Assessing the camera trap methodologies used to estimate density of unmarked populations

Obtaining accurate estimates of population density (i.e. the number individuals per area) continues to be a constant challenge for wildlife management and conservation (Nichols & Williams, 2006). It is widely recognized that estimating size costly, results are often not sufficiently precise well-informed purposes (Morellet et al., 2007). Feasible methods with which attain are, therefore, in great demand. The use remotely triggered cameras (camera traps) this purpose has substantially increased over last few years (Rovero Zimmermann, 2016). Camera trapping relatively low cost (except initial inversion), generates information on multiple species, minimally invasive makes it possible obtain highly cryptic species inhabiting wide range habitats (Steenweg 2017). As result, camera increasingly core tools monitoring (e.g. wild boar at European level, ENETWILD-consortium 2019). When densities target parameter, frequently used data spatially explicit capture–recapture models require individually identifiable animals (Royle 2013). However, many have no natural markings allow individual identification, so physical capture marking apply family trap data. In case, application such (in broad sense) can expensive, logistically challenging. Validating using traps absence broaden their applicability monitoring. context, there different estimate (number animals) without recognition: Time Event Model (Moeller 2018), Random Encounter (REM; Rowcliffe 2008), Spatial Counts (Chandler Royle, 2013; Evans Rittenhouse, Distance Sampling based (CT-DS; Howe 2017), Staying (REST; Nakashima 2018) and, more recently, model considers species’ space (Luo 2020). Most these modelling encounter rate, main divergence point between them procedure address effective sampling frame broader study area about one wishes make inference). While some abundance within an explicitly defined by accounting when where detected 2013), others collective field view (FOV) representative (Rowcliffe 2008). A in-depth theoretical comparison unmarked was described Gilbert al. (2020). empirical under conditions lacked demanded. We compared performance three compatible design: REM, REST CT-DS. CT-DS Sampling, framework considered as method (Buckland 2001; Thomas 2010). Considering robust framework, specific software advice design, testing could considerably increase populations. To date, been applied chimpanzee (Cappelle 2019), community rainforest (Bessone 2020, Cappelle 2021), bighorn sheep (Harris 2020) marmot (Corlatti REM is, doubt, most (Gilbert Over years, (Cusack 2015; 2019; Pfeffer 2018; Zero constantly developing (Caravaggi, 2017; Lucas 2015). Its originally limited need day distance travelled during day). approaches exclusively from were (Palencia 2021; 2016), significantly its applicability. (2018) extension but we would like highlight mathematical equivalence (Appendix S1). staying time amount remain FOV trap) instead range. This enhance relative time-consuming parameter obtained REM. forest ungulates (Nakashima tested human volunteers (Garland summarized assumptions Table 1. common feature they do spatial autocorrelation captures mark–recapture related animal does probability captured than trap). allows designs any spacing traps, therefore larger areas sampled single surveys. Moreover, share assumptions, analyses rescaling rate movement parameters (REM REST) or detectability process (CT-DS). respect, clarify key points. Regarding closure should noted if change survey, will provide average across period. difference relation methods, violations result detection too being small, generating positively biased (Obbard certain 0, negatively proportion p(0). For instance, p(0) = 0.30, only 30% true (Borchers 2002). Similarly, focal area, underestimated. minimize set appropriate height, activated faster possible. Second, spite all assumed behaviour affected each deals assumption way. those sequences react speed clear violation Those leaving zone suppose underestimation REST, CT-DS, consequence, density. stand zone, (REST) (CT-DS) inflated. Also, severe independence events expected because detections same considered, avoided variances nonparametric bootstrap, resampling points replacements; applying selection criteria choose best (Howe work, consistency, precision cost-effectiveness estimation Mediterranean environment, six populations (red deer Cervus elaphus, Sus scrofa red fox Vulpes vulpes), spanning behavioural traits densities. Additionally, populations, independent line transects drive counts. carried out two Spain distinct environmental conditions. One (site-A), Doñana Biological Reserve, territory approximately 6,800 ha located National Park (37°0?N, 6°30?W). Atlantic coast south-west Spain, dominated shrubland 300 marshland. climate thermomediterranean, marked seasons. altitude 15 m a.s.l. second (site-B) Montes de Toledo (39°23?N, 4°4?W), mountain chain central Spain. Vegetation woodland (Quercus spp.) ‘dehesas’ (savannah-like composed pastures mainly include oak trees). mesomediterranean, 900 all, 25 Bushnell Aggressor Trophy Cams site-A (from September till November 20 site-B 2018 April 2019) covering area. systematic design random origin. deployed facing north intersection grid 2 km (site-A) 1.5 spacing, 40–50 cm above ground angled parallel slope ground. baited. Realized locations deviated much 80 order mount trees avoid very unfavourable dense shrubland). Only had outside buffer. During installation trap, marks (rocks, branches, etc.) placed 2.5 intervals (Figure 1). These later locate position trap. operative day, record burst consecutive photos (rapid fire) activation, minimum triggering interval activations (0.6 s). date automatically stamped onto image. checked every 3 weeks batteries memory cards. consider recorded observation distances s 2, 4, …). Following