Towards good practice guidance in using camera-traps in ecology: influence of sampling design on validity of ecological inferences

Authors


Correspondence author: E-mail: sandra.hamel@uit.no

Summary

  1. The development of camera-traps has provided an opportunity to study ecological relationships and population dynamics of species that are rare, difficult to observe or capture. Their use has seen a major increase recently, particularly with the recent progress in methods adapted to species for which individuals cannot be identified.

  2. We took advantage of extensive camera-trap data sets from large spatiotemporal-scale studies of a diverse assemblage of avian and mammalian scavengers in subarctic/arctic tundra to determine sampling designs that minimize detection errors (false-negative) and to evaluate the influence of sampling design on estimation of site occupancy.

  3. Results showed that raw error rates in daily presence varied between 5 and 30% among species when using time-triggered cameras with a 5-min interval. Using movement-triggered cameras resulted in larger raw error rates, between 30 and 70%, as well as a lower number of daily presences detected. Increasing the time interval from 5 to 20 min greatly increased the raw error rate in daily presence, but it had negligible impacts on estimates and precision of occupancy and detection probability.

  4. Occupancy estimates were mostly influenced by variation in the number of days included during the sampling period. For most species, a threshold of between 20 and 30 problem-free days (i.e. without camera-related technical problems) was required to stabilize occupancy and detection probability, as well as to maximize their precision.

  5. Based on the results, we discuss guidelines for establishing sampling designs according to the different ecological questions researchers might want to answer. To our knowledge, our study is the first to directly test the influence of sampling design in camera-trap studies, providing guidelines that are likely to be directly applicable to a large range of species and ecosystems.

Introduction

The study of ecological relationships and population dynamics can be quite demanding for species that are difficult to observe or capture, like cryptic species or species living in a harsh environment. Even more challenging is the problem of working with species presenting conservation or ethical issues, such as rare species. Numerous non-invasive methods have recently been developed to promote advancement in ecological research on such species (Long et al. 2008; O'Connell, Nichols & Karanth 2011). One major advancement is the development of cameras, which have been used to ‘trap’ animals since their invention in the early years of the 20th century (O'Connell, Nichols & Karanth 2011). While early on wire-triggered pictures were mostly used to list species occurrence (Kucera & Barrett 2011), a wide variety of high-performance cameras with countless options are now available at a very low cost (Swann, Kawanishi & Palmer 2011). For instance, new systems have been developed for activating the camera, such as animal motion or heat triggers, as well as automatic delays or timers allowing pictures to be taken at specific times or intervals (Swann, Kawanishi & Palmer 2011).

The rapid development of camera technology has therefore provided access to undisturbed observations of a vast diversity of species, living under almost any kind of environmental conditions, and at any time of day or year (O'Connell, Nichols & Karanth 2011). While many analytical methods require individuals to be identified, methods have recently been developed for cases where individual identification is difficult or even impossible (Royle & Nichols 2003; MacKenzie, Bailey & Nichols 2004; Rowcliffe et al. 2008; Yamaura et al. 2011). The camera site is often considered as the sampling unit where presence/absence of species is recorded on repeated occasions over a certain period. The detection history specific to each species can then be used to estimate species or community status (e.g. species occurrence, species richness; MacKenzie et al. 2006; Cappo, De'ath & Speare 2007; Rowcliffe et al. 2008; O'Brien et al. 2010). Coupled with site-specific habitat characteristics, this species occurrence information can also be used to answer behavioural ecology questions (e.g. habitat use, activity patterns, foraging habits; Bowkett, Rovero & Marshall 2007; Bridges & Noss 2011; Linkie & Ridout 2011) at the species level or among species. Thus, owing to the evolution of both camera/computer technology and statistical methods, camera-traps are now used to answer questions from a wide range of ecological subjects that were before difficult or impossible to answer in many species (O'Connell, Nichols & Karanth 2011).

