Precipitation, vegetation productivity, and human impacts control home range size of elephants in dryland systems in northern Namibia

Abstract Climatic variability, resource availability, and anthropogenic impacts heavily influence an animal's home range. This makes home range size an effective metric for understanding how variation in environmental factors alter the behavior and spatial distribution of animals. In this study, we estimated home range size of African elephants (Loxodonta africana) across four sites in Namibia, along a gradient of precipitation and human impact, and investigated how these gradients influence the home range size on regional and site scales. Additionally, we estimated the time individuals spent within protected area boundaries. The mean 50% autocorrelated kernel density estimate for home range was 2200 km2 [95% CI:1500–3100 km2]. Regionally, precipitation and vegetation were the strongest predictors of home range size, accounting for a combined 53% of observed variation. However, different environmental covariates explained home range variation at each site. Precipitation predicted most variation (up to 74%) in home range sizes (n = 66) in the drier western sites, while human impacts explained 71% of the variation in home range sizes (n = 10) in Namibia's portion of the Kavango‐Zambezi Transfrontier Conservation Area. Elephants in all study areas maintained high fidelity to protected areas, spending an average of 85% of time tracked on protected lands. These results suggest that while most elephant space use in Namibia is driven by natural dynamics, some elephants are experiencing changes in space use due to human modification.


| INTRODUC TI ON
Home range is a fundamental concept of ecology, used to characterize space use patterns of animals and has been defined as the total area required to meet nutritional and reproductive needs throughout an animal's lifetime (Burt, 1943). Home-range estimation is a potentially useful metric for defining the appropriate size of protected areas or for understanding how environmental factors impact the behavior of individuals and the spatial distribution of populations (Börger et al., 2006).
However, there is a lack of studies which examine how ecological drivers contribute to individual variation in home range size within a region (Seigle-Ferrand et al., 2021). Understanding how these factors influence home range size and space use is important for informing the management and conservation of threatened species. This is especially true for large-bodied mammals, such as elephants (Loxodonta africana), which move long distances and are more susceptible to extinction as a result (Cardillo et al., 2005).
African savannah elephants are the world's largest terrestrial animal and are a species of high conservation concern (Thouless et al., 2016). Once widespread across the African continent, elephants are now largely restricted to isolated protected areas, with their distribution limited mostly by human encroachment rather than environmental conditions (Wall et al., 2021). This restriction of their range, along with poaching, has led to a steep decline in elephant numbers across Africa -a reduction to approximately 118,000 elephants in 10 years (2007)(2008)(2009)(2010)(2011)(2012)(2013)(2014)(2015)(2016); Thouless et al., 2016). The species was recently downgraded from Vulnerable to Endangered on the IUCN Red List (Gobush et al., 2021), with continental population estimates declining by more than 50% in the past two generations (50 years; Gobush et al., 2021). Elephants are important ecosystem engineers, found across a variety of habitats from deserts to tropical forests (Haynes, 2012). Because they inhabit many ecosystems, the size of elephant home ranges can differ dramatically between populations. For example, elephants in the deserts of Mali can have home ranges up to 32,000 km 2 (Wall et al., 2013), while the largest home ranges in the wet savannahs of Uganda are closer to 500 km 2 (Grogan et al., 2020). The disparity between populations highlights the need for population-level studies to assess the relationship between home range size and environmental factors to better understand factors driving differences in spatial requirements. It is especially important to compare populations at multiple scales and across gradients of land use and ecological conditions to best understand the scale at which elephants are influenced by anthropogenic and environmental factors.
In this study, we examined home range size of elephants in northern Namibia, an important stronghold for the species where elephant numbers have more than doubled since 1995 (Thouless et al., 2016). Elephants span a diverse mosaic of land uses and environmental conditions across Namibia, which consists of a vast network of protected areas that include both multiuse communal conservancies and formally protected national parks. Elephants within Namibia are found from the hyper-arid ecosystems in the west to the flooded grassland, savannah, and woodland habitats in the east. We expand upon past research by analyzing the largest dataset ever recorded (n = 86) on the movements of elephants across Namibian ecosystems. We tested the hypothesis that elephant home ranges correspond to extrinsic environmental factors which vary geographically across Namibia as found in other parts of Africa (Loarie et al., 2009).
We focused on home range size as a core ecological process to assess the space use needs of each population, with inference on the environmental factors that influence variability. While in some ways similar to a resource selection and/or step-selection function analysis (Boyce & McDonald, 1999;Manly et al., 2002;Roever et al., 2012;Thurfjell et al., 2014;Van Moorter et al., 2016), our analysis does not examine individual decisions (the points and turning angles) that animals make and builds upon previous work conducted across regional scales to assess elephant space use (Buchholtz et al., 2019;de Beer & van Aarde, 2008;Roever et al., 2012;Young et al., 2009).
To test our hypothesis, we incorporated Global Positioning System (GPS) telemetry data collected from elephants in four populations between 2008 and 2015. We combined these data with environmental variables, measured from remotely sensed data, to assess how each variable impacts home range size at regional and site-level scales.
Specifically, we tested the relative influence of precipitation, surface water, vegetation, human impact, and the amount of area protected on the variation in elephant home range size. Because of the overwhelming importance of water resources in arid systems (Wall et al., 2013), we hypothesized that differences between populations would primarily be driven by precipitation, while site-level variation would be best explained by the availability of forage resources and extent of anthropogenic footprint (Wall et al., 2021). By determining how natural and anthropogenic factors influence elephant space use at multiple scales, we aim to shed light on conservation successes and areas for concern for elephant management in Namibia.

