using occupancy models to predict grassland bird...
TRANSCRIPT
USING OCCUPANCY MODELS TO PREDICT GRASSLAND BIRD
DISTRIBUTIONS IN SOUTHEASTERN ALBERTA
A Thesis
Submitted to the Faculty of Graduate Studies and Research
In Partial Fulfillment of the Requirements
for the Degree of
Masters of Science
In
Biology
University of Regina
by
Nathan David Clements
Regina, Saskatchewan
October, 2014
Copyright 2014: N.D. Clements
UNIVERSITY OF REGINA
FACULTY OF GRADUATE STUDIES AND RESEARCH
SUPERVISORY AND EXAMINING COMMITTEE
Nathan David Clements, candidate for the degree of Master of Science in Biology, has presented a thesis titled, Using Occupancy Models to Predict Grassland Bird Distributions in Southeastern Alberta, in an oral examination held on August 29, 2014. The following committee members have found the thesis acceptable in form and content, and that the candidate demonstrated satisfactory knowledge of the subject material. External Examiner: Dr. Erin Bayne, University of Alberta
Co-Supervisor: Dr. Mark Brigham, Department of Biology
Co-Supervisor: Dr. Stephen Davis, Adjunct
Committee Member: Dr. Christopher Somers, Department of Biology
Chair of Defense: Dr. Katya Herman, Faculty of Kineiology & Health Studies
i
ABSTRACT
Widespread population declines of grassland birds have stimulated a variety of
conservation plans, many of which promote a landscape approach to conservation.
Identifying needs for survival and reproduction, and prioritizing key habitat requirements
that influence where species occur on the landscape is an essential first step to guiding
long term management efforts. I used single-species, single-season occupancy models to
(1) identify landscape factors influencing the distribution of grassland bird species, (2)
generate predictions of species distributions in southeastern Alberta, and (3) evaluate the
predictive performance of each species model with two sources of evaluation data. In
2012, I conducted three repeat surveys at each of 870 point count locations to generate
occupancy models, and evaluated the predictive performance of each model using a
partitioned data set from 2012 and an independent data set of 1398 point counts collected
in 2011. Records of detection/non-detection were collected for: Baird's Sparrow
(Ammodramus bairdii), Brown-headed Cowbird (Molothrus ater), Chestnut-collared
Longspur (Calcarius ornatus), Clay-colored Sparrow (Spizella pallida), Horned Lark
(Eremophila alpestris), Long-billed Curlew (Numenius americanus), Marbled Godwit
(Limosa fedoa), McCown's Longspur (Rhynchophanes mccownii), Savannah Sparrow
(Passerculus sandwichensis), Sprague's Pipit (Anthus spragueii), Upland Sandpiper
(Bartramia longicauda), Vesper Sparrow (Pooecetes gramineus), Western Meadowlark
(Sturnella neglecta), and Willet (Tringa semipalmata). Models that best explained
detectability typically included both date and time of day of the survey. The amount of
grassland on the landscape was a broadly useful measure in predicting occupancy.
Variables describing landscape topography and the geographic location (i.e., latitude and
ii
longitude) also aided in understanding where species occur and provided a more
complete explanation of resource selection. I show that the models generated for four
species were accurate predictors of occupancy. The remaining models require further
investigation to develop accurate predictions of occupancy, but are presented for
exploratory purposes and can be used to guide and refine future predictive modelling
efforts. I suggest that, depending on the application, these models and maps be used to
guide local conservation plans for birds, and can be used as a valuable reference for
prioritizing conservation activities in southeast Alberta. These models add to our
understanding of resource selection at the landscape level, and can assist in the process of
ensuring key habitats are identified and conserved.
iii
ACKNOWLEDGEMENTS
I would like to express my sincere appreciation to Dr. Stephen Davis whose
support and continual encouragement has significantly influenced my development as
both student of ecology, and more importantly, a student of life. I would also like to
thank my committee members, Drs. Mark Brigham and Christopher Somers for lending
their knowledge, time, and direction throughout this process. In addition, I would like to
thank the students in the Bird and Bat lab for the constant support, advice, and weekly
reminder that I was not alone in the pursuit of learning. Special thanks to Joe Poissant for
his help in improving my thesis. This has truly been a positive experience, and I cannot
thank them enough.
This project would not have been completed without the help of a great field
crew: S. James, J. Unruh, G. Foley, A. Stone, and especially D. Sawatzky. The early
morning commutes to transect, sunrises, click of the multi-coloured pens, and bird chatter
are the memories I will cherish most. I recognise that travel, intense point count surveys,
the nuances of data entry and field work in general can be demanding, and I would like to
thank my field crew for their commitment to this project. I thank the landowners in my
study area that provided access to over 1200 quarter sections of private land.
Both financial and in-kind support for this project was provided by: Canadian
Wildlife Federation, Alberta Conservation Association, University of Regina, and
Environment Canada - Canadian Wildlife Service. Without the aforementioned support,
the goals of this project would have been unattainable. The continuous support provided
by my employer, the Canadian Wildlife Federation, was extremely important from a
personal perspective; thank you.
iv
POST DEFENSE ACKNOWLEDGEMENT
I would like to thank my external examiner, Dr. Erin Bayne, for travelling to Regina for
my defense. His time and thoughtful insight were very helpful.
v
DEDICATION
I dedicate this thesis to my loving wife, Crystal, daughter, Aven, and son, Fisher.
They were unwavering in their continual encouragement and support for my endeavours.
The trade-offs associated with balancing education, work, and family life are difficult,
and I truly cannot underscore their importance for my research. All decisions have
consequences, and the cost of self-improvement by education was at the expense of our
time, which I can only hope to improve upon in our future.
I would also like to thank my family and friends for their support. This self-
fulfilling goal often led to me declining invites, and in most cases, you understood my
reasoning. To Mom and Dad, thank you. Growing up, I watched your every move and
observed the sacrifices you both made for your family. I understand more now than ever
just how difficult those decisions must have been. To Garnet and Gail, I feel truly
blessed to have you in my life and appreciate all the support you have shown to me and
my family.
It would be remiss of me if three of my mentors in ecology were not mentioned,
Paul Pryor, Stan Clements, and Glen Clements never passed up an opportunity to share
their knowledge. Thank you for helping foster my passion for conservation.
vi
TABLE OF CONTENTS
ABSTRACT ............................................................................................................................ I
ACKNOWLEDGEMENTS ............................................................................................... III
POST DEFENSE ACKNOWLEDGEMENT ................................................................. IV
DEDICATION ....................................................................................................................... V
TABLE OF CONTENTS ................................................................................................... VI
LIST OF TABLES .............................................................................................................. IX
LIST OF FIGURES ........................................................................................................... XII
LIST OF APPENDICES ................................................................................................... XV
LIST OF ABBREVIATIONS AND TERMINOLOGY ............................................. XVI
1. GENERAL INTRODUCTION .................................................................................... 1
1.1 Background ............................................................................................................... 1
1.2 Species Occurrence Records ................................................................................... 4
1.3 Accounting for Imperfect Detectability ................................................................ 5
1.4 Environmental Variables ........................................................................................ 8
1.5 Evaluation of Predictive Performance .................................................................. 9
1.6 Study Species ........................................................................................................... 10
1.7 Study Objectives ..................................................................................................... 13
1.8 Literature Cited ...................................................................................................... 14
2. PREDICTING GRASSLAND SONGBIRD DISTRIBUTION PATTERNS
USING OCCUPANCY MODELS ................................................................................. 30
vii
2.1 Introduction............................................................................................................. 30
2.2 Methodology ............................................................................................................ 33
2.2.1 Study Area ........................................................................................................ 33
2.2.2 Sampling Design and Data Collection, 2011 ............................................... 34
2.2.3 Sampling Design and Data Collection, 2012 ............................................... 36
2.2.4 GIS Methodology............................................................................................. 37
2.2.4.1 Explanatory Variables ............................................................................. 38
2.2.5 Statistical Analysis........................................................................................... 41
2.2.5.1 Determining Variable Scale and Collinearity ...................................... 41
2.2.5.2 Model Selection ......................................................................................... 42
2.2.5.3 Model Evaluation ..................................................................................... 43
2.2.5.4 Uninformative Parameters and Predictions......................................... 44
2.3 Results ...................................................................................................................... 45
2.3.1 General Results ................................................................................................ 45
2.3.2 Detection Probability ...................................................................................... 46
2.3.3 Occurrence Probability .................................................................................. 47
2.3.4 Model Evaluation ............................................................................................ 48
2.4 Discussion ................................................................................................................ 50
2.4.1 Accounting for Imperfect Detection ............................................................. 51
2.4.2 Resource Selection ........................................................................................... 53
2.4.2.1 Response to Cover Type .......................................................................... 54
2.4.2.2 Responses to Topographic Landscape Features .................................. 57
2.4.2.3 Influence of Geographic Location ......................................................... 59
viii
2.4.2.4 Influence of Well Density ........................................................................ 60
2.5 Conclusions .............................................................................................................. 61
2.6 Literature Cited ...................................................................................................... 63
3. CONCLUSIONS AND MANAGEMENT IMPLICATIONS .............................. 101
3.1 Summary................................................................................................................ 101
3.2 Management Implications................................................................................... 102
3.3 Literature Cited .................................................................................................... 104
ix
LIST OF TABLES
Table 2.1. Description of each variable used to build the species occupancy models with
the detection/non-detection avian survey records collected in the Special Areas,
2012. ........................................................................................................................... 87
Table 2.2. Pearson’s correlation coefficients for selected variables used to explain
species occurrence in the Special Areas, Alberta. Values are for the 400 m
landscape scale and were used to examine collinearity between predictor
variables. Variable codes are found in Table 2.1.................................................... 89
Table 2.3. Ranked species frequency of occurrence (%) based on the 2012 data used to
build species occupancy models, with mean occurrence rates (Ψ) and detection
probabilities (ρ) from surveys conducted in the Special Areas, Alberta. ............... 90
Table 2.4. Top-ranking models explaining the influence of spatial and temporal variables
on species detection (ρ) and occurrence (Ψ) in the Special Areas, 2012. Models
include a detection component (date [JDATE], time [TIME], wind [WIND]) and
occurrence component (grassland [GRASS], woody cover [WOODY], vegetative
greenness [MSAVI], surface water [WATER], solar radiation [HLI], wetness
[CTI], latitude [YLAT], longitude [XLONG], and well density [WD]). Only
those models with a cumulative Akaike Information Criterion (AIC) weight (wi)
of 0.95 for each species with corresponding number of parameters (K), value of
the negative log-likelihood function (LogLik), and the change in AIC value from
the top model (∆AIC) are presented. ........................................................................ 91
x
Table 2.5. Coefficient estimates (β) and unconditional standard errors (SE) for intercepts
and variables used to predict species occupancy in the Special Areas, 2012.
Variables used to explain the detection probability (ρ) and occurrence probability
(Ψ) are presented separately and zeros indicate that a covariate was considered but
not included in the final model. Species abbreviations are BAIS (Baird’s
Sparrow), BHCO (Brown-headed Cowbird), CCLO (Chestnut-collared
Longspur), CCSP (Clay-colored Sparrow), HOLA (Horned Lark), LBCU (Long
billed Curlew), MAGO (Marbled Godwit), MCLO (McCown’s Longspur), SAVS
(Savannah Sparrow), SPPI (Sprague’s Pipit), UPSA (Upland Sandpiper), VESP
(Vesper Sparrow), WEME (Western Meadowlark), and WILL (Willet). ............. 94
Table 2.6. Comparison of area under the curve (AUC) scores from the 2012 partitioned
testing data (n = 218). Testing data which were not adjusted to account for
imperfect detection are indicated by Unadjusted, and training data which accounts
for imperfect detection are indicated by Adjusted. Upper (Cl+) and lower (Cl-)
confidence limits are indicated. ................................................................................ 98
Table 2.7. Predictive performance of occupancy models used to estimate occurrence
probabilities of 14 grassland bird species within the Special Areas, Alberta.
Predictive performance was assessed using two sources of evaluation data, the
partitioned 2012 data (n=218) and the independent sample collected in 2011
(n=1498). Performance was measured based on area under the curve (AUC)
scores and Kappa qualitative scores. Frequency of occurrence, or observed
prevalence of species are indicated for each data set along with 95% confidence
limits (CI) for both measures. ................................................................................... 99
xi
Appendix A. Table A.1. Variable shrinkage estimates when model averaging was
incorporated to assist with model selection uncertainty in the Special Areas, 2012.
The relative variable importance values are on a scale from 0 to 1, giving an equal
weight in the global model used to predict species occupancy. Positive and
negative species-habitat relationships are indicated (+/-). Variables used to
explain the detection component, ρ, are (date [JDATE], time [TIME], wind
[WIND]) and variables used to explain the occurrence component, Ψ, are
(grassland [GRASS], woody cover [WOODY], vegetative greenness [MSAVI],
surface water [WATER], solar radiation [HLI], wetness [CTI], latitude [YLAT],
longitude [XLONG], and well density [WD]). ...................................................... 108
xii
LIST OF FIGURES
Figure 2.1. Geographic location of the Special Areas municipality (shaded grey) in
south-eastern Alberta, Canada. ................................................................................. 85
Figure 2.2. Point count survey locations conducted in the Special Areas, 2011-2012.
Black circles indicate 2012 survey locations (visited using a repeated survey
design) and grey circles indicate 2011 survey location (visited using a removal
sampling design). ....................................................................................................... 86
Figure 2.3. Species relationships between probability of detection (ρ) and date within the
Special Areas of Alberta, 2012. Each figure is based on the same standardised
scale of mean = 0 and variance = 1, where negative values represents survey
conducted earlier in the survey season and positive values represent surveys
conducted later in the survey season. The 95% confidence intervals are indicated
by the shaded grey area. ............................................................................................ 93
Figure 2.4. Species relationships between probability of occurrence (Ψ) and modified
Soil-adjusted Vegetation Index within a 400 m radius of the point count centre for
surveys conducted within the Special Areas of Alberta, 2012. Each figure is
based on the same standardised scale of mean = 0 and variance = 1, where
negative values indicate more bare ground and positive values indicate more
green vegetative biomass. The 95% confidence intervals are indicated by the
shaded grey area......................................................................................................... 95
xiii
Figure 2.5. Species relationships between probability of occurrence (Ψ) and Compound
Topographic Index within a 400 m radius of the point count centre for surveys
conducted within the Special Areas of Alberta, 2012. Each figure is based on the
same standardised scale of mean = 0 and variance = 1, where negative values
indicate more mesic areas and positive values indicate more xeric areas on the
landscape. The 95% confidence intervals are indicated by the shaded grey area.96
Figure 2.6. Species relationships between probability of occurrence (Ψ) and Heat Load
Index within a 400 m radius of the point count centre for surveys conducted
within the Special Areas of Alberta, 2012. Each figure is based on the same
standardised scale of mean = 0 and variance = 1, where negative values represent
less solar radiation and positive values represent higher rates of solar radiation on
the landscape. The 95% confidence intervals are indicated by the shaded grey
area.............................................................................................................................. 97
Appendix B. Figure B.1. Baird's Sparrow (Ammodramus bairdii) expected probability of
occurrence for the Special Areas, Alberta.............................................................. 109
Appendix B. Figure B.2. Brown-headed Cowbird (Molothrus ater) expected probability
of occurrence for the Special Areas, Alberta. ........................................................ 110
Appendix B. Figure B.3. Chestnut-collard Longspur (Calcarius ornatus) expected
probability of occurrence for the Special Areas, Alberta. ..................................... 111
Appendix B. Figure B.4. Clay-colored Sparrow (Spizella pallida) expected probability of
occurrence for the Special Areas, Alberta.............................................................. 112
Appendix B. Figure B.5. Horned Lark (Eremophila alpestris) expected probability of
occurrence for the Special Areas, Alberta.............................................................. 113
xiv
Appendix B. Figure B.6. Long-billed Curlew (Numenius americanus) expected
probability of occurrence for the Special Areas, Alberta. ..................................... 114
Appendix B. Figure B.7. Marbled Godwit (Limosa fedoa) expected probability of
occurrence for the Special Areas, Alberta.............................................................. 115
Appendix B. Figure B.8. McCown's Longspur (Rhynchophanes mccownii) expected
probability of occurrence for the Special Areas, Alberta. ..................................... 116
Appendix B. Figure B.9. Savannah Sparrow (Passerculus sandwichensis) expected
probability of occurrence for the Special Areas, Alberta. ..................................... 117
Appendix B. Figure B.10. Sprague's Pipit (Anthus spragueii) expected probability of
occurrence for the Special Areas, Alberta.............................................................. 118
Appendix B. Figure B.11. Upland Sandpiper (Bartramia longicauda) expected
probability of occurrence for the Special Areas, Alberta. ..................................... 119
Appendix B. Figure B.12. Vesper Sparrow (Pooecetes gramineus) expected probability
of occurrence for the Special Areas, Alberta. ........................................................ 120
Appendix B. Figure B.13. Western Meadowlark (Sturnella neglecta) expected
probability of occurrence for the Special Areas, Alberta. ..................................... 121
Appendix B. Figure B.14. Willet (Tringa semipalmata) expected probability of
occurrence for the Special Areas, Alberta.............................................................. 122
xv
LIST OF APPENDICES
APPENDIX A – Relative variable importance for grassland bird occupancy models
generated in the Special Areas, Alberta. ................................................................ 107
APPENDIX B – Probability of occurrence maps for bird species in the Special Areas,
Alberta. ..................................................................................................................... 109
xvi
LIST OF ABBREVIATIONS AND TERMINOLOGY
BIAS – difference between the expected value of an estimator and true value of the
parameter.
DETECTION PROBABILITY– relates to the probability a target species is present at the
time of the survey and detected at a site.
OCCUPANCY – Parameter of interest in ecological research (also referred to as
occurrence). The proportion of sites occupied calculated from the two processes
occurring in a general sampling situation (occurrence and detection probability).
OCCURRENCE PROBABILITY – relates to the probability of sites used by a target
species during the sampling period.
1
1. GENERAL INTRODUCTION
1.1 Background
All organisms require adequate quantities of usable resources to survive.
Understanding that animals select an area to live based on such resources as food, water,
and shelter have led conservation practitioners to develop a number of ways to identify
and prioritise areas that meet a given species’ basic requirements. Determining which
resources are selected at a higher propensity than others is of particular interest because it
may provide insight into species requirements for survival and reproduction (Manly et al.
2002; MacKenzie et al. 2006). By quantifying point count data with geographic and
environmental variables across the landscape, results have shown to be linked to
reproductive success and environmental suitability (Pearce and Ferrier 2001; Bock and
Jones 2004; Peterson 2006). Species Distribution Models (SDMs) typically combine
species occurrence or abundance data with geographical and environmental variables to
gain ecological insight about how species interact with the composition and spatial
arrangements of habitats, and thus to predict distributions across landscapes (Gusian and
Thuiller 2005; Elith and Leathwick 2009; Franklin 2010). SDMs are a commonly used
tool to describe and delineate the geographical distributions of flora and fauna (Guisan
and Thuiller 2005). These models are used in a wide variety of applications, including
assessing the impacts of land cover change, guiding and refining future research efforts,
projecting the impacts of climate change, supporting conservation prioritization and
reserve selection, and guiding species reintroduction (Elith and Leathwick 2009; Franklin
2009).
The emergence of qualitative mapping and distribution modelling has been
associated with advances in statistical techniques and geographic information technology
2
(Elith et al. 2006; Franklin 2010). Studies of species-habitat relationships combined with
advances in statistical analysis, such as generalised linear models, provide opportunities
to explain species-habitat relationships (e.g., Austin 1985; Guisan and Zimmermann
2000). Concurrently, technological advances in physical geography have provided new
spatial data including digital models of surface elevation, climate patterns, and remote
sensing of the surface condition of the earth. Geographic Information Systems (GIS)
have provided conservation practitioners with a platform to efficiently manipulate and
store spatial data (e.g., Laurent et al. 2005; Shirley et al. 2013). SDMs have been used
extensively over the past decade and are considered an essential tool in the fields of
ecology and conservation biology (Guisan and Thuiller 2005; Elith et al. 2006; Elith and
Leathwick 2009; Franklin 2010).
Species distribution patterns can be portrayed in relatively simple means, such as
habitat suitability indices and species range maps, which are practical tools used to
identify where a species is likely to occur. However, both tools are often criticised due to
limitations of statistical rigour (Boyce et al. 2002). Alternatively, species distribution
patterns may be developed using ecological niche theory, which is based on the idea that
each species has a unique set of requirements that must be provided by the habitat(s) to
persist (Hutchison 1957). Conservation practitioners use SDMs to identify key habitat
variables that species require, which in turn, can be used to predict species occurrence or
abundance for a given area. Accurate predictions are essential for identifying suitable
habitat for species, conservation planning/management, and forecasting biological
consequences of global change (Guisan and Thuiller 2005; Collinge 2009; Franklin
2009).