Several studies have been published to provide recommendations on the use of cameras in ecology. Many compared camera-traps with other sampling methods (Silveira, Jacomo & Diniz-Filho 2003; Rovero & Marshall 2009; De Bondi et al. 2010; Janecka et al. 2011), whereas others focused on comparing different camera types and their efficiency in trapping animals (Kelly & Holub 2008; Swann, Kawanishi & Palmer 2011). Several other studies also emphasized on the factors affecting trapping efficiency and sensitivity such as sex, species, season and camera position (Rowcliffe et al. 2008, 2011; Guil et al. 2010; Harmsen, Foster & Doncaster 2011; Foster & Harmsen 2012). Although some authors have discussed the importance of carefully selecting the sampling design of camera-traps (O'Connell, Nichols & Karanth 2011), no study was devoted at investigating the influence of the camera-sampling design on detection probability. Of the few studies discussing this topic, some only recommended a specific design without providing an assessment of its performance (e.g. O'Brien 2008; TEAM Network 2011), while others only compared a few specific sampling designs (e.g. O'Brien & Kinnaird 2011). For instance, no study used empirical camera-trap data to thoroughly test the influence of the length of a sampling period and the picture frequency on either raw detectability (when using raw detection rates), or estimator accuracy and precision (when using inference methods such as occupancy models). Detailed recommendations about these aspects of the camera-sampling design are therefore missing. Guidelines aimed to optimize sampling design are likely to depend on the question and the species of interest, and they are importantly needed because using an inappropriate sampling design can result in biased population or community estimations (Bailey et al. 2007). Furthermore, the lifetime of batteries and the size of the disc recording data are almost no more an issue, allowing researchers to collect a tremendous amount of information from millions of pictures. As a result, the large costs and the huge amount of time necessary to process these pictures is now becoming the limiting factor (Harris et al. 2010), and guidance on optimality of the sampling design would be extremely useful.

Here, we take advantage of extensive camera-trap data sets of a diverse assemblage of avian and mammalian scavengers in subarctic/arctic tundra (Killengreen et al. 2012) to evaluate the influence of the sampling design on false-negative detection errors (raw detections) and on estimates of occupancy (accounting for detection probability). These data sets provided species occurrence at baited sites for seven species that differ substantially in abundance and behaviour and for which individual identification is usually difficult. Our first goal was to determine whether we could take fewer pictures in a day without missing the daily presence of a species. Minimizing the error rate on daily presence is essential if the ecological question relates to daily variation in species presence, such as daily variation in habitat use, in activity budgets, and in daytime overlap among species. While we advocate the use of models that account for the probability of detection when detectability is <1, such fine-scale questions are often answered based on raw daily detections, and maximizing detectability is essential when using raw detections. Hence, we quantified the raw error rate in the daily presence of a species for different frequencies of pictures taken during a day. We also compared raw error rates in the daily presence between movement-triggered and time-triggered cameras. We expected to find a threshold after which increasing the time interval between pictures would lead to an increase in error rate. We expected this threshold to vary among species: species that only stay a short time on the bait (e.g. crow) should have a lower threshold than species usually spending a long time (e.g. fox). Secondly, we aimed to determine the suitable number of days required during a sampling period to capture the presence of a species at a site. Minimizing the error in knowing the presence of a species at a site is important if we are trying to determine population or community status, such as species relative density, abundance, occupancy and richness. Because detectability is accounted for in models used to estimate population abundance and species occupancy (MacKenzie et al. 2006), these models must be used rather than raw detections when evaluating the presence of a species at a site. We therefore evaluated the influence of the sampling design on the accuracy and precision of occupancy estimates (both occupancy and detection probability). We expected occupancy estimates to be less sensitive to the sampling design than the raw records of daily presence.

