dataset. Therefore, it may be a worthwhile endeavor to
recreate this dataset in a later study with more granular roads data. This road dataset cleaning and
augmentation is out of scope for this project.
54

3.4. Methods
Modeling methods for this study were loosely guided by the Ervin and Holly (2011)
study in which the authors tested the transferability of a Maxent model developed for cogongrass
location point data collected in the De Soto National Forest (NF) and Sandhill Crane National
Wildlife Refuge (NWR) areas in southeastern Mississippi to a site consisting primarily of
commercially managed pine timberlands in Clarke County, AL. This study piqued my interest in
transferability of Maxent models and prompted a more Alabama centric study of transferability.
3.4.1. Defining the Model
Detailed review of biologic and ecological requirements of cogongrass was conducted to
determine what environmental layers should be considered for this study. Ervin and Holly’s
(2011) study included soils related variables containing available water capacity, effective cation
exchange capacity, percent organic matter, pH, and percent Silt content; and Land Cover related
variables including percent canopy cover, and percent by ecological system (agriculture,
coniferous forest, deciduous forest, developed, harvested forest, managed forest, and other). In
the current study, available water capacity (AWC), effective cation exchange capacity (ECEC),
and percent organic matter (PctOM) were not used.
AWC was discarded for two reasons, the first being insight gained from the 2000 study
by King and Grace examined soil moisture content’s effect on cogongrass seedling germination
and growth and second, because in initial test run iterations AWC added little to no gain when
reviewing the Jackknife results of preliminary default Maxent runs.
ECEC was not included as it was a surrogate for total soil nutrient content and
availability in the Ervin and Holly (2011) study. For the current study, it was decided that this
55

was too broad of a valuation metric and other studies reviewed have sited cogongrass’ ability to
tolerate a broad range of soil nutrient levels.
Percent Organic Matter was not used in the current study as the heuristic review of soil
organic matter data within ArcGIS 10.6 revealed little difference in percent organic matter from
0 to 200cm of soil depth across the entire state of Alabama. This layer also added little to no gain
when reviewing the Jackknife results of preliminary default Maxent runs. The other
environmental covariates used in the Ervin and Holly (2011) study were included in the current
study as well as the addition of percent clay, percent sand, particle size, and distance to nearest
road.
The Maxent model was first run with the default settings in place as a baseline of model
fitness for use and to assist in the determination of which parameters would need to be tuned in
order to fit the model for the species and study location. Five replicates were run for the Model
Study Area in Maxent for the tuned model followed by five replicates each run against the test
areas with species and environmental layers masked to the Model Study Area using the
Projection layers directory/file setting.
3.4.2. Tuning the Model
Although Maxent default settings were determined by Phillips based on testing across a
wide range of species and environmental factors, it is suggested that models be tuned for the
specific species and location being modeled (Elith et al. 2011; Phillips 2017). Model tuning was
performed to maximize performance while minimizing the potential for overfitting. The Model
Study Area Maxent model was built using the biological environmental variables relevant to
AFC Work Unit 11 and Cogongrass in general (Table 6). Care was taken to select environmental
variables that both represented specific measures relevant to the biology and habitat preferences
56

of the species and were broad enough to be useful measures across the landscape. This same
thought process was given to tuning the model.
Several Maxent default settings were maintained in this study. The regularization
multiplier was left at 1 as was the case in the Ervin and Holly study. This value is a modifier to
help smooth the model in an attempt to avoid over-fitting and underfitting and helps to balance
fit and complexity within the model (Ethel et al 2010). Modifying this value was tested with
settings of 0.5, 0.8, and 2 (Table 8) with limited improvement when the regularization multiplier
was reduced and limited reduction in fitness when the regularization multiplier was doubled.
Table 8: Regularization Multiplier's effect on AUC. All other settings remaining equal.
Regularization Multiplier AUC (Training/Test)
0.5 0.712/0.717
0.8 0.709/0.715
1 0.708/0.715
2 0.699/0.705

The AUC calculation on one replicate run was used as the indicator of fitness while
running tuning tests. The decision to leave the default setting for these values was made as the
change in AUC due to modification of regularization multiplier alone did not significantly
change the modeled results. The number of background points, maximum number of iterations
per replicate run, convergence threshold, and default prevalence were all also left at their default
settings.
Parameters that were tuned in the model included values that resulted in modification of
the model itself and values that resulted in modification of the output from the model. Parameters
that resulted in modification of the model itself and were tuned in the Model Study Area Maxent
model were changing the output format to Logistic, modifying the replicate run type, setting the
57

number of replicates, selecting to add samples to the background, and selecting to use samples
with some missing data. Parameters that resulted in modification to the output files of the model
but not the model itself included: selecting to create response curves and run jackknife tests,
increasing the number of processor threads used by the model, selecting to write plot data,
selecting to add summary results to the Maxentresults.csv file, and selecting to write background
predictions. Appendix D: Maxent Model Settings Screen Captures shows how all of these were
set within Maxent. Some of these modifications are discussed below.
The Logistic output was selected rather than the newer default of Coglog as Logistic
output was the default in previous versions of Maxent and was the output format selected by
Ervin and Holly. Logistic output is also recommended in Phillips and Dudik 2007. Modifying
this setting increased the AUC of the resultant model marginally (0.698 to 0.708) but this minor
difference is most likely due to the nature of the random seed setting. Increasing the number of
processor threads allowed the model to use more of the computer’s processing capabilities thus
allowing some intensive processes such as jackknife creation to run faster. Checking the setting
to write background predictions was required in order to calculate TSS for each model replicate
run. Modifying the replicate run type involved setting the replicate run type to sub sample along
with setting the random test percentage to 50% and checking the random seed checkbox. These
three settings in conjunction provided a slightly better model AUC than using the default Cross
validate replicate run type (AUC improved from 0.708 to 0.725). The final model selected to use
in the study was the model with AUC of 0.725.
3.4.3. Gauging Fitness of the Model
The purpose of evaluating a model is to determine if it is useful, or fit, for the purpose the
model is being used for (O’Sullivan and Perry 2013). The Maxent model uses parameters and
58

constraints to modify the model output. Many studies using Maxent modify a few key
parameters but leave most parameters set to their default values. Depending on the study subject,
this may be an appropriate course of action. An analysis of the parameters and settings needed to
produce an appropriate model was performed in this analysis to ensure that model parameters
were appropriate to the study.
The fitness of the model in predicting the distribution of cogongrass was evaluated using
AUC, Sensitivity (Omission) and TSS. The relative contribution of each environmental variable
to the model was evaluated by review of the jackknife output tables as well as the plot graphs of
each individual environmental variable. In Section 3.4.1, analysis of these metrics allowed for
the removal of datasets that proved to be of little value to the study.
The Maxent model was set to five sub-sample replicates, withholding a randomly
selected 50% of the test data in each iteration. Setting the model to 10 replicates was also tested
however the statistics were not significantly different between the five replicate and 10 replicate
tests. The model results reported in Chapter 4 represent the resultant predicted potential
distribution of cogongrass given suitable conditions, averaged across five model replicates as
well as an averaged standard deviation.
3.4.4. Testing Transferability of the Model
To evaluate the effectiveness of the trained model, the predicted distributions produced
with new environmental data for the test areas were compared to additional verified presence
points of cogongrass locations in the test areas (test data). These test data were used to determine
the accuracy of the model’s predicted distribution, and therefore the viability of the model in
predicting distribution for cogongrass infestations when transferred to different geographic space
than where the model was originally trained. A model that was trained on a set of environmental
59

variables in one geographic space can be transferred by running the same model using the same
set of environmental variables that have been extracted to the new study area location (Phillips et
al. 2017).
When the model is transferred to a new geographic area, the Projection layers
directory/field in the Maxent user interface can be set to point the new model run (in this case
model runs for Test Area 1 and Test Area 2) to the original environmental layers (in this case the
Model Study Area’s environmental layers) so that the environmental layers used in the new
model runs are “clamped” to the original model layer ranges. Clamping essentially sets any layer
value in the new model run’s environmental layers that fall outside of the range of values for that
layer in the original model run to equal the outer bound of the original environmental layer. For
example, the range of percent silt in the PctSilt environmental layer for the Model Study Area
was 0 to 60%, for Test Area 1 the range was 0 to 56.1% and for Test Area 2 the range was 0 to
66%. Therefore, the new model run for Test Area 1 did not require clamping to the extent of the
range from the Model Study Area for the environmental variable, but PctSilt did require
clamping for the new model run for Test Area 2. The response to this variable in the new model
run for Test Area 2 is held constant for all values that fall outside of the training range (the range
of values found in this layer for the Model Study Area model run) essentially treating those
values as if they were at the limit of the range (in this case, 60%). According to Phillips in the
updated (2017) tutorial on Maxent, testing transferability by projecting the model in this manner
is appropriate when the goal is to evaluate a model at a set of test locations (Phillips 2017) which
is the goal of this study.
In this analysis, transferability of the model was tested by utilizing regional subsets of the
same environmental layers and maintaining the same parameters in both of the test areas. During
60