3
Although SDMs are considered a proficient tool for coalescing species-habitat
relationships, species responses to environmental gradients are often made with certain
assumptions. The primary assumptions of SDMs are that species are in equilibrium with
their environment, and that adequate sampling of all environmental gradients has been
conducted (Guisan and Thuiller 2005; Elith and Leathwick 2009). Many SDMs assume
that a species ecological niche is stable in both space and time, and in most cases,
researchers cannot directly sample the entire area of interest. Therefore, using
distribution models to predict species occurrence in an unbalanced ecosystem setting
(e.g., disease, invasions, or climate change), can be misleading when models do not
incorporate new combinations of environmental variables or responses to environmental
factors outside their original gradient (Elith and Leathwick 2009). Additional
complications with SDMs include: 1) environmental factors, or combinations thereof,
that may limit distributions or biotic interactions, 2) outcomes influenced by genetic
variability and evolutionary changes, and 3) dispersal pathways that are difficult to
predict (Midgley et al. 2006; Dormann 2007; De Marco et al. 2008). However, SDMs
are considered one of the best approaches for predicting species distributions when
assumptions are valid and results are rigorously assessed (Elith and Leathwick 2009;
Franklin 2010).
Generally, the following steps are taken to construct a SDM once the study area
has been defined (Guisan and Thuiller 2005; Elith and Leathwick 2009): (1) collect
sufficient records of known species occurrence and non-occurrence or abundance, (2)
extract values of environmental variables presumed to impact occurrence or abundance
for each site or area of interest, (3) use the environmental variables to fit an appropriate
4
modelling algorithm used to predict the likelihood of species occurrence or abundance,
(4) use the model to predict the likelihood of occurrence or abundance across an area of
interest, (5) evaluate the model for realism and uncertainty, and (6) evaluate the models
fitted response using independent or partitioned testing data to evaluate model predictions
to determine usefulness.
1.2 Species Occurrence Records
A common method of counting organisms is by surveying an area using point
counts (Hutto et al. 1986). Point count surveys involve surveyors standing at a particular
site and recording the occurrence or number of individuals within a given radius of the
observer for a set period of time (Johnson 1995). Point counts provide useful baseline or
“snapshot” information about local biodiversity and can lead to a better understanding of
ecosystem health (Gregory et al. 2005). However, counting organisms can be difficult, as
some organisms are mobile, some can be more cryptic or inhabit areas in low-density,
and others may avoid detection all together (Elphick 2008). Counting can be influenced
by many factors including the probability an observer detects an individual (MacKenzie
et al. 2002) and within-season dispersal after breeding failure (Betts et al. 2008).
Additionally, most counts have other associated errors such as portions of the population
being unavailable for detection, detection mistakes, sound transmission, or erroneous
counts (Elphick 2008; Lozier et al. 2009).
Primarily, there are two types of occurrence data used to develop SDMs;
presence-only and presence-absence, or more properly known as detection/non-detection
when there is a chance individuals are present but not detected (MacKenzie et al. 2006).
5
Each method performs better with more detection records (Wisz et al. 2008), and
detection/non-detection records that are unbiased and error free (Brotons et al. 2004;
Graham et al. 2007). In cases where species are more cryptic or occur less frequently, the
focus should be on improving the quality of detection/non-detection data (MacKenzie et
al. 2006; Lobo 2008). Presence-only records describe known detections, but lack
information about non-detections. Examples of presence-only data are those collected by
radio telemetry or historical records such as those available from museum and herbarium
collections. Despite the constraints and criticisms, presence-only data are effective when
used to map species distributions, and continue to receive much attention (Pearce and
Boyce 2006). Although some have acknowledged that predictions would be more robust
if non-detection data were included, most suggest the lack of systematic survey data
should not devalue presence-only data (Phillips et al. 2009). The advantage of
detection/non-detection data is that it conveys more information about each surveyed
location, allowing the examination of bias and species prevalence given the survey effort
(Brotons et al. 2004; Phillips et al. 2009). Non-detection data are also sometimes viewed
as misleading because some species are not easily detected, which can be especially
problematic for mobile species. Consequently, statistical advances have led to modelling
techniques that account for imperfect detection (hereafter, occupancy models; MacKenzie
et al. 2002; 2006). There is a growing acceptance that accounting for imperfect detection
using count data is essential to maximizing the predictive performance of SDMs
(MacKenzie 2006; Lobo et al. 2008).
1.3 Accounting for Imperfect Detectability
6
Virtually all detection/non-detection data can be affected by false absences
(species present but not detected) because they do not account for the possibility that a
species is present but may not elicit a cue for detection, or goes undetected by the
observer (MacKenzie et al. 2002; Wintle et al. 2004; Rocchini et al. 2011; Kery 2012).
When detection/non-detection data that do not account for imperfect detection are
analysed using logistic regression, a commonly used approach for modelling species
occurrence (Guisan and Zimmermann 2000), it is expected to yield biased results (Tyre et
al. 2003; Gu and Swihart 2004). Gu and Swihart (2004) determined that logistic
regression models were sensitive to low levels of imperfect detection. Logistic
regression models the relationship between habitat and species occurrence (a combination
of occurrence and detection probability), just not where a species might be (occurrence)
(MacKenzie et al. 2006). For example, MacKenzie (2006) compared modelling results of
similarly fitted resource selection probability functions and occupancy models. The
results showed that imperfectly detecting a species can lead to erroneous conclusions
about resource use. As such, failure to account for imperfect detections in count data can
lead to biased estimates of habitat relations, inappropriate management decisions, and
substantially change the predictive ability of a SDM (MacKenzie 2006; Kery and
Schmidt 2008; Marsh and Trenham 2008).
Methods used to account for detection errors include distance sampling (Buckland
2001), double-observer sampling (Nichols et al. 2000), removal sampling (Farnsworth et
al. 2002), and repeated count sampling (MacKenzie et al. 2002; 2006). These
methodologies all aim to minimise imperfect detection but have associated trade-offs
including cost, survey effort, and the type of detection biases addressed. Removal
7
sampling (specifically in the context of avian species), assumes that singing frequency
influences the probability of detection when birds are surveyed (Farnsworth et al. 2002).
Removal sampling is widely used because it is considered an efficient way to collect data
over large areas (Lele et al. 2012), but there are considerable arguments about bias
(Johnson 2008). The main assumption of removal sampling pertains to it being done on a
closed population. This means that immigration and emigration do not occur during the
sampling period, and that all species of interest are available for detection within the
count period (Johnson 2008). However, birds tend to be mobile and sing in irregular
bouts which may not satisfy the assumptions. Both distance and double observer
sampling address the probability that an individual is missed given it is present during the
survey. However, unlike removal sampling, neither approach effectively deals with the
probability that an individual elicits a cue during the counting period (Elphick 2008).
Repeated counts are often used to provide a robust evaluation of imperfect
detection, and are considered an optimal method to determine occupancy (MacKenzie
2005; MacKenzie et al. 2006; Bailey et al. 2007). The general assumption of repeated
surveys is that each count is conducted on a closed population (MacKenzie et al. 2006).
This means occurrence probability remains constant for the duration of the repeated
counts. Additional assumptions include: (1) the target species is identified correctly; (2)
occurrence and detection probabilities are similar across sites, or heterogeneity can be
accounted for using covariates; and (3) detection of species is independent for each site
visit (MacKenzie et al. 2006). Assuming true occupancy status does not change
throughout the sampling period, changes in detection and non-detection can be attributed
8
to detectability. Thus, a robust estimation of occurrence can be used to estimate species
occupancy, once detection biases have been accounted for.
There are three specific survey designs commonly employed to conduct repeated
counts; 1) “standard” design where each counting site is sampled on numerous occasions,
2) “double sampling” design where repeated surveys are conducted at a subset of
counting sites, and 3) “removal sampling” design where each counting site is sampled a
maximum amount of times, but once the target species is detected, no further surveys are
required (Mackenzie et al. 2006). A common criticism of repeated surveys is the
associated trade-off between feasibility and survey effort (MacKenzie et al. 2006; Lele et
al. 2012). Assessing the trade-offs between survey methodologies, such as increasing the
number of repeated site visits at the expense of decreasing the amount of sampling sites,
is an important consideration (MacKenzie and Royle 2005). On the other hand,
understanding the biology of a target species can provide further insight about the trade-
offs of an optimal design structure (MacKenzie et al. 2006; Bailey et al. 2007). A general
rule of thumb used is that rare species require more sampling sites and common species
require increased number of repeated surveys (MacKenzie et al. 2006). The removal
design is likely to be more efficient if detection probability is relatively constant;
however the standard method provides more flexibility when detection probability are
unknown, and multiple species are targeted for study (MacKenzie et al. 2006).
1.4 Environmental Variables
Early applications of SDMs concentrated on explaining the ecological drivers of
species distributions (Mac Nally 2000), whereas much of the current focus is to generate
9
predictions about distribution over space and time (Austin 2007; Elith and Leathwick
2009). Moreover, with the growing complexity of modelling algorithms, greater
availability of spatial data, and the demand for mapped conservation products, an
increasing number of authors have focused on predicting species distributions within a
given area (Elith and Leathwick 2009; Franklin 2009). Motivated by the availability of
spatial variables and the idea that a model will identify important variables to predict
distributions, some studies include many candidate predictors (Austin 2002; Franklin
2009). However, many investigators advise that variable selection be based on the
known or perceived drivers of ecological processes (Austin 2002; Cunningham and
Johnson 2006; Elith et al. 2006; Elith and Leathwick 2009; Lobo et al. 2008; Franklin
2009), as both the variable selection process and the resulting inferences will be
improved. While it appears logical to assess all variables that either indirectly (distal) or
directly (proximal) relate to a particular resource, Austin (2002) argued that the use of
indirect variables has a tendency to be more error prone over time and space compared
with directly correlated predictors. As such, indirect predictors such as elevation and
slope are often replaced with more functionally relevant proximal predictors such as
temperature, rainfall, water availability, soil moisture and solar radiation (Elith and
Leathwick 2009).
1.5 Evaluation of Predictive Performance
The application of a model to predict species distributions relies heavily on its
predictive accuracy. Model evaluations are necessary to test the predictive performance
of SDMs, and are essential to provide confidence in a distribution models predictive
10
ability (Fielding and Bell 1997; Guisan and Zimmermann 2000; Vaughan and Ormerod
2005; Barry and Elith 2006). In fact, ecological modelling may have little relevance if
the predictions are not assessed for accuracy (Verbyla and Litaitis 1989). Evaluations of
detection/non-detection data are limited to validation techniques such as partitioning,
resampling schemes, or tests with independent datasets (Araujo et al. 2005). However,
reliable independent validation data can be difficult and costly to obtain and most often
contain some level of inherent bias or uncertainty (Pearson et al. 2007). In cases where
independent data are unavailable, a dataset can be split into datasets used to build (or
train) and validate (test) models (Fielding and Bell 1997).
When prediction is the aim, model evaluation involves statistical analyses that
assess predictive performance (Liu et al. 2011). By comparing the models predictions
against reality, a confusion matrix can be built to determine false presence and false
absence errors (Fielding and Bell 1997), both of which assist in determining predictive
performance. Further evaluations of predictive performance are based on a small number
of statistical tests including area under the receiver operating characteristic curve, Kappa,
and correlation coefficients (Fielding and Bell 1997; Allouche et al. 2006; Liu et al.
2011). Once a species-habitat model has provided evidence of predictive ability, it can
be used to guide future management strategies, justify funding for research and
conservation activities, prioritise conservation activities and assess species status for
listing under endangered species legislation (Aldridge and Boyce 2007; COSEWIC
2011).
1.6 Study Species
11
Concern over the continent-wide population declines of grassland birds has
stimulated a variety of conservation plans, many of which either explicitly or implicitly
promote a landscape approach to conservation (Berlanga et al. 2010). Much of
conservation theory has been focused on species as the unit of interest, however a
landscape approach provides tools and concepts for managing lands to achieve social,
economic, and environmental objectives across a given area (Berlanga et al. 2010).
Current population trend data for grassland birds suggest that the causes of declines are
not locally isolated phenomena but involve the cumulative loss and degradation of
grassland habitat throughout the Americas (Vickery et al. 1999; Askins et al. 2007).
Approximately 85% of the world’s temperate grassland biome has disappeared since the
late 1800’s (Millennium Ecosystems Assessment 2005), and within Canada, only an
estimated 20% of native grassland remains (White et al. 2000). This substantial rate of
loss has reduced the capacity, or ability of natural areas to provide vital ecosystem
services (Millennium Ecosystem Assessment 2005). Therefore, efforts to either stabilise
or increase local grassland bird populations may only be possible once there is a clear
understanding of resource selection and species occupancy in a given area (Dunning et al
1995; Vickery et al. 1999; Brennan and Kuvlesky 2005).
Identifying influential drivers of habitat selection across the landscape has a
strong history in theoretical ecology (MacArthur 1960; Austin 2002) and is essential to
wildlife management (Leopold 1933; Franklin 2009). Knowledge of grassland bird
distributions on the landscape is important for conservation because birds tend to exhibit
different behaviour patterns, sometimes even at different scales, that likely influence both
occurrence and detection (Hutto 1985; Donovan et al. 2002; Chandler et al. 2009). There
12
is no one single scale that best measures ecological patterns (Levin 1992), thus the
appropriate scale is usually dictated by the objectives of the study, the ecosystem being
studied, and available data (Elith and Leathwick 2009). With birds, it is presumed that
habitat selection first occurs at broad geographic scales which are suggested to be
genetically linked, whereas selection at finer scales is likely a combination of learned
behaviour and the availability of food and shelter (Wiens 1972; Johnson 1980; Hutto
1985; Wiens 1989). Therefore, selection begins at the scale of the geographic range of
the species, then occurs at broad landscape scales for populations, followed by
establishing home ranges within regions containing a population, and then at a finer scale
to satisfy individual life history requirements (Johnson 1980). Consequently,
investigators often examine species responses to landscape scales based on the concept of
a selection orders to measure how species interact with spatial habitat arrangements on
the landscape (Addicott et al. 1987; Askins et al. 2007), and provide a context for species
requirements to survive and reproduce.
Much of the initial research on breeding bird responses to landscape cover
focused on species inhabiting forests (Robbins et al. 1989; Terborgh 1989; Finch 1991).
Recent studies of grassland birds suggest responses to landscape factors at scales from
200 to 1600 m (Ribic and Sample 2001; Bakker et al. 2002; Brontons et al. 2005;
Cunningham and Johnson 2006; Ribic et al. 2009). For example, Ribic and Sample
(2001) established that responses to landscape scales at both 200 m and 400 m radius
could be used to explain the distribution of Grasshopper Sparrow (Ammodramus
savannarum). Davis et al. (2013) found that landscapes within 400 m of grassland
parcels influenced abundance of grassland songbirds, but pointed out that the precise
13
scale is difficult to determine because of the strong correlations between landscape scales
(400 m – 1600 m radius) within the region.
Even with considerable research devoted to understanding habitat selection, it is a
poorly understood process (Jones 2001). No single measure can describe the many
factors that influence habitat selection, however Cunningham and Johnson (2011) suggest
that identifying a broadly influential measure such as the amount of a given land cover
type would aid analysis and reduce some of the confusion in interpreting overlapping
predictor variables. Calculating percent cover assumes that a targeted cover type can
accurately be defined as useful to a species and that this level of cover can be reliably
identified from spatial imagery or maps (Cunningham and Johnson 2011). For example,
Davis et al. (2013) found that abundance of two specialist species, Spragues’s Pipit
(Anthus spragueii) and Baird’s Sparrow (Ammodramus bairdii), were similarly
influenced by the amount of native grassland on the landscape between two sampling
seasons. However, generalist species such as Savannah Sparrow (Passerculus
sandwichensis) and Western Meadowlark (Sturnella neglecta) exhibited a mixed
response to the amount of habitat and landscape cover over the two seasons. Although
percent cover estimates from spatial data is not an exact representation of habitat in the
landscape, it has proven to be reliable and an easy means to conceptualise for
conservation managers relative to other measures (Cunningham and Johnson 2011).
1.7 Study Objectives
Distribution models provide conservation managers with GIS products (e.g., maps
and shape files) of where species typically occur in a given area, which in turn, provide a
14
context for species requirements to survive and reproduce (Manly et al. 2002; MacKenzie
et al. 2006), and reduce uncertainty in conservation planning (Elith and Leathwick 2009).
By modelling grassland bird distributions, my research will lead to a better understanding
of resource selection, and assist in the process of ensuring key habitats are identified and
conserved (Askins et al. 2007). My objectives are to: (1) determine landscape factors that
influence the distribution of grassland birds, (2) use occupancy modelling to generate
Species Distribution Models (SDMs) and predict probability of occurrence for grassland
birds in southeastern Alberta, and (3) evaluate the predictive performance of the SDM
models.
1.8 Literature Cited
ADDICOTT, J. F., J. M. AHO, M. F. ANTOLIN, D. K. PADILLA, J. S. RICHARDSON, AND
D. A. SOLUK. 1987. Ecological neighborhoods: scaling environmental patterns.
Oikos 49: 340-346.
ALDRIDGE, C. L. AND M. S. BOYCE. 2007. Linking occurrence and fitness to persistence:
habitat-based approach for endangered greater sage-grouse. Ecological
Applications 2: 508-526.
ALLOUCHE, O., A. TSOAR, R. KADMON. 2006. Assessing the accuracy of species
distribution models: prevalence, kappa, and the true skill statistic (TSS). Journal
of Applied Ecology 43(6): 1223–1232.
15
ARAUJO, M. B., R. G. PEARSON, W. THUILLERS, AND M. ERHARD. 2005. Validation of
species-climate impact models under climate change. Global Change Biology 11:
1504-1513.
ASKINS, R. A., F. CHAVEZ-RAMIEREZ, B. C. DALE, C. A. HAAS, J. R. HERKERT, F. L.
KNOPF, AND P. D. VICKERY. 2007. Conservation of grassland birds in North
America: understanding ecological processes in different regions. Ornithological
Monographs 64: 1-46.
AUSTIN, M. P. 1985. Continuum concept, ordination methods and niche theory. Annual
Review of Ecological Systems 16: 39–61.
AUSTIN, M. P. 2002. Spatial prediction of species distribution: an interface between
ecological theory and statistical modelling. Ecological Modelling 157: 101-18.
AUSTIN, M. P. 2007. Species distribution models and ecological theory: a critical
assessment and some possible new approaches. Ecological Modelling 200(1–2):
1–19.
BAILEY, L. L., J. E. HINES, J. D. NICHOLS, AND D. I. MACKENZIE. 2007. Sampling
design trade-offs in occupancy studies with imperfect detection: examples and
software. Ecological Applications 17(1): 281-290.
16
BAKKER, K. K., D. E. NAUGLE, AND K. F. HIGGINS. 2002. Incorporating landscape
attributes into models for migratory grassland bird conservation. Conservation
Biology 16: 1638-1646.
BARRY, S., AND J. ELITH. 2006. Error and uncertainty in habitat models. Journal of
Applied Ecology 43(3): 413-423.
BERLANGA, H. K., J. A. RICH, T. D. ARIZMENDI, M. C. BEARDMORE, C. J. BLANCHER,
P. J. BUTCHER, G. S. COUTURIER, A. R. DAYER, A. A. DEMAREST, D. W.
EASTON, W. E. GUSTAFSON, M. INIGO-ELIAS, E. KREBS, E. A. PANJABI, A. O.
RODRIGUEZ CONTRERAS, V. ROSENBERG, K. V. RUTH, J. M. SANTANA
CASTELLON, E. VIDAL, R. MA, AND T. WILL. 2010. Saving Our Shared Birds:
Partners in Flight Tri-National Vision for Landbird Conservation. Cornell Lab of
Ornithology: Ithaca, NY. Available at: www.savingoursharedbirds.org. Accessed
online October 5, 2012.
BETTS, M. G., A. S. HADLEY, N. RODENHOUSE, AND J. J. NOCERA. 2008. Social
information trumps vegetation structure in breeding-site selection by a migrant
songbird. Proceedings of the Royal Society B 275: 2257-2263.
BOCK, C. E., AND Z. F. JONES. 2004. Avian habitat evaluation: should counting birds
count? Frontiers in Ecology and the Environment 2(8): 403-410.
17
BOYCE, M.S., P. R. VERNIER, S. E. NIELSON, AND F.K.A. SCHMIEGLOW. 2002.
Evaluating resource selection functions. Ecological Modelling 157: 281-300.
BRENNAN, L. A. AND W. P. KUVLESKY. 2005. North American grassland birds: an
unfolding conservation crisis? Journal of Wildlife Management 69: 1-13.
BROTONS, L., W. THUILLER, M. B. ARAUJO, AND A.H. HIRZEL. 2004. Presence-absence
versus presence-only modelling methods for predicting bird habitat suitability.
Ecography 27: 437–448.
BROTONS, L., A. WOLFF, G. PAULUS, AND J-L. MARTIN. 2005. Effect of adjacent
agricultural habitat on the distribution of passerines in natural grasslands.
Biological Conservation 124: 407–414.
BUCKLAND, S. T. 2001. Introduction to distance sampling: estimating abundance of
biological populations. Oxford University Press, Oxford, UK.
CHANDLER, R. B., D. I. KING, AND S. DESTEFANO. 2009. Scrub–shrub bird habitat
associations at multiple spatial scales in beaver meadows in Massachusetts. Auk
126: 186–197.
COLLINGE, S. K. 2009. Ecology of fragmented landscapes. The Johns Hopkins
University Press, Baltimore, Maryland, USA.
