Study site

The 4,200-ha study site included the 3,386 ha of the Reserva Ecológica Michelin (REM) and adjacent properties in the municipalities of Ituberá and Igrapiúna, Bahia, Brazil (39°10’W 13°47’S to 39°13’W 13°54’S) (Fig 1). The study area is embedded in a regional landscape that retains approximately 40% forest cover and is part of a 9,000-ha rubber plantation established by Firestone in 1950s. The natural vegetation is lowland evergreen broadleaf moist forest. Average annual rainfall is about 2,000 mm with no distinct dry season [16]. Monthly rainfall ranges between 118 mm and 208 mm. Minimum average monthly temperatures range between 18.6 C and 21.9 C, while the maximum average monthly temperatures vary between 26.7 C and 30.8 C [17]. The studied mosaic consists of 1,800 ha of forest divided into four main blocks (one of them is part of a 13,000-ha forest block that extends beyond the study area), 1,000 ha of actively cultivated rubber groves, abandoned rubber groves overgrown with pioneer species, riparian forest corridors (30–50 m wide and up to 12 km long), cattail wetlands, streams and small rivers. The maximum distance between forest blocks is 1 km and the average distance between riparian corridors is ≤400 m. Although all of the forests were exploited for timber, firewood collecting, and Euterpe edulis palm extraction prior to the creation of the reserve in 2005, and some forest patches are manioc fallows, the flora is still highly diverse [18]. The small forest fragments and riparian forests are dominated by pioneer trees (principally, Cecropia, Schefflera, Byrsonima, Miconia, Henrietta, Tibouchina, Tapirira, Senna, Kielmeyera, Himatanthus, Inga, Stryphnodendron, Bauhinia), Rubiaceae, Melastomataceae, and Piperaceae shrubs/bushes, thin vines, and Cyperus sedges.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Study area and location of the sampling sites, Ituberá/Igrapiúna, Bahia, Brazil, 2013–2014.

https://doi.org/10.1371/journal.pone.0195311.g001

Rubber trees grow up to 15m and are planted in rows with a spacing of 8m (between rows) x 3m (between trees) (Fig 2). This allows a density of approximately 500 trees/ha with pioneer vegetation in the inter-rows. The floristic composition in well-developed inter-rows is similar to that of the riparian forests. Rubber tapping involves a diagonal cut in the bark from which the latex drips into a cup that is attached to the tree. Workers enter the groves by 5:30 a.m. and leave by 12:00 p.m., tapping 900 trees/day with each tree tapped on a 4-day rotation. This schedule is followed year-round. Rubber trees can be tapped for years by shifting the cuttings around the trunk so that, once planted, the landscape is stable for several decades. None of the areas sampled had been replanted since the creation of the plantation in 1950s.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Rubber trees as tall as 15 m with medium-developed inter-rows.

https://doi.org/10.1371/journal.pone.0195311.g002

The study fauna

The study fauna included all of the terrestrial and scansorial mammalian carnivores registered for the region: puma (Puma concolor), jaguarundi (Puma yagouaroundi), ocelot (Leopardus pardalis), margay (Leopardus wiedii), oncilla (Leopardus tigrinus), tayra (Eira barbara), South American coati (Nasua nasua), crab-eating raccoon (Procyon cancrivorus), crab-eating fox (Cerdocyon thous), domestic dog (Canis familiaris) and domestic cat (Felis catus) [17]. The domestic dog and the domestic cat are the only non-native carnivore species and it is estimated that nearly 70% of the rural households in the study area have dogs and/or cats (A. Dechner, personal communication). All species are mesocarnivores except for the puma. Other species of carnivores found in the landscape but not included in the study because the sampling methods were not adequate to capture their habitat use were: the southern river otter (Lontra longicaudis) and the kinkajou (Potos flavus). The jaguar was extirpated by hunters during the land conversion events of the 1950/60s [16]. The bush dog (Speothos venaticus) may have been historically present in the area [19] but further information on causes and approximate time of disappearance is unclear. In terms of prey species, three terrestrial species of mammals were extirpated between the 19^th century and the early 20^th century: lowland tapir (Tapirus terrestris), white-lipped peccary (Tayassu pecari) and giant armadillo (Priodontes maximus) (for full list of extant medium and large mammals see [16]). Long-term monitoring (1997–2015) through transect censusing showed that the relative abundance of the medium and large mammal fauna in our study area has increased by 117% since the creation of the REM in 2005 and the implementation of guard patrols (K.M. Flesher, personal communication), with the abundance of prey species reaching levels similar to well-protected reserves elsewhere in the neotropics [16].

