- Top of page
Camera trapping uses fixed cameras, triggered by infra-red sensors, to ‘trap’ images of passing animals. It is a quantitative technique that has relatively low labour costs, is non-invasive, incurs minimal environmental disturbance (Henschel & Ray 2003; Silveira, Jacomo & Diniz-Filho 2003), is robust to variation in ground conditions and climate and, most importantly, can be used to gain information on highly cryptic species and in difficult terrain where other field methods are likely to fail (Karanth & Nichols 1998; O’Brien, Kinnaird & Wibisono 2003; Silveira, Jacomo & Diniz-Filho 2003). Furthermore, camera traps are equally efficient at collecting data by day and night and provide the opportunity to collect additional information on species distribution and habitat use (Henschel & Ray 2003; Silveira, Jacomo & Diniz-Filho 2003), population structure and behaviour (Silveira, Jacomo & Diniz-Filho 2003; Wegge, Pokheral & Jnawali 2004).
In principle, the number of photographs taken per unit time (trapping rate) contains information about the density of a species. Supporting this expectation, a significant correlation between trapping rates and independent estimates of density has been demonstrated across species in Sumatra (O’Brien, Kinnaird & Wibisono 2003). However, the application of trapping rate as an index of abundance (Carbone et al. 2001; Silveira, Jacomo & Diniz-Filho 2003) is controversial, both on theoretical and practical grounds (Jennelle, Runge & MacKenzie 2002; Karanth et al. 2003). This is primarily because it does not estimate the probability of detection and might therefore be confounded by variation in this factor (MacKenzie et al. 2002; Pollock et al. 2002). In an extreme interpretation, this would mean that correlations between trapping rate and density are only reliable if they are recalibrated for every location and time period to which they are applied, thus negating the need for the index.
A well-established and effective way to get around this problem is to use capture–recapture models to estimate abundance, based on the retrapping of recognizable individuals by cameras (Karanth 1995; Karanth & Nichols 1998; Maffei et al. 2005). This method is preferable to trapping rate indices as it is capable of providing robust, unbiased density estimates that are comparable across sites (Jennelle, Runge & MacKenzie 2002; Wilson & Delahay 2001; Srbek-Araujo & Chiarello 2005). However, the method is restricted to species with individually unique natural markings or, in principle, to those for which a sample can be individually marked prior to camera trapping (Trolle & Kery 2003). Relatively few species have natural markings sufficiently variable to be individually recognizable, most camera trap studies focusing on spotted and striped felids (Karanth & Nichols 1998; Henschel & Ray 2003; Maffei et al. 2005). Furthermore, although a robust method to measure the area effectively sampled in capture–recapture studies has recently become available (Borchers & Efford, in press), most camera trapping studies have so far used ad hoc approaches to handle this problem, such as average distances moved between captures and independent measures of home range size (Soisalo & Cavalcanti 2006). These methods have no basis in theory and their reliability is questionable (Williams, Nichols & Conroy 2002).
Species without individual markings have been underrepresented in recent camera trapping research, consideration being restricted largely to presence in mammal inventories (Trolle 2003; Srbek-Araujo & Chiarello 2005) or as a supplement to the study of an individually identifiable target species (O’Brien, Kinnaird & Wibisono 2003). Models of occupancy (MacKenzie et al. 2002) and population size (Royle & Nichols 2003; Stanley & Royle 2005) that can estimate underlying detection probabilities from camera trapping data provide an important advance in this respect. However, currently these methods do not provide estimates of density and the extension of camera trapping methods to do this for species not individually identifiable would greatly extend the value of the technique.
Jennelle, Runge & MacKenzie (2002) suggest that a robust way to derive density estimates from camera trapping rates would be to model the underlying observation process. We present such a model, and provide field evidence of its reliability by comparing density estimates derived from camera trapping with known densities of several species in a large, enclosed animal park. Using simulations, we also explore how the precision of estimation is influenced by changing the amount and spatial allocation of sampling effort.