the initial environmental data layer creation process, the AFC Work Unit polygons were used as
boundary extent to split the original data layers into 18 separate raster layers. By using the Split
Raster tool in ArcGIS 10.6, transferability testing was substantially sped up, as data layer
manipulation requirements were lessened.
Maxent was run against Test Area 1 and Test Area 2 utilizing the same setting parameters
as were defined in the Model Study Area’s Maxent model. For each of the test areas, the
presence points .csv file was created by extracting only those presence point feature that fell
within the boundary of the selected AFC Work Unit. This file was then processed and converted
for use in Maxent as defined in Section 3.3. All environmental layers utilized in the study were
also extracted to the extent of both test areas as separate files following the same processes as
defined Section 3.3. The environmental layers folder was set to the folder housing the .asc files
for the test area included in the transferability test model and the output directory was set to the
test area folder’s output directory. The projection layers directory/file was set to the directory
that housed the environmental layers used in the original Model Study Area model run.
It is important to note that all layers, in each of the environmental layers directories
should use a common naming convention so that the Maxent model can determine appropriate
layer clamping for the Test Area model runs. For example, the drainage class layer is named
“DC” in all three directories (Modeled Study Area, Test 1, and Test 2). The model is then trained
using the environmental layers set in the environmental layers list for the Model Study Area
model and then transferred onto the Test Area environmental layers which clamps the
environmental variables for the current run to the bounds of the original model run to which the
layers are being transferred (Phillips 2017).
61

TSS is a special case of Kappa that reduces issues associated with prevalence that
prevents Kappa from being a useful metric for presence only data. TSS was used as a measure of
model fitness in this study. The formula for calculating TSS is shown in Equation 1.
TSS = Sensitivity + Specificity -1 (1)
A single TSS score for each model was determined by calculating the TSS for each of the
five replicates in a model run individually and selecting the run with the highest TSS score to be
the representative score for that test. TSS can be calculated from the Maxent output by selecting
the “write background predictions” selection on the Experimental tabs in Settings. This setting
tells Maxent to write background predictions files for each of the replicate runs.
Next, copy the “logistic” column from the background predictions file of replicate 0 and
paste it into column A of a spreadsheet (tab labeled 0). Then, open the sample predictions file for
replicate 0 and copy the “logistic prediction” column into column B of your spreadsheet. Step
three is to choose which threshold you want to use in the calculation. In this study I have chosen
to use the 10 percentile training presence logistic threshold. This value can be found in the
“Maxent Results” file as output by Maxent. For step 4, do a count of sample predictions test
values greater than the threshold for replicate 0, a count of sample predictions test values less
than the threshold of replicate 0, a count of background prediction test values greater than the
threshold for replicate 0, and a count of background prediction test values less than the threshold
for replicate 0. With these values in hand TSS for replicate 0 can be calculated.
Calculate Sensitivity as the count of cells where the sample predictions test values are
greater than the threshold, divided by the total count of sample prediction test values. Then,
calculate Specificity as the count of cells where the background predictions value is greater than
the threshold, divided by the total count of background prediction values. As defined above, the
62

TSS for replicate one is Sensitivity plus specificity minus one. This calculation is repeated for
each replicate in the Maxent run and the largest TSS from the replicate set was then used as the
TSS score for the model in this analysis.

63

Chapter 4 Results
Studies exploring the application of Maxent have indicated that there is no perfect metric to
evaluate all models for fitness to the study. Species such as cogongrass, which can tolerate a
broad range of ecological and environmental conditions, can produce model results with a large
area of predicted occurrence (Ervin and Holly 2011). It is suggested that each evaluation metric
be assessed in context with the specific species and variables in use and the desired use of the
model output. It is also suggested that a mix of evaluation metrics be used to determine model
suitability and fitness (Anderson 2012; Merrow, Smith, and Silander 2013; Radosavljevic and
Anderson 2014; Peterson et al. 2011). To this end, AUC, Sensitivity (Omission Rate), and TSS
were selected as measures of model suitability.
The Model Study Area Maxent model used in this analysis was evaluated using a 5-fold
sub-sample with 50% of the presence points set aside randomly for testing the model. It was
appropriate to use 50% of the presence points for testing due to the large number of presence
points included in the species presence data layer. The Model Study Area Maxent model results
were compared to the results from the Test Area 1 Maxent model and Test Area 2 Maxent model
transferability tests and the model suitability results for these three models (Model Study Area
Maxent model, Test Area 1 Maxent model, and Test Area 2 Maxent model) utilized in the study
are shown in Table 9and are discussed in greater detail below.
Table 9: Model Suitability Indicator Results
Indicator Model Study Area Test Area 1 Test Area 2
AUC (5 fold sub-sample) 0.7250 0.7460 0.8460
AUC std dev 0.0010 0.0020 0.0170
Test Omission 0.0832 0.0807 0.2941
TSS (highest of replicates) 0.4087 0.3944 0.2377

64

4.1. Area Under the Receiver Operating Characteristic Curve (AUC)
The area under the receiver operating characteristic curve (AUC) indicates fitness of the
model (Phillips et al. 2017). As discussed in Chapter 2, the AUC shows the average sensitivity
vs. specificity for the species being modeled and tells us how well the model can discriminate
between presence locations and background data. According to Elith et al. (2011), an AUC of
0.70 and above indicates sufficient fit for ecological niche study purposes. Therefore, the AUC
of 0.725 returned for the Model Study Area, along with its low standard deviation (0.0010),
indicates a stable model and a good fit for predicting the distribution of cogongrass within the
study area. Both of the transferability test areas returned AUC above 0.70, however the much
larger standard deviation in Test Area 2 warrants some concern.
The mean standard deviation for the AUC in the Model Study Area is 0.0010, which is
very low and a good indication of model stability. The AUC of Test Area 1 (0.746) is slightly
higher than that of the Model Study Area (0.725) with similarly low standard deviation (0.0020).
This is an indication of good transferability of the model to Test Area 1. The AUC of Test Area 2
(0.846) is even higher than that of Test Area 1, however the standard deviation is higher at 0.017.
Although this standard deviation is still within a valid range (95% of the replicate runs fall within
one standard deviation) it is much higher than the standard deviation of the AUC for the Model
Study Area and that of Test Area 1. Therefore, additional review of the results of the model in
Test Area 2 is required.
4.2. Sensitivity (Omission)
The omission rate for the model is depicted in the .html output file produced by Maxent.
The omission rates for test samples from the model run on each of the three study areas are
provided in Table 9. The omission rate shows model performance as a function of the predicted
65

occurrence. For the Model Study Area and Test Area 1, the modeled omission rate follows the
predicted omission closely with very low standard deviation. The Model Study Area omission
rate falls at 0.0832 and Test Area 1’s omission rate similarly falls at 0.0807. This shows a very
good match of the test data to the trained model predictions and is an indicator of a well fitted
model. For Test Area 2 the Omission rate was significantly higher (0.2941) again, warranting a
closer look.
4.3. Variable Contributions and Gain
An understanding of how the environmental variables selected for use within the model
effect the model outcome is important in understanding the statistics used to test model fitness.
Maxent produces very detailed output in the form of multipage html documents. In these
documents, tables showing the percent contribution and permutation importance of each
individual environmental variable included in the study assists in understanding the model
results. Also, in these documents, response curves provide a visual representation of the
predicted potential distribution of species occurrence in two graphs per variable.
The percent contribution indicates how much the individual variable contributes to the fit
of the model (gain). This value should be used with caution when variables are highly correlated
(Phillips et al. 2007; Phillips 2017) which is a potential issue with the particle size dataset used in
this analysis. The permutation importance shows the contribution of each variable via random
permutation (and does not rely on the path the model used to get to the final result) thus a larger
permutation importance indicates that the model depends heavily on the variable. For this
analysis we focus on the permutation importance of variables, as this value lessens the impact of
variable correlation. The percent contribution (Table 10) and permutation importance (Table 11)
for the Modeled Study Area as well as Test Area 1 and Test Area 2 are provided below.
66

Table 10: Percent contribution of environmental variables for the Model Study Area, Test Area
1, and Test Area 2.
Environmental
Variable
Model Study
Area
Test Area 1 Test Area 2
PctCanopy 61.1 38.7 20.1
EcolSys 11 7.7 9.4
Distance 10 2.1 56.4
pH 8.2 0.3 0.2
PartSize 3.5 19.3 1.1
DC 3.3 17.7 1.2
Bed 1.1 0.6 7.1
PctClay 0.7 3.2 1.3
PctSand 0.5 5.5 2.9
PctSilt 0.6 5 0.3

Table 11: Permutation importance of environmental variables for the Model Study Area, Test
Area 1, and Test Area 2.
Environmental
Variable
Model Study
Area
Test Area 1 Test Area 2
PctCanopy 56.9 30 36.6
EcolSys 14.7 8.8 4.1
Distance 8 2.4 43.5
pH 8.2 1.1 0.1
PartSize 4.4 16.3 2.3
DC 3 1.3 2.5
Bed 1.1 0.9 7.5
PctClay 0.7 13.9 1.5
PctSand 2.2 8.6 1.2
PctSilt 0.8 16.8 0.7

