University of Florida Thesis Or Dissertation Formatting

RESPONSE OF MID- AND LARGE-SIZED MAMMALS TO WOODY ENCROACHMENT IN A SOUTHERN AFRICAN SAVANNA

JOSE ROBERTO SOTO

A DISSERTATION PRESENTED TO THE GRADUATE SCHOOL OF THE UNIVERSITY OF FLORIDA IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF DOCTOR OF PHILOSOPHY

UNIVERSITY OF FLORIDA

2016

To God, who makes it all possible To the memory of my parents, Maria Consuelo Soto and Neftali Soto. To Nancy and Danny. You are my past, present and future.

ACKNOWLEDGMENTS

There are many people and organizations I need to thank for their support and without whom I would never have been able to accomplish this long and many times trying endeavor.

Special thanks to my major advisor Dr. Robert McCleery, first for introducing me to the fascinating and amazing world of savanna ecology and allowing me to work with him on this project and secondly for being such a committed and inspiring mentor. His hard work and commitment to student excellence is exemplary. My co-chair Dr. William Giuliano has guided and helped me since I first arrived in Gainesville for my master’s degree and I am indebted to him for all his help and support. I also need to express my sincere gratitude to the members of my committee. Each member shared their own area of expertise and knowledge with me in a patient and effective manner. Dr. Ara Monadjem patiently guided me as I learned about the ecology and wildlife of African savannas. Dr. Rob Ahrens was instrumental in my gaining an understanding and grasp of Bayesian Hierarchical models. Dr. Deborah Miller provided the expertise and knowledge I needed to better understand rangeland management and ecology. Dr.

Daniel Gwinn was an honorary member of my committee who provided immensely helpful and instrumental guidance in data analysis. Without Dr. Gwinn’s help, I would still be writing my first line of R code. Thanks for being such a good teacher. I would also like to thank James Colee and Mauricio Nunez-Regueiro for useful guidance and advice on statistical analyses.

I would also like to thank Dr. Martin Main for providing funding for the first three years of my doctoral program. Funding for my program also came from the Tropical Conservation and

Development program in the form of a Fellowship, and through teaching assistantships provided by the Department of Wildlife Ecology and Conservation and the Biology Department at the

University of Florida. Funding for the research project was provided by the University of

Florida. Staff members of All Out Africa and undergraduate students from the University of

Florida and Swaziland University participated in the data collection and entry process and this study could not have been successful without them. I would especially like to thank Henry

Brown and Alex Alford, undergraduate students at the University of Florida, for their invaluable help in sorting through the thousands of camera trap pictures obtained.

A special thanks to Claire Williams, Caprice McRae, Elaine Culpepper, and Monica

Lindberg. The rapid, kind and efficient manner in which they help and support WEC students through tedious paperwork is unique and makes graduate school so much easier for all of us. I can’t say enough about the manner and efficiency in which they support us. Thank you so much for this.

Last, but by no means least, my eternal, infinite and total gratitude to my muses: Nancy and Danny. Their endless support (logistical, emotional and in many other forms) got me through the tough times and I could have never made it without them. Danny: your simple presence and laughter gave me the fuel necessary to continue and not give up when all seemed lost. Nancy: where to begin? You and I know how difficult this journey has been, and I truly can’t imagine accomplishing what I have accomplished without you. You and I make the perfect team and there is no one else I would rather “play” with then you. Thank you and I love you will never be enough to express how I feel about you.

TABLE OF CONTENTS page

ACKNOWLEDGMENTS ...... 4

LIST OF TABLES ...... 8

LIST OF FIGURES ...... 9

ABSTRACT ...... 10

CHAPTER

1 INTRODUCTION ...... 12

2 SPATIO-TEMPORAL DYNAMICS OF HABITAT ASSOCIATIONS AND RELATIVE ABUNDANCE PATTERNS OF MID- AND LARGE-SIZED MAMMALS IN A WOODY ENCROACHED SOUTHERN AFRICAN SAVANNA ...... 15 Background ...... 15 Methods ...... 18 Study Site ...... 18 Research Design ...... 19 Vegetation Sampling and Plot Comparisons ...... 20 Mammal Surveys with Camera Traps ...... 21 Bayesian Hierarchical Multi-Species Abundance Model ...... 22 Temporal and Spatial Covariates ...... 24 Model Structure ...... 25 Community-Wide Model Selection Procedure ...... 26 Species-Specific Model Selection Procedure ...... 26 Results...... 27 Plot Comparisons Based on Habitat Covariates ...... 27 Species Richness Estimates ...... 28 Community-Level Effects of Covariates on Abundance and Probability of Detection ...... 28 Covariate Effects on Species-Specific Relative Abundance Estimates ...... 29 Covariate Effects on Species-Specific Probability of Detection ...... 29 Species-Specific Relative Abundance Estimates ...... 29 Discussion ...... 30 Species Richness and Community-Level Responses to Habitat Covariates ...... 30 Species-Specific Covariate Effects and Abundance Estimates ...... 31 Management and Conservation Implications ...... 33

3 ARE UNGULATE CO-OCCURRENCE PATTERNS INFLUENCED BY WOODY ENCROACHMENT? ...... 47 Background ...... 47 Methods ...... 50 Study Site ...... 50 Research Design ...... 50

Effects of Shrub Encroachment on Plant Structural Heterogeneity ...... 50 Mammal Surveys with Camera Traps ...... 51 Test of Spatial Autocorrelation ...... 52 Co-occurrence Models ...... 52 Effects of Shrub Encroachment on Spatial Overlap ...... 55 Results...... 56 Effects of Shrub Encroachment on Plant Structural Heterogeneity ...... 56 Effects of Shrub Encroachment on Spatial Overlap ...... 57 Discussion ...... 58

4 COMPARISON OF STATISTICAL MODELING FRAMEWORKS AND SAMPLING EFFORT SCHEMES TO GENERATE UNGULATE ABUNDANCE ESTIMATES FROM CAMERA TRAPPING SURVEYS IN A WOODY ENCROACHED SAVANNA ...... 70 Background ...... 70 Methods ...... 73 Study Site ...... 73 Research Design and Data Collection ...... 74 Royle and Nichols Heterogeneity Abundance Model ...... 75 Comparison of Abundance Estimates Using Different Modelling Frameworks ...... 76 Evaluating Sampling Effort Schemes Via Data Simulation ...... 79 Results...... 81 Comparison of Abundance Estimates Using Different Modelling Frameworks ...... 81 Evaluating Sampling Effort Schemes via Data Simulation ...... 82 Discussion ...... 84

5 CONCLUSIONS ...... 95

APPENDIX

A COVER SCALE USED TO MEASURE HABITAT COVARIATES ...... 97

B MULTI-SPECIES HIERARCHICAL ABUNDANCE MODEL CODE DESCRIPTION ....98

C LIFE-HISTORY TRAITS OF TERRESTRIAL MAMMALS DETECTED IN STUDY ...100

D TOP RANKED SINGLE-SPECIES OCCURRENCE MODELS FOR UNGULATES ...... 101

E CO-OCCURRENCE ESTIMATES FROM ROBUST TWO-SPECIES CONDITIONAL MODEL FOR UNGULATE SPECIES ...... 102

F DIAGNOSTIC PLOTS FOR MULTIPLE LINEAR REGRESSION USED TO MODEL THE INFLUENCE OF COVARIATES ON THE SPECIES INTERACTION FACTOR ..105

LIST OF REFERENCES ...... 106

BIOGRAPHICAL SKETCH ...... 118

LIST OF TABLES Table page

2-1 Information on camera trap surveys...... 35 2-2 Mean values and descriptive statistics of habitat covariates...... 35 2-3 Factor loadings from principle components analysis of habitat covariates...... 36 2-4 Random variables and derived parameters of the multi-species hierarchical Bayesian model...... 36 2-5 Community-level effects of habitat covariates on relative abundance...... 37 2-6 Community-level effects of habitat covariates on detection probability...... 37 2-7 Relative abundance and detection probability estimates and the direction of the effect of statistically significant covariates...... 38 3-1 Results of spatial autocorrelation test based on occurrence modeling residuals for ungulate species...... 62 3-2 List and description of parameters estimated in the robust design two-species conditional occupancy model in program MARK ...... 63 3-3 Estimates of the ungulate species interaction factor...... 64 3-4 Results from multiple regression analysis of the influence of predictor variables on the ungulate species interaction factor...... 65 3-5 Results from hierarchical partitioning analysis testing the influence of predictor variables on the ungulate species interaction factor...... 65 4-1 Naive occupancy estimates of large- and mid-sized mammals from camera-trapping surveys...... 89 4-2 Comparison of abundance estimates and standard errors from two modeling frameworks...... 89 4-3 Comparison of bias, precision, and accuracy of abundance estimates evaluating number of sites surveyed...... 90 4-4 Comparison of bias, precision, and accuracy of abundance estimates evaluating number of sampling occasions surveyed...... 91 A-1 Cover scale used to measure percent cover of different vegetation types ...... 97 C-1 Life-history traits of terrestrial mammals detected in camera trapping surveys ...... 100 D-1 Top ranked single-species occurrence models for ungulates ...... 101 E-1 Probability of occurrence estimated from robust two-species conditional model for ungulate species ...... 102

LIST OF FIGURES Figure page

2-1 Map of study area location...... 39

2-2 Map of research design...... 40

2-3 Spatial setup of sampling plots for camera trap and vegetation surveys...... 40

2-4 Results of principal components analysis of habitat covariates...... 41

2-5 Posterior probability distribution of total species richness of mid- and large-sized mammals...... 41

2-6 Plot-specific species richness versus vegetation cover measures...... 42

2-7 Total relative abundance across species versus vegetation cover measures...... 42

2-8 Comparison of species-specific predictions of mean relative abundance per plots across survey years...... 43

2-9 Comparison of species-specific predictions of mean relative abundance per plots between seasons...... 44

2-10 Predicted relative abundance as a function of habitat covariates...... 45

2-11 Predicted mean relative abundance per plot surveyed and 95% credible intervals...... 46

3-1 Association between shrub cover percentage and plant structural heterogeneity...... 66

3-2 Comparisons of the species interaction factor between ungulates in Swaziland...... 67

3-3 Comparisons of the species interaction factor between ungulates in Swaziland...... 68

3-4 Results of hierarchical partitioning analysis to determine the influence of covariates on the species interaction factor between ungulates...... 69

4-1 Comparison of estimates of average abundance per plots and their 95% confidence intervals for bushbucks, nyalas, and warthogs...... 92

4-2 Comparison of bias, precision and accuracy of abundance estimates for different combinations of detection probabilities and number of sites surveyed...... 93

4-3 Comparison of bias, precision and accuracy of abundance estimates for different combinations of detection probabilities and number of sampling occasions...... 94

F-1 Diagnostic plots for multiple linear regression used to model the influence of covariates on the species interaction factor for ungulates...... 105

Abstract of Dissertation Presented to the Graduate School of the University of Florida in Partial Fulfillment of the Requirements for the Degree of Doctor of Philosophy

RESPONSE OF MID- AND LARGE-SIZED MAMMALS TO WOODY ENCROACHMENT IN A SOUTHERN AFRICAN SAVANNA

Jose R. Soto

May 2016

Chair: Robert A. McCleery Cochair: William M. Giuliano Major: Wildlife Ecology and Conservation

The expansion of woody cover and subsequent loss of grass in tropical savannas is a global concern. Woody cover expansion is changing vegetation structure and composition and impacting animal species associated with open savannas. Although this phenomenon has been widely documented in sub-Saharan African savannas, few studies have examined how large mammals are affected by this global change. I examined species- and community-wide responses of mid- and large-sized mammals to expanding shrub and tree cover and a concomitant loss of grass cover in a southern African savanna in Swaziland. I first modeled the influence of temporal and spatial covariates on species richness and abundances across a gradient of woody cover for two years using camera traps and multi-species hierarchical Bayesian models. I also examined how woody cover expansion may be affecting co-occurrence patterns of ungulates in the study site. Finally, I compared statistical methods and sampling schemes to generate abundance estimates for animals without individual markings (i.e., ungulates) from presence/absence data to provide local managers with tools to monitor hard-to-detect wildlife populations in woody encroached areas.

Temporal covariates showed higher detection probabilities and abundances for most species during the second year of sampling and the rainy season. Habitat covariate results showed woody cover did not appear to be a strong predictor of mammalian species richness or mean community abundances, but percent grass cover was. Of the 23 species detected, the relative abundance of one showed a negative response to woody cover. Relative abundance of eight species showed a positive association to grass cover. My results indicate woody encroached areas are still used by a large number of species and therefore may provide some benefit to them.

However, low abundance estimates suggest some large and medium sized mammals may be vulnerability to woody encroachment. Most importantly, our study suggests that an indispensable condition for maximizing species richness and abundances of vulnerable species is the maintenance of a high number of areas with adequate grass cover.

Co-occurrence models showed higher spatial overlap of ungulates at woody encroached plots. Shrub cover altered spatial overlap more than other factors previously shown to affect overlap (i.e. guild, body mass differential, and season). My study suggests that plant structural heterogeneity and grass cover is reduced by woody cover and this concentrates resources and forces ungulate species to overlap more than expected.

Finally, I showed that presence/absence models (i.e., Royle and Nichols heterogeneity abundance model) used to estimate ungulate abundance in woody encroached savannas with camera trapping surveys can be less biased and more precise when sampling ≥ 50 sites and surveys consist of ≥ 10 sampling occasions. Researchers should also evaluate methods to increase detection probability for hard to detect species (p ≤0.10) to increase the predictive ability of camera trapping surveys in woody encroached savannas.

CHAPTER 1 INTRODUCTION

Grasslands cover over 20% of the world’s surface and provide essential ecosystem services to society as a whole and local communities across the globe (Gibson, 2009).

Collectively they contribute to 30% of the net global primary productivity (Grace et al., 2006), and their grass resources are the world’s most important food crops (Gibson, 2009). They also provide forage for livestock around the world, support the ecotourism industry and maintain diverse biological communities (Ratter et al., 1997; Uys, 2006; Lorenzen et al., 2012).

In grasslands across the world, grass cover is being replaced by shrubby (woody species

<2m) and tree (woody species > 2m) species (Archer, 1995; Van Auken, 2000; Roques et al.,

2001). Grass is being replaced by woody cover through distinct mechanisms such as reduced access to sunlight and competition for soil moisture. The potential drivers of this increase of woody species (woody encroachment hereafter) include: overgrazing, infrequent fire regimes, nitrogen deposition, increased carbon dioxide concentrations and climatic conditions (i.e., drought) (Milton and Dean, 1995; Roques et al., 2001; Wigley et al., 2010).

Woody encroachment in grasslands has consequences for rangeland productivity, biodiversity, community structure, and ecosystem function (Archer et al., 2001; Van Auken

2009). Woody encroachment may reduce animal species diversity and carrying capacity of grasslands through the loss of forage quantity and quality and through habitat fragmentation

(Milton and Dean, 1995; Blaum et al., 2007; Van Auken, 2009).

The impacts of woody encroachment have been documented for several taxonomic groups including arthropods (Steenkamp and Chown, 1996), reptiles (Meik et al., 2002), rodents

(Monadjem, 1997; Blaum et al., 2007b) and ungulates (Smit and Prins, 2015). Sirami and

Monadjem (2012) found increased woody cover influenced changes in bird species occurrence in

lowveld savannas of Swaziland, while Blaum et al. (2007a) also documented negative effects of woody cover on select carnivore species in South Africa. Similarly, Smit and Prins (2015) concluded that woody encroachment changed herbivore assemblages by favoring browsers over meso-grazers. These studies showed increasing woody cover affects wildlife species differently and leads to shifts in local community structure.

The changes in plant structure and composition associated to woody encroachment can have consequences for local mammalian diversity and abundances. Particularly those associated with open savanna (Smit and Prins, 2015). In this study, I examined patterns of habitat associations, abundances, and co-occurrence of mid- and large-sized mammals across a gradient of woody cover in a southern African savanna1 to determine how they are responding to woody encroachment.

In Chapter 2, I used mammal presence/absence data and multi-species hierarchical

Bayesian models to analyze small-scale patterns of habitat use and relative abundances associated to shrub, tree, and grass cover. I determined which habitat features influenced mammal use of plots across a gradient of woody encroachment by using estimates of species richness and relative abundance at each plot. I then obtained an average relative abundance measure across my study site for all species detected to the determine mammalian community composition of my study site and identify species that were potentially vulnerable to woody encroachment.

1 Savannas are defined as highly dynamic systems dominated by grass cover with a mosaic of scattered trees that transition between grass and tree dominated states (Scholes and Archer, 1997). Savannas typically occur along a continuum of woody cover that can range from 0 to 80% (Parr et al., 2014). This system is mostly driven by the interaction between disturbances such as fire and herbivory and climate (i.e., rainfall) and topo-edaphic characteristics of the landscape (Scholes and Archer, 1997).

In Chapter 3, I examined ungulate co-occurrence patterns across the gradient of woody encroachment. Ungulate species are known to partition resources and time, such that they can overlap in space without entering into direct competition (Bell, 1970; Jarman, 1974; Kleynhans et al., 2011; Gandiwa, 2013). Several factors have been proposed to explain these patterns of co- occurrence (see Chapter 3), yet the issue of how ungulate co-occurrence patterns may be influenced by increasing woody cover has not been examined. I used a robust design co- occurrence model to analyze patterns of spatial overlap between pairs of ungulate species and determine if woody cover influenced them.

In Chapter 4, I sought to provide guidelines and evaluate the use of camera traps for estimating ungulate abundance in woody encroached or forest sites. Managers need reliable methods to monitor ungulate populations and their response to management and/or natural habitat changes. In African savannas, traditional methods for estimating wildlife abundance have relied on direct observations and/or signs. These methods then generate abundance estimates from distance sampling methods or use indices of encounter rates as measures of relative abundance. Camera trapping in savannas is relatively recent, presents several advantages over traditional methods (e.g., sampling for 24-hr periods, non-invasiveness, etc.) and has not been used to obtain measures of abundance in savannas. I evaluated the use of the Royle-Nichols heterogeneity model (Royle and Nichols, 2003) applied to camera trapping data to provide researchers and managers with recommendations on sampling effort and study design. I also compared two statistical approaches that apply the Royle-Nichols heterogeneity model. Chapter

5 discusses general conclusions and management recommendations.

CHAPTER 2 SPATIO-TEMPORAL DYNAMICS OF HABITAT ASSOCIATIONS AND RELATIVE ABUNDANCE PATTERNS OF MID- AND LARGE-SIZED MAMMALS IN A WOODY ENCROACHED SOUTHERN AFRICAN SAVANNA

Background

The conversion of grasslands into systems dominated by woody species (i.e., woody and shrub encroachment) is a worldwide issue that is negatively impacting biodiversity and ecosystem function (Van Auken, 2009). Global change (i.e., altered rainfall patterns and increased atmospheric and soil CO2), along with human-induced changes to fire regimes and herbivory have resulted in the replacement of grass cover by shrubby (woody species <2m) and tree (woody species > 2m) species in rangelands across the world (Milton and Dean, 1995;

Roques et al., 2001; Wigley et al., 2010). This increase of woody species (woody encroachment hereafter) in grasslands may reduce animal species diversity and carrying capacity of grasslands through a loss of forage quantity and quality and through habitat fragmentation (Milton and

Dean, 1995; Blaum et al., 2007; Van Auken, 2009).

Woody encroachment is particularly widespread in tropical savannas of southern Africa

(Archer et al., 2001; Roques et al., 2001; Blaum et al., 2007a; Sirami and Monadjem, 2012). This pattern has been prevalent in the low-lying savannas of southern Africa where grass cover has decreased up to 30% and a shrub cover increase of 20% has been documented over a 20-year period in Swaziland (Sirami and Monadjem, 2012).

The drastic change in habitat structure and composition brought about by woody encroachment is likely to alter the occurrence and prevalence of savanna associated mammals in the region. Reduced grass cover due to an increase of unpalatable woody plants should result in the loss of forage to grazers (Kraaij and Ward, 2006; Meik et al., 2002; Lohman et al., 2014), while browser foraging may be limited by the increased prevalence of inaccessible thorny

thickets (Moleele et al., 2002). Reduced availability of prey (ungulates, reptiles, small mammals and invertebrates) in woody encroached savannas may lead to cascading effects on carnivores

(Steenkamp and Chown, 1996; Meik et al., 2002; Blaum et al., 2007a; Blaum et al., 2007b), especially for species such as the serval (Leptailurus serval), that is associated with open grassland habitats (Ramesh and Downs, 2013).

Most mid- and large-sized mammals in Swaziland are found within managed protected areas or reserves (Monadjem, 1998; Roques et al., 2001; Monadjem et al., 2003) where woody encroachment occurs (Bailey et al., 2016). Because mammals play an essential role in maintaining the structure and composition of vegetation in savannas (Dalerum, 2008; Goheen et al., 2010; Holdo et al., 2013), there is an urgent need for a better understanding of how mammal communities are responding to woody encroachment in southern African protected areas to inform conservation and management efforts in the region.

My goal for this Chapter was to understand the response of mid- and large-sized mammals to woody encroachment within protected areas. I had two specific objectives. First, I wanted to examine community-wide mammalian responses to woody encroachment. I compared mammalian use/avoidance patterns of sampling sites across a gradient of woody encroachment in protected areas. I used shrub and tree cover as indicators of woody encroachment. Certain thresholds of shrub and tree cover are used by wildlife as cover and food and thus, are beneficial to some species (Blaum et al., 2007a; Blaum et al., 2007c). Nevertheless, the overall quality of woody encroached plots may be reduced due to thick impenetrable layers of vegetation, less palatable and low nutrient vegetation, increased bare ground and invasive species (Lohman et al.,

2014). I used a species richness estimate (interpreted as number of species present in a plot) obtained from a hierarchical Bayesian multi-species abundance model (Yamaura et al., 2011) to

examine use/avoidance patterns and compare these to habitat covariates. I hypothesized that woody encroached areas with reduced grass cover would not provide adequate foraging resources to mammal communities and predicted that fewer species would make use of woody encroached areas.