Sampling design and data collection

We set up a 90-site grid with a spacing of approximately 600 m between sampling sites. Of the 90 sampling sites, 30 were in forests, 30 in riparian corridors and 30 in rubber groves (with medium to well-developed inter-rows, i.e. height > 2 m) (Fig 2). We used a trimestral random rotation sampling design in which 30 camera traps were placed at 30 randomly chosen sites (10 in forest, 10 in riparian corridors and 10 in rubber groves) each month, with a subsequent rotation to another 30 sites in the next month and so on until all 90 sites were sampled before repeating any. Thus, each of the 90 sites was sampled four times (once per trimester) between February 2013 and January 2014. Each camera was in place for an average of 22 days/month. The objective of the rotation was to maximize the sampled area, to reduce bait habituation by maximizing the time between bait placement, and to reduce spatial autocorrelation by maximizing the distance between sites sampled at any given time. We conducted this research with approval of the Animal Care and Use committee (IACUC) of Michigan State University under the AUF # 11/12-208-00. In addition, we collected the data with approval of the Brazilian government under the CNPq permit # 001308/2012-2.

We used two brands of cameras, Bushnell (Trophy Cam—Bushnell Corporation—Overland Park, KS) and Tigrinus (4.0D and 6.0D Tigrinus Equipamentos para Pesquisa Ltda—Santa Catarina, SC). Cameras were tested before the study began to gain familiarity with the different programming options offered by each brand and model. Our objectives were to 1) program each camera with equivalent settings (e.g. comparable time intervals between pictures), 2) maximize the chances of capturing all the animals that passed in front of all cameras (i.e. high sensitivity, short interval times between pictures), and 3) minimize possible differences between brands due to improper programming (Table 1). The cameras were randomly assigned to each site while guaranteeing that all sampling sites received both brands over the year. Given that the grid was laid out without reference to previous information on carnivore movements/presence and with the aim that the animals would pass in front of the cameras, we placed lures and baits one time at the beginning of the month at each one of the 30 sites to be sampled. We used a mix of scent lures and dead baits to attract all the different species in the carnivore community [20]. The lures consisted of 500 gr of mashed sardines in oil placed on the ground, 5 ml of banana oil applied on top of a stick placed on one side of the sardines, and 10 ml of a mix of lures (Raccoon Lure No. 1 by Russ Carman, Magna Gland Lure by Russ Carman, Mega Musk Lure by Russ Carman and Ausable Brand Catnip Oil Pint) applied on top of another stick placed on the other side of the sardines. All cameras were placed at a height of 25–35 cm and set to work 24 hours/day.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Information about the camera traps and programming used.

https://doi.org/10.1371/journal.pone.0195311.t001

Statistical analysis

To determine the habitat use of a carnivore community, and the effect that different brands of camera traps may have had on their detection, we used a Bayesian approach and a multispecies occupancy model with known species richness [21]. We fit a single-season model due to our limited number of sampling periods (i.e. 4). Occupancy has been defined as the proportion of area, patches, or sample units that is occupied or where a species is present [22]. Occupancy models can be described as a hierarchical, coupled logistic regression, with one regression describing the true occurrence (i.e. state or ecological process) and the other describing the detection probability (i.e. observation process), given that the species occurs in an area [23]. The effects of ignoring imperfect detection (i.e. assuming that a non-detection is equivalent to an absence) have been broadly discussed [24], and occupancy models represent a step forward in ecological studies by estimating occupancy while accounting for imperfect detection.

