Determining the efficacy of camera traps, live capture traps, and detection dogs for locating cryptic small mammal species

Abstract Metal box (e.g., Elliott, Sherman) traps and remote cameras are two of the most commonly employed methods presently used to survey terrestrial mammals. However, their relative efficacy at accurately detecting cryptic small mammals has not been adequately assessed. The present study therefore compared the effectiveness of metal box (Elliott) traps and vertically oriented, close range, white flash camera traps in detecting small mammals occurring in the Scenic Rim of eastern Australia. We also conducted a preliminary survey to determine effectiveness of a conservation detection dog (CDD) for identifying presence of a threatened carnivorous marsupial, Antechinus arktos, in present‐day and historical locations, using camera traps to corroborate detections. 200 Elliott traps and 20 white flash camera traps were set for four deployments per method, across a site where the target small mammals, including A. arktos, are known to occur. Camera traps produced higher detection probabilities than Elliott traps for all four species. Thus, vertically mounted white flash cameras were preferable for detecting the presence of cryptic small mammals in our survey. The CDD, which had been trained to detect A. arktos scat, indicated in total 31 times when deployed in the field survey area, with subsequent camera trap deployments specifically corroborating A. arktos presence at 100% (3) indication locations. Importantly, the dog indicated twice within Border Ranges National Park, where historical (1980s–1990s) specimen‐based records indicate the species was present, but extensive Elliott and camera trapping over the last 5–10 years have resulted in zero A. arktos captures. Camera traps subsequently corroborated A. arktos presence at these sites. This demonstrates that detection dogs can be a highly effective means of locating threatened, cryptic species, especially when traditional methods are unable to detect low‐density mammal populations.

