Day and night camera trap videos are effective for identifying individual wild Asian elephants

Study area

We began studying elephant behavior in January 2019, in the Salakpra Wildlife Sanctuary, a protected area in Kanchanaburi, Thailand, in collaboration with the Thai Department of National Parks, Wildlife and Plant Conservation (DNP), the government entity responsible for managing the sanctuary and all Thai national park lands. Salakpra is approximately 868 km² and is located within the 18,000 km² Western Forest Complex. It is a unique protected area in that it is completely closed to tourists and permission is required to enter. The Sanctuary contains areas of mixed deciduous forests (60%), dry dipterocarp forest (30%), and disturbed land (10%) (Chaiyarat, Youngpoy & Prempree, 2015). Data in this study were collected from four different locations as part of a larger elephant behavior project: Kaeng Kaeb (KK) and Khao Seua (KS) are located within the protected area, and Tha Manao (TMN) and Mae Plasoi (MPS) are located along the periphery of the protected area near crop fields (Fig. 1). KK and KS are ranger stations within the protected area of Salakpra where, except for park ranger patrols, human activity is at a minimum. TMN and MPS are villages (specifically, crop fields along the Sanctuary’s outside border) where chances of human-elephant interactions are high. Crop fields mainly consist of corn, pumpkin, sugar cane and cassava, depending on the season.

Permission

This study was approved by the Hunter College Institutional Animal Care and Use Committee (JP-Elephant Behavior 5/21), and permission was granted to collect data in Salakpra Wildlife Sanctuary by the National Research Council of Thailand (Plotnik 1/62) on behalf of the Thai Department of National Parks, Wildlife and Plant Conservation.

Camera trap installation

The videos analyzed in this study were recorded between February 2019 and January 2020. There were a total of 34 Browning Spec Ops Advantage remote-sensing cameras installed throughout the four sites as part of a larger behavioral monitoring project: eight in KK, 11 in KS, six in TMN, and nine in MPS (one camera from MPS was stolen in September, 2019, and was not replaced during this time period). To optimize detection and the recording of social behavior, cameras were installed around watering holes and salt licks in the protected area, and around crop fields and on pathways frequented by elephants in the villages. Camera traps were motion activated and set with a fast trigger (0.4 s) to capture 20-s high resolution video (30 frames/s) from up to ~25 m away. Videos were taken using natural light during the day and built-in infrared light at night. The cameras recorded the time, date and temperature during each recorded clip, which were automatically saved to SD cards our team collected and replaced approximately every 2 weeks.

Identifying individual elephants

In this study, there were 24 physical characteristics (Table S1) chosen to identify individual elephants and adapted from Goswami, Madhusudan & Karanth (2007), de Silva et al. (2013), and Vidya, Prasad & Ghosh (2014). These characteristics were re-defined to our specifications (see the ‘characteristics’ section below). Video clips from all four sites were first scanned and flagged for further investigation using VLC media player (version 3.0.10). For videos to be flagged, elephants had to be marked as visible and identifiable, meaning more than two characteristics were distinguishable (i.e., ear folds, tears, tail length, etc.).

Once an elephant was chosen in a flagged video, another video with the same elephant was found, primarily using videos from the same location (sanctuary or crop fields). However, in some rare instances, elephants were found to have traveled between locations (e.g., KK to KS). These videos were used to match the same characteristics, on a different date to qualify the viewed elephant as a unique individual (Fig. 2; see Montero (2020) where these procedures were originally outlined). We did not characterize an individual and add their traits to the database unless they were observed on more than one date and thus in more than one video. Elephants were also visually confirmed to be the same individual in a video from another date by looking at more specific, non-categorized details such as the shape and exact location of ear tears, shape of tail brushes, and bumps on the skin. In our study, in some instances, only one side of an elephant was visible. Therefore, when locating another instance/video, the whole view or the same side view of the elephant had to be visible in order to confidently label the images as representative of the same individual. This method was adapted from the methods used to identify individual elephants from handheld cameras in previous studies (Goswami, Madhusudan & Karanth, 2007; Goswami et al., 2019). As the purpose of the current study was to assess the efficacy of video camera trap data for identifying individual elephants, we selected footage where elephants were easily observable and thus did not need to exhaustively evaluate all of our video data.

Once an elephant was identified, it was then entered into an AirTable cloud-based database (San Francisco, CA, USA) with a unique code/number, screenshots, and associated characteristics. Characteristics were used to specify each area of an elephant’s body that could be described with different trait state options or specific features of that characteristic. For example, a characteristic such as back shape might have a trait state option such as humped to describe the characteristic. The code N/A was used when these areas of the body were not visible due to video quality or elephant body position. Video clips were matched to each elephant in the database and added continuously to record additional individual characteristics or previously unobserved trait states, as well as to monitor the elephant’s movement patterns between study areas. Male and female adult and sub-adult elephants were identified according to the age categorization outlined below. If an identified female was observed with juveniles or infants in two separate instances, the offspring were characterized and linked to the accompanying female(s) in the database.

