University of Southern California Dissertations and Theses

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

(USC Thesis Other)

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

PDF

Download a page range

Download transcript

Copy asset link

Request this asset

Transcript (if available)

Content

Species Distribution Modeling to Predict the Spread of Spartium junceum in
the Angeles National Forest

By

Jason I Martin

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

Copyright ® 2019 by Jason Martin
ii

I dedicate this to my wife, Christina Mireles Martin, who inspired me to continue my
education and never give up on pursuing the dreams and goals I had set for myself when
I was just a teenager. I also dedicate this to my beautiful children Elena Rose Martin and
Isaiah Eusevio Martin who have helped me regain life perspective and are the driving
force for me to persevere. To my amazing family, I love you all and I thank you for
empowering me in ways I never knew were possible.

iii
Table of Contents
List of Figures ................................................................................................................................. v
List of Tables ................................................................................................................................ vii
Acknowledgements ...................................................................................................................... viii
List of Abbreviations ..................................................................................................................... ix
Abstract ........................................................................................................................................... x
Chapter 1 Introduction .................................................................................................................... 1
1.1 Motivation ............................................................................................................................2
1.2 Study Area ...........................................................................................................................6
1.3 Surveyed Invasive Species ...................................................................................................7
1.4 Species Distribution Modeling and Maxent Explained .......................................................8
1.4.1. Multiple Maxent Models..........................................................................................10
1.5 Thesis Structure .................................................................................................................11
Chapter 2 Related Work................................................................................................................ 12
2.1 Spartium junceum and Invasive Species ............................................................................13
2.1.1. Spartium junceum History .......................................................................................13
2.1.2. Spartium junceum Characteristics ...........................................................................14
2.1.3. Spartium junceum Habitat Types and Plant Communities ......................................14
2.1.4. Impacts of Spartium junceum on ecosystems ..........................................................15
2.1.5. Spartium junceum Management and Control Methods ...........................................15
2.2 Species Distribution Modeling ..........................................................................................16
2.2.1. Presence-Absence Data Modeling ...........................................................................17
2.2.2. Presence-Only Data Modeling .................................................................................21
Chapter 3 Methodology ................................................................................................................ 29
3.1 Research Approach and Design .........................................................................................29
3.2 Research Sample Data .......................................................................................................31
3.2.1. Sample Biological Data ...........................................................................................32
3.2.2. Environmental Variables/Layers .............................................................................34
3.2.3. Data Processing and Standardization .......................................................................37

iv
Chapter 4 Results .......................................................................................................................... 43
4.1 Current Distribution of Spartium junceum in the Angeles National Forest .......................44
4.2 Maxent results using BiocClim Data at 1 km resolution ...................................................44
4.3 Maxent results using BioClim Data at 10 m resolution .....................................................49
4.4 Maxent results using DEM Raster layers at 10 m

resolution .............................................54
4.5 Maxent results using Euclidean Distance ANF features layers at 10 m resolution ...........57
4.6 Maxent results using combined 10 m BioClim and 10 m DEM layers .............................61
4.7 Maxent results using combined 10m BioClim, DEM, and Euclidean Distance ANF
Feature layers ................................................................................................................68
Chapter 5 Discussion and Conclusions ......................................................................................... 75
5.1 Strengths and Weaknesses .................................................................................................75
5.2 Interpretation of Maxent results for each set of layers.......................................................79
5.2.1 Interpretation of Variable Results .............................................................................83
5.3 Opportunities for future research .......................................................................................84
References ..................................................................................................................................... 87
Appendix A: List of Layers ......................................................................................................... 91

v
List of Figures
Figure 1 Angeles National Forest ................................................................................................... 6
Figure 2 Spartium junceum polygon data ..................................................................................... 33
Figure 3 Spartium junceum point data .......................................................................................... 34
Figure 4 Raster Conversion tools with Model Builder in ArcMap ............................................... 38
Figure 5 Maxent window with appropriate boxes checked .......................................................... 39
Figure 6 Maxent Basic Tab Settings ............................................................................................. 40
Figure 7 Maxent Advanced Tab Settings...................................................................................... 41
Figure 8 Spartium junceum point data in ANF ............................................................................. 45
Figure 9 Maxent suitability results using BioClim variables 1 km spatial resolution in ANF ..... 46
Figure 10 AUC-ROC for Spartium junceum ................................................................................ 47
Figure 11 Jackknife of regularized training gain for Spartium junceum ...................................... 47
Figure 12 Jackknife of regularized training gain for Spartium junceum ...................................... 48
Figure 13 Jackknife of AUC for Spartium junceum .................................................................... 48
Figure 14 Maxent results using BioClim 10 m resolution in ANF ............................................... 50
Figure 15 AUC-ROC for Spartium junceum .............................................................................. 51
Figure 16 Jackknife of regularized training gain for Spartium junceum ...................................... 51
Figure 17 Jackknife of regularized training gain for Spartium junceum ...................................... 52
Figure 18 Jackknife of AUC for Spartium junceum ..................................................................... 52
Figure 19 Maxent results using DEM at 10 m resolution in ANF ................................................ 54
Figure 20 AUC-ROC for Spartium junceum ................................................................................ 55
Figure 21 Jackknife of regularized training for Spartium junceum ............................................. 56
Figure 22 Jackknife of test gain for Spartium junceum ................................................................ 56
Figure 23 Jackknife of AUC for Spartium junceum ..................................................................... 56
Figure 24 Maxent results using Euclidean distance ANF features and 10m resoltuion in ANF .. 58
vi
Figure 25 AUC-ROC for Spartium junceum ................................................................................ 59
Figure 26 Jackknife of regularized training gain for Spartium junceum ...................................... 60
Figure 27 Jackknife of test gain for Spartium junceum ................................................................ 60
Figure 28 Jackknife of AUC gain for Spartium junceum ............................................................. 60
Figure 29 Maxent results in ANF using BioClim and DEM 10 m layers combined in ANF ....... 62
Figure 30 AUC-ROC using BioClim 10m and Dem 10m variables combined ............................ 63
Figure 31 Jackknife of regularized training for Spartium junceum .............................................. 64
Figure 32 Jackknife of test gain for Spartium junceum ................................................................ 65
Figure 33 Jackknife of AUC for Spartium junceum ..................................................................... 66
Figure 34 Maxent results using BioClim, DEM, and Euclidean distance ANF features 10 m
layers combined in ANF ........................................................................................................ 69
Figure 35 AUC-ROC Spartium junceum ...................................................................................... 70
Figure 36 Jackknife of regularized training for Spartium junceum .............................................. 71
Figure 37 Jackknife of test gain for Spartium junceum ................................................................ 72
Figure 38 Jackknife of AUC for Spartium junceum ..................................................................... 73
Figure 39 Spartium junceum polygon and point data after conversion ........................................ 78

vii
List of Tables
Table 1 WorldClim bioclimatic variables ..................................................................................... 35
Table 2 List of DEM Rasters ........................................................................................................ 36
Table 3 List of Feature Classes used to create Euclidean distance Rasters .................................. 36
Table 4 Variable contributions...................................................................................................... 49
Table 5 Variable contributions...................................................................................................... 53
Table 6 Variable contributions...................................................................................................... 57
Table 7 Variable contributions using Euclidean distance ANF features and 10 m resolution ..... 61
Table 8 Variable contributions using combined BioClim and DEM 10 m layers ........................ 67
Table 9 Variable contributions from the combined BioClim, DEM, and Euclidean distance ANF
features layers at 10 m resolution .......................................................................................... 74

viii

Acknowledgements
I would like to thank my advisor Dr. Travis Longcore, for his guidance, direction, and
for helping me get this process started on the right foot. I am especially grateful for his
ability to re-instill my confidence in this process and the ability to understand the
material in a very profound way. I would also like to thank Dr. John P. Wilson for
picking up where Dr. Travis Longcore left off and helping me get through the home
stretch and on to the finish line in one piece. Additionally, I would like to thank Drs. An-
Min Wu and Laura C. Loyola for coming on board at the eleventh hour and agreeing to
serve as my thesis committee. I would like to thank the faculty and staff of the Spatial
Sciences Institute for their patience with me throughout this entire process and their
willingness to work with me and accept me into this program.

ix
List of Abbreviations
ANF Angeles National Forest
AUC-ROC Area Under the curve Receiver Operating Characteristic
CAL-IPC California Invasive Plant Council
CT Classification Trees
CART Classification and Regression Trees
DEM Digital Elevation Model
ENFA Ecological Niche Factor Analysis
ENM Environmental Niche Modeling
GARP Genetic Algorithm for Rule-Set Prediction
GDB Geodatabase
GIST Geographic Information Science & Technology
GLM Generalized Linear Model
MAXENT Maximum Entropy
MLR Multiple Logistic Regression
OHV Off -highway Vehicle
SDM Species Distribution Modeling
SSI Spatial Sciences Institute
USC University of Southern California
USFS United States Forest Service
USGS United States Geological Survey
NOAA National Oceanic and Atmospheric Administration

x
Abstract
This study predicts the spreading pattern of an invasive plant species, Spartium junceum,
using Maxent, a type of Species Distribution Model (SDM). Species Distribution
Modeling estimates the relationship between species records at sites within a given study
area and the environmental or spatial characteristics of those sites. This study combines
environmental variables found at sites where species occurrence has been confirmed and
analyzes the results to predict future spreading patterns. A subset of occurrence data is
used for quality control with the intended purpose of validating the accuracy of the
model and its results. This study uses ArcGIS 10.6.1 and Maxent version 3.4.1 to
perform presence-only species distribution modeling of Spartium junceum from data
collected in 2011–2016 in the Angeles National Forest (ANF), which is managed by the
U.S. Forest Service. The primary study area is the ANF with an emphasis on the San
Gabriel Mountains which lie in the eastern portion of the national forest. The results will
increase the exposure of Maxent as a feasible, cost-effective species distribution model
that Federal land management agencies can incorporate into environmental analyses and
environmental impact studies that contribute to their land management plans. The
Angeles National Forest does not currently employ SDMs in its invasive species
management plan and Maxent modeling represents an option for studying species
distribution patterns where only presence data are available. Many studies use Maxent to
understand species distributions of flora and fauna, but few such efforts are updated
annually and used as a key indicator in a Federal Agency’s decision-making process in
their land management plans and actions.

1

Chapter 1 Introduction
Diverse topography, ecosystems, and flora and fauna characterize California’s
national forests. Native vegetation diversity is considered a key indicator of forest
health, but the spread of invasive species has complicated what interpreting that
diversity looks like and has thus become an important issue for land management
agencies such as the US Forest Service (USFS) and local communities that neighbor
national forests. The ability of invasive species to compete with native vegetation
for resources and alter the local ecosystem is a primary concern to the interested
parties.
The purpose of this research is to better understand the underlying factors that
contribute to the spread of Spartium junceum (Spanish Broom) in the Angeles National
Forest (ANF). Spartium junceum is a particularly mettlesome invasive species and is
very difficult to remove and therefore is emphasized as a species for removal and other
mitigation efforts by the USFS (Mullin et al. 2000). The operating premise of this
research is that Spartium junceum thrives and spreads under certain conditions and, by
identifying those conditions through species distribution modeling (SDM), accurate
estimates of the geographic regions to which they may spread as the weather conditions
under which they will thrive in the future evolve.
Some of the conditions used in Maxent are environmental while others are non-
environmental; both conditions will be referred to as variables in the remainder of the
thesis. The non-environmental variables that will be considered in this research are
linear distance to nearest water source, linear distance to recreation sites, linear distance
to local roads, linear distance to vegetation burn areas, and linear distance to the

2

wildland urban interface. The environmental variables that will be considered in this
research are biophysical and bioclimatic with the bioclimatic variables coming from the
worldclim dataset (www.worldclim.org). The biophysical variables are as follows:
elevation, slope, aspect, hill shade and visibility. The bioclimatic variables are as
follows: mean diurnal temperature range (mean of monthly, maximum-minimum
temperature) minimum temperature of the coldest month, annual precipitation,
precipitation of the driest month, precipitation seasonality, and precipitation of warmest
quarter (Qin et al., 2017) Datasets associated with each variable will be accessed in
Maxent and ArcGIS to perform predictive spatial analysis.
1.1 Motivation
Federal management of national public lands and the natural resources that make
up the diverse array of ecosystems in those lands is a major endeavor that requires a
large annual budget and workforce on the ground at the field level. With increasing
budget cuts to critical natural resource programs, it has become increasingly important
to be as efficient as possible with the workforce that is already in place and incorporate
new technologies in the land management process. The USFS’s 2018 discretionary
budget, for example, decreased by 16.5% while the amount of land management either
remained same or increased. The USFS discretionary budget for fiscal year 2017 was
$6.2 billion and the President’s approved budget for fiscal year 2018 was $4.7 billion
(USDA Forest Service, 2017). The combination of shrinking budgets and increasing
recreation pose a substantial challenge for the future of the agency and its land
management practices.

