The African large predator guild is one of the last intact large predator guilds globally, and interactions between its members influence ecosystem functioning. We conducted camera-trapping in the Maasai Mara Ecosystem (MME) to estimate leopard (Panthera leo) population density and investigate whether lion (Panthera pardus) and hyaena (Crocuta crocuta) impact leopard presence, while accounting for potential prey presence, and habitat. In 2019, we deployed cameras at 34 stations in the Mara Triangle within the MME for 63 nights. We estimated leopard density using a closed population spatially explicit capture recapture (SECR) framework and examined potential predictors of leopard presence using generalised linear mixed modelling. We recorded 725 leopard images and estimated population density at 1.90 ± 0.56 individuals 100 km² ⁻¹; relatively low compared to other areas and only slightly higher than previous MME estimates of cheetah, an ecologically subordinate competitor. The best model predicting leopard presence contained hyaena occurrence and showed a positive association, indicating “co-occurrence”. Hyaenas commonly kleptoparasitize leopard kills in MME, i.e., hyaenas may follow leopards for this reason. Although our preliminary results indicate that hyaena populations may limit leopard populations in the MME, further work is required to explicitly test hypotheses relating to hyaena-leopard interactions.

This study was conducted between April to June 2019 in the Mara Triangle conservancy (hereafter “the Triangle”) of south-west Kenya (1°24'23.58" S 34°54'23.58"). The Triangle comprises 510 km^2 ^of the MME, the latter also including the Maasai Mara National Reserve (MMNR), and six additional conservancies totalling approximately 2398 km² (Broekhuis and Gopalaswarmy, 2016). The dominant habitat is primarily open grassland interspersed with Vachellia gerrardii and Terminalia trees, along with shrub and riverine forest (Bhola et al., 2012). The Triangle is broadly grouped into open and closed habitats, including grasslands, wetland and balanite zones, Acacia and riparian woodland, and thickets (Table 1; Broekhuis et al., 2017).

Camera-trapping

We deployed 34 camera-trap stations for 63 consecutive nights between 22 April and 25 June 2019.  We considered this period long enough to optimize the probability of “capturing” leopards without violating closure assumptions inherent to the use of closed population capture-mark-recapture (CMR) models (Karanth and Nichols, 1998; Devens et al., 2018). Camera-trap stations (hereafter ‘stations’) were established in areas of likely leopard presence as determined by a variety of factors including prior knowledge, proximity to features commonly selected by leopards (e.g., permanent water sources or “prey trails”) (Karanth and Nichols, 2002), or the confirmation of additional physical evidence (e.g., scratch marks on trees, spoor, scat, territorial scent; Devens et al., 2018; 2021). Although leopards utilise road networks (du Preez et al., 2014; et al., 2015b), stations were purposefully not placed along roads to reduce their visibility to visitors. Each station comprised a pair of passive infrared triggered cameras of the following makes/models: Browning^TM Strike Force Pro X 1080 (One Browning Place, Morgan, UT 84050infra-red low-glow LED) [30 stations]; Browning^TM Dark Ops HD Pro X 1080 infra-red low-glow LED [1 station]; Spypoint^{TM Solar Dark (GTX 330, de la Jacques-Cartier Victoriaville, Quebec, Canada G6T 1Y3) [1 station]; and Bushnell}TM NatureView HD (Bushnell Outdoor Products, 9200, Cody Overland Park, KS 66214) [2 stations]. All Browning cameras additionally used low-glow LED to minimise potential disturbance to animals.

Camera-traps paired to a station were placed on opposite sides of a trail or path to capture both flanks of a passing leopard and improve the probability of individual identification; they were further angled offset from each other to prevent the flash of one camera from overexposing images recorded by the other (du Preez et al., 2015). Each camera was positioned approximately 2-6 m from where a leopard was expected to pass (Chapman and Balme, 2010), placed in protective metal cases to reduce the likelihood of theft or damage, and secured 30 to 70 cm above the ground to natural features (e.g., trees) or custom-made posts. Cameras were set to record images as three rapid-fire photographs in succession when triggered, with a minimum of one minute interval between subsequent “triggers”. As leopards are largely crepuscular, each trapping occasion was defined as the twenty-four-hours beginning and ending at noon on successive days (i.e., one “trap night”) (Karanth and Sunquist, 2000; Ramesh et al., 2009). Habitat within 25 m of each station was categorised as either open or closed (Table 1) based on digitised maps in ArcGIS 10.3.1. subject to ground verification. As small patches of thicket and riverine woodland along the Mara River represented the only “closed habitats” in the MME, the remaining “open habitats” consisted either of pure grassland, or grassland with scattered trees.

We followed Karanth and Nichols (2011) in using information regarding minimum leopard home ranges to determine the maximum spacing of stations, and thus avoid sampling gaps that could underestimate population density. In the absence of a specific published home range estimate for East African savannah habitats, we used the smallest home range of a female leopard recorded more broadly in the literature; this being 10 km² from Matopos, Zimbabwe (Smith, 1978) and supported by Parker et al., 2023 (from sites across Africa). A grid with cell size 5 km² (half home range diameter) was overlaid on the target area (see du Preez et al., 2014) to establish spacing criteria of approximately one station every two km (range 0.96 – 4.6) given the availability of suitable sites.  This spacing is similar to that used by Devens et al., (2021) (3-5 km) to allow even coverage of the sampling area and for avoiding gaps.