| Study sites
This study compares data from four sites along an east-west gradient within the arid to semi-arid savanna region of Southern by extension the wettest, site is the Zambezi region of Namibia (henceforth referred to as Zambezi; Figure 1). The site is part of the Kavango-Zambezi Transfrontier Conservation Area (KAZA) and is composed of several national parks, conservancies, and forest reserves, some of which connect to adjacent protected areas in Botswana, Angola, and Zambia. The site receives 607 ± 59 mm of rainfall annually (Funk et al., 2015; Appendix S2: Figure A1). An estimated 12,000 elephants reside in this part of Namibia (Craig & Gibson, 2019a, 2019b. The site has the highest density of humans (6.2 people/km 2 ; Namibia Statistics Agency, 2011) of our study areas and greatest human modification (Kennedy et al., 2019; Appendix S2: Figure A2).
Just west of Zambezi lies Khaudum National Park (henceforth Khaudum; Figure 1). The 3841 km 2 protected area is directly adjacent to the fenced Namibia-Botswana border and receives 540 ± 19 mm of rainfall annually (Funk et al., 2015; Appendix S2: Figure A1). Khaudum shares an open southern border with the Nyae Nyae community conservancy and is estimated to support approximately 8000 elephants (Craig & Gibson, 2019a, 2019b. The human population density in the Kavango region around Khaudum is approximately 4.6 people/km 2 (Namibia Statistics Agency, 2011).
Further west, Etosha National Park is a 22,270 km 2 protected area located in north-central Namibia (henceforth Etosha; Figure 1).
Etosha is a semi-arid savannah with approximately 394 ± 52 mm of rainfall annually (Funk et al., 2015; Appendix S2: Figure A1). The Park has been fenced since the early 1970's and is estimated to support an elephant population of approximately 2900 animals (Kilian, 2015).
There are limited influences from humans within the park, but Etosha borders some regions with high human population densities (>20 people/km 2 ), though most surrounding regions have low population density (<1 person/km 2 ; Namibia Statistics Agency, 2011).
Our driest study site was the Kunene region of northwestern Namibia (henceforth referred to as Kunene; Figure 1). Kunene is arid with much of its area lying within the Namib and pro-Namib desert.
The site receives 209 ± 119 mm of rainfall annually (Funk et al., 2015; Appendix S2: Figure A1), consisting of a patchwork of multiuse conservancies and more restricted concessions that support approximately 1100 elephants (Craig & Gibson, 2016). It has the lowest human population density of our study areas (0.8 people/km 2 ) due to its aridity and limited options for agriculture (Namibia Statistics Agency, 2011). The final dataset across all sites consisted of 3,669,784 GPS fixes spanning 8 years (2008)(2009)(2010)(2011)(2012)(2013)(2014)(2015). For individual elephants, the number of GPS fixes ranged from 5090 to 209,942, with a median value of 36,809 points. The first day of the tracking period was removed from each dataset to eliminate unusual movement behavior caused by collaring procedures (Northrup et al., 2014). A summary of the tracking data is provided in the Appendices S1 and S2.