I also obtained species-specific estimates of relative abundance across plots (from the same multi-species hierarchical model) and averaged these estimates across species and plots. I then used a model selection procedure (Burton et al., 2012) to determine the influence of habitat

(i.e., grass, shrub and tree cover) and temporal covariates on the community-wide averaged estimate of abundance per plot. I hypothesized that the community-wide abundance estimate would be negatively influenced by shrub and tree cover, but positively influenced by grass cover.

Second, I wanted to identify which mid- and large-sized mammalian species were potentially more vulnerable to woody encroachment. With low predation pressure due to reduced diversity and abundances of large predators in the study area (Monadjem, 1998), wildlife species’ population dynamics should be driven mostly by resource quality and availability. Thus, changes in species population parameters should reflect habitat conditions.

Vulnerability was assessed in two ways:

1) By examining the association between a species’ relative abundance and shrub, tree and grass cover. A negative influence of shrub and tree cover on a species’ relative abundance would suggest that the species avoids woody encroached areas, while a positive association to grass cover would imply that the species depends on this resource that is being lost by encroachment (Roques et al., 2001; Sirami and Monadjem, 2012; Bailey et al., 2016). Thus, a negative association of a species relative abundance to shrub and tree cover, and a positive association to grass cover would imply that a species is vulnerable to woody encroachment.

2) By qualitatively comparing the relative abundance measure obtained from the multi- species hierarchical model across species. The historical and continued increase of shrub and woody cover and concomitant decrease of grass cover in the study area (Roques et al., 2001;

Sirami and Monadjem, 2012; Bailey et al., 2016) is most likely leading to population reductions of species whose resources are negatively impacted by woody encroachment. Thus, current low abundance estimates across the study site for a species can be considered a negative response to this phenomenon, making this species vulnerable if this trend continues.

I hypothesized that grazers, grassland associated species, and smaller ungulate species that are more dependent on quality forage (McNaughton and Georgadis, 1986; Owen-Smith,

2002) would be more vulnerable to woody encroachment due to a decrease in habitat quality and availability at woody encroached plots. Thus, I predicted relative abundance of grazers, grassland associated species and small ungulates would be negatively associated with tree and shrub cover and positively associated with grass cover. I also predicted that the relative abundance estimates for these species would be lower than those of species with broader habitat and dietary requirements. In the case of ungulates, I predicted that browsers and mixed-feeders would be more abundant than grazers based on lower availability of the grass resource (Gordon and Prins, 2008).

Methods

Study Site

I conducted this study in the Mlawula Nature Reserve and the Mbuluzi Game Reserve, which cover 17,400 and 2,400 ha, respectively, and are located in the low lying savanna of north eastern Swaziland (Figure 2-1). The vegetation community on these two reserves is characterized as basalt sweet arid lowveld (Sirami and Monadjem, 2012) with patches of riparian forest along rivers (Sweet and Khumalo, 1994). The dominant large trees were Acacia nigrescens and

Sclerocarya birrea and the grass layer was dominated by Themeda triandra and Panicum maximum (Gertenbach and Potgieter 1975; Roques et al., 2001). Increases in shrub and tree cover, along with a decrease in grass cover, has been documented for the study area (Roques et al., 2001; Sirami and Monadjem, 2012; Bailey et al., 2016). The dominant species in the shrub layer is Dichrostachys cinerea (Roques et al., 2001). The average monthly temperature in the low-lying savannas of Swaziland is 18 ºC in July and 26 ºC in January (Goudie and Price

Williams, 1983). Rain falls predominantly during the summer months (October-April) and fluctuates about an annual mean of 675 mm (Swaziland Environmental Authority, 2001).

The protected areas of my study were surrounded by a matrix of sugar cane fields, commercial cattle ranches, and populated homesteads (Bailey et al., 2016). These reserves were fenced, with very few large predators (with the exception of some spotted hyena (Crocuta crocuta) and an occasional leopard (Panthera pardus)). As a result, wildlife immigration and emigration and most potential top-down effects on the mammalian species of the area were very limited. Roads and traffic were limited within these reserves and their impact on mammal habitat selection and use was most likely minimal.

Research Design

To evaluate the influence of woody encroachment on mammals’ use of space and relative abundances, we established research units along a gradient of grass, shrub (woody vegetation <

2m in height), and tree (woody vegetation > 2 m in height) cover across the Mlawula Nature and

Mbuluzi Game Reserves in Swaziland (Figures 2-1 and 2-2). We systematically established nine research units of 30.3 ha throughout these reserves (Figure 2-2). Each unit was further subdivided into nine sampling plots of 50 m2 with a distance of 250 m between plots. This resulted in a total of eighty-one plots. We conducted camera trap surveys and vegetation sampling at each

plot during the summer (December to February) and winter (June to August) seasons for two years (2012 and 2013).

Vegetation Sampling and Plot Comparisons

We sampled the composition and structure of vegetation once a year at thirteen points within each of the sampling plots per research unit (Figure 2-3). The values of each plot were averaged to obtain one measure of each variable per plot to be used as covariates in the multi- species abundance models. We measured grass, shrub, and bare ground cover percentage by placing a 1 m2 circular plot at each point. We expressed percentage per point as a cover class scale (Appendix A, Table A-1) (Daubenmire, 1959). We measured tree cover using the line intercept method (Canfield, 1941) and expressed this measure as the percentage of total length covered by tree cover.

We measured visual obstruction (vo), a measure of vegetation height and vertical density, using a robel pole (Vermeire et al., 2002). We placed the robel pole in the center of the sampling point and measured vo from a distance of 4m at a height of 1 m for each of the 4 cardinal directions. We obtained a measure of grass biomass per plot by using a disc pasture meter at each point and averaging across all points in a plot (Bransby and Tainton, 1977).

Additionally, I measured distance of the center of each plot to surface water using a vector layer of water obtained from 2003 Land Sat images in ARCGIS 10.2. I joined the vector layer with all the points of the center of each plot with the surface water layer. With this function, the center point of each plot was joined to the nearest body of water in the layer being analyzed and I obtained an output of the distance between the points being joined. It is important to note that some of these water sources were ephemeral and may only influence mammal use of research units and plots during the dry season when surface water is scarce. To account for this

potential seasonal effect of water availability at or near the sampling plots, I included an interaction term between season and distance to water as another covariate in the abundance model.

I ran a Pearson’s correlation and a Principal Components Analysis (PCA) following

McGarigal et al. (2000) on the habitat covariates measured to characterize the surveyed plots and test for significant correlation between covariates. The PCA allowed me to reduce correlated variables and retain variables that contribute more to the variation within my dataset. I ran the

PCA in program R (R Development Core Team, 2011). Two pairs of habitat covariates were highly correlated: grass cover percentage with grass biomass (r2 = 0.64; P < 0.001) and bare ground and shrub cover percentage (r2 = 0.4; P = 0.003). As a result, I removed grass biomass and bare ground cover percentage from subsequent analyses.

Mammal Surveys with Camera Traps

We conducted camera trap surveys during the summer (December to February) and winter (June to August) seasons for two years (2012 and 2013). Not all research units were sampled each year, but each unit was sampled at least once (Table 2-1). We used two camera traps ([Primos Truth Cam 35], Primos Hunting, St. Flora, Mississippi) per plot during all surveys with the exception of one survey period (November to December 2012) in which we only deployed one camera trap per survey due to logistical constraints. I accounted for the difference in the number of cameras deployed in the statistical models (see below). We tied the motion detection cameras to a tree 40- 60 cm above ground with a clear view of at least 10 m in front of the camera. We cleared the area in front of the camera of grass and shrubs and any obstructing feature. During each sampling period, we deployed the camera traps for a period of 5 days/nights in each plot; resulting in a total of five, 24-hour sampling occasions per survey period, per plot.

Bayesian Hierarchical Multi-Species Abundance Model

My study design consists of nested plots within larger research units. To account for random block effects (i.e., plots within research units) and the potential for spatial autocorrelation between plots (Royle and Dorazio, 2008) I used a modified version of the multi- species occupancy model (Royle and Dorazio, 2008), where individual species were treated as random effects taken from a higher hierarchical level (i.e., community level distribution). The model has been further extended to include the Royle and Nichols’ (2003) method that assumes a link between variations in abundance with variation in detection probability to derive an abundance estimator. This method was incorporated into the multi-species occupancy framework by Yamaura et al. (2011). Under this framework, my estimate of abundance is considered a measure of relative abundance per species at each plot (Yamaura et al., 2011; Beesley et al.,

2014). In the case of camera trapping studies, Tobler et al. (2015) defined the measure of abundance obtained from this model as the number of individuals using the area around the camera station, and also as an indicator of site preference by one or more individuals.

This model also incorporates estimates of detection probabilities to account for species specific imperfect detection that may introduce sampling biases (MacKenzie et al., 2006).

Incorporating detection probabilities is essential because differences in abundance may be due to sampling strategy rather than to the effects of the covariates we are testing for (Mackenzie et al.,

2006). Under the Royle-Nichols model (2003), detection probability (rij) is interpreted as the probability of detecting individuals of species i (average probability of detection of species i), while pij is the probability of detecting at least one individual of species i at plot j.

It is also important to note that the majority of the area’s study species have the ability to move in and out of the sampling plots within each study period, thereby violating the spatial closure assumption of occupancy modeling (Mackenzie et al., 2006). This implies that the

species richness estimates (with the exception of N, the total species richness of my study area) should be interpreted as number of species that visit or use a plot.

To model species richness and multi-species abundances, I divided each survey period in each plot into 5 one-day/night consecutive sampling occasions (MacKenzie et al., 2006; Royle and Dorazio, 2008) creating a matrix of 0’s and 1’s (0=absence, 1=presence). Where each column represented a species detected and the rows represented the plots surveyed.

I estimated the following posterior probability parameters with this model: species- specific detection probability (rij), total species richness for all plots (R), plot specific species richness (nj), an expected average abundance per species for all plots (1), and a species-specific abundance estimate per plot (λij). I estimated posterior means, standard deviations and 95% credible intervals for all parameters of interest (Table 2-4; Appendix B, Table B-1).

To estimate total number of mammalian species found in my study area, I included a data augmentation parameter (omega) to account for the fact that this parameter is unknown but the parameter space needs to be bound (Royle et al., 2007). Data augmentation is a process through which the existing data set is completed with all-zero detection histories based on a super community with a fixed upper bound M that comprises the n detected species. The augmented data set is modeled as a zero-inflated version of the complete data model (Royle et al., 2007) with a uniform prior distribution for the community size M and a Bernoulli distribution for omega (Yamaura et al., 2011). R, the community size or total species richness, is then estimated as a derived parameter. I augmented data to include a total of 40 species based on number of species reported for the area (Monadjem, 1998).

Temporal and Spatial Covariates

I used the multi-species hierarchical Bayesian abundance framework to model the effects of temporal (i.e., season, year) and environmental covariates (i.e., visual obstruction (vo), distance to water (hereafter dwat), shrub, grass, and tree cover) on detection probabilities and abundance (Yamaura et al., 2011; Beesley et al., 2014).

I standardized all continuous covariates measured to mean zero and a standard deviation of 1 to avoid issues arising (e.g., numerical algorithms may fail to find parameter estimates) when the difference between covariates ranges over several orders of magnitude (White, 2011). I converted categorical covariates (i.e., year, season and number of cameras per plot) to values of

0’s and 1’s. I coded plots sampled during 2012 as 0, which turned the 2013 surveyed plots to 1 and made this the default year in the analysis. In the case of season, I turned plots surveyed in

June-July (winter) of each year to 0, making November-December (summer) the default survey season. For cameras, I coded all plots with 2 cameras deployed as 1 and coded plots with 1 camera as 0.

I used the following covariates to model species relative abundance: shrub, tree and grass cover, season, year, distance to water and an interaction term for distance to water and season. I added this interaction term to account for seasonality in water availability due to the presence of seasonal rivers in the study site. I added distance to water to the models because some species surveyed have a strong dependence on water sources (e.g., waterbuck (Kobus ellipsiprymnus))

(Skinner and Chimimba, 2005), implying that this covariate would have a strong influence on some members of the mammalian community sampled and I needed to account for this effect. I included temporal covariates to account for potential inter-annual or seasonal variation. For example, ungulates have been shown to track seasonal shifts in resource abundance and quality

in African savannas (McNaughton and Georgadis, 1986; Owen-Smith, 2002; Kleynhans et al.,

2011). As a result, species abundance and plot use may vary according to season.

To model detection probability, I used the following covariates: visual obstruction, shrub, tree and grass cover percentage, number of cameras per plot, season and year. Although I expected habitat and temporal covariates (with the exception of number of cameras per plot and visual obstruction) to influence relative abundance estimates, these measures could also influence detection probability. For example, animals may be more easily detected in open (i.e., grassy) vs. closed (i.e., shrubby) habitats. Considering that resources are more abundant in the rainy season, and more concentrated during the dry season, more individuals would be available to sample (i.e., more widespread distribution throughout the study area) in the rainy season. Year was included in the model to account for the fact that different researchers deployed camera traps across the two years and this could lead to bias in camera placement per plot that increased or decreased detection and also to account for potential inter-annual variation (e.g., variation between years in rainfall patterns). Visual obstruction measures per plot were expected to correlate positively with probability of detection because high vo measures suggested an area was less concealed by vegetation in front of the camera traps. Finally, I expected that deploying two cameras instead of one increased the probability of detecting an individual in the plot.

Model Structure

My model describes the latent variable Nij (species abundance) with a Poisson distribution as:

Nij ~Poisson (ij),

Where ij is the Poisson mean for species i at plot j. Therefore, I assumed that species abundance could vary in time and space depending on covariates in the form:

Nij ~Poisson (ij) with log(ij) = 1i + xj 2i

Where 1i and 2i are the parameters to be estimated for species i and 1i (intercept) is a random species effect. Thus, our final abundance model was the following: log(ij) = 1i + 2i shrubj + 3i treej + 4i grassj + 5iseasonj + 6iyearj + 7idwatj +

8idwat* seasonj

Similarly, the detection probability (rij) model was: log (rij) = 1i + 2i shrubj + 3i treej + 4i grassj + 5ivoj + 6icameraj + 7iseasonj +

8iyearj.

Community-Wide Model Selection Procedure

Hierarchical Bayesian models, especially multi-species models, are too complex for standard model selection procedures such as Akaike Information Criterion (AIC) and even

Bayesian Deviance Information Criterion (DIC) and these approaches may provide unreliable results (Kery, 2010; Burton et al., 2012). For this reason, I used an approach followed by Burton et al. (2012), where the support for each covariate in the full model is evaluated by modeling the probability of inclusion using a mixture-modeling approach (Royle and Dorazio 2008). Each covariate is multiplied by an “inclusion parameter” (wi), which is a latent Bernoulli variable with uninformative prior probabilities equal to 0.5 (equal probability of covariate being included or not in the model). It is important to note that the inclusion probability parameter is estimated for the effect of the covariate at the community and not the species level (Burton et al., 2012).

Species-Specific Model Selection Procedure

To determine the species-level importance and direction of effect for the covariates modeled, I evaluated whether or not the credible interval of the posterior estimates overlapped

zero and the position of the distribution with regards to zero to assess direction of the association

(i.e., positive or negative).

I analyzed hierarchical models simultaneously using software JAGS (Plummer, 2003) and R (R Development Core Team, 2011). I ran three parallel chains of length 200,000, after a burn-in of 40,000 iterations and a thinning rate of 100. I assessed convergence using the Gelman-

Rubin diagnostic where an estimate >1.1 indicates low convergence rates between chains

(Brooks and Gelman 1998).

Results

We surveyed for a total of 1,980 trap nights. We recorded a total of 22 species on the sampling plots over the 2-year period (Table 2-6). Several species had < 10 detection over all surveys: serval, spotted hyena, side-striped jackal (Canis adustus), vervet monkey (Chlorocebus pygerythrus), and porcupine (Hystrix africaeaustralis). Despite this, the multi-species model parameters converged satisfactorily (Gelman-Rubin statistic ≤ 1.1: model median = 1.01, max =

1.19) (Brooks and Gelman 1998).

Plot Comparisons Based on Habitat Covariates

Plots surveyed varied in vegetation measurements (Table 2-2). The principal components analysis (Table 2-3 and Figure 2-4) shows that the first component, which explains most of the variation in the data (59.6%), groups plots according to values of shrub, tree and bare ground cover, while the rest of the plots are grouped according to grass cover percentage, visual obstruction and grass biomass values. Shrub cover was negatively correlated to grass cover and visual obstruction (r2 = -0.33, P = 0.003; r2 = -0.27, P = 0.01, respectively). These results indicate that plots with high grass cover and values of visual obstruction and grass biomass may differ in their overall structure and composition to those dominated by shrubs and trees, offering different resources to the mammalian community of the area.

Species Richness Estimates

Posterior probability distributions (Figure 2-5) for total species richness (N) across the study site had values between the number detected (22) and 23 total species, indicating that almost all species present in the study area were detected in my surveys. There was a nonlinear response of number of species per plot to shrub and tree cover that did not show a clear pattern

(Figure 2-6). Nonetheless, estimates of richness were higher at low or intermediate values of shrub and tree cover and the highest estimates of species richness occurred below 50% shrub and tree cover (Figure 2-6). The number of species per plot tended to increase with grass cover percentage and I found the higher estimates of richness associated with higher values of grass cover (Figure 2-6).

Community-Level Effects of Covariates on Abundance and Probability of Detection

Posterior probabilities for my model suggested grass cover, season and distance to water as important covariates influencing mean community abundance (Table 2-5) and grass cover, visual obstruction, number of cameras per plot, season and year as important covariates influencing the probabilities of detection (Table 2-6). Shrub and tree cover were not important predictors of mean community abundance or detection probabilities at a community-wide level, while grass cover received support for influencing both parameters (estimated inclusion probability parameter from Tables 2-5 and 2-6). An inclusion probability estimate of 0.5 (Table

2-5) for grass cover should be interpreted as grass cover being important for the overall community, but that the influence of this covariate is not consistent across all species. Examining plots of abundance across species, it is clear that abundance increases (Figure 2-7) with grass cover, but decreases with tree cover and, most markedly, with shrub cover.

Covariate Effects on Species-Specific Relative Abundance Estimates

Fourteen species showed a response to covariates used in the multi-species model (Table

2-7). The covariate that most species responded to positively was percentage grass cover (n=8), followed by season (n=6). Only waterbuck responded negatively to shrub cover, while Chacma baboon (Papio ursinus) showed a positive response to this covariate (Figure 2-10). Wildebeest

(Connochaetes taurinus) responded negatively to tree cover, but positively to grass cover. All effects of season were positive, suggesting higher abundances per plot during the summer

(November to December) survey for species where this covariate was important. (Figure 2-9).

The rusty-spotted genet (Genetta maculata), impala (Aepyceros melampus), and duiker

(Sylvicapra grimmia) responded negatively to the year covariate, showing a decrease in abundance between the first and second year of sampling. Finally, both zebra (Equus quagga) and impala abundances show a positive response to distance to water.

Covariate Effects on Species-Specific Probability of Detection

Visual obstruction negatively influenced the probability of detection for 10 species

(Table 2-7) and the probability of detection for five species was reduced during the winter months (June to July). Other covariates that influenced probability of detection were year, number of cameras per plot and grass cover.

Species-Specific Relative Abundance Estimates

Examining the graphical representation of species specific abundance, estimates were higher for several species during the second year of sampling (Figure 2-8) and also during the summer (November; Figure 2-9), although credible intervals of estimates mostly overlap, denoting similar estimates. My results show that mixed feeders (impala and nyala) and less selective browsers such as, bushbuck (Tragelaphus scriptus), duiker, and kudu (Tragelaphus strepsiceros) had higher relative abundance estimates than grazers (zebra, wildebeest, and

waterbuck) and the more selective browser giraffe (Giraffa camelopardalis). In the case of carnivores, the more generalist rusty-spotted was more abundant than other grassland associated species such as the serval, slender mongoose (Galerella sanguinea), spotted hyena and side- striped jackal. Species from other taxonomic groups associated with grasslands were also less abundant across the study site (porcupine, aardvark (Orycteropus afer), baboon and vervet monkey). This trend is also observed when relative abundance estimates are averaged across years and seasons (Figure 2-11).

Discussion

Species Richness and Community-Level Responses to Habitat Covariates

My results partially support my Hypothesis that woody encroached areas are used less frequently and by a fewer number of species. The number of species recorded and abundance estimates across all species were higher at low or intermediate values of shrub and tree cover

(≤50% shrub and tree cover), suggesting there is a threshold of shrub and tree cover that reduces diversity. Conversely, the number of species and abundance estimates across species increased with grass cover, suggesting a preference for this vegetation cover type.

Contrary to my expectations, I did not find a negative influence of shrub and tree cover at the community level, but grass cover was an important covariate included in the final community-wide model of abundance. These results suggest that grass cover was important for the overall community, but its influence was not consistent across all species. This coincides with the expected effects of woody encroachment on a guild or community. Studies on other taxa have found that woody encroachment does not increase species diversity, but instead alters community composition (Meik et al., 2002; Blaum et al., 2007a; Blaum et al., 2007b; Sirami and

Monadjem, 2012). More specifically, it benefits some species, particularly habitat and dietary generalists, over others. Consequently, my results suggest that certain species in the area still

utilize and possibly benefit from woody encroached plots. Nevertheless, my results indicate that grass cover constitutes an important component of the savanna system that helps drive species diversity. Therefore, it may not be increased of shrub and woody cover that affects mammalian species richness and cause diversity declines in the area as has been found in other studies

(Eldridge and Soliveres, 2015), but the loss of grass cover associated with woody encroachment.

Species-Specific Covariate Effects and Abundance Estimates

Contrary to my expectations and in support of my community-wide results, almost no species responded negatively to habitat covariates associated with woody encroachment.

Mammalian responses to habitat covariates indicated that neither shrub nor tree cover appear to restrict movement and space use of most species studied at the scale of my study. Studies on impacts of woody encroachment on other taxa (e.g., small mammals, meso-carnivores, lizards, birds) report a species-specific response to woody cover (Meik et al., 2002; Blaum et al., 2007a;

Blaum et al., 2007b; Sirami and Monadjem, 2012). I found only two species responded negatively to shrub and tree cover (waterbuck and wildebeest respectively) and only one species

(baboon) responded positively to shrub cover. Both waterbuck and wildebeest are large grazers with low habitat and dietary breadths (Barthelemy et al., 2008; Jones et al., 2009; O’Kane et al.,

2014), making them highly dependent on plots where grass is the dominant type of cover. Both these species also presented lower relative abundance measures in my study. Consequently, both these species should be classified as potentially vulnerable to the effects of grass loss due to woody encroachment.