By including the parameters at the species level as random effects governed by a common set of hyper-parameters or parameters at the community level, multispecies occupancy models allow for the estimation of covariate effects at the species level and of the aggregated effects at the community level (with both levels being related to each other) [25]. These models are valuable given that they increase the precision in occupancy estimates for most species in the community by borrowing strength from the entire community [26]. This allows for the inclusion of infrequently observed species, whose occupancy otherwise could not be estimated through single species occupancy models [25].

Camera data results were organized per species, in a 90 sites (rows) by 4 periods (columns) matrix, containing binary data where 1 represented when a species was detected at a specific site during a specific sampling period and 0 when it was not detected. Thus, we used a binary state model in which true occurrence z_ij is a latent state variable, specified as z_ij ~ Bernoulli (ψ_ij), where ψ_ij is the occurrence probability of species i at site j. The observation model was specified as x_ijk ~ Bernoulli (ρ_ijk·z_ij), where x_ijk is the detection/non-detection data for species i at site j and for the period k, and ρ_ijk is the detection probability for species i at site j and for the period k [25]. We included in the occupancy model the different types of habitat as covariates, and in the detection model the different brands of camera traps used. In both models, the covariates were included as indicators (i.e. 0 or 1). To avoid overparameterization, the ecological process model was specified as follows:

Where, ψ_ij is the probability that species i occurs at site j, α_0i is the occurrence probability for species i in rubber groves on a logit scale, α_1i is the difference (on a logit scale) in the occupancy probability between rubber groves and forest areas for species i, and α_2i is the difference (on a logit scale) in the occupancy probability between rubber groves and riparian areas for species i.

The observation model was as follows:

Where: ρ_ijk is the detection probability of species i at site j for the period k (if present, z = 1), β_0i is the detection probability (on a logit scale) of species i when using Bushnell cameras, and β_1i is the difference in the detection probability (on a logit scale) between Bushnell and Tigrinus cameras for species i (S1–S3 Supporting information).

We assumed a closed community (i.e. absence of local extinctions or colonization during one year); however, in order to account for the possibility of weak temporary emigration, and since it may have been confounded with the detection probability, we refer to the occupancy probability as the probability of use sometime during the study period as suggested by Kéry et al. [23]. We did all analyses using R [27] and R2OpenBUGS [28]. We used non-informative priors, and ran three chains each for 55,000 iterations, burning 15,000 and thinning the model by 3. We assessed convergence of the model by looking at the R-hat values, defined as the potential scale reduction factors [28], with values <1.1 being considered acceptable [29]. We tested the prior sensitivity through comparison with model results obtained with different priors. Finally, we assessed model fit by estimating the Bayesian p-value as calculated in Zipkin et al. [25]. The Bayesian p-value is a measure of the discrepancy between the observed data and data simulated under the proposed model, with values close to 0.50 indicating adequate fit and values near 0 or 1 indicating poor fit.

Finally, we explored the spatial autocorrelation between sampling sites, by computing experimental variograms [30]. Given that the locations of the cameras changed every month, we computed one variogram per month with all species’ captures aggregated to increase the number of sites with observations. Although experimental variograms do not require that the data follow a specific distribution [30], it has been demonstrated that asymmetric data, particularly in analyses with small sample sizes, can result in unreliable variograms [31]. Considering that our data are asymmetric, and only for the purpose of computing the variograms, we square-root transformed our count data to reduce the positive skewness cause by a large number of zeroes [32]. Variograms were calculated using the R package GeoR [33]. To observe how the semivariance changed as a function of the predetermined distances in the grid, we created lag intervals of 600 m (equivalent to the grid cell size). We set up a maximum lag distance of 4,800 m (roughly one third of the largest distance between pairs [30]).

Species habitat use in the rubber/forest mosaic

The studied landscape supports a nearly intact mammalian carnivore community seven decades after its conversion from continuous forest to a rubber/forest mosaic. Nonetheless, not all species use the new mosaic in a similar fashion. This differential habitat use by species is important as it not only determines their distribution in the landscape but may modulate their inter and intra-specific interactions [34].

