University of Southern California Dissertations and Theses

Distribution of Sonoran pronghorn (Antilocapra americana sonoriensis) on an active Air Force tactical range

(USC Thesis Other)

Distribution of Sonoran pronghorn (Antilocapra americana sonoriensis) on an active Air Force tactical range

PDF

Download a page range

Download transcript

Copy asset link

Request this asset

Transcript (if available)

Content DISTRIBUTION OF SONORAN PRONGHORN (ANTILOCAPRA AMERICANA
SONORIENSIS) ON AN ACTIVE AIR FORCE TACTICAL RANGE

by

Samuel C. Price

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)

May 2015

Copyright 2015 Samuel C. Price

ii

DEDICATION
I dedicate this document to my parents for their continuous encouragement throughout my
education, and to my fiancé for her sacrifice of time, and her dealing with a grouchy and tired
me.

ii

iii

ACKNOWLEDGMENTS
I thank my supervisor Nicole Tautfest for helping and recommending ideas while planning the
early stages of this project. I also thank those up at Luke AFB who agreed to let me use these data
for my thesis, specifically Aaron Alvidrez, whom I was in contact with. Much thanks as well to
my advisor, Dr. Travis Longcore, for being patient with my ever-changing schedule.

iii

iv

TABLE OF CONTENTS
DEDICATION ii
ACKNOWLEDGMENTS iii
LIST OF TABLES vi
LIST OF FIGURES vii
LIST OF ABBREVIATIONS x
ABSTRACT xi
CHAPTER 1: INTRODUCTION 1
CHAPTER 2: RELATED WORK 4
2.1 Sonoran Pronghorn 4
2.2 Areas of Intensive use by Sonoran Pronghorn 7
2.2.1 Kernel Density Estimation 7
2.2.2 Hotspots 9
2.3 Maximum Entropy Modelling 10
CHAPTER 3: METHODS 13
3.1 Study Area 13
3.2 Data Collection 14
3.3 Data Analysis 16
3.3.1 Hotspots 17
3.3.2 Maximum Entropy Modelling 18
3.3.2.1 Layer Preparation 18
3.3.2.2 Maxent Settings 21
iv

v

CHAPTER 4: RESULTS 27
4.1 Hotspots 28
4.2 Maximum Entropy Modelling 39
CHAPTER 5: CONCLUSION AND DISCUSSION 48
REFERENCES 52
APPENDIX A: HOTSPOT MAPS 59
APPENDIX B: SEASONAL HOTSPOT DISTANCE TABLE 75

v

vi

LIST OF TABLES
Table 1 Annual SPH Population Estimates and Maximum Individuals per Range 4
Table 2 Annual Hotspots’ Distance to Nearest Target 32
Table 3 Percent Contribution of Variables in the Two Model Runs 43
Table 4 Model Comparison of AUC and Standard Deviation Values 47
Table 5 Seasonal Hotspots’ Distance to Nearest Target 75

vi

vii

LIST OF FIGURES
Figure 1 SPH at water/feed station; male (back), fawns (front-left), and female (right) -
Photograph by USFWS Remote Camera ........................................................................................ 5
Figure 2 NTAC and STAC Study Area on the BMGR East......................................................... 14
Figure 3 Monitoring From One of the Observation Points (OP Echo) ......................................... 16
Figure 4 Euclidean Distance Calculation for the Target Feature Class in ArcMap 10.1.............. 20
Figure 5 Maxent Main Page .......................................................................................................... 22
Figure 6 Basic Settings Choices for Maxent................................................................................. 24
Figure 7 Advanced Settings Choices for Maxent ......................................................................... 25
Figure 8 Study Area With All SPH Sightings Displayed Over Targets and Roads ..................... 27
Figure 9 Fawning Season 2013 Hotspot Location Map with Underlying KDE
and Overlaid with a Distance from Nearest Target Table ............................................................ 28
Figure 10 Summer Season 2013 Hotspot Location Map with Underlying KDE
and Overlaid with a Distance from Nearest Target Table ............................................................ 29
Figure 11 Breeding Season 2013 Hotspot Location Map with Underlying KDE
and Overlaid with a Distance from Nearest Target Table ............................................................ 30
Figure 12 Hotspot Locations for the Entire Year of 2013 with Underlying KDE
and Overlaid with a Distance from Nearest Target Table ............................................................ 31
Figure 13 Boxplot for Seasonal Variation in Hotspot Distance from Nearest Target .................. 33
Figure 14 Annual Hotspots from 2010-2013 Above 1KM Target Buffer .................................... 35
Figure 15 Seasonal Hotspots for 2010 .......................................................................................... 35
Figure 16 Seasonal Hotspots for 2011 .......................................................................................... 36
Figure 17 Seasonal Hotspots for 2012 .......................................................................................... 37
vii

viii

Figure 18 Seasonal Hotspots for 2013 .......................................................................................... 38
Figure 19 Average AUC for 1,000 Replicates Including the Wetness Index ............................... 40
Figure 20 Variable Response Curves for 1,000 Replicates Including the Wetness
Index. From Left to Right the Variables Shown Include Observer Distance,
Distance from Road, Slope, Distance from Target, and Wetness Index....................................... 40
Figure 21 Jackknife Test on AUCs for 1,000 Replicates Including the Wetness
Index. From Top to Bottom, the Environmental Variables Include, Observer Distance,
Distance from Roads, Slope, Distance from Targets, and Wetness Index ................................... 41
Figure 22 Average AUC for 1,000 Replicates with the Wetness Index Removed ....................... 42
Figure 23 Jackknife Test on AUCs for 1,000 Replicates with the Wetness Index
Removed. From Top to Bottom, the Environmental Variables Include Observer Distance,
Distance from Roads, Slope, and Distance from Targets ............................................................. 42
Figure 24 Predictive Output Raster from the Second Model, Showing Probability
of a Sighting .................................................................................................................................. 44
Figure 25 Average AUC for 1,000 Replicates Using All Sightings with the 25 Percent
Subsample Statistical Method ....................................................................................................... 45
Figure 26 Predictive Output Raster from the Third Model, Showing Probability
of a Sighting .................................................................................................................................. 46
Figure 27 SPH 2010 Hotspot Map with Underlying KDE Raster ................................................ 59
Figure 28 SPH 2010 Breeding Season Hotspot Map with Underlying KDE Raster .................... 60
Figure 29 SPH 2010 Fawning Season Hotspot Map with Underlying KDE Raster ..................... 61
Figure 30 SPH 2010 Summer Season Hotspot Map with Underlying KDE Raster ..................... 62
Figure 31 SPH 2011 Hotspot Map with Underlying KDE Raster ................................................ 63
viii

ix

Figure 32 SPH 2011 Breeding Season Hotspot Map with Underlying KDE Raster .................... 64
Figure 33 SPH 2011 Fawning Season Hotspot Map with Underlying KDE Raster ..................... 65
Figure 34 SPH 2011 Summer Season Hotspot Map with Underlying KDE Raster ..................... 66
Figure 35 SPH 2012 Hotspot Map with Underlying KDE Raster ................................................ 67
Figure 36 SPH 2012 Breeding Season Hotspot Map with Underlying KDE Raster .................... 68
Figure 37 SPH 2012 Fawning Season Hotspot Map with Underlying KDE Raster ..................... 69
Figure 38 SPH 2012 Summer Season Hotspot Map with Underlying KDE Raster ..................... 70
Figure 39 SPH 2013 Hotspot Map with Underlying KDE Raster ................................................ 71
Figure 40 SPH 2013 Breeding Season Hotspot Map with Underlying KDE Raster .................... 72
Figure 41 SPH 2013 Fawning Season Hotspot Map with Underlying KDE Raster ..................... 73
Figure 42 SPH 2013 Summer Season Hotspot Map with Underlying KDE Raster ..................... 74

ix

x

LIST OF ABBREVIATIONS
AFB Air Force Base
ANOVA Analysis of Variance
AUC Area Under the Curve
BMGR Barry M. Goldwater Range
DEM Digital Elevation Model
GIS Geographic Information System
GME Geospatial Modeling Environment
KDE Kernel Density Estimate
MGRS Military Grid Reference System
NTAC North Tactical Range
ROCC Range Operations Control Center
SPH Sonoran Pronghorn
STAC South Tactical Range
US United States
USAF United States Air Force
USFWS United States Fish and Wildlife Service

x

xi

ABSTRACT
The population of Sonoran pronghorn (SPH; Antilocapra americana sonoriensis), an endangered
subspecies within the United States (US), has fluctuated from an estimated 282 individuals in
1994 to 21 in 2002 and back up to over 150 as of August 2014. As the population continues to
recover from drought-associated stressors, more SPH frequent the Barry M. Goldwater military
tactical range and the United States Air Force (USAF) closes more targets from training for
longer periods of time. In this thesis, hotspot analyses are combined with maximum entropy
distribution modeling to understand the geographic and seasonal variation in SPH distribution at
North and South Tactical Ranges (NTAC, STAC) in the Barry M. Goldwater Range East,
Arizona using data from a monitoring effort begun in 1997.
Results show hotspots of high densities of SPH near strafing and bombing targets,
supporting previous studies using fewer data. In Maxent-derived habitat models, distance from
targets had the strongest effect on model performance, followed by slope of the ground.
According to the models, distance from roads had no effect on the SPH locations, nor did
distance from observer. Prior studies attribute SPH preference for areas near targets to
attractiveness of forb growth following disturbance as forage, and high visibility resulting from
few tall shrubs or bushes. Output from the distribution model provides a predictive map of
habitat use that can be used to evaluate effects of range use on SPH in the future.