| Environmental predictors
Landscape information for vegetation, precipitation, surface water availability, protected area designation, and human impact were collected from globally available data layers, and processed in Google Earth Engine (Gorelick et al., 2017). Data from multiple sources and indices were used for each criterion to test which method best quantifies differences between the four sites. Mean and standard deviation values for variables included in resulting models are provided in Table 1 for each site.
We tested three MODIS-derived vegetation indices to quantify vegetation availability and variability: normalized difference vegetation index (NDVI), modified soil-adjusted vegetation index (MSAVI), and fraction of photosynthetically active radiation (FPAR). NDVI measures green biomass and vegetation productivity (Pettorelli, 2013). Because NDVI is less reliable in arid and semiarid areas due to the effect of bare soil (Boschetti et al., 2007), MSAVI was included as an alternative. MSAVI increases the dynamic range of vegetation signals and reduces the influence of soil background to better estimate vegetation in arid habitats (Qi et al., 1994). FPAR is a measure of the proportion of sun radiation received by a plant to the total available photosynthetically active wavelengths of radiation (Knyazikhin, 1999). For arid areas like Namibia, FPAR is expected to be a better predictor of herbaceous biomass as it encompasses both green and dry biomass (Tsalyuk et al., 2015).
The impact of water was examined both in terms of rainfall and available surface water. Precipitation estimates were extracted from the Climate Hazards Group InfraRed Precipitation with Station (CHIRPS) dataset, which estimates daily rainfall at 0.05° resolution (Funk et al., 2015). Annual mean precipitation was calculated from 2008 to 2015 for the study area. Surface-water availability was represented using the JRC Global Surface Water Mapping Layers (30-m resolution), providing data on the location and temporal distribution of surface water from 1984 to 2019 (Pekel et al., 2016). Bands for occurrence (the frequency with which water was present) and seasonality (how many months water is present) were used as variables in our analysis. In addition, we calculated the location of permanent and seasonal water sources by filtering the seasonality layer to pixels where water presence is greater than (permanent) and less than (seasonal) 9 months of the year.
Two data layers were used to represent human impact at 1-km resolution: Human Footprint (HF; Venter et al., 2016) and global Human Modification (HM; Kennedy et al., 2019). HF is an index of human pressures derived from the summation of eight data layers approximately representing human impact for 2009. HM is a metric for the proportion of a landscape that has been modified by humans and based on an existing threat classification system by Salafsky et al. (2008) for 2016. The layers differ significantly in how they are calculated (Oakleaf & Kennedy, 2018) and emphasize different aspects of human impacts (e.g., HF focuses more heavily on roads than HM). Data on road location and road type were included from the Global Roads Inventory Project (GRIP) dataset (Meijer et al., 2018). Protected area designations were derived from UNEP (UNEP-WCMC, 2018).

| Home range and movement
We calculated variograms, fit continuous-time movement models, and estimated home-range sizes using the ctmm package (Calabrese et al., 2016) in R (R Development Core Team, 2020).
Due to the amount of data being analyzed, we fit all models using the Smithsonian Institution's High Performance Computing Cluster (Smithsonian Institution, 2020). We first plotted the estimated semivariance function for each individual to assess the autocorrelation structure of the data (Fleming et al., 2014a). Resulting semivariograms were visually inspected for each animal to determine whether the data reached an asymptote, indicating whether animals met the range residency assumption (Calabrese et al., 2016). The semivariance function's curvature at short time lags was used to indicate whether or not the data could support velocity estimation.
Models were fit using residual maximum likelihood and ranked using AICc . We estimated home ranges conditional on the best-fit model for each individual using auto-correlated kernel density estimation (AKDE) at the 50% coverage level (Fleming et al., 2014b(Fleming et al., , 2018Fleming & Calabrese, 2017). Accounting for autocorrelation in home-range estimation is especially important for elephants, given the recognized underestimation of species area requirements due to the animals' large body size (Noonan et al., 2020).
For comparative purposes with historic range estimates, we also 20457758, 2022, 9, Downloaded

