Development of a camera trap for perching dragonflies: a new tool for freshwater environmental assessment

Sympetrum dragonflies

In Japan, dominant Sympetrum species, such as S. frequens, are perchers and are considered environmental indicators for agricultural landscapes consisting of paddy fields (Sprague, 2003; Inoue & Tani, 2010; Tanaka, 2016). Mature adult Sympetrum dragonflies have conspicuous red bodies in the autumn and are familiar to people in Japan (Inoue & Tani, 2010). There are concerns about the rapid decrease of these species in some regions in Japan due to pesticide use (Jinguji & Ueda, 2015). In addition, land abandonment can reduce aquatic ecosystems in Japan (Ikegami, Nishihiro & Washitani, 2011), which can impact these dragonflies.

Considering that the intensification and/or abandonment of paddy ecosystems are common in rice-producing regions, continuous broad-scale monitoring of these dragonflies is needed. However, monitoring methods for these species are fairly limited. Insects are usually observed by experts or trained citizens (Inoue & Tani, 2010; Tanaka, 2016). Therefore, examining the applicability of our camera trap to Sympetrum species has implications for conservation ecology. Considering that the genus is distributed worldwide, except in Oceania and south-central Africa (Sugimura et al., 1999), Sympetrum dragonflies were an ideal candidate for assessing our camera trap.

Detection parameters

In the field experiment, parameters for the detection algorithm were provisionally set based on empirical knowledge. The interval to reset the average ratio of signal values X was set to 300 s, and the threshold value Z for perching duration was set to 10 s. The threshold value for Y, the ratio of the instantaneous relative signal value to the average relative signal value, was set 1.05. The processor commanded the digital camera to reboot when the perching duration Td reached 7 s because the digital camera slept. Note that the microprocessor could not monitor the status of the camera and the reboot signal was similar to that for taking a photo. In addition, the microprocessor was set to periodically wake up after sleeping for 1 s to save power. Considering that three 1.2 V, 1,900 mAh batteries were used per detector and lasted at least 66 days during the preliminary study described below, power consumption was estimated to be less than 5 mW. The algorithm (Code S1) was described and compiled using MPLAB IDE v. 8.60 (Microchip Technology Inc.) and microprocessor programming was performed using the MPLAB ICD3 In-Circuit Debugger (Microchip Technology Inc.).

Field experiment

Study sites and camera trap setting. In a preliminarily experiment, a camera trap was set in the yard of the Fukushima branch, National Institute for Environmental Studies (NIES), Miharu town, Fukushima prefecture, Japan (37.434°N, 140.519°E) from October 11 to December 16, 2016 (Fig. 2A). Note that dragonflies are very sparse in the region after November (A Yoshioka, pers. obs., 2016).

A short-term field experiment was also conducted to examine the performance of the camera trap. From 8:05 on October 21 to 17:30 on October 22, 2016, three camera traps were set in the yard of the NIES, Tsukuba city, Ibaraki prefecture, Japan (36.049°N, 140.117°E; Figs. 4A–4C). The average temperatures on the two days were 14.5 °C and 14.9 °C, respectively, and precipitation was not recorded (Japan Meteorological Agency http://www.data.jma.go.jp/obd/stats/etrn/index.php?prec_no=44&block_no=&year=&month=&day=&view, accessed February 7, 2018). This season corresponds to the late stage of adult dragonflies (Inoue & Tani, 2010). Two camera traps were set in areas where the density of the dragonflies was relatively high, one in a field and one a pond. The field site consisted of small experimental paddy fields neighboring a greenhouse. The paddy fields were apparently not managed in the season and vegetation within the paddies was relatively short. Sympetrum dragonflies rarely perched on the vegetation within paddies. Instead, they perched on artificial poles and net-fences, or trees surrounding the site. The pond site consisted of a small pond and wet vegetation, in addition to paved ground. The Sympetrum dragonflies were frequently observed to perch on cattails (Typha spp.) and tall goldenrods (Solidago canadensis). The remaining trap was set on the rooftop of the main building of the institute, where no vegetation exists. There were some artificial poles, but dragonflies rarely perched at the site, which is relatively distant from aquatic habitats (thus, dragonflies were expected to be rarely detected). Each site was about 250 m² considering the homogeneity of land cover and accessibility for the dragonfly census described below. In addition, these sites were more than 150 m apart and intercepted by multiple buildings and trees.

Evaluation of auto-detected images. Images obtained from the three camera traps were used to count true positives and false positives (i.e., photos with and without Sympetrum dragonflies). The number of false positives indicated the extent of type I error. Each image was visually evaluated. When there were multiple photos within a few seconds, the first photo was removed from the analysis (see also the “not analyzed to avoid double count” column in Data S1). These images resulted from commands for rebooting the camera when it was already active and led to overestimates of true positives.