xi

1

CHAPTER 1: INTRODUCTION
Sonoran pronghorn (Antilocapra americana sonoriensis) is an endangered subspecies of
pronghorn that frequents an active Air Force tactical range in Southwestern Arizona. The
primary conservation effort is to increase the population size through captive breeding and
protection from military and recreational activities. Current SPH habitat encompasses
approximately 20,000 square kilometers on the Barry M. Goldwater Range (BMGR), Cabeza
Prieta National Wildlife Refuge and the Organ Pipe Cactus National Monument (Krausman et al.
2005b). In 2011, a re-introduction effort expanded the range into the King’s Valley of Kofa
National Wildlife Refuge (USFWS 2010). Consistently, 20-45 SPH utilize the BMRG range,
which is a heavily used training range for the Air Force and Marines. Since 2005, no analyses
have been undertaken of the data collected to examine continued trends. Using new spatial
analysis tools I can examine the data and compare it with prior studies’ results. Hotspot analyses
and maximum entropy modelling were used to describe range areas with high densities of use
and to identify potential variables important in predicting the species’ future use. Detailed
information regarding their spatial behavior in this habitat will both help the military plan future
training missions and will additionally enhance biologists’ ability to understand the subspecies.
The goal of this project was to assess where SPH are most likely to occur during different
portions of the year, potentially allowing the USAF to plan missions by focusing on areas of
lesser predicted use during particular time periods. This was accomplished through the use of
hotspots, using core use areas of a kernel density estimate to define each location (Blundell et al.
2001; Wal & Rodgers 2012). To examine why SPH preferred their observed locations, maximum
entropy modelling of distributions was applied to the sightings, using a number of environmental
variables (Phillips et al. 2004, 2006). The output of these models both define important variables,
1

2

and predict the probability of observing a SPH in a particular location. To accomplish these
goals, two related questions are asked. First, where in the study area are SPH frequenting, and
how does it change between seasons and between years? The answer to this question shows
where SPH were seen during a certain time period, though it does not answer why they were
there. This introduces the second question, what characteristics of the study area are important in
defining and predicting SPH location choice? These results showed what variables have high
predictive value when trying to determine future use by SPH.
Historically, SPH existed in the Sonoran desert of Mexico, up into southwestern Arizona
and as far west as southeastern California (O’Gara 1978). The United States Fish and Wildlife
Service (USFWS) listed SPH as endangered in 1967 (USFWS 1998; deVos and Miller 2005).
Since then, the population within the United States (US) has fluctuated due to drought from an
estimated 282 individuals in 1994, to fewer than 30 individuals in 2002 (Morgart et al. 2005;
Krausman et al. 2005a). The current population estimate is approximately 300 individuals within
the US. From a purely scientific basis, this limited population warrants special attention and
analysis.
SPH utilize all areas of the North (NTAC) and South Tactical Ranges (STAC) on the
BMGR in Southwestern Arizona, including areas directly surrounding live fire and high
explosive targets (Krausman et al. 2005b). Missions occur almost daily, including low, dry
overflights as well as actual releasing of bullets towards targets, and bombs towards high
explosive hills.
Studies have shown that this subspecies of pronghorn has a mean individual home-range
size of 511 km
2
with ranges between 43-2873 km
2
(Hervert et al. 2005). Habitat preferences
within this tactical range have been described, showing that pronghorn presence is mostly evenly
2

3

distributed throughout vegetation types, though there is a significant preference for disturbed
areas including both targets and burned patches (Krausman et al. 2005b). Between 1997 and
2005, the population in the study area was quite low, and steadily decreasing. Since 2005, no in-
depth analyses have updated the results from these studies, or examined changes in range use as
the population increased.
As the population continues to recover, and more SPH frequent the tactical ranges, the
USAF is required to close down a greater number of targets for longer periods of time. Current
operating instructions require normal targets to be shut down for a period of 24 hours following a
pronghorn sighting within a one-kilometer buffer. For high explosive targets, this buffer is
extended to 1.5 km. In a study area of between 10 – 25 km from West to East, and approximately
30 km North to South, a large area is often shut down during portions of the year because of
multiple, widespread sightings of SPH groups and individuals.
This study provides additional information to the range managers about areas of high
density use by SPH throughout the year. Perhaps target use priorities can be identified for
different times of the year. While predicting target use throughout the year will not change target
closures, mission planning can be made more efficient by focusing on historically less active
targets for a specific time period. Also, in the event of new target complexes being developed,
areas of low use can be chosen for target placement. Examining use of the ranges by SPH is
providing information about the effects of targets, roads and other variables on SPH movement
and area use.

3

4

CHAPTER 2: RELATED WORK
The SPH are one of five subspecies of pronghorn in the US (Hosack et al. 2002; Krausman et al.
2005a). The sizes of the two remaining US and Mexico populations of SPH are estimated every
other year (Table 1; ie. Mexico in 2012, US in 2013), a process that began in 2000 and continues
to present time (Pronghorn recovery meeting, 2013). Until 2002, the population was declining
sharply, with few sightings in the study area, though since then it has increased almost every
consecutive year (Table 1). Both populations are unstable, with little genetic diversity due to the
genetic bottleneck (Stephen et al. 2005). The majority of written work on SPH resides in
technical documents, though a few exist in peer reviewed journal articles as well as a number of
book sections (Krausman et al. 2005c). Most of these published articles or books refer to data
collected prior to 2003, with one or two as recent as 2005.

Table 1 Annual SPH Population Estimates and Maximum Individuals per Range
Year (20xx)
00 01 02 03 04 05 06 07 08 09 10 11 12 13
Max # on
NTAC
16 15 13 8 6 9 4 6 13 27 24 24 23 40
Max # on
STAC
14 21 13 9 9 6 23 20 25 26 41 32 19 44
Max Total on
Range
16 21 16 10 13 9 23 20 25 31 45 43 42 57
Population
Estimate
100 - 21 - 58 - 68 - 68 - 85 - 159 -
Source: Data provided by USFWS and Nicole Tautfest at Chiulista Services, Inc.
2.1 Sonoran Pronghorn
Pronghorn are a small deer-like ungulate living throughout parts of North America. Of the five
subspecies, the SPH is the rarest, existing in three separate populations in southwestern Arizona
and in northwestern Mexico (Stephen et al. 2005; Motsinger 2013). Adapted to dry conditions,
this subspecies is somewhat smaller and lighter than the other North American subspecies, with a
4

5

maximum weight between 48–60 kilograms (105-133 lbs). Because of widely distributed food
sources and drier conditions, home-ranges of the SPH are much larger than the other subspecies
(Yoakum 1978), at an average of 511 km
2
, compared to 22–170 km
2
of pronghorn from other
areas of the United States (Hervert et al. 2005). SPH have a characteristic sandy brown body,
with white stripes on the neck and a white belly and chest (Figure 1; Reid 2006). Males have a
distinct black patch behind the cheek.

Figure 1 SPH at water/feed station; male (back), fawns (front-left), and female (right) -
Photograph by USFWS Remote Camera
SPH prefer habitat dominated by palo verde (Parkinsonia spp) and chain fruit cholla
cactus (Opuntia fulgida), in part because the cholla fruit provides a source of stored water (Fox
1997). In years of high precipitation, creosote-bursage habitat communities are frequented, as
5

6

herbaceous plants grow abundantly (Hervert et al. 2005). Diet consists dominantly of forbs and
grasses, which are taken advantage of soon after precipitation, or in disturbed areas. Some
succulent shrubs are consumed during the winter months where forbs have been consumed or
have died out (Ochenfels 2014). Currently, hay is provided to supplement natural food sources in
captive pens and near artificial water sources in an effort to minimize mortality from starvation.
For many years it was thought that SPH required little to no freestanding water for
drinking, deriving all needed moisture from food sources (Tinker 1978). In 1987, a SPH was
photographed drinking from an artificial water source, and since then multiple similar sightings
have been documented using wildlife cameras (Morgart et al. 2005). Freestanding water is
important to SPH, and severe droughts limited water and forage, helping to cause the decline to
an estimated 21 individuals in the US by 2002 (Bright & Hervert 2005).
Observations of SPH on NTAC and STAC (the study area for this thesis) from 1999–
2002 indicate that approximately 73 percent of sightings occurred in military disturbed areas
(Krausman et al. 2005b). In a similar report, Hervert et al. (1997) demonstrated SPH were
attracted to target areas because they could see predators at a greater distance across the open
desert landscape, as well as the access to standing water and forb growth resulting from constant
disturbance. Another study focusing on the soil surrounding high explosive targets found that
standing water in bomb craters was most likely of limited use to SPH because of the high rates of
evapotranspiration (Tetra Tech 2005). This same study found that residue left by explosions
added insignificant amounts of sodium to the landscape, so it is unlikely the SPH are attracted
due to the conditions being similar to a salt lick. In general, little difference in SPH behavior
were observed on the BMGR between 1998-2000, when anthropogenic noise levels were
increased by aircraft (Krausman et al. 2001). For adult SPH, behavior changed less than 50
6