Species-specific results of my study suggest that taxonomic groups may respond differently to woody encroachment or that woody encroached plots are not completely degraded and may still provide some benefits (e.g., thermal and reproductive cover) to mammals,

including some grazers (i.e., zebras and warthogs). At the very least, they are still penetrable and act as corridors between preferred habitats within the study site.

In the case of carnivores, shrub cover can reduce prey availability but also provide suitable cover and may be beneficial to certain species until specific thresholds are met (Blaum et al., 2007a; Blaum et al., 2007c). Some ungulates may benefit from a diversity of vegetation provided by a landscape configuration made up of encroached and non-encroached patches that offer resources for browsers and mixed feeders, especially during drier periods when certain resources become scarce (Owen-Smith, 2002; Hobbs and Gordon, 2010). Nevertheless, as the level of encroachment increases, reductions in available foraging sites and competition should increase. Plots with mean shrub and tree cover percentage below 52 and 58% respectively were used by most species, suggesting these values should be considered as thresholds above which use of plots by species may decline.

Despite the weak response of species to shrub and tree cover, my study found that grass cover was an important positive predictor for eight species and important for the community as a whole, suggesting a strong dependence of certain species on this vegetation type. This is further evidence that woody cover may not be detrimental as long as grass cover can be maintained to support adequate populations and species diversity.

Relative abundance estimates for species which use areas covered with tall grass and riparian vegetation (e.g., servals and waterbucks) (Barthelemy et al., 2008; Ramesh and Downs,

2013) were low and may reflect vulnerability to woody encroachment. Thus, for the most part, patterns of abundance in my study reflect habitat and dietary preferences (generalists vs. specialists) in support of my second Hypothesis.

My results are comparable to Smit and Prins (2015), who found that increasing woody cover results in a shift in ungulate communities where browsers and mixed-feeders replace grazers. Ungulates with lower relative abundance estimates, and hence higher potential vulnerability to woody encroachment, are mostly grazers and also those with higher body mass values (i.e., zebra, wildebeest, and waterbuck). This is further supported by their positive association to grass cover, indicating that large-bodied ungulates with low niche breadths may be more vulnerable to increasing woody cover and the loss of grass cover in the area. Overall my study found similarities with other studies in that grassy habitats are an important driver of mammalian species abundance and distribution (Gandiwa, 2013).

Management and Conservation Implications

My results suggest a potential vulnerability of the medium- and large-sized mammal community to the loss of grasslands, possibly as a result of woody encroachment. Management strategies in woody encroached plots across the study area should focus more effort on the maintenance of adequate grass cover to increase the use and relative abundances of grazers and grassland associated species and maximize overall species richness. For example, managers should aim to reduce horizontal and vertical shrub and tree cover, such that grass cover has both space and access to light to regenerate. However, shrub removal programs should consider that I did not find a strong negative response to shrub and tree cover in my study. This is an indication that low to medium levels of these resources (<50%) can sustain diversity, potentially through an increase in structural heterogeneity (Tews et al., 2004).

The loss of ungulate grazers and selective browsers can have negative consequences for the savanna system in the area (McNaughton, 1992; Sankaran et al., 2008). Grazers such as the zebra and blue wildebeest were present in lower relative abundances and small bodied ungulates such as the steenbok, which are more selective feeders, were absent. The mammalian community

in my study was dominated by mid-sized ungulates (body mass range = 15 – 350 kg) classified as browsers (bushbuck, duiker and kudu) and mixed feeders (nyala and impala) coinciding with results from Smit and Prins (2015).

Carnivores also showed a similar trend. For example, the genet, a generalist species, was more abundant than species associated with open savanna (i.e., serval, slender mongoose, and side-striped jackal). As a result, there appeared to be a shift in the mammalian assemblage that results in the replacement of savanna associated species by browsers, mixed-feeders and generalists. This should be cause for concern as wildlife species that depend heavily on grass cover co-evolved with C4 grasses in savannas and are considered important drivers of the habitat composition and structure of this system (McNaughton, 1992; Sankaran et al., 2005; Sankaran et al., 2008). The potential consequences for vegetation structure and function due to a shift in mammalian community diversity should be the focus of future research and management efforts.

Table 2-1. Information on camera trap surveys. Surveys were conducted for mid- and large-sized mammals in Mlawula Nature and Mbuluzi Game Reserves, Swaziland, 2012-2013.

Year Season Research Units sampled Cameras per plot 2012 Winter 1, 2, 3, 4, 5, 6, 7 2 Summer 1, 2, 3, 4, 5, 6 1 2013 Winter 1, 3, 4, 7, 8, 9 2 Summer 1, 3, 4, 7, 8, 9 2

Table 2-2. Mean values and descriptive statistics of habitat covariates. Covariates were used in the Bayesian hierarchical multi-species abundance model used to test for their association to community-wide and species specific abundances of mid- and large- sized mammals in Mlawula Nature and Mbuluzi Game Reserves, Swaziland, 2012- 2013.

Variable measured Mean SE Range Mode Shrub cover (%) 31.05 2.03 2.33 – 86.25 17.33 Tree cover (%) 37.82 2.35 0.00 – 94.17 36.67 Grass cover (%) 26.57 1.46 2.00 – 60.04 43.00 Visual obstruction (cm) 12.04 2.25 1.00 – 81.09 1.65 Bare ground (%) 42.59 2.83 1.00 – 78.00 7.50 Grass biomass (kg) 6.69 2.51 2.51 - 15.15 4.30 Habitat variables: shrub cover = mean shrub cover percentage per plot, tree cover = mean tree cover percentage per plot, grass cover = mean grass cover percentage per plot, visual obstruction = mean measure of vegetation height per plot using robel pole, bare ground= mean bare ground cover percentage per plot and grass biomass = mean grass biomass per plot measured using a disc pasture meter.

Table 2-3. Factor loadings from principle components analysis of habitat covariates. Covariates were included in the Bayesian hierarchical multi-species abundance model used to test for their association to community-wide and species specific abundances of mid- and large-sized mammals in Mlawula Nature and Mbuluzi Game Reserves, Swaziland, 2012-2013. Total explained variance of axes = 83.35%.

Variable PCA Axis 1 PCA Axis 2 PCA Axis 3 Shrub cover (%) 0.35 -0.32 0.85 Tree cover (%) 0.37 0.64 0.17 Grass cover (%) -0.43 0.40 0.25 Bare ground cover (%) 0.39 -0.46 -0.31 Visual obstruction (cm) -0.45 -0.24 0.02 Grass biomass -0.45 -0.26 0.29 Eigenvalues 3.57 0.78 0.65 Variation explained (%) 59.57 12.93 10.85 Habitat variables: shrub cover = mean shrub cover percentage per plot, Tree cover = mean tree cover percentage per plot, grass cover = mean grass cover percentage per plot, bare ground= mean bare ground cover percentage per plot, visual obstruction = mean measure of vegetation height per plot using robel pole, and grass biomass = mean grass biomass per plot measured using a disc pasture meter.

Table 2-4. Random variables and derived parameters of the multi-species hierarchical Bayesian model.

Variable Distribution / Equation Description Nij Nij Poisson(λij) Abundance of species i at plot j, where λij is the Poisson mean for abundance of species i at plot j

N rij rij = 1-(1- pij) ij Average probability of detection of species i wi wi Bernoulli(0.5) Model inclusion parameter for each covariate included in models for λij and rij omega ω Uniform(0,1) Data augmentation parameter

Derived parameters pij Probability of detecting at least one individual of species i at plot j nj Plot-specific species richness

R Sum of species richness across all plots

Table 2-5. Community-level effects of habitat covariates on relative abundance. Posterior probability summaries, 95% credible intervals (CI) and inclusion probabilities of hyper-parameters from Bayesian hierarchical multi-species abundance model for mean community-level effects of plot habitat covariates on relative abundance coefficients in Mlawula Nature and Mbuluzi Game Reserves, Swaziland, 2012-2013.

Parameter Mean 95% CI Inclusion (covariate of abundance) probability Intercept -1.48 -2.20, -0.82 NA Shrub cover -0.06 -6.33, 6.22 0.00 Tree cover -0.04 -6.12, 6.14 0.09 *Grass cover 0.06 -5.22, 5.16 0.50 *Season 0.07 -5.33, 5.32 0.50 Year 0.00 -5.67, 5.81 0.27 *Distance to water 0.04 -5.10, 5.44 0.50 Distance to water*season -0.02 -6.14, 6.14 0.02 Parameters: shrub cover = mean shrub cover percentage per plot, Tree cover = mean tree cover percentage per plot, grass cover = mean grass cover percentage per plot, season = season survey occurred (rainy or dry season), year = year survey occurred (2012 or 2013), distance to water = distance of plot center to water source and an interaction term between the covariates distance to water and season. * Indicates covariate is included in final model.

Table 2-6. Community-level effects of habitat covariates on detection probability. Posterior probability summaries, 95% credible intervals (CI) and inclusion probabilities of hyper-parameters from Bayesian hierarchical multi-species abundance model for mean community-level effects of plot habitat covariates on probability of detection coefficients in Mlawula Nature and Mbuluzi Game Reserves, Swaziland, 2012-2013.

Parameter Mean 95% CI Inclusion (covariate of abundance) probability Intercept -2.84 -3.60, -2.26 NA Shrub cover 0.01 -3.65, 3.64 0.00 Tree cover 0.00 -3.82, 3.68 0.00 Grass cover 0.04 -2.91, 3.03 0.47 *Visual obstruction -0.62 -0.98, -0.30 1.00 *No. of cameras per plot 0.13 -0.53, 0.69 1.00 *Season -0.18 -2.10, 2.17 0.78 *Year -0.19 -2.41, 2.26 0.74 Parameters: shrub cover = mean shrub cover percentage per plot, tree cover = mean tree cover percentage per plot, grass cover = mean grass cover percentage per plot, visual obstruction = mean measure of vegetation height per plot using robel pole, no. of cameras = number of cameras (1 or 2) deployed per plot, season = season survey occurred (rainy or dry season), year = year survey occurred (2012 or 2013). * Indicates covariate is included in final model.

Table 2-7. Relative abundance (λ) and detection probability (p) estimates and the direction of the effect of statistically significant covariates. Measures for mammals detected in camera-trapping surveys conducted in Mlawula Nature and Mbuluzi Game Reserves, Swaziland, 2012-2013. Statistical significance is based on 90% credible interval non-overlap with zero.

Species Common Covariate Covariate λ (SD) p (SD) name effects on λ effects on p Orycteropus afer Aardvark 0.09 (0.07) 0.07 (0.04) season(+) vo(-) Papio ursinus Baboon 0.28 (0.13) 0.09 (0.04) shrub(+), grass(+) vo(-) Tragelaphus scriptus Bushbuck 0.83 (0.26) 0.07 (0.03) season(+) camera(+), season(-), year(+) Potamochoerus larvatus Bushpig 0.22 (0.15) 0.05 (0.03) Sylvicapra grimmia Duiker 0.92 (0.36) 0.06 (0.03) year(-) season(-) Genetta maculata Rusty-spotted genet 0.68 (0.28) 0.08 (0.04) season(+), year(-) grass(-), vo(-), season(-) Giraffa camelopardalis Giraffe 0.12 (0.12) 0.05 (0.04) grass(+) season(-) Crocuta crocuta Hyena 0.04 (0.09) 0.04 (0.04) Aepyceros melampus Impala 0.75 (0.18) 0.08 (0.04) season(+), year(-), dwater(+) vo(-), camera(+) Canis adustus Side-striped jackal 0.07 (0.10) 0.07 (0.05) grass(+) vo(-) Tragelaphus strepsiceros Kudu 0.72 (0.30) 0.09 (0.04) grass(+) vo(-) Tragelaphus angasii Nyala 0.67 (0.17) 0.12 (0.04) grass(+), season(+) vo(-), season(-), year(+) Hystrix africaeaustralis Porcupine 0.04 (0.08) 0.06 (0.04) Leptailurus serval Serval 0.08 (0.11) 0.05 (0.04) Galerella sanguinea Slender Mongoose 0.07 (0.10) 0.04 (0.04) grass(+) Chlorocebus pygerthrus Vervet Monkey 0.07 (0.09) 0.05 (0.04) Phacochoerus africanus Warthog 0.53 (0.16) 0.14 (0.05) vo(-) Kobus ellipsiprymnus Waterbuck 0.10 (0.12) 0.04 (0.04) shrub(-) Atilax paludinosus Water Mongoose 0.04 (0.09) 0.04 (0.04) Connochaetes taurinus Wildebeest 0.23 (0.10) 0.10 (0.05) can(-), grass(+), season(+) vo(-) Equus quagga Zebra 0.31 (0.23) 0.05 (0.03) grass(+), dwater(+) vo(-) Parameters: shrub = mean shrub cover percentage per plot, can = mean tree cover percentage per plot, grass = mean grass cover percentage per plot, vo = mean measure of visual obstruction per plot using robel pole, dwater = distance of center of plot to water source, camera = number of cameras (1 or 2) deployed per plot, season = season survey occurred (rainy or dry season), year = year survey occurred (2012 or 2013).

Figure 2-1. Map of study area location. Insert map shows location of Swaziland within the continent of Africa (red square), while main map shows location of nature and game reserves surveyed in Swaziland, 2012-2013.

Figure 2-2. Map of research design in Mlawula Nature and Mbuluzi Game Reserves, Swaziland. Larger map shows location of research sampling units (red squares) within the study area surveyed in 2012-2013. Upper right hand corner map denotes location of sampling plots (red circles) within two research units.

Figure 2-3. Spatial setup of sampling plots (circles) for camera trap and vegetation surveys. Plots were located within each research unit in Mlawula Nature and Mbuluzi Game Reserves, Swaziland, 2012-2013.

Figure 2-4. Results of principal components analysis of habitat covariates. Covariates are for Mlawula Nature and Mbuluzi Game Reserves, Swaziland (2012-2013). Variables measured per plot were: grass, shrub, tree and bare ground cover percentage, visual obstruction (VO) and grass biomass (DPM).

Figure 2-5. Posterior probability distribution of total species richness of mid- and large-sized mammals. Estimates are for Mlawula Nature and Mbuluzi Game Reserves, Swaziland (2012-2013) obtained from a Bayesian hierarchical multi-species abundance model.

Figure 2-6. Plot-specific species richness versus vegetation cover measures (mean value of vegetation type cover percentage in each plot). Estimates are for Mlawula Nature and Mbuluzi Game Reserves, Swaziland (2012-2013) obtained from Bayesian hierarchical multi-species abundance model.

Figure 2-7. Total relative abundance across species versus vegetation cover measures (mean value of vegetation type percentage in each plot). Estimates are for Mlawula Nature and Mbuluzi Game Reserves, Swaziland (2012-2013) obtained from Bayesian hierarchical multi-species abundance model.

Figure 2-8. Comparison of species-specific predictions of mean relative abundance per plots across survey years. Estimates for ungulates (A) and carnivores, primates and other taxa (B) for Mlawula Nature and Mbuluzi Game Reserves, Swaziland (2012-2013). Estimates obtained from Bayesian hierarchical multi-species abundance model. Bars represent 95% credible intervals. Ungulates are classified according to feeding guild and ordered in descending order according to body mass within guild.

Figure 2-9. Comparison of species-specific predictions of mean relative abundance per plots between seasons (June: winter and November: summer). Estimates for ungulates (A) and carnivores, primates and other taxa (B) for Mlawula Nature and Mbuluzi Game Reserves, Swaziland (2012-2013). Estimates obtained from Bayesian hierarchical multi-species abundance model. Bars represent 95% credible intervals. Ungulates are classified according to feeding guild and ordered in descending order according to body mass within guild.

Figure 2-10. Predicted relative abundance as a function of habitat covariates. Estimates obtained for mid- and large-sized mammals in Mlawula Nature and Mbuluzi Game Reserves, Swaziland (2012-2013) from a Bayesian hierarchical multi-species abundance model. Covariate effects shown for mean shrub, tree and grass cover percentage per plot. Only species with significant habitat effects are plotted.

Figure 2-11. Predicted mean relative abundance per plot surveyed and 95% credible intervals. Estimates obtained for mid- and large-sized mammals in Mlawula Nature and Mbuluzi Game Reserves, Swaziland (2012-2013) from a Bayesian hierarchical multi- species abundance model. Line divides savanna or open habitat associated species (left side) from species with wider niche breadths not necessarily associated to open or savanna habitat or that may use or prefer dense woodland or thickets to open habitats.

CHAPTER 3 ARE UNGULATE CO-OCCURRENCE PATTERNS INFLUENCED BY WOODY ENCROACHMENT?

Background

African savannas exhibit a higher diversity of ungulates than any other ecosystem in the world (Du Toit and Cumming, 1999). This high diversity is primarily driven by habitat structural heterogeneity and resource partitioning (Gandiwa, 2013). Species co-exist by consuming different plant species, and different parts and types of plant materials (grass, roots, leaves, fruits) (Bell, 1970; Jarman, 1974; Kleynhans et al., 2011; Hopcraft et al., 2012; Kartzinel et al.,

2015). African ungulates also have variation in an array of physiological traits (body mass, muzzle width, and digestive system [ruminant vs. non-ruminant]), and behaviours (e.g., seasonal patterns of resource selection) that allow them to partition resources (Jarman, 1974; Kleynhans et al., 2011; Hopcraft et al., 2012).

Despite resource partitioning strategies, the balance between African savannas and ungulates may be threatened by anthropogenic stresses. Along with other grassland systems across the world, African savannas are experiencing rapid changes in the structure and composition of their plant community. These changes appear to be driven by livestock overgrazing, increased atmospheric carbon dioxide, altered rainfall patterns, and altered fire regimes (Milton and Dean, 1995; Roques et al., 2001; Sankaran et al., 2005; Wigley et al., 2010).

Together these factors appear to be causing a reduction in grass cover, and an increase in shrubby species (i.e., woody encroachment).

As many grazing ungulate species depend on palatable grass, they are expected to respond negatively to increased woody vegetation. Selective browsers such as giraffes and small ungulates also rely on specific tree and shrub species that may be reduced with increases in certain woody vegetation (McNaughton and Georgadis, 1985). Herbivory also plays an

important role in limiting woody cover in savannas, thus the loss of selective browsers and grazers may create a positive feedback loop exacerbating woody encroachment (Du Toit and

Cumming, 1999; Augustine and McNaughton, 2004; Sankaran et al., 2005; Holdo et al., 2013;

Pringle et al., 2014).

Changes to plant communities such as woody encroachment may increase the potential for competition in species that have co-evolved to partition resources (Arsenault and Owen-

Smith, 2002; O’Kane et al., 2013; Ferretti et al, 2015). No study, to my knowledge, has examined whether African ungulate co-existence and spatial overlap are influenced by stressors such as woody encroachment. Furthermore, the effects of woody encroachment on ungulates has received little attention (Smit and Prins, 2015), with only one study that directly examined their response to woody encroachment (Smit and Prins, 2015), and a few others that focused on the response to habitat management interventions (Ben-Shahar, 1992; Isaacs et al., 2013). Thus, the potential mechanisms through which woody encroachment may alter abundances and community assemblages (Chapter 2) have not been identified. Understanding whether woody encroachment influences ungulate co-existence patterns can help us predict the direction of community assemblages and local abundances to inform management strategies in areas where woody encroachment is threatening wildlife species persistence.

My goal for this Chapter was to determine whether patterns of ungulate spatial co- occurrence were influenced by woody encroachment. Ungulate diversity and co-existence in

African savannas has been found to be influenced by plant structural heterogeneity (Owen-

Smith, 2002; Hobbs and Gordon, 2010; Gandiwa, 2013). The complex structure (i.e., tree-grass mosaic) of most African savannas provide a wide array of resources to sustain a diverse assemblage of ungulates. As a result, I first tested whether plant structural heterogeneity was

negatively affected by shrub cover. Sirami et al. (2009) found that certain levels of woody encroachment resulted in reduced plant structural heterogeneity in a South African reserve.

Therefore, structural heterogeneity may be the mechanism by which woody encroachment alters ungulate spatial distribution.

I then asked if woody encroachment influenced the degree of spatial overlap of species of similar body mass and foraging guild (i.e., grazers, browsers, and mixed feeders), and determined if the influence of shrub cover on spatial co-occurrence was stronger than other factors identified in the literature as drivers of ungulate co-occurrence patterns (i.e., season, foraging guild, and/or mass differential). Spatial overlap in ungulates should be stronger during the dry season when resources are scarce and for species belonging to the same foraging guild, and with larger body mass differential between them (Kleynhans et al., 2011). However, rapid changes in the plant community such as woody encroachment may lead to reduced and concentrated resources and are thus expected to increase spatial overlap among species depending on these resources.

I hypothesize that woody encroachment forces species of similar guilds to increase their spatial overlap due to a concentration of resources, regardless of mass differential or season. This occurs because woody encroachment reduces available space (by creating a denser, woody understory) and plant structural diversity, leading to depleted and spatially clustered resources at shrub encroached sites. Consequently, I predicted an increase in spatial overlap among similar and potentially competing species in areas of woody encroachment and a stronger influence of shrub cover on spatial overlap when compared to other factors.

Methods

Study Site

I conducted this study in the north-eastern region of Swaziland, in the Mlawula Nature

Reserve and the Mbuluzi Game Reserve (Figure 2-1). Both reserves combined cover an area of

19,800 ha. The study site mainly consists of a low lying savanna with large trees (i.e. Acacia nigrescens and Sclerocarya birrea) scattered along a grass layer dominated by Themeda triandra and Panicum maximum (Gertenbach and Potgieter 1975, Roques et al., 2001) with patches of riparian forest along rivers (Sweet and Khumalo, 1994). This plant type is known as basal sweet arid lowveld (Murcina and Rutherford, 2006). Woody encroachment in the reserves has led to an increasing shrub layer in the area that averages 30-40% across the study site (Roques et al., 2001;

Sirami and Monadjem, 2012). The dominant species in this shrub layer is Dichrostachys cinerea

(Roques et al., 2001). The average monthly temperature for the region is 18 ºC in July and 26 ºC in January (Goudie and Price Williams, 1983). Rain falls predominantly during the summer months (October-April) and fluctuates about an annual mean of 675 mm (Swaziland

Environmental Authority, 2001).