Leopard identification, presence, and detection histories

We identified individual leopards from photographs by their unique pelage spot patterns, facial scarring, and/or any other distinctive features (Swanepoel et al., 2015b; Fig. 1). Sex was determined by size, the presence of dewlaps and/or external genitalia (Balme et al., 2012) and/or a reference database, the latter particularly when individuals were only partially photographed, or captured from one angle. A detection history was then constructed at each station, consisting of binary values [‘1’ = presence; ‘0’ = absence (Otis et al., 1978)]. Each trap night was considered a single sampling occasion and each station an independent site (Linkie et al., 2007). Following Jenks et al., (2011) in estimating catch per unit effort, we calculated a Relative Abundance Index (RAI) per 100 trap nights by multiplying the total number of leopard detections by 100, and dividing this by the total number of trap nights.

Assessment of prey and intraguild competitors

We recorded the presence of lion, cheetah, and hyaena from photographs taken by camera-traps and used these as binary variables (i.e., “1” or “0” for presence and absence respectively). We similarly recorded the presence of seven species identified as potentially important leopard prey based on recent literature (Hayward et al., 2006a; Balme et al., 2019). We also used binary variables (1,0) to indicate the presence and absence of one or more of these species, to make inferences about the role of prey occurrence on leopard presence. Potentially important prey species were impala (Aepyceros melampus), Thomson’s gazelle (Eudorcas thomsonii), common duiker (Sylvicapra grimmia), Kirk’s dik-dik (Madoqua kirkii), bohor reedbuck (Redunca redunca), blue wildebeest (Connochaetes taurinis) and warthog (Phacochoerus africanus). 

Leopard population density estimation

To estimate leopard population density, we used a closed population spatially explicit capture recapture (SECR) framework in the secr.fit command of package secr 4.5.5 (Efford, 2022) within the R programming environment [R 4.1.3] (R Core Team, 2020), which is considered more reliable than traditional CMR methods (Borchers and Efford, 2008; Thapa et al., 2014; Broekhuis and Gopalaswarmy, 2016). We used a maximum likelihood approach to estimate population size by modelling animal movement and detection probability via estimation of two real parameters: the detection probability (, which declines with distance, the movement parameter sigma (and one derived parameter – Density () (Kalle et. al., 2011; Efford, 2022). We first tested the assumption of closure using a Stanley and Burnham closure test (Stanley and Burnham, 1999) in CloseTest (Fort Collins Science Center, US Geological Survey) before running secr models, which supported the assumptions of closure for our data (ꭕ₂= 12.125, df= 26, p = 0.990).

To estimate leopard population density, our stations (points) served as proximity detectors allowing for unrestricted movement, and multiple captures of the same individuals during the same capture occasion (Efford, 2022). We assumed a Poisson distribution and compared model fits between two commonly used detection functions: half-normal (), and negative exponential We also used AIC_c to control for small sample sizes. First, we fit the null model ( 1) as follows: because grass length in the Maasai Mara increased during the wet season (April to July), we expected leopard detection to decrease linearly ( (model 1). We then selected the best option for model 1 using the lowest AICc~, and compared this to model 2, which incorporated capture heterogeneity due to variations in behavioural response ( (model 2). This is because the “capture” of individual leopards may have varied for any number of reasons, including inherent behavioural differences, or sensitivity to human presence or activity. These models were contrasted with model 3, which combined the two previous models ((model 3), i.e., linear changes in detection over time and capture heterogeneity. We initially chose a buffer width of 2500 m to represent the average interval between stations, and then retrospectively fit a buffer using the function fit(suggest.buffer) to derive one that was wide enough to prevent truncation bias, but not too wide as to overestimate population density (Efford, 2022). This technique produced a buffer width of 9 km, a distance consistent with the wide-ranging nature of at least some resident leopards in our study area (SI, Figure 1). We further refined the buffer width by (a) running an effective sampling area (ESA) plot; and (b) by following Kalle et al., (2011) in varying buffer widths at intervals of 1 km in descending increments from 9 to 4 km until estimates stabilised at approximately 6km (SI, Figure 2). Thereafter, we used 6 km as our buffer for all subsequent analyses.

Predictors of leopard occurrence

Due to the relatively low proportion of locations with positive leopard presence, we decided occupancy modelling analytical frameworks were not appropriate for this study. Therefore, we modelled predictor variables associated with leopard presence on detection using standard generalised linear mixed models (GLMMs) in R (version 4.0.3; R Core Team, 2020) with Laplace approximation. These were then fitted using the ‘lme4’ package (Bates et al., 2012) and a binomial distribution with a logit link function. Next, we constructed a “global” mixed model that included all hypothesised variables (Table 1) as fixed effects and included location as a random effect that accounted for multiple trap days at the same camera stations, i.e., LEO_{LION+CHE+PREY+HAB+(1|LOC).  Models containing CHE were discarded due to the lack of variance associated with our recording of only two cheetah occurrences. We compared the subsequent global model (LEOLION+PREY+HAB+(1|LOC) and the intercept-only model (LEO1+(1|LOC) with each other, and with all nested iterations of the global model using the ‘dredge’ function in the ‘MuMIn’ package (Barton, 2016). Using the ‘arm’ package (Gelman et al., 2020), we then scaled the fixed effects in each model (where mean = 0, and standard deviation = 0.5) to standardise coefficients for direct comparison. We evaluated multicollinearity using variance inflation factors (VIF) from the ‘car’ package (Fox et al., 2012) and ranked models using the Akaike information criterion for small sample sizes (AIC}c~), where models with AIC_c < 2 were included in the best supported candidate model set (Burnham and Anderson 2002). We calculated model-averaged β-coefficients across all competitive models in this set (i.e., the top models), for which we derived 95% confidence intervals. We concluded that predictor variables were important in influencing leopard presence where the associated 95% confidence intervals of their averaged coefficients did not overlap zero.