- Top of page
In this paper, we have developed a technique for calculating animal densities from camera trapping rates by modelling the underlying detection process. By field-testing this novel approach, we have shown that it can provide reliable density estimates. By calculating the expected trapping rates for typical mammalian densities and day ranges (based on allometry), we have also shown that, in the majority of cases, the rate of return of records is likely to be sufficiently high to make the method efficient within realistic time frames under natural conditions. However, appropriate application of the technique requires a clear understanding of its constraints. First, there are three key assumptions, that: (i) animals conform adequately to the model used to describe the detection process; (ii) photographs represent independent contacts between animal and camera; and (iii) the population is closed. A second important consideration is the derivation of independent parameter estimates for modelling detection. These issues are discussed below, highlighting the implications for practical application of the method and considering possibilities for future statistical development.
A key underlying assumption is that animals behave like ideal gas particles, moving randomly and independently of one another. This is clearly unrealistic for animals in a natural setting, where individuals respond to one another and their physical environment. However, the purpose of the field test in this paper was to provide a preliminary assessment of whether the model can nevertheless provide an adequate approximation of the detection process. The fact that there was no apparent bias in density estimates for three out of the four species was encouraging, and it is also worth noting that an equivalent method used to model rates of capture by snares has proved successful across a wide range of species in natural settings (Rowcliffe, Cowlishaw & Long 2003). This suggests that the method is reasonably robust to typical behaviour patterns that may violate underlying model assumptions.
A second important assumption of the detection model is that animals move independently of the cameras, and this assumption will clearly be violated if camera placement strategies either avoid or target focal species. Indeed, inappropriate camera placement explains the dramatic underestimation of mara abundance in this study. During census counts, around 90% of this population was observed in the Central Park area, where they typically graze open lawns. These areas are also heavily frequented by visitors to the park and, in order to avoid clogging the cameras with photographs of people, and possible interference with cameras, cameras were placed in less crowded areas away from the open lawns. These areas were also less frequented by mara, hence the underestimation of abundance.
The mara example represents the opposite extreme to placement strategies typically used in camera trapping studies, which often attempt to maximize trapping rates of rare species by placing cameras near to signs of the animal, on trails, ridge tops and water holes known to be used by them, or by using baits and lures. These approaches violate the underlying observation model developed here and cannot be used to obtain unbiased density estimates. However, rigidly random placement is unlikely to be viable in many situations. A balanced approach is needed, in which the random ideal is followed as closely as possible, while giving each camera enough clear view to provide a reasonable chance of detecting animals. We anticipate that further experience with the method will yield more concrete practical guidelines on placement strategies.
Another possible way in which the assumption that animals move independently of cameras might be violated is trap shyness, caused by the avoidance of either the camera units themselves or of their flash (Séquin et al. 2003; Wegge, Pokheral & Jnawali 2004). The problem could in principle be detected by looking for signs of decline in trap rate over time, although such a trend could also be the result of a decline in abundance over time, or to a general decline in trapping probability caused, for example, by declining speed of movement. If suspected, trap shyness might be solved by using infra-red imaging instead of flash photography or, if the animal of interest is at least partly diurnal, by disabling the flash and relying on natural light and day-time photographs only. However, in some more extreme cases where animals detect and avoid any unusual objects associated with humans (Séquin et al. 2003), further work might be required to identify and eliminate the cue on which avoidance is based.
While the method described here does not allow camera placements to target focal species, a degree of directed sampling is possible through appropriate stratification. For example, if some areas of a study site are difficult to reach, so long as these areas can be defined and measured and at least some, albeit reduced, trapping effort takes place there, stratification of the kind described in this study can be used to obtain unbiased density estimates. Furthermore, whenever spatial variation in density is influenced by recognizable zonation on the ground, such as habitat type, stratification by these zones will be desirable in order to improve the precision of estimates.
The data for the method described must be in the form of numbers of independent contacts between animal (individual or group) and camera. This requires that an animal leaves the camera detection zone after a contact, and that the same or a different animal later re-enters in order to give a second, independent contact. If an animal remains within the detection zone for a long time, or a large group passes through over an extended period, several photographs may result from a single effective contact. In the field study presented here, we attempted to avoid this problem while limiting the amount of film used by setting the cameras to become inactive for 2 min after each photograph. However, using a long latency period runs the risk of missing independent contacts occurring in quick succession. Further work would be useful to refine advice on sensible latency periods but, in the meantime, it may be preferable to use little or no camera latency and assess which groups of photographs represent independent contacts based on their timing and content. This is obviously more feasible with digital cameras, for which picture storage is not generally a limiting factor.