Research Design

We conducted camera trapping and vegetation surveys along a gradient of woody encroachment at nine research units (25 ha), each subdivided into nine sampling plots (30 m2) separated by a distance of 250 m (Figures 2-2 and 2-3), resulting in a total of 81 sampling plots.

We sampled plots once a year per season (winter and summer, occurring from June to August and December to February respectively) for two years (2012-2013).

Effects of Shrub Encroachment on Plant Structural Heterogeneity

I used an estimate of structural evenness per plot derived from shrub (woody species < 2 m), tree (woody species > 2 m), and grass cover percentage values (see methods, Chapter 2) as a

measure of plant structural heterogeneity across plots. I categorized plots as being either shrub encroached or not, depending on a shrub cover percentage threshold of 30% shrub cover. I used a threshold of 30% because plant structural diversity has been found to decline above this level of shrub encroachment (Sirami et al., 2009).

I converted the values of each cover type per plot so they all summed to one. I then obtained Shannon’s diversity index (H) using package “vegan” in R (R Development Core

Team, 2011; Oksanen et al., 2015) and estimated Shannon’s equitability index (EH) as:

H/Hmax

Where Hmax = ln (tc) and tc represents the total number of cover types (n = 3).

The index values range from 0 to 1, with a value of 1 indicating complete evenness. High evenness estimates represented structurally heterogeneous plots because habitat cover values were more similar to each other and hence indicated that the plot was not dominated by a particular cover type.

I ran a Pearson’s product moment correlation in program R (R Development Core Team,

2011) between shrub cover percentage and the measure of structural evenness separately for low and high shrub cover plots to test whether shrub encroachment reduced plant structural heterogeneity.

Mammal Surveys with Camera Traps

We deployed camera traps at each plot for 5 days during the summer (December to

February) and winter (June to August) seasons for two years (2012 and 2013) throughout both reserves (Table 2-1). Not all research units were sampled each year, but each unit was sampled at least once. We used two camera traps ([Primos Truth Cam 35], Primos Hunting, St. Flora,

Mississippi) per plot during all surveys with the exception of one survey period (November to

December 2012) in which we only deployed one camera trap per month due to logistical

constraints. We mounted the motion detection cameras to a tree 40-60 cm above ground with a clear view of at least 10 m in front of the camera. Camera traps were placed in areas thought to have wildlife traffic (e.g., trails, open grazing swards) to increase probability of detection. We cleared the area in front of the camera of grass and shrubs, and any obstructing feature. During each sampling period, we deployed the camera traps for a period of 5 days/nights in each plot; resulting in a total of five sampling occasions per survey period, per plot.

Test of Spatial Autocorrelation

My study design consisted of sampling units located near each other, potentially violating assumptions of independence. I tested for spatial autocorrelation in residuals obtained from occurrence models following Moore and Swihart (2005). The process involved obtaining residuals from occurrence models estimated in program Mark (White and Burnham, 1999) for each species separately and estimating Moran’s Index on them using software GeoDa (Anselin et al., 2006) to test for spatial autocorrelation. I found no evidence for spatial autocorrelation for any of the species analyzed (Table 3-1) and concluded that co-occurrence models would not violate the assumption of independent sampling units.

Co-occurrence Models

I ran pair-wise co-occurrence models between ungulate species of similar feeding habits

(grazers, browsers and mixed feeders) to examine patterns of spatial overlap across the study site and to examine the influence of shrub cover on ungulate spatial overlap at the plot level. Species co-occurrence models test whether two species tend to occur more or less often together than expected under a hypothesis of spatial independence (Bailey et al., 2009). These models use presence/absence data to model species co-occurrence patterns based on occurrence probabilities and account for imperfect detection by incorporating estimates of detection probabilities (Bailey et al., 2009; Richmond et al., 2010; Waddle et al., 2010). All species in the study had larger

home ranges than the spacing between sampling plots, thus I interpreted occurrence estimates as the probability of using a site and not as number of sites or proportion of the study area occupied

(MacKenzie et al., 2004; MacKenzie et al., 2006).

I used a robust design two-species conditional occupancy model in program MARK

(White and Burnham, 1999; Mackenzie et al., 2004; Richmond et al., 2010). I used this model because it is stable when using covariates (Richmond et al., 2010) and it provides occurrence estimates for different sampling periods. Using a robust occupancy model allowed me to test for seasonal differences in occurrence and co-occurrence patterns where I divided the primary sampling periods by season and the secondary sampling periods by sampling occasions (survey days). Thus, I obtained occurrence and co-occurrence estimates for both seasons (i.e., dry and rainy).

The robust design two-species conditional model tests for occurrence patterns of a subordinate species given occurrence or absence of the dominant species (Richmond et al.,

2010). Larger ungulates have been found to exploit a wider range of food resources and depend more on quantity rather than quality resources, while smaller species are restricted to high quality diets (Illius and Gordon, 1992; Cromsigt et al., 2009; Owen-Smith, 2002). Additionally, because resources in African savannas are difficult to monopolize, exploitation competition is more common for ungulates in this system (Illius and Gordon, 1987). Furthermore, woody encroachment is expected to reduce resource quality (Milton and Dean, 1995; Blaum et al., 2007;

Van Auken, 2009), favoring larger ungulates. Thus, I assumed the larger ungulate of the pair to be the dominant species based on resource exploitation competition and their ability to subsist on low quality diets.

The two-species robust design model provides estimates for 16 parameters associated to probability of occurrence, detection, colonization and extinction for both species analyzed (Table

3-2). In this model, species A is assumed to be the dominant species over species B. The parameter A is the probability that species A occurs at a plot, regardless of the occurrence state of the other species; BA is the probability of occurrence of species B, given A is present; and

Ba is the occurrence probability of species B, given species A is absent. Probability of detection can also be estimated individually for each species in the absence of the other species (p) or as a conditional probability of detection (r). For example, rBA is the probability of detection for species B, given both species are present and species A is detected. The remaining parameterization of the model followed Richmond et al. (2010; [Table 3-2]).

The parameter I was particularly interested in was the species interaction factor (SIF or

). This parameter evaluates the degree of spatial overlap between the two species analyzed under a hypothesis of spatial independence (Bailey et al., 2009). The SIF is estimated as:

 = (A * BA) / (A (A*BA + (1 - A) Ba)

This factor is interpreted as follows:

 >1: The two species co-occur more frequently than expected

<1: The two species co-occur less often than expected

 = 1: The two species occur less often than expected and are said to be “spatially

independent” (also determined if the standard error overlaps 1.0)

Before running the two-species models, I evaluated best fit robust occurrence models for each species separately using habitat covariates measured per plot (Chapter 2). I then used the best occurrence and detection covariates for each species (Appendix D, Table D-1) in the robust design two-species occupancy models to estimate the species interaction factor.

Effects of Shrub Encroachment on Spatial Overlap

I estimated four species interaction factors for each pair-wise comparison, two per season for plots with shrub cover percentage < 30% and two per season for plots above this threshold. I first compared the SIF estimates to determine whether spatial overlap between pair-wise comparisons between low and high shrub cover differed.

I then compared the influence of shrub cover on the species interaction factor by season, weight differential, and foraging guild. I coded high (>30%) shrub cover percentage plots as “1” and low (<30%) shrub cover as “0”. I included season (dry or rainy) as a binary variable because

African savannas vary in quality and quantity across space and also across time (McNaughton and Georgadis, 1986; Owen-Smith, 2002; Kleynhans et al., 2011). I also included weight differential between each species as a continuous variable because ungulate size differential is also expected to lead to resource partitioning, where species of differing size are predicted to co- occur more often to avoid competition for resources (McNaughton and Georgiadis, 1986;

Arsenault and Owen-Smith, 2002; Kleynhans et al., 2011).

Finally, I used a categorical variable that indicated whether both species belonged to the same foraging guild or not. This covariate was used because spatial overlap should not be consistent among pairs examined and should depend on the degree to which they share resources

(i.e., higher spatial overlap between two grazers than between a grazer and a browser). Mixed feeders (i.e., nyala [Tragelaphus angasii] and impala [Aepyceros melampus]) have the potential to compete with both grazers and browsers. As a result, I ran pair-wise comparisons between mixed feeders and members of the other two guilds. I also included two suid species in the analyses (common warthog (Phacochoerus africanus) and bushpig (Potamochoerus larvatus)) because they are taxonomically associated and similar in size (Skinner and Chimimba, 2005).

Although both species exhibit different habitat, foraging, and temporal resource use patterns

(Skinner and Chimimba, 2005).

To avoid over-fitting the two-species co-occurrence model and possible convergence problems, I examined the influence of covariates on the species interaction factor using two methods: 1) a multiple linear regression, and 2) a hierarchical partitioning analysis (HP) (Chevan and Sutherland, 1991). I used the species interaction factor as the response variable and modeled the influence of the four explanatory variables (shrub cover, season, mass differential, and foraging guild) on this parameter. I used program R (R Development Core Team, 2011) to run the multiple linear regression using function “lm” and examined the model fit using the function

‘lmfit”, which provides plots of the residuals vs. the fitted values, a Q-Q plot of standardized residuals, a scale location plot, and a plot of residuals versus leverage to assess normality. The

HP analysis evaluates all possible models through a multiple regression approach and uses goodness-of-fit measures to estimate independent and joint contributions of each predictor

(Chevan and Sutherland, 1991). For this analysis I used package hier.part in program R (R

Development Core Team, 2011; Walsh and Mac Nally, 2013).

Results

We surveyed 1,980 trap nights and detected a total of 11 ungulate species. I could not examine patterns of resource partitioning for two species because of low number of detections, waterbuck (Kobus ellipsiprymnus) and giraffe (Giraffa camelopardalis).

Effects of Shrub Encroachment on Plant Structural Heterogeneity

I found a positive association between shrub cover percentage and heterogeneity for low shrub cover plots (R2 = 0.54; P = 0.0002) and an inverse association for high shrub cover plots

(R2 = -0.40; P = 0.0001) (Figure 3-1).

Effects of Shrub Encroachment on Spatial Overlap

Most pair-wise estimates of SIF increased with increasing shrub cover (indicating species

“attraction”) (Table 3-3). Grazers seemed to be the group most affected by shrub cover. In all but one pair-wise comparison including grazers, the species interaction factor increased with shrub cover for at least one of the two seasons (Table 3-3). The highest estimates of overlap were found for the zebra (Equus quagga) and wildebeest (Connochaetes taurinus) in shrub encroached plots. These were followed by SIF estimates for wildebeest and nyala, also in high shrub cover plots.

A decrease in the SIF (indicating low spatial overlap) as shrub cover increased occurred more often for browsers (i.e., kudu [Tragelaphus strepsiceros] and duiker [Sylvicapra grimmia]) and mixed feeders with browsers (nyala with kudu and duiker) in the dry season, with only the zebra (a grazer) and impala (a mixed feeder) presenting lower spatial overlap in high shrub cover plots.

Nyala was the species for which most pair-wise comparison estimates of SIF (n=5) differed between low and high shrub cover plots (Figures 3-2 and 3-3). This was followed by impala (n=3) and Kudu (n=3).

Diagnostic plots indicated the standardized residuals of the SIF regression behaved like a random sample from a normal distribution (Appendix F, Figure F-1), providing support for the fit of the model. Most estimates of SIF varied between seasons (Table 3-3), but the effect of season was most likely not unidirectional among all groups or not strong enough to be identified as an important covariate in the multiple linear regression model (MLR) (Table 3-4) or the hierarchical partitioning analysis (HP) (Table 3-5; Figure 3-4). The MLR model with all predictors explained 16% of the total variance (R2=0.16; P = 0.01). Shrub cover was found to be the only significant influential predictor variable for the SIF estimates. This variable was

positively associated to SIF, indicating that spatial overlap was higher between species pairs in areas with shrub cover over 30%. Furthermore, the HP analyses also confirmed these results, with shrub cover explaining over 80% of the total variance in the SIFs.

Discussion

Shrub cover consistently changed the interactions among pairs of ungulates in an African savanna. Increased spatial overlap was more common between grazers and between grazers and mixed feeders, suggesting that grass may have been a limited resource in the system (Chapter 2).

Although large herbivores can subsist on lower quality forage, they are negatively impacted by limited food resources (Illius and Gordon, 1992). Reduced grass cover and increased spatial overlap in light of woody encroachment has three potential implications for large grazers: 1) it reduces already limited space and resources which might not be sufficient to sustain them (i.e., preemptive competition [Gotelli, 2001]); 2) it increases grazing pressure, which has been shown to favor shrub encroachment (Roques et al., 2001); and 3) it forces grazers to depend on suboptimal forage, decreasing their fitness.

Unlike the pattern observed with grazers, several times I found some instances of low spatial overlap between species of other foraging guilds in areas of increasing shrub cover. This occurred more often for browsers and mixed feeders, and may suggest competition for resources between these groups (O’Kane et al., 2013). For example, duiker and kudus were found less often together in the dry season in both high and low shrub cover plots. However, spatial overlap for these two species increased in high shrub cover plots during the rainy season. Even though both duikers and kudus are browsers, their mass differential (Appendix C, Table C-1), should allow them to partition resources and thus overlap more extensively; but their low overlap during the dry season suggests resource partitioning to avoid competition when resources are scarce.

In areas of high shrub cover, spatial overlap of mixed feeders increased more with grazers than with browsers. Top single-species occurrence estimates for both mixed feeders included grass cover as an important covariate (Appendix D, Table D-1), suggesting that these species are relying on grass resources in the area. This indicates that mixed-feeders have more potential to compete with grazers than with browsers. This competition may increase as there is anecdotal evidence that nyalas are responding positively to woody encroachment and their abundances are increasing in the study area.

In contrast to my findings, African ungulates’ patterns of spatial overlap have been shown to be mediated by seasonal availability of resources, mass differential, and the foraging guild both species belong to (Du Toit, 2003). Alternatively, my study found that only woody encroachment helped to explain the patterns of spatial overlap among ungulates. I expected woody encroachment to limit and concentrate available resources, thereby promoting higher spatial overlap between similar species. In support of my Hypothesis, I found that species spatial overlap for most pair-wise comparisons was positively associated with shrub cover and that shrub cover was a stronger predictor of spatial overlap than other well established factors. This suggests that the changing vegetative structure within savannas appears to have a strong influence on spatial interactions of ungulates in areas of shrub encroachment. More specifically, my study found that similar species use the same areas and spatial overlap increases when shrub cover increases over 30%.

One explanation for the considerable changes in species spatial interactions in areas of increased shrub cover is the loss of plant structural heterogeneity. My results showed a negative association between plant structural heterogeneity and shrub cover, and are consistent with a previous study in the same area (Sirami et al., 2009). Consequently, low structural heterogeneity

across woody encroached plots may account for some of the patterns I observed in my study.

Additionally, there is also evidence that shrub cover was negatively associated with grass cover, which is essential for many African ungulates (Chapter 2). Woody shrubs in the region have been shown to extend from established clusters, leading to patchy patterns of resources (Bai et al., 2009). The dominant shrub in the area is Dichrostachys cinerea which is a small tree with rapid growth that tends to form dense, impenetrable thickets. Increasing cover of D. cinerea may be leading to reduced access to grass resources and spatial clustering of this resource throughout the study area.

In light of my results and predicted continued shrub encroachment in the region (Midgley and Thuiller, 2010; Bailey et al., 2016), we can expect species overlap to increase, leading to competition (Acebes and Malo, 2011) and higher pressure on areas with palatable resources. I recommend managers and researchers implement monitoring programs to examine whether continued spatial overlap in shrub encroached sites can lead to density dependent effects on less competitive species (Atickem and Loe, 2013; O’Kane et al., 2014). From my research, I would expect to see decreases in the numbers of meso-grazers (zebras and wildebeest). Continued shrub encroachment may also result in the decrease of smaller, less competitive species in the case of browsers and mixed-feeders. This suggests management and research should focus on these species. Researchers and managers should also try to determine whether the concentration of herbivores at shrub encroached sites will lead to further resource depletion and/or degradation.

If populations decrease and resources become degraded there are a number of potential management strategies that may be beneficial to species that are vulnerable to woody encroachment (Chapter 2) and increased competition. Localized mechanical removal of shrub cover has been shown to increase herbaceous plant productivity and create habitats for grassland

herbivores, while maintaining populations of ungulates associated to woodlands (Ben-Shahar,

1992). Similarly, increased burning frequency can result in higher densities of grassland associated species (Archibald and Bond, 2004; Klop and van Goethem, 2008) and may benefit browsers by allowing tree and shrub seedlings to establish in areas grazed after fire treatments

(Klop and van Goethem, 2008). Additionally, where it is feasible, the introduction of megafauna

(elephants) has been shown to suppress woody vegetation and open canopy cover in savanna environments (Wigley et al., 2010).

Table 3-1. Results of spatial autocorrelation test based on occurrence modeling residuals for ungulate species surveys in Swaziland, 2012-2013. Only species with a naive occurrence estimate () >0.10 were evaluated. NA indicates species was not detected or i<0.10 for that survey and thus was not included in analyses. Statistical significance was evaluated at p<0.05.

Species June 2012 November 2012 June 2013 November 2013 Common name Moran’s I p-value Moran’s I p-value Moran’s I p-value Moran’s I p-value Tragelaphus scriptus Bushbuck 0.13 0.88 0.03 0.93 0.17 0.79 0.05 0.79 Potamochoerus Bushpig 0.23 0.63 NA NA NA NA NA NA larvatus Sylvicapra grimmia Duiker 0.07 0.50 NA NA -0.07 0.99 NA NA Aepyceros melampus Impala -0.07 0.86 0.10 0.96 NA NA -0.12 0.99 Tragelaphus Kudu -0.01 0.79 0.17 0.19 0.15 0.86 NA NA strepsiceros Tragelaphus angasii Nyala 0.10 0.58 -0.08 0.34 0.14 0.71 0.38 0.72 Phacochoerus africanus Warthog 0.36 0.77 0.11 0.66 0.06 0.24 0.02 0.82 Connochaetes taurinus Wildebeest 0.15 0.58 0.07 0.45 -0.05 0.49 -0.02 0.57 Equus quagga Zebra 0.16 0.85 0.35 0.51 0.04 0.76 NA NA

Table 3-2. List and description of parameters estimated in the robust design two-species conditional occupancy model in program MARK. Species A is assumed to be dominant and species B subordinate. * Denotes parameters that are estimated for each primary session (season) separately.

Parameter Description

A Probability of occurrence for species A

BA Probability of occurrence for species B, given species A is present

Ba Probability of occurrence for species B, given species A is absent

*pA Probability of detection for species A, given species B is absent

*pB Probability of detection for species B, given species A is absent

*rA Probability of detection for species A, given both species are present

*rBA Probability of detection for species B, given both species are present and species A is detected *rBa Probability of detection for species B, given both species are present and species A is not detected A Probability of colonization by A, given species B was absent

B Probability of colonization by B, given species A was absent

AB Probability of colonization by A, given species B was present

BA Probability of colonization by B, given species A was present

A Probability of extinction of A, given species B was absent

B Probability of extinction of B, given species A was absent

AB Probability of extinction of A, given species B was present

BA Probability of extinction of B, given species A was present

Table 3-3. Estimates of the ungulate species interaction factor (SIF). The factor was estimated for pair-wise comparisons of ungulate species co- occurrence in Swaziland (2012- 2013). Last two columns correspond to covariates used to model their influence on the SIF estimate.

Species SIF (SE) for dry season SIF (SE) for rainy season Mass  Feeding Low shrub High Low shrub High (kg.) guild Species shrub shrub (lbs.) Bushbuck/duiker 1.29 (0.17) 1.33 (0.19) 0.57* (3.06) 0.91 (0.09) 13 Br

Kudu/bushbuck 0.86 (0.08) 1.29 (0.25) 0.98 (0.02) 1.15 (0.12) 74 Br Kudu/duiker 0.93 (0.07) 0.63 (0.16) 0.1 (0.08) 1.5 (0.34) 87 Br Nyala/bbuck 1.38 (0.23) 1.57 (0.25) 0.96 (0.07) 0.97 (0.06) 20 Br/Mf Nyala/kudu 1.14 (0.40) 0.91 (0.24) 0.98 (0.04) 0.79 (0.06) 54 Br/Mf Nyala/duiker 1.26 (0.35) 1.02 (0.24) 1.14 (0.70) 1.16 (0.19) 33 Br/Mf Bushbuck/impala 1.08 (0.11) 1.3 (0.2) 0.85 (0.08) 1.15 (0.09) 4 Br/Mf Kudu/impala 0.86 (0.08) 0.91 (0.18) 1.32 (0.23) 1.18 (0.11) 69 Br/Mf Impala/duiker 0.74 (0.12) 0.87 (0.23) 1.28 (0.29) 1.10 (0.19) 17 Br/Mf Nyala/impala 1.1 (0.32) 1.11 (0.23) 0.92 (0.10) 1.17 (0.14) 16 Mf Wildbeest/impala 1.14 (0.22) 1.11 (0.47) 0.68 (0.15) 0.99 (0.74) 66 Gr/Mf Zebra/impala 1.05 (0.15) 0.85 (0.57) 1.24 (0.37) 1.05 (0.13) 157 Gr/Mf Wildebeest/nyala 1.77 (0.30) 2.35 (0.49) 0.99 (0.08) 1.14 (0.11) 50 Gr/Mf Zebra/nyala 0.75 (0.22) 1.31 (0.46) 0.76 (0.30) 1.26 (0.22) 142 Gr/Mf Warthog/nyala 0.76 (0.20) 1.74 (0.39) 0.96 (0.10) 1.21 (0.13) 2 Gr/Mf Warthog/impala 1.24 (0.16) 1.62 (0.26) 1.24 (0.13) 1.04 (0.05) 14 Gr/Mf Zebra/wildebeest 0.91 (0.27) 1.17 (0.73) 1.67 (0.84) 3.21 (1.03) 91 Gr Zebra/warthog 1.02 (0.11) 0.93 (0.18) 0.83 (0.27) 1.17 (0.28) 144 Gr Wildebeest/warthog 1.28 (0.42) 1.60 (0.74) 0.78 (0.27) 1.30 (0.29) 53 Gr Warthog/bushpig 1.52 (0.44) 1.22 (0.12) 1.29 (0.80) 1.93 (0.38) 6 Sd * Indicates value not used in model because of convergence issues or inadequate standard error. Mass  indicates the mass differential between both animals for which SIF was estimated. For feeding guild: Br = browsers; Mf = mixed feeders; Gr = grazers; and Sd = suid species.