Materials and methods

Study area and design

The data came from three large spatiotemporal-scale studies conducted in Norway, where camera-traps were used to evaluate the structure of the arctic/subarctic scavenger guild. The main study was located in Finnmark county, Northern Norway, hereafter referred as the ‘main’ study/data set (see Killengreen et al. 2012 for details). Between 2005 and 2007, 48 camera-sites were followed each year during the late winter season. A digital camera was positioned at each site with a time trigger allowing pictures to be taken every 10 min for a period of between 72 and 82 days each year (see Appendix S1 for camera details). While still ongoing, we only included the first 3 years of the study as they included the longest number of days and more sites than in the years after 2007. Because the main study only collected information every 10 min, we were also interested in comparing the error rate in the daily presence with shorter time intervals as well as with movement-triggered cameras. Hence, we used data collected from two other large-scale camera-trap projects (hereafter referred as the ‘secondary’ study/data set): one at Børgefjell in Nordland county (17 camera-sites), and another at Knutshø, Dovrefjell and Sunndalsfjella in central Norway (19 camera-sites). As part of these projects, a parallel camera-design was used at 18 sites during the late winter season between 2009 and 2011. The parallel design consisted of two cameras positioned side by side: one camera taking pictures every 5-min and another using a movement sensor activated by the animals (see Appendix S1). In all three projects, a frozen block of reindeer/moose slaughter remains (60 × 40 × 10 cm) was placed in front of each camera between 2 and 4 times at each site. The bait replacement was not made systematically; most sites were visited every 2 weeks, and the bait was replaced when it had been consumed or buried under new snow. Overall, pictures provided presence data from seven arctic/subarctic species: hooded crow Corvus cornix, common raven Corvus corax, white-tailed eagle Haliaeetus albicilla, golden eagle Aquila chrysaetos, arctic fox Vulpes lagopus, red fox Vulpes vulpes and wolverine Gulo gulo. As the white-tailed eagle was almost never observed in the secondary data set, this species was excluded in the analyses conducted on the secondary study.

Because our goal was to determine the influence of the sampling design (i.e. picture frequency and number of days), we only included days without technical problems related to the camera (i.e. snow on the lens, no flash triggered at night, empty battery, etc.) for all data sets. As camera functionality will vary among the study areas due to differences in climatic conditions, our aim is to provide guidelines on how many problem-free days are required, so that researchers can adjust this with the percentage of days a camera is functional in their specific environment.

Data and statistical analyses

Raw detections: error rate in daily presence

Using the main data set, we reduced the recordings of individual species taken at 10-min interval to simple ‘detection/nondetection’ (1/0) for each day and site in each year of the study, and used this as the reference level for this data set. We then determined the daily presence/absence of each species at each site for longer time intervals: 20, 30, 60, 120, 180 and 240 min (where the daily presence based on the 20-min interval was determined using only every second picture, etc.). Although a 240-min interval can seem large, taking a picture every 4-h could be sufficient for large scavenger species that are expected to remain at a carcass in between foraging bouts to defend it until the carcass is less worthy. For each species, we calculated the error rate for each of the longer time intervals as the percentage of day-sites for which a species was classified as present based on the 10-min reference level but absent based on the longer interval.

To evaluate the daily error rate at a finer scale, we used the secondary data set and calculated the daily presence based on 5-min interval. Similarly as for the main data set, we used this 5-min interval as the reference level and then calculated the daily error rate between the 5-min interval and longer time intervals (10, 20, 30, 60, 120, 180 and 240). Furthermore, we computed the daily presence based on the movement-triggered cameras to compare it with the daily presence based on the 5-min interval. One problem with the movement-triggered cameras is that we can never be sure that an absence of a picture means an absence of species; no picture can also result from camera-related problems. For instance, if the sensitivity of the movement sensor is too high, too many pictures (e.g. crows) can be taken in a few days, filling up the memory card. If it is too low, then the movement of an animal in front of a camera might not trigger a picture. Time-triggered cameras, however, always record a picture, except when the camera has a technical problem. Therefore, to compare the two camera systems, we only included days for which the time-triggered camera worked, but starting only from the first day we recorded a picture on the movement-triggered camera. This way, we avoided including early or late days when it was obvious one of the two cameras did not function, as well as days where snow covered the lens of both cameras. Because we do not know which system provides the least raw error on daily presence of species, we computed two reference levels and performed two comparisons. The first computed the daily raw error rate for the 5-min interval using the movement-triggered cameras as the reference level. This means that only days with the presence of a species on the movement-triggered camera were used to assess whether the 5-min interval camera also recorded a presence or not. Therefore, this comparison allowed us to determine whether we are missing the presence of a species when using a 5-min interval. The second comparison computed the daily raw error rate for movement-triggered cameras using the 5-min interval as the reference level. Thus, only days with a presence based on the 5-min interval cameras were used to assess if the movement-triggered cameras also detected a presence. So, if the movement-triggered cameras had no logistic problems, the error rate in daily presence for each species should be null or very low. This comparison therefore allowed us to determine how reliable the movement-triggered cameras were.