Review of individual environmental variable’s response curves in conjunction with data
on percent contribution and permutation importance provides valuable information into the
impacts that each variable has on the model outcome. The response curves for each
environmental variable included in the model (Appendix E: Response Curves) indicate how each
variable affects the predicted probability of presence when all other variables are set to their
average value. This means that each curve shows the marginal impact of change to the predicted
67

potential distribution of cogongrass infestation given suitable conditions resulting from changing
just the one variable selected. Each curve shows the mean response of the 5-fold sub-set model
in red and +/- one standard deviation in blue. Below the four most significant variable’s response
curves for each model are explored.
In keeping with previous studies (Ervin and Holly 2011), percent canopy cover
(PctCanopy) had the highest permutation importance for both the Model Study Area (56.9%) and
Test Area 1 (30%) and fell to second highest for Test Area 2 (36.6%) (Figure 25). This is as
expected as cogongrass is somewhat intolerant to shade, and previous studies have shown shade
to be an important factor when predicting the distribution of the species (Ervin and Holly 2011;
Gaffney 1996; Patterson 1980). This is also consistent with the Ervin and Holly (2011) study in
which PctCanopy had a relative contribution of 77% (recall that their study focused on heavily
forested ecosystems in Mississippi). As the current study hypothesized, it was expected that
PctCanopy cover would have significant impact on predicted site suitability.

It is relevant to observe that in Test Area 2, where there is significantly more developed
and open (agricultural) land, the permutation importance of PctCanopy was lower than in the
other two model areas. As can be seen in the PctCanopy graphs above for both the Model Study
Area and Test Area 1, the impact of PctCanopy on the model remains relatively high and
Figure 25: Response Curves for percent canopy.
68

consistent with low standard deviation (thin blue area around the red response curve). The
impact of PctCanopy for Test Area 2 was much more variable across the replicate runs as
indicated by the thick blue area around the red response curve for the Test Area 2 graph. The
response curves for PctCanopy for all three models indicate a high impact of this variable on
each model as the curves all increase exponentially at the beginning of the range and decrease
just as dramatically at the end of the range.
Ecological System (Eco) was the second highest permutation importance in the Modeled
Study Area (14.7%). For Test Area 1, Eco was not amongst the top four in permutation
importance (8.8%) and for Test Area 2, eco was the fourth highest in permutation importance
(4.1%). This potentially shows some departure in consistency and thus transferability where
ecological system is concerned. Ecological system is a categorical data set and the importance of
each individual category plus or minus one standard deviation to the averaged marginal response
of the model to changing one variable is shown in Figure 26. In the graphs in Figure 26, the
missing columns represent ecological systems that have no impact on the potential distribution in
the model indicated. See Appendix F for definition of the categorical values for ecological
system for each model. Most ecological systems have average impact on the model (hovering
around 0.5 for the Model Study Area and Test Area 1 where over half of the systems have less
impact in Test Area 2 and the remainder’s impact is more volatile.
69

At first review of the Maxent output response curves for eco, it is unclear which
ecological systems have impact and which do not. In this instance it would be prudent to group
the data results from the three models into one graph to better review the response of cogongrass
to ecological system across the Model Study Area and the two transferability test areas.
Figure 27 provides a graph of cogongrass’ response to ecological system in each model with
consistent numbering for each ecological system present.

Figure 27: Consolidated graph of cogongrass' response to ecological system for each of the three
models.
This consolidated graph shows that not all ecological systems are present in all study
areas included in this test. Two such groups of ecological system are identified in Figure 27.
Note that ecological systems 11 through 23 (East Gulf Coastal Plain ecological systems) do not
have any orange bars associated with them indicating that these ecological systems are not
present in Test Area 2. Also note that ecological systems 39 through 49 (East Southern
Piedmont, East Southern Interior, and East Southern Ridge and Valley ecological systems) do
Figure 26: Response Curves for Ecological System
70

not have any blue or green bars associated with them indicating that these ecological systems are
not present in the Model Study Area or Test Area 1. These results suggest that the attribute
granularity of this layer may be too detailed for this study. Table 12 offers a potential grouping
of the ecological system into broader categories that are easier to consume. It is recommended in
any future studies utilizing this variable, that this data be grouped as indicated in Table 12 and
the models re-run to better gauge impacts of this variable on the models.
Table 12: Percent geographic area (km
2
) occupied by each grouped ecological system within the
Model Study Area, Test Area 1, and Test Area 2.
Ecological System
Group
Model Study Area
% of area
Test Area 1
% of area
Test Area 2
% of area
Forest/Woodlands 49.99 46.88 52.34
Floodplain Forest 26.54 22.24 0.14
Agriculture 8.07 14.64 20.84
Developed 3.18 3.09 15.05
Disturbed 10.99 12.39 8.19
Water 1.18 0.68 2.80
Other 0.06 0.07 0.64

For the Modeled Study Area, pH was the third highest in permutation importance (8.2%)
but did not rank in the top four for either of the two test areas. Cogongrass is tolerant of soils
with a range of pH values but has been shown to grow best in relatively acidic soils (pH of 4.7)
(Wilcut et al. 1888a). The graphs in Figure 28 show the impact that the pH covariate has on the
predicted probability of presence given that all other variables are kept at their average value.
The Model Study Area graph for pH shows a gradual increase in impact as pH increases from 0
to 8.3 and the impact decreases at pH values higher than 8.3. Test Area 1 exhibits a different
pattern in pH’s impact on the model. In Figure 28 pH graph for Test Area 1 shows that impact
remains high but relatively static until it reaches 5.0 then the impact due to pH decreases slightly
and the related standard deviation of impact increases as the pH approaches 8.0. For Test Area 2,
71

the pH runs a similar curve to the Model Study Area but with much greater variability to impact
between model run replicates.

Distance to road was the fourth highest in permutation importance (8.0%) for the Model
Study Area but did not rank in the top four in Test Area 1 and was the highest in permutation
importance for Test Area 2. As would be expected given the spikelet wind dispersed travel
distances discussed in the Yager, Miller, and Jones (2011) study briefly described in Section 2.3
and the physical seed dispersal distances as shown in the Rauschert, Mortensen and Bloser
(2017) study, Figure 29 shows that distance to road has its greatest impact in close proximity to
roads with stable to lessening impact as distance increases for the Model Study Area.

Figure 29: Response curves for distance to nearest road.
For Test Areas 1, impact to the model due to the presence of the distance to roads
variable with all other variables remaining at their average rate followed a similar curve through
Figure 28: Response curves for pH
72

roughly 2000 meters then increased in impact as distance increased. This could potentially be
due to a lack of interior woods roads in the roads layer dataset as noted via visual inspection of
the data in ArcGIS 10.6. For Test Area 2, the impact of distance to road was high at close
proximity then dropped exponentially and recovered only minimally on average over the course
of the range of distances for the layer. Test Area 2 also shows significant standard deviation
(greater than 1) as the curve approaches its maximum distance values. As discussed in Section
3.3.4, the road dataset could be improved with added local roads information in non-urban areas.
Test Area 2 contains significantly more local roads due to its higher percentage of developed
land compared to the Model Study Area and Test Area 1. It is recommended that a spatial data
creation project be launched to augment the roads dataset should this variable be considered for
future study.
For Test Area 1, percent silt (16.8%), particle size (16.3%) and percent clay (13.9%)
were the second, third, and fourth highest in permutation importance respectively in that model.
None of these variables ranked in the top four in the Model Study Area. These three variables are
correlated as the particle size categorical values are defined based on the soil texture percentages
defined by the percent of silt, clay, and sand. It was determined that including both the
categorical particle size variable and the component specific percentages was important to the
analysis. Percent silt provided a percent contribution of 10% to the Ervin and Holly study.
For Test Area 2, depth to soil restrictive layer ranked third in permutation importance
(7.2%) although this variable did not rank in the top four for either the Modeled Study Area or
Test Area 1. As Test Area 2 was selected to test the transferability of the model because of its
many dissimilarities to the Model Study Area, it was expected that this location would differ in
73

variables of importance. For a more detailed discussion of how the test sites were selected, refer
back to Section 3.1.
Finally, the jackknife tests run by Maxent are depicted in three charts in the output .html.
For this analysis, we focus on the Jackknife of test gain charts to judge variable impact on the
model as test data is used to judge model performance. The Jackknife of regularized training gain
and Jackknife of AUC could also be used to test model performance. Since we are focused on
testing the model and the model’s transferability to new geographic regions within the state, it
was determined that the best test would be to utilize the Jackknife of test gain. “Gain is closely
related to deviance, a measure of goodness of fit used in generalized additive and generalized
linear models” (Phillips 2017, 4). Data in these jackknife charts are normalized, so all study area
jackknife charts can be compared.
Figure 30 shows the Model Study Area’s jackknife chart. Two points of interest are
highlighted here. As was shown in the variable contributions table (Table 10), PctCanopy had the
highest percent contribution and highest permutation importance to the model in the Model
Study Area. The jackknife analysis seconds this conclusion. In the Jackknife of test gain, the
PctCanopy row shows that PctCanopy had the most significant information by itself about the
suitability of the environment for the species (blue bar) and had the most significant total impact
in the form of reduction in gain, when omitted from the analysis (red bar). Therefore, PctCanopy
provides the most independently important information that cannot be explained by use of other
variables included in the model.