Table 3-4. Results from multiple regression analysis of the influence of predictor variables on the ungulate species interaction factor. R2 = 0.16. * indicates significance (P < 0.05).

Predictor variables Parameter estimate SE t value Pr (>|t|) Intercept 1.07 0.11 9.39 < 0.01* Shrub cover 0.23 0.09 2.63 0.01* Mass dif. -0.0004 0.0004 -1.05 0.3 Season -0.05 0.09 -0.53 0.59 Feeding guild 0.0002 0.09 0.002 0.99 Variables: Mass dif. = mass differential between both species; hrub cover = categorical variable indicating whether plots were shrub encroached (>30% shrub cover percentage) or not (<30% shrub cover percentage); Season = categorical variable indicating data collected in the dry or rainy season; and feeding guild = categorical variable indicating whether both species belonged to the same feeding guild (i.e., grazers, browsers or mixed feeders) or not (e.g., SIF estimated for a grazer and a mixed feeder).

Table 3-5. Results from hierarchical partitioning analysis testing the influence of predictor variables on the ungulate species interaction factor (SIF). Results shown are for independent contribution of each variable (I), its conjoint contribution with all other variables (J), total contribution and the percentage of total explained variance (I.perc) by the variable.

Predictor variables I J Total I.perc Mass dif. -0.02 -0.001 -0.02 12.75 Shrub cover -0.15 0.001 -0.15 82.9 Season -0.007 -0.001 -0.008 3.88 Feeding guild -0.001 -0.001 -0.002 0.46 Variables: Mass dif. = mass differential between both species; Shrub cover = categorical variable indicating whether plots were shrub encroached (>30% shrub cover percentage) or not (<30% shrub cover percentage); Season = categorical variable indicating data collected in the dry or rainy season; and feeding guild = categorical variable indicating whether both species belonged to the same feeding guild (i.e., grazers, browsers or mixed feeders) or not (e.g., SIF estimated for a grazer and a mixed feeder).

Figure 3-1. Association between shrub cover percentage and plant structural heterogeneity for plots surveyed in Swaziland (2012-2013). Plots were grouped according to low (shrub cover < 30% [A]) and high (shrub cover > 30% [B]) shrub cover percentages. Plant structural heterogeneity was a measure of evenness between values of shrub, grass and canopy cover percentage. The heterogeneity measure ranges from 0 to 1. Pearson’s correlation estimate is shown.

Figure 3-2. Comparisons of the species interaction factor between ungulates in Swaziland (2012- 2013). Comparisons are between low and high shrub cover plots for browsers (first three comparisons), mixed feeders (fourth comparison), and browsers and mixed feeders (last two comparisons). Bars represent standard errors of the species interaction factors. Only pair-wise comparisons for which SE’s did not overlap are shown.

Figure 3-3. Comparisons of the species interaction factor between ungulates in Swaziland (2012- 2013). Comparisons are between low and high shrub cover plots for grazers and mixed feeders (first six comparisons) in Swaziland. Bars represent standard errors of the species interaction factor. Only pair-wise comparisons for which SE’s did not overlap are shown.

Figure 3-4. Results of hierarchical partitioning analysis to determine the influence of covariates on the species interaction factor between ungulates in Swaziland (2012-2013). Covariates included in model and shown in plot are the following: mass = mass differential between both species for which SIF was estimated; shrub = indicates whether plots were shrub encroached (>30% shrub cover percentage; coded as 1) or not (<30% shrub cover percentage; coded as 0); season = indicates whether SIF was estimated for the rainy (coded as 1) or the dry (code as 0) season; and guild = indicates whether SIF was estimated for species of the same feeding guild (i.e., grazers versus grazer, browser versus browser, or mixed feeder versus mixed feeder; coded as 0) or for species from two different guilds (i.e., grazer versus mixed feeder or browser versus mixed feeder; coded as 1).

CHAPTER 4 COMPARISON OF STATISTICAL MODELING FRAMEWORKS AND SAMPLING EFFORT SCHEMES TO GENERATE UNGULATE ABUNDANCE ESTIMATES FROM CAMERA TRAPPING SURVEYS IN A WOODY ENCROACHED SAVANNA

Background

The ability to rigorously and reliably estimate the abundance of animals is critical for the sound management of wildlife (Williams et al., 2002). In African savannas, wildlife abundance and density estimates are usually obtained from direct observations with line transects (Ben-

Shahar, 1992; Klop and van Goethem, 2008) or airplanes (Dunham, 2012; Gandiwa, 2013).

These methods can be highly effective in savannas with an open canopy. However, their application is limited in areas with thick vegetation such as forests, woodlands, or woody encroached savannas. Furthermore, these sight-based methods likely underrepresent mid-sized ungulates such as bushbuck (Tragelaphus scriptus), nyala (Tragelaphus angasii), and duiker

([Sylvicapra grimmia]; Coates and Downs, 2007; Collier et al., 2011) that are associated with dense cover.

Alternatively, some studies have relied on indirect methods such as sign surveys

(Atickem and Loe, 2013) and the use of camera trapping technology to survey mammal communities within areas of thick cover (Foster and Harmsen, 2012; Tambling et al., 2013;

Burton et al., 2015; Tobler et al., 2015). Roberts (2011) found that camera trapping detected more species per site than sign surveys. Furthermore, differences in surveyor identification of signs make this method less reliable than camera trapping (Roberts, 2011). Camera traps are ideal because they allow researchers to survey closed canopy sites for long time periods and detect rare and elusive species that otherwise may be undetected (Kucera and Barret, 2011).

Camera trapping has predominately been used to generate estimates of abundance of animals with distinctive individual markings (e.g., stripes and rosettes on tigers and jaguars)

(Foster and Harmsen, 2012), as well as to understand distribution and site occupancy patterns of mammals (O’Brian, 2011; Silva and Sieving, 2012; Sollmann et al., 2012; Zapata-Rios and

Branch, 2016). However, the complexity of analytical techniques and constraints of assumptions have limited the use of cameras to estimate the abundance of species when individual recognition was not possible (e.g., most ungulates). While there are established modeling frameworks for generating abundance estimates of unmarked animals with camera traps (Royle and Nichols,

2003), few studies have actually used this approach (Burton et al., 2015; Tobler et al., 2015). For managers and researchers to realize the potential of camera traps to estimate abundances of unmarked animals in areas of high cover, it is important to determine if this is a viable methodology and to generate recommendations for the use of these analytic tools and optimal sampling schemes (i.e., sampling effort based on detection probabilities).

Royle and Nichols (2003) developed a statistical method that accounts for imperfect detection and uses presence/absence data to obtain average abundance estimates per site surveyed. This method, which is ideal for camera trapping, is known as the Royle and Nichols heterogeneity model (RN model hereafter) and is less restricted than traditional mark-recapture models because it does not require marking individuals in a population. There are two statistical modeling frameworks that can be used to obtain abundance estimates from RN models: finite- mixture models (FM models hereafter) and Hierarchical multi-species Bayesian models (Msp hereafter).

Finite-mixture models are a type of hierarchical model based on the traditional frequentist approach (i.e., maximum likelihood estimation) that generates population parameters by treating variation in the observed data as being derived from an observation and a state process (Royle and Dorazio, 2008; Fiske and Chandler, 2011). FM models have been extensively used in the

literature and are readily available in user-friendly software such as MARK (White and

Burnham, 1999), PRESENCE (Hines, 2006a), and Unmarked (Fiske and Chandler, 2011). One drawback is they rely on large sample sizes to converge and are not spatially explicit, resulting in potential violation of model assumptions of spatial independence between sampling sites.

Multi-species models have been adapted to fit the RN model to generate abundance estimates (Yamaura, et al., 2011). Msp models for abundance estimates are a modified version of the multi-species occupancy model (Royle and Dorazio, 2008), where individual species parameters are treated as random effects derived from a community-wide distribution, and differences in the observation and state processes are also accounted for (Yamaura et al., 2011).

An important advantage of the the Msp model is that it can account for spatial autocorrelation commonly occurring in camera trapping data (Yamaura et al., 2011; Beesley et al., 2014). This advantage allows the Msp model to provide more robust and reliable estimates of abundance.

Another advantage of the Msp model is that it also provides estimates for multiple species and measures of community structure such as species richness (Tobler et al., 2015). A disadvantage of Bayesian models such as the Msp model is that they are computationally complex, requiring high capacity computers and knowledge of advanced programming languages (Gotelli and

Ellison, 2004). Managers and researchers without advanced statistical knowledge may shy away from Msp models, turning instead to FM models. However, it is not clear if the more commonly used method of FM models provide the performance and precision of Msp models. No studies, to my knowledge, compare RN model abundance estimates between different statistical modeling frameworks.

Managers and scientists not only have to choose an appropriate modeling framework for their data, but the accuracy of the estimates from these approaches will ultimately be a function

of study design and sampling schemes. Bailey et al. (2007) recommended evaluating study designs and tradeoffs in spatial and temporal replication of occupancy studies to improve inference and estimation of abundance. However, given the lack of empirical studies, study designs under different sampling effort schemes for surveys that aim to use the RN model with camera traps have not been evaluated properly.

To determine if camera trapping can be a viable option to estimate ungulate abundance using camera traps on unmarked animals in thick cover, we conducted camera trapping surveys for three ungulate species in a woody encroached savanna in Swaziland, southern Africa. Our study species included a grazer (warthog [Phacochoerus africanus]), a browser (bushbuck), and a mixed feeder (nyala). I compared different modelling frameworks and levels of sampling effort under different detection probabilities. First, I compared RN model abundance estimates obtained from finite-mixture models under maximum likelihood estimation in software MARK to the more robust hierarchical Bayesian multi-species models using software package JAGS

(Plummer, 2003). I then simulated data based on our camera trapping estimates to compare competing sampling efforts that varied according to number of sites, temporal replication, and expected detection probabilities.

Methods

Study Site

I conducted this study in Mlawula Nature Reserve and Mbuluzi Game Reserve, which cover 17,400 and 2,400 ha, respectively, and are located in the low-lying savannas of north- eastern Swaziland (Figure 2-1). The vegetation community on these two reserves is characterized as basalt sweet arid lowveld (Sirami and Monadjem, 2012) with patches of riparian forest along rivers (Sweet and Khumalo, 1994). The dominant large trees were Acacia nigrescens and

Sclerocarya birrea and the grass layer was dominated by Themeda triandra and Panicum

maximum (Gertenbach and Potgieter 1975; Roques et al., 2001). Increases in shrub and canopy cover, along with a decrease in grass cover, has been documented for the study area (Roques et al., 2001; Sirami and Monadjem, 2012; Bailey et al., 2016). The dominant species in this shrub layer is Dichrostachys cinerea (Roques et al., 2001). The average monthly temperature in the low-lying savannas of Swaziland is 18 ºC in July and 26 ºC in January (Goudie and Price

Williams, 1983). Rain falls predominantly during the summer months (October-April) and fluctuates about an annual mean of 675 mm (Swaziland Environmental Authority, 2001).

The protected areas of the study were surrounded by a matrix of sugar cane fields, commercial cattle ranches, and populated homesteads (Bailey et al., 2016). These reserves were fenced, with very few large predators with the exception of some spotted hyena (Crocuta crocuta) and an occasional leopard (Panthera pardus).

Research Design and Data Collection

To obtain abundance estimates of ungulates, we setup research grids with sampling plots across the Mlawula Nature and Mbuluzi Game Reserves (Figures 2-1 and 2-2). We established 9 research 30 ha research grids throughout the two reserves (Figure 2-2). We sub-divided each grid into 9 sampling plots of 30 m2 with a distance of 250 m between plots. This resulted in a total of

81 plots. In each plot, we conducted camera trap surveys during the summer (November to

February) and winter (June to August) seasons for two years (2012 and 2013). We deployed camera traps for 5 days at each sampling plot during each survey. In most surveys, we deployed two camera traps ([Primos Truth Cam 35], Primos Hunting, St. Flora, Mississippi) per plot.

Camera traps were setup in areas we believed could increase detection probability (e.g., based on wildlife signs, trails and foraging sites). The only exception was the November, 2012 survey, in which we only deployed one camera trap per plot due to logistical constraints. We mounted the motion detection cameras on a tree 40- 60 cm above ground with a clear view of at least 10 m in

front of the camera. We then cleared the area in front of the camera of grass and shrubs and any obstructing feature. Camera trap locations were chosen based on signs of wildlife and areas (e.g., trails, grazing lawns, etc.) where wildlife detection would be maximized.

Royle and Nichols Heterogeneity Abundance Model

I used the Royle and Nichols heterogeneity abundance model (RN) to obtain estimates of ungulate abundance and detection probabilities under imperfect detection (Royle and Nichols,

2003). The RN model assumes that heterogeneity in detection probability can be used as a surrogate to model heterogeneity in animal abundance and thus estimate the underlying distribution of abundances (Royle and Nichols, 2003). Abundance in these models is treated as a latent parameter, and integrated out of the following probability expression:

N p = 1-(1-r) i

Where: p = probability of observing 1 or more animals at a plot (i.e., the plot is occupied) r = detection probability for individuals

N = abundance of species at plots, also denoted as 

The distribution of N can then be modeled with a Poisson or negative binomial distribution model (Royle and Nichols, 2003). Abundance in this case is not observed directly but is inferred via detection histories made up of presence/absence data collected at study sites for the species of interest. This model also incorporates estimates of detection probabilities to account for species specific imperfect detection that may introduce sampling biases (MacKenzie et al., 2006). Incorporating detection probabilities is essential because differences in abundance may be due to sampling strategy rather than to the effects of covariates I may be testing for

(Mackenzie et al., 2006).

The RN model has a number of assumptions that must be met for abundance estimates to be valid. These assumptions are:

1) Animal detections must be independent of each other.

2) Abundance (N or ) is assumed constant across surveys of a plot (i.e., we are sampling a closed population).

3) Individual detection probability (r) is assumed constant across time.

I consider the 5-day sampling period a short enough period to meet the assumptions of a closed population. This should also ensure that conditions at each plot did not change enough to alter abundance estimates or detection probabilities throughout the sampling period (Royle and

Nichols, 2003). Additionally, we carried out the surveys across fenced reserves located within a landscape matrix that limited immigration and emigration. Our study species were mostly territorial and occurring at low densities (Skinner and Chimimba, 2005). Patterns of occurrence across the landscape for these types of species makes them amenable for presence/absence occupancy modelling, which forms the basis for the RN model (Royle and Dorazio, 2008). I also tested for and found no evidence of spatial auto-correlation in residuals derived from occupancy estimates for the species and surveys analysed (Table 3-1). Accordingly, I was confident that our study met the assumptions of the RN model.

Comparison of Abundance Estimates Using Different Modelling Frameworks

The RN model can be implemented in programs PRESENCE (Hines 2006a), MARK

(White and Burnham, 1999), and Unmarked, a statistical package in program R (Fiske and

Chandler, 2011). These programs model heterogeneity among sites in detection using finite mixture models that also fit ecological or state process and observation or detection process models separately under a maximum likelihood or frequentist framework (Royle and Dorazio,

2008; Fiske and Chandler, 2011). MARK allows the user to model abundance according to either a Poisson or negative binomial distribution (MARK user manual). This is a key advantage as the negative binomial distribution is a better descriptor of biological processes in some cases

(Bolker, 2008). More recently, the RN model was incorporated into multi-species hierarchical occupancy models (Msp) that model species as random effects taken from a higher hierarchical level or community-level distribution (Yamaura et al., 2011; Beesley et al., 2014). Msp models are analyzed simultaneously using software JAGS (Plummer, 2003) and R (R Development Core

Team, 2011).

I compared abundance estimates derived from finite-mixture models (FM) to those generated from the hierarchical Bayesian multi-species abundance model (Msp). I compared ungulate abundance estimates for a grazer (warthog), a browser (bushbuck), and a mixed feeder

(nyala) from surveys across seasons and years. These are all medium-sized ungulates that had adequate and consistent naive occupancy estimates across all surveys (naive occupancy estimate

> 20%; Table 4-1).

I generated abundance and detection probability estimates for the four surveys (one per season per year) using an FM and a Msp model. I used covariates in both models that were found to influence abundance and detection probabilities during a previous analysis that used the Msp model for all three species (Chapter 2). I evaluated models with covariates such as grass, shrub, tree cover and water availability that had the potential to influence species use and detection at the plots. I obtained averaged estimates from the top models based on AIC, AIC and AIC weight values following Burnham and Anderson (2002) (Appendix D, Table D-1).

For the Msp model, I estimated the following posterior probability parameters with this model for each species: species-specific detection probability (pi), and an abundance estimate per

plot (λi). I estimated posterior means, standard deviations and 95% confidence intervals for all parameters of interest (model detail in Appendix B).

The Msp model describes the latent variable Nij (species abundance) with a Poisson distribution as:

Nij ~Poisson (ij),

Where ij is the Poisson mean for species i at plot j. Therefore, I assumed that species abundance could vary in time and space depending on covariates in the form:

Nij ~Poisson (ij) with log(ij) = 1i + xj 2i

Where 1i and 2i are the parameters to be estimated for species i and 1i (intercept) is a random species effect. Thus, our final abundance model was the following:

log(ij) = 1i + 2i shrubj + 3i treej + 4i grassj + 5idwatj

Similarly, the detection probability (pij) model was:

log (pij) = 1i + 2i shrubj + 3i canopyj + 4i grassj + 5ivoj

Rubin diagnostic where an estimate >1.1 indicates low convergence rates between chains

(Brooks and Gelman 1998).

I compared precision of abundance estimates for each model and each species using 95% confidence intervals credible intervals (for the FM and Msp models respectively) for both models. Smaller intervals indicated higher precision. I also compared estimates from each model across years and seasons and determined if their 95% confidence intervals overlapped. I expected

abundance estimates from both models would be comparable, showing similar trends across seasons and years surveyed and overlapping within their 95% confidence and credible intervals.

Evaluating Sampling Effort Schemes Via Data Simulation

To determine ideal number of sampling plots for species with different detection probabilities, I simulated plot-level data for the RN model using program GENPRES (Hines,

2006b; Bailey et al., 2007). Program GENPRES generates detection histories under the assumed model (i.e., the RN model follows a Poisson distribution [Royle and Nichols; 2003]) for the parameters imposed by the user and then analyses the simulated data via program MARK to obtain estimates of abundance and standard errors (Bailey et al., 2007). The output consists of summaries and descriptive statistics of abundance estimates per simulation obtained from

MARK.

I used a combination of number of sites surveyed (n = 30, 50, 80, and 100) and individual detection probabilities (p = 0.1, 0.2, and 0.3) to evaluate sampling effort for species with differing detection probabilities. I based the proposed estimates of detection probabilities on a study of large ungulates (i.e., bushbuck, nyala, impala [Aepyceros melampus], greater kudu

[Tragelaphus strepsiceros], blue wildebeest [Connochaetes taurinus], common duiker

[Sylvicapra grimmia], waterbuck [Kobus ellipsiprymnus], giraffe [Giraffa Camelopardalis], and zebra [Equus quagga]) carried out in the same study sites that found probabilities to range between 0.22 and 0.57 (Collier et al., 2011). However, I used more conservative estimates to focus on more cryptic species that are harder to detect. I also based our estimates of detection probability on those derived from the previous analyses using program MARK, where detection probabilities ranged between 0.07 and 0.27. I then set  to 2 to simulate low abundance species typical of the area based on results from a previously run Msp model (Chapter 2). I used 5

sampling occasions per survey and ran 1,000 simulations for each combination of detection probability and number of sites surveyed.

I also evaluated the influences of the number of survey occasions on abundance estimate bias, precision and accuracy. I simulated three different sampling schemes consisting of different survey occasions (n= 5, 10, 15) typical of camera trapping studies (Burton et al., 2015). I set  to

2, p to 0.1 to mimic low detection rate species typical of the study site (Chapter 2), and number of sites at 50 to consider an intermediate and realistic survey effort.

I compared bias of the estimates under different sampling schemes using the percent deviation from the true  estimate for all simulations. I set an estimate of =2 for all simulations.

Unbiased estimates were those that approached or were no more than ±30% different than this pre-defined parameter estimate. I also tested precision and accuracy of simulated estimates.

Precision refers to how close to each other the repeated measurements were and accuracy combines both bias and precision (Zar, 1999; Williams et al., 2002). I compared precision between sampling schemes using the coefficient of variation ([CV]; Bolker, 2008). The coefficient of variation is defined as the ratio of the standard deviation to the mean (Gotelli and

Ellison, 2004). Higher CV estimates indicated less precise estimates. Finally, I used the root mean squared error (RMSE) as a measure of accuracy of varying number of sampling sites

(Williams et al., 2002). Smaller RMSE values point to more accurate (less biased and more precise) estimates. The RMSE is estimated as:

1 푅푀푆퐸 = √ ∑푛 (퐴 − 퐴 )2 푛−1 푖=1 푠 휏 Where: n = number of samples

As= Abundance estimate obtained from simulations

퐴휏 = true abundance

Results

We conducted a total of four surveys, two per season per year surveyed for a total of

1980 trap nights. Number of trap nights varied between surveys. We trapped 630 nights in June,

2012; 270 nights in November, 2012; and 540 nights for both the June and November, 2013 surveys. We detected a total of 11 ungulate species throughout the four surveys (Table 4-1).

Naive occupancy estimates ranged from 0 (giraffe and waterbuck) to 0.65 (bushbuck; Table 4-1).

Comparison of Abundance Estimates Using Different Modelling Frameworks

Most of the FM models provided biologically reasonable estimates (i.e., based on previous studies and what would be expected for each species). Most models converged satisfactorily except for the models run for the November, 2012 surveys for bushbuck and nyala, which provided unreasonable estimates (i.e.,  > 2,000). All the Msp model parameters converged satisfactorily (Gelman-Rubin statistic ≤ 1.1: model median range for all models =

1.01 – 1.06) (Brooks and Gelman 1998). Abundance estimates per plot for bushbuck ranged from 0.57 to 2.39 and 0.71 to 1.04 for the FM and the Msp models respectively (Table 4-2).

