Understanding the effect of landscape fragmentation and vegetation productivity on elephant habitat utilization in Amboseli ecosystem, Kenya

Abstract

Understanding factors affecting the distribution of the African elephant is important for its conservation in increasingly human-dominated savannah landscapes. However, understanding how landscape fragmentation and vegetation productivity affect elephant habitat utilization remains poorly understood. In this study, we tested whether landscape fragmentation and vegetation productivity explain elephant habitat utilization in the Amboseli ecosystem in Kenya. We used GPS (Global Positioning System) telemetry data from five elephants to quantify elephant habitat utilization. Habitat utilization was determined by calculating the time elephants spent within a unit area. We then used generalized additive models (GAMs) to model the relationship between time density and landscape fragmentation, as well as vegetation productivity. Results show that landscape fragmentation and vegetation productivity significantly (P < 0.05) explain elephant habitat utilization. A significant (P < 0.05) unimodal relationship between vegetation productivity and habitat utilization was observed. Results suggest that elephants spend much of their time in less fragmented landscapes of intermediate productivity. Bien comprendre les facteurs qui affectent la distribution de l'éléphant d'Afrique est important pour sa conservation dans des paysages de savane de plus en plus dominés par les hommes. Pourtant, on ne comprend pas encore très bien comment la fragmentation des paysages et la productivité de la végétation affectent la fréquentation du paysage par les éléphants. Dans cette étude, nous avons voulu savoir si la fragmentation du paysage et la productivité de la végétation expliquent la fréquentation de l'habitat par les éléphants dans l'écosystème d'Amboseli, au Kenya. Nous avons utilisé des données télémétriques GPS de cinq éléphants pour quantifier cette fréquentation de l'habitat. Celle-ci était déterminée en calculant le temps que les éléphants passent dans une unité de surface. Nous avons ensuite utilisé des Modèles Additifs Généralisés (MAG) pour modéliser la relation entre la densité de temps et la fragmentation du paysage, ainsi que la productivité de la végétation. Les résultats montrent que la fragmentation du paysage et la productivité de la végétation expliquent de manière significative (P < 0.05) la fréquentation de l'habitat par les éléphants. Nous avons observé une relation unimodale significative (P < 0.05) entre la productivité de la végétation et l'utilisation de l'habitat. Nos résultats suggèrent que les éléphants passent une grande partie du temps dans des paysages moins fragmentés où la productivité est intermédiaire. Understanding factors affecting the distribution and habitat use of the African elephant (Loxodonta africana) is key to its conservation in increasingly human-dominated savannah landscapes (Murwira & Skidmore, 2005; Evans & Adams, 2016). Although foraging resources have been widely hypothesized to be the major factor influencing elephant movement (Loarie, van Aarde & Pimm, 2009; Birkett et al., 2012; Chiyo et al., 2014), human-induced landscape fragmentation is also considered important (Blake et al., 2008; de Boer et al., 2013). Thus, selection of foraging areas by elephants has to take into account both foraging resources and landscape fragmentation. In this regard, understanding the combined effect of forage resources and landscape fragmentation on elephant movement is critical for predicting and managing species response to natural and anthropogenic changes in the landscape (Buij et al., 2007). The selection of foraging resources by large herbivores is primarily influenced by plant phenology which is closely linked to the nutritional quality/digestibility of the forage (Fryxell, 1991). Forage quality is known to decline with plant growth, and as the plant matures, there is also a reduction in forage digestibility due to changes in cell wall composition over time (Hebblewhite, Merrill & McDermid, 2008). Moreover, intake declines in low forage areas because of low net energy intake. In this regard, herbivores are known to prefer areas of intermediate forage productivity as they face forage quality–quantity trade-offs (Fryxell, 1991; Wilmshurst, Fryxell & Hudsonb, 1995; Hebblewhite, Merrill & McDermid, 2008) . Intermediate forage productivity landscapes are vegetation communities that are on the transition from open grasslands (low forage biomass m−2 − low intake rates) to woodlands or thicker vegetation (high forage biomass m−2 poor forage quality) (Hebblewhite, Merrill & McDermid, 2008). Elephants being bulk feeders are less likely to be influenced by forage quality (Owen-Smith & Chafota, 2012). Instead, elephants maximize their energy intake by making a trade-off between scarce high-quality forage and abundant low-quality forage (Bergman et al., 2001). It is therefore not known whether elephants respond to vegetation productivity in a similar manner to other herbivores. Herbivores do respond not only to the variability in forage resources but also to the patchiness of forage resources (Murwira & Skidmore, 2005). It therefore becomes important to understand the response of herbivores to forage resources and habitat fragmentation (Groom & Western, 2013). The existence of human infrastructure termed ‘fragmentation geometries’ such as roads, settlements and agricultural