Occupancy estimations

We used the main data set to compute estimates of species occupancy based on different combinations of number of days and of number of pictures per day (10, 20, 30, 60, 120, 180 and 240 min intervals) included in the analysis. We only performed this analysis with the main data set because the secondary data set only had a total of 18 sites spread over 3 years, and hence included too few sites per year to run these models. We used the R function ‘occu’ from the package ‘unmarked’ (Fiske & Chandler 2011), which fits the single-season occupancy model of MacKenzie et al. (2002). We only included an intercept for both detection and occupancy (psi[.]p[.] model), as our goal was simply to assess the difference in estimated occupancy for different sampling designs. We included the number of days to estimate occupancy in two different ways. First, we used the number of days starting from day 1 when the first bait was placed out, which represented the inclusion of consecutive days (4, 8, 12, …, 52) with an initial bait session followed by bait replacements at non-specific dates during the sampling period. Second, we used the number of days in groups of 4 days, where calculation for 4 days included the first day of each of the four bait periods, 8 days included the two-first days of the four bait periods, etc. The second calculation represented inclusion of blocks of days after each bait replacement (4, 8, 12, …, 40 bait-days), excluding a certain number of days prior to bait replacement that will vary depending on the time lapse between bait replacements. Therefore, the two calculations represented two distinct sampling strategies, and we expected that using consecutive blocks of bait-days would provide lower error rate than using consecutive days. Then, because it has been suggested that a large number of zeros could result in non-stable occupancy estimates (Sunarto et al. 2012), we also computed occupancy estimates by collapsing the presence/absence matrix into a single presence/absence value for each block of bait-days. Thus, we used 1 if the species was seen at least once, and 0 if it was never seen over the 4 days included in the block (i.e. the first day of each of the four bait sessions, the second day of each of the four bait sessions, etc.). So, instead of having 40 bait-days, we had 10 sampling occasions, each summarizing the presence/absence information of 4 days, 1 day in each of the four bait sessions. Finally, we also wanted to assess the influence of using fewer bait periods, so we also obtained occupancy estimates based on only two and three bait sessions. Note that occupancy models assume that the occupancy of species does not change over the time of the surveys, that is, the system is closed. This assumption is likely to be violated when the survey duration is long or when working with highly mobile species inhabiting large home ranges. In open systems like in our study, occupancy should be interpreted as the proportion of areas used by the species in a given time period rather than the proportion of areas occupied, that is, members of the species are found in the patch on at least one occasion rather than during the entire sampling period (MacKenzie, Bailey & Nichols 2004).

Results

Raw detections: error rate in daily presence

Contrary to our expectations, reducing only slightly the number of pictures taken per day resulted in high error rates in determining the raw daily presence for all species (Fig. 1a,b). Based on the main data set for which the shortest time interval was 10 min, we found that taking a picture every 20 min led to raw errors in daily presence ranging from 4% for crows and ravens to 27% for wolverines (Fig. 1a). Overall, the raw error rates with increase in time interval varied importantly among species and were lower for the four bird species than the three mammalian species (Fig. 1a). Using the secondary data set with 5-min time interval as the reference level, we found even larger raw error rates when going from 5 to 10-min intervals, ranging from 23% in crows to 58% in golden eagles (Fig. 1b). Using the movement-triggered cameras as the reference level, raw error rates in daily presence with a 5-min interval varied from 3% in crows to 33% in arctic foxes (Fig. 1c). When using the time-triggered cameras with a 5-min interval as the reference level, raw error rates in daily presence were much higher, ranging from about 30% in most species to over 60% in red foxes (Fig. 1d). Overall, the time-triggered cameras captured more daily presences than the movement-triggered cameras (compare sample sizes below the x-axis of Fig. 1d vs. c).

Figure 1.

Percent error on the daily presence of scavenger species in relation with different time lapse between pictures (a: main data set with the 10-min interval as the reference level for determining the daily presence of a species, b: secondary data set with the 5-min interval as the reference level), and with different camera systems (c: secondary data set with the movement-triggered camera as the reference level, d: secondary data set with the 5-min interval of the time-triggered camera as the reference level). Numbers in parentheses or below species names represent the sample size from which the percentage was calculated, that is the number of days a species was recorded as present.