A final assumption of the method is that the population surveyed is closed. In the field study presented here, there was no possibility of migration, and the trapping period was both relatively brief and outside the main breeding and mortality seasons, effectively giving a closed population. Ideally, trapping period should be targeted in this way to meet the closure assumption as far as possible. However, if abundance does change during a survey, the method will provide an estimate that simply averages across the trend.
Turning to the estimation of independent parameters, accurate measurement of trap-related parameters (radius and angle of detection) is relatively straightforward. These values vary to some degree with different cameras and environmental conditions (particularly temperature), and may be sensitive to animal size (Swann et al. 2004). Trials should therefore be carried out to define detection zones specific to each survey.
However, variation in detection zone dimensions is probably less important, and certainly easier to measure, than the animal-related parameters (group size and speed of movement). These are more problematic because they are generally difficult to measure without bias. For example, animal surveys frequently suffer from the fact that smaller groups are harder to detect, leading to overestimation of group size. Furthermore, both group size and speed of movement may be highly variable within species, for example changing with habitat, season, lunar cycle, levels of disturbance and population density. Ideally, group size and speeds of movement should therefore be estimated at the same time and place as the camera trap survey, using appropriate survey methods to avoid bias. In some circumstances it may be reasonable to use estimates from studies on the same or related species under similar conditions, and it is even possible to use allometric estimates of day range (Carbone et al. 2005). However, these indirect measures are likely to introduce a large degree of bias and should be used cautiously only in cases where rough approximations of density are fit for purpose. In order to understand the likely extent of errors inherent in using comparative species parameters, and to move towards an ability to control for these errors, we strongly encourage further empirical work on the determinants of spatial and temporal variation in group sizes and day range, both within and between species.
At this stage in the development of the method we have used bootstrapping to estimate variance in density. However, ideally variance would be estimated using a maximum likelihood approach. This would have the benefit of formally describing both observation error (variance in the number of photographs for a given local density) and process error (variance in local density between camera trap placements), and allowing classical hypothesis tests to be used when comparing densities. However, preliminary explorations suggest that trapping rate data may be unable to separate reliably the relative contributions of observation and process errors to overall variance. A possible solution would be to estimate independently the degree of skew in trapping rates expected given known patterns of animal movement, and to allow variance in local density to be estimated by the data. The ideal gas model on which the underlying trap rate is based assumes that particles move in straight lines, in which case the number of contacts is theoretically Poisson distributed (Hutchinson & Waser 2007). Animals clearly do not follow this pattern and Hutchinson & Waser (2007) show that, while this has no effect on the expected number of contacts, it leads to a more variable distribution of contacts. Despite this useful observation, the theoretical underpinning needed to support a maximum likelihood estimator is not yet sufficiently advanced and we therefore leave this development to a future publication.
Summing up the advantages and disadvantages, the method developed here represents an advance in circumventing two limitations of mark–recapture analyses of camera trap data. First, and most importantly, it does not require the study species to have individually recognizable markings, allowing the technique to be used for a far wider range of species. Secondly, the method is not sensitive to the spacing of cameras relative to the size of animal home ranges, and can therefore be applied more flexibly across a wide range of species. However, the method has a significant disadvantage in comparison with mark–recapture analyses in that the placement of camera traps must be carefully randomized relative to the distribution of animals. As a result, directed placement of traps in order to maximize trapping rates cannot be used, and records of extremely scarce species may therefore accumulate too slowly to be of use. Finally, we emphasize that the need for unbiased, independent estimates of group size and day range are the most significant constraints on the use of this approach.
In conclusion, we feel that careful application of this method, taking full account of its limitations, has the potential to provide a useful contribution to the animal survey tool box, either as a less labour-intensive alternative to existing methods or, for some highly cryptic species, perhaps the only practical way to estimate density. The underlying principles may also be applicable to other static methods such as acoustic monitoring. Priorities for further work are, on the theoretical side, to develop a maximum likelihood estimator and to quantify the degree and determinants of variation in the model parameters across space, time and animal taxa. On the practical side, further testing of the approach is needed on a larger scale, and with a wider range of species. For example, camera trapping while simultaneously monitoring individuals by intensive radio-telemetry would help greatly to show how variation in movement patterns might influence results.