fields not only impede animal movement but also fragment their habitats (Burnsilver, Worden & Boone, 2008; Western, Groom & Worden, 2009) . However, studies on large herbivores particularly elephant movement and habitat utilization have often considered different fragmentation geometries in isolation (Barnes et al., 1991; Blake et al., 2008). Habitat fragmentation is assumed to negatively influence elephant habitat utilization (Leimgruber et al., 2003). We therefore expect elephants to spend more time in less fragmented habitats compared to fragmented habitats. Thus, any meaningful prediction of animal movement in response to fragmentation geometries should consider their combined effect. It therefore becomes important to understand not only animal response to spatial and temporal changes in forage resources but also their response to landscape fragmentation. Hence, objective quantification of these factors becomes important in providing improved insights into elephant distribution and habitat utilization. Remote sensing could thus provide a meaningful approach to objectively quantify landscape fragmentation as well as forage resources due to its ability to provide a synoptic view of the landscape. The advancement of satellite remote sensing has allowed for the quantification of forage resources at large spatial extents and at high temporal resolution. To this end, spectral transform and derivatives such as vegetation indices have been developed that correlate with vegetation productivity and quality. For example, the satellite-derived normalized difference vegetation index (NDVI) has been used as a surrogate for forage greenness or abundance in explaining elephant movement (Loarie, van Aarde & Pimm, 2009; Matawa, Murwira & Schmidt, 2012). Although vegetation indices are useful as proxies of productivity, they are not able to account for short-term variations in productivity resulting from, for example changes in meteorological conditions (Monteith, 1972; Pachavo & Murwira, 2014). In this regard, there is need for the development of measures of forage resources that are sensitive to changes in meteorological conditions, that is temperature and humidity. The recent development of remotely sensed dry matter productivity (DMP) has allowed direct and precise estimates of productivity (Xu et al., 2012). DMP is proportional to net primary productivity and measures the growth rate of vegetation (dry mass increase) (Copernicus, 2013). The use of DMP as a direct measure of productivity could therefore improve understanding the link between vegetation productivity and animal movement (Xu et al., 2012). However, the application of DMP to wildlife studies has received limited attention despite its ability to provide a more direct measure of vegetation productivity. In this study, we test whether and how seasonal habitat utilization of GPS collared African elephant is influenced by landscape fragmentation and remotely sensed DMP. We hypothesize that elephants would spend more time in habitats with intermediate vegetation productivity while utilizing habitats that are less fragmented. The Amboseli ecosystem is located in Kajiado District, Rift Valley Province, Kenya (Fig. 1). The ecosystem covers an area ~ 8500 km2 (Burnsilver, Worden & Boone, 2008). The area is classified as arid to semi-arid savannah with an average annual rainfall of 340 mm. Rainfall is bimodal with most rains being received from March to April and from November to December (Tuqa et al., 2014). The area also experiences a dry period (June to September) and two transition periods (January–February and October–November) (Altmann et al., 2002). The transition season is the period of change from wet to dry season or vice versa. During the transition period, the mean month rainfall is low but rising or declining. Temperature ranges between 20°C and 30°C, while elevation varies between 850 and 1350 m above mean sea level. The dominant vegetation types include the broadleaf woodlands and dry tropical forests on the Kilimanjaro and Chyulu slopes, open grassland, riverine forest and scrubland in the Amboseli Basin as well as scattered Commiphora and Acacia woodlands (Western, 2007; Howe, Okello & Davis, 2013). The elephant population in Amboseli is estimated at about 1400 individuals (Chiyo et al., 2011). A total of five elephants were captured and fitted with GPS collars in Amboseli between 18th February and 15th March 2013 by Kenya Wildlife Service (KWS) and International Fund for Animal Welfare (IFAW). Of the five elephants, one was female and the rest were males and all the elephants belonged to different families (Table 1). The five GPS collared elephants represent five different elephant herds. Elephant herds can change numbers and composition, and examination of our tracking data indicated that these five collared animals belonged to distinct groups throughout the monitoring period. The fitted GPS collars were programmed to log the position of each individual after every 4 h for the period 20 February 2013 to 31 August 2013 resulting in a total of 4872 GPS fixes. The GPS collars had a relative error of 10 m which is marginal considering the minimal home range size of 10 km2 an adult elephant requires to meet its daily requirements in African open savannahs (Douglas-Hamilton, Krink & Vollrath, 2005). The GPS collars had a success fix rate ranging between 93% and 96.8% which is within acceptable range to characterize wildlife movement patterns and make sound inference (Frair