Occupancy estimations

Although occupancy models estimate and account for detection probability when estimating species occupancy, the number of days and the number of pictures taken per day influenced both detection and occupancy estimates, as well as their precision (Fig. 2). For species that were almost always absent (e.g. arctic fox) or present (e.g. raven), varying the number of pictures per day or the number of days had only a slight and negligible influence on occupancy estimate and its precision (Figs S1 and S2). For the other species, reducing the number of pictures taken per day (curves in lighter grey in Figs S1 and S2) generally decreased the occupancy estimate and its precision, but using a time lapse between 10 and 30-min did not have a large impact (Figs S1 and S2). Increasing the number of problem-free days increased the estimate and the precision of species occupancy (Figs S1 and S2). In general, a threshold between 20 and 30 problem-free days was necessary for the occupancy estimate to stabilize (Fig. S1), and its precision reached a maximum between 10 and 20 days when the time lapse between pictures remained below 30-min (Fig. S2). The detection estimate and its precision also required between 20 and 30 problem-free days to stabilize for most species (Figs S3 and S4). In general, as the detection probability decreased with increasing number of days, the occupancy estimated increased (Fig. 2). However, when comparing the same number of days but different time lapses between pictures, the probability of detection and the occupancy estimated were both lower with longer time interval (curves in lighter grey in Fig. 2) than shorter ones (Fig. 2; see also Figs S1 and S3).

Figure 2.

Variation in occupancy (1st row), precision of occupancy (width of the 95% CI; 2nd row), detection probability (3rd row), and precision of detection probability (4th row) estimated according to the number of problem-free days included in the analysis and the number of pictures taken per day. The latter is represented by the gradient from black to light grey (black = 10-min interval, light grey = 240-min interval). A value of 1 for occupancy means that the species was estimated as occupying all sites, whereas a value of 0 means it was estimated as not occupying any site. Missing circles indicate cases when standard error could not be estimated or when model did not converge. Estimates are based on 48 sites sampled in 2007 in Northern Norway and are presented for one relatively common and one relatively uncommon species (see Figs S1–S4 in Supporting information for results on all species).

For species and sites for which a bait was placed four times and data were available for at least 10 days at each bait session, we compared occupancy estimated according to consecutive number of days and consecutive blocks of bait-days (Fig. 3). Using consecutive blocks of bait-days markedly reduced the threshold required for the occupancy estimate to stabilize, being below 20 problem-free days based on 10- and 20-min time intervals (Fig. 3). Reducing the number of zeros in the presence/absence matrix using a single presence value for each block of bait-days did not influence estimation of occupancy (compare bottom panels in Fig. 3 with top panels in Fig. 4). Reducing the number of bait sessions included in the blocks did not importantly affect occupancy of raven, but it increased the number of blocks required to reach a plateau in golden eagle and red fox when a 10- or 20-min time interval was used (Fig. 4).

Figure 3.

Variation in occupancy estimated according to two different methods: consecutive days included irrespective of bait replacements (top panels) and consecutive blocks of bait-days (i.e. 4 represents the first day of each of the four bait sessions, 8 the first 2 days of each of the four bait sessions, etc.; bottom panels). Occupancy estimates are presented as in Fig. 2. Estimates are based on the 15 sites in 2005 at which bait was placed four times and data were available for at least 10 days at each bait session. There were not enough data available for the other species and years.

Figure 4.

Variation in occupancy estimated according to the number of blocks of problem-free days and the number of bait sessions included in the analysis. Each block represents the presence/absence of a species in a specific day after bait placement, based on 2, 3, or 4 days depending on the number of bait sessions included (e.g. block 1 in the four bait sessions represents the presence of a species in the first day following bait placement, which is calculated based on the 1st day of each of the four bait sessions; block 2 in the three bait sessions represents the presence of a species in the second day after bait placement, which is calculated based on the second day of each of the three bait sessions, etc.). Occupancy estimates are presented as in Fig. 2. Estimates are based on the 15 sites in 2005 for which bait was placed four times and data was available for at least 10 days at each bait session. There were not enough data available for the other species and years.