3

One of the most crucial aspects of doing applied scientific research lies in its
ability to affect the planet and the lives of those who live in it in a positive way and to
increase the capacity and efficiency for protecting and managing natural resources. That
is precisely the goal of this research. In this case, the science can benefit the USFS,
which is the second largest federal land management agency in the U.S. that is tasked
with caring for the largest portion of national public lands.
The USFS is responsible for managing the national forest system, including the
natural resources and recreation areas which lie within its boundaries. The USFS has
jurisdiction over 28.7% of the nation’s federally owned land and is responsible for all
the natural resource work that must be performed on that land (Vincent 2004). The ANF
is one of the smaller national forests in the U.S. at 668,887 acres (USDA Forest Service,
2012) but receives some of the highest recreational use. The USFS reported recreational
use on the ANF at 3,471,486 people during the 2000 calendar year (USDA Forest
Service, 2001). One of the biggest explanations for the recreational use per acre is the
proximity of the ANF to the second the largest metropolitan area in the U.S., Los
Angeles. The estimated population of Los Angeles County in 2017 was 10,105, 722
based on information provided by the U.S. Census Bureau (2017).
The ANF is facing a modern-day management crisis. As the population grows
and the size of forest stays the same, increasing levels of recreation use on the forest are
approaching capacity and are creating more natural resource work when the personnel
and resources available to do the work decline as budgets continue to shrink. One of the
biggest problems facing the USFS today is managing the increasing number of invasive
plant species found throughout the forest. This is a particularly challenging task because

4

recreation increases on the forest and unpermitted off-trail forest users act as vectors for
invasive plant seed dispersal. Invasive species compete with native species and, in the
process, alter the local ecosystems, which affects natural wildlife habitat, wildlife food
sources, recreational opportunities and buffer zones that are critical to the wildland-
urban interface (Stein et al., 2013).
The Natural Resources Department is an integral unit to the overall structure of
the forest and is responsible for surveying, documenting, reporting and treating invasive
plant populations on the ANF, including Spartium junceum. The Forest Service
Supervisor’s office in Arcadia, California is the headquarters for the entire ANF, and the
Natural Resources Department personnel.
The Natural Resources Department on the ANF uses GIS to analyze the data
they record in the field but could greatly benefit from SDM in their many projects and
programs. The lead biologist, botanist, and their subordinates use a combination of
Trimble GPS units and mobile smartphones and tablets to collect and document data on
the invasive species they study in the field. After collecting data in the field, each field
biologist and botanist use Esri’s ArcPad software for Trimble or Esri’s collector mobile
GIS app to digitize their data. The software they use is suitable for routine GIS field data
collection and standard analysis but can greatly benefit from more robust spatial analysis
tools such as SDM.
This study examines the spread of the invasive plant species Spartium junceum
in the ANF. The high levels of vehicle and foot traffic throughout the study area are the
result of its location in Los Angeles County and the presence of what is known as the

5

wildland-urban interface, which is the zone of transition “where houses and wildland
vegetation meet or intermingle” (Radeloff et al., 2018).
This project shows how the application of species distribution modeling can
serve as a powerful spatial analysis tool to support land management across areas that
range from small to large extents. The use of Maxent clarifies some of the variables that
influence the spread of invasive species to provide a better understanding of the
mechanisms that are most dominant in their distribution patterns.
Most of the data being used for this research has come internally from USFS but
also from other governmental agencies including: the U.S. Geological Survey (USGS),
and the National Oceanic and Atmospheric Administration (NOAA). All the
government agencies that are providing the data collect that data on an annual basis and
have large databases dedicated to storing and publishing that data once it is in a suitable
format for the public. The data provided for this work at the time it was gathered (2017)
was in a transactional phase and had yet to be made publicly available. Its use for this
thesis project was approved at the time of writing by ANF resource officer Jaime
Uyehara and collected and provided by Jason Martin.
The goal of this research is to use the Maxent, Esri’s ArcGIS software, and
different sets of unique environmental and non-environmental variables to better
understand the spatial distribution of Spartium junceum in the study area. In using more
robust and powerful spatial analysis, a predictive approach can be taken by the Forest
Service to better remove, monitor, and manage the invasive plant species problems
affecting the areas they manage.

6

1.2 Study Area
The ANF was chosen as the primary study area due to its relatively small extent
in comparison to other national forest and the high levels of recreation use found
throughout the forest (Figure 1).

Figure 1 Angeles National Forest
This study considers many different types of environmental variables at play in
the forest as well as the many recreational opportunities available to the public. The goal
is to portray how those opportunities are increasing the chances and likelihood that seeds
are transported and dispersed into far ranging areas where invasive species can establish
themselves and compete with native plants. The ANF provides access to numerous

7

hiking trails, off-highway vehicle (OHV) trails and access areas, scenic roadways,
campgrounds, day use picnic areas and stream and river access to the many visitors who
frequent the region during the hot summer months.
As a very highly frequented national forest, the ANF serves as prime location for
Maxent as the area provides a plethora of variables needed for the modeling itself and an
abundance of presence-based invasive plant species data which is the primary type of
data needed to run this species distribution model.
1.3 Surveyed Invasive Species
An invasive plant species is defined as a non-native (or alien) species to any
ecosystem that is under consideration. The ANF plays host to many different varieties of
invasive species some of which fall into its noxious weeds list as well. The noxious
weeds list indicates that a species is not only invasive but also harmful to the
environment or to wildlife in an environment and necessitates more extreme monitoring
and removal efforts than normal invasive plant species. This study focuses on one of
those major invasive species. Spartium junceum (Spanish Broom) is the target species of
this research and is one of highest priority invasive species in the ANF invasive species
and noxious weeds list. Spartium junceum is found in many areas throughout the forest
and notoriously difficult to remove. The difficulty in removing individual plants, the
resilience to approved chemical applicators and the longevity of seed life has led to
numerous annual exhaustive efforts to control population blooms. It has been estimated
that the seeds of Spartium junceum remain viable for 5 years and mature plants can
produce anywhere from 7,000-10,000 seeds per season (Zouhar, 2005).

8

This research focuses on the land contained within the administrative boundary
of the ANF with an emphasis on the eastern side of the forest where the San Gabriel
Mountains are located. The study area supports recreation in the form picnicking, day
use camping, hiking and OHV (Off highway Vehicle) use, which makes it an ideal study
area for this research.
1.4 Species Distribution Modeling and Maxent Explained
Species distribution modeling, also known as environmental niche modeling, is a
modeling technique that uses two different data types to estimate the potential locations a species
will likely inhabit in the future. The two different data types are integral in determining the type
of SDM that is most appropriate for individual cases studies; the two data types are presence-
only data and presence-absence data (Philips et al., 2004).
The required data types for a species distribution model are often determined on a case-
by-case basis but one of the main determining factors is the availability of data. In many
instances, insufficient records are available to support presence-absence species distribution
modeling and presence-only SDMs are more suitable (Philips et al., 2004). Furthermore, a key
issue with presence-absence is that a species may be declared absent from a landscape unit
simply because the species was not detected using the prescribed sampling methods. The effect
of this imperfect detection is that parameter estimates will be biased, and any modeling of the
data provides a description of the surveyor’s ability to find the species on the landscape, not
where the species is on the landscape (Mackenzie, 2005).
Maximum Entropy Modeling is a “species distribution model that uses presence-only
data to predict the distribution of a species based on the aforementioned theory of maximum
entropy” (Qin et al., 2017). One of the constraints of species distribution modeling is the

9

availability of data for a given study variable within a given study area. Maxent relies on large
datasets to be as accurate as possible (Philips et al., 2004). The more data that is available the
better the model runs as it allows for higher quality or more appropriate data to be selected and
used with the model, which can help produce more accurate predictions of a probability
distribution for species occurrence in each study area. Being able to obtain the necessary data is a
critical component of being able to use this type of model.
The availability of historical datasets is a limiting factor in presence-absence modeling,
as many of the necessary records are held in archives managed by museums or other archiving
entities and are not always easily accessible or readily available. Maxent is not as burdened by
this challenge as it relies on presence-only datasets (Philips et al., 2004). In this case the
presence-only data needed is readily available to the public in a spatial data clearinghouse and
internally via request from the USFS.
Maxent aims to predict a species’ distribution based on environmental predictors
(precipitation, temperature, etc.), background data and the number of individuals or samples in a
data set based on the underlying theory of maximum entropy (Philips et al., 2004). The theory of
maximum entropy relies heavily on the principle of being able to find what the best
approximation of a probability distribution is relative to the constraints on a given study variable
in a specific study area (Qin et al., 2017). Entropy itself can be looked at as a measure of how
much choice is involved in the selection of an event in each system. Combining the concepts of
entropy and species distribution modeling, Maxent essentially “agrees with everything that is
known but carefully avoids assuming anything that is not known when trying to find the optimal
probability distribution for a given study or study area (Qin et al., 2017). When using Maxent,
the study area of choice is divided into a grid and from that grid, Maxent extracts a sample of

10

background locations that it contrasts against the presence locations where a species (Spartium
junceum) is found and uses this with the environmental predictors to predict future areas of
presence (Merow 2013).
As is the case with presence-absence data, presence-only data is subject to sampling bias
(Qin et al., 2017). This occurs primarily with sample selection bias where areas that are
oversampled are given more consideration or weight than areas that are not. To account for this
sampling bias, careful consideration will be given to the choice of the overall study area and the
presence of multiple data points that fall near each to avoid spatial autocorrelation. Furthermore,
presence-only data from previous years will be used to validate occurrence of the target species.
The information resulting from this research will be useful in helping the USFS carry out their
land management practices and serve as a basis for future studies related to species distribution
as it relates to land management.
1.4.1. Multiple Maxent Models
Multiple Maxent models are run in this research in order to get a thorough understanding
of the factors that play the biggest role in the study area. In taking this step a more
comprehensive approach is used in understanding which variables play the biggest roles on their
own and which variables when combined paint the most accurate picture in predicting habitat
suitability. This is carried out by running Maxent multiple times with each unique set of
environmental variables (layers) to see which instance produces the highest statistical
significance and which individual variables within each unique set of layers are the most
important predictors of that significance. Lastly, Maxent is run a final time with a combination of
all the unique layers once they have been converted to the same spatial resolution. This final run
provides a complete and thorough use of all of the variables in all of the layers and produces

11

results that are compared to all other Maxent results which is discussed in chapter 4.
1.5 Thesis Structure
The next chapter discusses species distribution modeling, the different types of models
available for use, and various case studies where species distribution modeling has been used.
Chapter 3 details the procedures used to collect and process the data and the variables in Maxent.
Chapter 4 details the results from this study and provides insight into how the study can be used
by land management agencies such as the USFS. Chapter 5 discusses how this research can be
improved and other potential land management program areas where it can be employed.

12

Chapter 2 Related Work
National forests across the U.S. have been dealing with a major land management crisis
as a result of severe budget constraints, reoccurring furloughs and government shutdowns. The
impacts of these factors have had and continue to have major implications on the availability of
skilled personnel to tackle the diverse array of problems facing national forests today. The “2019
Forest Service budget for discretionary appropriations” was $4.77 billion, a decrease of $486
million from the FY 2018 budget in the annualized Continuing Resolution. It included $1.72
billion for the management of national forest system lands and $2.5 billion for wildland fire
management (U.S. Forest Service, 2018). With most of the upcoming FY 2019 budget being
allocated to wildfire management, other program areas must become more selective and efficient
in prioritizing their approach to land management.
The aim of this research is to use Maxent, to predict future spreading locations where
noxious and invasive plant species, Spartium junceum, may occur and use those results as a cost-
effective tool to help mitigate some of the land management challenges facing the USFS and
other agencies today. The purpose of this chapter is to examine and critically evaluate
information on Spartium junceum characteristics and habitat and to highlight several studies
where Environmental Niche Modeling (ENM), also referred to as SDM, was used for predictive
analysis of species like the test species in this research. The chapter first examines research and
reports relevant to the characteristics, habitat tendencies, and management practices for
controlling Spartium junceum infestations. The chapter then examines some of the different
types of SDM/ENM currently being used based on the data types required. Lastly, the chapter
discusses Maxent as the preferred method for SDM/ENM amongst the ecological community.

13

2.1 Spartium junceum and Invasive Species
Spartium junceum, also commonly known as Spanish broom, is an invasive plant with a
high-level pest rating based on the California Invasive Plant Council (CAL-IPC) pest rating
system. This rating means that “these species have severe ecological impacts on ecosystems,
plant and animal communities, and vegetation structure. Their reproductive biology and other
attributes are conducive to moderate to high rates of dispersal and establishment. These species
are usually widely distributed ecologically, both among and within ecosystems” (Warner et al.,
2003, 4). In addition, Spartium junceum is a noxious weed and is included in the California
Department of Food Agriculture’s noxious weed list. A noxious weed is a plant that has been
defined as such by law or regulation (CDFA, 2009). This is consistent with how the USFS
defines an invasive species and factors in the noxious weed designation.
According to the USFS, “invasive species are among the most significant environmental
and economic threats facing our Nation’s forest, grassland, and aquatic ecosystems. They
endanger native species and threaten ecosystem services and resources, including clean water,
recreational opportunities, sustained production of wood products, wildlife and grazing habitat,
and human health and safety” (USDA Forest Service, 2013). A species is invasive if it is non-
native to an ecosystem that is being studied (USDA Forest Service, 2013). This criterion applies
to this research and understanding the habitat characteristics is a critical component of
understanding the effects this invasive species has to ecosystems.
2.1.1. Spartium junceum History
Spartium junceum was initially introduced “into the California ornamental trade in 1848
in San Francisco. Beginning in the late 1930s, it was planted along mountain highways in
southern California” as a method to prevent soil erosion. Evidence of Spartium junceum

14

colonies can be seen along those same access routes and highways that run through the national
forest system today (Zouhar, 2005).
2.1.2. Spartium junceum Characteristics
Spartium junceum is a shrub that is 3 m tall with green stems and fragrant yellow pea-like
flowers that bloom from April to June. Spartium junceum have very deep branched taproots that
are difficult to remove and are associated with nitrogen fixing bacteria. The plant begins to
produce seeds when it reaches 2-3 years of age and is indicated by its pods turning brown. Each
pod can contain 10-18 seeds. The seeds can germinate readily without treatment and will
produce 7,000 to 10,000 seeds per plant per season (Zouhar, 2005). Furthermore, the pods and
the plants themselves can tolerate frost providing exceptional longevity and viability (CAL-IPC,
2004).
2.1.3. Spartium junceum Habitat Types and Plant Communities
Spartium junceum is a native of the southern Mediterranean region of Europe
and thus thrives in a Mediterranean climate like that found in southern California. A
Mediterranean climate is characterized by rainy winters and dry summers (University of
California, 2019).
Spartium junceum is most commonly found in disturbed areas such as along roadsides
where it was seeded in the early 1900s, but it has also spread to riparian and upland areas
throughout much of the southern California national forest system (Nickerman, 2009). There are
also other documented cases of Spanish Broom populations persisting in other disturbed areas
such as on eroding slope, and riverbanks. Spartium junceum has also managed to start invading
adjoining stands of chaparral that are near roadsides.

15

2.1.4. Impacts of Spartium junceum on ecosystems
Spartium junceum can rapidly colonize disturbed areas and develop thick shrub like
communities that can crowd out and prevent colonization by native chaparral plant species. It
also poses a particularly challenging problem during the fire season. Spartium junceum not only
serves as additional fuel during the fire season but it can survive if its roots are not burned.
Spartium junceum is also a nitrogen-fixing plant which means it can enrich the nitrogen
levels of the surrounding soils of the communities it inhabits. The ability of broom to fix
nitrogen affects the way in which the local nitrogen cycles behaves. This is unlikely to benefit
native plants who do not have the same nitrogen demands and may reduce species diversity in
ecosystems as a result (Zouhar, 2005).
2.1.5. Spartium junceum Management and Control Methods
Established infestations are difficult to eliminate due to the large and long-lived seed
banks that accumulate annually. Furthermore, the success of any control methods will vary based
on the site characteristics (topography, soil, and climate) where the species are found, the
condition and age of the stands or communities, and the human and technical resources available.
The main methods for trying to control Spanish Broom infestations are through prevention,
integrated management, chemical management, and physical/mechanical management.
2.1.5.1. Control Methods
There are the two primary control methods currently employed by the USFS for removal
of invasive plants species, integrated management, and physical/mechanical management.
Integrated management refers to an effective method of combining the use of a chainsaw
for larger mature plant shrubs or shears for smaller shrubs and the application of a chemical
herbicide such as a glyphosate applied as a 2-3% v/v foliar spray to the cut parts of the stem.