7

percent of the time. The conclusions drawn by Krausman, Harris, and Francine (2001) are that
military activities have minimal effects on SPH when appropriate monitoring and target closures
are in effect.
2.2 Areas of Intensive use by Sonoran Pronghorn
Hotspots are defined as spatial clusters of points (Liu and Brown 2003). These clusters are areas
with a high density of features or phenomena of interest (Nelson and Boots 2008). While a
number of methods are available to calculate and analyze hotspots, local spatial autocorrelation
and kernel estimators are the two most often used tools. Kernel estimators create density
surfaces, the highest values of which are distinguished as hotspots. Often these density surfaces
are used to define overall distributions, or home ranges and core use areas of animals (Kie et al.
2010; Worton 1989). For this thesis, only the core use areas were used when defining hotspots,
because highest use is the primary interest.
2.2.1 Kernel Density Estimation
Kernel density estimation (KDE) is applied to many fields of study, including but not limited to
medicine (Er et al. 2010), invasive species outbreaks (Nelson and Boots 2008), crime analysis
(Liu and Brown 2003), home-range studies (Blundell et al. 2001), and evaluation of hazardous
conditions (Mola-Yudego and Gritten 2010; Lin et al. 2011). A KDE creates a continuous raster
density surface over a two-dimensional XY coordinate plane (Downs and Horner 2008; Xie and
Yan 2008; Downs 2010). This density surface is useful for visualizing locations of spatial
abundance, as well as emphasizing areas with a lack of use. Because density is represented in
this way, not only are the focus points of abundance obvious, but a visible and calculable
relationship between spatial location and likelihood of an occurrence happening is established
(Anderson 2009).
7

8

When creating a KDE from a set of locations, two primary variables greatly affect the
output raster: the bandwidth (search radius/threshold) and the output cell size (Nelson and Boots
2008; Anderson et al. 2009; Krisp and Peters 2011). The bandwidth represents the search radius
of the kernel calculation while searching for nearby points. This variable is represented by a
distance chosen manually as a fixed kernel, or automatically as an adaptive kernel (Seaman and
Powell 2005). Larger bandwidth values produce smoother and larger KDE area calculations
(Ramp et al. 2005). The output cell size does not affect the calculation as much as the output
visualization grid. Cell size is the pixel size representation of the KDE (Nelson and Boots 2008).
Cell size should not be smaller than the minimum point accuracy being input into the KDE. To
create an accurate density estimation, Seaman et al. (1999) recommends the use of at minimum
30 geographic locations, and ideally 50 or more.
Fixed kernels for a KDE are chosen manually, relying on the judgment of the researcher.
Wand and Jones (1995) describe this method by examining various bandwidth values from
largest to smallest, and finding the one that is “most pleasing” to the researcher. This method is
subjective, allowing the researcher to decide how smooth the edges of the raster output will be.
For example, an edge with little smoothing has sharper edges, conforming more closely to the
actual input points while a smoother edge will incorporate more area, with fewer to no sharp
corners. Nelson and Boots (2008), studying mountain pine beetle hotspots, recommend using
values between 800–2000 meters because they return results that are smoothed, but not so
smoothed as to overestimate the distributions. A fixed kernel method may be useful when the
study purpose involves data exploration for hypotheses before another analysis (Silverman
1986).
8

9

Adaptive kernels are those that allow for local variation in smoothing of density
calculations. This means that areas of lower density are smoothed differently than those with
higher densities because of their nearness to other points (Wand and Jones 1995). While these
are useful in many situations, such as when working with data that are unknown to the
researcher, the approach has drawbacks. The main drawback is that comparability between
samples is lost because the local variation is calculating different results across the range, instead
of using a known, and comparable, search radius.
2.2.2 Hotspots
Kernel estimated hotspots are used for many types of analyses, both on a two dimensional plane,
and across a network. Nelson and Boots (2008) used KDE methods to locate hotspots of
mountain pine beetle infestations. They examined bandwidth values of 450, 800, 2000, and 9500
meters, and found that the middle two were best because they did not create as many fine scale
unconnected hotspots, nor as many large, over connected hotspots. Lin et al. (2011) examined
soil pollution hotspots using KDE methods and found it an appropriate method of delineating
areas of high pollution. A KDE output shows density of geographic location, and when these
densities are combined with geostatistics they show variables of interest within the attribute table
that coincide with those spatial densities. Anderson (2009) and Xie and Yan (2008) both
investigated traffic incident hotspots with KDEs. Anderson found that the continuous surface of
a KDE was useful in providing a spread of risk, showing predictive values at locations nearby
but not on top of known incidents. Xie and Yan (2008) took a standard KDE method and adapted
it for a network, concluding that their method was an improvement over the standard when using
a road network. In the wildlife and ecological field, hotspots are often used to examine road
mortality of animals, as is the case for a study by Ramp et al (2005), looking at fatalities along
9

10

the Snowy Mountain Highway in Australia. For this study, a bandwidth of 500 meters was
chosen a priori to be comparable with nearby analyses. Ramp et al. (2005) concluded that spatial
clustering using density is an appropriate method for identifying hotspots.
2.3 Maximum Entropy Modelling
Maximum entropy modelling is a method of distribution modelling using presence-only data
(Phillips, Dudik, and Schapire 2004; Elith et al. 2011). This method is useful because even
though an observer does not see an animal at a sample location, it may be hiding, or visibility
may be less than ideal, which introduces observer bias (Gu and Swihart 2004). In addition,
absence data are often not available, either because they were not collected, or because a project
did not have enough funding to sample enough sites to accurately or sufficiently delineate
absence locations. This modelling method evaluates the most uniform distribution across an area,
which allows the model to show how much better the data fit than a uniform distribution would
(Baldwin 2009). Though relatively new, this tool has been used extensively, with over one
thousand publications since 2006 (Merow, Smith, and Silander 2013).
Maxent is a software package designed to analyze data in this way, looking at
environmental variables to distinguish possible relationships between the location of animals and
the contributing variables (Phillips, Dudik, and Schapire 2004). This software allows a user to
input location points for a species or multiple species of interest, in a comma delimited file
(CSV), using three columns: “Species,” “X,” and “Y.” Environmental variables, which are any
potentially contributing variables, must be included in an ASCII format, and must be cropped to
the exact spatial dimensions. Many settings and tests can be specified within the options,
including the input feature method, the output format, and statistical calculations performed.
10

11

Although Maxent has many benefits, such as the ability to perform well with very limited
sample sizes, Hernandez et al. (2006) point out that model accuracy increases for smaller
geographic ranges, and for species with more restricted environmental tolerance. The potential
for sampling bias can negatively affect the usability of the results as well, though this is the case
in most studies dealing with wildlife. When choosing environmental variables for a particular
model, care should be taken to select variables that correspond well with the geographic
locations of the species. For example, an average precipitation layer should be from the same
temporal scale as that of the points being analyzed. Similarly, variables with little variation
across the study area extent should not be used, as they will not affect the model because they are
not at the appropriate scale (Phillips, Anderson, and Schapire 2006).
The ability to perform accurate analyses of distributions with only presence data has
made this method popular among the scientific community. Yost et al. (2008) used Maxent to
map sage grouse nesting habitat in southern Oregon. They tested their models by withholding 25
percent of their data from the training algorithm. They found that the type of vegetation was
most important in predicting nest sites, and visually, the output image of high probability areas
were accurate when comparing with the nest sites withheld for testing.
Suarez-Seoane et al. (2008) used Maxent to analyze seasonal change in little bustard
(Tetrax tetrax) distribution. They found that during the winter, lower elevations and more
wastelands were occupied, which differed from the spring use of higher elevation pasturelands.
The results showed that managers could potentially use spring habitat selection to predict winter
habitat, which is important when managing a rare species, or a sport hunted species.
Also, a study on predictive modeling of microhabitat for endemic Chilean birds, by
Moreno et al. (2011), used Maxent in a small geographic area (approximately 4.5 km
2
). They
11

12

showed one species’ preference for areas farther from roads in lower elevation forest, while
another species preferred higher percent slopes in more rugged terrain. Their study concludes
that when combining GIS to visualize and produce results for the analysis, maximum entropy is a
useful tool for forest management.
12

13

CHAPTER 3: METHODS
This chapter introduces the methods used for this thesis. First, the study area is defined more
clearly, then the data collection is explained. Because data collection started in 1997, some new
tools have been introduced, though the methods remain similar. The data analysis techniques are
described, beginning with the creation of hotspots, progressing to the use of Maxent to create
distribution models. The Maxent methods include management and creation of layers to work
within the software, and covers the settings used in this project.
3.1 Study Area
The data collected for this thesis come from NTAC and STAC, which are air-to-ground tactical
ranges on the BMGR East (Figure 2). NTAC and STAC consist of 777 km
2
of mostly flat land
with 813 individual targets that are combined into approximately 30 target complexes (USAF
Targets 2013 dataset). Three of these complexes consist of high explosive targets, which are of
extra concern and require more attention when monitoring. Bladed dirt and sand roads connect
targets to each other for access and maintenance. Four primary observation points designed to
provide the best coverage of all targets are used, and six secondary observation points are
available for special use when necessary. Bordering the study area on the South and Southeast
side is the Cabeza Prieta National Wildlife Refuge, and to the east is State Route 85, connecting
Gila Bend, Arizona to Ajo, Arizona.
13

14

Figure 2 NTAC and STAC Study Area on the BMGR East
3.2 Data Collection
Data were collected from February 1997 to April 2014 and continue to be collected as part of an
ongoing effort by the USAF to prevent SPH mortality caused by training activities in the study
area. Certain operating procedures have changed over the years; however most of the changes
pertain to buffer distances surrounding targets, which are closed from target practice following a
14