Estimates for Nyala ranged from 0.62 to 1.34 and 0.66 to 2.12 (for FM and Msp models respectively). While for warthogs, estimates ranged from 0.42 to 0.98 for the FM models and

0.43 to 1.46 for the Msp models.

Across the three ungulate species examined, all surveys yielded similar abundance estimates for the two models (Figure 4-1). However, estimates for the November surveys (i.e., rainy season [Figure 4-1]) were generally less precise for all three species. Estimates for bushbuck from the two models showed a varying trend across the four surveys, and no clear seasonal or yearly pattern could be discerned, except for a potential increase in abundance during the last survey (Figure 4-1). Nyala estimates were also similar across the two models and, unlike the bushbuck, a potential pattern of increased abundance estimates during the rainy season was

observed for both years surveyed (Figure 4-1). Finally, Abundance estimates for the warthog were more similar between models and seasons, with the exception of the November, 2012 survey. This last survey yielded the highest and least precise estimates (Figure 4-1). Overall, abundance remained constant across years and surveys for this species.

Evaluating Sampling Effort Schemes via Data Simulation

Evaluating simulations across different detection probabilities (p) and number of sites surveyed, I found bias (deviation from true abundance) decreased two-fold with each increase in number of sites for simulations with p = 0.1 (Figure 4-2; Table 4-3). Simulations with p = 0.1 also yielded the highest values of bias (186% and 321%) and the most variation in bias across number of sites simulated (from 38% to 321%). Bias of sampling schemes for animals with p 

0.2 and number of study sites greater than 30 did exceed 12% and were consistently less biased than schemes with p = 0.1 (Figure 4-2; Table 4-3). I also found that increasing detection probability reduced bias to less than 50% even for schemes with low number of sites.

Interestingly, sampling schemes with intermediate detection probability (p = 0.2) and low number of sites were the least biased (range from 15.5% to 39.5%). However, simulations for p =

0.3 were more consistent across number of sites surveyed (range from 48.5% to 52.5%).

Precision, as measured by the coefficient of variation, remained relatively constant and did not decrease as markedly when number of sites were increased and detection probability was

0.1 (Figure 4-5; Table 4-3). The most notable decreases in precision were for increased number of sites (>50) and p ≥ 0.3 and 100 sites with p = 0.2. In this case, precision decreased by more than 100%.

The most accurate estimates (evaluated with RMSE of estimates) were those with p = 0.3 and number of sites >50 (Table 4-3). Across all values of p evaluated, accuracy increased ≥ 50%

with number of sites sampled. The most notable increase was found for simulations with p = 0.2 and number of sites sampled increased from 30 (RMSE = 6.77) to 50 (RMSE = 2.97).

When testing for the effects of increasing number of sampling occasions, the most notable decrease in bias occurred when p = 0.1 and number of sampling occasions was increased

(Figure 4-3; Table 4-4). In this case bias decreased by 5 and 8 times when occasions were increased to 10 and 15 respectively. Surprisingly, for p = 0.2, bias decreased by half when number of sampling occasions was increased from 5 to 10 occasions. Bias for p = 0.3 only decreased by 6% and 8% percent with higher number of sampling occasions.

Simulations of surveys with 15 and 10 sampling occasions and p = 0.3 provided the most precise estimates of abundance (Figure 4-3; Table 4-4). Estimates for these conditions were over

100% more precise as all other estimates. Increasing number of sampling occasions at p = 0.1 had a more moderate effect, increasing precision by only 10% when number of sampling occasions were increased to 10 and 15. Surprisingly, precision increased at p = 0.2 when sampling occasions were increased from 5 to 10, but decreased again when occasions were increased to 15.

Finally, the least accurate estimates were found for p = 0.1 and 5 sampling occasions

(Figure 4-3; Table 4-4). Accuracy improved by more than half at p = 0.1, when sampling occasions were ≥ 10. Accuracy was highest for p = 0.2 and sampling occasions ≥ 10 and did not improve with increased p.

Discussion

I demonstrate that camera trapping is a viable option for estimating wildlife abundance for species without markings and in areas that may not be suitable for direct observation if certain guidelines are followed. Abundance estimates from our study were comparable to previous estimates for bushbucks in South Africa and Western Africa (Coates and Downs, 2007;

Marchal, et al., 2012), nyala in Zimbabwe (Dunham, 2012), and warthogs in Western Africa

(Marchal et al., 2012) generated from direct sighting methods (e.g., car and aerial transects) look for sites of studies. Nonetheless, some of these estimates were based on low numbers of sightings for species associated with denser cover such as bushbuck (Coates and Downs, 2007;

Klop and van Goethem, 2008). Camera trapping may address these problems because cameras typically survey an area more intensively and throughout longer periods and thus tend to increase detections for rare or elusive species, especially those associated to dense forest.

To compensate for low sightings of forest species, other studies have generated relative abundance indices from encounter rates with signs (Mugume, et al., 2015). Additionally, some camera trapping studies have also opted for using encounter rates (e.g., number of photographs/total trap nights) as indices of relative abundance (O;Brian, 2010). However, these encounter rates have been criticized on the grounds that their bias cannot be estimated (Guillera-

Arroita et al., 2014) and they do not account for imperfect detection (i.e., false absences), resulting in underestimates of animal occupancy and/or abundance (MacKenzie et al., 2006;

O’Brian, 2010). Using RN models, I was able to estimate bias and account for imperfect detection while obtaining reasonable estimates of abundance.

I also provided an evaluation of the modelling frameworks and sampling schemes that can be used to generate abundance estimates from camera traps in areas of thick cover. When comparing RN models under different modeling frameworks, I found that the more user-friendly

software options for the FM model generates similar estimates to the more complex Msp analyses. However, a drawback of the FM model is that it failed to converge during the

December 2012 surveys for bushbuck and for nyala. During the December 2012 surveys, we used only one camera trap per plot. The combination of less camera traps and overall less detectability during this period may have reduced the number of detections, thereby increasing number of capture histories or sites with zeros. Poisson distributions modeled with FM models, perform poorly with a large number of zeros (Denes et al., 2015). This may also be due to the inability to set a maximum and minimum value for parameter estimates for the FM models in

MARK, which results in unreasonable estimates for sparse datasets. When this occurs, it is likely that the optimization method used is “stuck” at a peak for the maxima or uses unreasonable starting values (Bolker, 2008). Program Unmarked provides the user with the option to set a limit on the parameter estimates (Fiske and Chandler, 2011) and thus may be preferable for cases when only sparse data with many zeros is available.

Both frameworks compared found a general trend of lower, but more precise and similar estimates in the dry season for all three species. This may be a reflection of species reducing their home ranges as a response to scarce resources (Keynhans et al., 2011) and thus being more available for detection during the drier periods. As a result, running camera trapping surveys to estimate abundance in the dry season may provide more conservative and precise estimates.

Regardless of the modelling framework used, the choice of sampling scheme can improve utility of camera trapping to obtain abundance estimates or examine abundance patterns through time and/or space (i.e., monitoring). Overall, my simulations show that ideal number of sites is ≥50, and sampling occasions ≥ 10 per survey, and that bias and precision are improved if p is increased. I also showed that precision and bias were strongly associated to detection

probability. Our results coincide with Denes et al. (2015) that found bias of the RN model to be relatively low unless detection probabilities are low or a small number of sites are sampled. They suggested increasing number of survey occasions (>5) to reduce bias from small sample sizes.

Precision of schemes with sites ≥ 50 and p = 0.3 yielded the most precise estimates, although sampling schemes with 100 sites and p = 0.2, were relatively similar and may be ideal when other conditions are difficult to meet (i.e., p = 0.3). Similarly, our results indicate that increasing sampling occasions with the highest p (0.3) could provide the most precise and least biased estimates of abundance. However, increasing number of sampling occasions seemed to have a stronger positive influence on abundance estimates for species with low p. When examining occupancy probabilities for mid- and large-sized mammals with camera traps, sampling periods of up to 16 days are common (Thornton et al., 2011; Silva-Rodriguez and

Sieving, 2012); however, it is important to consider that sampling for longer periods could lead to violation of the population closure assumption (Harmsen et al., 2010) and/or place further logistical or cost constraints on the study.

I found that the precision of abundance estimates improved as p increased, but from previous research conducted in our study site, I know that most ungulate species have low detection probabilities with camera traps (p < 0.2) (Chapter 2; Collier et al., 2011). Precision increased by > 100% when detection probability increased from 0.1 to 0.3 and > 50 sites were surveyed.

Our results coincide with recent work suggesting bias can be reduced for species with p 

0.3 (Derbes et al., 2015). In practice, camera trapping studies can be limited by logistical and funding constraints that preclude increasing survey length and/or number of sites sampled. When these conditions (sites > 50 or number of occasions > 10) cannot be met, I recommend evaluating

strategies to increase detection probabilities, as I found that increasing detection probabilities could reduce bias even for low number of sites. Strategies for increasing detection probabilities may include placing attractants in front of camera traps, increasing the number of cameras per site surveyed (Royle and Dorazio, 2008; Foster and Harmsen, 2012) or testing for locations that increase chances of detecting a species, as is done with spotted carnivores such as the jaguar

(Silver et al., 2004; Foster and Harmsen, 2012). Harmsen et al. (2010) recommend conducting longer surveys and pooling data across survey occasions to increase capture rates for capture/recapture studies. However, most of these strategies were recommended for capture/recapture studies for large carnivores (Harmsen et al., 2010; Foster and Harmsen, 2012) and need to be evaluated for their applicability in presence/absence surveys for ungulates in distinct environmental conditions.

Inevitably, researchers conducting wildlife surveys will have to make decisions regarding trade-offs relating to number of trapping occasions, sampling sites, and desired detection probabilities. For example, maximizing number of sites sampled is not always feasible or even the most effective or advantageous sampling scheme or design (Bailey et al., 2007; Foster and

Harmsen, 2012). As our study shows, other alternatives are available to obtain more reliable estimates. More specifically, in light of high costs and effort of setting up more sampling plots, our results show that sampling schemes should aim to increase survey length and occasions and find strategies to also increase detection probabilities.

In conclusion, camera trapping can be a viable option to estimate the abundance of

African ungulates and monitor their population trends in forests and woody encroached savannas when certain guidelines regarding number of sites and survey occasions are followed. Our analyses of modeling frameworks and sampling scheme simulations suggest that researchers and

managers should sample over longer survey periods (a minimum of 10 days per site) and across more sampling sites (a minimum of 50) to improve bias and precision of estimates. Strategies to increase detection probabilities should also be evaluated, as this may be a more practical alternative to increasing survey length and number of sampling sites. Our analyses of surveys under different analytical methods show that dry season surveys increased precision and produced more similar results between the three methods. However, I do not recommend using the FM model with sparse data and low number of detections, which should be expected for some species and under conditions typical of our study system.

Table 4-1. Naive occupancy estimates of large- and mid-sized mammals from camera-trapping surveys in Mlawula Nature and Mbuluzi Game Reserves, Swaziland, 2012-2013. Naïve occupancy estimates are defined as number of plots detected/total number of plots surveyed. *denotes no detections during survey.

Species Common June 2012 Nov. 2012 June 2013 Nov. 2013 name Tragelaphus scriptus Bushbuck 0.43 0.22 0.41 0.65 Potamochoerus larvatus Bushpig 0.11 0.02 0.04 0.06 Sylvicapra grimmia Duiker 0.44 0.02 0.11 0.04 Giraffa camelopardalis Giraffe 0.03 0.00* 0.09 0.00* Aepyceros melampus Impala 0.51 0.53 0.04 0.43 Tragelaphus strepsiceros Kudu 0.33 0.15 0.20 0.07 Tragelaphus angasii Nyala 0.36 0.46 0.35 0.41 Phacochoerus africanus Warthog 0.28 0.39 0.24 0.26 Kobus ellipsiprymnus Waterbuck 0.02 0.00* 0.09 0.07 Connochaetes taurinus Wildebeest 0.11 0.20 0.15 0.17 Equus quagga Zebra 0.11 0.07 0.07 0.04

Table 4-2. Comparison of abundance estimates (λ) and standard errors (SE) from two modeling frameworks for camera trapping data for ungulates in Swaziland (2012-2013).

Finite-mixture Model Multi-species Model Species λ (SE) λ (SE) λ (SE) λ (SE) λ (SE) λ (SE) λ (SE) λ (SE) Bushbuck 1.73 0.00 0.57 2.39 1.04 0.84 0.71 0.71 (1.21) (0.00) (0.14) (1.14) (0.01) (0.01) (0.004) (0.04) Nyala 0.62 0.00 0.74 1.34 0.68 2.12 0.66 1.3 (0.38) (0.00) (0.2) (0.45) (0.003) (0.02) (0.003) (0.01) Warthog 0.42 0.98 0.51 0.50 0.43 1.46 0.53 0.91 (0.14) (0.43) (0.41) (0.21) (0.002) (0.02) (0.003) (0.01) Abundance estimates were obtained from Royle and Nichols heterogeneity abundance model (RN) run using a finite-mixture and a hierarchical Bayesian multi-species abundance model using R and JAGS. The columns for each model represent the following surveys respectively: 1) June, 2012; 2) November, 2012; 3) June, 2013; and 4) November, 2013.

Table 4-3. Comparison of bias, precision, and accuracy of abundance () estimates from simulations using program GENPRES evaluating number of sites surveyed. Bias = % deviation from true estimate, precision = coefficient of variation (CV), and accuracy = root mean squared error (RMSE). p and number of  SE % Deviation CV RMSE sites per survey p = 0.1 100 2.76 0.18 38 2.04 6.27 80 3.81 0.27 91 2.04 9.22 50 5.72 0.37 186 2.21 13.04 30 8.42 0.48 321 2.74 17.27 p = 0.2 100 1.21 0.01 39.5 0.22 1.23 80 1.29 0.05 35.5 1.16 1.97 50 1.45 0.08 27.5 1.70 2.97 30 2.31 0.2 15.5 1.79 6.77 p = 0.3 100 0.95 0.003 52.5 0.09 0.96 80 0.96 0.003 52 0.11 0.96 50 0.97 0.005 51.5 0.15 0.98 30 1.02 0.04 48.5 1.38 1.7 I generated simulations in program GENPRES under different sampling schemes (number of sites surveyed) and detection probabilities (p) and analyzed with the Royle and Nichols heterogeneity abundance model in program MARK. Presented are estimates of abundance (λ) at each site surveyed, standard error (SE), percent deviation from estimate, coefficient of variation (CV), and root mean squared error (RMSE) summarized from 1,000 simulations for each sampling scheme. Abundance (λ) was set at 2 individuals per site for all simulations.

Table 4-4. Comparison of bias, precision, and accuracy of abundance () estimates from simulations using program GENPRES evaluating number of sampling occasions surveyed. Bias = % deviation from true estimate, precision = coefficient of variation (CV), and accuracy = root mean squared error (RMSE). p and number of  SE % Deviation CV RMSE sites per survey p = 0.1 5 5.72 0.37 186 2.30 12.29 10 2.73 0.18 36.7 2.10 5.8 15 1.54 0.01 23.13 2.04 0.64 p = 0.2 5 1.45 0.08 27.5 1.79 2.97 10 1.14 0.01 57.2 2.12 0.89 15 1.01 0.01 50.4 1.11 0.99 p = 0.3 5 0.97 0.005 51.5 1.50 0.98 10 0.92 0.002 45 0.08 1.08 15 0.87 0.002 43.73 0.06 1.13 I generated simulations in program GENPRES under different sampling scheme (number of occasions per survey) with true abundance = 2 detection probability = 0.1, number of sites sampled = 50. Simulated data was analyzed using the Royle and Nichols heterogeneity abundance model in program MARK. Presented are estimates of mean abundance (λ) at each site surveyed, standard error (SE), percent deviation from estimate, coefficient of variation (CV), and root mean squared error (RMSE) summarized from 1,000 simulations for each sampling scheme. Abundance (λ) was set at 2 individuals per site and detection probability (p) at 0.1 for all simulations.

Figure 4-1. Comparison of estimates of average abundance per plots and their 95% confidence intervals for bushbucks (A), nyalas (B), and warthogs (C). Estimates obtained in Swaziland across four different surveys (2012-2013). Estimates were obtained from RN heterogeneity models run under different modeling frameworks. FM refers to finite-mixture models and MSP to hierarchical Bayesian multi-species models.

Figure 4-2. Comparison of bias, precision and accuracy of abundance estimates for different combinations of detection probabilities (p) and number of sites surveyed. Estimates obtained from simulations run for the Royle and Nichols heterogeneity abundance model. Bias is measured as percent variation from true estimate, precision as the coefficient of variation (CV), and accuracy as the root mean squared error (RMSE) summarized from simulations (n = 1,000) run in program GENPRES and analyzed in program MARK using a true abundance estimate of  = 2.

Figure 4-3. Comparison of bias, precision and accuracy of abundance estimates for different combinations of detection probabilities (p) and number of sampling occasions. Estimates obtained from simulations run for the Royle and Nichols heterogeneity abundance model. Bias is measured as percent variation from true estimate, precision as the coefficient of variation (CV), and accuracy as the root mean squared error (RMSE) summarized from simulations (n = 1,000) run in program GENPRES and analyzed in program MARK using a true abundance estimate of  = 2.

CHAPTER 5 CONCLUSIONS

My results indicate that woody encroached plots are used by a large number of mammalian species and may provide benefits to some species of the study area. However, low relative abundance estimates suggest a vulnerability to woody encroachment in the area for some species. Most importantly, my study shows that sites with adequate grass cover are needed to maximize species richness and abundance of savanna associated species. My research suggests that instead of direct impacts, increased woody cover affects species richness and abundance indirectly through the loss of grass cover.

My study suggests that woody cover > 30% may reduce resources and/or available space.

My results showed lower species richness and overall abundances across plots with >30% shrub cover. In Chapter 3, I show that this threshold of woody cover is increasing the spatial overlap of ungulate species. This could ultimately lead to competitive interactions where one species is negatively impacted. In total, my study provides evidence for negative impacts of woody encroachment on large and medium mammals as has been found in other studies (Blaum et al.,

2007d; Sirami et al., 2009).

Managers should seek to recover grass cover across conservation areas. Woody plant cover and biomass can be reduced through mechanical means, herbicides and fire (Gibson,

2009). Areas where woody plant removal or restoration efforts are planned should be selected based on site history, soil attributes, and local acceptance and participation (Gibson, 2009).

However, most removal and restoration methods can be time-consuming and expensive, so alternative approaches should be evaluated. For example, decreasing canopy cover such that light can penetrate to the ground and promote grass cover growth should be tested. These management strategies may help reduce the probability of local extirpation of species sensitive to reduced

grass cover (i.e., wildebeest, waterbuck, zebra) and maximize overall species richness. These interventions are necessary because wildlife species that co-evolved with C4 grasses in savannas make this system particularly unique, and contribute to the maintenance of its structure and composition (Du Toit and Cumming, 1999; Augustine and McNaughton, 2004; Goheen et al.,

2010).

Finally, my study provides recommendations to use camera trapping and the Royle-

Nichols heterogeneity model as an effective tool to monitor wildlife populations in woody encroached savannas. By adjusting future surveys based on the results of my study, managers and researchers can implement long term studies to monitor population trends and responses to habitat changes and management interventions.

APPENDIX A COVER SCALE USED TO MEASURE HABITAT COVARIATES

Table A-1. Cover scale adapted from Daubenmire (1959) used to measure percent cover of different vegetation types as habitat covariates in Mlawula Nature and Mbuluzi Game Reserves, Swaziland, 2012-2013.