et al., 2010). The GPS data were captured in geographic coordinates. The geographic coordinates were then reprojected in ArcGIS GIS 10.1 (ESRI, 2011) to Universal Transverse Mercator (UTM) Zone 37 based on WGS 84 Spheroid. Next, we split the GPS location data for each elephant into three seasons, that is wet season (March and April), dry season (June to August) and transition season (February and May). This was based on the premise that seasonal variation in resources such as forage and water influences the distribution and movement patterns of the elephants differently. We used 10-day DMP data with a spatial resolution of 1 km freely downloaded from the Flemish Institute for Technological Research (VITO) (www.vito.be) as a proxy of vegetation productivity. DMP is directly proportional to net primary production (NPP). DMP is derived from SPOT-VEGETATION sensor on board of SPOT-4/5 satellite. DMP is a proxy for dry matter biomass increase, that is vegetation growth rate expressed in kilograms of dry matter per hectare per day (kg DM−1 ha−1 day−1) (Copernicus, 2013). It is directly proportional to net primary production (NPP), thus a measure of forage available for elephants. The DMP is derived by combining satellite data with meteorological data (solar radiation and temperature) following the classical Monteith approach (Monteith, 1972). The DMP data covered the period February 2013 to August 2013. We calculated the mean DMP for the three seasons based on the dekadal data using ENVI IDL (ITT Visual Information Solutions, 2009). Next, we extracted the mean DMP for each grid that coincided with an elephant GPS location using the average of a 3 × 3 km window. A 3 × 3 km window was selected because it is the minimum home range size required by an adult elephant (Douglas-Hamilton, Krink & Vollrath, 2005). Mean DMP was used in this study because it demonstrated to be highly correlated to elephant time density. To quantify landscape fragmentation, we first identified human infrastructure that leads to the subdivision and isolation of elephant habitat (Jaeger, 2000; Girvetz et al., 2008). Infrastructure considered were roads, agricultural fields, towns and human settlements. Data on human settlements were extracted from the Kenyan Wildlife Services (KWS) database. We verified the accuracy of KWS settlement geodatabase by overlaying the settlements on Google Earth (www.googleearth.com), and missing settlements were also digitized. The agricultural field layer was classified from MOD13Q1 MODIS NDVI (250 m) data set. The 16-day MODIS NDVI images for the period from August 2010 to August 2013 were downloaded from the USGS EROS Data Center (http://lpdaac.usgs.gov/). The NDVI data were then reprojected from the sinusoidal projection to UTM zone 37 based on WGS 84 Spheroid in ENVI 4.7 (ITT Visual Information Solutions, 2009). Prior to classification, we reduced noise in the NDVI images caused by remnants of clouds using a Savitzky–Golay filter (Jonsson & Eklundh, 2004) using the TIMESAT package. We used the maximum-likelihood classification method and 15 ground truth data to classify the NDVI images into three broad land cover types, that is agricultural fields, water and nonagriculture. The nonagriculture class was composed of bare ground, wooded grasslands, shrubland, woodland and riverine woodland. The overall classification accuracy using 42 test ground control points was 85% (kappa = 0.68). Fragmentation geometries influence elephant behaviour within a certain distance. Thus, areas close to landscapes used by humans become unavailable for elephants. We thus created a buffer of 500 m for roads (Blake et al., 2008) as well as settlements (Harris et al., 2008), while for towns, a buffer of 4 km from the town centre was created (Harris et al., 2008). Prior to testing whether time density was influenced by DMP and landscape fragmentation, we first randomly selected sample grid cells from the time density data layer in a GIS. Next, we tested the sampled data for spatial autocorrelation using Moran I (Tiefelsdorf, 2002) as a way to check for spatial independence. For all seasons, the time density data did not show any spatial independence (P < 0.05). The data were then randomly selected at increasing distance until spatial autocorrelation was no longer detected, that is at 6 km. Thus, for the dry season, a total of 60 grid cells were randomly selected, while for the transition and wet season, 47 grid cells selected, respectively. We then extracted DMP values and effective mesh size values that spatially coincided with the randomly selected time density sample grids using overlay analysis in a GIS. Next, we tested whether variation in time density could be explained by productivity and landscape fragmentation during different seasons. To achieve this, we used generalized additive models (GAMs) that allow for both linear and nonlinear responses (Venables & Dichmont, 2004). GAMs are able to identify nonlinearities using flexible nonlinear modelling approaches that utilize spline smoothing (Venables & Ripley, 1999). Time density was treated as the response variable, while mean DMP and effective mesh size were covariates in the models. Thus, we considered seasonal models which included either productivity or effective mesh size as well as models that included both covariates. The models were fit using the mgcv package in R (R Development Core Team, 2012). Prior to model development, we checked for multicollinearity between the covariates using