. Camera trapping ultimately removes the need to physically handle an individual and offers a means of detecting rare, elusive, or trap-shy individuals that may be missed by traditional, intensive, shorter-duration live trapping methods (Gray, Dennis, & Baker, 2017;Rendall, Sutherland, Cooke, & White, 2014).
Recently, several studies have demonstrated the success of a modified camera trap mounting design for use in small mammal surveys.
De Bondi et al. (2010) and Gray, Dennis, et al. (2017) demonstrated that infrared camera traps mounted vertically, with the lens oriented toward the ground at close range (<1.5 m), could be successfully used to detect and identify morphologically similar small mammal species. However, accurately identifying individuals to species, especially with infrared cameras, can still be problematic and time consuming, especially for species displaying morphological similarities and/or those with predominantly nocturnal behaviors (De Bondi et al., 2010;Meek & Pittet, 2012;Meek, Vernes, & Falzon, 2013).
An infrared camera flash is less detectable by species when images are taken compared with white flash (Meek & Pittet, 2012). The discreetness of infrared cameras is a result of scarcely observable light that is produced by a selection of light emitting diodes (LED) when an image is taken, triggered by changes in infrared light in the detection zone (Trolliet, Huynen, Vermeulen, & Hambuckers, 2014). However, infrared cameras only produce black and white, and potentially blurred photos, which may result in the inability to correctly identify the photographed species (Glen, Cockburn, Nichols, Ekanayake, & Warburton, 2013;Trolliet et al., 2014).
In such cases, the use of white flash cameras instead of infrared cameras may improve identification accuracy, since the former provide colored photographs (even at night) of a much higher quality (Meek & Pittet, 2012). White flash cameras also offer the prospect of improving speed (and therefore cost-effectiveness) of post-image analysis because of the clarity and color, which allows for faster identification of an individual, especially for species sharing similar traits. Processing time can be a significant drawback for indirect methods that require the user to search through thousands of images or videos (Glen et al., 2013).
Detection dogs have the potential to reduce reliance on other traditional methods such as live trapping. Detection dogs also provide an opportunity for a preliminary survey to determine presence or absence of a species and may be employed to locate the most appropriate specific locations for subsequent trapping and survey efforts (Reed, Bidlack, Hurt, & Getz, 2011). Detection dogs have been successfully used to detect a range of medium-large target mammals (e.g., bobcats, bears, koalas, etc.;Harrison, 2006;Long, Donovan, Mackay, Zielinski, & Buzas, 2007;Romane et al., 2015). However, their utility for detecting smaller, cryptic, and rare mammal species, such as threatened rodents or marsupials, is currently being explored in a range of settings to serve as models for the approach going forward.
The focus of the present study is the black-tailed Dusky Antechinus, Antechinus arktos, a small, recently discovered, federally endangered carnivorous marsupial mammal, located in subtropical rainforests of mid-eastern Australia (Baker, Mutton, Hines, & Dyck, 2014;Van Dyck & Strahan, 2008). The species' distribution is severely restricted, apparently limited to the highest altitude peaks (950+ m altitude) within the Tweed Shield Volcano caldera.
In a suite of foundational ecological studies, Gray, Baker, and Firn (2017), Gray, Burwell, and Baker (2016), and Gray, Dennis, et al. (2017) assessed the breeding synchrony, growth, and detectability of A. arktos. The latter study assessed the feasibility of close range, vertically oriented, baited infrared cameras to detect A. arktos. Gray, Dennis, et al. (2017) concluded that the camera orientation successfully provided close-up images that allowed for species-level discrimination. Both A. arktos and the co-occurring Brown Antechinus, Antechinus stuartii, can typically be distinguished by their differences in size (A. arktos are larger), pelage color (A. arktos are darker), and tail (A. arktos have a thicker, blacker tail). However, A. arktos is mostly nocturnal and returned images in the study of Gray, Dennis, et al. (2017) were black and white. Thus, a combination of stills and video was required to permit species-level discrimination between the rare A. arktos and its (orders of magnitude) more abundant congener, Antechinus stuartii. This process proved arduous and inefficient when multiplied across thousands of records. Gray, Dennis, et al. (2017) did not specifically compare efficacy of live trapping versus cameras for detecting small mammals, including A. arktos. They nevertheless suggested that although both methods had merit, live trapping may be preferable given the challenges inherent in post-deployment image handling and the need to rebait cameras every third day. Other recent studies have also suggested that a combination of detection methods may deliver the most effective and cost-efficient method for locating a target species (Garden, McAlpine, Possingham, & Jones, 2007;Paull et al., 2012).
The ability to accurately and effectively locate a species is imperative to the successful development of any conservation or management strategy, especially for rare or cryptic species. The current study therefore sought to extend previous detection studies by employing  We hypothesized that vertically mounted, white flash camera traps would provide a more effective means of detecting the presence of small mammal species at a given site, when compared with live trapping (Elliott) methods. Further, we hypothesized that the detection dog would effectively detect A. arktos in known locations ( Figure 1).

| MATERIAL S AND ME THODS
The study was conducted with approval from the Queensland  (Hunter, 2003).
The present study used a single overall site containing a pre-existing trapping grid from ongoing parallel research involving A. arktos because the species is known at that location.
Detection dog components of the present work were conducted at a broader scale throughout Springbrook National Park, around and within the live trapping/camera trapping regime areas, and within Border Ranges National Park. Border Ranges National Park is located in the scenic rim, ~140 km south of Brisbane, at the edge of the Tweed Shield Volcano caldera, to the southwest of Springbrook National Park. Border Ranges National Park also exhibits major remnants of UNESCO World Heritage listed "Gondwana Rainforests of Australia." Border Ranges National Park represents a historical site for A. arktos, where confirmed voucher specimen-based records were last retrieved in the late 1980s/early 1990s (Baker et al., 2014).
In the present study, detection dog surveys made use of existing trapping grids at both national parks, along with historical capture sites of A. arktos, and sites displaying ostensibly suitable habitat and/ or climate requirements (see Gray et al., 2016;Gray, Dennis, et al., 2017).

| Experimental design
Camera traps and Elliott traps were deployed across the same spatial scale (the pre-existing trapping grid in Springbrook National Park), which allowed for comparison of detection methods within similar environmental conditions. This experimental design was selected because it attempted to explore the effectiveness of each method, while also maximizing captures and minimizing stress on captured animals (De Bondi et al., 2010;Tasker & Dickman, 2001).
It also allowed for a direct comparison with previous live trapping and camera trapping efforts in the same location. This meant that although displaying different temporal scales, we could compare the effectiveness within the same area under the same conditions and attempt to account for the difference in sampling effort between the two methods.