| Statistical analyses
To assess variation in home-range sizes, we compared AKDE estimates at the 50% level using the meta function in ctmm, which estimates population-level parameters from individual-level parameter estimates while taking into account estimate uncertainty . We use this method to compare home range size between sexes and sites. An analysis of variance (ANOVA) was used to determine whether there were significant differences between groups (p < .05).
To summarize the environmental data layers, we used zonal statistics (function exactextractr, Baston, 2019) to calculate the mean and standard deviation within each individual's home-range polygon. The total length of roads and rivers within each home range were estimated and divided by home-range area to standardize estimates across individuals. Percentages of each protected area designation (national park, concession, communal conservancy, and forest reserve) within each home range were also calculated.
To determine which environmental variables drive variation in home-range size across elephant populations, we used generalized linear models (GLM). We eliminated highly correlated variables within each environmental category (vegetation, precipitation, surface water, human impact, and protected area) by first conducting univariate regressions, ranking individual models based on AIC to determine the best variable within each category to incorporate in subsequent analyses. All final variables were evaluated for correlation using a variance inflation factor (VIF) analysis (Hair et al., 1995). Once variable independence was determined, we combined all variables in a multivariate model after log transforming the dependent variable due to significant right skewness. We dredged the resulting model results using the MuMIn package in R to remove weakly correlated variables (Bartoń, 2012). The adjusted R 2 was calculated for the best model to estimate the proportion of explained variance. We conducted this two-step process because incorporating all variables into one model proved computationally difficult.
To determine the environmental drivers that predict homerange variation locally, we subset the data into the four sites and conducted separate GLM's with the variables from the full model for each site. These models were also dredged to determine the most parsimonious models. Sex was included as a variable in the Kunene and Etosha models, where data were available. The single male from Zambezi was removed from the site analysis models, representing a limitation of our dataset. Tracking period in days was initially TA B L E 1 The mean and standard deviations (in parentheses) of each environmental variable calculated using polygons derived by the combined 99% AKDE home ranges for individuals at each site included as a covariate in the GLM's but was removed as it showed no effects in the models.
We exported the probability mass function (PMF) calculated from each resulting AKDE home range to provide per pixel percentages of use in different habitats. We summed the PMF within protected areas boundaries to differentiate the percentage of space use between protected and nonprotected areas, and with a particular emphasis towards evaluating the percentage of space use within national parks. All analyses were completed using the R environment for statistical computing (Version 4.0.2; R Development Core Team, 2020).

| Variation in home range size
The mean 50% AKDE home range estimate was 2200 km 2 (95% CI: 1500-3100 km 2 ) for all elephants included in the study.

| Environmental predictors for regional variation
Of the five variables included in the full model, precipitation had the greatest effect on home range size (β = 0.71, SE = 0.11). The variation of annual precipitation demonstrated a strong positive correlation with home range size, meaning years with high rainfall variability were correlated with larger elephant home ranges (Figure 3). Neither variation in vegetation productivity (β = 0.18, SE = 0.11), the occurrence of surface water (β = 0.11, SE = 0.10), nor human modification had a significant effect with home range size (β = 0.15, SE = 0.10; Figure 3). National parks demonstrated the greatest correlation with home range size out of all the protected area designations (e.g., communal conservancies) and concessions. Lastly, there was a negative relationship between percentage of home range in a national park and home range size, but these results were nonsignificant (β = −0.14, SE = 0.11; Figure 3).

Our most parsimonious model included only precipitation and
vegetation (Table 3). These two variables explained approximately 53% of the regional variation in home range size. Home range size increased significantly with both variability in rainfall (β = 0.76, SE = 0.09) and vegetation (β = 0.28, SE = 0.09), though precipitation contributed more heavily to the trend.

| Environmental predictors for site variation
The variables that best explained home range size site variation differed between study sites. The most parsimonious model for Kunene included precipitation, surface water, human impact, and sex. This model explained approximately 82% of the variation in home range size across the site (Table 3). The variables with the greatest effects were precipitation (β = 0.60, SE = 0.10) and surface water variability (β = 0.38, SE = 0.08). Greater variation in human modification was positively correlated with home range size (β = 0.20, SE = 0.10), as was sex (β = 0.35, SE = 0.17; Table 3). The best model for Etosha included only precipitation, which explained approximately 74% of home range variation in home range size (