74

Figure 30: Jackknife of impact to gain by variables included within the model run for the Model
Study Area.

For the test areas, the jackknife tests provide key information about the differences in
variable impact on gain between the Model Study Area and the transferability test sites. Test
Area 1, which was the most similar to the Model Study Area, also indicates, through the
jackknife test, that PctCanopy has the most independently important information that cannot be
explained by use of other variables included in the model just as was the case with the Model
Study Area. It is important to note that the total percent contribution of PctCanopy droped for
Test Area 1 which is in large part due to the increase in importance of other variables in the
model outcome as seen in the jackknife results (Figure 31).
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Part Size
Pct Canopy
Pct Clay
Pct Sand
Pct Silt
Depth
Drainage Class
Ecological System
Distance
pH
Jackknife Results for Model Study Area
Gain without variable Gain with Variable
75

Figure 31: Jackknife of impact to gain by variables included within the model run for the
transferability Test Area 1.
The percentages of the individual soil textures (clay, silt, and sand) each independently
provide important information to the maxent model over the Modeled Study Area. The
importance of ecological system dropped significantly between the Model Study Area model and
the two test area models. Given the rural nature of Test Area 1 and the comparatively rural nature
of the Model Study Area, it was expected that these two study areas would have similar variables
of importance but from the results of the jackknife analysis, it is clear that soil texture plays a
larger role in site suitability in Test Area 1 than it did in the Modeled Study Area with PctSilt
providing 16.8% of permutation importance, PctClay providing 13.9% and PctSand providing
8.6% in Test Area 1. That being said, the inclusion of these layers in the original model was
advantageous as this allowed for the transferred model in Test Area 1 to utilize these important
factors. As discussed in Chapter 2, it is important for environmental variables selected for use in
Maxent to be broad enough in biological or ecological extent yet specific enough to be valuable
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Part Size
Pct Canopy
Pct Clay
Pct Sand
Pct Silt
Depth
Drainage Class
Ecological System
Distance
pH
Jackknife Results for Test Area 1
Series2 Series1
76

across the entire intended geographic region, when transferability of the model is a desired
outcome. The AUC for Test Area 1 (0.746) shows that the model is fit for use in testing the
probability of occurrence of cogongrass in this area.
The jackknife chart for Test Area 2 is quite interesting (Figure 32). It was expected that
the model environmental layers for the Model Study Area would differ significantly in results
from Test Area 2 as these two areas are quite different in many aspects. It was also hypothesized
that the model would not be as good a fit for Test area 2 as it was for the Model Study Area.
However, the AUC for Test Area 2 (0.846) was significantly higher than that of the Model Study
Area (0.725). Looking at AUC to gauge model fitness would inappropriately lead the modeler
astray in the assumption that the model is well suited in predicting probability of occurrence in
the Test area in this case. It is, as previous studies have suggested, important to use several
indices of fitness and to evaluate the model results thoroughly (MacDonald, 2004; Lobo et al.
2008).

Figure 32: Jackknife of impact to gain by variables included within the model run for Test Area
2.
-0.2 0 0.2 0.4 0.6 0.8 1
Part Size
Pct Canopy
Pct Clay
Pct Sand
Pct Silt
Depth
Drainage Class
Ecological System
Distance
pH
Jackknife Results for Test Area 2
Series2 Series1
77

The Jackknife test for Test Area 2 shows that all but three environmental variables
contribute a negative gain when determining the unique information contributed by that layer to
the model (blue bar). This can be an indicator of highly correlated data layers but not all of the
layers with negative test gain would be correlated. This draws into question the validity of use of
this model for Test Area 2 even though the AUC for Test Area 2 was high. PctCanopy
contributed useful unique information to the model (blue bar) and has the largest impact, after
distance to nearest road, to a reduction in gain when excluded (red bar). Distance to nearest road
had the bigest impact on the model as seen in both the jackknife test and the table of variable
contributions. The jackknife test shows significant reduction in gain when this variable is
excluded from the model. Test Area 2 has significantly more developed area and significantly
more mapped road features than the other two areas included in this study. This likely
contributes to the significance of road nearness to the resultant model
The three variables that contributed the most significant increase in gain in Test Area 2
when viewed in isolation acording to the jackknife of test gain, were distance to nearest road
feature (Distance), ecological system (EcolSys) and percent canopy (PctCanopy) in that order
acording to the jackknife. These are the three highest in permutation importance and percent
contribution for Test Area 2 as well.
4.4. True Skill Statistic (TSS)
The True Skill Statistic (TSS) is a form of Kappa that is not affected by prevalence or the
size of the validation set (Allouche, Tsoar, and Kadmon 2006). Allouche suggests that TSS
should be used over Kappa when a threshold-dependent measure is desired (Allouche, Tsoar, and
Kadmon 2006). TSS values range from -1 to +1 where values of 0 or less are no better than
random and the value of +1 is optimal (Allouche, Tsoar, and Kadmon 2006). TSS, as an
78

indicator of model fitness, was calculated for each replicate run within each study area’s model
results. The highest TSS value for each model was used as the score for that model in evaluating
model fitness using TSS (Table 13). Since TSS is a special case of Kappa and spans the same
value range, TSS can be gauged by the same degree of agreement assessment as would Kappa. A
value of +1 is perfect agreement, values of 0.75 to 1 represent excellent agreement, 0.4 to 0.75
indicate fair to good agreement and values less than 0.4 are an indication of poor agreement
(Monserud and Leemans 1992).
In this study, TSS values for all models were fairly low. TSS for the Model Study Area is
considered “Fair” at 0.4087 where TSS for Test Area 1 is just shy of “Fair” at 0.3944. the TSS
score for Test Area 2 was “Poor” at 0.2377. Given that the AUC values for the Model Study
Area and Test Area 1 were adequate but not stellar (0.725 and 0.746 respectively), a “Fair” TSS
value would be expected. For Test Area 2, the TSS is very low given the relatively good AUC
(0.846) however when taking the standard deviation of AUC and very low logistic threshold
(0.1462) returned from the Test Area 2 model run into account, in this instance TSS helps to
confirm that the transfer of the model to Test Area 2 is questionable.
Table 13: TSS for each replicate run for the Model Study Area, Test Area 1 and Test Area 2. The
highest TSS of the replicates for each model area was used.
Model Replicate 0 Replicate 1 Replicate 2 Replicate 3 Replicate 4
Model Study Area 0.4041 0.4087 0.4064 0.4079 0.4035
Test Area 1 0.3889 0.3938 0.3944 0.3925 0.3918
Test Area 2 0.2377 0.1837 0.1863 0.1401 0.2341

In summary, the Model Study Area Maxent model used in this analysis was evaluated
using a 5-fold sub-sample with 50% of the presence points set aside randomly for testing the
model. The Model Study Area model was then transferred to Test Area 1 and Test Area 2 and
79

the model suitability results for these three models (Model Study Area Maxent model, Test Area
1 Maxent model, and Test Area 2 Maxent model) were then compared. AUC, test omission rate,
TSS, and individual variable contributions were used as indicators of model fitness and
transferability success. The results of this study showed acceptable AUC (0.725) and fair TSS
(0.409) with a good omission rate (0.0832) for the Model Study Area. For Test Area 1 the AUC
was also acceptable (0.746) and TSS fair (0.394) with good omission rate (0.0807). Test Area 2
produced a good AUC (0.8460) but with a poor TSS (0.238) and poorer omission rate than the
other models tested (0.2941). Overall the covariates with the most influence on the model, as
determined by the permutation importance and review of the Jackknife of test gain, were
PctCanopy (56.9) followed by EcolSys (14.7) and soil pH (8.3) for the Model Study Area. Test
Area 1’s most influential covariates were PctCanopy (30). PctSilt (16.8), and PartSize (16.3).
Test Area 2’s most influential covariates were Distance (43.5), followed by PctCanopy (36.6),
and Bed (7.5).

