Camera trapping is a vital tool for wildlife monitoring. Accurately estimating a camera’s detection zone, the area where animals are detected, is essential, particularly for calculating population densities of unmarked species. However, obtaining enough detection events to estimate detection zones accurately remains difficult, particularly for rare species. Given that detection zones are influenced by species- and camera-specific traits, it may be possible to infer detection zones from these traits when data are scarce. We conducted a meta-analysis to assess how the number of detection events, species traits, and site-specific variables influence the estimation of the effective camera trap detection distance and angle. We reviewed published studies on detection zones, performed a power analysis to estimate the sample sizes required for accurate and precise estimates, and used mixed-effects models to test whether detection zones can be predicted from biological and technical traits. Our results show that approximately 50 detection events are needed to achieve error rates below 10%. The mixed-effects models explained 81% and 85% of the variation in effective detection distance and angle, respectively. Key predictors of detection distance included body mass, right-truncation distance, and camera brand, while angle was predicted by camera brand and installation height. Importantly, we demonstrate that combining model-based predictions with limited empirical data (fewer than 25 detections) can reduce estimation error to below 15% for rare species. This study highlights that detection zones can be predicted not only within, but also across, studies using shared traits, and that the right-truncation distance is a useful metric to account for habitat-specific visibility. These findings enhance the utility of detection zones in ecological studies and support better study design, especially for rare or understudied species.

Literature search

We searched for published articles reporting effective camera trap detection zones and the selected covariates on Web of Science using the following query: (camera trap* OR trail camera*) AND (effective radius OR detection zone OR detection distance OR Random Encounter Model OR Camera Trap Distance Sampling OR time-to-event model OR space-to-event model). The three density models for unmarked wildlife were included in the query as these always involve detection zone estimation to some degree. We excluded detection zones that were based on either human trials or manufacturer specifications, and detection zones that were estimated across multiple camera trap brands. We excluded two articles that already modeled detection zone parameters as a function of a covariate, but did not provide the raw data to do so independently. In total, 24 studies with data from 17 countries met our criteria (Table 1). Additionally, we included our own data set of 20 species from a study site in French Guiana, the collection of which is described in Wiegers et al. (2024).

Choice of covariates

We focused on six potential predictors that were consistently reported by papers or are accessible from the literature: 1) animal body mass; 2) animal phylogeny; 3) camera trap installation height; 4) camera trap brand; 5) the maximum observed angle per site (max 𝜃), and the maximum considered distance measurement (i.e. the right-truncation distance, hereafter max r) per site, as proxies for local vegetation density or ‘visibility’; and 6) the time, in seconds, between consecutive positional measurements to animals in the viewshed, also known as the ‘snapshot interval’ 𝑇 . The latter was included to account for differences arising from some studies only measuring distance to animals as they enter the camera viewshed (high 𝑇), versus studies that took measurements at many pictures per sequence at regular intervals (low 𝑇 ). The first approach is more common when using detection zones in the Random Encounter Model (Rowcliffe et al., 2008), and the latter for Camera Trap Distance Sampling (Howe et al., 2017), both of which are common models for estimating the abundance of unmarked populations. We did not include animal movement speed because such estimates are rarely reported. We note that movement speed is unlikely to be a useful detection zone predictor because it is known to vary markedly between sites, and even within species (Palencia et al., 2023). Site-specific speed data would therefore be required, which is likely unavailable for species for which detections are already limited. Additionally, we were not able to include animal skin surface temperature, because this parameter is virtually unknown for most if not all wildlife species. While internal body temperature measurements can often be found in literature, we note that these are not good proxies for emitted heat as observed by camera trap sensors (Welbourne et al., 2016). Finally, we did not include measures of camera trap sensitivity, because only few articles reported these.

Body mass estimates were acquired from the studies themselves when reported, and when not, from Jones et al. (2009), Tobias et al. (2022), Richard-Hansen et al. (1999), and Phillipps (2016). Camera trap installation heights and brands were either taken from the articles directly or by contacting the corresponding authors. Taxonomic ranks for the species were taken from the Wildlife Insights database (Ahumada et al., 2020). For most studies, max r (range 3 - 26 m) and max 𝜃 could be computed from the data or taken from the manuscript. For two studies that did not provide detection data but used marked poles to calibrate the distance to the camera, we approximated max r as the distance to the farthest used marker. For two unspecified max 𝜃 values, we took those as reported by the manufacturer or from other studies using the same model. Finally, we assigned studies that took one measurement for each ‘independent contact’, i.e. the first picture of an animal per photo sequence, we assigned a value of 𝑇 =1500 seconds, which was the otherwise highest reported value.

All analyses were performed in R v. 4.3.1 (R Core Team, 2021). We estimated the EDD and EDA using the Distance package (Miller et al., 2019) by fitting detection functions with up to 2 adjustments (Rowcliffe et al., 2011) for all species for which data were available. For these species, we first excluded detections of animals reacting to the camera, and based on a preliminary power analysis, further only considered detection zones with at least 50 trigger events. For the EDD, we left-truncated the data when the detection probability near the camera was lower than expected and right-truncated it when the detection probability became negligible (Buckland et al., 2015). For the EDA, we discarded the farthest 5% of measurements, as these are more likely to be influenced by distance rather than angle (Rowcliffe et al., 2011). We excluded detection zones with very poor fits to a detection function (n = 1 for EDD, n = 0 for EDA). EDD and EDA values from studies that did not provide data (n = 5 for EDD, n = 4 for EDA) were simply taken as stated in their respective articles. In total, the two final datasets included 167 EDDs from 24 studies, and 104 EDAs from 16 studies.