18
COSEWIC. 2011. Canadian wildlife species at risk. Committee on the status of
endangered wildlife in Canada. Available at:
http://www.cosewic.gc.ca/eng/sct0/rpt/rpt_csar_e.cfm. Accessed online May 4,
2011.
CUNNINGHAM, M. A., AND D. H. JOHNSON. 2006. Proximate and landscape factors
influence grassland bird distributions. Ecological Applications 16(3): 1062-1075.
CUNNINGHAM, M. A., AND D. H. JOHNSON. 2011. Seeking parsimony in landscape
metrics. Journal of Wildlife Management 75(3): 692-701.
DAVIS, S. K., R. J. FISHER, S. L. SKINNER, T. L. SHAFFER, AND R. M. BRIGHAM. 2013.
Songbird abundance in native and planted grassland varies with type and amount
of grassland in the surrounding landscape. Journal of Wildlife Management 77(5):
908-919.
DE MARCO, P., J. A. F. DINIZ-FILHO, AND L. M. BINI. 2008. Spatial analysis improves
species distribution modelling during range expansion. Biology Letters 4: 577-80.
DONOVAN, T. M., C. J. BEARDMORE, D. N. BONTER, J. D. BRAWN, R. J. COOPER, J. A.
FITZGERALD, R. FORD, S. A. GAUTHREAUX, T. L. GEORGE, W. C. HUNTER, T.
E. MARTIN, J. PRICE, K. V. ROSENBERG, P. D. VICKERY, AND T. B. WIGLEY.
19
2002. Priority research needs for the conservation of Neotropical migrant
landbirds. Journal of Field Ornithology 73: 329-339.
DORMANN, C. F. 2007. Promising the future? global change predictions of species
distributions. Basic and Applied Ecology 8: 387-397.
DUNNING, J. B. JR, D. J. STEWART, B. J. DANIELSON, B. R. NOON, T. L. ROOT, R. H.
LAMBERSON, AND E.E. STEVENS. 1995. Spatially explicit population models:
current forms and future uses. Ecological Applications 5: 3-11.
ELITH, J., C. H. GRAHAM, R. P. ANDERSON, M. DUDIK, S. FERRIER, A. GUISAN R. J.
HIJMANS, F. HUETTMANN, J. R. LEATHWICK, A, LEHMANN, J. LI, L. G.
LOHMANN, B. A. LOISELLE, G. MANION, C. MORITZ, M. NAKAMURA, Y.
NAKAZAWA, J. M. OVERTON, A. TOWNSEND PETERSON, S. J. PHILLIPS, K.
RICHARDSON, R. SCACHETTI-PEREIRA, R. E. SCHAPIRE, J. SOBERON, S.
WILLIAMS, M. S. WISZ, AND N. E. ZIMMERMANN. 2006. Novel methods improve
prediction of species’ distributions from occurrence data. Ecography 29: 129-151.
ELITH, J., AND J. R. LEATHWICK. 2009. Species distribution models: ecological
explanation and prediction across space and time. Annual Review of Ecology,
Evolution, and Systematics 40: 677-697.
20
ELPHICK, C. S. 2008. How you count counts: The importance of methods research in
applied ecology. Journal of Applied Ecology 45: 1313–1320.
FARNSWORTH, G. L., K. H. POLLOCK, J. D. NICHOLS, T. R. SIMONS, J. E. HINES, AND J.
R. SAUER. 2002. A removal model for estimating detection probabilities from
point-count surveys. Auk 119: 414-425.
FIELDING, A. H., AND J. F. BELL. 1997. A review of methods for the assessment of
prediction errors in conservation presence/absence models. Environmental
Conservation 24: 38–49.
FINCH, D. M. 1991. Population ecology, habitat requirements, and conservation of
neotropical migratory birds. USDA Forest Service General Technical Report RM-
205. Available at:
http://www.fws.gov/southwest/es/documents/R2ES/LitCited/LPC_2012/Finch_19
91.pdf. Accessed online August 28, 2012.
FRANKLIN, J. 2009. Mapping species distributions: spatial inference and prediction,
Cambridge University Press, Cambridge, UK.
FRANKLIN, J. 2010. Moving beyond static species distribution models in support of
conservation biogeography. School of Geographical Sciences and Urban Planning
and School of Life Sciences, Arizona State University, Tempe AZ USA.
21
GRAHAM, P. L., S. HAGEMANN, AND S. JAUN. 2007. On interpreting hydrological change
from regional climate models. Climate Change 81: 97-122.
GREGORY, R. D., A. J. VAN STRIEN, P. VORISEK, A. W. GMELIG MEYLING, D. G.
NOBLE, R. P. B. FOPPEN, AND D. W. GIBBONS. 2005. Developing indicators for
European birds. Philosophical Transactions of Royal Society B 360: 269–288.
GU, W., AND R. K. SWIHART. 2004. Absent or undetected? Effects of non-detection of
species occurrence on wildlife-habitat models. Biological Conservation 116(2):
195-203.
GUISAN, A., AND N. E. ZIMMERMAN. 2000. Predictive habitat distribution models in
ecology. Ecological Modelling 135: 147–186.
GUISAN, A., AND W. THUILLER. 2005. Predicting species distribution: offering more
than simple habitat models. Ecology Letters 8: 993–1009.
HUTCHINSON, G. E. 1957. Concluding remarks. Cold Spring Harbor Symposia on
Quantitative Biology 22: 415-427.
HUTTO, R. L. 1985. Habitat selection by nonbreeding, migratory land birds. pp 455-476
in M. L. Cody (ed.) Habitat selection in birds. Academic Press, Inc., Orlando.
22
HUTTO, R. L., S. M. PLETSCHET, AND P. HENDRICKS. 1986. A fixed-radius point count
method for nonbreeding and breeding season use. Auk 103: 593-602.
JOHNSON, D. H. 1980. The comparison of usage and availability measurements for
evaluating resource preference. Ecology 61: 65–71.
JOHNSON, D. H. 1995. Point counts of birds: What are we estimating? Monitoring bird
populations by point counts. U.S. Department of Agriculture, Forest Service
General Technical Report PSW-GTR-149: 117-123. Available at:
http://www.npwrc.usgs.gov/resource/birds/ptcounts/. Accessed online July 23,
2012.
JOHNSON, D. H. 2008. In defense of indices: The case of bird surveys. Journal of
Wildlife Management 72: 857–868.
JONES, J. 2001. Habitat selection studies in avian ecology: a critical review. Auk 118:
557-562.
KERY, M., AND B. R. SCHMIDT. 2008. Imperfect detection and its consequences for
monitoring for conservation. Community Ecology 9: 207-216.
23
KERY, M. 2012. Towards the modelling of true species distributions. Journal of
Biogeography 38: 617–618.
LAURENT, E. J., H. SHI, D. GATZIOLIS, J. P. LEBOUTON, M. B. WALTERS, AND J LIU.
2005. Using the spatial and spectral precision of satellite imagery to predict
wildlife occurrence patterns. Remote Sensing of Environment 97: 249-262.
LELE, S. R., M. MORENO. AND E. BAYNE. 2012. Dealing with detection error in site
occupancy surveys: What can we do with a single survey? Journal of Plant
Ecology 5: 22-31.
LEOPOLD, A. 1933. Game Management. Charles Scribner, New York.
LEVIN, S. A. 1992. The problem of pattern and scale in ecology. Ecology 73: 1947-1967.
LIU, C., M. WHITE, AND G. NEWELL. 2011. Measuring and comparing the accuracy of
species distribution models with presence-absence data. Ecography 34: 232-243.
LOBO, J. M. 2008. More complex distribution models or more representative data?
Biodiversity Informatics 5:14-19.
24
LOBO, J. M., A. JIMÉNEZ-VALVERDE, AND R. REAL. 2008. AUC: a misleading measure
of the performance of predictive distribution models. Global Ecology and
Biogeography 17: 145-51.
LOZIER, J. D., P. ANIELLO, AND M. J. HICKERSON. 2009. Predicting the distribution of
Sasquatch in western North America: anything goes with ecological niche
modelling. Journal of Biogeography 36: 1623-1627.
MAC NALLY, R. 2000. Regression and model-building in conservation biology,
biogeography and ecology: the distinction between—and reconciliation of—
‘predictive’ and ‘explanatory’ models. Biodiversity and Conservation 9: 655–71.
MACARTHUR, R. 1960. On the relative abundance of species. American Naturalist 94:
25-36.
MACKENZIE, D. I., J. D. NICHOLS, G. B. LACHMAN, S. DROEGE, J. A. ROYLE, AND C.
A. LANGTIMM. 2002. Estimating site occupancy rates when detection
probabilities are less than one. Ecology 83: 2248–2255.
MACKENZIE, D. I. 2005. What are the issues with presence-absence data for wildlife
managers? Journal of Wildlife Management 69(3): 849-860.
25
MACKENZIE, D. I., AND J. A ROYLE. 2005. Designing efficient occupancy studies:
General advice and tips on allocation of survey effort. Journal of Applied Ecology
42: 1105-1114.
MACKENZIE, D. I. 2006. Modeling the probability of resource use: the effect of, and
dealing with, detecting a species imperfectly. Journal of Wildlife Management
70(2) Invited Paper.
MACKENZIE, D. I., J. D. NICHOLS, J. A. ROYLE, K. P. POLLOCK, L. L. BAILEY, AND J.
E. HINES. 2006. Occupancy estimation and modeling: inferring patterns and
dynamics of species occurrence. Academic Press, San Diego, California, USA.
MANLY, B. F. J., L. L. MCDONALD, D. L. THOMAS, T. L. MCDONALD, AND W. P.
ERICKSON. 2002. Resource selection by animals: statistical design and analysis
for field studies. Second Edition. Kluwer, Boston, Massachusetts, USA.
MARSH, D. M., AND P. C. TRENHAM. 2008. Current trends in monitoring programs for
animal and plant populations. Conservation Biology 22: 647-655.
MIDGLEY, G. F, G. O. HUGHES, W. THUILLER, AND A. G. REBELO. 2006. Migration rate
limitations on climate change induced range shifts in Cape proteaceae. Diversity
and Distributions 12: 555.
26
MILLENNIUM ECOSYSTEM ASSESSMENT. 2005. Ecosystems and human well-being:
Biodiversity synthesis. World Resource Institute, Washington, DC.
NICHOLS, J. D., J. E. HINES, J. R. SAUER, F. W. FALLON, J. E. FALLON, AND P.J.
HEGLUND. 2000. A double-observer approach for estimating detection probability
and abundance from point counts. Auk 117: 393-408.
PEARCE, J. L., AND S. FERRIER. 2001. The practical value of modelling relative
abundance of species for regional conservation planning: a case study. Biological
Conservation 98: 33-43.
PEARCE, J. L., AND M. S. BOYCE. 2006. Modeling distribution and abundance with
presence-only data. Journal of Applied Ecology 43: 405-412.
PEARSON, R. G., C. J. RAXWORTHY, M. NAKAMURA, AND A. T. PETERSON. 2007.
Predicting species distributions from small numbers of occurrence records: a test
case using cryptic geckos in Madagascar. Journal of Biogeography 34: 102-117.
PETERSON, A. T. 2006. Uses and requirements of ecological niche models and related
distributional models. Biodiversity Informatics 3: 59-72.
27
PHILLIPS, S. J., M. DUDIK, J. ELITH, C. GRAHAM, A. LEHMANN, J. LEATHWICK, AND S.
FERRIER. 2009. Sample selection bias and presence only models of species
distributions. Ecological Applications 19:181-197.
RIBIC, C. A., AND D. W. SAMPLE. 2001. Associations of grassland birds with landscape
factors in southern Wisconsin. American Midland Naturalist 146: 105-121.
RIBIC, C. A., R. R. KOFORD, J. R. HERKERT, D. H. JOHNSON, N. D. NIEMUTH, D. E.
NAUGLE, K. K. BAKKER, D. W. SAMPLE, AND R. B. RENFREW. 2009. Area
sensitivity in North American grassland birds: patterns and processes. Auk 126:
233–244.
ROBBINS, C. S., D. K. DAWSON, AND B. A. DOWELL. 1989. Habitat area requirements of
breeding forest birds of the middle Atlantic states. Wildlife Monographs 103: 1–
34.
ROCCHINI, D., J. HORTAL, S. LENGYEL, J. M. LOBO, A. JIMENEZ-VALVERDE, C.
RICOTTA, G. BACARO, AND A. CHIARUCCI. 2011. Accounting for uncertainty
when mapping species distributions: the need for maps of ignorance. Progress in
Physical Geography 35(2): 211-226.
SHIRLEY, S. M., Z. YANG, R. A. HUTCHINSON, J. D. ALEXANDER, K. MCGARIGAL, AND
M. G. BETTS. 2013. Species distribution modelling for the people: unclassified
28
landsat TM imagery predicts bird occurrence at fine resolutions. Diversity and
Distributions 1: 1-12.
TERBORGH, J. 1989. Where have all the birds gone? Princeton University Press,
Princeton, New Jersey, USA.
TYRE, A. J., B. TENHUMBERG, S. A. FIELD, D. NIEJALKE, K. PARRIS, AND H. P.
POSSINGHAM. 2003. Improving precision and reducing bias in biological surveys:
Estimating false-negative error rates. Ecological Applications 13: 1790-1801.
VAUGHAN, I. P., AND S. J. ORMEROD. 2005. The continuing challenges of testing species
distribution models. Journal of Applied Ecology 42: 720–730.
VERBYLA, D. L., AND J. A. LITAITIS. 1989. Resampling methods for evaluating
classification accuracy of wildlife habitat models. Environmental Management
13: 783–7.
VICKERY, P. D., P. L. TUBARO, J. M. CARDOSO DA SILVA, B. G. PETERJOHN, J. R.
HERKERT, AND R. B. CAVALCANTI. 1999. Conservation of grassland birds in the
western hemisphere. Studies in Avian Biology 19: 2-26.
WHITE, R. P., S. MURRAY, AND M. ROWEDER. 2000. Pilot analysis of global
ecosystems: grassland ecosystems. World Resource Institute, Washington, D.C.
29
WIENS, J. A. 1972. Anuran habitat selection: early experience and substrate selection in
Ranacascadae tadpoles. Animal Behavior 20: 218-220.
WIENS, J. A. 1989. Spatial scaling in ecology. Functional Ecology 3: 385-397.
WINTLE, B. A., M. A. MCCARTHY, K. M. PARRIS, AND M. A. BURGMAN. 2004.
Precision and bias of methods for estimating point survey detection probabilities.
Ecological Applications 14: 703–712.
WISZ, M. S., R. J. HIJMANS, J. LI, A. T. PETERSON, C. H. GRAHAM AND A. GUISAN.
2008. Effect of sample size on the performance of species distribution models.
Diversity and Distributions 14: 763–773.
30
2. PREDICTING GRASSLAND SONGBIRD DISTRIBUTION PATTERNS
USING OCCUPANCY MODELS
2.1 Introduction
Concern over population declines of grassland birds has stimulated a variety of
conservation plans, many of which explicitly promote a landscape approach to
conservation (Berlanga et al. 2010). Population trend data for grassland birds suggest
that the causes of declines are not locally isolated phenomena but involve the cumulative
loss and degradation of grassland habitat throughout the Americas (Vickery et al. 1999;
Askins et al. 2007). Approximately 85% of the world’s temperate grassland biome has
disappeared since the late 1800’s (Millennium Ecosystems Assessment 2005), and within
Canada, only 20% of native grassland remains (White et al. 2000). This substantial loss
of habitat has reduced the ability of native grasslands to provide vital ecosystem services
(Millennium Ecosystem Assessment 2005). As a result, many grassland-dependent taxa
are now of conservation concern, with grassland birds having declined more than any
other group in North America (Samson and Knopf 1994). However, predicting
occupancy patterns on the landscape can provide great value in assessing the impacts of
environmental change, and give the ability to stabilise or increase grassland bird
populations once a clear understanding of resource selection has been achieved (Dunning
et al. 1995; Vickery et al. 1999; Brennan and Kuvlesky 2005).
Determining which resources are selected at a higher propensity than others is of
particular interest because it may provide insight into species requirements for survival
and reproduction (Manly et al. 2002; MacKenzie et al. 2006). For birds, resource
selection begins at the scale of the geographic range of the species, then at broad
landscape scales for populations, followed by individual levels to establish home ranges,
31
and finally at a finer scale to satisfy individual life history requirements (Johnson 1980).
Much of the initial resource selection research was focused on local habitat attributes
such as vegetation structure (Wiens 1973; Fisher and Davis 2010), and investigations of
breeding bird responses to environmental conditions primarily targeted forest-dwelling
species (Robbins et al. 1989; Terborgh 1989; Finch 1991). Studies of nesting grassland
birds suggest that birds respond to landscape factors at scales from 200 to 1600 m
(Bakker et al. 2002; Brontons et al. 2005; Ribic et al. 2009). For example, Ribic and
Sample (2001) found that responses to landscape scales at both 200 and 400 m radius
could be used to explain the distribution of Grasshopper Sparrow (Ammodramus
savannarum). Davis et al. (2013) found that landscapes within 400 m of grassland
parcels influenced abundance of grassland songbirds, but pointed out that determining the
precise scale was difficult because of the strong correlations between scales (400 – 1600
m radius) within the region. No single measure can describe the many factors that
influence resource selection, however Cunningham and Johnson (2011) suggest that
identifying a broadly influential measure such as the amount of a given land cover type
has proven to be reliable and easy to conceptualise relative to other measures for
conservation managers.
Identifying that birds select habitats based on a set of geographic and
environmental characteristics is fundamental to understanding resource selection (Manly
et al. 2002; MacKenzie et al. 2006). Species distribution models typically combine count
data with geographical and environmental variables to understand how species interact
with the composition and spatial arrangements of habitats, which in turn, can be used to
predict species distributions across an area of interest (Gusian and Thuiller 2005; Elith
32
and Leathwick 2009; Franklin 2010). Using functionally relevant variables such as the
amount of a particular habitat and spatially explicit covariates describing topography are
essential to pinpointing the ecological drivers of resource selection and predicting
distributions across a larger area of interest (Austin 2002; Elith and Leathwick 2009;
Cunningham and Johnson 2011). Accurate detections of individuals are also an
important consideration for distribution modelling because birds are mobile, and tend to
exhibit different activity patterns which likely influence both occurrence and detection
probability across the landscape (Hutto 1985; Donovan et al. 2002; Chandler et al. 2009).
Virtually all count data can be affected by false absences (species present but not
detected; MacKenzie et al. 2006), and recent investigations have demonstrated that not
accounting for imperfect detection will yield biased results (Tyre et al. 2003; Gu and
Swihart 2004; Marsh and Trenham 2008). Occupancy models are a type of species
distribution model which enable researchers to account for imperfect detection
(MacKenzie et al. 2002; 2006). Models that account for detectability bias can lead to
better management decisions, assist in ensuring key habitats are identified and conserved
by improving our understanding of resource selection and the predictive ability of
distribution models. A reliable distribution model can be used to guide future
management strategies, justify funding for research and conservation activities, prioritise
conservation activities and assess species status with confidence (Aldridge and Boyce
2007; COSEWIC 2011).
My objectives were to determine landscape factors that influence the distribution
of grassland birds, use occupancy models to generate predictions of species distribution,
33
and evaluate the predictive performance of 14 grassland bird species models constructed
based on data from southeastern Alberta, Canada.
2.2 Methodology
2.2.1 Study Area
I conducted my study within the boundary of a municipality named the Special
Areas (www.specialareas.ab.ca), in southeastern Alberta (51°18’31”N and 110°56’19”
W; Figure 2.1). Bordered to the South by the Red Deer River, the Special Areas
encompass 2 million ha of land predominately used for livestock grazing, annual
cropping, and oil and gas extraction. Grazing is the focal agricultural activity, although
roughly 25% of the study area is under dryland and irrigational farming practices (mainly
wheat/fallow; AAFC 2012). The area is characterised by undulating plains with
hummocky uplands occurring towards the northern portions. Soil characteristics are
primarily Brown Chernozemic, although Dark Brown Chernozemic soils become more
prominent as the moisture gradient increases from south to north (Natural Regions
Committee 2006). Characterised by short hot summers and cold winters, the area
typically follows a continental summer-high weather pattern with maximum precipitation
occurring in June. Mean annual temperature for the warmest month (July) is
approximately 18° Celsius (Natural Regions Committee 2006). Drying winds, low
summer precipitation, high summer temperatures, and intense sunshine contribute to a
mean annual precipitation of approximately 333 mm (Natural Regions Committee 2006).
The Special Areas is located in the dry mixed-grass prairie and northern fescue
grassland natural subregions of Alberta, and extends into the transitional zone of the
34
central parkland (Natural Regions Committee 2006). Native vegetation consists
primarily of blue grama (Bouteloua gracilis), green needlegrass (Nassella viridula),
needle-and-thread grass (Hesperostipa comata), june grass (Bromus tectorum), northern
wheatgrass (Elymus macrourus), western wheatgrass (Pascopyrum smithii), northern
rough fescue (Festuca altaica), and plains rough fescue (Festuca hallii). Sedges (Carex
spp), sage (Artemsia spp.), and various other forbs commonly occur throughout the area.