Cover-Class Scale Cover percentage

1 one or a few individuals 2 0-4%

3 5-10%

4 11-20%

5 21-35%

6 36-50%

7 51-60%

8 61-75% 9 76-90%

10 91-100%

APPENDIX B MULTI-SPECIES HIERARCHICAL ABUNDANCE MODEL CODE DESCRIPTION

# Priors betmu[i] ~ dnorm(0,.1) #Abundance bettau[i] ~ dgamma(.1,.1) #Precision parameter for abundance phimu[i] ~ dt(0,0.4076321,7.63179) #Detection probability phitau[i] ~ dgamma(.1,.1) #Precision parameter for detection probability bet1[i] ~ dnorm(betmu[1],bettau[1]) #Covariates on abundance bet2[i] ~ dnorm(betmu[2],bettau[2]) #Covariates on detection probability omega ~ dunif(0,1) #Data augmentation parameter w[i] ~ dbern(.5) #Model inclusion parameter

for(i in 1:7){ betmu[i] ~ dnorm(0,.1) bettau[i] ~ dgamma(.1,.1) } for(i in 1:8){ phimu[i] ~ dt(0,0.4076321,7.63179) phitau[i] ~ dgamma(.1,.1) }

for(i in 1:nsp){ bet1[i] ~ dnorm(betmu[1],bettau[1]) bet2[i] ~ dnorm(betmu[2],bettau[2]) etc. phi1[i] ~ dnorm(phimu[1],phitau[1]) phi2[i] ~ dnorm(phimu[2],phitau[2]) etc. } omega ~ dunif(0,1) for(i in 1:13){ w[i] ~dbern(.5) }

for(i in 1:nsp){ o[i] ~ dbern(omega) for(k in 1:nplot){ N[i,k]~ dpois(lam[i,k]) #Species-specific abundance per plot lam[i,k]<-exp(log.lam[i,k])*o[i] log.lam[i,k] <- bet1[i] + w[1]*bet2[i]*shrub[k] + w[2]*bet3[i]*tree[k] + w[3]*bet4[i]*grass[k] + w[4]*bet5[i]*season[k] + w[5]*bet6[i]*year[k] + w[6]*bet7[i]*dwat[k]

ymat[i,k]~dbin(p[i,k],5) p[i,k] <- 1-(1-r[i,k])^N[i,k] r[i,k] <- exp(logit.r [i,k])/(1+exp(logit.r[i,k])) logit.r[i,k]<-phi1[i] + w[7]*phi2[i]*shrub[k] + w[8]*phi3[i]*tree[k] + w[9]*phi4[i]*grass[k] + w[10]*phi5[i]*visobs[k] + w[11]*phi6[i]*camera[k] + w[12]*phi7[i]*season[k] + w[13]*phi8[i]*year[k] }}

R <- sum(o[]) #Sum of number of species across all plots for(i in 1:nplot){ rplot[i] <- sum(N[,i]>0) #Plot-specific estimate of species richness abund[k] <- sum(N[,k]) }

APPENDIX C LIFE-HISTORY TRAITS OF TERRESTRIAL MAMMALS DETECTED IN STUDY

Table C-1. Life-history traits of terrestrial mammals detected in camera trapping surveys in Mlawula Nature and Mbuluzi Game Reserves, Swaziland, 2012-2013. Mean Body Home range Geographic Dietary Trophic Feeding guild Digestive strategy Groups Species mass (kg) (km2) Range (km2) breadth level (ungulates) (ungulates) Ungulates Bushbuck 43 0.02 43,250 4 1 browser ruminant Common Duiker 15 0.21 15,639 3 1 browser ruminant Greater Kudu 206 835 206,056 3 1 browser ruminant Giraffe 964 82 964,655 4 1 selective browser ruminant Warthog 82 1.55 82,500 4 1 grazer non-ruminant Waterbuck 204 1.28 204,393 3 1 grazer ruminant Common Wildebeest 199 2.5 198,620 1 1 grazer ruminant Zebra 400 164 400,000 2 1 grazer non-ruminant Impala 53 1.97 52,592 2 1 mixed feeder ruminant Nyala 88 1.63 87,617 4 1 mixed feeder ruminant Bushpig 69 4.47 69,064 4 2 omnivorous non-ruminant Carnivores Rusty-spotted Genet 2 11 1,756 2 2 NA NA Spotted Hyena 63 22 63,370 1 3 NA NA Side-striped Jackal 10 1 10,581,413 4 2 NA NA Serval 12 2 12,526,197 2 3 NA NA Slender Mongoose 0.54 0.66 544 3 2 NA NA Water Mongoose 4 2 3,600 2 2 NA NA Primates Vervet Monkey 8 2 4,577,375 6 1 NA NA Baboon 18 9 17,729 7 1 NA NA Other Aardvark 56 4 18,035,652 2 3 NA NA Taxa Porcupine 15 1.3 14,936 3 1 NA NA Life history traits obtained from Skinner and Chimimba (2005) and Jones et. al. (2009). Dietary breadth is defined as number of categories consumed by species. Categories were: vertebrates, invertebrates, fruit, flowers/nectar/pollen, leaves/branches/bark, seeds, grass and roots/tubers. Trophic levels were defined as (1) herbivore (not vertebrate and/or invertebrate), (2) omnivore (vertebrate and/or invertebrate plus any of the other categories) and (3) carnivore (vertebrate and/or invertebrate only).

100

APPENDIX D TOP RANKED SINGLE-SPECIES OCCURRENCE MODELS FOR UNGULATES

Table D-1. Top ranked single-species occurrence models for ungulates detected in camera trapping surveys in Mlawula Nature and Mbuluzi Game Reserves, Swaziland, 2012-2013.

Top ranked robust occurrence models for ungulates in Swaziland (2012-2013). Models were run in program MARK to determine covariates included in best fit models to describe ungulate occurrence () and detection probabilities (p). Models were run for each species for both low and high shrub cover plots.Where: AICC (Akaike’s Information Criterion corrected for small sample size), K= number of parameters, ΔAICC = the difference of AICC values for the most supported model and the given model, and wi=AICC weight for each model. Covariates: grass=grass cover percentage, can= canopy cover percentage, dwat= distance to water, het= structural heterogeneity, and vo=visual obstruction.

101

APPENDIX E CO-OCCURRENCE ESTIMATES FROM ROBUST TWO-SPECIES CONDITIONAL MODEL FOR UNGULATE SPECIES

Table E-1. Probability of occurrence estimated from robust two-species conditional model for ungulate species detected in camera trapping surveys in Mlawula Nature and Mbuluzi Game Reserves, Swaziland, 2012-2013. Included are standard errors (SE) and upper and lower 95% confidence intervals (CI) for pair-wise comparisons where: A = estimate for first species, B = estimate for second species, BA = estimate of species B given species A is present, and Ba = estimate of species B given A is absent. Species A is the first species mentioned in column 1. Low shrub cover percentage High shrub cover percentage Species Parameter  SE Lower CI Upper CI  SE Lower CI Upper CI Bushbuck/Duiker A  0.59 0.11 0.37 0.77 0.59 0.10 0.39 0.76 B  0.73 0.11 0.49 0.88 0.55 0.11 0.34 0.75 BA  0.95 0.09 0.34 0.99 0.74 0.16 0.37 0.93 Ba 0.43 0.17 0.16 0.74 0.28 0.17 0.08 0.66 Impala/Bushbuck A  0.69 0.09 0.49 0.84 0.48 0.09 0.31 0.66 B  0.62 0.09 0.43 0.77 0.56 0.09 0.38 0.72 BA  0.67 0.11 0.44 0.85 0.73 0.12 0.45 0.90 Ba 0.49 0.19 0.18 0.81 0.40 0.13 0.18 0.66 Nyala/Impala A  0.50 0.14 0.24 0.76 0.42 0.09 0.26 0.59 B  0.52 0.09 0.34 0.70 0.64 0.12 0.39 0.83 BA  0.58 0.18 0.24 0.86 0.72 0.14 0.39 0.91 Ba 0.47 0.19 0.17 0.80 0.59 0.19 0.24 0.87 Impala/Duiker A  0.76 0.11 0.49 0.92 0.46 0.10 0.28 0.65 B  0.54 0.13 0.30 0.76 0.57 0.12 0.33 0.77 BA  0.40 0.12 0.20 0.64 0.49 0.15 0.23 0.76 Ba 1.00 0.00 1.00 1.00 0.63 0.19 0.25 0.89 Kudu/Duiker A  0.95 0.05 0.70 0.99 0.44 0.11 0.25 0.66 B  0.42 0.10 0.25 0.62 0.77 0.09 0.56 0.90 BA  0.39 0.10 0.22 0.60 0.48 0.16 0.22 0.76 Ba 0.90 0.20 0.60 1.00 0.80 0.20 0.30 0.96 Kudu/Impala A  0.77 0.10 0.53 0.91 0.47 0.15 0.22 0.73 B  0.68 0.11 0.45 0.85 0.65 0.15 0.33 0.87 BA  0.59 0.13 0.34 0.80 0.59 0.21 0.22 0.88 Ba 1.00 0.00 0.00 1.00 0.70 0.15 0.36 0.91

102

Low shrub cover percentage High shrub cover percentage Species Parameter  SE Lower CI Upper CI  SE Lower CI Upper CI Kudu/Nyala A  0.54 0.18 0.23 0.83 0.62 0.14 0.33 0.84 B  0.48 0.13 0.26 0.72 0.42 0.09 0.16 0.82 BA  0.55 0.19 0.22 0.85 0.38 0.13 0.18 0.65 Ba 0.40 0.26 0.07 0.85 0.49 0.20 0.16 0.82 Kudu/Bushbuck A  0.67 0.15 0.35 0.88 0.41 0.10 0.23 0.62 B  0.76 0.14 0.41 0.94 0.61 0.09 0.42 0.77 BA  0.66 0.16 0.32 0.89 0.79 0.16 0.37 0.96 Ba 1.00 0.00 1.00 1.00 0.49 0.13 0.25 0.73 Nyala/Bushbuck A  0.42 0.08 0.27 0.58 0.43 0.09 0.27 0.61 B  0.58 0.10 0.38 0.75 0.57 0.10 0.38 0.74 BA  0.80 0.12 0.47 0.95 0.89 0.13 0.36 0.99 Ba 0.42 0.14 0.19 0.69 0.32 0.12 0.13 0.59 Nyala/Bushbuck A  0.42 0.08 0.27 0.58 0.43 0.09 0.27 0.61 B  0.58 0.10 0.38 0.75 0.57 0.10 0.38 0.74 BA  0.80 0.12 0.47 0.95 0.89 0.13 0.36 0.99 Ba 0.42 0.14 0.19 0.69 0.32 0.12 0.13 0.59 Nyala/Duiker A  0.44 0.09 0.28 0.61 0.43 0.09 0.27 0.61 B  0.70 0.19 0.28 0.93 0.66 0.13 0.38 0.85 BA  0.88 0.09 0.59 0.97 0.67 0.18 0.28 0.91 Ba 0.55 0.33 0.08 0.95 0.64 0.19 0.27 0.90 Nyala/Impala A  0.50 0.14 0.24 0.76 0.42 0.09 0.26 0.59 B  0.52 0.09 0.34 0.70 0.64 0.12 0.39 0.84 BA  0.58 0.18 0.24 0.86 0.72 0.14 0.39 0.91 Ba 0.47 0.19 0.17 0.80 0.59 0.19 0.24 0.87 Nyala/Warthog A  0.55 0.13 0.30 0.77 0.41 0.08 0.26 0.58 B  0.50 0.10 0.31 0.69 0.34 0.09 0.19 0.53 BA  0.38 0.13 0.17 0.64 0.59 0.15 0.29 0.83 Ba 0.64 0.19 0.26 0.90 0.16 0.10 0.04 0.45 Warthog/Bushpig A  0.66 0.19 0.27 0.91 0.50 0.09 0.33 0.68 B  0.50 0.20 0.17 0.83 0.84 0.07 0.65 0.93 BA  0.77 0.14 0.42 0.94 0.68 0.13 0.40 0.87 Ba 0.00 0.00 0.00 0.00 1.00 0.00 1.00 1.00

103

Low shrub cover percentage High shrub cover percentage Species Parameter  SE Lower CI Upper CI  SE Lower CI Upper CI Warthog/Impala A  0.64 0.09 0.44 0.79 0.56 0.10 0.37 0.73 B  0.54 0.10 0.36 0.71 0.54 0.10 0.34 0.72 BA  0.67 0.11 0.43 0.84 0.87 0.10 0.56 0.97 Ba 0.31 0.16 0.09 0.67 0.12 0.08 0.03 0.37 Wildebeest/Impala A  0.37 0.10 0.21 0.57 0.20 0.07 0.10 0.38 B  0.69 0.13 0.41 0.88 0.62 0.23 0.19 0.92 BA  0.79 0.13 0.46 0.95 0.70 0.19 0.29 0.93 Ba 0.64 0.19 0.26 0.90 0.61 0.29 0.13 0.94 Wildebeest/Nyala A  0.48 0.10 0.30 0.66 0.16 0.07 0.06 0.34 B  0.49 0.10 0.31 0.67 0.43 0.09 0.27 0.60 BA  0.87 0.09 0.59 0.97 1.00 0.00 1.00 1.00 Ba 0.14 0.08 0.04 0.38 0.32 0.10 0.16 0.53 Wildebeest/Warthog A  0.23 0.07 0.12 0.41 0.16 0.07 0.06 0.36 B  0.53 0.15 0.26 0.79 0.33 0.09 0.19 0.52 BA  0.68 0.18 0.30 0.91 0.53 0.25 0.14 0.89 Ba 0.49 0.19 0.18 0.81 0.30 0.10 0.14 0.51 Zebra/Impala A  0.58 0.12 0.34 0.78 0.17 0.10 0.05 0.44 B  0.79 0.09 0.56 0.91 0.47 0.09 0.31 0.64 BA  0.94 0.06 0.68 0.99 0.40 0.28 0.06 0.87 Ba 0.58 0.15 0.28 0.82 0.49 0.11 0.29 0.68 Zebra/Nyala A  0.62 0.19 0.26 0.88 0.48 0.20 0.17 0.81 B  0.56 0.14 0.30 0.80 0.44 0.09 0.27 0.62 BA  0.42 0.16 0.16 0.73 0.58 0.23 0.18 0.89 Ba 0.79 0.33 0.07 1.00 0.31 0.20 0.07 0.74 Zebra/Warthog A  0.94 0.11 0.27 1.00 0.59 1.07 0.00 1.00 B  0.51 0.12 0.28 0.73 0.55 1.07 0.00 0.99 BA  0.52 0.12 0.30 0.73 0.51 0.95 0.00 1.00 Ba 0.35 0.75 0.00 1.00 0.60 1.03 0.00 1.00 Zebra/Wildebe est A  0.42 0.03 0.35 0.48 0.09 0.01 0.07 0.11 B  0.31 0.06 0.21 0.43 0.17 0.06 0.08 0.31 BA  0.28 0.13 0.10 0.58 0.38 0.19 0.11 0.75 Ba 0.33 0.04 0.25 0.41 0.15 0.06 0.07 0.29

104

APPENDIX F DIAGNOSTIC PLOTS FOR MULTIPLE LINEAR REGRESSION USED TO MODEL THE INFLUENCE OF COVARIATES ON THE SPECIES INTERACTION FACTOR

Figure F-1. Diagnostic plots for multiple linear regression used to model the influence of covariates on the species interaction factor estimated for pairwise co-occurrence comparisons of ungulates detected in camera trapping surveys in Mlawula Nature and Mbuluzi Game Reserves, Swaziland, 2012-2013.

105

LIST OF REFERENCES

Acebes, P., Traba, J., and Malo, J.E. 2011. Co-occurrence and potential for competition between wild and domestic large herbivores in a South American desert. Journal of Arid Environments 77, 39-44.

Anselin, L., Syabri, I., and Kho, Y. 2006. GeoDa: An introduction to spatial data analysis. Geographical Analysis 38, 5-22.

Archer, S. 1995. Tree-grass dynamics in a Prosopis-thornscrub savanna parkland: reconstructing the past and predicting the future. Ecoscience 2, 83-99.

Archer, S., Boutton, T.W., and Hibard, K.A. 2001. Trees in grasslands: biogeochemical consequences of woody plant expansion. In Schulze, M., Heimann, S., Harrison, E., Holland J., and Lloyd, I. (Eds.). Global biogeochemical cycles in the climate system. Academic Press, San Diego, California, pp. 115 – 138.

Archibald, S., and Bond, W.J. 2004. Grazer movements: spatial and temporal responses to burning in a tall-grass African savanna. International Journal of Wildland Fire 13, 377-385.

Arsenault, R., and Owen-Smith, N. 2002. Facilitation versus competition in grazing herbivore assemblages. Oikos 97, 313-318.

Asner, G.P., Elmore, A.J., Olander, L.P., Martin, R.E., and Harris, A.T. 2004. Grazing systems, ecosystem responses and global change. Annual Review of Environmental Resources, 29: 261-299.

Atickem, A., and Loe, L.E. 2013. Livestock-wildlife conflicts in the Ethiopian highlands: assessing the dietary and spatial overlap between mountain nyala and cattle. African Journal of Ecology 52, 343-351.

Augustine, D.J., and McNaughton, S.J. 2004. Regulation of shrub dynamics by native browsing ungulates on East African rangeland. Journal of Applied Ecology, 41, 45-58.

Bai, E., Boutton, T.W., Wu, X. B., Liu, F., and Archer, S.R. 2009. Landscape-scale vegetation dynamics inferred from spatial patterns of soil 13 C in a subtropical savanna parkland. Journal of Geophysical Research 114, G01019.

Bailey, K.M., McCleery, R.A., Binford, M.W., and Zweig, C. 2016. Land-cover change within and around protected areas in a biodiversity hotspot. Journal of Land Use Science 11, 154- 176.

Bailey, L.L., Hines, J.E., Nichols, J.D., and MacKenzie, D.I. 2007. Sampling design tradeoffs in occupancy studies with imperfect detection: examples and software. Ecological Applications 17, 281-290.

106

Barthelemy, K., Roland, L., and Brice, S. 2008. Diet and food preference of the waterbuck (Kobus ellipsiprymnus defassa) in the Pendjari National Park, Benin. African Journal of Ecology 46, 303-310.

Beesley, L.S., Gwinn, D.C., Price, A., King, A.J., Gawne, B., Koehn, J.D., and Nielsen, D.L. 2014. Juvenile fish response to wetland inundation: how antecedent conditions can inform environmental flow policies for native fish. Journal of Applied Ecology 51, 1613-1621.

Bell, R.H.V. 1970. The use of the herb layer by grazing ungulates in the Serengeti. In: Watson, A. (ed) Animal populations in relation to their food resources (Symposium of the British Ecological Society, Aberdeen 1969). Blackwell, Oxford, pp 111-124.

Ben-Shahar, R. 1992. The effects of bush clearance on African ungulates in a semi-arid nature reserve. Ecological Application 2, 95-101.

Besag, J., York, J., and Mollie, A. 1991. Bayesian image restoration with two applications in spatial statistics. Annals of the Institute of Statistical Maths 43, 1-59.

Blaum, N., Rossmanith, E., Popp, A., and Jeltsch, F. 2007a. Woody encroachment affects mammalian carnivore abundance and species richness in semiarid rangelands. Acta Oecologica 31:86-92.

Blaum, N., Rossmanith, E., Popp, A., and Jeltsch, F. 2007b. Land use affects rodent communities in Kalahari savanna rangelands. African Journal of Ecology 45, 189–195.

Blaum, N., Rossmanith, E., Fleissner, G., and Jeltsch, F. 2007c. The conflicting importance of shrubby landscape structures for the reproductive success of the yellow mongoose (Cynictis penicillata). Journal of Mammalogy 88, 194-200.

Blaum, N., Rossmanith, E., Schwager, M., and Jeltsch, F. 2007d. Responses of mammalian carnivores to land use in arid savanna rangelands. Basic and Applied Ecology 8, 552-564.

Bolker, B.M. 2008. Ecological models and data in R. 1st. edition. Princeton University Press. Princeton, NJ.

Boycott, R., Forrester, B., Loffler, L., and Monadjem, A. 2007. Wild Swaziland: Common animals and plants. Natural History Society of Swaziland and Swaziland Conservation Trust. Swaziland.

Bransby, D.I., and Tainton, N.M. 1977. The disc pasture meter: possible applications in grazing management. Proceedings of the Grassland Society of Southern Africa, 12:115-118.

Brooks, S.P., and Gelman, A. 1998. General methods for monitoring convergence of iterative simulations. Journal of Computational and Graphical Statistics 7, 434-455.

107

Burnham, K.P., and Anderson, D.R. 2002. Model selection and multimodel inference: A practical information theoretic approach. 2nd. edition. Springer-Verlag, New York, New York, USA.

Burton, A.C., Sam, M.K., Balangtaa, C., and Brashares, J.S. 2012. Hierarchical multi-species modeling of carnivore responses to hunting, habitat and prey in a West African protected area. PLoS One 7, e38007.

Canfield, R.H. 1941. Application of the line interception method in sampling range vegetation. Journal of Forestry 39, 388-394.

Caro, T. 2008. Decline of large mammals in the Katavi-Rukwa ecosystem of western Tanzania. African Zoology 43, 99-116.

Caro, T.M., and Stoner, C.J. 2003. The potential for interspecific competition among African carnivores. Biological Conservation 110, 67-75.

Chandler, R.B., and Royle, J.A. 2013. Spatially explicit models for inference about density in unmarked or partially marked populations. The Annals of Applied Statistics 7, 936 – 954.

Chevan, A. and Sutherland, M. 1991. Hierarchical Partitioning. The American Statistician 45, 90-96.

Coates, G.D., and Downs, C.T. 2007. Population estimates of bushbuck (Tragelaphus scriptus) in valley thicket and coastal bushveld-grassland habitat. South African Journal of Wildlife Research 37, 91-95.

Collier, B.A., McCleery, R.A., Clahouh, K.W., Roques, K.G., and Monadjem, A. 2011. Detection probabilities of ungulates in the eastern Swaziland lowveld. South African Journal of Wildlife Research 41, 61-67.

Craigie, I. D., Baillie, J. E., Balmford, A., Carbone, C., Collen, B., Green, R. E., and Hutton, J. M. 2010. Large mammal population declines in Africa’s protected areas. Biological Conservation 143, 2221-2228.

Cromsigt, J.P.G.M., Prins, H.H.T., and Olff, H. 2009. Habitat heterogeneity as a driver of ungulate diversity and distribution patterns: Interaction of body mass and digestive strategy. Diversity and Distributions 15, 513-522.

Dalerum, F., Somers, M.J., Kunkel, K.E., Cameron, E.Z., 2008. The potential for large carnivores to act as biodiversity surrogates in southern Africa. Biodiversity and Conservation 17, 2939-2949.

Daubenmire, R.F. 1959. A tree-coverage method of vegetational analysis. Northwest Science 33, 43-64.

108

Denes, F.V., Silveira, L.F., and Beissinger, S.R. 2015. Estimating abundance of unmarked animal populations: accounting for imperfect detection and other sources of zero inflation. Methods in Ecology and Evolution 6, 543 – 556.

Du Toit, J. T., and Cumming, D. H. M. 1999. Functional significance of ungulate diversity in African savannas and the ecological implications of the spread of pastoralism. Biodiversity and Conservation 8, 1643-1661.

Du Toit, J.T. 2003. Large herbivores and savanna heterogeneity. In: Du Toit, J, Rogers, K., and Biggs, H. (eds) The Kruger experience: Ecology and management of savanna heterogeneity. Island Press, London, pp 292-309.

Dunham, K.M. 2012. Trends in populations of elephants and other large herbivores in Gonarezhou National Park, Zimbabwe, as revealed by sample aerial surveys. African Journal of Ecology 50, 476-488.

Eldridge, D.J., and Soliveres, S. 2014. Are shrubs really a sign of a declining ecosystem function? Disentangling the myths and truths of woody encroachment in Australia. Australian Journal of Botany 62, 594-608.

Ferretti, F., Corazza, M., Campana, I., Petrocini, V., Brunetti, C., Scornavacca, D., and Lovari, S. 2015. Competition between wild herbivores: reintroduced red deer and Apennine chamois. Behavioral Ecology 26, 550-559.

Fisher, J.T., Wheatley, M., and Mackenzie, D. 2014. Spatial patterns of breeding success of grizzly bears derived from hierarchical multistate models. Conservation Biology 28, 1249-1259.

Fisher, N., and Lee, A. 1992. Regression models for an angular response. Biometrics 48, 665–677.

Fiske, I, and Chandler, R. 2011. Unmarked: An R Package for Fitting Hierarchical Models of Wildlife Occurrence and Abundance. Journal of Statistical Software 43, 1-23.