| Live trapping
Live trapping was conducted across four deployments between 24 June and 18 August 2017 to target the period prior to breeding (September) for A. arktos . A total of 200 type A aluminum Elliott folding traps (Elliott Scientific Equipment, Upwey, Vic.) were employed for the live capture of individuals per deployment, as per previous work on the species Gray Burwell & Baker, 2016). Elliott traps were stationed over a single pre-existing trapping grid at Springbrook National Park, which consisted of four parallel transects (200 m total ~20 m apart). Each of the four transects had 25 tags placed ~8 m apart along their length, and each of these 25 tag locations was considered an individual site during analysis Gray, Dennis, et al., 2017).
Two traps were set at each of the 25 tags, no <1 m apart from each other, along each transect (50 traps per transect) to increase chances of detection in an area typically displaying high levels of small mammal diversity (Dickman, 1986). Best of All Lookout is a known habitat of A. arktos, and the adopted trapping grid has been employed in previous studies involving live capture and camera trapping of A. arktos, making it ideal for the present work .
Antechinus arktos is predominantly nocturnal (Gray, Dennis, et al., 2017), and thus, each individual deployment consisted of three consecutive nights. A mixture of peanut butter, oats, vegetable oil, and bacon was used for bait, which was replenished daily. Traps were set at sundown and cleared as close to sunrise as possible to minimize containment time of individuals and reduce potential overheating/ exposure. Traps were left closed during daylight hours, and bait was replenished each night as necessary. All individuals were identified to species level and released immediately at point of capture. All A. arktos were weighed to the nearest 0.5 g using a Pesola spring balance, sexed, ear clipped, assessed for reproductive condition, and microchipped with a passive integrated transponder (PIT) tag for recapture ID. Antechinus stuartii were also weighed, sexed, and assessed for reproductive condition.

| Camera trapping
Camera trapping was conducted across four deployments between 30 June and 21 August 2017 to target the period prior to breeding for A. arktos , as for the live trapping component of this study. Each site was trapped commencing one night after live trapping events to allow individuals a period of regular foraging. Modified Reconyx PC850 Hyperfire Professional White Flash Camera Traps were used. Camera traps were set for three consecutive nights per deployment, as in the live trapping component of the study. Camera trap focal length was factory pre-set to 70 cm and the detection grid altered to match this shorter focal length. One camera was set at every fifth tag along each of the same four transects used in the live trapping component of the study, and thus, a total of 5 cameras were spread evenly for each of the four transects, constituting 20 cameras in total across the trapping grid. This represents a small spatial scale to deploy cameras; however, A. arktos displays a very patchy and limited distribution in steep, difficult-to-traverse terrain . The utilization of the single trapping grid composed of four transects at Springbrook National Park for both components also facilitated comparison of detection methods, as they were deployed in the same area, within days of each other.
In previous studies, cameras were typically mounted horizontally, often attached to structures such as trees, to effectively capture animal profiles (Sweitzer, Vuren, Gardner, Boyce, & Waithman, 2000). For the current study, however, we employed a revised camera mounting strategy to effectively detect small terrestrial mammal species, modifying the design of Gray, Dennis, et al. (2017). Camera units were mounted vertically on trees at a height of 70 cm, to produce photographs of the zone of ground directly underneath the camera ( Figure 2) and attempt to correlate camera height with species body size/mass. To reduce the potential for false triggers and ensure unobstructed photographs, vegetation and dense leaf litter around the detection grid were removed prior to deployment.
Camera units were set (running) for the entire three-day deployment, recording three photographs for every trigger, with a one-second interval between each photograph in the same trigger event.
These three photos were subjectively defined as a single photographic "event." These settings were similarly selected to provide photographs of the individual in different positions and ultimately increase identification accuracy. Camera units were operational 24 hr per day, for each day of field deployment to account for potential diurnal activity . When processing photographs, once a species had been recorded, individuals of the same species were not recorded for another 10 min to minimize the possibility of detecting the same individual again and maximize the potential to record individuals that rarely visited camera units (Gray, Dennis, et al., 2017;Weerakoon et al., 2014).
Photographic events were recorded as binary response variables: "1" was used to indicate that the species had been observed and "0" used if the species was not observed. Data were collected in this manner for each camera, for each of the days across the entire three-day deployment period (i.e., all photographic events taken were recorded this way except for individuals seen within 10 min of each other). Data were then pooled for each deployment at the trap level such that all data from day one to three of each deployment were pooled to indicate that the species was detected or not detected for that deployment. We acknowledge that this does not fully account for potential camera "recaptures" (i.e., individuals that may reappear at various times in the night under the same or different cameras). Similarly, our analysis could not account for recaptures of individuals of various species (other than PITtagged A. arktos in Elliott traps), either between nights or deployments.
Bait used for the camera units was identical to that used in live trapping methods (see above). Bait containers were composed of one 50-mm mosquito proof vent cowl (W: 62 mm, H: 84 mm) and one greywater hose adaptor (W: 57 mm, H: 62 mm) purchased from Bunnings Warehouse (Brisbane, Australia). These were spraypainted black (to prevent under-exposed photos) and secured to the ground using three, large tent pegs. This bait container set up allowed for individuals to view and smell the bait but not remove it; therefore, bait did not have to be replenished during each deployment. It also assisted with identification of individuals because of the known dimension of each bait container (62 mm diameter).
Photographs were processed, and all individuals were identified to species level based on pelage color/appearance, body size and shape, head size and shape, and tail length/appearance. If identification was not possible because of blurred images or lack of distinguishable features, it was classified as "unknown" and not used in subsequent analyses. A second examiner ensured identification accuracy.