exploratory analyses, left-truncated fewer near 2001). right-truncated lower 0.1. pass through estimated replacement followed (2019) select models. evaluate method, taken images estimation) carry analysis method. Finally, values statistically Wald test, test statistic W assessed chi-squared distribution degree freedom (Wald Wolfowitz, 1940). line-transect sampling, gold standard reed (Acevedo Surveys rutting season (September), began hr before sunset four-wheel-drive vehicle 10 km/hr A, foot B. total surveyed considering 71 51 B, respectively. Each transect repeated twice. group detected, observer telemeter, well composition sex age classes stratified convectional sampling. strata both characterized high good visibility (open habitats), other poor (woody habitats). Data eliminate 10%–15% (approx.) furthest observations Half-normal, hazard uniform cosine, hermite polynomial simple fitted function. adjustment term AIC population, counts (ENETWILD-consortium Concretely, January scrubland zones overlapping hunting activities. 697.21 ± 74.42 (SE), 37.33 observers 5.37 (SE). Observers fixed open firebreaks). count duration 4 11:00 15:00), while locations, 44 beaters 5.29 (SE) 440 dogs 52.92 moving area; assuming detected. coordinator collected likelihood double counting. Density dividing observed survey moment (midday), woodlands areas, grassland (dehesas) 0. supported telemetry tagged (E. Laguna, unpubl. data). ranged 0.23 individuals/km2 fox—CT-DS—site-B) 34.87 deer—REM—site-B; Figure 2). shown 2. found significant positive correlation (per site, n 6) (Pearson correlation: REM-REST: R 0.87, p 0.025, 6; REM-CTDS: 0.93, 0.0063, REST-CTDS: 0.88, 0.02, 6). differences test: vs. REM: 5.63, 0.02; REST: 4.90, 0.03). Significant among general, tend higher estimates, did find coefficients variation (CV) (ANOVA measurements, p: 0.698). CV 0.28, 0.36 0.42 effort required (Table 2), least terms image processing carrying analysis. processing, surveys, groups 259 observed. Best half normal cosine hazard-rate (site-B). 10.68 2.31 22.94 2.98 sites counts, 66.33 22.81. After discarding 8.24 2.18 methods. site-B, 0.50, 0.48), 8.33, 0.004) (W 9.13, 0.003). possibility increases situations obtained. study, methodologies (REM, CT-DS) either identification captures. Two them, far little (but see Bessone 2020; Our show correlation, five monitored. comparing available population. suggests consistency especially suggest may underestimating currently hypothesis why is. general tendency generate generated equivalent results. pattern Previous underestimate 2) reinforces partially explained consequence malfunction problem traps. Despite bursts photos, around 12% took further 9% activations, longer S4). Furthermore, evaluations concluded recovery manufacturer's rating s) consistent (https://www.trailcampro.com/products/bushnell-trophy-cam-hd-low-glow?_pos=40&_sid=6a15c66a4&_ss=r). problems case because, sequences, effectively complete trajectories view. (Y, Equation 3). (if times entry exit recorded) entered left photographed). influence (a) photo enough sequence (b) time-lapse travel estimation. recommended reliably fast trigger rates registration inside FOV, Any compromise (McIntyre precision, previous studies uncertainty associated variance including activity level. CVs ranging 0.28 attributable reduced challenging small covered camera; means strongly influenced microsite larger-scale habitat characteristics pre-stratification. covariates also improve precision. increasing effort, placement (Schaus Eventually, protocol adaptive surveys applied. consists conducting additional recorded, useful distributed patchily sparsely (Buckland, 2004). Simulations situation 20–25 yield coefficient 0.40 (figure our (mean estimates). Accordingly, 0.20, (Williams 2002), 100 (2020) 0.37 750 locations. (2021) variety achieve 0.10 0.20 days 50 placements. Future needed improvement regarding place then move new location) cost-effective approach reduces devices. explored here free-standing, derived necessary references auxiliary habitual practice (Caravaggi 2016; Cusack Manzo 2012). data, usually underestimates Derive identify reference objects known 1), costs similar distinctly least-time consuming, analysis, crossed triangle wider Although strong limit low-density monitored, efforts, estimates. On hand, 2020), snapshot s, analysing peak pattern, active proposed monitor records spends front analyse goodness-of-fit critical discuss how deal since reactions spent camera, bias. adjust evident estimations, respectively (see ‘Shared features methods’ section Materials methods), consensus attraction avoidance induce bias Because this, authors discard first period become accustomed 2017); discarded indicated reaction reasonable function fitting opinion, none totally solves induced reactive animals, research solve conclusion, compared. showed sampled), precision; clearly enhances priority future development placements) notably utility optimization records, scenarios abundant rapid response time, preferable less risk case. But Authors thank Eric Yoshihiro help kindly comments acknowledge Park, Ángel Moreno Amanda García providing logistic support study. work partly funded ENETWILD project (http://www.enetwild.com). P.P. received MINECO-UCLM FPU grant (FPU16/00039) Society Spanish Researches United Kingdom UCLM ‘On Move' grant. Comments S.T. Buckland, anonymous reviewer editors helped us upon earlier drafts. P.P., J.M.R., J.V. P.A. conceived designed methodology; conducted survey; J.M.R. analysed data; managed writing manuscript. All contributed critically drafts gave final approval publication. via Zenodo https://zenodo.org/record/4745594#.YJkJ1j_tbIU (Palencia, 2021). case-by-case basis request corresponding author. Please note: publisher responsible content functionality supporting supplied authors. queries (other missing content) directed author article.