Shrubs such as western snowberry (Symphoricarpos occidentalis), silverberry (Elaeagnus
commutata) and rose (Rosa spp.) occur sporadically in mesic sites. The area supports a
diverse community of grassland birds, including several species listed by the Species at
Risk Act in Canada, such as Sprague’s Pipit (Anthus spragueii), McCown's Longspur
(Rhynchophanes mccownii), Chestnut-collared Longspur (Calcarius ornatus),
Loggerhead Shrike (Lanius ludovicianus), Burrowing Owl (Athene cunicularia), Long-
billed Curlew (Numenius americanus), and Ferruginous Hawk (Buteo regalis).
2.2.2 Sampling Design and Data Collection, 2011
I used results from avian surveys conducted in 2011 as an independent source of
species occurrence records to evaluate the predictive power of the occupancy models
developed using point count data I collected in 2012.
In 2011, I used the national draft Sprague’s Pipit Resource Selection Function
model developed by the Canadian Wildlife Service (S. Davis unpubl. data) to randomly
stratify the sampling area for surveys. Although the model is specific to Sprague’s Pipit,
it is largely based on the amount of grassland in the study area and thus relevant to most
birds that require grassland habitat. I used a Geographic Information System (ArcMap
35
10.0; ESRI 2011) to designate each legal section of land (256 ha; McKercher and Wolfe
1986) in the Special Areas as either suitable or unsuitable habitat according to the model.
The ratio of suitable to unsuitable habitat was approximately 2:1; therefore, I sampled
twice as much suitable habitat as unsuitable habitat. I randomly selected legal sections
[n=100 suitable and n=50 unsuitable] to use as starting points for each transect. The
closest road or trail to the northeast corner of the legal section was used as the starting
point for each transect. I delineated a 16 km transect by randomly selecting a direction
from each starting point and maintained that direction until I encountered a road
intersection. I used a random number generator to select a number between 1 and 4 (1 =
North, 2 = East, 3=South, and 4=West) and assigned the direction to the next intersection.
Point count locations were plotted every 800 m along each transect to maintain a
reasonable distance between sampling locations (Cunningham and Johnson 2006). All
surveys were conducted either along gravelled grid roads and trails (e.g. those
constructed by the oil and gas industry) or at sites > 150 m from roads. On-road and off-
road transects were sampled at a 1:1 ratio to address roadside sampling bias (Ralph et al.
1995).
I conducted 1398 avian point counts (mid-May to early-July; Figure 2.2), using a
removal sampling design (Farnsworth et al. 2002). Each location was surveyed by a
trained observer for three equal intervals of 2.5 minutes, comprising an overall point
count duration of 7.5 minutes. The number and species of bird were recorded within both
a 100 m and 400 m radius of the observer (Hutto et al 1986). Point count locations were
located using a hand-held global positioning system unit. All point counts were
conducted between sunrise and 0930 CST on days with winds <20 km/hr and no
36
precipitation or unseasonal weather events. Since the principal objective of this study
was to determine occupancy, I converted abundance records to reflect species that were
detected or not detected within 400 m of the survey point.
2.2.3 Sampling Design and Data Collection, 2012
I used ArcMap 10.0 and classified satellite imagery from Agriculture and Agri-
foods Canada 2011 crop mapping inventory (AAFC 2012) to calculate the ratio of
grassland to cropland cover within the study area. Distinguishing native grassland from
planted grassland, and accurately identifying shrub cover using the satellite imagery was
impossible. Therefore, I calculated the amount of grassland habitat by combining
grassland, pasture/hayland, and shrubland cover types, which resulted in approximately
75% of the study area being considered grassland cover. The ratio of grassland to
cropland was approximately 3:1; therefore, I used this ratio to stratify my random
sampling effort.
I classified each quarter section (64 ha; McKercher and Wolfe 1986) in the
Special Areas as either Low (0-25%), Moderate (25.01-75%), or High (75.01-100%)
based on the percentage of grassland cover. I randomly selected 150 quarter sections
designated “high” and 50 quarter sections designated “moderate and low” to assign
starting locations for transect surveys. Point counts were conducted along a 24 km
transect following the methods used in 2011. However, during the 2012 breeding season
(mid-May to early-July), I employed a standard repeated survey design to account for
potential detection bias (Mackenzie et al. 2002; 2006). Each point count was surveyed by
trained observers for a 5 minute period, and the number and species of bird were recorded
37
within both a 100 m and 400 m radius of the observer (Figure 2.2). Surveyors recorded
date, time of day and wind speed at the start of each point count. Each of the 870 point
count locations was visited three times within a five day period by a different observer to
reduce surveyor bias (MacKenzie et al. 2006). This repeated sampling design provides
less biased estimates of occupancy given heterogeneous detection probabilities
(MacKenzie and Royle 2005; MacKenzie et al. 2006). Observers alternated starting
points from either end of the transects to ensure that sampling at each site occurred at
different times throughout the survey period (MacKenzie et al. 2006). I converted
abundance records to reflect species that were detected or not detected within 400 m of
the survey point.
2.2.4 GIS Methodology
All explanatory variables were constructed using ArcMap 10.0 (ESRI 2011). I
applied a 5 km buffer to the boundary of my study area to ensure that all landscape
information could be extracted for counts near the study area boundary. I used the
polygon to raster tool to generate a 30 x 30 m grid of the study area, encompassing
2,771,922 cells which I used to match the processing extent and resolution for each
spatially explicit explanatory variable (see below). I selected twelve explanatory
variables based on existing knowledge of grassland bird distributions and categorised
these into five model classes; detection, land cover, geographic location, topography, and
well density (Table 2.1). All digital data were projected by the Universal Transverse
Mercator (UTM Zone 12N) Coordinate system using the World Geodetic System WGS
1984.
38
2.2.4.1 Explanatory Variables
Most bird detections during point count surveys are based on auditory cues, and
consequently song rate activity can greatly influence detection probability (Alldredge et
al. 2007). Therefore, I included date, time of day, and wind speed as visit-specific
covariates to assess variation in species detection rates.
I calculated the amount of grassland, woody cover, surface water, and vegetative
greenness to help explain land cover influences on occurrence probability. Using the
AAFC land cover classification imagery, I independently derived the percentage of
grassland (grassland and hay/pastureland areas combined) and percentage of woody
cover (shrubland, coniferous, and deciduous tree areas combined) variables in raster
format. Surface water was also derived by combining the watercourses, wetlands, and
water-bodies vector data (Natural Resources Canada 2007). The surface water data were
in vector format, therefore I used the polygon to raster tool to represent surface water in a
single binary raster format. The new raster files for grassland, woody vegetation, and
surface water were reclassified into a binary raster datasets, where each 30 x 30 m raster
cell containing the target variable was set equal to “1” and everything else was set equal
to “0”. I built a modified Soil-adjusted Vegetation Index (MSAVI2), which quantified
plant biomass and provided an index to vegetative greenness on the landscape from
satellite imagery obtained from GeoBase.ca (Natural Resources Canada 2004b). Since
models which include several remote sensing predictors from the same area tend to be
highly correlated (Zimmerman et al. 2007), I chose the MSAVI2 over other indices like
Normalized Difference Vegetative Index for my analyses as it is better suited to regions
39
with a high degree of exposed soil (Qi et al. 1994). To derive the MSAVI2, I built two
raster files composed of red and near-infrared multispectral band data. Both
multispectral bands were constructed by mosaicking the 1:50,000 Landsat 7 ortho-image
map-sheets that intersected the study area, using the model builder tool. Using the red
band and near-infrared band indices, I calculated an index for vegetative greenness
ranging between -1 and +1 for each 30 x 30 m in raster calculator (Qi et al. 1994).
I used the geographic location of a point count (i.e., latitude and longitude) within
the modelling process to model spatial autocorrelation (Rangel et al. 2006) and identify
the geographic influence on species occurrence across the study area. Both the latitude
and longitude variables (i.e., geographic location) were constructed by using the 30 x 30
m study area raster file as a base layer. Each cell was converted from cell to centroid
point, and each point was assigned an arbitrary value representing cell geometry in
increasing values for latitude and longitude. Both latitude and longitude shape files were
then converted to a raster format using the point to raster tool.
I derived two topographic variables to assist in explaining the influence of local
habitat conditions on occurrence probability. Both the Compound Topographic Index
(CTI) and Heat Load Index (HLI) variables were constructed from a Digital Elevation
Model (DEM). The CTI is a function of both slope and upstream contribution to
determine wetness accumulation based on flow direction. This index reflects the state of
soil moisture across the landscape (Gessler et al. 1995). The HLI is a function of folding
the aspect so that the highest values are southwest and the lowest values are northeast
while also accounting for steepness of slope, and reflects the state of solar radiation
across the landscape (McCune and Keon 2002). I constructed the DEM by first
40
mosaicking the 1:50,000 National Topographic System map sheets tiles that intersected
the study area (Natural Resources Canada 2004a), using the model builder tool. I then
used the Fill DEM tool to smooth out imperfections in the DEM. To derive the CTI
gradient, I constructed raster files for flow direction, flow accumulation and contributing
areas based on the 30 x 30 m pixel resolution using the raster calculator tool. I used
Gessler et al.’s (1995) equation for calculating CTI in raster calculator to derive a new
raster reflecting an index of soil moisture for the study area. I used the HLI equation
provided by McCune and Keon (2002) in raster calculator to derive the Heat Load Index
based on the 30 x 30 m pixel resolution.
I used oil and gas well density as a proxy for the amount of oil and gas
disturbance on the landscape (Kalyn-Bogard 2011). I built the well density layer by first
determining the number of well sites, both active and inactive, occurring in the study area
(IHS Energy Group 2012). I then calculated how many well sites occurred within a circle
of 2.59 km2 (radius = 908 m which represented 1 square mile) for each cell in the 30 x 30
m study area raster file using the point density tool.
For each explanatory variable, I ran a moving window analysis at varying radii
(200 m, 400 m, 800 m, and 900 m) using a circular focal window with the focal statistics
tool. The varying radii represented the percentage of each variable in raster format
within the circular areas, which I then used to determine which scale was most relevant to
proceed with for analysis (Cunningham and Johnson 2006).
41
2.2.5 Statistical Analysis
All statistical analyses were conducted in R statistical software program (R
Development Core Team 2013). Explanatory variables were standardised using a z-score
procedure, normalizing values to a mean of 0 and variance of 1, which optimised the
maximum likelihood estimators for use by the modelling algorithm (Fiske and Chandler
2011).
2.2.5.1 Determining Variable Scale and Collinearity
I calculated correlation coefficients for each covariate at the four scales (i.e.,
grass200 m, grass400 m, grass800 m, and grass900 m…woody200 m, woody400 m,
woody800 m, and woody900 m…) to examine correlation between spatial scales and
determine if multiple scales should be considered for investigation (Porter and Church
1987; Wiens 1989). Using the mean focal window estimates from the point count
locations conducted in 2012 (n=870), I determined that Pearson correlation coefficient at
all scales were highly correlated (r>.80), and therefore proceeded with analysis using data
from the 400 m scale (Fletcher and Koford 2002). While I acknowledge that bird
occurrence may be influenced by multiple landscape scales, the 400 m scale consistently
led to lower AIC values. Furthermore, the 400 m scale is consistent with the bird surveys
I conducted, closely resembles the area of a legal quarter section (the common unit of
land management and ownership; McKercher and Wolfe 1986), and has been identified
as a relevant scale in previous studies for the region (Davis et al. 2013). I then tested for
multicollinearity among each of the covariates at the 400 m extent. I found that CTI and
HLI were moderately correlated (r=-.72), but retained both to determine their relative
42
importance for each species (Table 2.2). All variables were inspected for non-linear
relationships.
2.2.5.2 Model Selection
I used occupancy models (unmarked package; Fiske and Chandler 2011) to
explain variation in grassland bird occupancy as a function of the five model classes (i.e.,
detection, land cover, geographic location, topography, and well density). Akaike’s
Information Criterion (AIC; Akaike 1974) was used to evaluate and rank all candidate
models (Burnham and Anderson 2002). Models with the lowest AIC score and highest
AIC weight (Wi) were considered to be the best to fit the data of those considered
(Burnham and Anderson 1998), and model averaging was used to deal with model
selection uncertainty for species with competing models.
I initially fit each species occupancy model for detection probability (i.e., date,
time of day, and wind), which was then kept constant to model the occurrence probability
component (Tyre et al. 2003; Mackenzie et al. 2006). I examined all subsets and ranked
performance independently for each model class against a null model. The top candidate
model set from the detection, land cover, topography, geographic location, and well
density model classes were combined to form the global model.
I identified all models within a cumulative weight of 95% and only included
models which contained the grassland variable (MuMIN package; Barton 2013). I based
all predictions on the model averaged parameter estimates, which were weighted with a
variable shrinkage factor depending on how important each variable was across
competing models.
43
2.2.5.3 Model Evaluation
I evaluated each species model using two approaches: 1) using the independent
dataset collected in 2011, and 2) by randomly partitioning point count data from 2012. In
both cases, I created Receiver Operator Characteristic (ROC) plots, calculated Area
Under the Curve (AUC), and calculated Kappa statistics (PresenceAbsence package;
Freeman and Moisen 2008). The AUC score for each species model represents the
likelihood that the model correctly identifies a true positive detection for a species that is
higher than a false positive detection. According to Pearce and Ferrier (2000), models
are deemed “useful” for predicting occupancy if AUC values > 0.7, “good” if values >
0.8, and “excellent” if AUC values > 0.9. Unlike the ROC approach, the Kappa statistic
is a threshold-dependent performance measure (Pearce and Ferrier 2002). For Kappa, I
used the true skill statistic recommended by Allouche et al. (2006) to select a threshold
for model evaluation and comparison because it minimises the mean error rate for
positive and negative observations. According to Landis and Koch’s (1977) guidelines,
models with Kappa values k<0.2 are considered “poor”, 0.2<k>0.4 are considered “fair”,
0.4<k>0.75 are considered “good”, and k>0.75 are considered “excellent”. I deemed
species occupancy models “reliable” if AUC > 0.7 and Kappa > 0.2 under both
evaluation data sets, “marginal” in a models ability to predict if they were lacking in a
performance test, and “unreliable” if no test result was successful.
I used Huberty’s (1994) rule to determine the partitioning ratio of training to
testing data needed for evaluation purposes with the 2012 data set. I used their equation
and found that 25% was needed to evaluate models with the 2012 data and therefore
44
partitioned the count data into 75% for building the models (n = 652) and 25% for
evaluating (n = 218) the models. Based on the 25% ratio for testing data, I specified 25%
of the points from each transect using a random number generator.
Imperfect detection is a common problem associated with most evaluation data
sets. For example, the point count records I collected in 2011 do not explicitly account
for imperfect detection, and therefore are not considered the same currency as the data I
used to build my occupancy models. In the case of partitioned data, Rota et al. (2011)
recommended adjusting the testing data for imperfect detection as it yields better
performance results. Therefore, since the problem of imperfect detection was still present
with my 2012 testing data, I used equation 1 to adjust predictions of occupancy for
imperfect detection, where Ψ represents occurrence probability and ρ is detection
probability.
EQUATION 1
The adjustments were calculated by using the model averaged fixed effects from the
detectability model to predict detection probability for each visit to survey sites. I then
multiplied the predicted probability of occurrence by the predicted probability of
detection to arrive at a prediction of occupancy at each site, assuming heterogeneous
detection probability.
2.2.5.4 Uninformative Parameters and Predictions
Since the primary goal of my research was to predict occupancy, I identified and
then discarded variables that had no discernible effect on occupancy predictions for my
final models (e.g., Pagano and Arnold 2009). I sequentially eliminated the least
45
important covariate from each species global model identified using the model parameter
estimate and standard error. If eliminating a covariate led to a reduction in AIC and no
net loss of the models predictive performance (i.e., Kappa and AUC), I considered the
variable uninformative and discarded it from the model (Arnold 2010). This approach
was continued until no additional covariate could be eliminated without leading to an
increase in AIC or decrease in predictive performance. I used 95% confidence intervals
(CI) to evaluate global models and only discuss parameters whose slopes do not include
zero. I used the parameter estimates from each species final model to predict species
occurrence for each 30 x 30 m pixel across the study area (raster package; Hijmans and
van Etten 2013).
2.3 Results
2.3.1 General Results
I conducted 870 point counts in 2012, and analysed detection/non-detection
records for 14 grassland bird species detected at 40 or more locations (Table 2.3). I made
an exception for McCown’s Longspur, which was modelled given its status as a priority
species. The most frequently occurring species was Western Meadowlark, with the
highest probability of occurrence (Ψ=.99) and probability of detection (ρ =.91), followed
by Vesper Sparrow, Savannah Sparrow, Brown-headed Cowbird, Clay-colored Sparrow,
Horned Lark, Sprague’s Pipit, Baird’s Sparrow, Marbled Godwit, Willet, Chestnut-
collared Longspur, Long billed Curlew, Upland Sandpiper, and McCown’s Longspur
(Table 2.3).
I used the top model to estimate occupancy for Chestnut-collared Longspur,
Upland Sandpiper, and Western Meadowlark, as the model selection process indicated no
46
competitive models within a cumulative AIC weight (wi) <0.95. I used the model
averaging results for the remaining species to estimate occupancy because of model
selection uncertainty (Table 2.4). Variable shrinkage factors on coefficient weights used
to model occupancy are provided in Appendix A. Table A.1. The detection models that
best explained the data typically included both date and time of day (Table 2.5). In
addition to the amount of grassland in the landscape, occurrence probability was best
explained by topographic and geographic location variables. In particular, grassland
songbird occurrence was often best explained by a combination of the modified soil-
adjusted vegetative index, compound topographic index, and heat load index, whereas
shorebird occurrence was most often associated with surface water and the compound
topographic index.
2.3.2 Detection Probability
Date influenced detection rates positively for Baird’s Sparrow, Chestnut-collared
Longspur, Horned Lark, Sprague’s Pipit, and negatively for Western Meadowlark and
Willet (Figure 2.3). Surveyors were more likely to detect Western Meadowlark and
Willet earlier in the season whereas Baird’s Sparrow, Chestnut-collared Longspur,
Horned Lark, and Sprague’s Pipit were more readily detected as the breeding season
progressed. Baird’s Sparrow, Clay-colored Sparrow, Long-billed Curlew, Savannah
Sparrow, Vesper Sparrow, and Western Meadowlark were most likely to be detected near
sunrise (Table 2.5). Conversely, detection of Horned Lark, Marbled Godwit, Sprague’s
Pipit, and Upland Sandpiper increased later in the morning. Increasing wind speeds
47
negatively influenced detection rates for Clay-colored Sparrow, Marbled Godwit,
Savannah Sparrow, Sprague’s Pipit, and Vesper Sparrow (Table 2.5).
2.3.3 Occurrence Probability
Brown-headed Cowbird, Horned Lark, Long-billed Curlew, Marbled Godwit, and
Vesper Sparrow probability of occurrence was negatively associated with the amount of
grassland habitat on the landscape (Table 2.5). This suggests that the occurrence of these
species was positively associated with the amount of cropland in the landscape since both
were inversely correlated (r=-0.87). In contrast, Baird’s Sparrow, Chestnut-collared
Longspur, and Sprague’s Pipit exhibited a strong positive association with increasing
grassland cover in the landscape (Table 2.5). Long-billed Curlew was the only species
that was negatively influenced by increasing proportion of woody vegetation on the
landscape (Table 2.5). Brown-headed Cowbird and Clay-colored Sparrow occurrence
was positively correlated with the modified soil-adjusted vegetation index, selecting areas
with increased green vegetation (Figure 2.4). Long-billed Curlew, McCown’s Longspur,
and Vesper Sparrow occurrence was negatively influenced by the presence of surface
water on the landscape (Table 2.5).
Occurrence of Baird’s Sparrow, Long-billed Curlew, Marbled Godwit, Sprague’s
Pipit, Upland Sandpiper, and Western Meadowlark was negatively correlated with the
compound topographic index, indicating they selected more xeric areas (Figure 2.5). In
contrast, Clay-colored and Vesper sparrows selected areas that were more mesic. The
occurrence of Baird’s Sparrow, Brown-headed Cowbird, Chestnut-collared Longspur,
Horned Lark, and Savannah Sparrow were negatively correlated with the heat load index,
48
indicating they were more likely to occur in areas that were exposed to less solar
radiation (Figure 2.6).
Latitude and longitude influenced occurrence probability for all species except
Brown-headed Cowbird and Upland Sandpiper (Table 2.5; Appendix B. Figure B.1, B.3-
B.10, B.12-B.14). Horned Lark, McCown’s Longspur, and Sprague’s Pipit occurrence
was most common towards the eastern portion of the study area (Appendix B. Figure B.5,
B.8, B.10) whereas Savannah Sparrow was most common towards the western portion
(Appendix B. Figure B.9). Baird’s Sparrow, Chestnut-collared Longspur, Horned Lark,
Long-billed Curlew, Marbled Godwit, McCown’s Longspur, Sprague’s Pipit, and
Western Meadowlark occurrence decreased towards the northern portion of the study
area whereas Clay-colored and Vesper sparrow occupancy increased northwards
(Appendix B. Figure B.1, B.3-8, B.10, B.12, B.13). Well density did not influence
occurrence of any of the grassland songbird species, but the occurrence of Long-billed
Curlew and Marbled Godwit were positively correlated with well density and Willet
occurrence was negatively correlated (Table 2.5).