Foster, R.J., and Harmsen, B.J. 2012. A critique of density estimation from camera-trap data. Journal of Wildlife Management 76, 224-236.

Gandiwa, E. 2013. Vegetation factors influencing density and distribution of wild large herbivores in a southern African savanna. African Journal of Ecology 52, 274-283.

Gertenbach, W.P.D., and Potgieter, A.L.F. 1975. Veldbrandnavorsing in die knoppiesdoring- maroelaveld van die sentrale distrik, Kruger Nationale Park. Unpublished burning trial report for the South African National Parks Board, Pretoria.

Gibson, D.J. 2009. Grasses and grassland ecology. 1st. edition, Oxford University Press, New York.

109

Goheen, J. R., Palmer, T. M., Keesing, F., Riginos, C., and Young, T. P. 2010. Large herbivores facilitate savanna tree establishment via diverse and indirect pathways. Journal of Animal Ecology 79, 372-382.

Gordon, I.J., and Prins, H.H.T. 2008. Grazers and browsers in a changing world: Conclusions. In: Gordon, I.J., and Prins, H.H.T (Ed.). The ecology of browsing and grazing. Ecological Studies, Volume 195. Springer, Berlin. pp. 309-321.

Gotelli, N.J., and Ellison, A.M. 2004. A primer of ecological statistics. 1st. edition. Sinaur Associates, Inc., Sunderland, MA.

Goudie, A., and Price Williams, D. 1983. The Atlas of Swaziland. Swaziland Trust Commission. 90pp.

Grace, J., San Jose, J., Meir, P., Miranda, H.S., and Montes, R.A. 2006. Productivity and carbon fluxes of tropical savanna. Journal of Biogeography 33, 387-400.

Guillera-Arroita, G., Lahoz-Monfort, J.J., MacKenzie, D.I., Wintle, B.A., and McCarthy, M.A. 2014. Ignoring Imperfect Detection in Biological Surveys Is Dangerous: A Response to ‘Fitting and Interpreting Occupancy Models'. PLoS ONE 9, e99571. doi:10.1371/journal.pone.0099571.

Hines, J.E. 2006a. PRESENCE – software to estimate patch occupancy and related parameters. USGS-PWRC. http://www.mbr-pwrc.usgs.gov/software.html

Hines, J. E. 2006b. GENPRES- Software to Generates patch occupancy data and analyze using programs MARK or PRESENCE. USGS-PWRC. http://www.mbr-pwrc.usgs.gov/software/presence.html.

Hobbs, N.T., and Gordon, L.J. 2010. How does landscape heterogeneity shape dynamics of large herbivore populations? In: Owen-Smith, N. (Ed.) Dynamics of large herbivore populations in changing environments, towards appropriate models. Wiley-Blackwell, John Wiley and Sons Ltd. pp. 141-164.

Holdo, R. M., and Mack, M. C. 2014. Functional attributes of savanna soils: contrasting effects of tree canopies and herbivores on bulk density, nutrients and moisture dynamics. Journal of Ecology 102, 1171-1182.

Holdo, R. M., Holt, R. D., and Fryxell, J. M. 2013. Herbivore–vegetation feedbacks can expand the range of savanna persistence: insights from a simple theoretical model. Oikos, 122(3), 441-453.

Hopcraft, J.G., Anderson, T.M, Perez-Vila, S., Mayemba, E., and Olff, H. 2012. Body size and the division of niche space: food and predation differentially shape the distribution of Serengeti grazers. Journal of Animal Ecology 81, 201-213.

110

Illius, A.W., and Gordon, I.J. 1992. Modeling the nutritional ecology of ungulate herbivores:evolution of body size and competitive interactions. Oecologia 89, 428-434.

Isaacs, L., Somers, M.J., and Dalerum, F. 2013. Effects of prescribed burning and mechanical bush clearing on ungulate space use in an African savannah. Restoration Ecology 21, 260-266.

Jarman, P.J. 1972. Seasonal distribution of large mammal populations in the unflooded Middel Zambezi Valley. Journal of Applied Ecology 9, 283-299.

Jarman, P. J. 1974. The social organisation of antelope in relation to their ecology. Behaviour 48, 215–267.

Jarman, P.J., and Sinclair, A.R.E. 1979. Feeding strategy and the pattern of resource partitioning in ungulates. In: Sinclair, A.R.E. and Norton Griffiths, N. (eds) Serengeti: dynamics of an ecosystem. Chicago University Press, Chicago, pp 130-163.

Jones, K.E., Bielby, J., Cardillo, M., Fritz, S.A., O'Dell, J., David, C., Orme, L., Safi, K., Sechrest, W., Boakes, E.H., Carbone, C., Connolly, C., Cutts, M.J., Foster, J.K., Grenyer, R., Habib, M., Plaster, C.A., Price, S.A., Rigby, S.E., Rist, J., Teacher, A., Bininda-Emonds, O.R.P., Gittleman, J.L., Mace, G.M., and Purvis, A. 2009. PanTHERIA: a species-level database of life history, ecology, and geography of extant and recently extinct mammals. Ecology 90:2648.

Kartzinel, T.R., Chen, P.A., Coverdale, T.C., Erickson, D.L., Kress, W.J., Kuzmina, M.L., Rubenstein, D.I., Wang, W., and Pringle, R.M. 2015. DNA metabarcoding illuminates dietary niche partitioning by African large herbivores. Proceedings of the National Academy of Sciences 112, 8019-8024.

Kate E. Jones, Jon Bielby, Marcel Cardillo, Susanne A. Fritz, Justin O'Dell, C. David L. Orme, Kamran Safi, Wes Sechrest, Elizabeth H. Boakes, Chris Carbone, Christina Connolly, Michael J. Cutts, Janine K. Foster, Richard Grenyer, Michael Habib, Christopher A. Plaster, Samantha A. Price, Elizabeth A. Rigby, Janna Rist, Amber Teacher, Olaf R. P. Bininda- Emonds, John L. Gittleman, Georgina M. Mace, and Andy Purvis. 2009. PanTHERIA: a species-level database of life history, ecology, and geography of extant and recently extinct mammals. Ecology 90:2648.

Kery, M. 2010. Introduction to WinBUGS for ecologists: A Bayesian approach to regression, ANOVA, mixed models and related analyses. Academic Press, San Diego.

Kleynhans, E.J., Jolles, A.E., Bos, M.R.E., and Olff, H. 2011. Resource partitioning along multiple niche dimensions in differently sized African savanna grazers. Oikos 120, 591-600.

Klop, E., and Goethem, J.V. 2008. Savanna fires govern community structure of ungulates in Bénoué National Park, Cameroon. Journal of Tropical Ecology, 24, 39-47.

111

Kraaij, T., and Ward, D., 2006. Effects of rain, nitrogen, fire and grazing on tree recruitment and early survival in bush-encroached savanna, South Africa. Plant Ecology, 186, 235-246.

Levick, S.R., Asner, G.P., Kennedy-Bowdoin, T., and Knapp, D.E. 2009. The relative influence of fire and herbivory on savanna three-dimensional vegetation structure. Biological Conservation 142, 1693-1700.

Lohmann, D., Tietjen, B., Blaum, N., Joubert, D.F., and Jeltsch, F. 2014. Prescribed fire as a tool for managing shrub encroachment in semi-arid savanna rangelands. Journal of Arid Environments 107, 49-56.

Lorenzen, E. D., R. Heller, R., and Siegismund, H.R. 2012. Comparative phylogeography of African savannah ungulates. Molecular Ecology 21, 3656-3670.

MacArthur, R.W., and MacArthur, J.W. 1961. On bird species diversity. Ecology 42, 594–598.

MacKenzie, D. I., Bailey, L.L., and Nichols, J. D. 2004. Investigating species co-occurrence patterns when species are detected imperfectly. Journal of Animal Ecology 73, 546-555.

MacKenzie, D. I., Nichols, J. D., Royle, J. A., Pollock, K. H., Bailey, L. L., and Hines, J. E. 2006. Occupancy estimation and modeling. Elsevier, Amsterdam, The Netherlands.

Marchal, A., Lejeune, P., Bouche, P. Ouedraogo, M., Sawadogo, P., Yameogo, D.D., Yanogo, F., and Vemeulen, C. 2012. Status of medium-sized ungulate populations in 2010, at the Nazinga Game Ranch, Burkina Faso (Western Africa). Biotechnology, Agronomy, Society and Environment 16, 307-315.

McGarigal, K., Cushman, S., and Stafford, S., 2000. Multivariate Statistics for Wildlife and Ecology Research. Springer-Verlag, New York, Inc.

McNaughton, S. J. 1992. The propagation of disturbance in savannas through food webs. Journal of Vegetation Science 3, 301–-314.

McNaughton, S.J., and Georgadis, N.J. 1986. Ecology of African grazing and browsing mammals. Annual Review of Ecology and Systematics 17, 39-65.

Meik, J.M., Jeo, R.M., Mendelson III, J.R., and Jenks, K.E. 2002. Effects of bush encroachment on an assemblage of diurnal lizard species in central Namibia. Biological Conservation 106, 29-36.

Midgley, G.F., and Thuiller, W. 2010. Potential responses of terrestrial biodiversity in Southern Africa to anthropogenic climate change. Regional Environmental 11, 127-135 . Milton, S.J., and Dean, W.R.J. 1995. South Africa’s arid and semiarid rangelands: Why are they changing and can they be restored? Environmental Monitoring and Assessment 37, 245-264.

112

Monadjem, A. 1997. Habitat preferences and biomasses of small mammals in Swaziland. African Journal of Ecology 35, 64–72.

Monadjem, A. 1998. Distributional patterns and conservation status of mammals of Swaziland, southern Africa. Koedoe 41, 45-59.

Monadjem, A., Boycott, R.C., and Culverwell, J. 2003. Threatened vertebrates of Swaziland. Swaziland Red Data Book: Fishes, amphibians, reptiles, birds and mammals.

Moore, J.E., and Swihart, R.K. 2005. Modeling patch occupancy by forest rodents: incorporating detectability spatial autocorrelation with hierarchically structured data. Journal of Wildlife Management 69, 933-949.

Mugume, S., Isabirye-Basuta, G., Otali, E., Reyna-Hurtado, R., and Chapman, C.A. 2015. How do human activities influence the status and distribution of terrestrial mammals in forest reserves? Journal of Mammalogy 96, 998–1004.

Norman, G.W., Conner, M.M., Pack, J.C., and White, G.C. 2004. The effects of fall hunting on survival of male wild turkeys in Virginia and West Virginia. Journal of Wildlife Management 68, 393-404.

O’Brian, T.G. 2010. Abundance, density and relative abundance: A conceptual framework. In O’Connell, A.F., Nichols, J.D. and Karanth, K.U (Eds.). Camera traps in animal ecology: methods and analyses. Springer, New York.

O’Kane, C.A.J., Duffy, K.J., Page, B.R., and Macdonald, D.W. 2013. Effects of resource limitation on habitat usage by the browser guild in Hluhluwe-iMfolozi Park, South Africa. Journal of Tropical Ecology 29, 39-47.

O’Kane, C.A.J., Page, B.R., and MacDonald, D.W. 2014. Differing influences of resource availability on the demographics and habitat selection of wildebeest compared with impala. Journal of Tropical Ecology 30, 189-198.

Oksanen, J., Guillaume Blanchet,F., Kindt, R., Legendre, P., Minchin, P.R., O'Hara, R.B., Simpson, G.L., Solymos, P., Stevens, M.H., and Wagner, H. 2015. Vegan: Community Ecology Package. R package version 2.2-1. http://CRAN.R-project.org/package=vegan

Owen-Smith, N. 2002. Adaptive herbivore ecology: from resources to populations in variable environments. Cambridge University Press, United Kingdom. 374 pp.

Parr, C.L., Lehmann, C.E.R., Bond, W.J., Horrmann, W.A., and Andersen, A.N. 2014. Tropical grassy biomes: misunderstood, neglected, and under threat. Trends in Ecology and Evolution 29, 205-213.

113

Plummer, M. 2003. JAGS: A program for analysis of Bayesian graphical models using Gibbs sampling. Proceedings of the 3rd International Workshop on Distributed Statistical Computing (Eds Leisch, F., and Zeleis, A.), Vienna, Austria.

Pringle, R.M., Goheen, J.R., Palmer, T.M., Charles, G.K., DeFranco, E., Hohbein, R., Ford, A.T., Tarnita, C.E. 2014. Low functional redundancy among mammalian browsers in regulating an encroaching shrub (Solanum campylacanthum) in African savannah. Proceedings of the Royal Society of Biology 281: 20140390. http://dx.doi.org/10.1098/rspb.2014.0390

Prins, H.H.T. 2000. Competition between wildlife and livestock in Africa. In Prins, H.H.T., Grootenhuis, J.G., and Dolan, T.T. (eds). Kluwer Academic, Boston. pp 51-80.

R Development Core Team. 2011. R: A language and environment for statistical computing. R Foundation for statistical computing, Vienna, Austria.

Ramesh, T., and Downs, C.T. 2013. Impact of farmland use on population density and activity patterns of serval in South Africa. Journal of Mammalogy 94, 1460-1470.

Ratter, J. A., Ribeiro, J.F., and Bridgewater, S. 1997. The Brazilian Cerrado vegetation and threats to biodiversity. Annuals of Botany 80, 223-230.

Richmond, O.M.W, Hines, J.E., and Beissinger, S.R. 2010. Two-species occupancy models: a new parameterization applied to co-occurrence of secretive rails. Ecological Applications 20, 2036-2046.

Roberts, N.J. 2011. Investigation into survey techniques of large mammals: surveyor competence and camera-trapping vs. transect-sampling. Bioscience Horizons 4, 40-49.

Roques, K.G., O’Connor, T.G., and Watkinson, A.R. 2001. Dynamics of woody encroachment in an African savanna: Relative influence of fire, herbivory, rainfall and density dependence. Journal of Applied Ecology 38, 268-280.

Royle, J. A., and Dorazio, R.M. 2008. Hierarchical modeling and inference in ecology. Academic Press, Amsterdam, The Netherlands.

Royle, J.A., and Nichols, J.D. 2003. Estimating abundance from repeated presence-absence data or point counts. Ecology 84, 777–790.

Sankaran, M., Hanan, N.P., Scholes, R.J., Ratnam, J., Augustine, D.J., Cade, B.S., Gignoux, J., Higgins, S.I., Le Roux, X., Ludwig, F., Ardo, J., Banyikwa, F., Bronn, A., Bucini, G., Caylor, K.K., Coughenour, M.B., Diouf, A., Ekaya, W., Feral, C.J., February, E.C., Frost, P.G.H., Hiernaux, P., Hrabar, H., Metzger, K.L., Prins, H.H.T., Ringrose, S., Sea, W., Tews, J., Worden, J., and Zambatis, N., 2005. Determinants of woody cover in African savannas. Nature 438, 846-849.

114

Sankaran, M., Ratnam, J. and Hanan, N. 2008. Woody cover in African savannas: the role of resources, fire and herbivory. Global Ecology and Biogeography 17, 236-245.

Scholes, R.J., and Archer, S.R. 1997. Tree-grass interactions in savannas. Annual Review of Ecological Systematics 28, 517-544.

Scholte, P. 2011. Towards understanding large mammal population declines in Africa’s protected areas: A West-Central African perspective. Tropical Conservation Science 4, 1-11.

Silva, E.A., and Sieving, K.E. 2012. Domestic dogs shape the landscape-scale distribution of a threatened forest ungulate. Biological Conservation 150, 103 – 110.

Silver, S.C., Ostro, L.E.T., Marsh, L.K., Maffei, L., Noss, A.J., and M. Kelly. 2004. The use of camera traps for estimating jaguar (Panthera onca) abundance and density using capture/ recapture analysis. Oryx 28,148–154.

Sirami, C., and Monadjem, A. 2012. Changes in bird communities in Swaziland savannas between 1998 and 2008 owing to woody encroachment. Diversity and Distributions 18, 390-400.

Skinner, J.D., and Chimimba, C.T. 2005. The mammals of the southern African subregion. 3rd edition. Cambridge University Press.

Smit, I.P.J., and Prins, H.H.T. 2015. Predicting the effects of woody encroachment on mammal communities, grazing biomass and fire frequency in African savannas. PLoS ONE 10, e0137857, doi: 10.1371/journal.pone.0137857.

Sollmann, R., Malzoni Furtado, M., Hofer, H., Jacomo, A.T.A., Mundim Torres, N., and Silveira, L. 2012. Using occupancy models to investigate space partitioning between two sympatric large predators, the jaguar and puma in central Brazil. Mammalian Biology 77, 41-46.

Sponheimer, C.C., De Ruitter, D.J., Lee-Thorp, J.A., Codron, D.M and Codron, J. 2003. Diets of impala from Kruger National Park: evidence from stable carbon isotopes. Koedoe 46, 101- 106.

Steenkamp, H.E., and Chown, S.L. 1996. Influence of dense stands of an exotic tree, Prosopis glandulosa Benson, on a savanna dung beetle (Coleoptera: Scarabaeinae) assemblage in southern Africa. Biological Conservation 78, 305–311.

Struhsaker, T.T. 1967. Ecology of vervet monkeys (Cerocphitecus aethiops) in the Masai- Amboseli Game Reserve, Kenya. Ecology 48, 891-904.

Swaziland Environmental Authority. 2001. Swaziland National Biodiversity Strategy and Action Plan. Unpublished report.

Sweet, R.J., and Khumalo S. 1994. Range Resources and Grazing Potentials in Swaziland.

115

FAO Report, Ministry of Agriculture and Co-operatives, Mbabane.

Tambling, C.J., Minnie, L., Adendorff, J., and Kerley, G.I.H. 2013. Elephants facilitate impact of large predators on small ungulate prey species. Basic and Applied Ecology 14, 694-701.

Tews, J., Brose, U., Grimm, V., Tielborger, K., Wichmann, M.C., Schwager, M., and Jeltsch, F. 2004. Animal species diverstiy driven by habitat heterogeneity/diversity: the importance of keystone structures. Journal of Biogeography 31, 79-92.

Thornton, D., Branch, L., and Sunquist, M. 2011. Passive sampling effects and landscape location alter associations between species traits and response to fragmentation. Ecological Applications 21, 817 – 829.

Tobler, M.W., Zuniga Hartley, A., Carrillo-Percastegui, S.E., and Powell, G.V.N. 2015. Spatiotemporal hierarchical modelling of species richness and occupancy using camera trap data. Journal of Applied Ecology 52, 413-421.

Uys, R. G. 2006. Patterns of plant diversity and their management across South African rangelands. Doctoral thesis. Botany Department, University of Cape Town.

Van Auken, O.W. 2000. Shrub invasions of North American semiarid grasslands. Annual Review of Ecology and Systematics 31, 197-215.

Vermeire, L.T., Ganguli, A.C., and Gillen, R.L. 2002. A robust model for estimating standing crop across vegetation types. J. Range Manage. 55, 494–497.

Waddle, J.H., Dorazio, R.M., Walls, S.C., Rice, K.G., Beuachamp, J., Schuman, and Mazzotti, F.J. 2010. A new parameterization for estimating co-occurrence of interacting species. Ecological Applications 20, 1467-1475.

Walsh, C. and Mac Nally, R. 2013. hier.part: Hierarchical Partitioning. R package version 1.0-4. http://CRAN.R-project.org/package=hier.part

Western, D., Russell, S., and Cuthill, I. 2009. The status of wildlife in protected areas compared to non-protected areas of Kenya. PLoS One 4, e6140

White, G.C., and Burnham, K.P. 1999. Program MARK: Survival estimation from populations of marked animals. Bird Study 46 Supplement, 120-138.

Wigley, B.J., Bond, W.J., and Hoffman, M.T. 2010. Thicket expansion in a South African savanna under divergent land use: local vs. global drivers? Global Change Biology 16, 964- 976.

Williams, B.K., Nichols, J.D., and Conroy, M.J. 2002. Analysis and management of animal populations. Academic Press, San Diego CA. pp. 817.

116

Yamaura, Y., Royle, A., Kuboi, J., Tada, T., Ikeno, S., and Makino, S. 2011. Modelling community dynamics based on species-level abundance models from detection/nondetection data. Journal of Applied Ecology 48, 67-75 . Zapata-Rios, G., and Branch, L.C. 2016. Altered activity patterns and reduced abundance of native mammals in sites with feral dogs in the high Andes. Biological Conservation 193, 9- 16.

Zar, J.H. 1999. Biostatistical analysis. 4th edition. Prentice Hall, Upper Saddle River, NJ.

117

BIOGRAPHICAL SKETCH

Jose was born in the small town of Chimaltenango, approximately 54 km from

Guatemala City. Shortly after graduating high school in Guatemala, he began to teach English, an occupation that he practiced for 15 years. Having been raised most of his life in the city, it was only during a trip to the Mayan temples of Tikal in Peten, that he realized how rich and abundant the natural resources of his country were. He was fascinated with the idea that he was walking through a jungle where such majestic animals as jaguars roamed. This trip was pivotal in his decision to pursue a degree in wildlife biology. The next year he enrolled in the San Carlos

University of Guatemala, where he obtained his bachelor’s degree in biology. This is where he met his beautiful wife, Nancy, with whom he has a wonderful son named Danny. For his bachelor’s thesis, Jose examined the impacts of subsistence hunting on terrestrial vertebrates in the Maya Forest. Soon after, Jose began working in wildlife research projects with the Wildlife

Conservation Society, Program for Guatemala in Peten in 2001. Since then he has worked in research and outreach projects involving endangered Neotropical vertebrates. In 2008, Jose completed his master’s degree at the Department of Wildlife Ecology and Conservation in the

University of Florida. For his research project, he examined determinants and spatial and temporal patterns of human-carnivore conflicts in and around the Maya Biosphere Reserve in

Guatemala. An important contribution of this work was the establishment of a national program for monitoring and mitigating human-carnivore conflicts that involved many local stakeholders.

After this, he started his doctoral program where he studied the impacts of global change on savanna associated mammals in southern Africa. Jose plans to return to his home country to apply and share the knowledge and experience he has obtained throughout his graduate studies with the local conservation community. He dicho!

"ID Y ENSEÑAD A TODOS"

118