15

SPH sighting. Also, from 1997–2007, monitoring sessions were performed twice, rather than
once, in a 24-hour period (Nicole Tautfest, face-to-face, December, 2014). The second
monitoring session was primarily for double-checking the high explosive hills.
These methods are reported from the monitoring protocol set forth by the USAF and
pronghorn biologists for monitoring SPH on the ranges (Pronghorn Biologists 2012). Starting at
sunrise, biological monitors scan NTAC and STAC for at least two hours from primary
observation points on each day with training missions (a minimum of two monitors on each
range). When missions are late and monitoring is more difficult due to many animals or difficult
environmental conditions, monitoring may go longer than the minimum two hours. Secondary
observation points are used when one range is closed for maintenance, or when road or weather
conditions limit access to the primaries.
Observation point locations were selected because they provided vantage points to see
target complexes clearly, so SPH sightings are potentially biased towards locations near targets.
Scanning is performed using 10x50 binoculars of various brands, as well as with 20-60x Leica
and Swarovski spotting scopes (Figure 3). Scans of the target areas are prioritized over areas
without targets, although surrounding areas are always scanned. In addition, telemetry is used to
locate radio collared individuals across the landscape and to place them within visible groups.
Yagi antennas (Communication Specialists, Orange, CA, USA) are used in combination with
Communication Specialists (Orange, CA, USA) R-1000 receivers. A minimum of two crossing
bearings is required for a location to be established by telemetry alone.
When a SPH or group of SPH is observed, location is recorded in military grid
reference system (MGRS) format to the nearest 100 meters (e.g. TA946997). Basic behavioral
observations are recorded, taking note of number of SPH in the group, sex, age, and general
15

16

direction of travel. At the end of the monitoring window, locations are reported to the range
operations control center (ROCC), where they are input into ArcMap to determine target
closures. Targets where inert ammunition is used are closed for a period of 24 hours if the
sighting is within a one kilometer buffer. For the three high explosive target complexes, this
buffer is extended to 1.5 kilometers.

Figure 3 Monitoring From One of the Observation Points (OP Echo)
3.3 Data Analysis
All analyses were conducted using a combination of ArcMap (version 10.1), Geographic
Modelling Environment (GME; version 0.7.2.1), R (version 3.1.0), and Maxent (version 3.3.3k).
GME works as a standalone program, though it requires an installation of ArcGIS and R to
function. It replaces an older ArcGIS plugin called “Hawths Tools,” which was created for
ArcGIS 9.3.x, and specifically designed for advanced spatial analysis and modeling. Maxent is a
standalone software package developed to study geographic distributions of species and to
16

17

predict probability of presence (Phillips, Anderson, and Schapire 2006; Phillips, Dudik, and
Schapire 2006; Kumar and Stohlgren 2009).
3.3.1 Hotspots
This section prepares to answer the first question presented in this thesis: where are hotspots of
SPH use across the study area? There is a temporal aspect within the question, in that variation
occurring between seasons and throughout the year is important. For example, if biologists are
wanting to predict SPH use during the month of November, they must examine a hotspot
analysis of recent sightings and one from the previous year’s coinciding time period. In this way,
they will account for prior year’s areas of high use while incorporating the current year’s trends
to predict important points within the study area for the future time period of interest. To achieve
this task, hotspots were calculated for the years 2010-2014, and for each breeding, fawning, and
summer season within each of those years. This was accomplished using ArcMap 10.1, creating
a KDE with 100-meter cell size and a bandwidth of 1,000 meters. The bandwidth value was
chosen because it represents the buffer distance for the majority of the targets throughout the
range. Points within that distance of each other are important when calculating the density, and
will be significant to the monitoring effort. Only time periods with a minimum of 30 sighting
locations were used for the analysis, as recommended by Seaman et al. (1999). GME was then
used to create polygons around the highest 20 percent of the density values. These polygons were
brought back into ArcMap, and center points within each polygon were distinguished using the
Feature to Point tool, specifying inside features. These were output into a new feature class
named for the time period and sighting type (e.g. All SPH, year 2013 = cent_2013). The
visualizations were symbolized using large red or green dots over the KDE raster.
17

18

Hotspots were calculated for many different temporal scales, to examine variation at
different times of the year, and between years. One hotspot map was produced for all sightings,
one for each year’s sightings, and one for each season from 2010-2014. For this analysis and the
others, seasons are defined as: breeding season (Oct-Jan), fawning season (Feb-May), and
summer season (Jun-Sep). Once calculated, these hotspots were put into a table and the distance
from the nearest target was added. Both the annual hotspot table and the seasonal hotspot table
were then analyzed using an analysis of variance (ANOVA) test using R, to test for significant
differences in distance from nearest targets between years and between seasons.
3.3.2 Maximum Entropy Modelling
Maxent is a small but powerful program, with many features that are constantly being updated.
These options allow users to customize the input features, as well as output formats and
statistical analyses. Though customization is good for researchers with a good understanding of
the methods involved, the default settings also produce accurate and usable results (Phillips and
Dudik 2008; Baldwin 2009). Many of the following methods were derived from a tutorial
(Phillips 2005).
3.3.2.1 Layer Preparation
The first step in initiating this analysis consisted of preparing a plan to address the potential
sampling bias inherent in any study. Creating an effort layer to use as a bias file within Maxent
was not feasible because the only way to do so was very subjective. Therefore, the SPH dataset
was split into two. The first dataset consisted only of telemetry sightings. These sightings were
not visible to the observer’s eyes, relying solely on telemetry bearings for previously radio
collared SPH, and as such cannot be misplaced due to bad judgment of distance, or other
visibility related environmental variables such as wind or haze and heat distortion. The second
18

19

dataset was created using only visual sightings, those seen without any telemetry aid. In this way,
instead of setting aside a percentage of the overall sightings for training and testing, I created the
telemetry layer for training the algorithms, and the visual layer for testing of that algorithm.
Next, I assembled and created the environmental layers. Though I accounted for some
potential bias in observer ability by using telemetry sightings for training, I created a distance
raster of distance from observation point to test the overall contribution of observation distance
to sighting location. I created this layer using the Euclidean Distance tool within ArcMap, using
100 meter cell sizes. The 100 meter cell size was chosen because it coincides with the scale of
the SPH sightings collected. Smaller sizes would potentially introduce artifacts, while larger
sizes would limit the accuracy and utility of the output when considering such a small study area.
This layer was called “Observer Distance.”
As distance from roads and distance from targets are considered important because of
their inherent cause of disturbance to the SPH, I created a distance layer for each of these factors,
with 100 meter cell size. Two layers were created, again using the Euclidean Distance tool in
ArcMap, one from the target feature class, and one from the road feature class (Figure 4). Output
raster layers were labeled “Target Distance,” and “Road Distance.”
19

20

Figure 4 Euclidean Distance Calculation for the Target Feature Class in ArcMap 10.1
Observation of SPH show they prefer areas of flat ground, rarely venturing up slopes
(Nicole Tautfest, face-to-face, July, 2014). Because access to some locations requires climbing
hills or small mountains, it is not always the case that steep slopes are avoided. Because of this, I
created a slope layer from a 90 meter Arizona digital elevation model (DEM), which was
provided by biologists at Luke Air Force Base (AFB). I cropped the DEM to the general area,
ran the Slope calculation tool within ArcMap, then resampled the raster to 100 meter cell sizes.
This layer was named “Slope.”
Lastly, I created a wetness index layer, to see if soil water content had an effect on SPH
location preference. A precipitation layer could be used, though because variation across such a
small area is often quite low, the wetness index allows the study to look at overall drainage and
20

21

water storage rather than amount of precipitation directly. This layer was called “Wetness
Index.”
To make sure each layer had the same geographic bounds, I created an extent layer,
which I proceeded to use when clipping all the environmental layers to the same size. I clipped
both input SPH layers as well, to make sure the overall extents matched. Once clipped, each
layer was exported to ASCII format, using the Raster to ASCII tool for raster layers, and the
Export Feature Attribute to ASCII for the SPH point features. The two SPH feature classes were
organized into a single folder, while each raster layer and coinciding ASCII file were put into
another folder, as required by Maxent.
3.3.2.2 Maxent Settings
The main screen for Maxent is relatively simple, though without reading through the tutorial and
the descriptive papers on the software, choosing options would be quite complicated (Figure 5).
On the top left, the samples used for the analysis are chosen from a single comma delimited file.
Multiple species or groupings can be included in a single file, but in this case only one was
needed. For my analysis, this was the telemetry only file. The top right of the screen is where the
user must choose the folder containing the environmental layers, and then differentiate between
each layer being of a continuous or categorical type. All of the environmental variables in this
model were continuous.
21

22

Figure 5 Maxent Main Page
Maxent has many options for features that vary the complexity of the model (Figure 5).
The default setting is Auto, which lets the software choose the feature type based on the input
variables. Merow, Smith, and Silander (2013) explain that when auto is checked, if there are
more than 80 presence locations, then all the feature classes will be used. These features define
how the model will be fit. I accepted the default of Auto, because it would run all of the feature
types, choosing those that are most useful when building the model.
Underneath the environmental layers box, on the right, are check boxes to create response
curves, picture predictions, and do jackknife tests to measure variable importance within the
22

23

model. Response curves show how input environmental variables affect the model prediction
(Phillips 2005). The pictures of predictions creates an image of the output grid in a format easily
readable by standard image viewers, in case access to ArcMap is limited. The Jackknife test
examines variable importance to the overall model, by removing the variable and running the
model again, to see the overall difference in results. In this way, a variable is assigned an
importance to the model in the form of a percent contribution.
An output format of logistic was chosen because it produces a value within cells across
the study area consistent with the probability of having a sighting at that location. It is easily
understandable and provides a predictive capacity (Phillips, Anderson, and Schapire 2006;
Phillips and Dudik 2008; Elith et al. 2011). I chose an output type of asc, for its ability to be
incorporated into ArcMap for integration with future sighting information. Finally, I chose an
output folder, which included all the output information produced during a model run.
After the main settings are chosen, clicking on the settings tab at the bottom brings up a
new window with additional options (Figure 6). This is the page that allowed me to specify a test
dataset of visual only sightings. When filling this, and specifying a test dataset, all sightings from
the previous page’s sighting input is used to train the model algorithm. Most options were left as
default, except I increased the maximum number of background points from 10,000 to 20,000.
Merow, Smith and Silander (2013) mention that the default or similar value is acceptable for this
if all areas within the extent are potentially accessible to the species. I chose a larger maximum
background points because of the large number of presence locations available (Phillip and
Dudik 2008). Below that, I chose to run 1,000 replicates for statistical purposes, with
crossvalidate as the replicated run type. This method leaves out “folds” of data, which are used
23

24

for testing against each other, providing evaluation for the model. In this way, it utilizes all the
available data for testing each replicate run (Phillips 2005).