16

Physical/mechanical management refers to the human use of tools to physically remove both
young and mature plants. While young Spartium junceum can be removed by hand rather easily
more mature broom must be removed with the use of a weed wrench which can be an exhausting
process and is only effective for small broom infestations where the previous control method is
impractical or too expensive (Zouhar, 2005). Control methods represent a crucial aspect of
invasive species management alongside prevention.
2.1.5.2. Prevention
According to USFS documentation the most effective way of managing an invasive
species is by preventing the establishment and spread of that species. Some of the ways this is
accomplished is by limiting seed dispersal, minimizing soil disturbance, introducing more native
plants into areas where infestations have been known to occur and, removing current plants to
limit the spread of infestation.
While there are numerous control methods and prevention measures currently being used
for removal and mitigation efforts there is no predictive analysis currently being used for
Spartium junceum and this represents an opportunity for the incorporation or utilization of SDM
as a cost-effective way to anticipate future locations of infestation and prioritize management
activities.
2.2 Species Distribution Modeling
SDM is a machine learning technique that has become a popular analytical
framework for predicting species distributions based on species specific geo-located
occurrence information and the environmental variables that contribute to that species’
survival rate (Vaclavik and Meentemeyer, 2009, 3248). There are two primary types of
SDMs that are differentiated based on the types of data they require and thus cater to

17

certain types of research and analysis. The two primary types of data required to perform
any type of predictive modeling are presence-absence data and presence-only data. The
merits of each data type and the availability of each data type are the primary
determining factors for model selection in each community or industry field. The
following sub-section details species distribution modeling that rely on presence-absence
data and explores some of the different types that are commonly used.
2.2.1. Presence-Absence Data Modeling
Presence-absence species distribution modeling relies on presence-absence data,
which means it is heavily dependent on the availability of current presence and absence
data for the case subject and historical data for that same subject. Historical data can be
from prior studies or from prior field collection provided the data is in a similar or
usable format that can be converted into the format that is the same as that of the current
presence data being used and the same format used by a particular model of choice.
There is a third category known as pseudo-absence data modeling which relies
on modelers generating pseudo-absence data by sampling environmental conditions at
locations where the species is not recorded and hence absent (Vaclavik and
Meentemeyer, 2009, 3249). There is a risk of introducing false or negatives errors when
engaging in this process because no actual records of absence data exist but rather, they
are being created. In addition, the results from using pseudo-absence data are not as
accurate as reflected in the area under the curve receiver operating characteristic (AUC-
ROC) status that will be explained in the following paragraph. As a result of the type of
data available and the prevalence of true presence-absence data model types used by the
ecological community and suitable for comparison with presence-only data models,

18

pseudo-absence data modeling techniques were not considered. For this research three
types of presence-absence models were examined.
Three presence-absence species distribution models commonly used in the
ecological community are multiple logistic regression (MLR), generalized linear models
(GLM), and classification and regression trees /classification trees (CART/CT). The
primary means of evaluating how effective each presence-absence model is in predicting
potential future distributions of a case species is by comparing the results of each model.
This can be accomplished by using similar test species amongst the three models and
evaluating the results of each by using AUC-ROC, which offers a global performance
measure for classification problems based on various threshold settings (Pepe et al.,
2007, 928). The closer to 1.0 the AUC-ROC value is the more accurate the result is
since a value of 1.0 indicates perfect discrimination whereas a value of 0.50 indicates
that the model performed no better than random.
The three models are explored in more detail by referencing the work of Vaclavik and
Meentemeyer (2009) on invasive species distribution modeling and Bedia et al. (2011) on
predicting plant species distribution across an alpine rangeland in northern Spain as a means of
further explaining the merits of this type of modeling and for providing the background for
choosing presence-only data modeling for the actual research performed in this thesis. GLM and
CT models were evaluated using an extensive dataset on the distribution of the invasive forest
pathogen Phytophhthora ramorum in California and the AUC-ROC (Vaclavik and
Meentemeyer, 2009, 3248). MLR and a CART model were evaluated using presence-absence
data from 15 different plant species found on the Riofrio rangeland in northern Spain.

19

2.2.1.1. Multiple Logistic Regression
MLR is a parametric technique that relies on the assumption that the response of
a species to environmental variables is consistent across space and time. This technique
attempts to model a relationship between two or more independent variables and a
dependent variable by fitting a linear equation to the data. MLR is considered one of the
simplest and easiest SDMs to work with and is frequently used by ecologists who have
access to presence-absence data.
Bedia et al. (2011) describe how MLR performed compared to the other SDMs for
modeling plant distributions across northern Spain. MLR achieved mixed results when compared
with the other models in the study. MLR provided high AUC-ROC values for half of the 15
samples in the study and low values for the other half with the highest AUC-ROC value being
0.92 and the lowest being 0.64. The results based on presence-absence data for non-invasive
species data were somewhat inconsistent across the board for all the models. However, the MLR
models produced some of the poorest resolution results amongst all the models in the study and
were ultimately ruled out as an option for predicting future distributions of Spartium junceum
(Bedia et al., 2011).
2.2.1.2. Classification Trees/Classification and Regression Trees
CTs and CARTS offer “a non-parametric, data driven method that recursively
partitions data into homogeneous groups based on identification of a specific threshold
for each environmental predictor variable” (Vaclavik and Meentemeyer, 2009, 3252). In
this method a hierarchical decision tree of rules is created to split data into “present and
absent” categories. According to Bedia et al. (2011), CARTs are better than MLRs for
predicting the distribution of oak species in California. For most species in the study,

20

CART produced good predictive resolutions, but also low AUC-ROC values. The
predictive resolution is a value that quantifies the “deviation of the prediction from the
true species prevalence” (Bedia et al., 2011). The highest AUC-ROC value for a CART
was 0.89 and the lowest was 0.67. This is consistent with the results from Vaclavik and
Meentemeyer (2009) where CT produced an AUC-ROC of 0.89 with true presence-
absence data and 0.65 when using true presence-pseudo absence data created from
environmental sampling.
Both cases highlight research performed on native species distributions which
does not fit invasive species distribution patterns nor address the fact that invasive
species are not necessarily in equilibrium with their environments as is more typical
with native plant species. When using true presence-absence data, CART/CT models
produced some high AUC-ROC values but when using pseudo-absence data, the AUC-
ROC values dropped significantly and as a result were ultimately ruled out as an option
for predicting future distributions of Spartium junceum (Bedia et al., 2011).
2.2.1.3. Generalized Linear Models
A GLM “is an extension of common multiple regression that allows for the
modeling of non-normal response variables” (Vaclavik and Meentemeyer, 2009, 3252).
This means that GLM is well suited for SDMs focused on invasive species where there
is available presence-absence data with variables that may vary tremendously such as
more localized weather or climate factors that might be influenced by subtle topography
changes. This was the case with the test species Phytophhthora ramorum, which is the
invasive pathogen, studied by Vaclavik and Meentemeyer (2009). This sample species is
like Spartium junceum in that it is invasive but differs in that it is a pathogen and not a

21

plant species and because there was available absence data to use in the model. With
their data GLM is able produces binary responses that specify the relationship between a
dependent (or response) variable Y, and a set of predictor variables, X.
When using true presence-absence data GLM produced an AUC-ROC value of
0.90 which was the best out of the models tested including Maxent, which was the
model used in this research (Vaclavik and Meentemeyer, 2009, 3254). There is plenty
of evidence to support using CART/CT or GLM for invasive species distribution
modeling but neither meet the requirements necessary for studying the invasive species
in this research given the lack of absence data. There are no test cases of Spartium
junceum being used as sample data for analysis with GLM or any of the previously
mentioned models and that may be due to the lack of available absence data as well.
For a more in-depth explanation of how AUC-ROC is used to evaluate the statistical
significance of the results produced by each model and thus the effectiveness of each model, the
reader is referred Hanley and McNeil (1982) and Zou et al. (2007).
2.2.2. Presence-Only Data Modeling
Presence-only SDM relies on presence-only data, which means that it is not
hampered by a reliance on the availability of absence data but rather only needs presence
data on the test species being analyzed during the time a study is being conducted.
Absence data is a very selective data type and can be very difficult to acquire. In many
cases absence data records are archived by herbaria and are either not available to the
public for use or are only obtainable by formal request which may or may not be
granted. This may occur because researchers are unwilling to share their findings with
the public or other researchers or as a result of that data being too sensitive for

22

governmental or other reasons. Furthermore, absence data for a sample species may not
even exist and this is often the case with invasive plant species. The very nature of
invasive plant species makes it difficult to acquire their absence data because
management and mitigation efforts require eradication and removal rather than studying
patterns on a seasonal or annual basis.
One of the biggest differences between presence-absence modeling and
presence-only modeling is having to account for sampling bias and an often-encountered
lack of information on a species’ prevalence.
Three commonly used presence only SDMs used by the ecological community
are Ecological Niche Factor Analysis (ENFA), General Algorithm Rule-Set Prediction
(GARP), and Maximum Entropy (Maxent) (Pearson et al., 2007). Like the examples of
the presence-absence SDMs in Section 2.2.1, these three presence-only models were
evaluated on how effective each one is in predicting potential future distributions of a
case species and by comparing the results of each model. This was accomplished by
using similar test species with the three models and evaluating the results of each by
using the area under the AUC-ROC. The performance of the three models has been
explored in a more detail by Vaclavik and Meentemeyer (2009). Maxent and ENFA
have been evaluated by using the same invasive forest pathogen Phytophhthora
ramorum data as the presence-absence models and AUC-ROC (Vaclavik and
Meentemeyer, 2009, 3248). GARP and Maxent have also been evaluated using
presence-only data from two invasive plant species, a generalist (Bromus tectorum) and
a specialist (Tamarix chinensi). In this context, a generalist is one with characteristics or
traits that allow them to inhabit a broad range of habitats whereas a specialist is one with

23

traits that require specific habitats or specific ecosystem conditions for them to thrive.
Lastly, additional examples of Maxent are referenced in Section 2.2.2.3.
2.2.2.1. Ecological Niche Factor Analysis
ENFA is a model that compares the distribution of locations where a sample species was
identified to a reference set describing the whole area. It does this by computing variables that
explain a major part of the ecological distribution of the sample species. The model produces two
factors with biological significance: marginality and specialization. Marginality describes how
the sample species’ optimum conditions differ from the global mean of environmental conditions
in a study area. This allows for comparisons between the environmental variables that may be
found at the site location and study area and the variables found at the site location but not in the
study area. This helps to identify correlated variables and spatial autocorrelation, which measures
the correlation between the values of one variable and the values at those locations that are
closest to it. Specialization refers to factors that are sorted by the decreasing amount of explained
variance, which helps describe how species variance compares to global variance.
When using presence-only Phytophhthora ramorum data ENFA produced an AUC-ROC
value of 0.80 which was lower than any of the AUC-ROC values produced by presence-absence
models which used the same presence data but also incorporated true absence data as well.
Furthermore, ENFA had a poorer efficiency (commission/omission error rate) than Maxent, CT
and GLM and tended to overestimate species distributions (Vaclavik and Meentemeyer, 2009,
3255). ENFA provided a good test case and positive results for presence-only modeling but was
not as accurate as Maxent and does not have as many positive test case reviews as Maxent in the
ecological literature. Therefore, it was not used for predicting future distributions of Spartium
junceum in this thesis research project.

24

2.2.2.2. General Algorithm Rule Set Prediction
GARP models relate ecological characteristics of known species-specific
occurrence locations (point data in GIS) to points that have been randomly selected from
the remainder of the study region. This helps the model to develop a series of decision
rules that best summarize the factors that are associated with the species’ presence in a
study region (Adjemian et al., 2006, 94). GARP uses presence-only data in the form of
sample species occurrence data along with environmental data to enable the model to
determine or predict locations where the species should be found. The output results take
the form of multiple predictive distribution maps that can be used for comparison
amongst each other or in combination with one another to provide a comprehensive
picture of where a study sample species might be expected to be found. Like Maxent,
GARP performs very well when dealing with small sample sizes and provides a cost-
effective means of rapidly generating spatial data that can be analyzed in GIS.
When using the generalist species Bromus tectorum, GARP produced an AUC-
ROC value of 0.58 for training data and a value of 0.50 for validation data. When using
the specialist species Tamarix chinensi, GARP produced an AUC-ROC of 0.72 for
training data and a value of 0.830 for validation data (Evangelista et al., 2008, 811).
These values are not as high as any of the AUC-ROC values produced by presence-
absence models, but these are the result of a different data set and upon closer inspection
show how the choice of generalist or specialist species produces major differences in
overall model performance. In general, specialist species were easier to predict.
The type of data used by Evangelista et al. (2008) is more reflective of the
Spartium junceum data being used in this thesis research project as both species are
invasive plant species and Spartium junceum would be considered a specialist species.