Coatis showed a significantly higher probability of using forests and riparian corridors than rubber groves which is consistent with other studies that have described coatis as a forest species [35–38]. On the other hand, camera trap studies conducted exclusively in neotropical lowland moist forests have frequently recorded few coatis’ captures [39] and/or have reported moderate occupancy values for this species (≤0.42 [40]) compared to our mean occupancy value (0.98 in forests). We believe that the difference may be due to the fact that coatis in those areas may be spending more time in the canopy than on the ground to avoid predation, as the carnivores with the highest capture frequencies and/or occupancy values in those studies were ocelots, jaguars and pumas, all predators of coatis [41]. The relatively high detection probability of coatis, in comparison to other species, was likely because it was the only species that showed a consistently strong attraction towards the lures. Not surprisingly, we found that 91% of the captures of coatis were registered during the day (i.e. from 6:00 a.m. to 5:59 p.m.), and that the difference between day and night captures was statistically significant (one-way chi-square test, x² = 138.94, df = 1, p< 0.001).

Domestic dogs are non-native carnivores but are nonetheless an important part of the ecosystem due to their abundance, widespread distribution, and the fact that they prey on native species throughout the Atlantic forest [42–44]. We found that domestic dogs have significantly lower probabilities of using forests (mean probability of 0.27) than riparian corridors and rubber crops, indicating that the forest may serve as a partial refuge for wildlife from this non-native predator. The preference of dogs for the rubber groves is not surprising as they often accompany their owners during their work in the rubber groves. Although they also enter the forest to pursue prey on their own or in the company of hunters, the higher use probabilities of dogs in the rubber groves and the frequent camera trap registers of dogs with rubber tappers, indicate that dogs generally preferred to stay close to their owners rather than to enter the forest on their own.

There was high spatial overlap in habitat use between dogs and foxes, but their use of these areas was temporally separated; 92% of dog captures were diurnal while 73% of fox captures were nocturnal (two-way chi-square, x² = 98.69, df = 1, p< 0.001). Foxes’ preference for rubber groves and riparian corridors and their low probability of using forest is consistent with the results from other studies showing that foxes prefer open, forest edge and agricultural habitats in mosaic landscapes [2, 35, 36, 45].

The wide posterior intervals for the habitat use probabilities of tayras and wildcats, allowed us to highlight a pattern in the precision of the occupancy parameter of species with low detection probabilities and moderate capture values, which we discuss further in the next section. We found that 96% of the captures of tayras were registered during the day (one-way chi-square test, x² = 22.16, df = 1, p< 0.001), while 96% of the captures of wildcats were registered at night (one-way chi-square test, x² = 18.18, df = 1, p< 0.001). In spite of the fact that we only captured pumas twice, once in the forest and once in a riparian corridor, we regularly found scats, tracks, and rakings in all forest habitats and, less frequently, in the rubber groves. The small size of the study area relative to the home range of pumas suggests that there are few individuals inhabiting the area. The low number of captures may be related to the fact that pumas regularly use trails and unpaved roads [46], and thus the placement of the cameras independent of the trail network may have made our sampling design inadequate for determining the habitat use of this species. The low number of captures of crab-eating raccoons was unexpected as we found tracks of this species and saw it along the roads throughout the landscape. Other scientists in Bahia found raccoons to be common in other agroforestry landscapes [36, 47] and on a larger regional scale raccoons were widely reported and recorded in all landscape types; they are wetland habitat specialists, but generalists in terms of broader habitat classifications (i.e. they exploit wetland habitats wherever these occur in the regional landscape) [17].

There were two terrestrial species that were not captured by the camera traps, the jaguarundi and the domestic cat. We did see the jaguarundi several times in the rubber groves and riparian corridors but sightings were infrequent and interviews in the broader region indicate that the jaguarundi is rarely seen [17]. These results are consistent with other studies in areas with differing levels of human disturbance in which the species was not recorded or recorded with very low capture rates [36, 39, 48]. Domestic cats, although found in human settlements, were not captured in any of our sampled sites which is consistent with other studies in the region where domestic cats were rarely recorded [36], suggesting that cats do not wander far from human settlements.