Figure 6 Basic Settings Choices for Maxent
Clicking to the right of the basic tab brings up the advanced options (Figure 7). All
defaults were accepted except with an additional check mark for the “Write plot data” box, to
provide access to raw data used in creating the response curves, in a format importable to
peripheral software. After checking these settings, closing the window brings back the main
page, where the run button can be clicked to start the analysis.
24

25

Figure 7 Advanced Settings Choices for Maxent
After these main analyses were completed, I ran three more models using mostly the
same settings, except a more traditional approach to calculating statistics on the replicated runs.
In the third model, I used all the data for SPH, combining all telemetry and visual sightings.
Rather than cross-validation, I used the subsample feature which holds back a random
percentage, in this case 25 percent, of points for testing, and the remaining 75 percent is for
algorithm training. This model run is indicative of an ideal dataset with all points being equal in
quality, with little to no observer bias. It was then compared to the model created to account for
observer bias. The fourth model consisted of the same settings, but run on telemetry only
sightings. The fifth model again used the same settings, but on the visual only sightings. These
25

26

last two models were run to see if there were any differences in model performance when the
training and testing data were from one type of sighting, instead of both like model three, or
separate for models one and two.
26

27

CHAPTER 4: RESULTS
The analysis included 8,092 SPH sightings between March 1997 and March 2014 (Figure 8). The
telemetry only sightings numbered 1,464 and visual only sightings were 3,919. The remaining
2,709 sightings were seen and located with telemetry, with no information as to whether they
were seen before an animal with telemetry was heard, or whether the telemetry assisted in the
observation. Because of this, these sightings were not used for the Maxent portion of the
analysis, though they were all included when creating hotspots.

Figure 8 Study Area With All SPH Sightings Displayed Over Targets and Roads
27

28

4.1 Hotspots
Hotspot calculations using the 100-meter cell size and 1,000-meter bandwidth produced
acceptable results that correspond visually with the locations of SPH sightings. Results are
visually appealing, and portray the important details of high-density areas. To provide an
example of the KDE output, a figure is shown of the KDE raster with isopleth polylines as well
as hotspot locations (Figure 9). For all consecutive maps, only the hotspot locations are shown,
along with coordinates for those locations and distance from nearest target. For the 2013 fawning
season, the average distance from target was 466 meters. Ten of the eleven hotspots were <615
meters from targets, while a single hotspot was at 2.1 kilometers.

Figure 9 Fawning Season 2013 Hotspot Location Map with Underlying KDE and Overlaid
with a Distance from Nearest Target Table
28

29

Summer and breeding season sample maps show similar hotspots, though the summer
season map has hotspots located further south, towards the artificial water station (Figure 10,
Figure 11). In the summer season, the average distance from targets was 1,127 meters, though
six of the seven points were <670 meters, and one was at almost five kilometers. This farthest
distance hotspot coincides with an artificial water and feed station, which is heavily used during
the summer months, or other seasons during a dry year. The breeding season hotspots had an
average distance of 317 meters to the nearest target, and the farthest was 600 meters.

Figure 10 Summer Season 2013 Hotspot Location Map with Underlying KDE and Overlaid
with a Distance from Nearest Target Table
29

30

Figure 11 Breeding Season 2013 Hotspot Location Map with Underlying KDE and
Overlaid with a Distance from Nearest Target Table
When taken as a whole, the 2013 sightings have hotspots in similar locations, though
spread out over the entire area, rather than clumped in one part of the study area as are certain
seasons. Looking at annual hotspots for 2013, few of them are different from any of the seasonal
breakdown maps, they are just spread out over the entire area rather than clumped (Figure 12).
Over the year, the average hotspot distance from a target comes to 1,254 meters, with four of the
six locations under 1000 meters, one at 1,092 meters, and the farthest at five kilometers. Again,
30

31

this farthest location is at an artificial feeding station, which draws the SPH to it during hot and
dry times of the year (location 1).

Figure 12 Hotspot Locations for the Entire Year of 2013 with Underlying KDE and
Overlaid with a Distance from Nearest Target Table
A summary of hotspot distances emphasizes the nearness of these high use locations to
targets. Annual hotspots are comprised of the entire year’s SPH sightings, so must be tested
separately from the seasonal hotspots, which split the locations into groups. Using an analysis of
31

32

variance (ANOVA) on the annual hotspots’ distance from targets, there is no significant
difference between years (F=0.006, p=0.999). There is an average of 1,305 meters from the
nearest target, with a range of 36–5016 meters (Table 3). Each year has one large distance, near
the artificial water source, and when this location is removed from each year, the average
decreases to 388 meters.
Table 2 Annual Hotspots’ Distance to Nearest Target
Year Hotspot Distance
2010 1* 4959
2010 2 737
2010 3 504
2010 4 316
2010 5 306
2011 1* 5016
2011 2 36
2011 3 341
2011 4 162
2012 1* 4980
2012 2 36
2012 3 332
2012 4 685
2012 5 171
2013 1* 4947
2013 2 1092
2013 3 337
2013 4 953
2013 5 64
2013 6 134

*Located at artificial food and water station.
Seasonal hotspots show a different pattern, where there is a significant difference
between the distance from targets during the summer season from the breeding and fawning
seasons (F=3.978, p=0.0236, Table 5-Appendix B). When the food and water station hotspots are
removed from the data, the difference in target distance between seasons becomes insignificant
32

33

(F=0.266, p=0.7674). The breeding seasons have an average distance from nearest target of 459
meters, while the fawning and summer season are 380 meters and 1,227 meters respectively. The
summer season’s average distance is much greater because of the artificial water source drawing
SPH, and when these outliers are removed, the average distance to targets for the summer season
decreases to 477 meters (Figure 13).

Figure 13 Boxplot for Seasonal Variation in Hotspot Distance from Nearest Target
These hotspot results and others, which are included in Appendix A, agree with a simple
look at the number of sightings and hotspots within one kilometer of targets (Figure 14).
Performing a quick buffer analysis shows that 58 percent of all sightings are within a one
kilometer buffer. Based on these analyses, it confirms that SPH congregate near targets, though
33

34

the specific targets change throughout the year and between years. In general, the hotspots move
more southward during the summer months each year, and move slightly northward during the
winter months (Figure 15, Figure 16, Figure 17, Figure 18). Also, far fewer sightings are made
throughout the range during the summer months. Results of each hotspot calculation for the last
four years were produced as maps, and are included in Appendix A of this thesis.

34

35

Figure 14 Annual Hotspots from 2010-2013 Above 1KM Target Buffer

Figure 15 Seasonal Hotspots for 2010
35

36

Figure 16 Seasonal Hotspots for 2011
36

37

Figure 17 Seasonal Hotspots for 2012
37

38

Figure 18 Seasonal Hotspots for 2013
38

39

4.2 Maximum Entropy Modelling
Results and model evaluation are based on the area under the curve (AUC), which represents
model performance (Phillips, Anderson, and Schapire 2006). Separate models can then be
compared to each other when evaluating which model to pursue further. Individual variables
within a model run can be examined using the percent contribution, as calculated by the
jackknife tests. Baldwin (2009) used an AUC evaluation scale where a value of less than 0.7 was
considered uninformative, while values between 0.7 and 0.9 were considered good, and values
greater than 0.9 were considered very good. Two models were run to examine the effect of
including a wetness index within the model. The first model, including the wetness index, also
included the target distance, road distance, observer distance, and slope. The second model
removed the wetness index, but was otherwise the same.
The 1,000 replicates of the first model, which incorporated the wetness index, produced
an average AUC of 0.803, with a standard deviation of 0.166 (Figure 19). When examining
individual variable response curves, observer distance and road distance had a certain amount of
variation present between replicates, while the remaining three variables had almost no variation
between replicates (Figure 20). The jackknife test results for each variable on AUC values show
that targets had the most impact as a standalone variable, as the model AUC went down
significantly when it was removed from the model (Figure 21). The most important variable
shown is the target distance, as it had the highest AUC value when it was the only variable
considered, and the model AUC drops significantly when it was removed. The three variables
above it (observer distance, road distance, and slope) had moderate AUC results if they were the
only variable, but when removed from the overall model, there was little change. The wetness
39

40

index variable had very little impact on its own, and when removed, increased the model AUC,
prompting the second model run, excluding the wetness index.