80

Chapter 5 Conclusions
The goal of this study was two-fold, to evaluate the fitness for use of Maxent in predicting the
potential distribution of cogongrass infestation given suitable conditions within the Model Study
Area and to test the transferability of that model to other study areas within the state of Alabama.
The guiding objective, beyond the generation of an appropriate model that is transferable across
various areas of the state, was the hope that the resulting model and transferability tests would be
useful in guiding future survey efforts and funding allocation decisions. In this chapter, we
provide a general overview of study concerns as well as results of the study. This chapter
concludes by providing a brief narrative on inferences gleaned from the study and potential
future work related hereto.
To evaluate the fitness for use of Maxent in predicting the probability of presence
distribution of cogongrass within the Model Study Area and to test the transferability of that
model to other study areas within the state of Alabama, it was important to thoroughly review the
species’ biological, climatic and ecological requirements. Cogongrass is highly tolerant to a wide
range of conditions and therefore determining the best environmental covariates to use within the
model was time consuming. Cogongrass’ range of habitat with relation to geographic location
(Latitudes 45ºN to 45ºS), rainfall (75 – 500cm average annual), elevation (sea level to 2000m),
soil organic matter, habitable sites, and temperature (tolerant to -14ºC) were all generally met
within the geographic boundary of the state of Alabama. Review of previous research on the
species was used to guide the environmental covariates used in the study, with a focus on land
use and soils related variables.
A review of the test areas’ output data was performed to determine which environmental
factors play the biggest role in transferability hit or miss. The percent contribution and
81

permutation importance of each variable was reviewed along with modeled response curves and
the results of the jackknife of regularized test gain. In this study, the most relevant environmental
covariate for all three study sites was percent canopy. Percent canopy was the variable with the
highest level of permutation importance and percent contribution for both the Model Study Area
and Test Area 1 and was the second highest in these factors for Test Area 2. Percent canopy was
also in the top three for effect on gain according to the jackknife of regularized test gain graph
included in the Maxent output dataset. Therefore, this environmental variable should be included
in any future work related to the species. Ecological system, distance to road, percent silt and
percent clay also showed significance in this study.
Several of the layers selected for use in the study empirically have some degree of
correlation, for instance the particle size layer is a general classification (grouping) of the soil
texture as determined by grain size for the topmost horizon of soil using the standards used by
the U.S. Department of Agriculture. The individual soil texture layers (percent clay, percent
sand, and percent silt) are considered in the particle size layer to some degree. However, since
particle size is a classified categorical dataset and the three soil texture layers are discrete
measurable values, it was determined that both types of data could be included without issue.
Correlation between datasets should be taken into consideration when analyzing SDMs such as
that produced in this study.
5.1. Uncertainty in the Model
This model may be used to support decisions related to where to survey for cogongrass
locations and what counties to focus on for eradication efforts. Therefore, it is important that the
uncertainty in the model be clearly understood so that the value of the model results can be
articulated to stakeholders. Sample selection bias is a fundamental limitation of presence only
82

modeling such as is the case in this study using Maxent. This bias can have a significant impact
on the model outcome (Elith 2011; Phillips et al. 2009). Examples of sample selection bias can
be found in this study and in the Ervin and Holly (2011) study, which specifically focused on a
biased sample by sampling along roadways.
In the current study, sample selection bias is introduced by the method of discovery and
subsequent reporting of suspected cogongrass location points to the AFC. The AFC relies
heavily on landowner and public reporting of suspected point locations and then investigates and
verifies those locations. This bias cannot be removed due to the nature of infestation reporting;
however, it is prudent to take it into account when analyzing the model result. Also, some
sampling bias can be removed from the study in areas where the species presence point data is
not uniformly scattered across the entire extent of the test area’s geography space by use of a
minimum bounding geometry layer that will ensure that any test or background pseudo-presence
data predictions will use the same geographic boundary as the training data.
5.2. Proposed Future Work
Future work related to this study should include the testing of transferability across
additional AFC Work Units and potentially recalibrating and retesting models as new data
becomes available (Stohlgren and Schnase 2006). It would be appropriate to test the model
against all AFC Work Units in a future study as differences in model performance was noted
between the areas included in this study. Further, in the Stohlgren and Schnase study, it was
recommended that the modeling process be an iterative process in which the model is
recalibrated as new species presence data become available (Stohlgren and Schnase 2006; Crall
et al. 2013). Thus, if a model is used to prompt a guided species survey, the survey results can
then be added to the volume of existing species presence point data and the model can be rerun
83

to create a new model with this updated sample layer to better inform the next guided survey.
This would be especially important if the output of the model was to be used to guide funding for
control and eradication efforts in the future.
Additionally, field verification of model output to determine if the predicted locations do,
in fact, support cogongrass infestations would be useful. And finally, further work into
transportation corridor related factors on the distribution of cogongrass should also be
considered. Specifically, the roads layer used in this study did not contain all interior local roads,
especially in heavily timbered and rural locations. Since Alabama has a high percentage of
forested area, it would be prudent to launch a project to update the roads layer used in the
distance to roads calculation or pursue the purchase or construction of a better suited roads layer.
5.3. Findings
Given the acceptable AUC, omission rate and TSS values of the original Model Study
Area’s Maxent model output, the model produced for this study can be considered to be an
appropriate model for predicting the presence of cogongrass in the Model Study Area. The
transferability test for the model leaves some open questions, however. The results of this study
showed that when the area targeted for transfer is similar environmentally and geographically to
the Modeled Study Area, this model can perform sufficiently well to be used to inform the
analyst on predicted probability in the target area. When the target area is highly dissimilar, as is
the case with Test Area 2 in this study, caution should be taken when transferring the model to
this new geographic space. It would perhaps be more valid to re-evaluate the model against the
new geographic area and re-run with a modified set of covariates as appropriate.
In summary, the model produced by Maxent for the Model Study Area had an AUC of
0.725 which is considered to be acceptable for use in conservation planning (Elith et al. 2006).
84

The environmental covariates selected for the study were suitably broad in their biological and
ecological suitability to the species being studied to allow for successful transfer of the model to
two other AFC Work Units within the state, however detailed review of the model results using
multiple metrics for testing fitness should be employed when verifying model transferability
success.
This study adds to the body of work related to species distribution modeling using
Maxent for cogongrass as well as transferability studies of Maxent models for invasive species in
general. Although additional work is suggested to further this study of transferability of Maxent
model for cogongrass, the findings of this study suggest that Maxent is potentially a suitable tool
for modeling the predicted potential distribution of cogongrass infestation given suitable
biological and ecological variables are utilized. This study also suggests that a suitably trained
Maxent model can be successfully projected to similar geographic areas within a limited extent,
such as a state as was tested here. The transfer of a suitably trained Maxent model to an area of
dissimilar geographic or environmental conditions, should be accepted with caution.

85

References
Akobundu, I. O., and F. E. Ekeleme. 2000. “Effect of method of Imperata cylindrica
management on maize grain yield in the derived savanna of south-western Nigeria.”
Weed Research. 40(4): 335–341.
Alabama Forestry Commission. n.d. “Cogongrass.” Accessed April 18
th
, 2018.
http://www.forestry.state.al.us/Pages/Informational/Invasive/Cogongrass.aspx.
Alabama Forestry Commission. n.d. “Cogongrass viewer.” Accessed April 18
th
, 2018.
http://www.forestry.state.al.us/viewers/afc_cogongrass_viewer.aspx
Allouche, Omri, Asaf Tsoar, and Ronen Kadmon. 2006. “Assessing the accuracy of species
distribution models: prevalence, Kappa and the true skill statistic (TSS).” Journal of
Applied Ecology 43 (6):1223-1232. doi: 10.1111/j.1365-2664.2006.01214.x.
Anderson, R.P. 2012. “Harnessing the world’s biodiversity data: promise and peril in ecological
niche modeling of species distributions.” Annals of the New York Academy of Sciences
1260: 66-80.
Ayeni, A.0. 1985. “Observations on the vegetative growth pattern of speargrass (Imperata
cylindrica (L.) Beauv.).” Agriculture. Ecosystems and Environment 13 (3-4): 301-307.
doi: 10.1016/0167-8809(85)90017-9.
Clemson University. N.d. “Clemson Regulatory Services.” Accessed November 20, 2018.
https://www.clemson.edu/public/regulatory/plant-
protection/invasive/cogongrass/index.html
Coulston, J. W., Gretchen G. Moisen, B.T. Wilson, M.V. Finco, W.B. Cohen, and C.K. Brewer,
2012. “Modeling percent tree canopy cover: a pilot study.” Photogrammetric
Engineering & Remote Sensing 78(7): 715-727.
Crall, Alycia W., Catherine S. Jarnevich, Brendon Panke, Nick Young, Mark Renz, and Jeffrey
Morisette. 2013. “Using habitat suitability models to target invasive plant species
surveys.” Ecological Applications 23: 60-72. doi: 10.1890/12-0465.1.
Databasin. September 3, 2014. “GAP Land Cover Data for Alabama, USA” Last Accessed
March 14, 2019. https://databasin.org/datasets/e6c2c82715be44bba3579fa6010acfd5.
Dozier, Hallie. Sandra K. Gaffney, Eric McDonald, R.R.L. Johnson, and Donn G. Shilling. 1998.
“Cogongrass in the United States: History, Ecology, Impacts, and Management.” Weed
Technology. 12: 737-743.
Dickens, R. 1974. “Cogongrass in Alabama after sixty years.” Weed Science 22: 177-179.
86

