The use of camera traps has drastically increased over the last years (Rowcliffe & Carbone, 2008), and a large amount of data are being collected from sites all over the world. However, most studies only publish a small subset of the collected data, usually focusing on only one or two species. The example of capture–recapture studies using camera traps shows how the availability of a clear methodology and guidelines for study designs (Karanth & Nichols, 1998) can result in a large number of researchers adopting the methodology for a range of species (e.g. Noss, Peña & Rumiz, 2004; Silver et al., 2004, Jackson et al., 2006; Trolle et al., 2008). At the same time, the use of a standard methodology across studies allows for easy comparisons between sites (e.g. Karanth & Nichols, 1998; Silver et al., 2004; Kelly et al., 2008). Our article (Tobler et al., 2008) presents different methods for analyzing and presenting species inventory data collected with camera traps and addresses various design issues. We hope that this will encourage others to analyze and publish newly collected or already available data on mammal communities from different sites.
Following the suggestion by O'Brien (2008), we tested the performance of occupancy models for species richness estimation with our data. We followed the guidelines given in MacKenzie et al. (2006: pp. 250–253) and tested four different models (1) a single season model with constant detection probability (SS); (2) a single season model with survey-specific detection probabilities (SS time); (3) a two-group finite mixture model (FM2); (4) the Royle–Nichols model (RN). Occupancy models use presence/absence data from multiple surveys to estimate detection probabilities. We combined data from six camera trap days for one survey occasion which resulted in 10 survey occasions used in the analysis of each dataset. The total number of species possibly present was set to 28, according to a species list based on direct and indirect observations made by various researchers during or shortly before and after the camera trap surveys. We evaluated the effect of sample size by using subsamples of the data. We analyzed the data in the order they were collected without randomization, so that results differ from the ones given in table 3 in Tobler et al. (2008). All occupancy models were calculated with Presence (Hines, 2007) and Jackknife estimators were calculated with EstimateS (Colwell, 2006). Model selection based on the Akaike information criterion (AIC) selected the FM2 model as best model for both datasets, followed by the RN model (Table 1). Selection of these particular models indicates that the data contains a high level of heterogeneity, something we would expect with camera trap data from a wide range of species. The FM2 model gave the best estimates of the three occupancy models while the SS model hardly differed from the observed number of species (Sobs) (Table 2). The Jackknife estimators performed better than Sobs as we have shown before (Tobler et al., 2008) but showed more variance than the FM2 model. One drawback of the occupancy models tested here is that the total number of species that can possibly occur at a site needs to be known from a regional species list. The model then estimates the fraction of species that is actually present at the site. However, regional species lists are not always available or complete. Future studies need to evaluate how sensitive these models are with regard to the value chosen for the total number of species.
The usefulness of capture frequencies as an index for abundance was not the topic of our article; however, we briefly addressed the issue in the discussion of our results. Kelly et al. (2008) correctly pointed out that capture frequencies were highly correlated between surveys and suggests that the usefulness of capture frequencies as an index should be explored. However, this analysis includes frequencies from all species spanning two orders of magnitude (0.4–66) and does not give any information on the relationship between abundance and capture frequencies. It cannot be determined whether an increase of the capture frequency from 5.6 to 13.2 for the red brocket deer Mazama americana really reflects a population increase, nor if the decline of the capture frequency from 16.0 to 8.1 for the collared peccary Pecary tajacu reflects a population decline.
We agree that there exists a relationship between abundance and capture frequencies, however, as pointed out by Tobler et al. (2008) and Rowcliffe & Carbone (2008), there are several other factors that influence capture probability and that can confound the relationship. In order to compare data between species, surveys or study sites the capture probability needs to be estimated from the data. Rowcliffe & Carbone (2008) refer to a new method they recently developed to address this problem. Another promising approach is to use occupancy as a surrogate for abundance (MacKenzie & Nichols, 2004), and to use occupancy models to estimate capture probabilities (MacKenzie et al., 2006). Occupancy models are very flexible and allow the inclusion of covariates to test for difference in occupancy rates between sites or habitats and multi-season models can be used to look at extinction and colonization rate. However, despite their great potential, so far they have only rarely been applied to camera trap data (e.g. MacKenzie et al., 2005; Linkie et al., 2007).
We hope that our results will encourage others to investigate possible new ways of analyzing camera trap data. As pointed out by Rowcliffe & Carbone (2008), many interesting questions could be addressed based on existing data if all these data were organized and presented in a more standardized way. We are convinced that advances in camera technology together with better designs and new methods and software for data analysis will allow camera trap studies to address an even broader range of questions, thus making them one of the most versatile tools available to biologists.