Figure 19 Average AUC for 1,000 Replicates Including the Wetness Index

Figure 20 Variable Response Curves for 1,000 Replicates Including the Wetness Index.
From Left to Right the Variables Shown Include Observer Distance, Distance from Road,
Slope, Distance from Target, and Wetness Index
40

41

Figure 21 Jackknife Test on AUCs for 1,000 Replicates Including the Wetness Index. From
Top to Bottom, the Environmental Variables Include, Observer Distance, Distance from
Roads, Slope, Distance from Targets, and Wetness Index
The second model, excluding the wetness index variable, increased the average AUC
value to 0.805 and lowered the standard deviation to 0.150 (Figure 22). The increase, though
minimal, was an improvement. Variable response curves remained the same, as did the jackknife
test results. Without the wetness index, the jackknife test is more obvious in showing that the
target layer is most important. What is interesting is that while both observation distance and
road distance perform better than slope when considered on their own, slope causes the next
largest decrease in overall model AUC when it is removed (Figure 23).
41

42

Figure 22 Average AUC for 1,000 Replicates with the Wetness Index Removed

Figure 23 Jackknife Test on AUCs for 1,000 Replicates with the Wetness Index Removed.
From Top to Bottom, the Environmental Variables Include Observer Distance, Distance
from Roads, Slope, and Distance from Targets
Comparing the percent contribution of variables from the two runs shows an increase for
target distance and slope, but a decrease in observer distance (Table 3). Even though the percent
contribution was higher for observer distance, it was the slope that had a larger effect on the
42

43

model performance. Importantly, no variable increased the overall model AUC when removed
from the run.
Table 3 Percent Contribution of Variables in the Two Model Runs
Variable
Percent contribution
(With Wetness Index)
Percent contribution
(NO Wetness Index)
Target Distance 71.0 78.1
Observer Distance 22.2 10.8
Slope 5.2 7.9
Wetness Index 1.0 NA
Road Distance 0.5 3.1

Once the second model was identified as performing better, with a higher AUC value, the
output raster probability layer was imported into ArcMap and examined (Figure 24). Each cell of
the raster has a value related to the probability of detecting a SPH at its location. This raster was
produced using all of the telemetry sightings for training, and the visual only sightings for
testing. Seasonal or even monthly predictions could be made in order to more accurately predict
locations for the future based on recent data only.
43

44

Figure 24 Predictive Output Raster from the Second Model, Showing Probability of a
Sighting
The third model, using all the sighting data and the subsample method for calculating
statistics produced a slightly lower AUC value of 0.791 (Figure 25). Where this model improved
significantly was in the standard deviation, which came in at 0.004, as compared to the 0.166 and
0.150 of the first and second models. The. response curves and jackknife tests were no different
44

45

from the first two models. The third model’s output map shows a smoother distribution of high
predictive values, though the overall predictive ability is lower (Figure 26).

Figure 25 Average AUC for 1,000 Replicates Using All Sightings with the 25 Percent
Subsample Statistical Method
45

46

Figure 26 Predictive Output Raster from the Third Model, Showing Probability of a
Sighting
Models four and five used telemetry only, then visual only data respectively for input
locations, and the 25 percent subsample to calculate the AUC. Both AUC values were higher
than for model two and model three, with standard deviations much lower than model two, and
comparable to model three (Table 4). Though the overall AUC values were higher, the jackknife
test and residuals remained the same when compared to the previous models.
46

47

Table 4 Model Comparison of AUC and Standard Deviation Values

Model 1
*
Model 2
**
Model 3
***
Model 4
****
Model 5
*****

Variables 5 4 4 4 4
Target Distance x x x x x
Observer Distance x x x x x
Slope x x x x x
Road x x x x x
Wetness Index x - - - -

AUC Value 0.803 0.805 0.791 0.814 0.816
Standard Deviation 0.166 0.150 0.004 0.008 0.005

* All variables, telemetry for training, visual for testing
** No wetness index, telemetry for training, visual for testing
*** No wetness index, all sightings, 75% training, 25% testing

**** No wetness index, telemetry only sightings, 75% training, 25% testing
***** No wetness index, visual only sightings, 75% training, 25% testing

47

48

CHAPTER 5: CONCLUSION AND DISCUSSION
The results from this study confirm those from previous studies that used pre-2005 data, showing
high densities of use, and high probabilities of future use, surrounding targets and disturbed areas
(Krausman et al. 2005b). This thesis provides a demonstration of a different tool’s ability to
analyze and portray use of the target range by SPH. Unlike the study by Krausman et al. (2005b),
it allows users to examine various environmental variables at any particular scale to decide
which are most important when considering the future preference of SPH for a particular
location.
Both the hotspot analysis and distribution model indicate that SPH have strong preference
for areas at and around target structures. The few times a hotspot location was more than one
kilometer from a target, it was located at an artificial water tank just east of the southeastern
corner of target range. SPH used all of the range, but visual inspection makes apparent locations
where more sightings occur (Figure 9). The majority of hotspots from 2010 to 2014 are located
among four areas. These locations are more or less central within the range, including the
northernmost targets on STAC and the southernmost targets on NTAC. The maps with more
hotspot locations show that SPH were more spread out, and therefore there were more locations
where the highest 20 percent density values occurred. Fewer hotspots show less spreading, and
more clumping.
When the hotspots were examined on a seasonal time scale, the summer season stood out
from the breeding and fawning seasons in that the distance from these hotspots to the nearest
target was greater. This agrees with observations and sightings during the summer months,
because many SPH move southwest towards the food and water station, while others leave the
study area entirely. When the food station hotspots were removed from the analysis for each
48

49

season (they only occurred in the summer season of each year), the results of the ANOVA
determined there was no significant difference in distance from targets between seasons. This
emphasizes the effect of the food and water station on the SPH locations. During the breeding
and fawning seasons, the SPH have the same access to that location as during the summer
months, but choose to forage elsewhere. In this way, the food and water station provides a
resource that has a significant impact on SPH location during the summer season of each year.
Before this study, the biologists monitoring the SPH had not identified the influence of
roads on SPH sightings. Krausman et al. (2005b) examined roads and disturbance, though at a
much coarser scale, by marking a one kilometer block as either having a road or a target or
undisturbed. If there was a road anywhere within that block then it was given that attribute. Their
study showed a high use of roads by SPH, with 70 percent of the road blocks having SPH
sightings. These results show a high usage of road blocks by SPH, though I would argue that this
method is not useful because roads are interspersed practically everywhere throughout the range,
so it is difficult to get anywhere far from a road. Because of this, the one kilometer scale will
inaccurately show roads as being important. By looking at the roads layer at a closer scale in this
thesis, the effect of roads on SPH sightings is minimal, while the effect of targets remains high.
There are times when SPH are seen walking along roads, traveling or foraging, though not often.
The results from the Maxent analysis show that the roads have very little effect on the
pronghorn locations, in that when removed from the analysis, model performance did not
decrease. Gavin and Komers (2006) examined the effects of various intensities of human
disturbance on American pronghorn and found that while there did not appear to be any strong
preference towards or away from roads, there was increased vigilance and less foraging time
within groups near roads. The roads in this study consisted of public roads with average traffic
49

50

values estimated by the department of transportation between <10 to >300 vehicles per day.
They also noticed that larger groups were affected less in this way than individuals or small
groups. Observation distance had the same result within the model’s jackknife tests, though
when considering the percent contribution to the model, it takes a much larger percentage than
the road distance or the other variables. This is interesting because when examining the jackknife
test, they are almost exactly the same in AUC model performance values. Prior studies always
mention potential bias caused by observation locations (Krausman et al. (2005b), though none
look at observation distance or type of sighting as a way to mediate this bias.
Interestingly, slope appears to have the most negative effect on the model AUC when
removed, after target distance. Unfortunately, because most of the study area is relatively flat,
this is hard to corroborate through observation in the field. It is true that the SPH are rarely on
the hills themselves, but they are sometimes seen in between hills, where access is limited to
being on the hill’s steep slope at some point while traveling.
The third model results showed that while the predictive ability decreased slightly, the
variance present decreased greatly. This shows that there are differences between the telemetry
locations and the visual locations. SPH are seen more easily close to observers. The third model
included all sightings, which means that there were many groups that were not specified as
having been seen first, or whether their sighting was due to a telemetry animal being within the
group. The fact that a random 25 percent of the sightings were chosen to test with, I believe,
explains the lower AUC value because it is unknown which locations were used for training or
testing. Models four and five had higher AUC values and lower standard deviations because each
25 percent test dataset was taken from the same dataset as the 75 percent used in training the
algorithm. This is what I expected, and why I don’t think they are the best option for the primary
50

51

model. Observer bias is an issue, in this study and others, and should be taken into account. That
the models using only one type of sighting performed better, with less variation in runs, supports
the idea that the two sightings are different enough from each other to affect the model’s
performance. I believe that by using telemetry only sightings for training, and visual sightings for
testing, a portion of observer bias is accounted for, even though the model AUC is lower and the
standard variation is larger.
Future analyses of this type should explore the use of precipitation interpolation layers, as
well as detailed vegetation classifications for supplemental environmental variables. Because the
study area is small relative to the scale at which these variables are currently available, they were
not included in the analysis. The addition of a bias layer could potentially help as well, these
methods attempted to account for visibility bias by using telemetry only sightings for training
and visual only for testing. As observation points are positioned to provide maximum visibility
to targets, this is definitely an important consideration when interpreting the results. Vegetation
types also have different cover values, as small trees are harder to see through than short grass or
shrubs. Perhaps a layer built from the vegetation type could represent an index of visibility
within the model.
Other than the addition of new variables into the model, the time scale being analyzed
will definitely affect the predictive capacity of the output. If the SPH biologists and monitors are
interested in a prediction of sighting locations for the next month, the analysis should be run
using recent sightings only, to provide the level of variation required.
51