2.3.4 Model Evaluation
I found that predictive performance typically increased once the 2012 testing data
were adjusted to account for imperfect detection (Table 2.6). I therefore evaluated
species models for predictive power using the adjusted testing data from 2012, and
determined that 8 of the 14 models were considered reliable based on receiver operating
characteristic curves and the AUC scores (Table 2.7). The model performances for
Western Meadowlark and Chestnut-collared Longspur were considered excellent
49
(AUC>0.90). Predictive performance of the McCown’s Longspur and Baird’s Sparrow
models were considered good (0.80<AUC>0.90), and the Sprague’s Pipit, Vesper
Sparrow, Horned Lark, and Savannah Sparrow models were considered useful
(0.70<AUC>0.80). Models for the remaining species did not adequately predict
occupancy (AUC<0.70). I found 11 of 14 species models were suitable predictors of
occupancy (k>0.20) when evaluated with the Kappa statistic. Baird’s Sparrow, Chestnut-
collared Longspur, Western Meadowlark, Sprague’s Pipit, and Vesper Sparrow were
considered good (0.40<k>0.75). Horned Lark, Willet, Savannah Sparrow, Clay-colored
Sparrow, Brown-headed Cowbird, and Marbled Godwit models were considered fair
(0.20<k>0.40). McCown’s Longspur, Long-billed Curlew and Upland sandpiper models
were considered poor and not useful predictors of occupancy (k<0.20).
I found that 6 of the 14 species models, when evaluated with the 2011 data that
was not corrected for detection bias, had suitable predictive performance based on
receiver operating characteristic curves and the AUC scores (Table 2.7). McCown’s and
Chestnut-collared longspurs models were considered good (0.80<AUC>0.90). Sprague’s
Pipit, Western Meadowlark, Baird’s Sparrow, and Clay-colored Sparrow models were
considered useful (0.70<AUC>0.80) but models for the remaining species did not
adequately predict occupancy. I found that 7 of 14 species models were suitable
predictors of occupancy when evaluated for Kappa. Chestnut-collared Longspur,
Sprague’s Pipit, Baird’s Sparrow, Clay-colored Sparrow, Western Meadowlark, Horned
Lark, Savannah Sparrow models were all considered fair (0.20<k>0.40) but the models
for the remaining species did not adequately predict species occupancy (k<0.20).
50
Overall, occupancy models tended to have the greatest predictive performance
when detection probability was moderate to high (ρ >0.5). Based on the combined
evaluation results of Kappa and AUC statistics, Baird’s Sparrow, Chestnut-collared
Longspur, Sprague’s Pipit, and Western Meadowlark models adequately predicted
occupancy across the study area (Table 2.7). However, Clay-colored Sparrow, Horned
Lark, Savannah Sparrow, and Vesper Sparrow models were marginal in their ability to
adequately predict occupancy, and the Brown-headed Cowbird, Long-billed Curlew,
Marbled Godwit, McCown’s Longspur, Upland Sandpiper, and Willet models were
unreliable.
2.4 Discussion
The usefulness of a species distribution model relies heavily on its predictive
performance (Fielding and Bell 1997; Guisan and Zimmermann 2000; Barry and Elith
2006). Tests for both AUC and Kappa are often used to measure the predictive
performance of a species distribution model, though AUC has received more attention
(Fielding and Bell 1997; Liu et al 2011). I evaluated the predictive performance of each
species occupancy model using two sets of testing data and tests for AUC and Kappa, as
both statistical measures are useful in comparing a model’s predictive performance, but
have known drawbacks. My results indicate that models for four species - Baird’s
Sparrow, Chestnut-collared Longspur, Sprague’s Pipit, and Western Meadowlark - were
“reliable” predictors of occupancy, and thus useful as conservation tools. Although the
McCown’s Longspur model performed within reliable limits when evaluated for AUC,
the model was deemed unsuccessful at predicting occupancy when evaluated using the
51
Kappa statistic; possibly because the statistic is strongly influenced by species prevalence
(Fielding and Bell 1997). This underscores the need to test for both discrimination
capacity and reliability (Austin 2007; Lobo et al. 2008). Unfortunately, the ten remaining
distribution models remain unreliable and require further research to develop accurate
predictions of occupancy.
2.4.1 Accounting for Imperfect Detection
I attempted to reduce the effects of detection variability by restricting counts to
certain calendar periods, time of day, and weather conditions. I also minimised the
effects of observer bias, roadside sampling bias, and within-season breeding dispersal.
Although I acknowledge that not all factors known to affect detectability were accounted
for (e.g., habitat features, predation, and social interaction), I further reduced detection
variability statistically, by adjusting point count data for imperfect detection (i.e., date,
time of day, wind). Thus, robust estimates of detection probability were used to build the
occupancy models, with many aspects of detection bias accounted for.
My results show that the detectability of seven species was influenced by survey
date. For temperate North American grasslands, Ralph et al. (1995) suggested that late
May to the first week in July was the period where most species are vocal and therefore
detected, but acknowledged that peak breeding activity often overlaps with this period.
Variability in singing rates have been hypothesised to be correlated with the breeding
stage of individuals (Diehl 1981; Wilson and Bart 1985), and males typically sing more
frequently during the mate selection stage or while establishing and defending a breeding
territory. As the reproductive season progresses towards incubation, and brood rearing,
52
males are less likely to defend breeding territories. For instance, I found that the
detectability of Western Meadowlark and Willet decreased throughout the survey season,
likely following the progression from egg-laying to brood rearing or lack of territory and
mate defence due to reproductive failure. Whereas, I found Baird’s Sparrow, Chestnut-
collared Longspur, and Sprague’s Pipit detectability increased as the survey season
progressed, indicative of later arriving migratory species that are in the mate selection
stages, or establishing and defending territories. I also established that the detectability
of Horned Lark increased as the survey season progressed, suggesting this early
migratory species may be re-nesting after nest failure or initiating multiple clutches and
still defending breeding territories.
My results indicate that the time of day a survey was conducted influenced the
detectability of nine species. Many species display temporal variation in singing
behaviour during the day that influences detectability (Robbins 1981; Johnson 1995;
Diefenbach et al. 2007; Elphick 2008). For example, most songbirds concentrate singing
activity during a three to five hour period after sunrise (Ralph et al. 1995, Staicer et al.
1996). My findings are consistent with this as the detectability of Baird’s Sparrow, Clay-
colored Sparrow, Long-billed Curlew, Savannah Sparrow, and Western Meadowlark was
greater earlier in the morning, a time when songbirds sing most intensely and frequently
(Staicer et al. 1996). Horned Lark, Marbled Godwit, Sprague’s Pipit, and Upland
Sandpiper peak vocalization periods occurred later in the morning, perhaps due to
increased temperature or enhanced visual display.
Wind speed can also influence detectability (Anderson and Ohmart 1977; Johnson
2008). I found that wind speed influenced the detection of six species despite restricting
53
my surveys to days and times when wind speed was <20 km/h. The degree to which
wind affects acoustic surveys depends on a combination of factors including the species,
habitat, and an observers hearing ability (Johnson 2008). Standardised count
methodology recommends that surveys should be discontinued if wind speeds reach 20
km/hr (Ralph et al. 1995), although the slightest of breezes can influence the ability of an
observer to detect singing males (Johnson 2008). My results suggest that the detectability
of Sprague’s Pipit, a species that primarily sings while conducting aerial displays
(Robbins and Dale 1999), was negatively influenced by wind. On the other hand, I found
a positive relationship between wind speed and the detectability of McCown’s Longspur,
a species which also sings while conducting aerial displays. Although McCown’s
Longspur displays are much shorter and closer to the ground (~10 m; With 2010), the
positive relationship to wind I found for this species is likely influenced by an observer
effect. Another conceivable explanation is that the small number of McCown’s Longspur
occurrences (detected at 3% of the point count locations in 2012) was correlated simply
by chance with higher amounts of wind. I also found that wind negatively influenced
detectability of Clay-colored, Savannah, and Vesper sparrows and highlights the
importance of ensuring that surveys are conducted during relatively calm conditions.
2.4.2 Resource Selection
My results show that integrating count data with landscape level features is an
effective way to understand resource selection by grassland birds. Occupancy models
have the implicit assumption that a species distribution can be accurately characterised,
and therefore predicted from spatial data (MacKenzie et al. 2006; Elith and Leathwick
54
2009). This relationship allows for distributions to be mapped without having to collect
fine-scale habitat data, which can be costly to acquire and often increases complexity and
uncertainty in the analysis (Cunningham and Johnson 2006). A substantial number of
metrics have proven to be useful in describing avian resource selection at the landscape
scale (Fahrig 2003). Identifying the amount of a targeted land cover type provides clear
recommendation for conservation management (Cunningham and Johnson 2011). I used
the percentage of grassland in a landscape as a broadly influential metric to explain
resource selection. Some grassland birds are known to be area sensitive (Johnson and Igl
2001; Ribic et al. 2009), and have also been shown to respond to metrics representing
vegetative structure and composition of plant communities across the landscape (Davis
2004). Habitat diversity or variation in the vegetation structure across the landscape is
related to moisture, aspect, topography, and disturbance patterns (Ribic et al. 2009).
Therefore, I further attempted to explain within-cover type variability by modelling
additional factors with known ecological relationships at the landscape scale. Explaining
the complex factors that influence where a bird occurs on the landscape is difficult;
however the occupancy estimates I generated provide a more complete explanation of
resource selection relative to cover type alone, and can aid in future conservation
management.
2.4.2.1 Response to Cover Type
The models I evaluated show that grassland bird occurrence varied with the
amount of grassland cover within 400 m of the survey location. The amount of grassland
cover had a positive influence for nine species indicating that these species were more
55
common in native and tame grassland landscapes. The selection of grassland habitats for
breeding by most of these species has been well documented (Owen and Myers 1973;
Johnson et al. 2004; Davis et al. 2013). Many species of birds found in grassland habitats
are uniquely adapted to the grassland biome and depend on it for various stages of their
life histories. However, some grassland birds now primarily depend on
anthropogenically altered habitats within the ecosystem to survive and reproduce. I
found that Brown-headed Cowbird, Horned Lark, Long-billed Curlew, and Vesper
Sparrow were negatively correlated with grassland cover, suggesting these species select
a variety of habitats to breed in, including cropland (Davis and Duncan 1999; Johnson et
al. 2004; Coppedge et al. 2006; Davis et al. 2013).
I found that occurrence patterns for six species were influenced by the modified
soil-adjusted vegetation index. Identifying where species more commonly occur in
response to changing plant communities across the landscape can be valuable for
providing a more complete explanation of resource selection, relative to cover type alone
(Cunningham and Johnson 2011). The modified soil-adjusted vegetation index I
constructed is correlated with the photosynthetic activity of vegetation, total plant cover,
bare ground, and plant biomass on the landscape, which was represented as the amount of
green vegetation within 400 m. Relationships between occupancy and amount of green
vegetation are not particularly well documented, but the modified soil-adjusted
vegetation index identifies similar vegetative characteristics to the known breeding
habitat requirements of grassland birds. For instance, Baird’s Sparrow, Chestnut-collard
Longspur, and Sprague’s Pipit are associated with grassland cover with short to moderate
vegetation (i.e., 20 - 100 cm) and low to moderate litter accumulation (Johnson et al.
56
2004), areas that the modified soil-adjusted vegetation index characterises across the
landscape as having lower amounts of green vegetation. A positive response to green
vegetation was identified for Clay-colored Sparrow, a species which tends to occur in
areas with high cover diversity, and moderate to high amounts of herbaceous biomass
(Davis and Duncan 1999).
I found little evidence that the amount of woody cover influenced grassland bird
resource selection. The negative response I found for Long-billed Curlew is similar to
that from other studies (Dugger and Dugger 2002). Long-billed Curlew prefers open
nesting areas with short vegetative cover, and also tend to use cropland dominated
landscapes for foraging, roosting, and nesting (Prescott 1997; Dechant et al. 2003). For
birds in grassland habitats, authors have often reported that the amount of woody cover
can be an important predictor of species occurrence (McMaster and Davis 2001: Bakker
et al. 2002; Fletcher and Koford 2002). For example, Prescott et al. (1995) and Davis and
Duncan (1999) determined Clay-colored Sparrows select habitats in areas with increased
woody cover. Conversely, McMaster and Davis (2001) and Davis et al. (1999) reported
that Chestnut-collared Longspur, McCown’s Longspur, and Sprague’s Pipit occur more
frequently in areas with low visual obstruction and little or no woody vegetation, perhaps
as a predator avoidance strategy. For instance, nest predation of McCown’s Longspur in
Colorado was six times higher when shrub cover was present within 1 m of the nest
(With 1994). Therefore, my findings, or lack thereof in some cases, suggest that woody
cover within 400 m may not be as strong of a predictor of species occurrence at the
landscape level, and examining smaller scale responses may better explain resource
selection. Another conceivable explanation for the weak response to woody cover that I
57
found is that the spatial data used to generate the values for the woody cover variable
were not accurate.
My models show that the amount of surface water within 400 m influenced
occurrence of four species. I hypothesised that the amount of surface water on the
landscape would positively influence water-dependent species. I found negative
responses to surface water for Long-billed Curlew and Vesper Sparrow, species that
usually occur in drier sites (Prescott et al. 1995; Gratto-Trevor 2000). My results show
Willet was positively influenced by the amount of surface water on the landscape,
possibly reflecting their occurrence at foraging sites. I also found McCown’s Longspur
occurrence was positively influenced by surface water, even though this species has been
said to prefer drier, sandier sites (Wershler et al. 1991). This longspur may be selecting
marginal breeding areas in my study area because this is the edge of its range. Another
plausible explanation is that the small number of McCown’s Longspur detections
(detected at 3 % of the point count locations in 2012) was correlated simply by chance
with higher amounts of surface water (i.e., location of the Red Deer River).
2.4.2.2 Responses to Topographic Landscape Features
I found evidence that eight of fourteen species responded to the compound
topographic index I used to assess variation in levels of soil moisture across the study
area. The compound topographic index takes into account topographic features including
slope, flow accumulation, flow direction, and contributing area, to form a representation
of the amount of soil moisture across the landscape. Soil moisture is known to influence
structural changes in plant communities and avian food sources (Wiens 1974; Cody 1985;
58
Wiens 1989). I found that six species selected breeding areas that were low lying and
thus able to retain more moisture. My results further show that two species, Clay-colored
and Vesper sparrows, selected areas that are less likely to accumulate moisture.
Determining why species respond to the compound topographic index at the landscape
level is difficult. My findings are not consistent between species, and broad scale
relationships between grassland birds and levels of soil moisture may also reflect
occurrence responses to topographic landscape features, structural changes in plant
communities, and annual precipitation events. Therefore, grassland bird responses to soil
moisture at the landscape level deserve further research, which will help to more
accurately describe species response patterns and lead to a better understanding of these
relationships.
Six of the species for whom I had enough data, occurred more frequently in
locations exposed to lower amounts of solar radiation. The heat load index I constructed
uses slope, aspect, and latitude to provide a measure of the variation in solar radiation
across the landscape. Most often, solar radiation is measured using proxy variables such
as slope or aspect, which often have indirect effects on the distribution of a species
(Austin 2002; Elith and Leathwick 2009). I found Baird’s Sparrow, Brown-headed
Cowbird, Chestnut-collared Longspur, Horned Lark, Savannah Sparrow, and Willet to
select sites in areas that received less solar radiation, perhaps due to the resulting
structural changes in plant communities produced by variation in the amount of solar
radiation. There are few published examples of bird distributions in relation to solar
radiation. This can be interpreted to mean that solar radiation is not a primary driver of
resource selection but it could also mean that it simply has not been considered to date.
59
Yet, high exposure to solar radiation has been shown to influence habitat diversity on the
landscape (McCune and Keon 2002; Natural Regions Committee 2006) and has been
hypothesised to reduce cold stresses associated with breeding birds (With and Webb
1993). Local variation in solar radiation across the landscape has also been suggested to
provide critical climate refugia (Austin and Niel 2010), perhaps influencing where
species occur. Therefore, the relationships I found between species occurrence and solar
radiation deserve further attention.
2.4.2.3 Influence of Geographic Location
I found that latitude and longitude were useful in predicting the geographic
location of where a species was likely to occur. Considering the large study area (20,000
km2), transition from prairie to parkland, and various anthropogenic activities occuring, I
predicted species occurrence would vary North-to-South and East-to-West. In addition to
broad-scale landscape changes in the study area, occurrence responses to the latitude and
longitude variables may also reflect gradients in precipitation, population distributions,
and other landscape characteristics (Niemuth et al. 2005; Elith and Leathwick 2009).
Higher occurrence was more common in the drier and less woody southern portion of the
study area for eight species, as compared to the more mesic and wooded areas to the
north. With the exception of Horned Lark, the distributions of these eight species are
near to their known northern limits in Alberta (Semenchuk 2007). In particular,
McCown’s Longspur was restricted to the south-eastern corner of the study area, likely
delineating the northwest edge of this species North American range. I also found that
Vesper and Clay-colored sparrow occurrence increased towards the northern portion of
60
the study area. This is consistent with their known distributions that extend into more
mesic and wooded areas (Semenchuk 2007). My models show that Horned Lark,
McCown’s Longspur, and Sprague’s Pipit occurrence increase towards the eastern
boundary of the study area, whereas Savannah Sparrow increased towards the western
boundary. Both Horned Lark and Savannah Sparrow occur further north and west than
my study area; therefore I consider my findings for these species to be indicative of a
regional selection within their overall range, perhaps due to the land cover matrix within
the study area. For example, Savannah Sparrows occur more commonly in areas without
extensive cultivation (McMaster and Davis 2001), whereas Horned Larks tend to occupy
cultivated areas (Owen and Myers 1973). The responses to longitude I found for
McCown’s Longspur and Sprague’s Pipit are consistent with these birds, likely reaching
the western extent of their ranges within the study area (Semenchuk 2007).
2.4.2.4 Influence of Well Density
Anthropogenic disturbances are likely to influence where a bird may occur on the
landscape (Askins et al. 2007), and our ability to characterise grassland bird responses to
these disturbances is limited. Oil and gas wells have a cumulative disturbance on the
landscape which includes the construction of roads and well pads, noise, vehicular traffic,
invasive plants, and soil compaction that may reduce both habitat quality and quantity
(Askins et al. 2007; Leu et al. 2008). I predicted that species would exhibit a strong
negative response to oil and gas well densities, especially in the southwest portion of my
study area where the degree of disturbance is high. However, the relationship between
species occurrence and well density was weak, and in the case of some shorebirds,
61
opposite to what I expected. Both Long-billed Curlew and Marbled Godwit were more
common in areas with higher well densities. Willet occurred more in areas with lower
well densities. Although I found that some species do respond negatively with increasing
well densities, the responses were not as strong as expected. Other studies in the mixed
northern grasslands (Dale et al. 2009; Hamilton et al. 2011; Kalyn Bogard and Davis
2014) have reported correlations between avian abundance and well density to vary
between species. For example, both Hamilton et al. (2011) and Kalyn Bogard and Davis
(2014) suggest that the abundance of the generalist Savannah Sparrow tends to increase
in areas with higher gas well densities. Sprague’s Pipit response varied between studies
as Hamilton et al. (2011) found a weak negative effect of well density on abundance
while Kayln Bogard and Davis (2014) found that abundance was not influenced by well
proximity or density. It is possible that some of the species I investigated respond to well
density in a specific manner, however I suggest that other factors, or combinations
thereof, need to be explored. For instance, refining the oil and gas well data by
separating active from inactive well sites may lead to improved estimates of occurrence
and identify specific ecological response patterns due to noise and increased human
activity on the landscape.
2.5 Conclusions
Understanding and accounting for potentially complex species-habitat
relationships is challenging. However, my results demonstrate that occupancy modelling
holds promise because it accounts for imperfect detection, while providing quantitative
insights into the magnitude and significance of species-habitat relationships. I found that
62
variation in seasonal breeding behaviour exists, and adjusting count data for imperfect
detection led to improved estimates of occupancy. Among the geographic and
environmental variables I used to predict occupancy, the composition of grassland cover
was a strong predictor of occurrence across species. Identifying species-specific
responses to grassland composition is useful for understanding where species occur on
the landscape, which provides a tool that conservation practitioners can use for
protecting, managing, and restoring habitats. By incorporating additional location-
specific landscape level factors associated with vegetation composition, topography, and
geographic location, my results provide a more complete explanation of resource
selection by the species I studied. My study also demonstrates that occupancy modelling
approaches hold promise for making justifiable inferences about species distribution. At
the same time, some of my predictions of species occupancy were weak and require
further investigation to develop more accurate predictions.
Overall, occupancy models provide intuitive, spatially referenced predictions that
conservation practitioners can incorporate into future conservation planning. The
modelling equations can be used to create maps of species distributions on the landscape,
which can be used to guide future management strategies, justify funding for research,
and prioritise conservation activities. Lastly, the modelling algorithm is extremely
flexible and can be extended to include occupancy based surveys and estimates of other
taxa. For example, if detection/non-detection data were readily available for other avian
species-at-risk such as Ferruginous Hawk (Buteo regalis) or Loggerhead Shrikes (Lanius
ludovicianus), conservation practitioners could apply occupancy modelling to minimise
63
the effects of imperfect detection and better understand these species distributions within
their range.