EDDMapS. 2019. “Early Detection & Distribution Mapping System.” Accessed August 4, 2019.
The University of Georgia - Center for Invasive Species and Ecosystem Health.
http://www.eddmaps.org/.
Elith, Jane, C.H. Graham, Robert P. Anderson, Miroslav Dudík, Simon Ferrier, Antoine Guisan,
Robert J. Hijmans, et al. 2006. “Novel Methods Improve Prediction of Species’
Distributions from Occurrence Data.” Ecography 29, no. 2(April): 129-151.
Elith, Jane, Steven J. Phillips, Trevor Hastie, Miroslav Dudik, Yung En Chee, and Colin J. Yates.
2011. “A statistical explanation of Maxent for ecologists.” Diversity and Distributions
17: 43-57.
Enloe, S.F., D.K. Lauer, N.J. Loewenstein, and R.D. Lucardi. 2018. “Response of twelve Florida
cogongrass (Imperata cylindrica) populations to herbicide treatment.” Invasive Plant
Science and Management 11(2): 82-88.
Ervin, Gary N., and D.C. Holly. 2011. “Examining Local Transferability of Predictive Species
Distribution Models for Invasive Plants: An Example with Cogongrass (Imperata
cylindrica).” Invasive Plant Science and Management 4 (4): 390-401.
Esri. n.d. “ArcMap: Extract by Mask” Accessed August 1
st
, 2019.
http://desktop.arcgis.com/en/arcmap/10.6/tools/spatial-analyst-toolbox/extract-by-
mask.htm.
Estrada, James A., and S. Luke Flory. 2014. “Cogongrass (Imperata cylindrica) invasions in the
US: Mechanisms, impacts, and threats to biodiversity.” Global Ecology and
Conservation 3: 1-10.
Eussen, J.H.H., and S. Wirjahardja. 1973. “Studies of an alang-alang, Imperata cylindrica (L.)
Beauv. vegetation.” Biotropica Bullitan. no. 6.
Gaffney, J.F. 1996. “Ecophysiological and Technical Factors Influencing the Management of
Cogongrass (Imperata cylindrica).” Ph.D. dissertation, University of Florida.
Halvorsen, Rune, Sabrina Mazzoni, John Wirkola Dirksen, Erik Næsset, Terje Gobakken, and
Mikael Ohlson. 2016. “How Important Are Choice of Model Selection Method and
Spatial Autocorrelation of Presence Data for Distribution Modelling by Maxent?”
Ecological Modelling 328: 108–118.
Hijmans, R.J., S.E. Cameron, J.L. Parra, P.G. Jones and A. Jarvis. 2005. “Very high-resolution
interpolated climate surfaces for global land areas.” International Journal of Climatology
25: 1965-1978.
Holm, L.G., D.L. Pucknett, J.B. Pancho, and J.P. Herberger. 1977. “The World’s Worst Weeds.
Distribution and Biology”. Honolulu HI: University of Hawaii Press.
Howard, Janet L. 2005. “Imperata brasiliensis, I. cylindrica. In: Fire Effects Information
System.” Accessed March 24
th
, 2019. U.S. Department of Agriculture, Forest Service,
87

Rocky Mountain Research Station, Fire Sciences Laboratory. https://www.fs.fed.us
/database/feis/plants/graminoid/impspp/all.html.
Hubbard, C.E. 1944. “Imperata cylindrica. Taxonomy, Distribution, Economic significance, and
Control.” Agricultural Bureau Joint Publication. No. 7, Imperial Bureau Pastures and
Forage Crops, Aberystwyth, Wales. Great Britton.
King, Sharon E., and James B. Grace, 2000. “The Effects of Soil Flooding on the Establishment
of Cogongrass (Imperata cylindrica), a Nonindigenous Invader of the Southeastern
United States.” Wetlands 20(2): 300-306.
Lee, S.A. 1977. “Germination, rhizome survival, and control of Imperata cylindrica(L.) Beauv.
on peat.” MARDI Research. Bulletin 5(2): 1-9.
Lippincott, Carol L. 1997. “Ecological consequences of Imperata cylindrica (cogongrass)
invasion in Florida sandhill.” Dissertation, University of Florida.
Lippincott, C.L. 2000. “Effects of I. cylindrica (cogongrass) invasions on fire regimes in Florida
sandhill.” Natural Area Journal 20: 140–149.
Livingston, M.J., and C. Osteen. 2008. “Integrating Invasive Species Prevention and Control
Policies.” Economic Brief No. 11. USDA, Economic Research Service.
Lucardi, Rima, Lisa Wallace, and Gary Ervin. 2014. “Invasion Success in Cogongrass (Imperata
cylindrica): A Population Genetic Approach Exploring Genetic Diversity and Historical
Introductions.” Invasive Plant Science and Management 7, no. 1: 59–75.
MacDonald, G.E. 2004. “Cogongrass (Imperata cylindrica)-Biology, Ecology, and
Management.” Critical Reviews in Plant Sciences. 23(5): 367-380.
McNeely, Jeffrey A. (ed.). 2001. The Great Reshuffling: Human dimensions of invasive alien
species. IUCN, Gland, Switzerland and Cambridge, UK.
Merow, Cory, Matthew J. Smith, and John A. Silander. 2013. “A Practical Guide to Maxent for
Modeling Species’ distributions: What it Does, and Why Inputs and Settings Matter.”
Echography 36(10): 1058-1069. doi: 10.1111/j.1600-0587.2013.07872.x.
Monserud, Robert A., and Rik Leemans. 1992. “Comparing global vegetation maps with the
Kappa statistic.” Ecological Modelling, 62: 275-293.
Narkhede, Sarang. 2018. Understanding AUC-ROC Curve. Towards Data Science. Posted June
26th, 2018. Accessed 11/5/2018. https://towardsdatascience.com/understanding-auc-roc-
curve-68b2303cc9c5
O‘Sullivan, David, and George W. Perry. 2013. Spatial Simulation: Exploring Pattern and
Process. London and New York: John Wiley & Sons.
88

Patterson, D.T. 1980. “Shading effects on growth and partitioning of plant biomass in
cogongrass (Imperata cylindrica) from shaded and exposed habitats.” Weed Science. 28:
735-740.
Peterson, A.T., J. Soberon, R.G. Pearson, R.P. Anderson, E. Ma r tıne z-Meyer, M. Nakamura, and
M.B. Araujo. 2011. Ecological niches and geographic distributions. Princeton University
Press, Princeton, NJ.
Phillips, Steven J. 2017. “A Brief Tutorial on Maxent.” Accessed 1/5/2019.
http://biodiversityinformatics.amnh.org/open_source/Maxent/.
Phillips, Steven J., and Miroslav Dudik. 2007. “Modeling of species distributions with Maxent:
new extensions and a comprehensive evaluation.” Ecography 31: 161-175
Phillips, Steven J., Miroslav Dudík, and Robert E. Schapire, “Maxent software for modeling
species niches and distributions” (Version 3.4.1). Accessed on 1/5/2019.
http://biodiversityinformatics.amnh.org/open_source/Maxent.
Radosavljevic, Aleksandar and Robert P. Anderson. 2014. “Making better MAXENT models of
species distributions: complexity, overfitting and evaluation.” Journal of Biogeography
41: 629-643.
Rauschert, Emily S.J., David A. Mortensen, and Steven M. Bloser. 2017. “Human-mediated
dispersal via rural road maintenance can move invasive propagules.” Biological
Invasions 19: 2047-2058. doi: 10.1007/s10530-017-1416-2.
Sajise, P.E. 1976. “Evaluation of cogon (Imperata cylindrica) as a serial stage in Philippine
vegetational succession. 1. The cogonal seral stage and plant succession. 2. Autecological
studies on cogon. Dissertation Abstracts International B: 3040-3041. Weed Abstracts no.
1339.
Stohlgren, T.J., and J.L. Schnase. 2006. “Risk analysis for biological hazards: What we need to
know about invasive species.” Risk Analysis 26: 163-173.
Tabor, P. 1949. “Cogongrass, Imperata cylindrica (L.) Beauv., in the southeastern United
States.” Agronomic Journal. 41: 270.
Tabor, P. 1952. “Comments on cogon and torpedograsses: A challenge to weed workers.”
Weeds. 1: 374-375.
Terry, P.J., G. Adjiers, I.O. Akobundu, A.U. Anoka, M.E. Drilling, S. Tjitrosemito, and M.
Utomo. 1997. “Herbicides and mechanical control of Imperata cylindrica as a first step in
grassland rehabilitation.” Agroforestry Systems. 36:151–179
West, Amanda M., Sunil Kumar, Cynthia S. Brown, Thomas J. Stohlgren, and Jim Bromberg.
2016. “Field validation of an invasive species Maxent model.” Ecological Informatics.
36: 126-134. doi:10.1016/j.ecoinf.2016.11.001
89