52

REFERENCES
Anderson, Tessa K. 2009. “Kernel Density Estimation and K-means Clustering to Profile Road
Accident Hotspots.” Accident Analysis and Prevention 41: 359-364.
Baldwin, Roger A. 2009. “Use of Maximum Entropy Modeling in Wildlife Research.” Entropy
11: 854-866.
Beyer, H. L. 2012. Geospatial Modelling Environment (Version 0.7.2.1). URL
http://www.spatialecology.com/gme.
Blundell, Gail M., Julie A. K. Maier, and Edward M. Debevec. 2001. “Linear Home Ranges:
Effects of Smoothing, Sample Size, and Autocorrelation on Kernel Estimates.”
Ecological Monographs 71: 469-489.
Bright, Jill L. and John J. Hervert. 2005. “Adult and Fawn Mortality of Sonoran Pronghorn.”
Wildlife Society Bulletin 33: 43-50.
deVos, James C. and William H. Miller. 2005. “Habitat Use and Survival of Sonoran Pronghorn
in Years with Above-average Rainfall.” Wildlife Society Bulletin 33: 35-42.
Downs, Joni A. 2010. “Time-geographic Density Estimation for Moving Point Objects.”
Geographic Information Science 6292: 16-26.
Downs, Joni A. and Mark W. Horner. 2008. “Effects of Point Pattern Shape on Home-range
Estimates.” Journal of Wildlife Management 72: 1813-1818.
Elith, Jane, Steven J. Phillips, Trevor Hastie, Miroslav Dudik, Yung E. Chee, and Colin J. Yates.
2011. “A Statistical Explanation of MaxEnt for Ecologists.” Diversity and Distributions
14: 43-57.
Er, A. C., M. H. Rosli, A. Asmahani, M. R. Mohamad Naim, and M. Harsuzilawati. 2010.
“Spatial Mapping of Dengue Incidence: A Case Study in Hulu Langat District, Selangor,
52

53

Malaysia.” International Journal of Environmental, Ecological, Geological and Mining
Engineering 4: 25-29.
ESRI. 2012. ArcGIS Desktop: Release 10.1. Redlands, CA: Environmental Systems Research
Institute.
Fox, L. M. “Nutritional content of forage in Sonoran pronghorn habitat, Arizona.” MS Thesis,
University of Arizona, Tucson, 1997.
Gavin, Shannon D. and P. E. Komers. 2006. “Do Pronghorn (Antilocapra americana) Percieve
Roads as a Predation Risk?” Canadian Journal of Zoology 84: 1775-1780.
Gu, Weidong and Robert K. Swihart. 2004. “Absent or Undetected? Effects of Non-detection of
Species Occurrence on Wildlife-habitat Models.” Biological Conservation 116: 195-203.
Hernandez, Pilar A., Catherine H. Graham, Lawrence L. Master, and Deborah L. Albert. 2006.
“The Effect of Sample Size and Species Characteristics on Performance of Different
Species Distribution Modeling Methods.” Ecography 29: 773-785.
Hervert, John J., Jill L. Bright, Robert S. Henry, Linden A. Piest, and Mark T. Brown. 2005.
“Home-range and Habitat-use Patterns of Sonoran Pronghorn in Arizona.” Wildlife
Society Bulletin 33: 8-15.
Hervert, John J., Linden A. Piest, W. Ballard, R. S. Henry, Mark T. Brown, and S. Boe. 1997.
“Sonoran Pronghorn Population Monitoring: Progress Report.” Arizona Game and Fish
Department, Technical Report, 126. Phoenix, AZ.
Hosack, Dennis A., P. S. Miller, John J. Hervert, and R. C. Lacy. 2002. “A Population Viability
Analysis for the Endangered Sonoran Pronghorn, Antilocapra americana sonoriensis.”
Mammalia 66: 207-229.
53

54

Kie, John G., Jason Matthiopoulos, John Fieberg, Roger A. Powell, Francesca Cagnacci, Michael
S. Mitchell, Jean-Michel Gaillard, and Paul R. Moorcroft. 2010. “The Home-range
Concept: are Traditional Estimators Still Relevant with Modern Telemetry Technology?”
Philosophical Transactions of the Royal Society B 365: 2221-2231.
Krausman, Paul R., John R. Morgart, Lisa K. Harris, Chantal S. O’Brien, James W. Cain III, and
Steven S. Rosenstock. 2005. “Introduction: Management for the Survival of Sonoran
Pronghorn in the United States.” Wildlife Society Bulletin 33: 5-7.
Krausman, Paul R., Lisa K. Harris, Sarah K. Haas, Kiana K. G. Koenen, Pat Devers, Daniel
Bunting, and Mark Barb. 2005. “Sonoran Pronghorn Habitat Use on Landscapes
Disturbed by Military Activities.” Wildlife Society Bulletin 33: 16-23.
Krausman, Paul R., Lisa K. Harris, and Jon Francine. 2001. “Noise Effects of Military
Overflights on Sonoran Pronghorn: Final Report.” The University of Arizona, Tucson,
AZ.
Krausman, Paul R., John R. Morgart, Lisa K. Harris, Chantal S. O’Brien, James W. Cain, III,
Steven S. Rosenstock. 2005. “Sonoran Pronghorn Literature: an Annotated
Bibliography.” USGS Open-File Report 2005-1265. U.S. Geological Survey, Southwest
Biological Science Center, Sonoran Desert Research Station, The University of Arizona,
Tucson, AZ.
Krisp, Jukka M. and Stefan Peters. 2011. “Directed Kernel Density Estimation (DKDE) for Time
Series Visualization.” Annals of GIS 17: 155-162.
Kumar, Sunil and Thomas J. Stohlgren. 2009. “Maxent Modeling for Predicting Suitable Habitat
for Threatened and Endangered Tree Canacomyrica monticola in New Caledonia.”
Journal of Ecology and Natural Environment 1: 94-98.
54

55

Lin, Yu-Pin, Hone-Jay Chu, Chen-Fa Wu, Tsun-Kuo Chang, and ChiuYang Chen. 2011.
“Hotspot Analysis of Spatial Environmental Pollutants Using Kernel Density Estimation
and Geostatistical Techniques.” International Journal of Environmental Research and
Public Health 8: 75-88.
Liu, Hua and Donald E. Brown. 2003. “Criminal Incident Prediction Using a Point-pattern-based
Density Model.” International Journal of Forecasting 19: 603-622.
Merow, Cory, Matthew J. Smith, and John A. Silander, Jr. 2013. “A Practical Guide to MaxEnt
for Modeling Species’ Distributions: What it Does, and Why Inputs and Settings Matter.”
Ecography 36: 1058-1069.
Mola-Yudego, Blas and David Gritten. 2010. “Determining Forest Conflict Hotspots According
to Academic and Environmental Groups.” Forest Policy and Economics 12: 575-580.
Moreno, Roberto, Ricardo Zamora, Juan R. Molina, Angelica Vasquez, and Miguel A. Harrera.
2011. “Predictive Modeling of Microhabitats for Endemic Birds in South Chilean
Temperate Forests Using Maximum Entropy (Maxent).” Ecological Informatics 6: 364-
370.
Morgart, John R., John J. Hervert, Paul R. Krausman, Jill L. Bright, and Robert S. Henry. 2005.
“Sonoran Pronghorn Use of Anthropogenic and Natural Water Sources.” Wildlife Society
Bulletin 33: 51-60.
Motsinger, J. 2013. Defenders of Wildlife Blog. “Road to Recovery: Sonoran Pronghorn.”
http://www.defendersblog.org/2013/07/road-to-recovery-sonoran-pronghorn (accessed 9
July 2014).
Nelson, Trisalyn A. and Barry Boots. 2008. “Detecting Spatial Hot Spots in Landscape
Ecology.” Ecography 35: 556-566.
55

56

Ochenfels, R. Pronghorn antelope. Arizona Antelope Foundation.
http://azantelope.org/Facts___Research/About_Pronghorn/about_pronghorn.html
(accessed 10 July 2014).
O’Gara, Bart W. 1978. “Antilocapra americana.” Mammalian Species 90: 1-7.
Phillips, Steven. 2005. "A Brief Tutorial on Maxent." ATandT Research.
Phillips, Steven J. and Miroslav Dudik. 2008. “Modeling of Species Distributions with Maxent:
New Extensions and a Comprehensive Evaluation.” Ecography 31:161-175.
Phillips, Steven J., Miroslav Dudik, and Robert E. Schapire. “A Maximum Entropy Approach to
Species Distribution Modeling.” In Proceedings of the 21st International Conference on
Machine Learning, 655-662. ACM Press, New York, 2004.
Phillips, Steven J., Robert P. Anderson, and Robert E. Schapire. 2006. “Maximum Entropy
Modeling of Species Geographic Distributions.” Ecological Modelling 190: 231-259.
Pronghorn Biologists. Sonoran Pronghorn Monitoring Protocol. Gila Bend and Barry M.
Goldwater Range Complex, Gila Bend, AZ, 2012.
R Development Core Team. 2011. R: A language and environment for statistical computing
(Version 3.1.0). R Foundation for Statistical Computing, Vienna, Austria. URL
http://www.R-project.org/.
Ramp, Daniel, Joanne Caldwell, Kathryn A. Edwards, David Warton, and David B. Croft. 2005.
“Modelling of Wildlife Fatality Hotspots Along the Snowy Mountain Highway in New
South Wales, Australia.” Biological Conservation 126: 474-490.
Reid, F. A. A field guide to mammals of North America, 500-501. New York, NY: Houghton
Mifflin Company, 2006.
56

