Background

Abundance, or population size, is a primary parameter in wildlife conservation and management used to prioritize conservation actions and assess the conservation effectiveness (Seber 1973; McCarthy et al. 2008; Jenks et al. 2011; Jones 2011; Rovero et al. 2014). Camera-trapping, in combination with the traditional capture-recapture or spatially explicit capture-recapture methods, has been widely applied to estimate the absolute abundance of individually identifiable species (Seber 1973; Nichols 1992; Karanth et al. 2004; Royle and Young 2008; Russell et al. 2012; Karki et al. 2013; Avgan et al. 2014; Qi et al. 2015; Linden et al. 2017). The camera-trapping is efficient for surveying wide-ranging, cryptic and elusive animals in inhospitable environment such as tropical rainforest (Tobler et al. 2008, Rovero et al. 2014). Unlike many survey techniques, it is non-invasive, thereby reducing observer bias and influence on animal’s behavior and requires less manpower (Gompper et al. 2006, Alonso et al. 2015, Caravaggi et al. 2017). Natural body markings, such as stripes, spots, and blotches, from camera trap photos are some indicators to identify individuals (Sharma et al. 2013; Boron et al. 2016; Mettouris et al. 2016). Therefore, the capture-recapture method using camera trap images for estimation of population size has primarily targeted big cats with distinctive coat patterns such as tigers (Panthera tigris), leopards (Panthera pardus), jaguars (Panthera onca), and snow leopards (Panthera uncia) (Karanth and Nichols 1998; Henschel and Ray 2003; Cullen et al. 2005; Soisalo and Cavalcanti 2006; Maffei and Noss 2008; McCarthy et al. 2008; Lynam et al. 2009). However, natural body markings include not only coat patterns but also whisker spot patterns and permanent scars (Creel and Creel 1997; Anderson et al. 2007; Osterrieder et al. 2015).

As monitoring technology using camera traps improved, individual identification of small carnivores including leopard cats (Prionailurus bengalensis bengalensis) in South Asia and American martens (Martes americana) in the USA, became also feasible (Bashir et al. 2013; Sirén et al. 2016). However, compared to leopard cats in South or Southeast Asia, adult Amur leopard cats (P. b. euptilurus) in East Asia have dim fur patterns on flanks, which makes photo-identification extremely difficult. The Amur leopard cat is a Class II endangered species in South Korea (Ministry of Environment 2011) while leopard cats are generally least concerned in most other countries (Ross et al. 2015). However, very little is known about its population and behavior (Park et al. 2017), and capture-recapture method has not been applied yet to estimate the population. Therefore, we aimed to test whether the Amur leopard cat has natural body markings to develop a reliable noninvasive identification system for use in studies of behavior and population parameter estimates.

Methods

In order to identify reliable natural markings, we analyzed Amur leopard cat images from two different sources—facial images of seven captive individuals in two wildlife rescue centers, Gyungnam Wildlife Rescue Center and Korean Avian Conservation Society in Cheolwon, in South Korea, taken with high-resolution (18 Mpxl) Digital Single-Lense Relfex cameras (EOS 7D, Canon Inc, DSC-HX400V, Sony Inc.), and images of 11 wild individuals taken with low-resolution (3.1 Mpxl) camera traps (Reconyx HC550). To acquire wild leopard cat images, we deployed 16 camera traps in two different riparian parks, Hwapocheon in Gimhae and Eulsukdo in Busan between May and November 2015 (8 camera traps per riparian park, Fig. 1). To increase detection rate, we placed a lure stick scented with Valerian oil in front of the camera traps (Burki et al. 2010; Anile et al. 2012; Steyer et al. 2013) (Fig. 2). We selected a total of 11 high-resolution images of 7 captive leopard cats (4 females and 3 males) and 15 low-resolution images of 11 wild leopard cats (sex undetermined) that contained less noise and clutter among the 63 low-resolution images that we obtained. We then further divided those images into three different groups based on featured body parts and image resolution: high-resolution facial images (11 images of 7 leopard cats), low-resolution facial images (8 images of 7 individuals), and low-resolution ventral images (7 images of 4 individuals). One low-resolution image was used for matching both facial and ventral parts. We did not have enough high-resolution ventral images to compare individuals, and thus excluded the group in our analysis.

