Effectiveness of camera traps for quantifying daytime and nighttime visitation by vertebrate pollinators

Abstract Identification of pollen vectors is a fundamental objective of pollination biology. The foraging and social behavior of these pollinators has profound effects on plant mating, making quantification of their behavior critical for understanding the ecological and evolutionary consequences of different pollinators for the plants they visit. However, accurate quantification of visitation may be problematic, especially for shy animals and/or when the temporal and spatial scale of observation desired is large. Sophisticated heat‐ and movement‐triggered motion‐sensor cameras (“camera trapping”) provide new, underutilized tools to address these challenges. However, to date, there has been no rigorous evaluation of the sampling considerations needed for using camera trapping in pollination research. We measured the effectiveness of camera trapping for identifying vertebrate visitors and quantifying their visitation rates and foraging behavior on Banksia menziesii (Proteaceae). Multiple still cameras (Reconyx HC 500) and a video camera (Little Acorn LTL5210A) were deployed. From 2,753 recorded visits by vertebrates, we identified five species of nectarivorous honeyeater (Meliphagidae) and the honey possum (Tarsipedidae), with significant variation in the species composition of visitors among inflorescences. Species of floral visitor showed significant variation in their time of peak activity, duration of visits, and numbers of flowers probed per visit. Where multiple cameras were deployed on individual inflorescences, effectiveness of individual still cameras varied from 15% to 86% of all recorded visits. Methodological issues and solutions, and the future uses of camera traps in pollination biology, are discussed. Conclusions and wider implications: Motion‐triggered cameras are promising tools for the quantification of vertebrate visitation and some aspects of behavior on flowers. However, researchers need to be mindful of the variation in effectiveness of individual camera traps in detecting animals. Pollinator studies using camera traps are in their infancy, and the full potential of this developing technology is yet to be realized.

In pollination research, camera trapping may be particularly effective at overcoming limitations in the capacity of direct human observation to detect reclusive pollinators. Further, the time for which observations can be made is greatly increased, allowing constant monitoring of flowers over many days, weeks, or even months.
Importantly, camera traps also provide untested potential for a detailed quantification of visitation rates and behavior, enabling new insight into the consequences of pollinator behavior for plant mating.
The variation in foraging strategies and social behaviors that affect pollinator movements can have profound effects on plant mating (Krauss et al., 2017;Mitchell, Irwin, Flanagan, & Karron, 2009), making quantification of pollinator behavior important for understanding the ecological and evolutionary consequences of different pollinator groups. In this way, camera trap data can be fully utilized to explicitly test ideas or hypotheses, rather than merely estimate abundance or density.
While the potential of camera trapping as a method for pollinator detection is clearly vast, there has been no rigorous evaluation of the sampling considerations needed for this type of study. Such an evaluation will be important for understanding potential issues such as the number of replicate camera traps needed, their effectiveness at detecting different groups of organism, the effect of camera setup (e.g., distance, angle), and ambient conditions (e.g., time of day, temperature), and the data they can reliably collect (Jumeau, Petrod, & Handrich, 2017). Here, we aimed to (a) measure the effectiveness of camera trapping as a method of identifying vertebrate visitors and (b) quantify visitation rates and timing of pollinator visits and (c) resolve which aspects of foraging behavior could be quantified.
We focused on Banksia menziesii (Proteaceae), a species visited by multiple bird, mammal and insect species, but primarily reliant on vertebrates for pollination (Ramsey, 1988a(Ramsey, ,b, 1989. Our approach also addresses the recent call (Burton et al., 2015) for more thorough reporting of methodological details to facilitate efforts to evaluate and improve the reliability of camera trapping surveys.

| Study system
Banksia menziesii (Proteaceae) is a common tree or woody shrub of Banksia woodlands, a threatened ecological community endemic to sandy soils of southern Western Australia (Collins, Collins, & George, 2008). Flowering occurs from February to October with a peak in June. Inflorescences are most commonly red, but yellow or pink variants occur in some populations. Each inflorescence has 600-1,400 nectar-producing flowers arranged orthogonally around a central woody axis of up to 12 cm in length, with ca. 40-60 flowers open on any single day (Figure 1; Ramsey, 1988a).
Banksia menziesii is self-incompatible, so dependent on pollinators facilitating cross-pollination for seed set to occur (Ramsey & Vaughton, 1991 (Brown et al., 1997;Houston, 2000;Ramsey, 1988bRamsey, , 1989). For still images, we used the RECONYX Hyperfire HC500 (http:// www.reconyx.com/product/HC500-HyperFire-High-Output-Covert-IR), which is a mid-price-range camera capable of detecting small animals. This camera has an image resolution of 1080P high definition and passive infrared sensor to detect a differential in heatand-motion between a subject and the background temperature, and a "low-glow" infrared flash array. Recorded temperatures ranged from −3 to 32°C, with a mean of 18°C. Cameras were mounted at varying heights up to 1.5 m on star pickets using zip ties and positioned at the same height as the inflorescence approximately 60 cm away ( Figure 1). Inflorescences just beginning to bloom were arbitrarily chosen for monitoring so as to collect visitation data throughout the entire life-span of each inflorescence.