25

Turning next to the results from Maxent, the results were lower for Bromus tectorum
and the same for Tamarix chinensi. The overall results for GARP were comparable to
Maxent but not as strong and therefore was not used for predicting future distributions of
Spartium junceum in this thesis research project.
2.2.2.3. Maximum Entropy Modeling
Maximum entropy modeling techniques use presence-only data to predict the
distribution of a species based on the theory of maximum entropy (Qin et al., 2017,
140). The principle of maximum entropy states that the most appropriate “probability
distribution to model a given set of data is the one with the highest degree of entropy
(i.e., that which is most spread out, or closest to uniform)” among all those that satisfy
the constraints of our prior knowledge (Philips et al., 2006). In information theory,
entropy refers to randomness or unpredictability. The idea behind this is that the
information that does not fall within the probability distribution created by the model
has maximum entropy (randomness or unpredictability) and thus those results are
discarded and those that fall within the probability distribution remain. The remaining
information represents the best possible description of the distribution of data that was
fed into the model prior to it being run (Kemp, 2012, 4). Maxent creates this probability
distribution based on environmental variables spread across the entire study area
(Pearson et al., 2007, 106). The modeling process in Maxent uses an iterative heuristic
approach where the first instance of the model being run represents the first solution that
is tested. Once the first solution is tested the results from that first test are permuted and
run over again and again, sometimes through 500 iterations, until the model is trained
with data that improves with each iteration. The data that does not improve the model

26

with each successive iteration is discarded until ultimately the model moves to a stable
solution that was better than the first one obtained.
Maxent offers many advantages but also a few drawbacks. In addition to
utilizing easier to obtain presence-only data, Maxent can run without needing large
sample datasets, it can also utilize both continuous and categorical data, and can
incorporate interactions between different variables. It also has efficient deterministic
algorithms that have been developed, and a probability distribution that has a concise
mathematical definition. Over-fitting can be avoided by using l1-regularization and
because dependence of the Maxent probability distribution on the distribution of
occurrence localities is explicit, there is the potential (in future work) to address the
issue of sampling bias formally” (Philips et al., 2006, 234). Maxent requires special
software in order to be run, it is relatively new statistical model in comparison to other
SDMs such as GLM and GARP and thus has less guidelines and fewer methods for
estimating the amount of error in a prediction. Furthermore, the amount of regularization
requires more study as does the model’s effectiveness in avoiding over-fitting compared
to other models. When considering the use of Maxent for this thesis research project it
was also important to consider another major factor in this type of modeling and that is
sampling bias.
One of the biggest differences between presence-absence and presence-only modeling is
having to account for sampling bias. Occurrence data are frequently biased, for example, data
that is closer to access routes will be better sampled and better documented and thus better
represented in the results of the model. When the bias is large, presence-only models mimic the
biased sampling distribution as much as they do the species distribution. This can be corrected by

27

having the background data reflect the same bias and thus essentially cancel it out. This is
accomplished by using as a background the set of occurrence data for an entire group of species
that may have been captured or observed using the same methods. In this research, this was
accomplished by randomly selecting background points, but this can also lead to situations where
samples are counted more than once. This problem can be mitigated by keeping track of which
sites have been sampled in order to not reselect sites that have already been sampled. Lastly,
another way to account for sampling bias is to keep track of which environmental variables have
been better sampled and where they have been sampled in order to weight the value of those
samples (Philips et al., 2008, 162). There is a large body of research performed using Maxent
that shows how accurate it is in predicating species distribution patterns with the results being
evaluated with the same AUC-ROC values as the models in the previous sections.
When using a generalist species such as Bromus tectorum, Maxent produced an AUC-
ROC value of 0.55 for training data and 0.50 for validation data. When using a specialist species
such as Tamarix chinensi, Maxent produced an AUC-ROC of 0.77 for training data and 0.830 for
validation data (Evangelista et al., 2008, 811). These AUC-ROC values are very similar to those
produced by GARP on the same dataset and indicate a strong predictive outcome. Maxent
produced strong results predicting distributions of the endangered conifer Thuja sutchuenensis in
southwestern China with an AUC-ROC value of 0.99. This study used sample data from a 3-year
long field survey in conjunction with global BioClim data (Qin et al., 2017, 142).
When using presence-only Phytophhthora ramorum data, Maxent produced an AUC-
ROC value of 0.85 which was lower than any of the AUC-ROC values produced by all but one
of presence-absence models that used the same data (ENFA). Lastly, when Maxent was used on
the following invasive species Common reed (Phragmites australis), musk thistle (Carduus

28

nutans), salt cedar (Tamarix), and Russian olive (Elaeagnus angustifolia) it produced positive
results. For Phragmites australis it produced an AUC-ROC of 0.98 for training data and 0.97 for
validation data. For Carduus nutans it produced an AUC-ROC of 0.978 for training data and
0.965 for validation data. For Tamarix it produced an AUC-ROC of 0.988 for training data and
0.983 for validation data. For Elaeagnus angustifolia it produced an AUC-ROC of 0.976 for
training data and 0.945 for validation data (Hoffman et al., 2008, 362-364).
The results from the last study which included multiple invasive plant species datasets
show how Maxent produced strong AUC-ROC results. These results also show that Maxent
performed well when used with both native and invasive species data. The combination of the
results and the similarities of the different data types to the data being used in this research effort
serve as an additional rationale for using Maxent in this thesis research project.
Maxent was ultimately chosen for this thesis research project because of the many
advantages of the software, the ability to successfully address the challenges of sampling bias,
the availability of presence-only data for Spartium junceum, and the positive AUC-ROC results
generated from multiple related research studies on native and non-native invasive species in the
last two decades.

29

Chapter 3 Methodology
The main purpose of this research is to evaluate how effective Maxent is as a tool for
predicting future locations and distribution patterns of the invasive plant species Spartium
junceum. The primary source of sample data came from the USFS, with a secondary source
coming from the Cal Flora/Jepson herbarium, which was used as a quality check. The third
source of data came from the WorldClim global dataset. The last source of data came from a
USGS digital elevation model that was then used to derive other layers in ArcMap. Section 3.1
outlines how this research was approached and designed. Section 3.2 details the different
datasets used in this research. Sections 3.2.1 and 3.2.2 discuss the different data sets in more
detail and how they were obtained. Lastly, Section 3.3 details how the data was processed and
formatted for use with Maxent and ArcGIS.
3.1 Research Approach and Design
The basis for this research revolves around the understanding that environmental
conditions at locations where a plant species is found and considered to be thriving is consistent
with the conditions that are expected to be found in other areas in the study extent where the
same species might be found at some point in the future. This fundamental notion was critical to
how this research was approached and designed. Multiple instances of the Maxent model are run
with the same biological sample data but with different environmental layers to test this
fundamental notion. Running the model with different environmental layers and comparing the
results allows for a more comprehensive dive into the many factors at play including climate,
topography and the scale at which the underlying environmental data is recorded and then used
in the model.

30

This thesis research project relies on the use of ArcGIS for Desktop version 10.6 and
Maxent 3.4.1 (Phillips et al., 2006). The Maxent software is available for free and can be
downloaded while ArcGIS for Desktop version 10.6 needs to be purchased. Both were required
for this thesis research project.
Maxent uses data in the form of raster and CSV files that have been converted from Excel
spreadsheets that contain latitude and longitude coordinates for the occurrence locations of the
sample data being used. Vector data can be used in Maxent, but this can be accomplished
indirectly. Vector data but must be converted into rasters in ArcMap before being used in
Maxent. This approach was used for the ANF features vector dataset.
Maxent uses categorical (biological samples) and continuous data (environmental and
proximity variables) as inputs in order to generate a summary html file which contains several
charts, images, and tables with information about AUC-ROC values, the variables that contribute
the most to the predicted outcome, and a series of jackknife tests.
The AUC-ROC results are particularly important because they indicate if the model
produces results that are random or are better than random, which reveals how much confidence
can be had in the results. The level of confidence is indicated by the AUC value, which ranges
from 0.0 to 1.0. A median AUC value of 0.5 indicates that the results from the model are no
better than random while a value of 1.0 indicates the highest level of confidence. A higher AUC
value in Maxent translates to higher confidence in Maxent’s ability to accurately predict the
species distribution patterns. In addition, there are two different sets of results produced by
Maxent in the output html file which explain the contributions each individual environmental
variable has on the outcome of the model: the analysis of variable contribution and three
jackknife tests.

31

The Jackknife of regularized training gain shows the training gain of each variable if it
was run in isolation and compares that to the training gain of all of other variables (Phillips et al.,
2006). This helps show which variables contribute the most individually while considering how
they contribute amongst all of the other variables in the dataset. The jackknife of test gain is
interpreted the same way as the jackknife of regularized training game but uses the test gain data
from Maxent as opposed to the training gain data. Lastly, the jackknife of AUC uses the same
jackknife test but with AUC values on the test data. The three jackknife tests combine to create
the values behind the analysis of variable contribution. The analysis of variable contribution
gives estimates of the relative contributions of each environmental variable used in Maxent and
shows their percent contribution to the overall results in the form of a table.
Maxent’s default settings produce an ASCII output file that shows the study area and
displays the predicted areas of habitat suitability using values that range from 0 to 1. Warmer
colors (reds, oranges, and yellows) show areas with better predicted conditions for habitat
suitability while cooler colors indicate the opposite (blues and greens). The dots/points that are
found scattered throughout the ASCII file represent the presence locations (sample data) that are
used when training the model. The ASCII file that is produced by Maxent must be imported into
ArcMap and converted from ASCII format back to its original raster format with the ASCII to
Raster tool so that further analysis of the results can be done.
3.2 Research Sample Data
There are two primary data types needed by Maxent to perform this research. The first
type is the sample biological species data. The second type is environmental data, which when
combined with the sample data is used to train Maxent. The environmental data can be further
broken up into three sub-groups: bioclimatic data, data derived from a USGS DEM, and ANF

32

feature vector data converted to raster data. Maxent will be run a total of six times using these
environmental layers in isolation at their original recorded spatial resolutions, in isolation at a
resampled resolution and lastly, in combination with one another. The results of each Maxent run
are discussed and compared in Chapter 4.
3.2.1. Sample Biological Data
The Spartium junceum sample data was obtained from the USFS. A geodatabase (GDB)
containing ANF vector data was downloaded as a zip file and then extracted into a local
workspace in ArcCatalog. Necessary data from the GDB in ArcCatalog for this thesis project
was filtered out by using a series of queries. The results of those queries were then exported into
a new project-specific file geodatabase for use in ArcMap and then Maxent.
The Spartium junceum sample data, which was in the project GDB, was in polygon
format and had be converted into a point format. The Create Random Points tool was used to
create a large set of random points in the study extent with unique x, y coordinates. Those points
were then filtered out with the Select by Location tool, which allowed for the selection of all of
the points that fell inside the polygon boundaries of the original Spartium junceum layer. Figures
2 and 3 show the Spartium junceum feature class shapefile before and after the Create Random
Points and S elect by Location tools were run.

33

Figure 2 Spartium junceum polygon data
Once the Spartium junceum polygon feature class shapefile was converted into a point
feature class shapefile and x, y coordinates were created, the data was then saved as a database
file (.dbf) in the same project GDB. That database file was converted into an Excel file where
only the necessary headers for species, longitude and latitude were kept. The Excel file was
saved as a comma separated values (.csv) file and was then ready for use in Maxent.

34

Figure 3 Spartium junceum point data.
3.2.2. Environmental Variables/Layers
The sample data used in Maxent must be combined with raster layers containing
information on the environmental conditions that are present at locations where the sample data
are found in order to train the model to predict future locations for the species of interest. For this
study those environmental raster layers come in the form of three groups. The first group
contains environmental layers related to climate. This study uses 19 bioclimatic variables from
the WorldClim version 2.0 dataset to create a comprehensive understanding of climate factors
The 19 bioclimatic variables used in this study reflect the most current data available (Table 1).

35

Table 1 WorldClim bioclimatic variables

Code Variable
BIO1 Annual Mean Temperature
BIO2 Mean Diurnal Range (Mean of monthly (max temp - min temp))
BIO3 Isothermality (BIO2/BIO7) (* 100)
BIO4 Temperature Seasonality (standard deviation *100)
BIO5 Max Temperature of Warmest Month
BIO6 Min Temperature of Coldest Month
BIO7 Temperature Annual Range (BIO5-BIO6)
BIO8 Mean Temperature of Wettest Quarter
BIO9 Mean Temperature of Driest Quarter
BIO10 Mean Temperature of Warmest Quarter
BIO11 Mean Temperature of Coldest Quarter
BIO12 Annual Precipitation
BIO13 Precipitation of Wettest Month
BIO14 Precipitation of Driest Month
BIO15 Precipitation Seasonality (Coefficient of Variation)
BIO16 Precipitation of Wettest Quarter
BIO17 Precipitation of Driest Quarter
BIO18 Precipitation of Warmest Quarter
BIO19 Precipitation of Coldest Quarter

The second group of environmental variables that is used in this thesis research project
are derived from a series of DEMs that were obtained from the USGS. These DEMs have a
spatial resolution of 1/3 arc second (90 m) whereas the 19 bioclimatic variables have a resolution
of 30 arc seconds (~1 km). The digital elevation model has a substantially higher spatial
resolution than the bioclimatic data and was used to derive a secondary set of environmental
variables using tools in ArcMap. Elevation, hill shade, aspect, slope and visibility rasters were

36

generated using the Surface tool kit that is stored within the Spatial Analyst Toolbox. Table 2
lists the DEM layers that were created in ArcMap and used in the third Maxent run.
Table 2 List of DEM Rasters
Name of feature class Type of feature class
DEM ANF Study Area (Elevation) Raster
DEM Hillshade Raster
DEM Slope Raster
DEM Aspect Raster
DEM Visibility Raster

The third group of environmental variables used in this thesis research project were
created by converting ANF features vector data into raster data by using the Euclidean distance
tool, which is a type of proximity tool found in the Spatial Analyst Toolbox. Proximity tools are
used to indicate the distance between presence data points (Spartium junceum sample data) and
select features (ANF feature data) that are believed to influence species distribution patterns.
Table 3 lists the ANF features data used to create the Euclidean distance rasters for use in the
second run of Maxent.
Table 3 List of feature classes used to create Euclidean distance rasters
Name of feature class Type of feature class
Angeles National Forest Service Roads Line
Angeles National Forest Recreation Facilities Point
Water Bodies Polygon
California Flow lines (Hydrography) Line
Angeles National Forest Burn Severity Dataset Polygon

37

3.2.3. Data Processing and Standardization
Once all of the presence and raster data were downloaded and/or created it was necessary
to standardize the spatial resolutions, projections and boundaries of these datasets. The presence-
only data was already converted into csv format and needed no further changes; however, the
raster data required substantial conversions. The first step was converting the rasters into the
same projection, which was WGS84 for this study. This was accomplished using the Project
Raster tool found in the Data Management Toolbox. The newly projected rasters were then
clipped to the study boundary using the Clip tool found in the Data Management Toolbox. The
clipped rasters are then converted into ASCII files, the required raster format for Maxent, by
using the Raster to ASCII tool found in the Conversion Toolbox. There is one extra step in the
case of the bioclimatic variables, which precedes the raster to ASCII step and that was to
resample the rasters into the same format as the DEM rasters. This was accomplished with the
Resample tool found in the Data Management Toolbox. This workflow is the same for all of the
raster files and can be run quickly using the Model Builder function in ArcMap. Figure 4 details
the conversion process from raster file to ASCII file in Model Builder.

38

Figure 4 Raster Conversion tools with Model Builder in ArcMap.
Once all of the data is converted into the necessary format it is placed into the appropriate
destination windows in Maxent. The Spartium junceum sample data is entered in the samples
window, study area versions of the BioClim, Euclidean distance and DEM rasters are entered in
the environmental layers window, and study extent versions of the BioClim, Euclidean distance
and DEM rasters are entered in the projection layers window. The study area versions of the
environmental layers reflect the areas where the actual sample data lie within the ANF boundary
and the study extent environmental layers reflect the entire area within the ANF boundary. All of
the environmental layers come from the same raster source files. The outputs of the model are
sent to a designated folder created by the user, which is set in the Maxent output directory
window. The ‘create response variables’, ‘make predictions’, and ‘do jackknife to measure

39

variable importance’ boxes are checked as well. Lastly, the ‘logistic’ output format was chosen
per the consensus recommendations from Young et al. (2011) and Philips (2017). Figure 5 shows
the first Maxent window with all of the sections filled out.