To investigate the required sample size for accurate detection zone estimation, we estimated EDDs and EDAs with a similar approach for subsets of the complete dataset for all species with at least 200 detections (n = 75 for EDD, n = 35 for EDA). To this end, we created measurement subsets of varying sample sizes for each species-site combination, ranging from 10 to 200 through sampling with replacement. For each subset, we then estimated both the EDD and EDA by fitting hazard rate and half-normal functions. (Rowcliffe et al., 2011). For this power analysis, we did not fit adjustments to reduce computing time. This process was repeated 100 times for each species and subset sample size combination. Then, to investigate how the estimation error relates to sample size, we computed the mean relative error (MRE) between the EDD and EDA estimates of the subsets [𝐸𝑠𝑡𝑖𝑚𝑎𝑡𝑒^𝑠𝑢𝑏^−𝐸𝑠𝑡𝑖𝑚𝑎𝑡𝑒^𝑓𝑢𝑙𝑙^] and the estimates calculated from the full datasets as [𝐸𝑠𝑡𝑖𝑚𝑎𝑡𝑒^𝑠𝑢𝑏^−𝐸𝑠𝑡𝑖𝑚𝑎𝑡𝑒^𝑓𝑢𝑙𝑙^] / 𝐸𝑠𝑡𝑖𝑚𝑎𝑡𝑒^{𝑓𝑢𝑙𝑙}.

Mixed-effects models

We created two linear mixed-effects models with the nlme package (Pinheiro et al., 2007) to determine the relationship between the selected covariates and the detection zone parameters. For both the EDD and EDA, we constructed a full model with the logarithm of body mass, camera trap brand, camera trap height, species group, the logarithm of the snapshot interval, and either max r or max 𝜃 as fixed effects, and study site as a random effect. We considered the following species groups which were grouped based on similar morphology and/or lifestyles: 1) ungulates (odd- and even-toed ungulates), 2) carnivores/insectivores (carnivorans, opossums, pangolins, armadillos, and the aardvark); 3) small mammals (rodents, hares, and hedgehogs); 4) primates; and 5) birds. We used the approach from Nakagawa and Schielzeth (2013), calculating the conditional and marginal R² to demonstrate the variation explained by respectively all effects (𝑅_𝑐²) and the fixed effects (𝑅_𝑚² ). Then, to investigate the predictive power of the covariates, we assessed the performance of these models at predicting the EDD and EDA for three common scenarios that practitioners interested in detection zones may be faced with:

• estimating the detection zone for a species at a study site, where detection zones for more common species are available, but not for the target species;
• estimating the detection zone for a species at a study site where detection zones for more common species are available, but where the practitioner also has a limited amount of measurements for the target species;
• estimating the detection zone for any species at a novel study site for which no detection zone measurements are available at all.

To test the within-sample predictive performance of the EDD and EDA models (scenario 1), we conducted leave-one-out cross-validation (LOOCV), in which we iteratively trained the model with the data for all detection zones except for one, and then predicted the EDD/EDA for the removed entry. We assessed model performance of the two models and reported the MRE value and the R² between the observed and predicted detection zone estimates. We then considered scenario 2, in which practitioners have not met the required number of measurements for accurate detection zone estimation, but they did collect a limited dataset ranging from 1 to 49 measurements. For this case, we investigated whether prediction accuracy can be improved by combining the model output from scenario 1 with an estimate produced by fitting detection functions to the limited data set. To this end, we used the output from the fitted detection functions that were already computed for the power analysis, for which we subsampled the data for each species-and-site specific detection zone 100 times for sample sizes 𝑛 = 10, 20, 30, … 200. Each detection zone produced by such a subsample was first combined with the predicted estimate provided by the scenario 1 model by taking the mean of both. We then computed the relative error between this joint estimate and the ‘true’ estimate that was produced by fitting a detection function to the complete dataset. Then, across all 100 subsamples per species and per site, we took the mean relative error of this joint estimate at each value for 𝑛 to show how, on average, its accuracy improves by increasing 𝑛. Finally, instead of simply taking the mean of the scenario 1 prediction and the estimate produced by fitting detection functions to the subsample, we also tested the effect of assigning different relative weights to either estimate on predictive performance. We therefore calculated the relative error per 𝑛 for eleven different weight distributions, in which we incrementally reduced the weight assigned to one estimate from 100% to 0% in steps of 10%. Finally, for scenario 3 we also conducted leave-group-out cross-validation (LGOCV), in which we iteratively trained the model on the detection zones from all studies but one. The predictive performance was then tested on the detection zones from the study left out to evaluate how the models fare when no site-specific trigger events are available at all. For the LGOCV, we only considered the brands Cuddeback, Browning, Bushnell, and Reconyx, since only these were used by more than one study. Similarly, we excluded primates for the EDA LGOCV due to this taxon being only present in one study that reported EDAs.