Individual behavior. Studies aiming to report on specific behaviors (feeding, reproduction, territoriality, social interaction, etc.) must direct sampling efforts to places of interests (e.g., salt licks uses: Blake et al., 2010; carcass scavenging: Bauer et al., 2005; specific habitat use: Sequin et al., 2003). To date, only few studies use camera traps data to study individual ranging behavior and estimate home range size (e.g., Gil Sanchez et al., 2011). Those often have to be completed with data collected using other protocols such as telemetry or indirect animal clues (feeding residuals, latrines, nests, etc.), which could explain the relatively small number of studies estimating home range size.
Population level studies. Studies dealing with population monitoring usually need stronger sampling effort and more complex sampling design. To do so, camera traps are increasingly used as an alternative to other more traditional methods. However, Gompper et al. (2006) proved camera traps to be inefficient at detecting small canids, which were otherwise detected by scat surveys, DNA analysis and/or snowtracking. When comparing different methodologies for the census of population diversity and abundance, camera trapping appear to be the most appropriate method in difficult to access areas compared to line transect or animal track survey (Silveira et al., 2003). Using camera traps to estimating population density can involve complex sampling design and be subject to numerous biases. Firstly, it is important to consider the bias of disproportionally samples more easily accessible or more attractive places for wildlife where detection probability is increased (Foster et al., 2011). The typical procedure to characterize an animal population in a given habitat consists of setting up the sampling effort (camera traps) in a random or systematic way (Foster et al., 2011). As explained by Rowcliffe et al. (2013), cameras can be positioned in less or more attractive places to animals as long as those are proportionally sampled in regards to their relative occurrence in the studied ecosystem. Thus, using a grid and a random number generator, or following a stratified design allow ones to select positions where to install the cameras at random in regards to the animals (Rowcliffe et al., 2013). However, some researchers have set up their cameras in specific places where the targeted elusive species are likely to pass, hoping to maximize encounter rate (e.g., Sanderson, 2004; Weckel et al., 2006); some have even tried to lure animals with attractive smells or baits (e.g., Trolle et al., 2003; Garrote et al., 2012). Indeed, placing camera traps in a non-random way is not necessarily an issue as “it is the animal population within an area that is the subject of sampling by observation stations, not the area itself” as observed by Bengsen et al. (2011). Secondly, one needs to consider variations in detection probability due to the material used. The use of incandescent flash at night can easily spook the target animals and negatively influence future visitation rates in the vicinity of the camera (Sequin et al., 2003; Wegge et al., 2004). Thus, in the case of capture-recapture sampling or studies on habitat use of nocturnal species, it is preferable to avoid using camera models with an incandescent flash. In addition, it is important to make sure all set up cameras have sufficient battery life for a given sampling period. Due to spatial variations in animal community or to different camera ( for example: MOULTRIE A-30 GAME CAMERA ) models, the number of pictures taken can greatly vary between cameras and some can see their batteries getting empty much more rapidly than others do. Cameras running out of batteries possibly miss information (animals passing in the field of detection without being photographed) and lead to underestimated wildlife estimation. Apart from sampling bias, population estimates with low precision is a common issue when using camera traps data. Sampling design with low detection probability, due to a low number of camera traps, a short duration of a study or inadequate material can only permit to obtain low sample size, which itself limits our ability to obtain precise parameters and strongly affects the strength of population estimates (Foster et al., 2011). As a mean to increase sample size, setting up two cameras at a same station allows obtaining pictures of both flanks for marked animals and can facilitate the identification of individuals (Kalle et al., 2011; Negrões et al., 2012).