Figure 5 Maxent window with appropriate boxes checked
In the settings section of the Maxent window there are further parameters which could be
checked and/or selected. In the Basics tab (Figure 6) settings all of the boxes have been left as
they are but a random test percentage of 25 has been added to the window and the number 15 has
been added to the replicates window. The test percentage indicates that 25% of the sample data
will be used for training the model and the other 75% will be used as the actual data. The number
15 in the replicate window signifies that the model will run through 15 replications and the

40

results from the model will come from an average of all of the results. The replication run type
selected for this thesis research project is sub-sample. These parameters in Table 6 are used each
time the model is run.

Figure 6 Maxent Basic tab settings
All boxes in the advanced window in Figure 7 except for the ‘write output grids’ have
been left unchanged. The ‘write output grids’ box has been unchecked in order to speed up
model performance. This eliminates the need to produce results from each specific replication in
the output folder and instead produces and exports the average results from all the replications
into the folder. A value of 5,000 maximum iterations is used over the default value of 500. This
allows for a more accurate model as each replication runs through, 5,000 iterations to create the

41

best and most accurate model possible. There is no value selected for the default setting in the
‘apply threshold’ rule box but for this research, ‘minimum training presence’ is selected.

Figure 7 Maxent Advanced tab settings
Maxent is run six different times with each run using a unique set of individual
environmental variables or a unique combination of environmental variables. The results Maxent
produces are in the form of an HTML file, which is placed in the outputs folder. The HTML file
contains a graphic image that is created by the parameters set in the Maxent windows and
produced an ACSII file with habitat suitability information in the form of continuous values
ranging 0 to 1. These values can be changed into discrete values by using the Reclassify Tool

42

found in the Spatial Analyst Toolbox. The output ASCII file is imported into ArcMap and
converted back into a raster file using the ASCII to Raster tool. The raster is then classified into
an appropriate number of breaks using a 90% threshold of sensitivity. This 90% sensitivity
threshold can be found in the outputs folder in the maxentResults.csv file under the filed header
“10 percentile training presence logistic threshold” (Young et al., 2011). Three breaks are used
for this thesis research project to indicate low, medium and high levels of habitat suitability. The
results from all six Maxent runs are presented in Chapter 4.

43

Chapter 4 Results
The results from running Maxent with the sample data come in the form of
tables, graphic images and charts that explain which variables contribute to the model,
and the overall performance of the model based on the AUC. These results are important
because they allow for a well-informed evaluation of how well the model is performing
with the given sample data and environmental layers.
In this thesis research project Maxent was run numerous times with different
environmental layers and the same sample data. The idea behind using different
environmental layers in each run is to see if one set of layers is more indicative of
species presence compared to the other layers. This was accomplished by running
Maxent numerous times with each set of unique environmental layers and systematically
modifying the parameters based on the results. This process entails combining, or in
some cases, eliminating environmental layers that are deemed insignificant contributors
to the overall performance of the model and rerunning the model with those deemed
significant contributors; this information is found in the ‘Analysis of variable
contributions’ in the results section after Maxent is successfully run.
The results from running Maxent with each unique set of environmental layers
are compared to each other as a way of evaluating which has the most significant
influence on the presence of Spartium junceum and to see how well suited Maxent is as
an SDM. The following subsections report the Maxent results for each unique set of
environmental variables. The comparison of the results and the final evaluation of
Maxent are discussed Chapter 5.

44

4.1 Current Distribution of Spartium junceum in the Angeles National Forest
Figure 8 represents the current distribution pattern of Spartium junceum in the ANF as
collected by USFS natural resources staff and contracted partner, the Rancho Santa Ana Botanic
Garden. The forest encompasses 655,598 acres (1,024 square miles). The aggregate acreage of
all Spartium junceum samples collected is 3,163.23 acres or approximately 4.95 square miles.
The Spartium junceum point data reproduced in Figure 8 does not give an area value for the
samples but the original source of the Spartium junceum point data does. The source data is the
Spartium junceum polygon layer that was converted to point data in ArcMap using the Create
random points tool. The Spartium junceum polygon layer does have values for acreage and area
and that is what is used to calculate the aggregate values. Sections 4.2 through 4.7 detail the
results from running Maxent with different environmental layers and spatial resolutions.
4.2 Maxent results using BioClim Data at 1 km resolution
This section examines the results that come from using the 19 bioclimatic environmental
layers that come from worldclim.org at a maximum spatial resolution of 30 arc seconds (~1 km).
The parameters that are used for this set of environmental variables are the same as
those that are displayed in Figures 5-7 in Section 3.2.3 and the same parameters are
used for each ensuing set of environmental variables in the following subsections. The
parameters that were ultimately selected result from trial and error and the
recommendations from Young et al. (2011) and Philips (2017).
The results in Figure 9 indicate that the forest has roughly equal amounts of high,
medium and low suitability areas for Spartium junceum.

45

Figure 8 Spartium junceum point data in ANF

Figures 11-13 show the results of the jackknife test which indicate that
ascii_wc2_0_bio_30s_15 (precipitation seasonality) had the highest gain when run
independently for all three of the jackknife tests of variable importance, which means it has the
most useful information when run by itself.
Table 4 shows the variable contribution of each environmental layer in Maxent as a
percentage. Under normal circumstances the variables that contribute very little or contribute 0%
are taken out and the model is re-run but for this dataset which has such a large spatial resolution
it has been deemed unnecessary to do so.

46

Figure 9 Maxent suitability results using BioClim 1 km spatial resolution in ANF
The world BioClim environmental layers come in a maximum spatial resolution of 1 km

which is the equivalent of 0.386 mi. The entire aggregated area of all the Spartium junceum
sample data is 4.95 mi
2
. The spatial resolution of this data set is large in comparison to the study
area boundary and the aggregated value for all of the sample data and thus it is not the most
indicative of how much influence each variable has on the sample data set. In Section 4.3
Maxent is run with the BioClim variables after they have been resampled to a spatial resolution
of 10 m.

47

Figure 10 AUC-ROC Curve for Spartium junceum

Figure 11 Jackknife of regularized training gain for Spartium junceum

48

Figure 12 Jackknife for test gain for Spartium junceum

Figure 13 Jackknife of AUC for Spartium junceum

49

Table 4 Variable contributions

4.3 Maxent results using BioClim Data at 10 m resolution
The results in Figure 14 indicate that a very large portion of the forest can be deemed as
suitable habitat for Spartium junceum with most of the suitable habitat located in the eastern
portion of the San Gabriel Mountains.

50

Figure 14 Maxent results using BioClim 10 m resolution in ANF
Figure 15 shows the AUC-ROC value for the BioClim environmental layers after they
have been resampled to 10 m spatial resolution and have run through Maxent. The model has an
AUC of 0.664 and a standard deviation of 0.006 which means it is significant, but it is slightly
less significant than the BioClim data at its original spatial resolution.
Figures 16-18 show the results of the jackknife test which show that
ascii_wc2_0_bio_30s_15 (precipitation seasonality) has the highest gain when run independently
for all three jackknife tests of variable importance, which means it has the most useful
information when run by itself.

51

Figure 15 AUC-ROC for Spartium junceum

Figure 16 Jackknife regularized training gain for Spartium junceum

52

Figure 17 Jackknife of test gain for Spartium junceum

Figure 18 Jackknife of AUC for Spartium junceum

53

Table 5 shows the variable contribution for the BioClim data at 10 m spatial resolution.
The results are different than those of the BioClim data at its original spatial resolution. While
the variable with the highest contribution is still ascii_wc2_0_bio_30s_15, the percentage
increased with the resampled variables from 26.4 to 40.6%. In addition, the lowest contributing
variable changed from ascii_wc2_0_bio_30s_16 (precipitation of the wettest quarter) in the
original BioClim set to ascii_wc2_0_bio_30s_08 (mean temperature of the wettest quarter) in the
resampled set with both contributing variables contributing 0%.
Table 5 Variable contributions

54

4.4 Maxent results using DEM Raster layers at 10 m

resolution
The map reproduced in Figure 19 uses a DEM and other layers created from that DEM by
using the Spatial Analyst toolbox as mentioned in Section 3.2.1. The DEM file used for this
Maxent run came at a 10 m

spatial resolution and did not need to be resampled. The 10 m

spatial
resolution is well suited to the size of the study region. As seen in the two previous results maps,
most of the suitable habitat is found in the eastern portion of the mountain range. In addition,
there are signs of increased occurrence and more pronounced dispersion of highly suitable areas
in the study area when using the DEM variables in comparison to the BioClim variables.

Figure 19 Maxent results using DEM at 10 m resolution in the ANF

55

Figure 20 shows the AUC-ROC value for the DEM environmental layers at their
recorded 10 m resolution. The model has an AUC of 0.707 and a standard deviation of 0.007
which means it is significant but is slightly less significant than the AUC values (0.751)
produced when running the BioClim layers at their original 1 km spatial resolution in Maxent.

Figure 20 AUC-ROC for Spartium junceum

Figures 21-23 show the results of the jackknife test which shows that
ascii_dem_anf_visbility (visibility or Spartium junceum) had the highest gain when run
independently for all three jackknife tests for variable importance, which means it has the most
useful information when run by itself.

.

56

Figure 21 Jackknife of regularized training gain curve for Spartium junceum

Figure 22 Jackknife of regularized training for Spartium junceum

Figure 23 Jackknife of AUC for Spartium junceum

57

Table 6 shows the variable contribution for the DEM layers at a 10 m resolution. The
layer with the highest variable contribution is supported by the results of the jackknife tests
above.
Table 6 Variable Contribution

The DEM layers were recorded at the 10 m

scale and thus the spatial resolution is more
appropriate for the size of study area. The DEM layers benefit from being recorded at this scale
and do not have to be interpolated whereas the BioClim layers do. While the original set of
BioClim layers have a higher AUC-ROC value than the DEM layers it does not necessarily mean
that they are more accurate or are better indicators of suitable habitat for the sample species. This
will be discussed further in Chapter 5.
4.5 Maxent results using Euclidean Distance ANF features layers at 10 m
resolution
The map reproduced in Figure 24 uses environmental layers that were created by
converting USFS vector data into raster data using the Euclidean distance tool in ArcMap. The
rasters were created using the same 10 m spatial resolution as the DEM environmental layers in
order to maintain consistency. As is seen in the three previous maps, most of the suitable habitat
is found in the eastern portion of the mountain range although there is increased suitability in the

58

western portions of the mountain range as well. The results from running Maxent with this set of
environmental layers are the most different. The predicted suitability areas are very concentrated
along a visible network. This visible network is derived from the converted vector files with
USFS roads producing the most visible results in Figure 24. Furthermore, the presence of USFS
roads helps explain why there is a noticeable increase in suitability areas in the western portion
of the study area, and similarly helps explain the large patches of unsuitable areas in the eastern
portion of the study area.

Figure 24 Maxent results using Euclidean distance ANF features and 10 m resolution in ANF
Figure 25 shows the AUC-ROC value for the Euclidean distance proximity
environmental layers at their recorded 10 m spatial resolution. The model has an AUC of 0.692

59

and a standard deviation of 0.005 which means it is significant but less significant than the
Maxent results from the DEM and original BioClim layers.

Figure 25 AUC-ROC for Spartium junceum

Figures 26-28 show the results of the jackknife test in which the anf_roads (Euclidean
distance Angeles National Forest roads) had the highest gain when run independently for all
three jackknife tests for variable importance, which means it has the most useful information
when run by itself.
Table 7 shows the variable contribution for the Euclidean distance ANF features layers at
10 m resolution. The Euclidean distance layer of Spartium junceum to local forest service roads
has the highest variable contribution and is supported by the results of the jackknife tests above.

60

Figure 26 Jackknife of regularized training gain for Spartium junceum

Figure 27 Jackknife of test gain for Spartium junceum

Figure 28 Jackknife of AUC for Spartium junceum

61

Table 7 Variable contributions using Euclidean distance ANF features and 10 m
resolution

4.6 Maxent results using combined 10 m BioClim and DEM layers
The map reproduced in Figure 29 uses a combination of environmental layers from
Sections 4.3 and 4.4. The spatial resolution of both sets of environmental layers is 10 m and the
underlying values of the individual ASCII files are the same which allowed Maxent to run both
layers combined without any issues. As seen in the four previous maps, most of the suitable
habitat is again found in the eastern portion of the forest but there are some increased areas
concentrated at the eastern end of the western portion of the forest. The results from running
Maxent with these combined environmental layers shows higher levels of suitability located in
the south facing aspects in the forest and low suitability on the north facing aspects in the forest.
Figure 30 shows the AUC-ROC value for the BioClim and DEM 10 m layers. The model
has an AUC of 0.730 and a standard deviation of 0.005 which means it is significant and has a
higher AUC-ROC value than the BioClim 10 m, DEM 10 m, and Euclidean distance ANF
features layers when run on their own. However, it is still lower than the results produced with
the original BioClim 1 km layers when run on their own.

62

Figure 29 Maxent results using BioClim and DEM 10 m layers combined in ANF
Figures 31-33 show the results of the jackknife tests which show that
ascii_dem_anf_visbility has the highest gain for variable importance when run independently for
all three jackknife tests, which means it has the most useful information when run by itself. This
is the same result found in Section 4.4, which is a further indicator of how critical this variable is
to Spartium junceum habitat.
Table 8 shows the variable contribution of the BioClim and DEM 10 m layers. The
ascii_dem_anf_visbility has the highest variable contribution from these input layers. This is
consistent with the results of the jackknife tests above and like the results reported in Section 4.4.

63

Figure 30 AUC-ROC BioClim 10 m and DEM 10 m variables combined

The BioClim layers were resampled to the 10 m scale with the intention of combining
them with the DEM layers in this section in order to strengthen the number and quality of
environmental layers being used in Maxent. Resampling the BioClim layers is significant
because it allows these layers to be combined in a way that the BioClim 1 km layers cannot
while maintaining spatial resolution and integrity. The intention of this is to provide Maxent with
a set of variables that produces results from a set environmental layers that are spatially relevant
and have a spatial resolution to the study area. The combined BioClim and DEM 10 m layers
produce an AUC-ROC value that is higher than all but the BioClim layers at their original 1 km
spatial resolution. While the combined layers in this section may have a lower AUC-ROC value
than the BioClim 1 km layers this does not necessarily mean that they are less reliable or less
preferred. The significance of this argument will be discussed further in Chapter 5.