| Camera trapping
Camera settings were as follows: sensitivity = high; pictures per trigger = 10; picture interval = rapid-fire; quiet period = no delay.
These settings armed the camera to take 10 photos over ca. 9 s when triggered by motion, with a trigger speed of 1/5th second. The cameras continue to capture bursts of photographs as long as there is movement detected, and capture images day and night. On average, cameras (or batteries -we used rechargeable AA Panasonic eneloop batteries) were changed every 2 weeks, and digitally stored photographs downloaded to a computer. Overall, 20 cameras were used (on some inflorescences we employed multiple camerassee below), and dates and location of cameras and inflorescences recorded.
Downloaded photographs were scored manually for the presence of vertebrate visitors to inflorescences. Individual photographs were imprinted with date, time, photograph number in the series of 10, temperature, and camera number. For each visit captured by cameras, species, date, start time, finish time, duration of visit, number of flowers probed, inflorescence flowering stage, and temperature, were recorded. Inflorescence flowering stage identifies the cumulative proportion of flowers that have opened on an inflorescence, so, for example, a proportion of 0.1 indicates that approximately 10% of flowers have opened from the base of the inflorescence. For Western Spinebills, males and females were distinguished by clear differences in plumage. When a visit was longer than one series of 10 photographs (ca 9 s), duration of visit was estimated from the arrival time (typically photo 1 of 10) and then the time of departure, which often included a short but variable lag period (typically 2-30 s) between the final photograph in the first series and the first triggered photograph of the next series.
We tested for variation in visitor composition within and among inflorescences with χ 2 tests using SYSTAT v13 software. The number of probes was estimated from photographs and contrasted to accurate estimates of probe rate obtained from additional video camera footage (see below). Differences in the mean duration of visits by each bird species to inflorescences were assessed by oneway Analysis of Variance and post hoc Tukey tests using SYSTAT v13 software.

| Quantification of camera effectiveness in visitor detection
At most inflorescences, multiple cameras (2, 3 or 4) were deployed at equal distances (60 cm) from the same side of an inflorescence. This overlap in monitoring enabled an assessment of the accuracy of cameras based on the number of known visits that went undetected by a given camera. Data from multiple cameras on the three inflorescences with the greatest overlap in recording was assessed to generate a relative effectiveness index for each camera, calculated as the number of visits captured by a single camera divided by the total number of visitors captured across all cameras at that inflorescence, and multiplied by 100 to convert it to relative percentage effectiveness. Thus, relative effectiveness index measures the percentage of known visits recorded by a single camera.
For each video, date, start time, finish time, duration of visit, number of floral probes and the species visiting was recorded. Mean visit length and mean number of floral probes per species per second was estimated.
When videos and cameras captured the same visit, we contrasted duration of visit and mean probe rate for all birds to assess whether duration and probe rates captured by cameras underestimated the values as determined from videos, and assessed the significance of differences by dependent t-tests for paired samples using SYSTAT v13 software.
On individual inflorescences, bird visits increased gradually to a maximum visitation rate (18% of total visits) at mid-inflorescence flowering (i.e., when 50% of the flowers had opened), and declined  Honey possums were recorded throughout the night, with a peak in the hour after sunset (Figure 4).

| The behavior of floral visitors
Only three intra-or interspecies aggressive interactions between birds were recorded. All vertebrate visitors were recorded probing flowers and therefore assumed to transfer pollen. Mean recorded duration of visits by birds to inflorescences differed among species (ANOVA; F = 29.7; p < 0.001; df = 4 (note White-cheeked Honeyeater was excluded from the analysis due to too few data (N = 4)). From still photos, the mean (±SE; N) recorded duration of visit to an inflorescence per species was significantly greater   (Newland & Wooller, 1985), but contrasted to another study that found Brown Honeyeaters, New Holland Honeyeaters, and Western Wattlebirds were frequent visitors to B. menziesii, with the Western F I G U R E 3 The mean (±SE) proportion of total bird (n = 1,911) and honey possum (n = 353) visits that occurred at each of 10 flowering stages for 5 inflorescences of Banksia menziesii that were monitored over their entire duration. The inflorescence flowering stage is ranked from 0 to 1, where, for example, 0 indicates no flowers opened and 0.5 represents 50% of the inflorescence has flowered Spinebill and Red Wattlebird relatively uncommon (Ramsey, 1989).