Intra-community interactions. In the case of seed dispersal studies, the camera is often set up so that the visual field includes the fruits or seeds of interest to maximize the chances of photographing frugivores (Seufert et al., 2010; Nyiramana et al., 2011). Variables of interest here are frequency of visits and the relative contribution of different animal species to seed removal. From personal experience, two remaining limitations can, however, be identified. The first limitation occurs when the camera is positioned close to a fruit/seed sample so observers can easily quantify the number of items manipulated by animals. Here, the focal distance might be too close to being able to photograph all the animals visiting the area. The camera would then record a limited number of visiting species and individual animals. By contrast, the second limitation occurs when the camera is positioned to sample the widest area possible below a fruiting tree canopy, in order to systematically record all visiting animals. In this scenario, the focal distance might be too high to allow observers to see accurately the number of fruits/seeds manipulated. An alternative could be to set up two or more cameras at a same location to sample both the tree canopy’s shadow and a fruit sample on the floor. In the latter case, an alternative to evaluate species-specific contribution to seed removal could be to consider visit frequencies per species in the area. Additionally, seed removal rate can be indirectly assessed with an exclusion experiment (Culot et al., 2009).
Data analysis. The identification of individual animals is generally made by natural fur marks, injuries, and coloration patterns (dots, bands). This identification is, however, always subjective and likely to vary according to the observer and thus likely to affect the precision of estimates. To diminish the risk of mistaken identification, different computer models are able to help matching pictures of marked individuals (Kelly, 2001; Mendoza et al., 2011). Such tools allow observers to improve their ability to recognize individual animals and to be more precise in making population density estimates.
Individual identification is a crucial step in making population estimate. The spatially explicit capture-recapture technique is increasingly used for this purpose (e.g., Efford, 2011; Kalle et al., 2011; O’Brien et al., 2011). This technique assumes that animals are independently distributed in space and that they use defined home ranges. Thus, a model must be run, which considers, on the one hand, a population parameter (population density) and, on the other hand, a process of individual recognition. The detection process is itself driven by a mathematical function describing the probability of detecting an animal, which decreases as the center of a given home range gets further away from a camera trap (Kalle et al., 2011).
Camera trap data are also used to generate abundance indices and get quick insight into population size. However, the power of such indices is limited compared to true estimates of population density for different reasons. Firstly, variations in indices cannot necessarily be attributed to true variations in population size. Indeed, to use and be able to compare such indices one needs to make the assumption that wildlife detectability is constant over time, space and between species, however, this is either not tested, nor true (Sollman et al., 2012). Secondly, those indices are rarely calibrated with the actual population and thus only give little information on the true dynamic of population size (Sollman et al., 2012). Moreover, a too low number of traps set up (replicas) does not allow the calculation of a confidence interval (variance) necessary to estimate the exactitude of indices (Azlan et al., 2006), though Bengsen et al. (2011) adapted a General Index Model able to account for variance when calculating population abundance indices.
Camera traps data such as species detection/non-detection can also be used in occupancy model (e.g., MacKenzie et al., 2003; Long et al., 2011) to predict species occurrence and determine population dynamic parameters. Such models generate detection probability data and thus prevent the recording of false absence. This has very helpful implications for monitoring elusive species for which observations are scarce.
4. Conclusion
Depending on the data to be collected, the target animal species and the type of ecosystem, it is essential to first choose the appropriate equipment to collect the data needed, as not all camera models will be suitable for a specific research objective. Given the increasing use of camera trapping by scientists, we believe that the available technologies should and will know improvements in the future. Higher image resolution resulting from larger sensor and more efficient infrared beam would allow a better identification of individuals, especially for marked nocturnal species. Even more discrete and faster cameras would prevent spooking animals and get more unblurred pictures. Next, the implementation of appropriate sampling protocols must be seriously considered. In a general way, we believe that homogenization of detection probability could improve the use of camera traps data by diminishing biases and allowing stronger inter-site and inter-species data comparison. This could be done:
– at the camera scale, by using camera models having similar features (detection zone, field of view, trigger speed, etc.),
– at the ecosystem scale by implementing standardized sampling scheme (number of cameras, spacing, and placement).
Having a standard sampling protocol would also permit more solid use of statistical models and interpretation of results. The use of computer tools to improve the scientific value of pictures is increasingly common but all does still not agree basic assumption requirements. Future development of computer tools for population density, abundance and site occupancy estimates would need to rely on empirical validated results on individual habitat use behavior and population dynamics.