64

Figure 31 Jackknife of regularized training gain for Spartium junceum

65

Figure 32 Jackknife of test gain for Spartium junceum

66

Figure 33 Jackknife of AUC for Spartium junceum

67

Table 8 Variable contributions using combined BioClim and DEM 10 m layers

68

4.7 Maxent results using combined 10m BioClim, DEM, and Euclidean
Distance ANF Feature layers
The map reproduced in Figure 34 uses the environmental layers from Sections 4.3, 4.4
and 4.5. The spatial resolution of all three sets of environmental layers is 10 m and the
underlying values of the individual ASCII files are the same which allowed Maxent once again
to run all three layers combined without any issues. As seen in the results reproduced in the five
previous sections, most of the suitable habitat is found in the San Gabriel Mountains in the
eastern portion of the forest but some suitable habitat occurs in the western portion of the forest.
In addition, the results show higher levels of suitability located on south facing aspects of the
forest and low suitability on north facing aspects of the forest like the results reported in Section
4.6.
The results are better defined here than in any other section and that is the result of
having so many environmental layers available for Maxent to use. Figure 35 shows the AUC-
ROC value for the BioClim, DEM, and Euclidean Distance ANF features 10 m layers combined.
The model has an AUC of 0.762 and a standard deviation of 0.004 which means it is significant
and has the highest AUC-ROC value of all of the Maxent runs, indicating it the best predictor of
Spartium junceum habitat suitability.
Figures 36-38 show the results of the jackknife tests which show that anf_roads have the
highest gain when run independently for all three jackknife tests for variable importance,
meaning it has the most useful information by itself. The variable that decreases the gain the
most when it is omitted is ascii_dem_slope, which indicates it has the most information that is
not present in any of the other variables. The results from this set of jackknife tests indicate that

69

the proximity of Spartium junceum to forest service roads is a key indicator of habitat suitability
which is like the results in Section 4.5.
Table 9 shows the variable contributions of the combined BioClim, DEM, and Euclidean
Distance ANF features 10 m layers. The ascii_dem_anf_visbility has the highest variable
contribution at 31.9% which is only slightly higher than the variable contribution of anf_roads at
30.1 %. The next highest variable contribution (19.65%) is ascii_dem_slope and the remaining
variable contributions from the remainder of the layers are 2.4% or less.

Figure 34 Maxent results using the BioClim, DEM, and Euclidean distance ANF features 10 m
layers combined in ANF

70

Figure 35 AUC for Spartium junceum
The BioClim layers that were resampled at a 10 m scale as described in Section 4.6 have
been combined with the Euclidean distance ANF features layer, which were created at the same
10 m scale. The Euclidean distance ANF features were created by first converting the original
source vector data into raster data and then using the Euclidean distance tool with a 10 m spatial
resolution. This is accomplished with the intention of combining all the layers to provide the
most complete set of environmental layers possible for use in Maxent.
The combined BioClim, DEM, and Euclidean distance ANF features layers at a 10 m
resolution produce the highest AUC-ROC values of all of the Maxent results in this research and
that is possible because all of the layers have the same spatial resolution. The significance of this
result is discussed in the next chapter.

71

Figure 36 Jackknife of regularized training gain for Spartium junceum

72

Figure 37 Jackknife of test gain for Spartium junceum

73

Figure 38 Jackknife of AUC for Spartium junceum

74

Table 9 Variable contributions from the combined BioClim, DEM, and Euclidean Distance ANF
features layers at 10 m resolution

75

Chapter 5 Discussion and Conclusions
The Maxent results for each run indicate larger areas of suitable habitat than actual
surveys of documented locations where Spartium junceum has been found at this time. While the
results from each run are different, they do reflect an increased habitat suitability range where
Spartium junceum may spread or may already be colonizing. This reflects a trend that USFS
should monitor.
This research helps fill in some of the gaps as it pertains to Spartium junceum and
invasive species management options in the national forest system by adding to the body of
knowledge that is currently available. The results from this research show how SDMs can serve
as a very low-cost option for surveying and monitoring invasive species that is worth exploring
further in national forests.
The remainder of this chapter is divided into two sections. The first discusses the
strengths and weaknesses of the Maxent modeling approach, and the second discusses the
opportunities for future work
5.1 Strengths and Weaknesses
Maxent appears to be a capable modeling tool for evaluating invasive species habitat
suitability over large areas. The results produced by running the model with different
environmental variables showed that the model can handle layers from different sources and at
different resolutions without incurring any issues. The model was able to produce results that
were consistent in terms of the type of trends that were highlighted. The model’s ability to use
presence-only data is one of its biggest strengths because there is no real way to accumulate true
presence-absence data for this invasive species. There is no real way to survey areas where the
species is absent without documenting areas where it is present first and then removing the

76

species and continually monitoring where it pops up again as a control measure for comparison
which runs counter to the primary land management goal of removing the species altogether.
The study area for this research was selected at the forest level whereas much of the
research that was conducted previously has been at either the regional or the continental level.
The bulk of SDM research has used the bioclimatic data layers at a 1 km resolution (Elith e t al.,
2010) or layers recorded at a more regionally friendly 30 m resolution (e.g. Vaclavik and
Meentemyer, 2009). Maxent was able to handle all of the environmental layers at their different
resolutions and produce results that reflected an overall trend of increased habitat suitability. The
AUC value for each model run was over 0.60 which means that each was statistically significant
with results that are better than random (0.50 or less). There is some debate about how accurate
AUC is as a measure of statistical significance in SDMs but the variable results from the DEM
layers and the Euclidean Distance ANF features layers are consistent with what was expected to
be the most influential factors in predicting suitable habitat (i.e., proximity to forest service
roads, elevation, slope and visibility).
There were no real challenges in terms of obtaining the sample data from the Forest
Service as I was able to obtain the data directly while I was employed by them as an intern when
this thesis project research was launched. However, there were significant issues with the format
of the sample data, the introduction of uncertainty from using the Create Random Points tool in
ArcMap, the collection methods of the sample data, sampling bias, and the resolution of the
environmental variables from the BioClim dataset.
The sample data and x, y coordinates associated with them were in polygon form and
needed to be converted into point form in ArcMap using the Create random points tool. The
results from that step produced a large dataset that needed to be filtered by selecting only those

77

points that fell within the original recorded polygon locations and eliminating all those that did
not. Using the Create Random Points tool was a necessary step in order to get the sample data
into a usable format in Maxent but as a result almost certainly would have introduced uncertainty
into the data set. That is because the original source polygons included numerous 10 m grid cells
and therefore the random points produced by the tool may have acquired one of a variety of
values for one or more of the pertinent co-variates.
There were 1,285 records in the original Spartium junceum polygon file and 6,245,000
records after running the create random points tool. This large number was necessary to ensure
that enough point features fell within every polygon in the study area. That number was reduced
to 222,614 after selecting point features that fell within the Spartium junceum polygons from the
source feature class. The two important things to understand is that the number of records
substantially increased because of the need to have individual x, y coordinates for each
occurrence of Spartium junceum and that uncertainty was almost certainly introduced into the
dataset as a result. Each instance of a point feature that was assigned an x, y coordinate pair
corresponded to a location where Spartium junceum should be found but did not necessarily
reflect the amount of plants actually found at each location (introduced uncertainty) nor the
number of coordinate pairs that were required to define the entire polygon boundary from the
source file.
Data collection methods serve as a major limitation for Maxent because the sample data
need to have one x, y coordinate pair per record for the model to be run. The ANF does not
currently have a clear and established protocol for capturing presence data in the field that relies
on point sampling which limits the confidence in the accuracy and truth of the data. Given this
lack of confidence, we cannot be certain how many actual plants are found at an actual location

78

where a polygon boundary was created, nor can we be certain of the actual coordinate pair count
that would define the entire boundary for each polygon with sample data. Figure 39 illustrates
this point.

Figure 39 Spartium junceum polygon data and point data after conversion
There also existed a realistic and likely possibility that the spartium junceum data
included sampling bias due the increased likelihood that mapped instances where spartium
junceum was found was closer to Forest Service roads and thus more likely to be included in the
USFS dataset than those instances of spartium junceum that were further away from the Forest
Service road network.
There were some difficulties when working with the BioClim data due to the resolution at
which it was recorded, and the efforts needed to resample the data. There was significant loading
time required to build the pyramids needed to display the BioClim raster data in ArcMap and
significant processing time needed to clip the raster to the study boundary. This proved to be a
very limiting factor in terms of using the model because any changes or errors in the final
versions of the rasters necessitated reloading and reprocessing of the files, which required

79

substantial time. The time required increased significantly when the BioClim data was resampled
at a 10 m

spatial resolution as that required the tool to create substantially more pixels than were
in the original file.
Running Maxent with the higher resolution raster files takes substantial memory and
required more than the 524 MB that is preset in the Maxent bat File. The available RAM was
increased to 2,096 MB to allow Maxent to run using the larger raster files without any issues.
This is an important factor to consider when working with data on thumb or hard drives which
will be used on multiple computers or machines. The amount of RAM available will need to be
the same on all machines that run the software. Systems resources are a crucial factor when
working with Maxent and that is something the USFS will need to consider if they want to use
the software even if on a trial basis only. Furthermore, storage space is a critical issue that should
be considered. The outputs from running the model for this thesis project using 15 replications
consumed 80 GB of storage space. If larger files are used or more replications are needed or
desired this can become a significant factor that would need to be accounted for.
5.2 Interpretation of Maxent results for each set of layers
When looking at the results from running Maxent with each set of environmental layers it
is important to consider two major factors, the resolution at which the data was captured and the
overall trend of the results.
The overall trend from each Maxent run is the increase in the areas that can be deemed as
suitable habitat for Spartium junceum; however, there were significant differences in the actual
size of those areas predicted by the various runs. This can be attributed to the resolution at which
the data was recorded and the appropriateness of that data relative to the size of the study area.
The spatial resolution of the BioClim layers is 1 km, which is the equivalent of 247.11 acres in a

80

study area spanning 655,387 acres. The resolution does not seem to be as much of an issue when
looking at the entire forest, but it was when selection was focused on the area occupied by
Spartium junceum. At this resolution, it takes just only 12.8 cells of BioClim 1 km data to cover
the entire Spartium junceum habitat. The BioClim data were recorded at the global scale to track
global climate trends. The BioClim layers are better suited for predicting habitat suitability
across a larger study area such as the entire national forest system rather than one forest by itself.
To account for the problems caused by the spatial resolution the original BioClim layers
were resampled in ArcMap to 10 m to match the DEM layers which were recorded at that spatial
resolution. At this resolution, a 10 m

cell is the equivalent of 0.00247 acres which is well suited
to the thesis research given the total area occupied by Spartium junceum (3,168 acres). When
looking at the results from Figures 9 and 14 the predicted areas of habitat suitability are smoother
and more defined in the resampled set of layers than in the original raster layer and match the
habitat area in the eastern and western parts of the forest better than the original BioClim layers.
The DEM environmental layers seem to be the best suited for the study area and that is
most likely because they were initially recorded at a 10 m

spatial resolution. The study area is
relatively small, and the DEM layers are better suited to account for the sizes of both the study
area and the area occupied by Spartium junceum. The results from running the DEM layers with
Maxent reflect this as they show habitat suitability areas that are much more detailed than those
produced with the BioClim layers. The AUC value 0.707 shows strong statistical significance
and reflects how elevation, slope, aspect and visibility tend to influence fauna in various ways in
forest ecosystems. The AUC values for the BioClim variables and resampled BioClim variables
were 0.751 and 0.664, which offered justification for combining layers from those pair of
sources. It is important to emphasize that each cell or pixel in the original BioClim layers

81

accounts for 1/13 of the sample species occurrence area, which is misleading with such a small
study area and does raise questions about the efficacy of the 0.751 AUC value in this instance.
Maxent was run a fourth time with a set of vector data feature classes that were converted
into raster data using the Euclidean Distance tool in order to indicate the proximity of those
geographic features to each other. The reason for using this tool was to account for spatial
autocorrelation and to indicate the influence that the proximity of each layer has on the sample
data. The results in Figure 24 look very similar to the sample species locations shown in Figure
9. The suitable habitat area is smaller and more delineated in this instance and that can be
attributed to the underlying vector data being the most detailed and spatially relevant in terms of
the scale at which it was recorded. The AUC value when running Maxent with the Euclidean
Distance rasters was 0.692 which means it is statistically significant but not as significant as the
results generated using the DEM layers or the BioClim 1 km layers. This can be attributed to the
fact that the data in these layers were originally recorded as vector features and needed to be
converted to raster features and then Euclidean distances. This major factor was considered when
performing the final Maxent run with the combination of all three of the 10 m spatial resolution
layers.
The fifth Maxent run combined the BioClim and DEM 10 m layers. This was done
because both layers were either recorded or converted to 10 m spatial resolution and had AUC
values that were significant. The AUC value for this run was 0.730, which was higher than the
AUC values produced when the BioClim and DEM layers when run on their own. This AUC
value reflects a high degree of statistical significance indicating that the result from running
Maxent with these two data sets is significantly better than random. This can be attributed to the
strength of combining both layers together to provide a stronger understanding of the many

82

variables at play that influence habitat suitability. The variable contribution percentages are
critical indicators of which variables play the largest role in habitat suitability and hence
potential future occupancy. The two biggest contributors from each data set are in line with
some of the key variables found in the historical background information for this species
(Zouhar, 2005). Section 5.2.1 discusses the variable contributions in a little more depth.
The sixth and final Maxent run incorporates the BioClim 10 m layer, the DEM 10 m
layers, and the Euclidean Distance ANF features layers. This was done because all layers were
either recorded or converted to 10 m spatial resolution and had AUC values that were significant.
The AUC value for this run was 0.762, which was the highest AUC value produced in all of the
Maxent runs. This can be attributed to two primary factors. The first being that in combining the
three layers a robust final layer set was created which can better account for the many variables
and factors at play that influence habitat suitability. The second factor is that in combining the
three different layers some the inconsistencies of a given individual layer, such as a source being
a vector layer, or a layer being recorded at a different spatial resolution, can be accounted for by
combining the three sets of layers.
The AUC values were used as the primary means of evaluating how well a model
performed at predicting habitat suitability. The AUC is “a common metric for assessing the
predictive ability and hence utility of a habitat suitability model” (Glover-Kapfer (2015). The
AUC assesses whether model predictions are better than random by interpreting the value scores
on a 0 to 1 scale. AUC values with a score of 1 indicate the model predicts presence perfectly
and a value of 0.5 indicates model predictions are equivalent to random guesses (Philips et al.,
2006). Interpreting the results from the six Maxent runs indicates that all models are better than
random and the BioClim, DEM and ANF features 10 m combination layers present a good level