| Detecting pollinator species and their behavior
Our camera traps showed that honeyeater visits to inflorescences occurred throughout the day but peaked mid-to late-morning, lasted on average 34 s during which an average of 8.6 flowers was probed, strongly suggesting they all contribute to effecting pollen transfer, much of it within inflorescences. Pollinator exclusion experiments confirm that these honeyeaters are effective floral visitors for pollen removal and deposition on stigmas (Ramsey, 1988b). Combined, these results highlight that the importance of different honeyeater species for B. menziesii is likely to vary within and between sites depending on the composition of the local honeyeater community.  Our extra video footage was a critical supplement to photos for assessment of still cameras and documenting visitor behavior. While probe rate was accurately recorded by still cameras, time spent on inflorescences, and therefore number of floral probes per visit, were underestimated compared to video. Video data is much more memory intensive, and the capacity of memory cards was a limiting factor to the potential recording life -in some cases, capacity was filled after only 3 days of recording. However, videos are undoubtedly more informative than still images, and will be increasingly utilized for camera trapping as the technology continues to develop and costs decrease (Caravaggi et al., 2017). For our purposes, videos of 60 s duration nearly always captured an entire visit by birds to an inflorescence.

An important limitation of both video and photograph-based
camera trapping is the ability to detect aggression between individuals, a characteristic of nectarivorous honeyeaters that leads to frequent disruption to foraging (e.g., Phillips, Steinmeyer, Menz, Erickson, & Dixon, 2014). Aggression often takes the form of prolonged pursuits, suggesting that camera trapping is unlikely to replace direct observation for understanding the significance of aggression for pollination and pollen dispersal.

| Methodological issues -technical
Despite significant advances in the quality of camera traps (Rovero et al., 2013), the use of camera trapping to accurately quantify behavior and visitation rates of vertebrate pollinators poses significant challenges. The ultimate scenario is one where each camera captures the entire duration of every visit during the day and night.
Clearly, the cameras we employed fell short of this objective, despite identical settings on the same model of camera similarly positioned relative to the inflorescence. Detection at night in particular was highly variable among cameras. Consequently, multiple cameras were necessary for accurate quantification of visitation rates.
Routine cleaning of the infrared detection array window, including the mask, lens and light meter, may improve the effectiveness of cameras, as is the routine use of a moisture-absorbing desiccant system within the camera housing.
False triggers, where the camera was triggered despite no visitor, was a significant issue that varied depending on camera, weather conditions and physical setup. False triggers were typically a result of vegetation movement caused by wind, particularly on sunny days.
Here, as leaves warm, cameras cannot distinguish between warmed leaves moving with the wind and warm-blooded animals moving in the scene. In some cases, as many as 99% of many thousands of photos were generated by false triggers for individual cameras, inefficiently consuming memory and battery power, and necessitating a time-consuming screening process of elimination. Choosing relatively sheltered inflorescences that were up to one meter above the ground, and avoiding the sun shining directly on the face of the camera, appeared to help minimize false triggers. At the other extreme, detection lags or failures impacted effectiveness and the accuracy of estimates of the length of pollinator visits, resulting in gaps of 2-30 (-60) seconds between multiple series of photos of the same visit. Color marking of birds to enable the recognition of individuals in photos would help to address this issue.
Other aspects of the camera trapping method demonstrated excellent performance. The quality of photos was such that all vertebrate species could be identified. With 12 AA rechargeable batteries per camera, camera life is claimed to be up to 40,000 images, so cameras can be left in the field for many weeks or even months, as long as the memory card has sufficient capacity (32GB is possible, our photos were ca 200-700 KB each). The use of an external TA B L E 1 Floral probe rates recorded by motion-triggered videos of bird visitors to three inflorescences (labelled 2, 10 and 11) of Banksia menziesii an on-going resource library of objective visitation data that can be returned to at any time for checking and/or extracting additional information, a clear advantage over field observational data. Rovero et al. (2013) provide a review of multiple cameras (see their Table 2) split into high-end ($550-$1,000), mid range (ca $450), and low end (ca $200). We used a mid-range camera, and it may be that high-end cameras have better effectiveness than these mid-range cameras, and where more accurate quantification of the number of visitors is required, it might be necessary to invest in these more expensive cameras. For our purposes, having multiple cameras on each inflorescence was critical, potentially leading to a higher cost per inflorescence than one superior camera. As such, before larger scale implementation of camera trapping of floral visitors, further testing of other models is required.
One option for collecting visitation data that does not rely on the ability of the camera to detect movement is to utilize cameras with time-lapse capacity, in a similar way to phenocam networks that are monitoring vegetation status and environmental changes (Brown et al., 2016;https://phenocam.org.au