Sampling and statistical methods

Camera traps are being increasingly used in ecological studies as they make it possible to collect information on elusive species in a minimally invasive manner and are cost-effective [49, 50]. While today there is a wide selection in the market of camera traps, some studies have shown variability in performance among brands, models [51], and even between cameras of the same model [12]. Protocols for monitoring of birds and mammals recommend the use of one or two high-end models of camera traps, possibly to reduce the bias derived from this variability [52, 53]. However, high-end camera traps are costly. This is critical considering that most biodiversity hotspots are in developing countries where the resources to obtain high-end camera traps may not be readily available.

We found that the detection varied among species and that tayras and wildcats had the lowest detection probabilities. Low detection probabilities make it difficult for occupancy models to distinguish between a site where the species is truly absent and a site where it is simply not detected. Thus, occupancy estimates produced with low detection probabilities (<0.15) are unreliable [24, 54]. We found that there is a significantly higher detection probability for the studied carnivore species when using Bushnell as opposed to Tigrinus cameras. The difference in detection probabilities may be due to differences in sensitivity, trigger speed, and noise produced by the equipment (as it may scare carnivores away). The Tigrinus cameras are composed of a Sony Cyber-shot embedded in a rigid box which, when activated, produces a louder noise than the Bushnell cameras. This difference in the detection probability between the different brands of camera traps show the necessity of accounting and/or controlling for such factors when necessary. This can be done by estimating ecological parameters while accounting for imperfect detection. Although obtaining the data needed to adjust statistically for imperfect detection may require more effort to be able to provide reliable parameter estimations (e.g. extending the sample period to increase detection rates [54, 55]), ignoring the effect of the cameras on the detection probability may seriously bias the results and therefore mislead conservation actions. In terms of sampling, the rotation of the camera traps used in this study proved to be effective at controlling spatial autocorrelation.

One of the benefits of multispecies occupancy modeling is derived from the potential these models offer for increasing the precision of coefficients of rarely detected species by borrowing strength from the entire community [26]. However, the wide posterior intervals for habitat use probability, reaching 1.00 in all cases for tayras and wildcats seem to be an artifact of the model. This artifact appears to occur when the species exhibit a moderate level of captures and a low detection probability (Fig 5). This indicates that when using these models, the occupancy estimation of a species that is moderately common but difficult to detect is highly uncertain with a tendency for the model to suggest that the species could be everywhere.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5. Relationship between the number of detections (counts or captures) and the predicted width, high end, and low end of the 95% interval of the occupancy parameter for species with detection probability <0.15 in a multispecies occupancy model.

Data taken from Zipkin et al. 2009.

https://doi.org/10.1371/journal.pone.0195311.g005

Finally, we call for caution when drawing conclusions about the ecological requirements of carnivore species derived from observations/captures as these always depend on the species detection probabilities. We agree with Banks-Leite et al. [56] that in tropical areas where some species are naturally rare and logistics frequently complex, it can be difficult to implement models that adjust for imperfect detection. However, it is then the ethical obligation of researchers to control the effect of known factors on the detection probability, or to be conservative when drawing conclusions as in many cases there may be unknown factors affecting the species detections. Detection probability may be affected by logistic and environmental factors, as well as by interspecific interactions and by the ecology of the species per se. Providing conclusions about species with low samples sizes, low numbers of sampled sites, low sampling effort, low detection probabilities, or ignoring the fact that sometimes species present higher detection probabilities in disturbed ecosystems (either because of concentrated foraging activities or due to visibility), may lead to erroneous conclusions and as a result lead to inappropriate conservation actions.

Although in this paper we described the habitat use of carnivores after 70-years of conversion from continuous forest to a rubber/forest mosaic, information related to the temporal dynamics of the species, as well as their intra/inter specific interactions is lacking. This information is essential to fully understand the effects of landscape modification on the ecology of the species. In addition, although it is frequently assumed that the capture of an animal in a specific habitat is an indicator of long-term habitat suitability, other factors affecting populations at different temporal and spatial scales, which are usually not considered in ecological studies (e.g. habitat change and its effects on the physical and reproductive health of species), should be considered.