2.6 Literature Cited
AGRICULTURE AND AGRI-FOOD CANADA [AAFC]. 2012. 2011 AAFC crop type map of
Canada, ftp://ftp.agr.gc.ca/pub/outgoing/aesb-eos-gg/Crop_Inventory/2011/.
Accessed online February 16, 2012.
AKAIKE, H. 1974. A new look at the statistical model identification. IEEE Transactions
on Automatic Control AC 19: 716-723.
ALDRIDGE, C. L., AND M. S. BOYCE. 2007. Linking occurrence and fitness to
persistence: habitat-based approach for endangered greater sage-grouse.
Ecological Applications 2: 508-526.
ALLDREDGE, M. W., T. R. SIMONS, AND K. H. POLLOCK. 2007. Factors affecting aural
detections of songbirds. Ecological Applications 17: 948–955.
ALLOUCHE, O., A. TSOAR, AND R. KADMON. 2006. Assessing the accuracy of species
distribution models: prevalence, kappa, and the true skill statistic (TSS). Journal
of Applied Ecology 43(6): 1223–1232.
64
ANDERSON, B. W., AND R. D. OHMART. 1977. Climatological and physical
characteristics affecting avian population estimates in southwestern riparian
communities using transect counts. Pages 193–200 in R. R. Johnson and D. A.
Jones, editors. Importance, preservation and management of riparian habitat: a
symposium. U.S. Department of Agriculture Forest Service General Technical
Report RM-43, Fort Collins, Colorado, USA.
ARNOLD, T. W. 2010. Uninformative parameters and model selection using akaike’s
information criterion. Journal of Wildlife Management 74(6): 1175-1178.
ASKINS, R. A., F. CHAVEZ-RAMIEREZ, B. C. DALE, C. A. HAAS, J. R. HERKERT, F. L.
KNOPF, AND P. D. VICKERY. 2007. Conservation of grassland birds in North
America: understanding ecological processes in different regions. Ornithological
Monographs 64: 1-46.
AUSTIN, M. P. 2002. Spatial prediction of species distribution: an interface between
ecological theory and statistical modelling. Ecological Modelling 157: 101-18.
AUSTIN, M. P. 2007. Species distribution models and ecological theory: a critical
assessment and some possible new approaches. Ecological Modelling 200(1–2):
1–19.
65
AUSTIN, M. P., AND K. P. VAN NIEL 2010. Improving species distribution models for
climate change studies: variable selection and scale. Journal of Biogeography
38(1): 1-8.
BAKKER, K. K., D. E. NAUGLE, AND K. F. HIGGINS. 2002. Incorporating landscape
attributes into models for migratory grassland bird conservation. Conservation
Biology 16: 1638-1646.
BARRY, S., AND J. ELITH. 2006. Error and uncertainty in habitat models. Journal of
Applied Ecology 43(3): 413-423.
BARTON, K. 2013. MuMIn: muti-model inference. R package version 1.9.5.
http://CRAN.R-project.org/package=MuMIn.
BERLANGA, H. K., J. A. RICH, T. D. ARIZMENDI, M. C. BEARDMORE, C. J. BLANCHER,
P. J. BUTCHER, G. S. COUTURIER, A. R. DAYER, A. A. DEMAREST, D. W.
EASTON, W. E. GUSTAFSON, M. INIGO-ELIAS, E. KREBS, E. A. PANJABI, A. O.
RODRIGUEZ CONTRERAS, V. ROSENBERG, K. V. RUTH, J. M. SANTANA
CASTELLÓN, E. VIDAL, R. MA, AND T. WILL. 2010. Saving Our Shared Birds:
Partners in Flight Tri-National Vision for Landbird Conservation. Cornell Lab of
Ornithology: Ithaca, NY. Available at: www.savingoursharedbirds.org. Accessed
online October 5, 2012.
66
BRENNAN, L. A. AND W. P. KUVLESKY. 2005. North American grassland birds: An
unfolding conservation crisis? Journal of Wildlife Management 69: 1-13.
BROTONS, L., A. WOLFF, G. PAULUS, AND J-L. MARTIN. 2005. Effect of adjacent
agricultural habitat on the distribution of passerines in natural grasslands.
Biological Conservation 124: 407–414.
BURNHAM, K. P., AND D. R. ANDERSON. 1998. Model selection and inference: a
practical information - theoretic approach./Springer-Verlag, New York, USA.
BURNHAM, K. P., AND D. R. ANDERSON. 2002. Model selection and multimodel
inference: a practical information - theoretic approach, 2nd ed. -/ Springer-Verlag,
New York, USA.
CHANDLER, R. B., D. I. KING, AND S. DESTEFANO. 2009. Scrub–shrub bird habitat
associations at multiple spatial scales in beaver meadows in Massachusetts. Auk
126: 186–197.
CODY, M. L. 1985. Habitat selection in grassland and open-country birds. Pp. 191-226 in
M. L. Cody (editor). Habitat selection in birds. Academic Press, Orlando, FL.
67
COPPEDGE, B. R., D. M. ENGLE, R. E. MASTERS, AND M. S. GREGORY. 2006.
Development of a grassland integrity index based on breeding bird assemblages.
Environmental Monitoring and Assessment 118:125–145.
COSEWIC. 2011. Canadian wildlife species at risk. Committee on the status of
endangered wildlife in Canada. Available at:
http://www.cosewic.gc.ca/eng/sct0/rpt/rpt_csar_e.cfm. Accessed online May 4,
2011.
CUNNINGHAM, M. A., AND D. H. JOHNSON. 2006. Proximate and landscape factors
influence grassland bird distributions. Ecological Applications 16(3): 1062-1075.
CUNNINGHAM, M. A., AND D. H. JOHNSON. 2011. Seeking parsimony in landscape
metrics. Journal of Wildlife Management 75(3): 692-701.
DALE, B. C., T. S. WIENS, AND L. E. HAMILTON. 2009. Abundance of three grassland
songbirds in an area of natural gas infill drilling in Alberta, Canada. Proceedings
of the Fourth International Partners in Flight Conference: Tundra to Tropics 194–
204.
DAVIS, S. K. 2004. Area sensitivity in grassland passerines: effects of patch size, patch
shape, and vegetation structure on bird abundance and occurrence in southern
Saskatchewan. Auk 4: 1130-1145.
68
DAVIS, S. K. 2011. Canadian Wildlife Service draft Sprague’s Pipit (Anthus spragueii)
resource selection function model. Unpublished raw data.
DAVIS, S. K., D. C. DUNCAN, AND M. A. SKEEL. 1999. Distribution and habitat
associations of three endemic grassland songbirds in southern Saskatchewan.
Wilson Bulletin 111: 389-396.
DAVIS, S. K. AND D. C. DUNCAN. 1999. Grassland songbird occurrence in native and
crested wheatgrass pastures of southern Saskatchewan. Studies in Avian Biology
19: 211-218.
DAVIS, S. K., R. J. FISHER, S. L. SKINNER, T. L. SHAFFER, AND R. M. BRIGHAM. 2013.
Songbird abundance in native and planted grassland varies with type and amount
of grassland in the surrounding landscape. Journal of Wildlife Management 77(5):
908-919.
DECHANT, J. A., M. L. SONDREAL, D. H. JOHNSON, L. D. IGL, C. M. GOLDADE, P. A.
RABIE, AND B. R. EULISS. 2003. Effects of management practices on grassland
birds: Long-billed Curlew. Northern Prairie Wildlife Research Center,
Jamestown, ND. Northern Prairie Wildlife Research Center
Online. http://www.npwrc.usgs.gov/resource/literatr/grasbird/lbcu/lbcu.htm
(Version 12DEC2003). Accessed online July 5, 2013.
69
DIEFENBACH, D. R., M. R. MARSHALL, J. A. MATTICE, AND D. W. BRAUNING. 2007.
Incorporating availability for detection in estimates of bird abundance. Auk 124:
96-106.
DIEHL, B. 1981. Bird populations consist of individuals differing in many respects. Pages
225–229 in C. J. Ralph and J. M. Scott, editors. Estimating numbers of terrestrial
birds. Studies in Avian Biology 6. Cooper Ornithological Society, Camarillo,
California, USA.
DONOVAN, T. M., C. J. BEARDMORE, D. N. BONTER, J. D. BRAWN, R. J. COOPER, J. A.
FITZGERALD, R. FORD, S. A. GAUTHREAUX, T. L. GEORGE, W. C. HUNTER, T.
E. MARTIN, J. PRICE, K. V. ROSENBERG, P. D. VICKERY, AND T. B. WIGLEY.
2002. Priority research needs for the conservation of Neotropical migrant
landbirds. Journal of Field Ornithology 73: 329-339.
DUGGER, B. D., AND K. M. DUGGER. 2002. Long-billed Curlew (Numenius americanus),
The Birds of North America Online (A. Poole, Ed.). Ithaca: Cornell Lab of
Ornithology; Retrieved from the Birds of North America Online:
http://bna.birds.cornell.edu/bna/species/628. Accessed online June 21, 2013.
DUNNING, J. B. JR, D. J. STEWART, B. J. DANIELSON, B. R. NOON, T. L. ROOT, R. H.
LAMBERSON, AND E. E. STEVENS. 1995. Spatially explicit population models:
current forms and future uses. Ecological Applications 5: 3-11.
70
ELITH, J., AND J. R. LEATHWICK. 2009. Species distribution models: ecological
explanation and prediction across space and time. Annual Review of Ecology,
Evolution, and Systematics 40: 677-697.
ELPHICK, C. S. 2008. How you count counts: The importance of methods research in
applied ecology. Journal of Applied Ecology 45: 1313–1320.
ESRI. 2011. ArcGIS Desktop: Release 10. Redlands, CA: Environmental Systems
Research Institute.
FAHRIG, L. 2003. Effects of habitat fragmentation on biodiversity. Annual Review of
Ecology, Evolution, and Systematics 34:487–515.
FARNSWORTH, G. L., K. H. POLLOCK, J. D. NICHOLS, T. R. SIMONS, J. E. HINES, AND J.
R. SAUER. 2002. A removal model for estimating detection probabilities from
point-count surveys. Auk 119: 414-425.
FIELDING, A. H., AND J. F. BELL. 1997. A review of methods for the assessment of
prediction errors in conservation presence/absence models. Environmental
Conservation 24: 38–49.
71
FINCH, D. M. 1991. Population ecology, habitat requirements, and conservation of
neotropical migratory birds. USDA Forest Service General Technical Report RM-
205. Available at:
http://www.fws.gov/southwest/es/documents/R2ES/LitCited/LPC_2012/Finch_19
91.pdf. Accessed online August 28, 2012.
FISKE, I. J. AND R. B. CHANDLER. 2011. unmarked: An R package for fitting hierarchical
models of wildlife occurrence and abundance. Journal of Statistical Software
43:10 1-23. http://www.jstatsoft.org/v43/i10/.
FISHER, R. J. AND S. K. DAVIS. 2010. From Wiens to Robel: a review of grassland bird
habitat selection. Journal of Wildlife Management 74: 265-243.
FLETCHER, R. J., AND R. R. KOFORD. 2002. Habitat and landscape associations of
breeding birds in native and restored grasslands. Journal of Wildlife Management
66: 1011-1022.
FRANKLIN, J. 2010. Moving beyond static species distribution models in support of
conservation biogeography. School of Geographical Sciences and Urban Planning
and School of Life Sciences, Arizona State University, Tempe AZ USA.
72
FREEMAN, E. A., AND G. G. MOISEN. 2008. PresenceAbsence: An R package for
presence-absence model analysis. Journal of Statistical Software 23(11):1-31.
http://www.jstatsoft.org/v23/i11.
GESSLER, P. E., I. D. MOORE, N. J. MCKENZIE, AND P. J. RYAN. 1995. Soil-landscape
modeling and spatial prediction of soil attributes. International Journal of GIS
9(4):421-432.
GRATTO-TREVOR, C. 2000. Use of managed and natural wetlands by upland breeding
shorebirds in southern Alberta. Proceedings of the fifth prairie conservation and
endangered species conference, Saskatoon, Saskatchewan.
GU, W., AND R. K. SWIHART. 2004. Absent or undetected? Effects of non-detection of
species occurrence on wildlife-habitat models. Biological Conservation 116(2):
195-203.
GUISAN, A., AND N. E. ZIMMERMAN. 2000. Predictive habitat distribution models in
ecology, Ecological Modelling 135: 147–186.
GUISAN, A., AND W. THUILLER. 2005. Predicting species distribution: offering more
than simple habitat models. Ecological Letters 8: 993–1009.
73
HAMILTON, L. E., B. C. DALE, AND C. A. PASZKOWSKI. 2011. Effects of disturbance
associated with natural gas extraction on the occurrence of three grassland
songbirds. Avian Conservation and Ecology 6(1): 7.
HIJMANS, R. J., AND J. VAN ETTEN. 2013. Raster: Geographic data analysis and
modeling. R package version 2.1-25. http://CRAN.R-project.org/package=raster.
HUBERTY, C.J. 1994. Applied Discriminant Analysis. New York, USA: Wiley
Interscience: 466.
HUTTO, R. L. 1985. Habitat selection by nonbreeding, migratory land birds. Pp. 455-476
in M. L. Cody (ed.) Habitat selection in birds. Academic Press, Inc., Orlando.
HUTTO, R. L., S. M. PLETSCHET, AND P. HENDRICKS. 1986. A fixed-radius point count
method for nonbreeding and breeding season use. Auk 103:593-602.
IHS ENERGY GROUP. 2012. IHS Well Data (Canada): IHS Energy Group; online
database available from IHS Energy Regional Headquarters, 1331 MacLeod Trail,
Suite 200, Calgary, AB T2G 0K3, Canada. Accessed online December 12, 2012.
JOHNSON, D. H. 1980. The comparison of usage and availability measurements for
evaluating resource preference. Ecology 61: 65–71.
74
JOHNSON, D. H. 1995. Point counts of birds: What are we estimating? Monitoring bird
populations by point counts. U.S. Department of Agriculture, Forest Service
General Technical Report PSW-GTR-149: 117-123. Available at:
http://www.npwrc.usgs.gov/resource/birds/ptcounts/. Accessed online July 23,
2012.
JOHNSON, D. H. 2008. In defense of indices: the case of bird surveys. Journal of Wildlife
Management 72: 857–868.
JOHNSON, D. H., AND L. D. IGL. 2001. Area requirements of grassland birds: a regional
perspective. Auk 118: 24-34.
JOHNSON, D. H., L. D. IGL, AND J. A. DECHANT SHAFFER (SERIES COORDINATORS).
2004. Effects of management practices on grassland birds. Northern Prairie
Wildlife Research Center, Jamestown, ND. Jamestown, ND: Northern Prairie
Wildlife Research Center Online.
http://www.npwrc.usgs.gov/resource/literatr/grasbird/index.htm
(Version 12AUG2004). Accessed online August 1, 2012.
KALYN-BOGARD, H. J. 2011. Natural gas development and grassland songbird
abundance in south-western Saskatchewan: the impact of gas wells and
cumulative disturbance. M.Sc. thesis. University of Regina, Regina.
75
KALYN-BOGARD, H. J., AND S. K. DAVIS. 2014. Grassland songbirds exhibit variable
responses to the proximity and density of natural gas wells. The Journal of
Wildlife Management 78(3): 471-482.
LANDIS, J. R., AND G. G. KOCH. 1977. The measurement of observer agreement for
categorical data. Biometrics. 33: 159–174.
LEU, M., S. E. HANSER, AND S. T. KNICK. 2008. The human footprint in the west: a
largescale analysis of anthropogenic impacts. Ecological Applications 5: 1119-
1139.
LIU, C., M. WHITE, AND G. NEWELL. 2011. Measuring and comparing the accuracy of
species distribution models with presence-absence data. Ecography 34: 232-243.
LOBO, J. M., A. JIMÉNEZ-VALVERDE, AND R. REAL. 2008. AUC: a misleading measure
of the performance of predictive distribution models. Global Ecology and
Biogeography 17: 145-51.
MACKENZIE, D. I., J. D. NICHOLS, G. B. LACHMAN, S. DROEGE, J. A. ROYLE, AND C.
A. LANGTIMM. 2002. Estimating site occupancy rates when detection
probabilities are less than one. Ecology 83: 2248–2255.
76
MACKENZIE, D. I., AND J. A. ROYLE. 2005. Designing efficient occupancy studies:
General advice and tips on allocation of survey effort. Journal of Applied Ecology
42: 1105-1114.
MACKENZIE, D. I., J. D. NICHOLS, J. A. ROYLE, K. P. POLLOCK, L. L. BAILEY, AND J.
E. HINES. 2006. Occupancy estimation and modeling: inferring patterns and
dynamics of species occurrence. Academic Press, San Diego, California, USA.
MANLY, B. F. J., L. L. MCDONALD, D. L. THOMAS, T. L. MCDONALD, AND W. P.
ERICKSON. 2002. Resource selection by animals: statistical design and analysis
for field studies. Second Edition. Kluwer, Boston, Massachusetts, USA.
MARSH, D. M., AND P. C. TRENHAM. 2008. Current trends in monitoring programs for
animal and plant populations. Conservation Biology 22: 647-655.
MCCUNE, B. AND D. KEON. 2002. Equations for potential annual direct incident
radiation and heat load. Journal of Vegetation Science 13: 603-606.
MCKERCHER, R. B., AND B. WOLFE. 1986. Understanding Western Canada's Dominion
Land Survey System. Division of Extension and Community Relations report,
University of Saskatchewan, Saskatoon. 26 pp.
77
MCMASTER, D. G. AND S. K. DAVIS. 2001. An evaluation of Canada's Permanent Cover
Program: habitat for grassland birds? Journal of Field Ornithology 72:195-210.
MILLENNIUM ECOSYSTEM ASSESSMENT. 2005. Ecosystems and human well-being:
Biodiversity synthesis. World Resource Institute, Washington, DC.
NATURAL REGIONS COMMITTEE. 2006. Natural regions and subregions of Alberta.
Compiled by D.J. Downing and W.W. Pettapiece. Government of Alberta. Pub.
No. T/852.
NATURAL RESOURCES CANADA. 2004a. Canadian Digital Elevation Data, Earth Science
Sector, Centre for Topographic Information,
http://www.geobase.ca/geobase/en/data/cded/index.html. Accessed online
November 28, 2012.
NATURAL RESOURCES CANADA. 2004b. Landsat 7 Orthorectified Imagery over Canada,
Earth Science Sector, Centre for Topographic Information,
ftp://ftp.geogratis.gc.ca/landsat_7/ortho/ Accessed online November 26, 2012.
NATURAL RESOURCES CANADA. 2007. National Hydro Network, Canada, Earth Science
Sector, Centre for Topographic Information,
http://geobase.ca/geobase/en/data/nhn/index.html. Accessed online December 10,
2012.
78
NIEMUTH, N. D., M. E. ESTEY, AND C. R. LOESCH. 2005. Developing spatially explicit
habitat models for grassland bird conservation planning in the Prairie Pothole
Region of North Dakota. Pages 469-477 in Bird Conservation Implementation
and Integration in the Americas: Proceedings of the Third International Partners
in Flight Conference 2002, C.J. Ralph and T.D. Rich, editors. USDA Forest
Service PSW-GTR-191, Albany, Califonia, USA.
OWENS, R. A., AND M. T. MYRES. 1973. Effects of agriculture upon populations of
native passerine birds of an Alberta fescue grassland. Canadian Journal of
Zoology 51:697-713.
PAGANO, A. M., AND T. W. ARNOLD. 2009. Estimating detection probabilities of
waterfowl broods from ground-based surveys. Journal of Wildlife Management
73: 686-694.
PEARCE, J. L., AND S. FERRIER. 2000. Evaluating the predictive performance of habitat
models developed using logistic regression. Ecological Modelling 133: 225-/245.
PEARCE, J. L., L. A. VENIER, S. FERRIER AND D. W. MCKENNEY. 2002. Measuring
prediction uncertainty in models of species distribution. In: Predicting Species
Occurrences: Issues of Accuracy and Scale (eds Scott, J.M., Heglund, P.J.,
Morrison, M.L., Haufler, J.B., Raphael, M.G., Wall, W.A. & Samson, F.B.).
Island Press, Covelo, CA, pp. 383–397.
79
PORTER, W. F., AND K. E. CHURCH. 1987. Effects of environmental patterns on habitat
preference analysis. Journal of Wildlife Management 51: 681-685.
PRESCOTT, D. R. C. 1997. Avian communities and NAWMP habitat priorities in the
northern prairie biome of Alberta. NAWMP-029. Land Stewardship Centre of
Canada, St. Albert, Alberta. 41 pages.
PRESCOTT, D. R. C., A. J. MURPHY, AND E. EWASCHUK. 1995. An avian community
approach to determining biodiversity values of NAWMP habitats in the aspen
parkland of Alberta. NAWMP-012. Alberta NAWMP Centre, Edmonton, Alberta.
58 pages.