83

of performance in terms of habitat suitability. Although there is some criticism of using AUC as
the primary means of evaluating how well SDMs such as Maxent perform (e.g., Lobo, Jimenez-
Valverde & Real, 2008) it still “remains the most common means of measuring model
performance to date” (Glover-Kapfer, 2015)
5.2.1 Interpretation of Variable Results
The relevancy of the BioClim variables in this study needs to be carefully assessed. The
original BioClim variables were recorded at a scale that is too large to be considered a good
predictor of suitable habitat. This is because the resolution reflects large-scale climate trends
rather than localized weather factors, which are more appropriate for the size of the study area.
Resampling the BioClim data did help adjust the resolution and make the data more applicable to
the study region but did not necessarily represent localized weather or climate as well as it could
have had it been modeled at a 10 m resolution at the onset. It is worth mentioning that
“precipitation seasonality” had the highest variable contribution in Maxent for both the original
and resampled BioClim data. This could indicate that even though the data was resampled it is
likely that it still does not accurately depict the local climate or microclimate expected or
experienced in the study area.
The most reliable indicators of habitat suitability in this study are the DEM and Euclidean
Distance ANF features layers because they were recorded at a scale that is appropriate for the
study area. The largest variable contributors to the DEM layers are ascii_dem_anf_visbility at
56% and ascii_dem_slope_10 m at 39.1%. The largest variable contributors for the Euclidean
distance ANF features layer are anf_roads at 70% and and_recreation_facilities at 15.8%. For
the combined the BioClim and DEM 10 m layer data set, the two highest variable contributions
are ascii_dem_anf_visbility at 47% and ascii_dem_slope_10 m at 33.3%. The next highest

84

variable contribution came from ascii_wc2_0_bio_30s_15 (precipitation seasonality) at 3.7%.
This is a significant result because this was the same variable that had the highest variable
contribution in both BioClim Maxent runs. This is another indication that BioClim on its own
may not be a good indicator of habitat suitability at the local scale.
In the combined BioClim, DEM, and Euclidean Distance ANF features combined 10 m
layer dataset, the top three variable contributors are ascii_dem_anf_visbility at 31.9%, anf_roads
at 30.1%, and ascii_dem_slope_10 m at 19.6%. These three variables account for 81.6% of the
total variable contribution while the remaining 27 variables account for only 18.4%. The top
three variables come from the DEM and Euclidean distance ANF features layers while 19 of
remaining 27 variables come from the BioClim data. This is another strong indication that
despite the high AUC values found in Sections 4.2 and 4.3, the results are not as a strong an
indicator as one might initially think. This further supports the merits of creating a more
comprehensive variable dataset by combining the three 10 m layer data sets. The top three
variables can be considered the most critical elements that need to be present when assessing an
area for Spartium junceum habitat suitability. The remainder of the variables should be
considered but not with the same weight as the top three. Furthermore, while the BioClim
variables should not be relied on in this case, they do represent a key set of variables that could
and should be incorporated in future models if they are captured at finder resolutions than 1 km.
5.3 Opportunities for future research and Model improvement
The overall goal of this study was to test the feasibility of using Maxent as a tool
for aiding land management agencies such as the USFS in their invasive species
management plans. There is no real cost associated with using the software as it is free
to use and comes with tutorial information. This makes Maxent a great tool for land

85

management use but based on the results of this research there are a few more steps that
need to be taken before Maxent can be confidently used in everyday land management.
While this research did demonstrate that with presence-only data and some initial
reading and training, agencies can easily get started with Maxent, it also revealed some
issues that need to be addressed before confidently using Maxent as an everyday tool.
The USFS should work to establish clear protocols for capturing presence data in
the field that rely on point sampling to avoid some of the aforementioned problems.
Additionally, the USFS can ensure the enforcement of a standardized workflow for data
entry into invasive species data dictionaries. Furthermore, one or more environmental
variables focused on geology and/or soil should be added to Maxent to account for
spartium junceum’s characteristically large root system that enables it to penetrate
substrate which can promote long life and further spreading (invasion) to new habitat
areas. Lastly, resampling of the BioClim data represents another key area for
improvement.
Resampling of the BioClim data can be improved by incorporating the use of a
more precise resampling scheme that includes some form of interpolation such as block
kriging or by locating regional weather and/or climate data sets that were recorded at
finer resolutions. There are a large number of studies that now utilize the ANUCLIM
software (Xu and Hutchinson, 2013) to generate finer scale representations of climate
(e.g. solar radiation, precipitation, daily maximum temperature, isothermality), which
can be used to improve on the methods and results of this research.
There is no shortage of opportunities for future research with the abundance of
different plant and animal species found throughout the forest. Maxent is not only

86

applicable to invasive species management but also to other areas such as native species
management, forest health assessment and forecasting, vector born disease prediction,
and regional weather and climate forecasting (Antoine and Thuiller, 2005). The USFS
could benefit greatly from incorporating SDMs such as Maxent into their everyday land
management plans but this research only serves as a starting point for having those
conversations. The issues of sampling bias introduced uncertainty from using the Create
Random Points tool, data collection and data entry methods, incorporation of substrate
variables, and selecting appropriately scaled climate data would all need to be addressed
first before Maxent would be ready for everyday use.

REFERENCES

Adjemian, Jennifer C Z, Girvetz, Evan H, Beckett, Laurel, and Foley, Janet E. “Analysis of
Genetic Algorithm for Rule-Set Production (GARP) Modeling Approach for Predicting
Distributions of Fleas Implicated as Vectors of Plague, Yersinia Pestis, in
California.” Journal of medical entomology 43, no. 1 (January 2006): 93–103.

Bedia, J., J. Busqué, and J.M. Gutiérrez. “Predicting Plant Species Distribution across an Alpine
Rangeland in Northern Spain. A Comparison of Probabilistic Methods.” Applied
Vegetation Science14, no. 3 (2011): 415-32. Accessed Feburary 2, 2018.
Doi:10.1111/j.1654-109x.2011.01128.x.

CAL-IPC. Match 9, 2004. “Plant Assessment Form Spartium junceum.” Accessed December 13,
2017. https://www.cal-uipc.org/plants/profile/spartium-junceum-profile/

CDFA. n.d. “Encyclopedia: Weed Ratings” Accessed January 06, 2018.
https://www.cdfa.ca.gov/plant/IPC/encycloweedia/weedinfo/winfo_table-sciname.html

Elith, Jane, Steven J. Phillips, Trevor Hastie, Miroslav Dudík, Yung En Chee, and Colin
J. Yates. “A Statistical Explanation of MaxEnt for Ecologists.” Diversity and
Distributions17, no. 1 (2011): 43-57. Doi:10.1111/j.1472-4642.2010.00725.x.
Evangelista, Paul H., Sunil Kumar, Thomas J. Stohlgren, Catherine S. Jarnevich, Alycia
W. Crall, John B. Norman Iii, and David T. Barnett. “Modelling Invasion for a
Habitat Generalist and a Specialist Plant Species.” Diversity and Distributions
14, no. 5 (2008): 808-17. Doi:10.1111/j.1472-4642.2008.00486.x.
Glover-Kapfer, Paul. 2015. “A training manual for habitat suitability and connectivity
modeling using tigers (Panthera tigris) in Bhutan as example”.
10.13140/RG.2.2.34804.86409. Accessed August 26, 2019.
Hanley, J. A., and B. J. McNeil. “The Meaning and Use of the Area under a Receiver Operating
Characteristic (ROC) Curve.” Radiology143, no. 1 (1982): 29-36.
Doi:10.1148/radiology.143.1.7063747.

Hoffman, Justin D., Sunil Narumalani, Deepak R. Mishra, Paul Merani, and Robert G. Wilson.
“Predicting Potential Occurrence and Spread of Invasive Plant Species along the North
Platte River, Nebraska.” Invasive Plant Science and Management1, no. 04 (2008): 359-
67. Accessed January 12, 2019. Doi:10.1614/ipsm-07-048.

Kemp, Karen. 2012.“The Hawai`i Island Crop Probability Map: An Update of the Crop Growth
Parameters for the Hawai`i County Crop Model.” Accessed December 18, 2018.
https://spatial.usc.edu/wp-content/uploads/2017/09/CropProbabilityMap-Final-Report.pdf

Lobo, J., Jiménez ‐Valverde, A., & Real, R. (2008). AUC: a misleading measure of the
performance of predictive distribution models. Global Ecology and
Biogeography, 17(2), 145–151.Accessed June 12, 2018.
https://doi.org/10.1111/j.1466-8238.2007.00358.x

Mackenzie, Darryl I., and Royle, J. Andrew. “Designing Occupancy Studies: General Advice
and Allocating Survey Effort.” Journal of Applied Ecology 42, no. 6 (December 2005):
1105–1114.

Merow, Cory, Smith, Matthew J., and Silander, John A. “A Practical Guide to MaxEnt for
Modeling Species’ Distributions: What It Does, and Why Inputs and Settings
Matter.” Ecography 36, no. 10 (October 2013): 1058–1069.

Mullin, Barbara H., Lars W.J. Anderson, Joseph M. DiTomaso, Robert E. Eplee, and
Kurt D. Getsinger. 2000. Invasive Plant Species. Council for Agricultural
Science and Technology, no. 13. Accessed September 14, 2016.
http://www.iatp.org/files/Invasive_Plant_Species.html
Nickerman, Janet., and Joanna Clines. September 2009.
“Noxious and Invasive, non-native weeds Specialist Report and Noxious Weed Detection
Survey Plan.” Accessed December 13, 2017.
https://www.fs.usda.gov/Internet/FSE_DOCUMENTS/stelprdb5167066.pdf

Pearson, Richard G., Raxworthy, Christopher J., Nakamura, Miguel, and Townsend Peterson, A.
“ORIGINAL ARTICLE: Predicting Species Distributions from Small Numbers of
Occurrence Records: a Test Case Using Cryptic Geckos in Madagascar.” Journal of
Biogeography 34, no. 1 (January 2007): 102–117.

Pepe, Margaret S., Holly Janes, and Jessie Wen Gu. “Letter by Pepe Et Al Regarding
Article, “Use and Misuse of the Receiver Operating Characteristic Curve in Risk
Prediction”.” Circulation116, no. 6 (2007). Accessed October 23, 2018
doi:10.1161/circulationaha.107.709253.
Phillips, S. J. “A Brief Tutorial on Maxent.” 2017. Accessed on October 27, 2018.
http://biodiversityinformatics.amnh.org/open_source/maxent/.
Phillips, Steven J., and Dudík, Miroslav. “Modeling of Species Distributions with Maxent: New
Extensions and a Comprehensive Evaluation.” Ecography 31, no. 2 (April 2008): 161–
175.

Phillips, Steven J., Miroslav Dudík, and Robert E. Schapire. “A Maximum Entropy Approach to
Species Distribution Modeling.” Twenty-first International Conference on Machine
Learning – ICML 04, 2004. Doi:10.1145/1015330.1015412.

Phillips, Steven J., Anderson, Robert P., and Schapire, Robert E. “Maximum Entropy Modeling
of Species Geographic Distributions.” Ecological Modelling 190, no. 3 (2006): 231–259.

Qin, Aili, Bo Liu, Quanshui Guo, Rainer W. Bussmann, Fanqiang Ma, Zunji Jian, Gexi Xu, and
Shunxiang Pei. “Maxent Modeling for Predicting Impacts of Climate Change on the
Potential Distribution of Thuja Sutchuenensis Franch., an Extremely Endangered Conifer
from Southwestern China.” Global Ecology and Conservation10 (2017): 139-46.
Doi:10.1016/j.gecco.2017.02.004.

Radeloff, Volker C., David P. Helmers, H. Anu Kramer, Miranda H. Mockrin, Patricia M.
Alexandre, Avi Bar-Massada, Van Butsic, Todd J. Hawbaker, Sebastián Martinuzzi,
Alexandra D. Syphard, and Susan I. Stewart. “Rapid Growth of the US Wildland-urban
Interface Raises Wildfire Risk.” PNAS. March 27, 2018. Accessed October 13, 2018.
http://www.pnas.org/content/115/13/3314.

Stein, Susan M., James P. Menakis, Mary A. Carr, Sara J. Comas, Susan I. Stewart,
Helene Cleveland, Lincoln Bramwell, Volker C. Radeloff. “Wildfire, wildlands,
and people: understanding and preparing for wildfire in the wildland-urban
interface—a Forests on the Edge report. January 2013. Accessed September 30,
2017. https://www.fs.fed.us/openspace/fote/reports/GTR-299.pdf

Thuiller, Wilfried, Araújo, Miguel B., and Lavorel, Sandra. “Generalized Models Vs.
Classification Tree Analysis: Predicting Spatial Distributions of Plant Species at
Different Scales.” Journal of Vegetation Science 14, no. 5 (October 2003): 669–680.

University of California. n.d. “Mediterranean Climate.” UC Rangelands Archive. Accessed
January 11, 2019.
http://rangelandarchive.ucdavis.edu/Annual_Rangeland_Handbook/Mediterranean_Clima
te/.

US Census Bureau. 2017. “ACS Demographics and Housing Estimates”. Accessed September
28, 2017.
https://factfinder.census.gov/faces/tableservices/jsf/pages/productview.xhtml?src=CF

USDA Forest Service. February 2018. “FY 2019 Budget Justification.” Budget &
Performance. Accessed January 5, 2019.
https://www.fs.fed.us/sites/default/files/usfs-fy19-budget-justification.pdf

USDA Forest Service. May 2017. “Fiscal Year 2018 Budget Overview”. Accessed
October 6, 2017.
https://www.fs.fed.us/sites/default/files/usfs-fy18-budget-overview.pdf

USDA Forest Service. August 2013. “Forest Service National Strategic Framework for
Invasive Species Management.” Accessed January 10, 2019.
https://www.fs.fed.us/foresthealth/publications/Framework_for_Invasive_Specie
s_FS-1017.pdf

USDA Forest Service. n.d. “History & Culture.” Accessed October 13, 2018.
https://www.fs.usda.gov/main/angeles/learning/history-culture.

USDA Forest Service. January 2012. “Land Areas of the National Forest System”.
Accessed September 25, 2016.
http://www.fs.fed.us/land/staff/lar/LAR2011/LAR2011_Book_A5.pdf.

USDA Forest Service. 2001. “National Visitor Use Monitoring Results August 2001
USDA Forest Service Region 5.” Accessed September 25, 2016.
https://www.fs.fed.us/recreation/programs/nvum/reports/year1/R5_Angeles_final
.%09htm#_Toc522596885

Václavík, Tomáš, and Ross K. Meentemeyer. “Invasive Species Distribution Modeling
(iSDM): Are Absence Data and Dispersal Constraints Needed to Predict Actual
Distributions?” Ecological Modelling 220, no. 23 (2009): 3248-258.
doi:10.1016/j.ecolmodel.2009.08.013.