| Future uses of camera traps
The strengths of camera traps as a methodology for pollination biol- However, exceptions occur even in well-supported syndromes (e.g., Quintero, Genzoni, Mann, Nuttman, & Anderson, 2017), and clear associations between pollinator groups and floral traits are not evident in some plant communities or taxonomic groups (Ollerton et al., 2009), particularly those with more generalist species. As such, pollination syndromes should be considered working hypotheses until tested, and camera traps provide an efficient tool to test syndrome predictions.
Following detection of vertebrate pollinators with camera traps, experiments and/or quantification of pollen loads are critical to confirm the effectiveness of the floral visitor (e.g., Ramsey, 1988b).
Camera traps assessing pollination extend their use from a conservation perspective (Caravaggi et al., 2017). Globally, there F I G U R E 5 Diagrammatic representation of recorded vertebrate visits for multiple cameras on each of three Banksia menziesii inflorescences. Figure demonstrates relative effectiveness of each camera, where nonoverlapping bars show failure of one or more cameras to detect a visit recorded by at least one camera. Black indicates a nocturnal honey possum visit is clear evidence of recent declines in bird and mammal pollinators (Potts et al., 2010;Regan et al., 2015). For example, South Western Australia is a Global Biodiversity Hotspot where ca. 15% of 8,379 native vascular plant taxa (Gioia & Hopper, 2017) and ca. 40% of species listed as threatened flora are pollinated by vertebrates (Brown et al., 1997;Keighery, 1982 (Bluthgen, 2010). For species with low visitation rates, camera trapping provides the potential to drastically increase the number of floral visitors observed and avoid the bias toward specialization due to small sample sizes. As such, at least for communities of vertebrates, camera trapping has the potential to make an important contribution to understanding the mechanisms underpinning community structure.
Camera traps document visitation but not visitor movements to and from an inflorescence. However, knowledge of interflower movements within and among plants is critical for an understanding of the consequences of visitor behavior on plant mating (Krauss et al., 2009;Krauss et al., 2017). Observational studies could be optimized from preliminary camera trapping to determine the period of peak pollinator activity. A powerful complement to observational studies is to employ an extensive network of camera traps synchronized for time and date, with banded birds so that individuals, rather than just species, can be identified on camera images. In this way, movement maps can be constructed for individuals based on time, date and location from photos or videos documenting visits.

| CON CLUS IONS
Our study has demonstrated that camera trapping is an exceptional tool for pollination biology studies that not only seek to identify vertebrate visitors, but also to quantify some aspects of behavior such as visitation patterns. In this way, camera traps provide a powerful addition to observation, especially when complemented with individual bird identification through banding, the use of trackers to document movement, and genetic markers for paternity assignment to document realized pollen dispersal and paternal diversity within and among fruits.
However, given the inconsistency between the cameras we used, multiple cameras on individual flowers/inflorescences/ plants are recommended, and these could be complemented by motion-triggered digital video recorders and/or time-lapse photography for further detail on visitation behavior. Increasingly, sophisticated cameras employing time-lapse photography perhaps currently provide the most powerful capacity for accurate quantification of visitation by vertebrates at flowers, although even these bring their own set of challenges that include prolonged data scoring and a reduced ability to quantify behavior.
Pollinator studies using camera traps are in their infancy, and the full potential of this developing technology is yet to be realized. These new tools offer exciting new insights into potentially novel ecological and evolutionary consequences for plants pollinated by vertebrates.

CO N FLI C T O F I NTE R E S T
Authors declare no conflict of interest.

AUTH O R S' CO NTR I B UTI O N S
SLK and DGR conceived and executed the study and collected the data; SLK and CE analyzed the data; SLK led the writing of the manuscript. All authors contributed significantly to the drafts and gave final approval for publication.

DATA ACCE SS I B I LIT Y
Data available from the Dryad digital repository https://doi. org/10.5061/dryad.tj7pp1b