Prey population estimation

We conducted prey population surveys of the top four prey species: wild boar (Sus scrofa), bezoar goat (C. aegagrus), urial (O. vignei) and small livestock, which contributed over 85% of PO in leopard diet (see above). There was insufficient data on abundance of other potential prey species, which occur seldom in leopard diet in our study area. Because of variations in detection probability of prey due to different habitat structure or species ecology, we applied different population estimation methods for each species (see below). We used a stratified random sampling approach for estimation of wild prey abundance.

As there is no reported seasonal migration of wild prey to or from GNP, temporal availability of wild prey was considered as constant. There is no significant difference in the overall diet of leopards between steppe and forest parts of the park [29] and the ranging pattern of leopards in GNP is not restricted to a specific habitat [24]. Therefore, variations in spatial availability of wild prey were not considered as well. However, we acknowledge that there are subtle variations in spatiotemporal availability of wild prey within ranges of different leopard individuals, which we were unable to measure. On the other hand, as domestic prey is available to leopard predation only when grazing outside villages and conflict cases are spatially explicit [37], we considered the spatiotemporal availability of livestock (see below).

We applied Distance sampling using line transects to estimate the density of urial in GNP [38]. A detailed description of the urial line transect sampling design and modeling is provided elsewhere [25]. We sampled a 340 km² steppe area of the park as the main urial habitat by surveying 17 transects (Fig 1). We used Distance 6.0 software [39] for estimation of urial abundance.

We used the random encounter model (REM) for estimation of wild boar abundance [40], using camera-trapping data from January to December 2011 provided by the Persian Wildlife Heritage Foundation [24]. As we attempted to set up camera traps with a minimum distance of 2 km apart in areas favored by leopards across the park (i.e. wherever leopard signs such as scrapes or scats were detected), we assumed that these locations are random to wild boar movements [41]. Our camera trapping procedure has been described in the literature before [24]. Estimation of wild boar density using the REM method incorporates the number of independent photographic events of the species (y), total camera trapping effort (t), average daily movement of the species when active (v) and average group size (g), as well as camera trap-related parameters such as detection distance (r) and angle (θ) using the following equation [40]:

We retrieved r and θ values from the published literature [40] that used the same brand of camera traps (Deercam DC300; Non Typical Inc., Wisconsin, USA) at 12 m and 0.175 radians, respectively. Since wild boars are considered large animals (50–300 kg in GNP) [35], there is little difficulty in their detection by camera traps and classical approaches in estimation of camera trap parameters seem sufficient [42]. Daily movement of wild boars (v) was extracted from a radio-tracking study in a primeval temperate forest, which is ecologically comparable to GNP, as 6.8 ± standard error (SE) 0.57 km.day^-1 [43]. Average group size of wild boars (g) was calculated through the encounters of this species on line transects in GNP (S1 File). We estimated the variance of density using the delta method, as the squared variance of each independently estimated REM parameter added to the squared variance of bootstrapped wild boar encounter rate (y/t) [40]. We conducted bootstrapping by resampling camera locations 10,000 times with replacement [40].

To assess the abundance of bezoar goats, which inhabit hardly accessible rocky terrain, we applied a double-observer point-count approach based on mark-recapture theory [44–45]. We followed habitat descriptions in the literature to identify the rugged habitat of bezoar goats in GNP [46–47]. Namely, we used a threshold of 0.03 for the ruggedness index [48] and 40° for slope using the digital elevation model (DEM) in ArcGIS 10.1 (ESRI, Redlands, CA). Moreover, we added a buffer of 200 m as bezoar goats graze in areas near cliffs as well [47]. We excluded isolated habitat patches of less than 3.5 km², where we did not expect any animals to exist. Thereby, in the remaining 53.6 km² of bezoar goat habitat we selected 20 random sampling points with a minimum distance of 3 km (Fig 1). We selected vantage points at 200–500 m away from the hardly accessible bezoar goat habitat using the viewshed function in ArcGIS 10.1 for scanning the sampling points. Each vantage point was visited once by two trained observers (usually one park ranger and one team member: AG, AKH, LE or MS), each equipped with a rangefinder, binoculars and compass. Each observer conducted four alternating independent scans of 15 minutes and recorded sightings up to a maximum distance of 1000 m to equalize detection changes for both observers [44]. For each sighting, observers mapped the location and recorded the number of animals (cluster size) and the distance to the center of the cluster. After finishing the survey, both observers compared their data sheets to identify ‘captures’ (clusters detected by one observer) and ‘recaptures’ (clusters detected by both observers). Due to the relatively low density of the target species, it was unproblematic to distinguish groups and avoid double counting. Point counts were conducted during six fieldwork days (17–19 November and 2–4 December 2014). Due to bad weather conditions and difficult accessibility, we omitted surveying four points. We used the program DOBSERV [49] to model detection probability of bezoar goat clusters based on two capture-recapture models: equal or unequal detection probability between the observers. Methodological details underlying the program are described elsewhere [49]. We used the Akaike information criterion corrected for small sample sizes (AIC_c) to find the most parsimonious model(s) and selected the best models as those having Δ AIC_c < 2 [50]. To translate this number into density, we multiplied it by the average group size and divided by the sampling area. The sampling area was calculated from the overlap of areas visible from vantage points and the identified bezoar goat habitat using the viewshed function in ArcGIS 10.1. The density was then extrapolated to the total bezoar goat habitat to calculate the abundance.

We conducted structured questionnaire surveys in March and May 2013 among 136 council members and village heads from 34 villages within the GNP watershed to obtain data on small livestock depredation by leopards [22]. Although leopard depredation on small livestock has been reported in 12 out of 34 studied villages, we considered only villages < 2.5 km away from GNP boundaries (spatial availability). This was done to include only villages with the highest likelihood of attacks by leopards from GNP, according to the earlier results on Persian leopard movements [51] and the highest intensity of conflicts in the vicinity of reserves [21–22, 52]. Details on our interview survey procedure have been provided elsewhere [22]. As spatial characteristics of carnivore attacks on livestock may play an important role in the conflict [4, 21, 53], we used only small livestock data from villages with confirmed leopard attacks in our analysis (spatial availability). In GNP, small livestock are available to leopard predation only during free grazing or corralling at night in the pastures and are not killed while in pens inside villages [2–3, 22, 52]. Having interviewed 18 shepherds, we developed a small livestock availability coefficient (C) as the proportion of the average number of grazing days to the number of all days in a month (temporal availability). So, C ranged from 0 to 1 per month, where 0 means that small livestock do not graze outside and stay only in village pens and 1 means staying overnight in the fields away from villages all month. When small livestock were grazed during the day and returned to village pens in the evening, we considered C as 0.5. Livestock numbers in each conflict village were corrected by multiplying by their corresponding value of C. We acknowledge that this is a simplification of predation risk, which can be affected by other husbandry factors as well [11]. However, as small livestock are within the leopard’s preferred prey body mass range [19], cause no injury threat to leopards, and graze mostly inside GNP or its surroundings, we believe their predation probability is determined by availability [3, 52]. Cattle predation by leopards also occurs in the villages around GNP [22]. However, as cattle grazing outside pens and consequently predation by leopard depends on cattle age and breed [2, 13, 22], we were unable to incorporate its availability and excluded cattle from this study.