| DISCUSS ION
Elephant home ranges in northern Namibia varied widely across local and regional scales. Our results highlight, however, that much of this variation can be explained by a few key environmental variables. On a regional scale, we found that precipitation and vegetation explained 53% of the variation in home range size. At the site level, home range differences were also influenced by the interac- Our MCP estimates were smaller on average in Kunene and Etosha than previously published range sizes (Lindeque & Lindeque, 1991;Leggett, 2006aLeggett, , 2006b; Table 2), but similar in that our estimates varied widely between individuals from the same site (i.e., Etosha: 240-13,000 km 2 ). Our 95% AKDE estimates were more than double MCP estimates for the same individuals ( Table 2). This is consistent with studies which have compared AKDE to traditional metrics (Moßbrucker et al., 2016;Noonan et al., 2020). Our AKDE and MCP estimates were also larger than local convex hull estimates for the same populations from Roever et al. (2012), which is consistent with comparison of these methods in Noonan et al. (2019).
Despite high individual variation, regional differences in home range size were clearly correlated with precipitation and NDVI, which we anticipated given Namibia's pronounced rainfall gradient (Appendix S2: Figure A1). This finding is consistent with other studies of megaherbivores in Africa (e.g., Knüsel et al., 2019), in which mean annual rainfall explained 74% of the variation in giraffe  (2010), for example, found that the daily displacement distance of elephants decreased with increased rainfall across 13 southern African study sites (including Etosha), while Grogan et al. (2020) found that annual precipitation was the only variable found to be negatively correlated with annual home range size.
While we did not hypothesize that vegetation would be important regionally, the inclusion of NDVI in the best regional model is not surprising. High-quality vegetation is known to be an important grazing resource for herbivores, which impacts their space use (McLoughlin & Ferguson, 2000;Tufto et al., 1996;van Beest et al., 2011). Across our study sites, vegetation productivity follows a similar West-East gradient in relation to the precipitation gradient in Namibia. Elephants are known to be particularly adept at seeking out highly productive patches of vegetation throughout the year (Loarie et al., 2009). Other studies, however, have found that elephant movements cannot be solely attributed to vegetation productivity, with individuals having complex foraging strategies which are not uniform in space or time (Boettiger et al., 2011). This may explain why elephants prefer areas with higher landscape heterogeneity ( (Schnegg & Kiaka, 2018). The cost/benefit ratios are highly variable between conservancies with some experiencing large profit margins while others suffer disproportionate losses from human-wildlife conflict (Brown, 2011). While Kunene elephant numbers have increased overall in the past decades (Schnegg & Kiaka, 2018), evidence exists that elephants in some Kunene conservancies experience higher levels of stress and potentially lower calf recruitment compared with those in Etosha (Hunninck et al., 2017).
Anthropogenic disturbances have caused significant changes in vegetation structure and composition in Kunene, which will only further degrade the landscape without intervention (Inman et al., 2020). Similar results were found for giraffes (Giraffa camelopardalis) whose home range sizes were negatively correlated with distance to densely populated towns (Knüsel et al., 2019). Our findings also support previous studies which specify Zambezi as an area of high human-wildlife conflict with restrictions on animal movement (Stoldt et al., 2020). Despite high human modification, relatively high elephant numbers are sustained. Occupancy of this area by elephants and other large mammals has increased in recent decades but is more heavily constrained and fragmented by agricultural expansion and fences (Stoldt et al., 2020). Existing corridors should be carefully monitored, maintained and protected to preserve connectivity in the face of human pressures (Brennan et al., 2020). This is especially key because Zambezi sits at the heart of the KAZA Transfrontier In conclusion, our results demonstrate the variation in drivers of elephant range size across ecological and human modification gradients. In arid sites, which tended to have larger home ranges and lower human density, human activity became more influential to recorded range sizes. In the highest human density area, human activity was the sole correlate of elephant range size in our top model. Interestingly, home range estimates of elephants have not altered drastically from estimates 30 years ago. A key concern going forward is the interaction and competition for space between growing human and elephant populations. Our results highlight the critical role government-and community-run protected areas play in the current Namibian elephant distribution. Maintaining healthy populations of this wide-ranging megaherbivore is no easy feat, but Namibia's success should be acknowledged in the face of continentwide declines of this endangered species. writing -review and editing (equal).

ACK N OWLED G M ENTS
We thank the Namibian Ministry of Environment, Forestry and

CO N FLI C T O F I NTE R E S T
The authors declare that there is no conflict of interest.

DATA AVA I L A B I L I T Y S TAT E M E N T
In view of poaching concerns, the datasets generated or analyzed during the current study are not publicly available but can be made available upon reasonable request via Movebank (ID: 1807299477).