Wilcut, J.W., R.D. Dute, B. Truelove, and D.E. Davis. 1988a. “Factors limiting the distribution
of cogongrass, Imperata cylindrica, and torpedograss, Panicum repens.” Weed Science.
36: 577-582.
Wilcut, J.W., B. Truelove, D.E. Davis, and J.C. Williams. 1988b. “Temperature factors limiting
the spread of cogongrass (Imperata cylindrica) and torpedograss (Panicum repens).”
Weed Science. 36: 49-55.
Willard, T.R., D.W. Hall, D.G. Shilling, J.A. Lewis, and W.L. Currey. 1990. “Cogongrass
(Imperata cylindrica) distribution on Florida highway rights-of-way.” Weed Technology.
4: 658-660.
U.S. Department of Agriculture. n.d. “SSURGO Soil Map Coverage versus the U.S. General Soil
Map Coverage.” Accessed April 10, 2019.
https://www.nrcs.usda.gov/wps/portal/nrcs/detail/soils/survey/geo/?cid=nrcs142p2_0536
26
U.S. Department of Agriculture. n.d. “NRCS Plants Database”. Accessed April 10, 2019.
https://plants.usda.gov/core/profile?symbol=IMCY
U.S. Fish and Wildlife Service. n.d. “Invasive Species.” www.fws.gov/invasives/faq.html#q2
U.S. Forest Service. n.d. “Invasive Species Profile.”
https://www.fs.fed.us/invasivespecies/speciesprofiles/documents/cogon-grass.pdf
Yager, Lisa, Deborah Miller, and Jeanne Jones. 2011. “Woody Shrubs as a Barrier to Invasion by
Cogongrass (Imperata cylindrica)” Invasive Plant Science and Management Apr-Jun
2011, Vol.4(2), pp. 207-211. Doi: 10.1614/IPSM-D-10-00052.1.
Young, Nick, Carter Lane, and Paul Evangelista. 2011. “A Maxent Model v3.3.3e Tutorial
(ArcGISv10).” Natural Resource Ecology Laboratory at Colorado State University and
the National Institute of Invasive Species Science.
Yu, Feng. 2013. “Improving Model Performance For Invasive Plant Species Distribution Using
Global-Scale Presence-Only Data: Parameterization And Data Quality”. Master of
Science Thesis. Purdue University. https://docs.lib.purdue.edu/open_access_theses/112

90

Appendix A: Soils Related Environmental Covariate Maps

91

92

93

94

95

96

97

98

99

100

101

102

103

104

105

106

107

108

109

110

111

Appendix B: Other Environmental Covariate Maps

112

113

114

115

116

117

118

119

120

Appendix C: Data Layer Conversion Steps
This table represents the step for converting the presence point data to .csv for use in Maxent
Step Description
1 Open the cogongrass points shapefile in ArcGIS
2 Add new columns for LatYDD and LongXDD to the shapefile table
3 Calculate the Latitude and Longitude geometry as decimal degrees
4 Use the Table to Excel tool in ArcGIS to dump the data into Excel format
5 Open the file in Microsoft Excel
6 Convert the .xls file to .csv
7 Open the .csv file
8 Remove all columns except Species, LatYDD, and LongXDD
9 Save the file

121

This table of conversion steps includes data sources and layers that were ultimately not used in
the final model; however, it may be useful for the reader to review how data from these sources
were prepared before their usefulness was determined to be insignificant to the study.
Environmental Layer Description and Conversion Steps
PRISM climate data • Download climate data from the PRISM Climate Group website
(http://www.prism.oregonstate.edu/)
• Each dataset is a raster dataset at 800M resolution (roughly ½ mile
grid cells). The values in the dataset are presented in millimeters and
the rasters are classified in 5-inch increments. These datasets are in
Nad83.
• Data conversion steps:
• In ArcGIS
• Open the AVG Precipitation dataset
• Clip AVG Precipitation to the study area geometry (the
state of Alabama)
• Save as “AVGPrecipClip”
• Open the AVG Min Temp dataset
• Clip AVG Min Temp to the study area geometry (the
state of Alabama)
• Save as “AVGMinTempClip”
• Open the AVG Max Temp dataset
• Clip AVG Max Temp to the study area geometry (the
state of Alabama)
• Save as “AVGMaxTempClip”
Digital Elevation Model • Download digital elevation models for the state of Alabama
• In ArcGIS
• Use the Merge to New Raster tool to merge the DEMs together into
one raster
• Use Clip Raster tool to clip the new raster to the study area

122

Appendix D: Maxent Model Settings Screen Captures
Model Study Area Maxent Model Settings:

123

Test Study Area 1 (Work Unit 12) Maxent Model Settings:

124

Test Study Area 2 (Work Unit 8) Maxent Model Settings:

125

Appendix E: Response Curves
Response Curves for Imperata cylindrica to each environmental variable included in the models
for each of the three study areas.
126

Appendix F: Ecological Systems with Category Groupings
Ecological Systems for Model Study Area:
ID Ecological System % of
Total
Category
1 Cultivated Cropland 1.21% Agriculture
2 Developed, High Intensity 0.03% Developed
3 Developed, Low Intensity 0.26% Developed
4 Developed, Medium Intensity 0.09% Developed
5 Developed, Open Space 2.80% Developed
6 Disturbed/Successional - Shrub Regeneration 1.83% Disturbed
7 East Gulf Coastal Plain Black Belt Calcareous Prairie and
Woodland - Herbaceous Modifier
0.02% Forest/Woodlands
8 East Gulf Coastal Plain Black Belt Calcareous Prairie and
Woodland - Woodland Modifier
0.08% Forest/Woodlands
9 East Gulf Coastal Plain Dry Chalk Bluff 0.00% other
10 East Gulf Coastal Plain Interior Upland Longleaf Pine
Woodland - Loblolly Modifier
34.12% Forest/Woodlands
11 East Gulf Coastal Plain Interior Upland Longleaf Pine
Woodland - Offsite Hardwood Modifier
8.37% Forest/Woodlands
12 East Gulf Coastal Plain Interior Upland Longleaf Pine
Woodland - Open Understory Modifier
0.71% Forest/Woodlands
13 East Gulf Coastal Plain Large River Floodplain Forest -
Forest Modifier
8.83% Floodplain forest
14 East Gulf Coastal Plain Large River Floodplain Forest -
Herbaceous Modifier
0.16% Floodplain forest
15 East Gulf Coastal Plain Limestone Forest 0.09% Forest/Woodlands
16 East Gulf Coastal Plain Northern Mesic Hardwood Forest 0.10% Floodplain forest
17 East Gulf Coastal Plain Small Stream and River Floodplain
Forest
8.11% Floodplain forest
18 East Gulf Coastal Plain Southern Loblolly-Hardwood
Flatwoods
0.94% Forest/Woodlands
19 East Gulf Coastal Plain Southern Mesic Slope Forest 5.72% Floodplain forest
20 Evergreen Plantation or Managed Pine 5.66% Forest/Woodlands
21 Harvested Forest-Shrub Regeneration 7.54% Disturbed
22 Harvested Forest - Grass/Forb Regeneration 1.61% Disturbed
23 Open Water (Aquaculture) 0.03% water
24 Open Water (Fresh) 1.15% water
25 Pasture/Hay 6.86% Agriculture
26 Quarries, Mines, Gravel Pits and Oil Wells 0.01% Developed
27 Southern Coastal Plain Blackwater River Floodplain Forest 3.61% Floodplain forest
28 Unconsolidated Shore 0.01% other
29 Undifferentiated Barren Land 0.06% other
127

Ecological Systems for Study Area 1:
ID Ecological System Description % of
Total
Category
1 Cultivated Cropland 7.67% Agriculture
2 Developed, High Intensity 0.03% Developed
3 Developed, Low Intensity 0.53% Developed
4 Developed, Medium Intensity 0.10% Developed
5 Developed, Open Space 2.36% Developed
6 Disturbed/Successional - Shrub Regeneration 1.99% Disturbed
7 East Gulf Coastal Plain Interior Upland Longleaf Pine
Woodland - Loblolly Modifier
29.35% Forest/Woodlands
8 East Gulf Coastal Plain Interior Upland Longleaf Pine
Woodland - Offsite Hardwood Modifier
8.78% Forest/Woodlands
9 East Gulf Coastal Plain Interior Upland Longleaf Pine
Woodland - Open Understory Modifier
2.68% Forest/Woodlands
10 East Gulf Coastal Plain Large River Floodplain Forest -
Forest Modifier
4.33% Floodplain forest
11 East Gulf Coastal Plain Large River Floodplain Forest -
Herbaceous Modifier
0.16% Floodplain forest
12 East Gulf Coastal Plain Small Stream and River Floodplain
Forest
5.12% Floodplain forest
13 East Gulf Coastal Plain Southern Mesic Slope Forest 6.94% Floodplain forest
14 Evergreen Plantation or Managed Pine 6.07% Forest/Woodlands
15 Harvested Forest-Shrub Regeneration 7.25% Disturbed
16 Harvested Forest - Grass/Forb Regeneration 3.15% Disturbed
17 Open Water (Fresh) 0.68% Water
18 Pasture/Hay 6.97% Agriculture
19 Quarries, Mines, Gravel Pits and Oil Wells 0.07% Developed
20 Southern Coastal Plain Blackwater River Floodplain Forest 5.56% Floodplain forest
21 Southern Coastal Plain Nonriverine Cypress Dome 0.13% Floodplain forest
22 Unconsolidated Shore 0.01% other
23 Undifferentiated Barren Land 0.07% other