Vincent, Carol Hardy. 2004. “Federal Land Management Agencies: Background on
Land and Resources Management”. Accessed September 30, 2017.
http://nationalaglawcenter.org/wp-content/uploads/assets/crs/RL32393.pdf.

Warner, Peter J., Carla C. Bossard, Matthew L. Brooks, Joseph M. DiTomaso, John A. Hall, Ann
M. Howald, Douglas W. Johnson, John M. Randall, Cynthia L. Roye, Maria M. Ryan,
and Alison E. Stanton. “Criteria for Categorizing Invasive Non-Native Plants that
Threaten Wildlands.” 2003. Accessed January 10, 2019.
1.
Xu, Tingbao, Hutchinson, Michael F. New developments and applications in the ANUCLIM
spatial climatic and bioclimatic modelling package. Environmental Modelling and
Software. 2013;40(C):267-279. doi:10.1016/j.envsoft.2012.10.003

Young, Nick, Lane Carter, Paul Evangelista. “A MaxEnt Model v3.3.3e Tutorial (ArcGIS v10).”
2011. Accessed February 17, 2019.
http://ibis.colostate.edu/webcontent/ws/coloradoview/tutorialsdownloads/a_maxent_mod
el_v7.pdf

Zou, Kelly H., A. James O’Malley, and Laura Mauri. "Receiver-Operating Characteristic
Analysis for Evaluating Diagnostic Tests and Predictive Models." Circulation115, no. 5
(2007): 54-57. Accessed January 10, 2019. doi:10.1161/circulationaha.105.594929.

Zouhar, Kris. "Fire Effects Information System (FEIS).” 2005.
Accessed October 12, 2018.
https://www.fs.fed.us/database/feis/plants/shrub/spajun/all.html.

Appendix A: List of Layers
BioClim 1 km Raster Layers
Short
Name
Description Projection Source URL
Spatial
Resolution
BIO1
Annual Mean
Temperature
WGS 1984 https://www.worldclim.org/bioclim 1km
BIO2
Mean Diurnal Range
(Mean of monthly
(max temp - min
temp))
WGS 1984 https://www.worldclim.org/bioclim 1km
BIO3
Isothermality
(BIO2/BIO7) (* 100)
WGS 1984 https://www.worldclim.org/bioclim 1km
BIO4
Temperature
Seasonality (standard
deviation *100)
WGS 1984 https://www.worldclim.org/bioclim 1km
BIO5
Max Temperature of
Warmest Month
WGS 1984 https://www.worldclim.org/bioclim 1km
BIO6
Min Temperature of
Coldest Month
WGS 1984 https://www.worldclim.org/bioclim 1km
BIO7
Temperature Annual
Range (BIO5-BIO6)
WGS 1984 https://www.worldclim.org/bioclim 1km
BIO8
Mean Temperature of
Wettest Quarter
WGS 1984 https://www.worldclim.org/bioclim 1km
BIO9
Mean Temperature of
Driest Quarter
WGS 1984 https://www.worldclim.org/bioclim 1km
BIO10
Mean Temperature of
Warmest Quarter
WGS 1984 https://www.worldclim.org/bioclim 1km
BIO11
Mean Temperature of
Coldest Quarter
WGS 1984 https://www.worldclim.org/bioclim 1km
BIO12 Annual Precipitation WGS 1984 https://www.worldclim.org/bioclim 1km
BIO13
Precipitation of
Wettest Month
WGS 1984 https://www.worldclim.org/bioclim 1km
BIO14
Precipitation of
Driest Month
WGS 1984 https://www.worldclim.org/bioclim 1km
BIO15
Precipitation
Seasonality
(Coefficient of
Variation)
WGS 1984 https://www.worldclim.org/bioclim 1km
BIO16
Precipitation of
Wettest Quarter
WGS 1984 https://www.worldclim.org/bioclim 1km
BIO17
Precipitation of
Driest Quarter
WGS 1984 https://www.worldclim.org/bioclim 1km
BIO18
Precipitation of
Warmest Quarter
WGS 1984 https://www.worldclim.org/bioclim 1km
BIO19
Precipitation of
Coldest Quarter
WGS 1984 https://www.worldclim.org/bioclim 1km

BioClim 10 m Raster Layers
Short
Name
Description Projection Source URL
Spatial
Resolution
BIO1
Annual Mean
Temperature
WGS 1984 https://www.worldclim.org/bioclim 10 m
BIO2
Mean Diurnal Range
(Mean of monthly
(max temp - min
temp))
WGS 1984 https://www.worldclim.org/bioclim 10 m
BIO3
Isothermality
(BIO2/BIO7) (* 100)
WGS 1984 https://www.worldclim.org/bioclim 10 m
BIO4
Temperature
Seasonality (standard
deviation *100)
WGS 1984 https://www.worldclim.org/bioclim 10 m
BIO5
Max Temperature of
Warmest Month
WGS 1984 https://www.worldclim.org/bioclim 10 m
BIO6
Min Temperature of
Coldest Month
WGS 1984 https://www.worldclim.org/bioclim 10 m
BIO7
Temperature Annual
Range (BIO5-BIO6)
WGS 1984 https://www.worldclim.org/bioclim 10 m
BIO8
Mean Temperature of
Wettest Quarter
WGS 1984 https://www.worldclim.org/bioclim 10 m
BIO9
Mean Temperature of
Driest Quarter
WGS 1984 https://www.worldclim.org/bioclim 10 m
BIO10
Mean Temperature of
Warmest Quarter
WGS 1984 https://www.worldclim.org/bioclim 10 m
BIO11
Mean Temperature of
Coldest Quarter
WGS 1984 https://www.worldclim.org/bioclim 10 m
BIO12 Annual Precipitation WGS 1984 https://www.worldclim.org/bioclim 10 m
BIO13
Precipitation of
Wettest Month
WGS 1984 https://www.worldclim.org/bioclim 10 m
BIO14
Precipitation of
Driest Month
WGS 1984 https://www.worldclim.org/bioclim 10 m
BIO15
Precipitation
Seasonality
(Coefficient of
Variation)
WGS 1984 https://www.worldclim.org/bioclim 10 m
BIO16
Precipitation of
Wettest Quarter
WGS 1984 https://www.worldclim.org/bioclim 10 m
BIO17
Precipitation of
Driest Quarter
WGS 1984 https://www.worldclim.org/bioclim 10 m
BIO18
Precipitation of
Warmest Quarter
WGS 1984 https://www.worldclim.org/bioclim 10 m
BIO19
Precipitation of
Coldest Quarter
WGS 1984 https://www.worldclim.org/bioclim 10 m

Angeles National Forest Digital Elevation Model Raster Layers
Short
Name
Description Projection Source Source URL
Spatial
Resolution
DEM
ANF
Study
Angeles National
Forest Elevation Data
WGS 1984
United States
Geological
Survey
https://viewer.
nationalmap.go
v/basic/
10m
DEM
ANF
Visibility
Angeles National
Forest Spanish
Broom Visibility data
derived from DEM
and Spanish Broom
Sample Points
WGS 1984
Created using
Arc Map from
source DEM file
via the United
States
Geological
Survey
N/A 10m
DEM
ANF
Aspect
Angeles National
Forest Aspect/
downslope direction
of the maximum rate
of change derived
from DEM file
WGS 1984
Created using
Arc Map from
source DEM file
via the United
States
Geological
Survey
N/A 10m
DEM
ANF Hill
shade
Angeles National
Forest Hill
shade/shaded relief
derived from DEM
file considering the
illumination source
angle and shadows
WGS 1984
Created using
Arc Map from
source DEM file
via the United
States
Geological
Survey
N/A 10m
DEM
ANF
Slope
Angeles National
Forest Slope/
gradient derived
from DEM
WGS 1984
Created using
Arc Map from
source DEM file
via the United
States
Geological
Survey
N/A 10m

Angeles National Forest features converted to Rasters from Vector Layers
Short
Name
Description Projection Source
Spatial
Resolution
ANF CA
Flow lines
California Hydrography
Data clipped to Angeles
National Forest Boundary
WGS 1984
Initial internal data request
via Jason Martin, GIS Intern
USFS. Data now currently
available on USFS
Geospatial Data
Clearinghouse website.
10m
ANF
Recreation
Facilities
Designated recreation
facilities/sites located
through the Angeles
National Forest
WGS 1984
Initial internal data request
via Jason Martin, GIS Intern
USFS. Data now currently
available on USFS
Geospatial Data
Clearinghouse website.
10m
ANF Roads
Angeles National Forest
Service roads managed
by the US Forest Service
WGS 1984
Initial internal data request
via Jason Martin, GIS Intern
USFS. Data now currently
available on USFS
Geospatial Data
Clearinghouse website.
10m
ANF
Vegetation
Burn
Severity
Areas burned in the
Angeles National Forest
by wildfire including the
year of the fire and the
name of the fire.
WGS 1984
Initial internal data request
via Jason Martin, GIS Intern
USFS. Data now currently
available on USFS
Geospatial Data
Clearinghouse website.
10m
ANF Water
Bodies
Standing water bodies
(Lakes, Small Lakes,
ponds, etc...) in the
Angeles National Forest.
WGS 1984
Initial internal data request
via Jason Martin, GIS Intern
USFS. Data now currently
available on USFS
Geospatial Data
Clearinghouse website.
10m
ANF
Wildland
Urban
Intermix
Buffered feature class
detailing areas where
Wildland habitat, flora
and fauna mix with
Urban interface and
human interaction.
WGS 1984
Initial internal data request
via Jason Martin, GIS Intern
USFS. Data now currently
available on USFS
Geospatial Data
Clearinghouse website.
10m

Angeles National Forest Vector layers list
Short
Name
Description Projection Source
Source
URL
ANF CA
Flow lines
California
Hydrography Data
clipped to Angeles
National Forest
Boundary
WGS 1984
Initial internal data request via
Jason Martin, GIS Intern USFS.
Data now currently available on
USFS Geospatial Data
Clearinghouse website.
https://data.
fs.usda.gov/
geodata/
ANF
Recreation
Facilities
Designated recreation
facilities/sites located
through the Angeles
National Forest
WGS 1984
Initial internal data request via
Jason Martin, GIS Intern USFS.
Data now currently available on
USFS Geospatial Data
Clearinghouse website.
https://data.
fs.usda.gov/
geodata/
ANF Roads
Angeles National
Forest Service roads
managed by the US
Forest Service
WGS 1984
Initial internal data request via
Jason Martin, GIS Intern USFS.
Data now currently available on
USFS Geospatial Data
Clearinghouse website.
https://data.
fs.usda.gov/
geodata/
ANF
Vegetation
Burn
Severity
Areas burned in the
Angeles National
Forest by wildfire
including the year of
the fire and the name
of the fire.
WGS 1984
Initial internal data request via
Jason Martin, GIS Intern USFS.
Data now currently available on
USFS Geospatial Data
Clearinghouse website.
https://data.
fs.usda.gov/
geodata/
ANF Water
Bodies
Standing water bodies
(Lakes, Small Lakes,
ponds, etc...) in the
Angeles National
Forest.
WGS 1984
Initial internal data request via
Jason Martin, GIS Intern USFS.
Data now currently available on
USFS Geospatial Data
Clearinghouse website.
https://data.
fs.usda.gov/
geodata/
ANF
Wildland
Urban
Intermix
Buffered feature class
detailing areas where
Wildland habitat, flora
and fauna mix with
Urban interface and
human interaction.
WGS 1984
Initial internal data request via
Jason Martin, GIS Intern USFS.
Data now currently available on
USFS Geospatial Data
Clearinghouse website.
https://data.
fs.usda.gov/
geodata/

Angeles National Forest Vector layers list (cont.)
Short
Name
Description Projection Source
Source
URL
Spanish
Broom
Sample
Points

Spanish Broom
features class created
in ArcMap using
random point
generating tool and
filtered to .5 meters
distance of actual
locations of Spanish
Broom from Spanish
Broom polygon file

WGS 1984
Initial internal data request via
Jason Martin, GIS Intern USFS.
Data now currently available on
USFS Geospatial Data
Clearinghouse website.

https://data.
fs.usda.gov/
geodata/
Spanish
Broom
Polygon

Spanish Broom
invasive species data
in the form of a
polygon feature class
resulting from field
crew collection

WGS 1984
Initial internal data request via
Jason Martin, GIS Intern USFS.
Data now currently available on
USFS Geospatial Data
Clearinghouse website.
https://data.
fs.usda.gov/
geodata/
ANF
Boundary

Administrative
Boundary of the
Angeles National
Forest as administered
by the US Forest
Service

WGS 1984
Initial internal data request via
Jason Martin, GIS Intern USFS.
Data now currently available on
USFS Geospatial Data
Clearinghouse website.
https://data.
fs.usda.gov/
geodata/

Abstract (if available)

Linked assets

University of Southern California Dissertations and Theses

Conceptually similar

PDF

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

PDF

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

PDF

Predicting Hydromantes shastae occurrences in Shasta County, California

PDF

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

PDF

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

PDF

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

PDF

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

PDF

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

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

Spatial analysis of human activities and wildfires in the Willamette National Forest

PDF

Mixed forest image classification of paper birch: using AVIRIS bandwidths ranging from 530 to 745 nm

PDF

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

Asset Metadata

Creator Martin, Jason Isaiah (author)

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

School College of Letters, Arts and Sciences

Degree Master of Science

Degree Program Geographic Information Science and Technology

Publication Date 10/30/2019

Defense Date 10/30/2019

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

Tag Angeles National Forest,area under the curve receiver operating characteristic,California Invasive Plant Council,classification and regression trees,classification trees,digital elevation model,Ecological Niche Factor Analysis,environmental niche modeling,generalized linear model,Genetic Algorithm for Rule-Set Prediction,maximum entropy,multiple logistic regression,National Oceanic and Atmospheric Administration,OAI-PMH Harvest,off-highway vehicle,species distribution modeling,United States Forest Service,United States Geological Survey

Language English

Contributor Electronically uploaded by the author (provenance)

Advisor Wilson, John (committee chair), Loyola, Laura (committee member), Wu, An-Min (committee member)

Creator Email jimartin@usc.edu,jimartin016@gmail.com

Permanent Link (DOI) https://doi.org/10.25549/usctheses-c89-229643

Unique identifier UC11673729

Identifier etd-MartinJaso-7894.pdf (filename),usctheses-c89-229643 (legacy record id)

Legacy Identifier etd-MartinJaso-7894.pdf

Dmrecord 229643

Document Type Thesis

Rights Martin, Jason Isaiah

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