Discussion

Camera-traps have become a valuable methodological tool that enables researchers to answer a vast number of questions related to both behaviour and population dynamics of species without too much disturbance, effort or cost (O'Connell, Nichols & Karanth 2011). The limitation today is often not the amount of data, but to gather the adequate information to answer the question at hand. We took advantage of large spatiotemporal camera-trap studies to evaluate the influence of camera-trap sampling design on detection mistakes and precision of estimates. We first evaluated the raw error rate on the daily presence of a species according to variation in the time interval between each picture taken. We then evaluated how occupancy estimates and their precision changed with variation in the time interval between pictures and in the number of days included. Our results clearly show that the best sampling strategy will vary according to the ecological questions the researchers want to answer. With that in mind, we provide some important guidelines on the sampling design required for camera-trap studies to minimize raw detection errors and maximize estimator precision. To our best knowledge, such guidelines have not been thoroughly discussed or provided in the literature, although they will be useful for designing the increasing number of studies aiming to use camera-traps in animal ecology.

Movement- or time-triggered cameras?

Both the movement- and the time-triggered cameras offer advantages and disadvantages (see Swann, Kawanishi & Palmer 2011 for a detailed discussion). The movement-triggered camera ensures capture of an animal as soon as it arrives on sight, as long as the factors affecting the efficiency of the trigger are met (e.g. light conditions, temperature, innate trigger time of the camera) and there are no technical failures with the camera. Yet, unless a delay after a trigger is set, it can take numerous pictures of the same animal. More importantly, one major problem with movement-triggered cameras is that an absence of pictures does not necessarily mean an absence of animals in front of the camera; it could have resulted from the animal movement not triggering the camera or from a technical failure. Time-triggered cameras do not share this problem because pictures are taken at specific times even when no animal is located in front of the camera. You are thus assured that the camera will be activated at each time interval unless it has any technical problems, and the latter will be easily determined from missing pictures. Using a time interval between pictures, time-triggered cameras can control for the non-independence of sequential pictures of the same animals, but it also means that an animal can be missed if the interval between the pictures is inappropriate.

Using a parallel camera system, we showed that we are missing some daily presences when using a time interval of 5 min. Indeed, the time-triggered cameras had recorded an absence up to 30% of the days that the movement triggered cameras had recorded a presence. Nevertheless, the movement-triggered cameras presented more logistic problems, missing up to over 60% of the presences recorded by the time-triggered cameras. Although most studies usually use movement- instead of time-triggered cameras (e.g. Bowkett, Rovero & Marshall 2007; De Bondi et al. 2010; Linkie & Ridout 2011; Royle et al. 2011), it is important to note that movement-triggered cameras recorded overall a lower number of daily presences than time-triggered cameras based on raw detections. Both systems might provide similar accuracy and precision of occupancy estimates, but we could not evaluate this possibility because yearly data sets from the parallel system were not sufficient to run these models. We therefore recommend opting for time-triggered cameras (with a low time interval), unless the aim is to capture rare animals that are likely to pass too quickly in front of the camera to be captured (e.g. ocelots Leopardus pardalis, Di Bitetti, Paviolo & De Angelo 2006; tiger Panthera tigris, Karanth 1995). Furthermore, if the question of the study relates to a specific species, it might be possible to find the best settings to minimize the technical problems related to using movement-triggered cameras, such as finding the best sensitivity level for the trigger. This is unlikely to be possible for studies aimed at multiple species because they will necessarily differ in behaviour, which will result in a trade-off for setting the sensitivity level of the trigger.

What is the best time interval?

The best time interval will depend on the type of inference sought. With respect to raw detections, the 5-min interval cameras recorded the highest number of daily presences compared with the movement-triggered cameras, but raw error rates on daily presence still occurred and they varied importantly among species. Therefore, use of raw detections for direct inference should be avoided because they are very sensitive to the design and are difficult to compare among species and studies. If one does use raw detections, we highly recommend using a short time interval, such as 1-min for instance, or even using video cameras, to ensure reaching a detectability very close to 1. On the other hand, if the aim is to determine the presence of species at a specific site during a certain period, for assessing for instance species occupancy, abundance or richness, then using a time interval of up to 20 min, and even 30 min in some species, provided reliable detection information. Indeed, estimates and precision of occupancy and detection probability were more affected by variation in the number of days than in the time intervals below 30 min. Therefore, if the time to process the pictures is a considerable concern, we recommend using a larger time interval, that is, 20-min, recording pictures over more days rather than using a short interval with fewer days.

How many problem-free days are required?