128

Ecological Systems for Study Area 2:
ID Ecological System Description % of
Total
Category
1 Allegheny-Cumberland Dry Oak Forest and Woodland -
Hardwood
7.20% Forest/Woodlands
2 Allegheny-Cumberland Dry Oak Forest and Woodland -
Pine Modifier
0.92% Forest/Woodlands
3 Cultivated Cropland 3.38% Agriculture
4 Cumberland Riverscour 0.16% water
5 Developed, High Intensity 0.53% Developed
6 Developed, Low Intensity 4.48% Developed
7 Developed, Medium Intensity 1.32% Developed
8 Developed, Open Space 8.73% Developed
9 Disturbed/Successional - Grass/Forb Regeneration 1.87% Disturbed
10 Disturbed/Successional - Shrub Regeneration 3.05% Disturbed
11 Evergreen Plantation or Managed Pine 5.36% Forest/Woodlands
12 Harvested Forest-Shrub Regeneration 1.72% Disturbed
13 Harvested Forest - Grass/Forb Regeneration 1.56% Disturbed
14 Northeastern Interior Dry Oak Forest - Mixed Modifier 0.00% Forest/Woodlands
15 Open Water (Fresh) 1.65% water
16 Pasture/Hay 17.46% Agriculture
17 South-Central Interior Large Floodplain - Forest Modifier 0.09% floodplain forest
18 South-Central Interior Mesophytic Forest 6.88% Forest/Woodlands
19 South-Central Interior Small Stream and Riparian 0.99% water
20 Southeastern Interior Longleaf Pine Woodland 0.29% Forest/Woodlands
21 Southern Appalachian Low Mountain Pine Forest 8.70% Forest/Woodlands
22 Southern Interior Acid Cliff 0.00% other
23 Southern Interior Calcareous Cliff 0.00% other
24 Southern Interior Low Plateau Dry-Mesic Oak Forest 0.00% Forest/Woodlands
25 Southern Piedmont Cliff 0.00% Other
26 Southern Piedmont Dry Oak-(Pine) Forest - Hardwood
Modifier
0.64% Forest/Woodlands
27 Southern Piedmont Dry Oak-(Pine) Forest - Loblolly Pine
Modifier
0.08% Forest/Woodlands
28 Southern Piedmont Dry Oak-(Pine) Forest - Mixed
Modifier
0.09% Forest/Woodlands
29 Southern Piedmont Mesic Forest 0.11% Forest/Woodlands
30 Southern Piedmont Small Floodplain and Riparian Forest 0.04% floodplain forest
31 Southern Ridge and Valley Dry Calcareous Forest 20.74% Forest/Woodlands
32 Southern Ridge and Valley Dry Calcareous Forest - Pine
modifier
1.32% Forest/Woodlands
33 Undifferentiated Barren Land 0.64% Other

129

Ecological Systems for the Model Study Area, Test Area 1, and Test Area 2 with
consolidated groupings.

Abstract (if available)

Abstract As of April 19th, 2018, there were 34,771 verified locations of cogongrass (Imperata cylindrica (L.) Beauv.) infestations within the state of Alabama. Cogongrass is a highly invasive non-native species of rhizomatous grass that is considered one of the ten worst weeds worldwide. This highly invasive and environmentally destructive species has caused significant damage throughout its current distribution and efforts to control and eradicate the threat have been underway for almost a decade. This study utilized the Maximum Entropy (Maxent) model to predict the location of invasive cogongrass within the state of Alabama. The model developed using the presence locations and environmental data for the Model Study Area, one Alabama Forest Commission (AFC) Work Unit, was applied to two additional AFC Work Units to test transferability of the model to areas of similar and dissimilar ecological and geographic makeup. The Model Study Area’s Maxent model resulted in an acceptable AUC (0.725 with sd = 0.0010) and fair TSS score (0.4087) with a test omission rate of 0.0832. Transferability test results differed between the two test areas. Using the Model Study Area’s model on Test Area 1, an area similar in most aspects to the Model Study Area, resulted in an AUC of 0.746 with a standard deviation of 0.002, a TSS score of 0.3944 and a test omission rate of 0.0807. These results indicated that the original model was sufficiently transferable to the similar Test Area 1. Test Area 2 was dissimilar from the Model Study Area in most environmental covariates as well as number of verified presence point locations. Applying the model to Test Area 2 resulted in an AUC of 0.846 with a standard deviation of 0.017, a TSS score of 0.2377 and a test omission rate of 0.2941. These results suggest the need for some concern about the suitability of the transferred model to Test Area 2.

Linked assets

University of Southern California Dissertations and Theses

Conceptually similar

PDF

Predicting archaeological site locations in northeastern California’s High Desert using the Maxent model

PDF

Building better species distribution models with machine learning: assessing the role of covariate scale and tuning in Maxent models

PDF

Predicting the presence of historic and prehistoric campsites in Virginia’s Chesapeake Bay counties

PDF

A Maxent-based model for identifying local-scale tree species richness patch boundaries in the Lake Tahoe Basin of California and Nevada

PDF

Species distribution modeling to predict the spread of Spartium junceum in the Angeles National Forest

PDF

Selection of bridge location over the Merrimack River in southern New Hampshire: a comparison of site suitability assessments

PDF

Predicting Hydromantes shastae occurrences in Shasta County, California

PDF

Modeling burn probability: a Maxent approach to estimating California's wildfire potential

PDF

Using Maxent to model the distribution of prehistoric agricultural features in a portion of the Hōkūli‘a subdivision in Kona, Hawai‘i

PDF

Habitat suitability modeling of Mexican spotted owl (Strix occidentalis lucida) in Gila National Forest, New Mexico

PDF

A critical assessment of the green sea turtle central west Pacific distinct population segment utilizing maxent modeling on nesting site locations

PDF

Using volunteered geographic information to model blue whale foraging habitat, Southern California Bight

PDF

Evaluating predator prey dynamics and site utilization patterns of golden eagles using resource selection modeling and spatiotemporal pattern mining

PDF

A comparison of GLM, GAM, and GWR modeling of fish distribution and abundance in Lake Ontario

PDF

A model for emergency logistical resource requirements: supporting socially vulnerable populations affected by the (M) 7.8 San Andreas earthquake scenario in Los Angeles County, California

PDF

Modeling nitrate contamination of groundwater in Mountain Home, Idaho using the DRASTIC method

PDF

Integration of topographic and bathymetric digital elevation model using ArcGIS interpolation methods: a case study of the Klamath River Estuary

PDF

Utilizing GIS and remote sensing to determine sheep grazing patterns for best practices in land management protocols

PDF

Preparing for immigration reform: a spatial analysis of unauthorized immigrants

PDF

Archaeological least cost path modeling: a behavioral study of Middle Bronze Age merchant travel routes across the Amanus Mountains, Turkey

Asset Metadata

Creator Shanks, Rachel Eagle (author)

Core Title Assessing the transferability of a species distribution model for predicting the distribution of invasive cogongrass in Alabama

School College of Letters, Arts and Sciences

Degree Master of Science

Degree Program Geographic Information Science and Technology

Publication Date 10/11/2019

Defense Date 08/30/2019

Publisher University of Southern California (original), University of Southern California. Libraries (digital)

Tag Alabama,AUC,cogongrass,Imperata cylindrica,invasive species management,Maxent,maximum entropy,model transferability,OAI-PMH Harvest,SDM,species distribution model,TSS

Language English

Contributor Electronically uploaded by the author (provenance)

Advisor Kemp, Karen (committee chair), Lee, Su Jin (committee member), Oda, Katsuhiko (committee member)

Creator Email rshanks@usc.edu,wohalirke@yahoo.com

Permanent Link (DOI) https://doi.org/10.25549/usctheses-c89-225262

Unique identifier UC11673383

Identifier etd-ShanksRach-7854.pdf (filename),usctheses-c89-225262 (legacy record id)

Legacy Identifier etd-ShanksRach-7854.pdf

Dmrecord 225262

Document Type Thesis

Rights Shanks, Rachel Eagle

Type texts

Source University of Southern California (contributing entity), University of Southern California Dissertations and Theses (collection)

Access Conditions The author retains rights to his/her dissertation, thesis or other graduate work according to U.S. copyright law. Electronic access is being provided by the USC Libraries in agreement with the a...

Repository Name University of Southern California Digital Library

Repository Location USC Digital Library, University of Southern California, University Park Campus MC 2810, 3434 South Grand Avenue, 2nd Floor, Los Angeles, California 90089-2810, USA