| Detection dog
The conservation detection dog was trained by a certified professional dog trainer (Certification Council for Professional Dog Trainers (CCPDT)). Dog obedience and socialization training began at nine weeks old, with odor detection training on A. arktos scat commencing at 12 months of age. The dog was trained to associate the scent of A. arktos with a reward (e.g., food, tennis ball) and to freeze ("indicate") at the location of the odor. The dog was trained on a few A. arktos scats from previous live trapping studies at the current study site, and again trained for one month after fresh scat, hair and live trap bedding material used by captured  Field trials in Springbrook National Park were located within the trapping grid employed for live and camera trapping aspects of this study and nearby areas (within 1 km of the trapping grid), which displayed presumed suitable habitat or climate requirements for A. arktos (Baker et al., 2014;.
The detection dog surveyed all four transects, and the surrounding areas off-leash.
The remaining locations used for canine field trials occurred throughout Border Ranges National Park. These locations represented areas where A. arktos is known to have historically occurred, including historical capture sites (Appendix 1) and where recent (unsuccessful) live trapping and camera surveys have been undertaken semi-regularly (Baker et al., 2014;Gray Burwell & Baker, 2016;Gray, Dennis, et al., 2017), as well as areas that displayed presumably suitable habitat and climatic requirements from which the species was unknown. As for Springbrook National Park, where possible, in Border Ranges National Park the detection dog team followed transects from a trapping grid employed in previous studies or surveyed freely in areas away from the trapping grid. The search focus was in and around several locations where voucher specimens had been historically retrieved (see Baker et al., 2014).
The detection team recorded the GPS location of each "indication" (i.e., each time the dog indicated that it had detected the odor and displayed a response), which included latitude, longitude, time, date, and elevation. During the field deployment, the detection dog was fitted with a GPS tracking collar to review search patterns after field deployments. The dog's indications were also marked by flagging tape for later confirmation of accuracy using camera traps. Any positive A. arktos indication by the dog was immediately investigated for the purposes of locating the scat or animal source of the odor.