Our analyses included species that vary considerably in their use of bait and habitat. Although camera-trap performance has been shown to be sensitive to species size, with smaller species being harder to detect than larger ones (Tobler et al. 2008), detection probability was unrelated to species size in our study system. The behaviour of a species rather than its size might be more influential in a white, open habitat like the Arctic tundra during winter, compared with forests (e.g. Tobler et al. 2008). Furthermore, occupancy and detection estimates, as well as their precision, stabilized generally within 20–30 problem-free days for all species. Therefore, we recommend using a sampling design of similar length to maximize occupancy estimation and precision. The sampling strategy could be closer to 20 days if the research is aimed at species that are either abundant or usually seen very early and often during a sampling session, like crow and raven, but it should be closer to 30 days for species that are either less abundant or cover large home ranges, such as red fox and wolverine. Nevertheless, we recommend using a sampling strategy aiming at recording 30 problem-free days if the research is aimed at more than one species, to avoid estimation biases among species because species requiring closer to 30 days will be underestimated compared with species requiring only 20 days.

Occupancy models estimate the detection probability of species and account for it when estimating occupancy (MacKenzie et al. 2006). Assuming a constant detection probability through time, our study showed that the precision of occupancy and detection estimates increased when more problem-free days were included. This reinforces the need to be conservative and include at least enough problem-free days to maximize precision of estimates. With more problem-free days included, however, there was an increase in the occupancy estimated for most species. This likely resulted from the violation of the closure assumption. Indeed, it is important to note that the occupancy estimated in our study represent the percentage of the area used by a species because we worked in an open system (MacKenzie, Bailey & Nichols 2004). To estimate the actual proportion of areas occupied by a species at a specific time, we would need to use only some surveys performed in a short period of time to ensure we would estimate occupancy in a closed system, and thereby avoid overestimating occupancy (see MacKenzie & Royle 2005 for a discussion on allocation of survey effort in such cases). For instance, if a species is rare but covers large areas, the proportion of areas used measured over a long time will be higher than the proportion of areas occupied at a specific time. Thus, the number of days required for modelling occupancy will depend on the goal of the research. Our results clearly illustrate the compromise between the violation of the closure assumption and the precision of occupancy estimates. If the goal of the research is to determine the actual occupancy (i.e. the proportion of areas occupied at a specific time by a species), increasing the number of sites might help increasing the precision of parameters estimated from only a few surveys.

What is the effect of bait?

For studies using bait to attract species, the sampling strategy can consist of using a certain number of days following the placement of baits or using continuous days irrespective of bait replacements. Using continuous days is likely to include fewer presences because the attraction to the bait diminishes with time as it is consumed. Indeed, using certain days after bait placements did stabilize the occupancy estimates over many fewer days compared with using continuous days when enough bait sessions were included. Thus, if the research aims at estimating occupancy for species that are attracted to bait, we recommend using more bait sessions and fewer days in each session. Although we could only test this sampling strategy in three species, it seemed that using four bait sessions every two or 3 weeks and including only the first 5/6 days of each bait session would be an appropriate sampling strategy. Finally, collapsing the presence/absence matrix to reduce the number of zeros did not affect occupancy estimations, but this could not be done for species that were really rare and thereby would have included more zeros. Since reducing the number of zeros by collapsing the presence matrix seems to stabilize the occupancy estimations in other studies on rare species (Sunarto et al. 2012) and because doing so did not affect the estimates for the more common species in our study, this strategy could be useful when working with rare species.

To sum up, we assessed how the sampling design in camera-trap studies can be tailored to maximize raw detectability and optimize estimation of species occupancy. Our research thoroughly tests the influence of the sampling design and provides detailed guidelines in the context of camera-trap studies. The sort of considerations made explicit in our analyses and discussion are likely to be relevant to most species and research questions. Some of our results are also likely to be directly applicable to other species and ecosystems. However, we stress the importance of prior knowledge on the ecology of the species and the ecosystems of interest to devise efficient and robust sampling strategies. When such knowledge is missing, our work highlights the need for empirically assessing different design options before it is possible to arrive at an optimal sampling strategy for the focal question or study system.

Acknowledgements

This research was financed by the Directorate for Nature Management and The Norwegian Research Council. We thank all the fieldworkers that made this research possible over the years, as well as all the biologists that processed the millions of pictures collected each year. We thank Jim Nichols and an anonymous reviewer for constructive comments on a previous version of this manuscript.

Ancillary