QI, J., A. CHEHBOUNI, A. R. HUETE, Y. H. KEER, AND S. SOROOSHIAN . 1994. A
modified soil vegetation adjusted index. Remote Sensing of Environment 48:
119–126.
R DEVELOPMENT CORE TEAM. 2013. R: a language and environment for statistical
computing. R Foundation for Statistical Computing, Vienna, Austria, www.R-
project.org. R version 3.0.0 (2013-04-03) -- "Masked Marvel" Copyright (C) 2013
The R Foundation for Statistical Computing Platform: x86_64-w64-mingw32/x64
(64-bit).
80
RALPH, C. J., J. R. SAUER, AND S. DROEGE (EDITORS). 1995. Monitoring bird
populations by point counts. U.S. Department of Agriculture Forest Service,
Pacific Southwest Research Station, General Technical Report PSWGTR-149,
Albany, California, USA.
RANGEL, T. F. L. V. B., J. A. F DINIZ-FILHO, AND L. M. BINI. 2006.Towards an
integrated computational tool for spatial analysis in macroecology and
biogeography. Global Ecology and Biogeography 15: 321–27.
RIBIC, C. A., AND D. W. SAMPLE. 2001. Associations of grassland birds with landscape
factors in southern Wisconsin. American Midland Naturalist 146: 105-121.
RIBIC, C. A., R. R. KOFORD, J. R. HERKERT, D. H. JOHNSON, N. D. NIEMUTH, D. E.
NAUGLE, K. K. BAKKER, D. W. SAMPLE, AND R. B. RENFREW. 2009. Area
sensitivity in North American grassland birds: patterns and processes. Auk 126:
233–244.
ROBBINS, C. S. 1981. Effect of time of day on bird activity. Pages 275–286 in C. J. Ralph
and J. M. Scott, editors. Estimating numbers of terrestrial birds. Studies in Avian
Biology 6. Cooper Ornithological Society, Camarillo, California, USA.
81
ROBBINS, C. S., D. K. DAWSON, AND B. A. DOWELL. 1989. Habitat area requirements of
breeding forest birds of the middle Atlantic states. Wildlife Monographs 103: 1–
34.
ROBBINS, M. B., AND B. C. DALE. 1999. Sprague's Pipit (Anthus spragueii), The Birds
of North America Online (A. Poole, Ed.). Ithaca: Cornell Lab of Ornithology;
Retrieved from the Birds of North America Online:
http://bna.birds.cornell.edu/bna/species/439. Accessed online August 14, 2013.
ROTA, C. T., R. J. FLETCHER, J. M. EVANS, AND R.L. HUTTO. 2011. Does accounting for
imperfect detection improve species distribution models? Ecography 34: 659-670.
SAMSON, F. AND F. KNOPF. 1994. Prairie Conservation in North America. BioScience
44: 418-421.
SEMENCHUK, G. P. 2007. The atlas of breeding birds of Alberta: a second look.
Federation of Alberta Naturalists, Edmonton, Alberta, Canada.
STAICER, C. A., D. A. SPECTOR, AND A. G. HORN. 1996. The dawn chorus and other diel
patterns in acoustic signaling. In Ecology and Evolution of Acoustic
Communication in Birds (D. E. Kroodsma and E. H. Miller, eds.), pp. 426–453.
Ithaca, NY: Cornell University Press.
82
TERBORGH, J. 1989. Where have all the birds gone? Princeton University Press,
Princeton, New Jersey, USA.
TYRE, A. J., B. TENHUMBERG, S. A. FIELD, D. NIEJALKE, K. PARRIS, AND H. P.
POSSINGHAM. 2003. Improving precision and reducing bias in biological surveys:
estimating false-negative error rates. Ecological Applications 13: 1790-1801.
VICKERY, P. D., P. L. TUBARO, J. M. CARDOSO DA SILVA, B. G. PETERJOHN, J. R.
HERKERT, AND R. B. CAVALCANTI. 1999. Conservation of grassland birds in the
western hemisphere. Studies in Avian Biology 19: 2-26.
WERSHLER, C., W. W. SMITH, AND C. WALLIS. 1991. Status of the Baird’s sparrow in
Alberta: 1987/1988 update with notes on other grassland sparrows and Sprague’s
pipit. Proceedings of the Endangered Species and Prairie Conservation Workshop
2: 87–89.
WHITE, R. P., S. MURRAY AND M. ROWEDER. 2000. Pilot analysis of global ecosystems:
grassland ecosystems. World Resource Institute, Washington, D.C.
WIENS, J. A. 1973. Interterritorial habitat variation in grasshopper and Savannah
sparrows. Ecology 54: 877–884.
83
WIENS, J. A. 1974. Climatic instability and the "ecological saturation" of bird
communities in North American grasslands. Condor 76: 385-400.
WIENS, J.A. 1989. Spatial scaling in ecology. Functional Ecology 3: 385-397.
WILSON, D. M., AND J. BART. 1985. Reliability of singing bird surveys: effects of song
phenology during the breeding season. Condor 87: 69–73.
WITH, K. A. 1994. The hazards of nesting near shrubs for a grassland bird, the McCown's
Longspur. Condor 96: 1009-1019.
WITH, K. A. 2010. McCown's Longspur (Rhynchophanes mccownii), The Birds of North
America Online (A. Poole, Ed.). Ithaca: Cornell Lab of Ornithology; Retrieved
from the Birds of North America Online:
http://bna.birds.cornell.edu/bna/species/096. Accessed online August 15, 2013.
WITH, K. A., AND D. R. WEBB. 1993. Microclimate of ground nests: the relative
importance of radiative cover and wind breaks for three grassland species. Condor
95: 401-413.
ZIMMERMAN, N. E., T. C. EDWARDS, G. G. MOISEN, T. S. FRESCINO, AND J. A.
BLACKARD. 2007. Remote sensing-based predictors improve distribution models
84
for rare, early successional and broadleaf tree species in Utah. Journal of Applied
Ecology 44: 1057-1067.
85
Figure 2.1. Geographic location of the Special Areas municipality (shaded grey) in
south-eastern Alberta, Canada.
86
Figure 2.2. Point count survey locations conducted in the Special Areas, 2011-2012.
Black circles indicate 2012 survey locations (visited using a repeated survey design) and
grey circles indicate 2011 survey location (visited using a removal sampling design).
87
Table 2.1. Description of each variable used to build the species occupancy models with
the detection/non-detection avian survey records collected in the Special Areas, 2012.
CLASS VARIABLE VARIABLE
CODE
VARIABLE
DESCRIPTION SOURCE REFERENCE
Detection
Date JDATE Ordinal date of
survey during the
survey period
Visit-
specific
Repeated
surveys, 2012
Time TIME Time of day survey
was conducted
Visit-
specific
Repeated
surveys, 2012
Wind WIND Wind (km/h) during
point count visit
Visit-
specific
Repeated
surveys, 2012
Land Cover
Percentage of
grassland
(400m radius)
GRASS Grassland and
Pastureland/Hayland
categories combined
Agriculture
and Agri-
foods
Canada
AAFC 2012
Percentage of
woody cover (400m radius)
WOODY Shrubland,
Coniferous, and Deciduous
categories combined
Agriculture
and Agri-foods
Canada
AAFC 2012
Percentage of
surface water
(400m radius)
WATER Watercourses,
Wetlands, and
Waterbodies
categories combined
GeoBase.ca Natural
Resources
Canada 2007
Percentage of
soil-adjusted
vegetative
index (400m
radius)
MSAVI Calculated from Red
band and Near-
infrared band data
GeoBase.ca Natural
Resources
Canada
2004b
Topography
Compound
topographic
index (400m
radius)
CTI Index of soil
moisture (DEM)
GeoBase.ca
(1:50,000
map sheets)
Natural
Resources
Canada 2004a
Heat load
index (400m
radius)
HLI Index of solar
radiation (DEM)
GeoBase.ca
(1:50,000
map sheets)
Natural
Resources
Canada 2004a
Geographic
location
Latitude YLAT North and South
geographic
distribution
30 x 30 m
study area
grid
Longitude XLONG West and East
geographic
distribution
30 x 30 m
study area
grid
88
Density
Well density
in km/2 (400m
radius)
WD Oil and Gas well
density
IHS Well
Data
(Canada):
online database
IHS Energy
Group 2012
89
Table 2.2. Pearson’s correlation coefficients for selected variables used to explain
species occurrence in the Special Areas, Alberta. Values are for the 400 m landscape
scale and were used to examine collinearity between predictor variables. Variable codes
are found in Table 2.1.
WD GRASS WOODY CTI HLI MSAVI WATER XLONG YLAT
WD 1.00
GRASS 0.10 1.00
WOODY 0.00 -0.14 1.00
CTI -0.04 0.14 -0.11 1.00
HLI 0.00 -0.03 0.22 -0.72 1.00
MSAVI -0.11 -0.11 0.18 -0.18 0.18 1.00
WATER -0.03 0.01 -0.04 0.14 -0.08 0.02 1.00
XLONG -0.02 -0.30 0.04 -0.27 0.20 0.05 -0.12 1.00
YLAT -0.11 -0.24 0.18 -0.04 0.12 0.48 0.04 0.08 1.00
90
Table 2.3. Ranked species frequency of occurrence (%) based on the 2012 data used to
build species occupancy models, with mean occurrence rates (Ψ) and detection
probabilities (ρ) from surveys conducted in the Special Areas, Alberta.
SPECIES
FREQUENCY
(% of sites
detected)
OCCURRENCE
PROBABILITY
(Ψ)
DETECTION
PROBABILITY
(ρ)
Western Meadowlark (Sturnella neglecta) 0.96 0.99 0.91
Vesper Sparrow (Pooecetes gramineus) 0.82 0.93 0.70
Savannah Sparrow (Passerculus sandwichensis) 0.79 0.84 0.76
Brown-headed Cowbird (Molothrus ater) 0.70 0.83 0.70
Clay-colored Sparrow (Spizella pallida) 0.69 0.76 0.74
Horned Lark (Eremophila alpestris) 0.62 0.75 0.59
Sprague's Pipit (Anthus spragueii) 0.48 0.45 0.67
Baird's Sparrow (Ammodramus bairdii) 0.39 0.33 0.61
Marbled Godwit (Limosa fedoa) 0.34 0.64 0.26
Willet (Tringa semipalmata) 0.27 0.55 0.20
Chestnut-collared Longspur (Calcarius ornatus) 0.21 0.07 0.59
Long-billed Curlew (Numenius americanus) 0.19 0.27 0.24
Upland Sandpiper (Bartramia longicauda) 0.11 0.18 0.67
McCown's Longspur (Rhynchophanes mccownii) 0.03 0.0002 0.19
91
Table 2.4. Top-ranking models explaining the influence of spatial and temporal variables
on species detection (ρ) and occurrence (Ψ) in the Special Areas, 2012. Models include a
detection component (date [JDATE], time [TIME], wind [WIND]) and occurrence
component (grassland [GRASS], woody cover [WOODY], vegetative greenness
[MSAVI], surface water [WATER], solar radiation [HLI], wetness [CTI], latitude
[YLAT], longitude [XLONG], and well density [WD]). Only those models with a
cumulative Akaike Information Criterion (AIC) weight (wi) of 0.95 for each species with
corresponding number of parameters (K), value of the negative log-likelihood function
(LogLik), and the change in AIC value from the top model (∆AIC) are presented.
SPECIES MODEL K LogLik ∆AIC wi
Baird's
Sparrow ρ(JDATE-TIME)Ψ(GRASS-HLI-MSAVI-YLAT) 8 -744.75 0.00 0.27
ρ(JDATE-TIME)Ψ(GRASS+CTI-HLI-MSAVI-YLAT) 9 -743.77 0.04 0.27
ρ(JDATE-TIME)Ψ(GRASS+CTI-HLI-YLAT) 8 -745.24 0.98 0.17
ρ(JDATE-TIME)Ψ(GRASS+CTI-MSAVI-YLAT) 8 -745.28 1.07 0.16
ρ(JDATE-TIME)Ψ(GRASS-HLI-YLAT) 7 -746.49 1.47 0.13
Brown-
headed
Cowbird
ρ(-JDATE)Ψ(-GRASS-HLI+MSAVI) 6 -1259.60 0.00 0.81
ρ(-JDATE)Ψ(-GRASS+MSAVI) 5 -1262.05 2.92 0.19
Chestnut-
collared
Longspur
ρ(JDATE)Ψ(GRASS-MSAVI-HLI-YLAT) 7 -445.78 0.00 1.00
Clay-
colored
Sparrow
ρ(-TIME-WIND)Ψ(GRASS-CTI+MSAVI+YLAT) 8 -1073.84 0.00 0.77
ρ(-TIME-WIND)Ψ(GRASS-CTI+YLAT) 7 -1076.60 3.53 0.13
ρ(-TIME)Ψ(GRASS-CTI+MSAVI+YLAT) 7 -1076.92 4.17 0.10
Horned
Lark ρ(JDATE+TIME)Ψ(-GRASS-HLI-MSAVI+XLONG-YLAT) 9 -1127.31 0.00 0.58
ρ(TIME)Ψ(-GRASS-HLI-MSAVI+XLONG-YLAT) 8 -1128.65 0.68 0.42
Long-billed Curlew
ρ(-TIME)Ψ(-GRASS+CTI-WATER+WD-WOODY-YLAT) 9 -481.91 0.00 0.35
ρ(-TIME)Ψ(-GRASS+CTI-WATER-WOODY-YLAT) 8 -483.53 1.23 0.19
ρ(-TIME)Ψ(-GRASS+CTI-WATER+WD-YLAT) 8 -484.20 2.58 0.10
ρ(-TIME)Ψ(-GRASS+CTI+WD-WOODY-YLAT) 8 -484.27 2.71 0.09
ρ(-TIME)Ψ(-GRASS+CTI-WATER-YLAT) 7 -485.59 3.36 0.07
ρ(.)Ψ(-GRASS+CTI-WATER+WD-WOODY-YLAT) 8 -484.77 3.71 0.05
ρ(-TIME)Ψ(-GRASS+CTI-WOODY-YLAT) 7 -485.94 4.06 0.05
92
ρ(.)Ψ(-GRASS+CTI-WATER-WOODY-YLAT) 7 -486.33 4.84 0.03
ρ(-TIME)Ψ(-GRASS+CTI+WD-YLAT) 7 -486.59 5.35 0.02
ρ(-TIME)Ψ(-GRASS-WATER+WD-WOODY-YLAT) 8 -485.74 5.65 0.02
ρ(-TIME)Ψ(-GRASS+CTI-YLAT) 6 -487.99 6.16 0.02
ρ(.)Ψ(-GRASS+CTI+WD-WOODY-YLAT) 7 -487.05 6.27 0.02
Marbled
Godwit ρ(TIME-WIND)Ψ(-GRASS+CTI-MSAVI+WD-YLAT) 9 -777.18 0.00 0.44
ρ(TIME)Ψ(-GRASS+CTI-MSAVI+WD-YLAT) 8 -779.12 1.89 0.17
ρ(WIND)Ψ(-GRASS+CTI-MSAVI+WD-YLAT) 8 -779.38 2.40 0.13
ρ(.)Ψ(-GRASS+CTI-MSAVI+WD-YLAT) 7 -780.53 2.71 0.11
ρ(TIME-WIND)Ψ(-GRASS+CTI+WD-YLAT) 8 -779.87 3.39 0.08
ρ(TIME)Ψ(-GRASS+CTI+WD-YLAT) 7 -781.74 5.13 0.03
ρ(-WIND)Ψ(-GRASS+CTI+WD-YLAT) 7 -782.11 5.87 0.02
McCown's
Longspur ρ(WIND)Ψ(GRASS+XLONG-YLAT) 5 -67.92 0.00 0.46
ρ(WIND)Ψ(GRASS-WATER+XLONG-YLAT) 7 -67.47 1.11 0.26
ρ(.)Ψ(GRASS+XLONG-YLAT) 5 -69.84 1.84 0.18
ρ(.)Ψ(GRASS-WATER+XLONG-YLAT) 6 -69.41 2.99 0.10
Savannah
Sparrow ρ(-TIME-WIND)Ψ(GRASS-HLI-XLONG) 7 -1105.64 0.00 0.52
ρ(-TIME)Ψ(GRASS-HLI-XLONG) 6 -1107.33 1.38 0.26
ρ(-TIME-WIND)Ψ(GRASS-HLI) 6 -1107.48 1.68 0.22
Sprague's Pipit
ρ(JDATE+TIME-WIND)Ψ(GRASS+CTI+XLONG-YLAT) 9 -862.64 0.00 0.38
ρ(JDATE+TIME)Ψ(GRASS+CTI+XLONG-YLAT) 8 -864.02 0.75 0.26
ρ(JDATE+TIME-WIND)Ψ(GRASS+CTI-YLAT) 8 -864.61 1.94 0.14
ρ(JDATE+TIME)Ψ(GRASS+CTI-YLAT) 7 -865.97 2.65 0.10
ρ(JDATE+TIME-WIND)Ψ(GRASS-YLAT) 7 -866.55 3.81 0.06
ρ(JDATE+TIME-WIND)Ψ(GRASS+XLONG-YLAT) 8 -865.61 3.94 0.05
Upland
Sandpiper ρ(TIME)Ψ(GRASS+CTI) 5 -315.86 0.00 1.00
Vesper
Sparrow ρ(-TIME-WIND)Ψ(-GRASS-CTI-WATER+YLAT) 8 -1183.51 0.00 0.56
ρ(-TIME)Ψ(-GRASS-CTI-WATER+YLAT) 7 -1184.75 0.47 0.44
Western
Meadowlark ρ(-JDATE-TIME)Ψ(GRASS+CTI-YLAT) 7 -652.94 0.00 1.00
Willet ρ(-JDATE)Ψ(GRASS+CTI-HLI+WATER-WD) 8 -651.93 0.00 0.21
ρ(-JDATE)Ψ(GRASS+CTI-HLI-WD) 7 -653.20 0.55 0.16
ρ(-JDATE)Ψ(GRASS-HLI+WATER-WD) 7 -653.27 0.68 0.15
ρ(-JDATE)Ψ(GRASS+CTI-HLI+WATER) 7 -653.75 1.64 0.09
ρ(-JDATE)Ψ(GRASS+CTI-WD) 6 -654.80 1.75 0.09
ρ(-JDATE)Ψ(GRASS+CTI+WATER-WD) 7 -653.80 1.75 0.06
ρ(-JDATE)Ψ(GRASS-HLI+WATER) 6 -655.14 2.43 0.06
ρ(-JDATE)Ψ(GRASS-HLI-WD) 6 -655.18 2.52 0.06
ρ(-JDATE)Ψ(GRASS+CTI-HLI) 6 -655.23 2.61 0.06
ρ(-JDATE)Ψ(GRASS+CTI+WATER) 6 -655.48 3.10 0.04
93
Baird’s Sparrow Chestnut-collared Longspur
Sprague’s Pipit Western Meadowlark
Figure 2.3. Species relationships between probability of detection (ρ) and date within the
Special Areas of Alberta, 2012. Each figure is based on the same standardised scale of
mean = 0 and variance = 1, where negative values represents survey conducted earlier in
the survey season and positive values represent surveys conducted later in the survey
season. The 95% confidence intervals are indicated by the shaded grey area.
94
Table 2.5. Coefficient estimates (β) and unconditional standard errors (SE) for intercepts and variables used to predict species
occupancy in the Special Areas, 2012. Variables used to explain the detection probability (ρ) and occurrence probability (Ψ)
are presented separately and zeros indicate that a covariate was considered but not included in the final model. Species
abbreviations are BAIS (Baird’s Sparrow), BHCO (Brown-headed Cowbird), CCLO (Chestnut-collared Longspur), CCSP
(Clay-colored Sparrow), HOLA (Horned Lark), LBCU (Long billed Curlew), MAGO (Marbled Godwit), MCLO (McCown’s
Longspur), SAVS (Savannah Sparrow), SPPI (Sprague’s Pipit), UPSA (Upland Sandpiper), VESP (Vesper Sparrow), WEME
(Western Meadowlark), and WILL (Willet).