| Data analyses
Both camera trapping and live capture data were recorded in the form of repeated occurrence observations of a species, at fixed sites.
A single site was defined as a local (1-2 m) region around a detection method, that is, either a fixed camera or Elliott trap. Therefore, multiple "sites" existed within the trapping grid. Sites were sampled for three nights over each of the four deployments. Data were pooled for each deployment at the trap level (i.e., three days of observations were pooled per deployment for each individual trap).
Occupancy modeling was conducted using the program PRESENCE version 2.12.24 (Hines 2006) for comparison of camera and live trapping methods for all species. Single-season occupancy models (MacKenzie et al., 2002;Nichols et al., 2008) were employed in the current analysis to estimate the probability of detecting each species for each deployment with each method. Covariates used were as follows: deployment, species, and method.
Two models were run in PRESENCE to determine the effect of detection method, deployment variation, and species on detection probability. The first model constrained detection to be the same for both sampling methods, whereas for the second model detection was dependent on sampling method. The first model was, and the second model was, where, is the logit function, p is the probability of detection, β i is the estimated parameter for the ith deployment, β j for the jth species and β k for the kth method, and, is the indicator function, which is used to encode the different categories from the data into the model. AICs were compared between models to determine the model that minimized the amount of information lost from the data.

| RE SULTS
The model that accounted for the variation in method was selected for analysis, as the AIC value (1,827.9) was lower than the AIC (2,163.04) for the model that did not (Table 1). This indicated that there was a better fit to the data for the model where detection was dependent on sampling method.

| Camera trapping general findings
The camera trapping survey period consisted of 240 trap-nights, across the total of four deployments (Table 3). Over all four deployments, mammal captures consisted of 3,893 photo events representing eight different mammal species, including four-target species: Antechinus arktos, A. stuartii, Rattus fuscipes, and Melomys cervinipes (Figure 4).
The remaining nontarget species were as follows: Perameles nasuta, Trichosurus caninus, Thylogale thetis, Felis catus, and multiple rainforest birds. Because of the inability to live trap these larger nontarget species, they were excluded from subsequent comparative analyses.
Camera trapping for most species displayed high detection probabilities ( Figure 3). Antechinus stuartii was captured most often of all four species, contributing 30.1% of all mammal captures ( and 60.7% across the four deployments [ Figure 3]). Because of the high detection probabilities observed during the study, the camera trapping aspect appears to display an adequate sampling duration to detect a rare/ cryptic species such as A. arktos.

| Detection dog findings
Between August and September 2017, the detection dog team sur-  Capture refers to detecting the species in at least one photograph out of the three photographs taken in each "event."

| Mounting design and type of remote cameras
Camera-based species detection relies implicitly on the ability to correctly identify a species and be confident of distinguishing it from similar taxa. Several studies have demonstrated the difficulties associated with identifying coexisting small mammal species from photographs taken using horizontally oriented camera traps (e.g., Glen et al., 2013;Meek & Vernes, 2016). One recent study found that horizontally positioned white flash digital camera traps produced photographs that did not provide enough precision of certain diagnostic features to enable accurate identification of some rodent species (Meek & Vernes, 2016 Similar issues were encountered by Meek et al. (2013), resulting in a lack of clarity when differentiating certain aspects of an individual from the background, which ultimately affected identification ability.
However, a previous study employing a vertical mounting design found that even after clearing vegetation and leaf litter around the detection grid, difficulties remained in discriminating the tail of individuals from blades of grass or twigs (De Bondi et al., 2010). The current study avoided these problems by carefully clearing most grasses, leaf litter, and vegetation within the detection grid, leaving the ground relatively bare, which facilitated identification of individuals, especially in cases where only part of the animal's body was in frame. Such issues are habitat-specific, and we were fortunate that rainforest leaf litter was easily manipulated, reducing handling time during camera deployment.