Validation by time-lapse camera monitoring. To check perching dragonflies undetected by the camera traps (i.e., false negative or type II errors) and to evaluate the detection sensitivity, an additional digital camera was placed near each camera trap. At the field and pond sites, time-lapse cameras (TLC 200) captured movies at 10-s intervals. At the rooftop site where dragonflies were expected to be rare, a video camera captured movies at 30 fps, instead of a time-lapse camera. After the experiment, the movies were thoroughly checked and the presence or absence of Sympetrum dragonflies in each frame was recorded. Movies were initially used to screen for files that include perching dragonflies. Then, movie files including perching dragonflies were split to jpeg image frames using FFmpeg ver. 2015-04-02 (https://ffmpeg.org/) to evaluate frame numbers. No dragonflies perched on the camera trap at the rooftop site during the experiment. To calculate sensitivity, defining a false-negative case based on movie files obtained from the cameras for validation was required. Because perching dragonflies often showed instantaneous takeoff and landing, a “perching event” was defined as an array of movie frames including fewer than two continuous frames without a dragonfly (Fig. 5A). That is, if a dragonfly left the camera trap for more than two frames (20 s), the perching event was closed. Note that 20 s corresponds to the bout criterion interval (Slater, 1974) and the number of “perching events” did not change substantially by increasing the intervals to longer than 20 s. Based on the timestamps for the image frames, the start and end times of each perching event were obtained. Then, the detection of each perching event by camera traps was verified by matching timestamps of autodetected images. If at least one image capturing a Sympetrum dragonfly was obtained by the camera trap during a perching event period, the perching event was recorded as a true positive. If no image of a dragonfly was detected during an event, this was recorded as a false negative. Then, sensitivity was estimated for each site as (true positives ÷ (true positives + false negatives)). A caveat is that an individual could contribute to multiple perching events, and the replacement of perching individuals within 20 s is possible. Thus, sensitivity for a camera trap may reflect the activity of dragonflies, but not necessarily the population density. In addition, Fisher’s exact test was used to compare the frequencies of true positives and false negatives among sites using the fisher.test function implemented in R version 4.0.0 (R core Team, 2020).

Comparison with the dragonfly census. To examine whether detection by the camera traps reflects the dragonfly population density, a comparison among survey methods is important. In parallel with the short-term experiment, census data for Sympetrum dragonflies were obtained at each site. As a standardized census method, a timed survey (Balzan, 2012; Harabis & Dolny, 2015) was adopted because a transect survey was not suitable for the non-linear configurations of sites. At each site, an investigator counted and recorded the number of Sympetrum dragonflies for 10 min three times a day (six observations in total). Previous studies have reported that marked individuals of some perching dragonfly species could maintain reproductive territories for about 1.5–2.5 h (Koenig & Albano, 1985; Moore & Martin, 2016). Thus, three 3-hour time periods were established (9:00–12:00, 12:00–15:00, and 15:00–18:00) and all of the sites were visited for the census for each time period (note that dragonflies were rarely observed from 8:00 to 9:00). For each time period, the first site visited at first was randomly determined and the remaining two sites were surveyed within 1 h from the start of the first survey (see Data S3 for details). Therefore, the effect of sites was not confounded with variation among time periods. During the timed survey, the investigator slowly and randomly walked while searching for dragonflies within the site, because the dragonfly density was relatively low and counting individuals at fixed points was not likely to be effective. To minimize double counting, perching individuals at the same point were counted only once for each 10-min survey, unless the investigator observed individuals leaving the point. For each site, the numbers of Sympetrum dragonflies observed per 10 min period were summed to obtain the “total frequency of dragonflies observed”.

To obtain robust inferences on the relationship between camera detection and census observations given the small sample size, two statistical analyses were used (see Code S2). Note that the number of images of Sympetrum dragonflies was used as an index for camera detection rather than the perching events owing to the low values at the field and rooftop sites (one and zero, respectively). First, differences in Sympetrum dragonfly detection by the camera trap among sites and differences in the total frequency of dragonflies observed among sites were evaluated. If camera detection indicates the dragonfly population density, similar to census observation, statistical inferences should be similar for the two approaches. To test the differences among the sites, a goodness-of-fit test was applied to each pair of sites under the null hypothesis that the probability of dragonfly occurrence is equal between sites (i.e., 0.5). The probabilities of obtaining data (or a more extreme difference than the observed value) under the null hypothesis were computed by Monte Carlo simulations (Hope, 1968) with 1,000,000 replicates using the chisq.test function implemented in R version 4.0.0 (R Core Team, 2020). The random seed was set to 0. The significance probabilities were adjusted by Holm’s method (Holm, 1979) to account for multiple comparisons. Next, a direct approach was used to examine relationships between camera detection and census observation. Because the site number was small, the sample unit was set to dragonfly detection over a 3-hour time period per day per site and census observation corresponding to the time period (i.e., n = 18). This assumption was based on the previous studies that perching individuals are usually replaced within 3 h (Koenig & Albano, 1985; Moore & Martin, 2016). Then, a generalized linear mixed model (GLMM) with a quasi-Poisson error distribution was used to evaluate whether the number of dragonflies detected per 3-hour period (the response variable) reflected the number of dragonflies observed per 10 min (the explanatory variable). In the model, site, day, and time period were included as random variables. The GLMM was implemented using the glmmTMB function in the glmmTMB package for R version 4.0.0 (Brooks et al., 2017; R Core Team, 2020). Similarly, the GLMM was calculated using the number of dragonflies detected as the explanatory variable and the number of dragonflies observed as the response variable. The caveat is that the adherence of an individual Sympetrum dragonfly (mainly S. frequens) to a territory is unclear. Nevertheless, the territorial behavior of S. frequens is not remarkable (Sugimura et al., 1999) and their adherence to a specific perch may be relatively weak. In this study, the only individual that perched on the camera trap in the field site held its territory for 25 min (from 13:30 to 13:55 on October 22) at most (see Data S2). At the pond site, an individual was considered to hold its territory for 2.4 h (from 12:08 to 14:32 on 21th October) at most, if we assumed that the replacement of individuals was indicated by leaving the perch for more than 180 frames (30 min), which is an adequate period of time for copulation and tandem oviposition (Sugimura et al., 1999; Ishizawa, 2012). Thus, a 3-hour time period should be sufficient for the replacement of perching individuals.