57

Seaman, D. Erran, Joshua J. Millspaugh, Brian J. Kernohan, Gary C. Brundige, Kenneth J.
Raedeke, and Robert A. Gitzen. 1999. “Effects of Sample Size on Kernel Home Range
Estimates.” Journal of Wildlife Management 63: 739-747.
Seaman D. Erran and Roger A. Powell. 2005. “An Evaluation of the Accuracy of Kernel Density
Estimators for Home Range Analysis.” Ecology 77: 2075-2075.
Silverman, Bernard W. Density Estimation for Statistics and Data Analysis. London, UK:
Chapman and Hall, 1986.
Stephen, Catherine L., James C. Devos, Jr., Thomas E. Lee, Jr., John W. Bickham, James R.
Heffelfinger, and Olin E. Rhodes Jr. 2005. “Population Genetic Analysis of Sonoran
Pronghorn (Antilocapra americana sonoriensis).” Journal of Mammalogy 86: 782-792.
Suarez-Seoane, Susana, Eladio L. Garcia de la Morena, Manuel B. Morales Prieto, Patrick E.
Osborne, and Eduardo de Juana. 2008. “Maximum Entropy Niche-based Modelling of
Seasonal Changes in Little Bustard (Tetrax tetrax) Distribution.” Ecological Modelling
219: 17-29.
Tetra Tech, Inc. Ecological Factors that may Influence Sonoran Pronghorn: North and South
Tactical Ranges, Barry M. Goldwater Range, Luke AFB, Arizona. By Tetra Tech, Inc.
San Bernardino, CA, 2005.
Tinker, B. Mexican Wilderness and Wildlife. University of Texas Press, Austin, USA, 1978.
United States Fish and Wildlife Service. Final Revised Sonoran Pronghorn Recovery Plan.
Albuquerque, NM, 1998.
United States Fish and Wildlife Service. Fact Sheet: Sonoran Pronghorn. Albuquerque, NM,
2003.
57

58

United States Fish and Wildlife Service. Final Environmental Assessment for Reestablishment of
Sonoran Pronghorn. Albuquerque, NM, 2010.
Vander Wal, Eric and A. R. Rodgers. 2012. “An Individual-based Quantitative Approach for
Delineating Core Areas of Animal Space Use.” Ecological Modelling 224: 48-53.
Wand, M. P. and M. C. Jones. Kernel Smoothing. London, UK: Chapman and Hall, 1995.
Worton, B. J. 1989. “Kernel Methods for Estimating the Utilization Distribution in Home-range
Studies.” Ecology 70: 164-168.
Xie, Zhixiao and Jun Yan. 2008. “Kernel Density Estimation of Traffic Accidents in a Network
Space.” Computers, Environment and Urban Systems 32:396-406.
Yoakum, J. D. “Pronghorn.” In Big game of North America: ecology and management, edited by
J. L. Schmidt and D. L. Gilbert, 103-121. Washington, D.C.: Wildlife Management
Institute and Harrisburg, PA: Stackpole Books, 1978.
Yost, Andrew C., Steven L. Petersen, Michael Gregg, and Richard Miller. 2008. “Predictive
Modeling and Mapping Sage Grouse (Centrocercus urophasianus) Nesting Habitat Using
Maximum Entropy and a Long-term Dataset from Southern Oregon.” Ecological Informatics 3:
375-386.
58

59

APPENDIX A: HOTSPOT MAPS

Figure 27 SPH 2010 Hotspot Map with Underlying KDE Raster
59

60

Figure 28 SPH 2010 Breeding Season Hotspot Map with Underlying KDE Raster
60

61

Figure 29 SPH 2010 Fawning Season Hotspot Map with Underlying KDE Raster
61

62

Figure 30 SPH 2010 Summer Season Hotspot Map with Underlying KDE Raster
62

63

Figure 31 SPH 2011 Hotspot Map with Underlying KDE Raster
63

64

Figure 32 SPH 2011 Breeding Season Hotspot Map with Underlying KDE Raster
64

65

Figure 33 SPH 2011 Fawning Season Hotspot Map with Underlying KDE Raster
65

66

Figure 34 SPH 2011 Summer Season Hotspot Map with Underlying KDE Raster
66

67

Figure 35 SPH 2012 Hotspot Map with Underlying KDE Raster
67

68

Figure 36 SPH 2012 Breeding Season Hotspot Map with Underlying KDE Raster
68

69

Figure 37 SPH 2012 Fawning Season Hotspot Map with Underlying KDE Raster
69

70

Figure 38 SPH 2012 Summer Season Hotspot Map with Underlying KDE Raster
70

71

Figure 39 SPH 2013 Hotspot Map with Underlying KDE Raster
71

72

Figure 40 SPH 2013 Breeding Season Hotspot Map with Underlying KDE Raster
72

73

Figure 41 SPH 2013 Fawning Season Hotspot Map with Underlying KDE Raster
73

74

Figure 42 SPH 2013 Summer Season Hotspot Map with Underlying KDE Raster
74

75

APPENDIX B: SEASONAL HOTSPOT DISTANCE TABLE
Table 5 Seasonal Hotspots’ Distance to Nearest Target
Year Season Hotspot Distance
2010 breeding 1 66
2010 breeding 2 245
2010 fawning 1 190
2010 fawning 2 640
2010 fawning 3 475
2010 fawning 4 301
2010 summer 1 4942
2010 summer 2 495
2010 summer 3 255
2010 summer 4 309
2010 summer 5 543
2010 summer 6 176
2011 breeding 1 170
2011 fawning 1 48
2011 fawning 2 88
2011 fawning 3 325
2011 fawning 4 267
2011 fawning 5 449
2011 fawning 6 61
2011 summer 1 5015
2011 summer 2 419
2011 summer 3 173
2011 summer 4 1356
2011 summer 5 188
2012 breeding 1 168
2012 breeding 2 217
2012 breeding 3 71
2012 breeding 4 970
2012 breeding 5 875
2012 breeding 6 351
2012 breeding 7 355
2012 breeding 8 90
2012 breeding 9 2628
2012 fawning 1 84
2012 fawning 2 296
2012 fawning 3 865
2012 fawning 4 276
2012 summer 1 4992
2012 summer 2 1256
2012 summer 3 137
2012 summer 4 498
75

76

2012 summer 5 776
2012 summer 6 17
2013 breeding 1 600
2013 breeding 2 198
2013 breeding 3 527
2013 breeding 4 18
2013 breeding 5 247
2013 fawning 1 2141
2013 fawning 2 438
2013 fawning 3 415
2013 fawning 4 73
2013 fawning 5 378
2013 fawning 6 584
2013 fawning 7 611
2013 fawning 8 84
2013 fawning 9 101
2013 fawning 10 239
2013 fawning 11 66
2013 summer 1 4954
2013 summer 2 1084
2013 summer 3 108
2013 summer 4 512
2013 summer 5 309
2013 summer 6 664
2013 summer 7 260

76

Abstract (if available)

Abstract The population of Sonoran pronghorn (SPH

Linked assets

University of Southern California Dissertations and Theses

Conceptually similar

PDF

Spatial distribution of the Nile crocodile (Crocodylus niloticus) in the Mariarano River system, Northwestern Madagascar

PDF

Disparities in food access: an empirical analysis of neighborhoods in the Atlanta metropolitan statistical area

PDF

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

PDF

Historical temperature trends in Los Angeles County, California

PDF

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

PDF

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

PDF

Distribution and correlates of feral cat trapping permits in Los Angeles, California

PDF

Spatial distribution of the endangered Pacific pocket mouse (Perognathus ssp. pacificus) within coastal sage scrub habitat at Dana Point Headlands Conservation Area

PDF

Spatio-temporal analysis of Western snowy plover nesting at Vandenberg Air Force Base

PDF

A comparison of urban land cover change: a study of Pasadena and Inglewood, California, 1992‐2011

PDF

Effect of spatial patterns on sampling design performance in a vegetation map accuracy assessment

PDF

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

PDF

Using Maxent modeling to predict habitat of mountain pine beetle in response to climate change

PDF

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

PDF

An evaluation of soil sampling methods in support of precision agriculture in Northeastern North Carolina

PDF

A geospatial analysis of income level, food deserts and urban agriculture hot spots

PDF

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

PDF

The wildland-urban interface in Lassen County, California: a change analysis 2000–2015

PDF

Estimating populations at risk in data-poor environments: a geographically disaggregated analysis of Boko Haram terrorism 2009-2014

PDF

Modeling patient access to point-of-care diagnostic resources in a healthcare small-world network in rural Isaan, Thailand

Asset Metadata

Creator Price, Samuel C. (author)

Core Title Distribution of Sonoran pronghorn (Antilocapra americana sonoriensis) on an active Air Force tactical range

School College of Letters, Arts and Sciences

Degree Master of Science

Degree Program Geographic Information Science and Technology

Publication Date 02/10/2015

Defense Date 01/13/2015

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

Tag distribution modeling,hotspots,kernel density estimate,Maxent,OAI-PMH Harvest,Sonoran pronghorn,United States Air Force

Format application/pdf (imt)

Language English

Contributor Electronically uploaded by the author (provenance)

Advisor Longcore, Travis R. (committee chair), Lee, Su Jin (committee member), Ruddell, Darren M. (committee member)

Creator Email Samemi@gmail.com,samuelp@usc.edu

Permanent Link (DOI) https://doi.org/10.25549/usctheses-c3-531085

Unique identifier UC11298242

Identifier etd-PriceSamue-3172.pdf (filename),usctheses-c3-531085 (legacy record id)

Legacy Identifier etd-PriceSamue-3172.pdf

Dmrecord 531085

Document Type Thesis

Format application/pdf (imt)

Rights Price, Samuel C.

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