F I G U R E 4
Cropped white flash camera photographs documenting all four small mammal species encountered during survey periods. Images were cropped for the purpose of the paper to ensure individual differences were observable. Equally, to continue the improvement and development of camera trapping surveys, researchers must ensure that survey methods involve a clear emphasis on the fundamental processes of animal abundance, movement, and detection and should incorporate a more detailed treatment of methodological characteristics and assumptions. This clarity will enable future researchers to assess, compare, and develop the effectiveness and reliability of detection camera trapping surveys, further strengthening our ability to answer vital questions and close knowledge gaps (Burton et al., 2015).

| Factors influencing trap success
Detection probabilities varied notably between detection methods.
The probability of an individual being successfully detected ultimately depends on a multitude of factors. Effectiveness of a trap, the probability of an individual locating a trap, and the individual's reaction to the trap, among other factors, all have significant influences on trap success (Tasker & Dickman, 2001). In the current study, the observed variation in detection probabilities between species may have similarly been caused by additional factors such as bait strength, attraction to bait, trap-shyness, or various biological factors influenced by population dynamics (Gray, Dennis, et al., 2017).
Antechinus stuartii, M. cervinipes, and R. fuscipes all displayed high detection probabilities within and across deployments for both trapping methods, relative to A. arktos. All three of the common species occur in high density within the trapping site and taken together represent the major components of terrestrial, small mammal taxa within subtropical rainforests of southeast Queensland (Wood, 1971). Thus, the relatively high probabilities of detection, regardless of trapping method, for these three small mammal species compared with that of A. arktos are unsurprising.
However, there were some differences in relative detected species abundance between our study and previous work at the same site.
In previous studies that employed live trapping or camera trapping methods within the same area, R. fuscipes produced the highest detection probabilities Gray, Dennis, et al., 2017). Live trapping components of the study performed by  occurred across a two-year period, between April and October, and the camera trapping component of the study by Gray, Dennis, et al. (2017) Leung, 1999;Wood, 1970).  found that A. stuartii exhibited no notable difference in trap success between years in their study, but they observed the highest trap success of this species in July. Field work for the current study occurred around the same period (all field work being completed during July, or within a few weeks before or after July).
Antechinus stuartii also displayed a decrease in % total mammal captures between the two detection methods in the current study. Antechinus stuartii was the smallest species to appear in both detection methods and may have sometimes been deterred from investigating bait containers, or chased away by larger, more dominant species. In previous studies, A. stuartii has exhibited a heightened alertness when more dominant species are present and even fled when certain species approached (Dickman, 1991;Gray, Dennis, et al., 2017 This suggests some small mammal species persist at very low density and may remain undetected even after intensive and strategic surveys using concerted live and camera trapping. Presumably, at Bar Mountain Picnic Area, our dog detected an A. arktos either at its subterranean nest, or based on scent overlaying, a recently used foraging trail. Such site-specific microhabitat use proved the difference in targeting the exact camera location to capture footage of an A. arktos the following night.

| Management implications
The development and growing understanding of detection methods has provided a variety of novel opportunities for gathering a plethora of ecological and biological information. Information provided by these detection methods may consequently prompt changes to management or policy. Therefore, without strategic selection of method and experimental design, researchers may encourage the use of survey methods or monitoring designs that fail to effectively complete the objective or detect the target species (Clare et al., 2017). The choice of detection method will primarily depend upon the management objectives, as each method exhibits its own specific benefits and drawbacks, so the detection method used should be survey specific. Our findings demonstrate the utility of white flash, vertically oriented, remote wildlife camera traps and conservation detection dogs when compared to the traditional method of live trapping (Elliott traps) for locating rare and cryptic species.
Detection dogs provide a unique opportunity to locate target species, which may be overlooked by traditional methods. The ability of dogs to provide more detailed, fine-scale information (e.g., between sexes or even individuals) warrants further exploration. If we are to mitigate the risks that endanger our mammals, it is imperative that we successfully survey rare species to enable the development of effective conservation, management, and monitoring programs.
Indirect methods, such as detection dogs, will become an increasingly important tool in achieving this goal.

ACK N OWLED G M ENTS
The current project was generously supported by funding from and Artie Ziff, who all graciously volunteered their time to help with certain components of field work, quantitative methods, or provided other contributions to the study.

CO N FLI C T O F I NTE R E S T
None declared.

DATA AVA I L A B I L I T Y S TAT E M E N T
All observational (Camera trap/Elliott trap) data for this study will be available from the Dryad Digital Repository, at: https ://doi. org/10.5061/dryad.7h44j 0zqr.