DETECTION
PROBABILITY (ρ) BAIS BHCO CCLO CCSP HOLA LBCU MAGO MCLO SAVS SPPI UPSA VESP WEME WILL
INTERCEPT (β) 0.429(0.09) -0.063(0.07) 0.349(0.12) 1.044(0.07) 0.367(0.07) -1.163(0.81) -1.058(0.11) -1.452(0.52) 1.157(0.06) 0.723(0.08) -1.392(0.25) 0.831(0.06) 2.350(0.09) -1.389(0.15)
JULIAN DATE 0.440(0.09) -0.300(0.06) 0.821(0.12) 0 0.109(0.07) 0 0 0 0 0.354(0.08) 0 0 -0.420(0.08) -0.283(0.09)
TIME -0.310(0.08) 0 0 -0.198(0.07) 0.310(0.06) -0.218(0.09) 0.139(0.07) 0 -0.208(0.06) 0.292(0.08) 0.304(0.14) -0.331(0.06) -0.350(0.08) 0
WIND 0 0 0 -0.157(0.06) 0 0 -0.137(0.08) 0.443(0.23) -0.113(0.06) -0.128(0.08) 0 -0.086(0.05) 0
OCCURRENCE
PROBABILITY (Ψ) BAIS BHCO CCLO CCSP HOLA LBCU MAGO MCLO SAVS SPPI UPSA VESP WEME WILL
INTERCEPT (β) -0.717(0.13) 1.56(0.20) -2.633(0.27) 1.133(0.12) 1.088(0.14) -1.018(0.24) 0.587(0.29) -8.310(2.19) 1.694(0.13) -0.196(0.11) -1.499(0.26) 2.644(0.34) 4.397(0.44) 0.194(0.27)
GRASSLAND 1.424(0.16) -0.428(0.21) 1.643(0.28) 0.388(0.11) -1.124(0.17) -0.679(0.20) -0.047(0.16) 0.449(0.61) 0.839(0.11) 1.507(0.15) 0.201(0.17 -1.367(0.42) 0.591(0.21) 0.122(0.16)
SHRUBLAND 0 0 0 0 0 -0.731(0.43) 0 0 0 0 0 0 0 0
VEGETATIVE GREENNESS -0.247(0.14) 0.379(0.17) -0.531(0.23) 0.275(0.12) -0.371(0.14) 0 -0.418(0.19) 0 0 0 0 0 0 0
SURFACE WATER 0 0 0 0 0 -0.344(0.17) 0 -0.613(0.72) 0 0 0 -0.353(0.12) 0 0.298(0.19)
SURFACE WATER2
0 0 0 0 0 0 0 0 0 0 0 0 0 0
WETNESS 0.294(0.18) 0 0 -0.502(0.11) 0 0.4756(0.17) 0.937(0.19) 0 0 0.239(0.11) 0.766(0.17) -0.494(0.14) 0.424(0.25) 0.522(0.28)
HEAT LOAD -0.393(0.18) -0.261(0.11) -0.846(0.21) 0 -475(0.11) 0 0 0 -0.454(0.10) 0 0 0 0 -0.528(0.27)
LONGITUDE 0 0 0 0 0.507(0.13) 0 0 2.112(1.04) -0.220(0.12) 0.221(0.12) 0 0 0 0
LATITUDE -0.880(0.14) 0 -1.809(0.26) 1.080(0.15) -0.677(0.14) -1.333(0.26) -0.622(0.22) -4.187(1.55) 0 -0.457(0.11) 0 0.779(0.17) -1.444(0.33) 0
WELL DENSITY 0 0 0 0 0 0.698(0.39) 1.126(0.46) 0 0 0 0 0 0 -0.396(0.24)
95
Chestnut-collared Longspur Clay-colored Sparrow
Horned Lark Marbled Godwit
Figure 2.4. Species relationships between probability of occurrence (Ψ) and modified
Soil-adjusted Vegetation Index within a 400 m radius of the point count centre for
surveys conducted within the Special Areas of Alberta, 2012. Each figure is based on the
same standardised scale of mean = 0 and variance = 1, where negative values indicate
more bare ground and positive values indicate more green vegetative biomass. The 95%
confidence intervals are indicated by the shaded grey area.
96
Clay-colored Sparrow Long-billed Curlew Marbled Godwit
Sprague’s Pipit Upland Sandpiper
Figure 2.5. Species relationships between probability of occurrence (Ψ) and Compound Topographic Index within a 400 m radius of the
point count centre for surveys conducted within the Special Areas of Alberta, 2012. Each figure is based on the same standardised scale of
mean = 0 and variance = 1, where negative values indicate more mesic areas and positive values indicate more xeric areas on the
landscape. The 95% confidence intervals are indicated by the shaded grey area.
97
Baird’s Sparrow Chestnut-collared Longspur
Horned Lark Savannah Sparrow
Figure 2.6. Species relationships between probability of occurrence (Ψ) and Heat Load Index
within a 400 m radius of the point count centre for surveys conducted within the Special Areas of
Alberta, 2012. Each figure is based on the same standardised scale of mean = 0 and variance =
1, where negative values represent less solar radiation and positive values represent higher rates
of solar radiation on the landscape. The 95% confidence intervals are indicated by the shaded
grey area.
98
Table 2.6. Comparison of area under the curve (AUC) scores from the 2012 partitioned testing
data (n = 218). Testing data which were not adjusted to account for imperfect detection are
indicated by Unadjusted, and training data which accounts for imperfect detection are indicated
by Adjusted. Upper (Cl+) and lower (Cl-) confidence limits are indicated.
SPECIES
TESTING
DATA
(2012)
AUC CI (+) CI (-)
Baird's Sparrow Unadjusted 0.84 0.90 0.79
Adjusted 0.85 0.90 0.80
Brown-headed Cowbird Unadjusted 0.61 0.70 0.52
Adjusted 0.61 0.70 0.53
Chestnut-collared Longspur Unadjusted 0.88 0.93 0.83
Adjusted 0.90 0.95 0.86
Clay-colored Sparrow Unadjusted 0.68 0.76 0.60
Adjusted 0.68 0.76 0.61
Horned Lark Unadjusted 0.75 0.81 0.68
Adjusted 0.75 0.81 0.68
Long-billed Curlew Unadjusted 0.66 0.75 0.58
Adjusted 0.67 0.75 0.58
Marbled Godwit Unadjusted 0.68 0.75 0.60
Adjusted 0.67 0.74 0.60
McCown's Longspur Unadjusted 0.88 0.96 0.80
Adjusted 0.88 0.97 0.80
Savannah Sparrow Unadjusted 0.74 0.82 0.66
Adjusted 0.74 0.82 0.66
Sprague's Pipit Unadjusted 0.78 0.84 0.72
Adjusted 0.79 0.85 0.73
Upland Sandpiper Unadjusted 0.58 0.70 0.47
Adjusted 0.58 0.70 0.46
Vesper Sparrow Unadjusted 0.78 0.86 0.71
Adjusted 0.78 0.86 0.71
Western Meadowlark Unadjusted 0.98 1.00 0.96
Adjusted 0.98 1.00 0.96
Willet Unadjusted 0.66 0.75 0.58
Adjusted 0.69 0.77 0.61
99
Table 2.7. Predictive performance of occupancy models used to estimate occurrence
probabilities of 14 grassland bird species within the Special Areas, Alberta. Predictive
performance was assessed using two sources of evaluation data, the partitioned 2012 data
(n=218) and the independent sample collected in 2011 (n=1498). Performance was measured
based on area under the curve (AUC) scores and Kappa qualitative scores. Frequency of
occurrence, or observed prevalence of species are indicated for each data set along with 95%
confidence limits (CI) for both measures.
SPECIES TESTING
DATA
OBSERVED
PREVALENCE
AUC
value
95%
(Cl)
AUC
Ranking
Kappa
value
95%
(Cl)
Kappa
Ranking
Baird's Sparrow 2012 0.41 0.85 0.05 Good 0.56 0.11 Good
2011 0.38 0.76 0.03 Useful 0.34 0.04 Fair
Brown-headed
Cowbird
2012 0.72 0.61 0.09 Poor 0.23 0.14 Fair
2011 0.55 0.60 0.03 Poor 0.16 0.05 Poor
Chestnut-collared
Longspur
2012 0.20 0.90 0.05 Excellent 0.55 0.12 Good
2011 0.17 0.84 0.03 Good 0.45 0.05 Fair
Clay-colored Sparrow 2012 0.71 0.68 0.08 Poor 0.26 0.14 Fair
2011 0.56 0.71 0.03 Useful 0.31 0.05 Fair
Horned Lark 2012 0.58 0.75 0.06 Useful 0.39 0.12 Fair
2011 0.53 0.68 0.03 Poor 0.25 0.05 Fair
Long-billed Curlew 2012 0.18 0.67 0.08 Poor 0.15 0.08 Poor
2011 0.12 0.64 0.04 Poor 0.09 0.03 Poor
Marbled Godwit 2012 0.31 0.67 0.07 Poor 0.22 0.10 Fair
2011 0.19 0.57 0.04 Poor 0.07 0.03 Poor
McCown's Longspur 2012 0.03 0.88 0.08 Good 0.17 0.15 Poor
2011 0.01 0.88 0.07 Good 0.08 0.04 Poor
Savannah Sparrow 2012 0.81 0.74 0.08 Useful 0.27 0.12 Fair
2011 0.64 0.65 0.03 Poor 0.23 0.05 Fair
Sprague's Pipit 2012 0.49 0.79 0.06 Useful 0.46 0.12 Good
2011 0.37 0.77 0.02 Useful 0.38 0.05 Fair
Upland Sandpiper 2012 0.12 0.58 0.12 Poor 0.09 0.08 Poor
2011 0.07 0.59 0.06 Poor 0.02 0.01 Poor
100
Vesper Sparrow 2012 0.82 0.78 0.07 Useful 0.42 0.13 Good
2011 0.79 0.64 0.04 Poor 0.19 0.06 Poor
Western Meadowlark 2012 0.96 0.98 0.02 Excellent 0.47 0.21 Good
2011 0.90 0.77 0.05 Useful 0.27 0.06 Fair
Willet 2012 0.25 0.69 0.08 Poor 0.30 0.14 Fair
2011 0.19 0.56 0.04 Poor 0.06 0.03 Poor
101
3. CONCLUSIONS AND MANAGEMENT IMPLICATIONS
3.1 Summary
The objectives of my research were to: (1) determine factors that influence the
distribution of grassland birds on the landscape, (2) use occupancy modelling to generate Species
Distribution Models (SDMs) and predict the probability of occurrence for grassland birds within
the Special Areas of southeastern Alberta, and (3) evaluate the predictive performance of the
occupancy models.
In chapter 2, I show that using count data, in conjunction with occupancy models and the
geographical and environmental variables across the landscape, is useful for understanding how
species interact with the composition of cover types. My results show that occurrence data are
influenced by detectability, and a repeated survey design is useful in that it allows for statistically
adjusting occurrence data for imperfect detection. I found identifying species occurrence
patterns to the amount of grassland cover type was a useful measure for explaining occurrence,
and corresponded strongly with previous studies (Owen and Myers 1973; Johnson et al. 2004;
Davis et al. 2013). The additional landscape features I used to describe variation in species
occurrence further helped to explain resource selection. Both topography and the geographic
location on the landscape were important for explaining where species occur. Bird species
exhibited mixed responses to variation in vegetative biomass across the landscape, as explained
by the modified soil-adjusted vegetation index. The surface water, woody cover, and well
density variables also influenced occurrence, but again I found the results varied among species.
Clearly, further research is needed to fully understand the complex features that influence where
species occur on the landscape, but these models provide an objective, quantitative method for
evaluating landscapes for conservation.
102
Predictive distribution models provide a graphical means to interpret patterns of resource
selection across an area of interest. The spatially explicit maps produced by my analysis are
considered an index of a species probability of occurrence (MacKenzie et al. 2006) based on the
coarse-grained landscape variables identified in each model. Both the maps and models can be
used in a wide variety of applications, including assessing the impacts of land cover change,
guiding and refining future research efforts, supporting conservation prioritization and reserve
selection, and guiding species reintroduction (Elith and Leathwick 2009; Franklin 2010).
Regardless of the application, evaluation is a necessary step to provide confidence in the
predictive ability of a distribution model (Fielding and Bell 1997; Elith and Leathwick 2009).
By comparing model predictions with two evaluation datasets, I demonstrated that my models
hold promise in making justifiable inferences about species distribution. I showed that
distribution models for four species provide evidence for use as a conservation tool. The
remaining species models, which had poor predictive performance, require further investigation
to determine whether, or the extent to which they can be improved. Overall, my models
provided quantitative insights into the magnitude and importance of species-habitat relationships
at the landscape level, and a level of confidence in how useful the distribution models are to
predict species occurrence.
3.2 Management Implications
Given that conservation resources will always be limited, strategic approaches to
developing management actions for ecological communities at the landscape level are necessary
(Berlanga et al. 2010). The models and Geographic Information System products (e.g., maps
103
and shape files) that I generated will allow conservation practitioners to pinpoint areas on the
landscape that have the highest likelihood for successful conservation management actions.
Additionally, occupancy models are applicable in multi-species management actions, as the
single-species, single-season models I produced can be refined into multi-species occupancy
models (Royle and Dorazio 2008). For example, if Baird’s Sparrow, Chestnut-collared
Longspur, and Sprague’s Pipit are collectively targeted as species of conservation concern, a
multi-species model can be used to address explicit questions about what, where, and how to
manage an area for these grassland bird species
Distribution models provide a foundation for policymakers, conservation planners and
land managers to make informed decisions about conservation strategies. However, like any
predictive modelling approach, the models I present here are imperfect for a variety of reasons.
Given the range of detection and occurrence probabilities I observed, employing the standard
survey design with three repeated visits proved to be flexible for the primary species
investigated. However, the low number of McCown’s Longspur detections suggests that more
sampling sites are needed to improve occupancy estimates. The precision of my occupancy
estimates for Long-billed Curlew, Willet, and Marbled Godwit were also low, and increasing the
number of repeated visits to each site would improve precision. For example, with no
consideration of cost, MacKenzie and Royle (2005) suggest that increasing the amount of
repeated surveys to seven would reach optimal precision. Predictive models should incorporate
annual variation in occurrence and demographic patterns (MacKenzie et al. 2006) to increase
their rigour for conservation management and planning efforts. Incorporating interactions
between variables within a multi-scale analysis may explain the data better than a single spatial
104
scale I used (i.e., 400 m radius), but there is no guarantee this will provide more insight into
explaining occupancy patterns. Additionally, my models are limited by the availability of
accurate spatial data within the study area. More accurate environmental variables should result
in better resource selection models. Clearly improvements can be made to each species model,
yet in the absence of more accurate spatial data, the predictive distribution models I generated
provide a strong basis for making informed management decisions. These models and maps
should be used to guide regional conservation planning for birds, and can be used as a valuable
reference to guide and refine future research efforts in other regions.
3.3 Literature Cited
BERLANGA, H. K., J. A. RICH, T. D. ARIZMENDI, M. C. BEARDMORE, C. J. BLANCHER, P. J.
BUTCHER, G. S. COUTURIER, A. R. DAYER, A. A. DEMAREST, D. W. EASTON, W. E.
GUSTAFSON, M. INIGO-ELIAS, E. KREBS, E. A. PANJABI, A. O. RODRIGUEZ
CONTRERAS, V. ROSENBERG, K. V. RUTH, J. M. SANTANA CASTELLÓN, E. VIDAL, R.
MA, AND T. WILL. 2010. Saving Our Shared Birds: Partners in Flight Tri-National
Vision for Landbird Conservation. Cornell Lab of Ornithology: Ithaca, NY. Available at:
www.savingoursharedbirds.org. Accessed online October 5, 2012.
DAVIS, S. K., R. J. FISHER, S. L. SKINNER, T. L. SHAFFER, AND R. M. BRIGHAM. 2013.
Songbird abundance in native and planted grassland varies with type and amount of
grassland in the surrounding landscape. Journal of Wildlife Management 77(5): 908-919.
105
ELITH, J., AND J. R. LEATHWICK. 2009. Species distribution models: ecological explanation and
prediction across space and time. Annual Review of Ecology, Evolution, and Systematics
40: 677-697.
FIELDING, A. H., AND J. F. BELL. 1997. A review of methods for the assessment of prediction
errors in conservation presence/absence models. Environmental Conservation 24: 38–49.
FRANKLIN, J. 2010. Moving beyond static species distribution models in support of conservation
biogeography. School of Geographical Sciences and Urban Planning and School of Life
Sciences, Arizona State University, Tempe AZ USA.
JOHNSON, D. H., L. D. IGL, AND J. A. DECHANT (Series Coordinators). 2004. Effects of
management practices on grassland birds. Northern Prairie Wildlife Research Center,
Jamestown, ND. Jamestown, ND: Northern Prairie Wildlife Research Center Online.
http://www.npwrc.usgs.gov/resource/literatr/grasbird/index.htm (Version 12AUG2004).
Accessed online December 12, 2012.
MACKENZIE, D. I., AND J. A. ROYLE. 2005. Designing efficient occupancy studies: General
advice and tips on allocation of survey effort. Journal of Applied Ecology 42: 1105-1114.
106
MACKENZIE, D. I., J. D. NICHOLS, J. A. ROYLE, K. P. POLLOCK, L. L. BAILEY, AND J. E.
HINES. 2006. Occupancy estimation and modeling: inferring patterns and dynamics of
species occurrence. Academic Press, San Diego, California, USA.
OWENS, R. A., AND M. T. MYRES. 1973. Effects of agriculture upon populations of native
passerine birds of an Alberta fescue grassland. Canadian Journal of Zoology 51:697-713.
ROYLE, J. A., AND R. M. DORAZIO. 2008. Hierarchical modeling and inference in ecology: the
analysis of data from populations, metapopulations and communities. Academic Press,
New York, New York, USA.
107
APPENDICIES
APPENDIX A – Relative variable importance for grassland bird occupancy models generated in
the Special Areas, Alberta.
108
Appendix A. Table A.1. Variable shrinkage estimates when model averaging was incorporated to assist with model selection
uncertainty in the Special Areas, 2012. The relative variable importance values are on a scale from 0 to 1, giving an equal
weight in the global model used to predict species occupancy. Positive and negative species-habitat relationships are indicated
(+/-). Variables used to explain the detection component, ρ, are (date [JDATE], time [TIME], wind [WIND]) and variables
used to explain the occurrence component, Ψ, are (grassland [GRASS], woody cover [WOODY], vegetative greenness
[MSAVI], surface water [WATER], solar radiation [HLI], wetness [CTI], latitude [YLAT], longitude [XLONG], and well
density [WD]).
SPECIES ρ(JDATE) ρ(TIME) ρ(WIND)
Ψ(GRASS) Ψ(WOODY) Ψ(WATER) Ψ(MSAVI) Ψ(YLAT) Ψ(XLONG) Ψ(HLI) Ψ(CTI) Ψ(WD)
Baird's Sparrow 1.00 (-)1.00 1.00 (-)0.70 (-)1.00 (-)0.84 0.60
Brown-headed Cowbird (-)1.00 (-)1.00 1.00 (-)0.81
Chestnut-collared Longspur* 1.00 1.00 (-)1.00 (-)1.00 (-)1.00
Clay-colored Sparrow (-)1.00 (-)0.90 1.00 0.87 1.00 (-)1.00
Horned Lark 0.58 1.00 (-)1.00 (-)1.00 (-)1.00 1.00 (-)1.00
Long-billed Curlew (-)0.90 (-)1.00 (-)0.80 (-)0.81 (-)1.00 0.98 0.65
Marbled Godwit 0.73 (-)0.68 (-)1.00 (-)0.86 (-)1.00 1.00
McCown's Longspur 0.72 1.00 0.36 (-)1.00 1.00
Savannah Sparrow (-)1.00 (-)0.74 1.00 (-)0.78 (-)1.00
Sprague's Pipit 1.00 1.00 (-)0.64 1.00 (-)1.00 0.70 0.89
Upland Sandpiper* 1.00 1.00 1.00
Vesper Sparrow (-)0.56 (-)1.00 (-)1.00 1.00 (-)1.00
Western Meadowlark* (-)1.00 (-)1.00 1.00 (-)1.00 1.00
Willet (-)1.00 1.00 0.64 (-)0.78 0.73 (-)0.75
* indicates species with single top model
109
APPENDIX B – Probability of occurrence maps for bird species in the Special Areas,
Alberta.
Appendix B. Figure B.1. Baird's Sparrow (Ammodramus bairdii) expected probability of
occurrence for the Special Areas, Alberta.
110
Appendix B. Figure B.2. Brown-headed Cowbird (Molothrus ater) expected probability
of occurrence for the Special Areas, Alberta.
111
Appendix B. Figure B.3. Chestnut-collard Longspur (Calcarius ornatus) expected
probability of occurrence for the Special Areas, Alberta.
112
Appendix B. Figure B.4. Clay-colored Sparrow (Spizella pallida) expected probability of
occurrence for the Special Areas, Alberta.
113
Appendix B. Figure B.5. Horned Lark (Eremophila alpestris) expected probability of
occurrence for the Special Areas, Alberta.
114
Appendix B. Figure B.6. Long-billed Curlew (Numenius americanus) expected
probability of occurrence for the Special Areas, Alberta.
115
Appendix B. Figure B.7. Marbled Godwit (Limosa fedoa) expected probability of
occurrence for the Special Areas, Alberta.
116
Appendix B. Figure B.8. McCown's Longspur (Rhynchophanes mccownii) expected
probability of occurrence for the Special Areas, Alberta.
117
Appendix B. Figure B.9. Savannah Sparrow (Passerculus sandwichensis) expected
probability of occurrence for the Special Areas, Alberta.
118
Appendix B. Figure B.10. Sprague's Pipit (Anthus spragueii) expected probability of
occurrence for the Special Areas, Alberta.
119
Appendix B. Figure B.11. Upland Sandpiper (Bartramia longicauda) expected
probability of occurrence for the Special Areas, Alberta.
120
Appendix B. Figure B.12. Vesper Sparrow (Pooecetes gramineus) expected probability
of occurrence for the Special Areas, Alberta.
121
Appendix B. Figure B.13. Western Meadowlark (Sturnella neglecta) expected
probability of occurrence for the Special Areas, Alberta.
122
Appendix B. Figure B.14. Willet (Tringa semipalmata) expected probability of
occurrence for the Special Areas, Alberta.
*******************************