Fig. 1
figure 1

Camera trap locations within sampling blocks of two study sites (Hwapocheon and Eulsukdo) in the Nakdong river basin

Full size image
Fig. 2
figure 2

Survey design of a baited camera trap to effectively capture leopard cat images

Full size image

To match individuals, we used HotSpotter, an algorithm for identifying individual animals against a labeled database based on extracting and matching key points and descriptors in the region of interests (Crall et al. 2013). HotSpotter returns the matching result with a score and rank when a reference image is compared with multiple images and correctly matched individuals for 95% of the queries in the previous study which tested the matching individuals of zebras, jaguars, giraffes, and lionfish (Crall et al. 2013). The score, based on the Local Naive Bayes Nearest Neighbor methods, presents the similarity among descriptors of the images and is as low as 0 for non-distinct images (Crall et al. 2013). We first tested the validity of this algorithm by matching a sample facial image (LM.1) of a reference cat with facial images of different species that included a tiger and a leopard. We then matched the sample facial images of a reference leopard cat (LM.1) with a different image (LM.1-1) of that same leopard cat. Further, we matched different leopard cat individuals by comparing coat patterns of their body parts (face and venter) from the selected leopard cat images to each other (Fig. 3). We also tested different image combinations such as containing 0, 1, and 2 images of the same individual to check the changes in scores and ranks in various situations.

Fig. 3
figure 3

Individual identification of leopard cats through comparing fur patterns of a a face from high-resolution images and b a venter from low-resolution camera trap images using HotSpotter

Full size image

Results and discussion

We confirmed that matching scores of a leopard cat facial image with other species was as low as 0 while the score of matching images of the same leopard cat was 4,236,180 in one vs. one matching. In one vs. many matching, HotSpotter correctly top-ranked (rank 0) an image of the same individual leopard cat with the reference leopard cat image 100% by matching facial and ventral part although matching scores varied from 10,838 to 228,151 (n = 7, mean = 56,743, median = 27,181, SD = 77,404). When there was no matching image with a reference image, rank 0 scores were smaller (n = 11, score range = 2,571 ~ 28,136, mean = 12,011, median = 8,870, SD = 8,870) than the scores of having a matched individual (Mann-Whitney U test, Z = 2.75, p < 0.01) and often smaller than 10,000 (n = 8). Having two images, which matched with a reference image (n = 1), those two images were top-ranked, rank 0 and rank 1 respectively. (Note: see the Additional file 1 for the complete matching scores).

Our results indicate that facial and ventral fur patterns of the Amur leopard cat are good matching points to be used reliably to identify individuals although we admit that our small sample size may provide some uncertainty. Comparing the difference in scores between correct matching and incorrect matching, 10,000 in rank 0 was a useful reference score to determine the correct matching in HotSpotter. That is, a matching score around 10,000 needs additional attention before categorizing images. Furthermore, differences in images due to being taken at different angles did not affect the algorithm to match individuals correctly. However, the quality of the image affects the matching success (Gowans and Whitehead 2001; Kelly 2001; Halloran et al. 2015). In computer-based matching, low resolution refers to a small number of pixels available to compare images. Accordingly, low-resolution images, especially when the matching area is small, tended to show low matching scores in our study. Thus, to increase low-resolution image accuracy, we used an anti-weed sheet to reduce image blur and background noise which negatively affect image quality (Halloran et al. 2015). Moreover, we used a bait, Valerian scent stick, to maximize the chance to attract more leopard cats, extend staying time in front of a camera trap, and acquire a clear image of face and venter of leopard cats. These efforts produced 63 camera trap images of which only 15 images were suitable for matching. Even though the numbers of suitable images were low, our study suggests that using ventral and facial fur patterns is a promising approach for individual identification of Amur leopard cats. However, further testing with increased sample size is beneficial to improve the reliability of the study. Also, research on permanency of facial and ventral fur patterns is necessary to examine the feasibility of using the fur patterns as a lasting natural marking for individual identification because our study proceeded for 6 months. We anticipate that the study result will be useful to researchers interested in studying behavior or population parameter estimates of Amur leopard cats based on capture-recapture models.