the variance inflation factor (VIF) (O'Brien, 2007; Dormann et al., 2013). We included both covariates if VIF<10. In all cases, the VIF between DMP and effective mesh size for also seasons was less than 10 indicating that collinearity did not exist between the explanatory variables. Selection of the best model was based on the corrected Akaike information criterion (AICc) adjusted for small sample size (Burnham & Andersen, 1998). We retained models with the lowest AICc score. Although our model combinations resulted in <10 candidate models, we only report models that are within ΔAIC≤10 as these have substantial support (Burnham & Anderson, 2002). Retention of competing models within 2 ΔAIC was based on the significance of the additional parameter/s (Arnold, 2010). In addition, we assessed the relative of each model by considering the Akaike which measure the that model is the best the of models (Burnham & Andersen, 1998). with than are considered as candidate models providing the best fit et al., 2009). During the wet season, the most model that explained elephant time density based on AICc included landscape fragmentation as the only (Table Elephant time density with a in landscape fragmentation = < (Fig. However, the additive model with vegetation productivity and landscape fragmentation as covariates had a < 2 that the model was also in explaining elephant time density. However, we selected the model with effective mesh size as the of vegetation productivity in the model did not improve model = = For the transition season, the most model explaining time density of elephants based on AICc vegetation productivity and landscape fragmentation (Table A nonlinear relationship is between time density and productivity = < during the transition season (Fig. while time density with a in fragmentation = < (Table It is to that time density is high at intermediate of productivity and at high effective mesh size fragmented (Fig. For the dry season, the most model that explained the time density of elephants both the of productivity and effective mesh size (Table A nonlinear relationship was also between time density and productivity = < (Fig. while time density with an in effective mesh size = < (Table Time density is high at intermediate of productivity and at high effective mesh size (Fig. Results from this study suggest that habitat utilization by elephants in human-dominated landscapes is most explained by a of landscape fragmentation and vegetation productivity during the dry and transition seasons, than each of the factors (Table However, during the wet season, fragmentation was the most important The is not as forage quality and are known to be high during the wet season, and forage is during the wet season et al., 2012; & Murwira, 2014). This is also with behaviour of the African elephant in human-dominated landscapes elephants often their from humans by more time in than in (Barnes et al., 1991; et al., 2009). In addition, this is with other studies (Fryxell, 1991; et al., de Boer et al., that have a by herbivores for areas with intermediate For example, a high of herbivores at intermediate of biomass in the We thus a of landscape fragmentation and productivity our of habitat utilization in human-dominated landscapes (Fig. To the best of our studies have not the combined of fragmentation and productivity in a model to explain The relationship between elephant time density and effective mesh size is with known elephant habitat that is that elephants to spend more time in than For example, et a relationship between human infrastructure and elephant density. In this study, we used an landscape fragmentation that is the effective mesh size landscape to the effect of fragmentation on elephants, and the with is of elephant response to fragmentation. Thus, we that the effective mesh size landscape meaningful of landscape fragmentation useful for predicting elephant The unimodal or relationship between DMP and time density of elephants in this study has been for herbivores other than elephants. For example, de Boer et a unimodal relationship on the relationship between NDVI and elephant that herbivores often make a trade-off between quality and areas of low productivity low habitat use & as they not forage while areas of high productivity are limited by plant and low forage quality (Fryxell, 1991; et al., 2008). van de & also that high productivity areas are less likely to be used due to of This is also with van Aarde & that elephants areas with high productivity the relationship between time density and productivity in this study the of DMP in understanding To the best of our satellite remote DMP has not been used to understand wildlife habitat utilization. The objective of this study was to test whether landscape fragmentation and vegetation productivity explain habitat utilization of African elephants. on the we that elephants spend much of their time in landscapes that are less fragmented and by intermediate vegetation productivity. We also that remotely sensed DMP can be used to explain Results of this study that if the of the African elephant is to be large areas of intermediate vegetation productivity need to be within savannah landscapes. We are to the Kenya Wildlife Services (KWS) for to this study in the Amboseli ecosystem and to elephant GPS tracking fragmentation and their during We also the International Fund for Animal Welfare for providing support for elephants. We also the for at and The Amboseli for Elephants for their