Distinguishing individual characteristics by category

We categorized elephants into four age classes (A—adult, B—subadult, C—calves, D—infants; see Table S2 for details). All relative height differences and estimated age ranges were adapted from de Silva, Ranjeewa & Weerakoon (2011). Solitary bulls were usually coded as adults (A), as they typically leave their natal herd once mature (Sukumar, 1989; Fernando & Lande, 2000). However, in some instances, particularly when individuals we observed were among other bull elephants and height comparisons could be made, males were coded as subadult (B). In social groups, we categorized adult females by their enlarged breasts, if they were observable, or the presence of calves (de Silva, Ranjeewa & Weerakoon, 2011). Although the present study utilized age as a characteristic, the age classes mentioned are only estimates based on the trait state definitions (Table S2); we were not able to determine the exact age of individuals.

To determine the body conditions of each individual, we assessed the pelvic, shoulder, and back bones as elephants moved in a video. Body condition definitions were adapted from Fernando et al. (2009) and simplified to three categories: Zero for the underweight condition if ribs were visible, one for the normal condition if pelvic and shoulder bones were visible but ribs were not, and two for the overweight condition where pelvic and shoulder bones were not prominent (Table S3). The prominence of the backbone was also used as an indicator of body condition (Wemmer et al., 2006).

We categorized whether the individual had either tusks or tushes (incisors that are much smaller and thinner than tusks) (Kurt, Hartl & Tiedemann, 1995). In Asian elephants, only males have tusks—although not all do—while both males and females can have tushes (i.e., short tusk-like protrusions from the top of the mouth), but again, not all do (Sukumar, 1989; Kurt, Hartl & Tiedemann, 1995; Chelliah & Sukumar, 2013). When tusks were present, tusk symmetry, arrangement, and angle were recorded accordingly (Table S4). We categorized tusk symmetry based on whether the tusk length was symmetrical and tusk arrangement based on the tusk growth direction of both tusks compared to each other. We also categorized tusk angle based on the direction of the tusks in reference to a horizontal plane. We used side views of the elephant to best determine tusk angle, and the position of the trunk to help guide the decision (Fig. S1).

We described characteristics of the elephants’ ears, focusing on the left and right ear separately. Also, we considered top folds and side folds (labeled as primary and secondary fold, respectively, in de Silva et al. (2013)) as separate characteristics. We described top ear folds based on the degree to which the top ear was folded on both sides, and side folds by the way each side fold was positioned. We used the angular shape of the bottom of the ear (or ear lobes) to describe them (Fig. 3). We described other characteristics of the ears (ear tears, holes and depigmentation) when possible (Table 1). When ear tears and holes were present, we categorized their locations starting from the top to the bottom of the ear; if there were multiple tears or holes in one ear, we recorded the location with the most tears or holes. The location of other tears or holes along the ear were added as a note in the database.

The back shape of each individual was organized into three categories (Fig. S2 and Table S5). We did not observe a ‘concave back’ in this study, but because it was described in the population studied by Vidya, Prasad & Ghosh (2014) in India, it was included as a possible category.

There were two different tail characteristics used to identify the elephants: tail length and brush type (Fig. S3 and Table S6). We categorized tail length based on the distance from the rump to the tip of the tail, not including the ‘tail brush’ or hair. We described the tail-brush type based on its length and location (i.e., the anterior side closest to the body, the posterior side farthest from the body, or both sides of the tail).

Finally, we categorized depigmentation on parts of the elephants’ bodies other than the ears if and when it occurred (Table S7). Figure S4 shows an example of an elephant with depigmentation on various parts of his body.

Interrater reliability

We assessed reliability of the categorization of elephant characteristics between the first author (SM-DLT) and another trained coder in a subset of thirty video observations, each of a different individual elephant (19 recorded during the daytime, 11 at night). Cohen’s kappas for each of 17 characteristics varied from poor to excellent between the two coders, although most characteristics had moderate or better agreement (Table S8). Although some kappas were poor, we did not exclude those characteristics in our initial analyses because ultimately the categorization used was based on multiple videos of an individual rather than the single video used in our reliability assessment. In addition, the verification of an individual’s identity relied on multiple characteristics (never just one) as well as a separate visual confirmation using other physical traits described previously. For the purposes of illustrating these points, we have included relevant analyses that include and exclude the characteristics for which there were kappas that represented poor agreement—back shape and body depigmentation—in the results below.

We further assessed the reliability of more than one coder correctly identifying the same individual in multiple videos—which aims to demonstrate the effectiveness of the ID protocol—by testing a naïve coder. This coder reviewed the written protocol but was not trained on identifying elephants, and was tasked with identifying 20.8% of the 72 adult elephants (15 individuals: eight males and seven females) from a video dataset. The coder was provided with the identification protocol including definitions of characteristics and trait states as well as each of the 15 elephants’ characteristic profiles and a reference video for each of the 15 individuals. This test mimicked the procedure that the first author followed when originally determining that a video did or did not match an elephant already identified. The coder was instructed to use the characteristic profiles to identify the 15 elephants in the video set, which included two videos of each of the 15 elephants and six videos that included only other elephants for which no other information or reference videos were provided. However, they were not told how many videos of each of the 15 elephants were in the set nor how many videos contained elephants other than those characterized. They did know that some of the videos in the set were elephants other than the 15 and were asked to identify which videos contained these “other” elephants. The first author and the additional coder agreed on the identity of the elephant, or that the video contained an elephant of undetermined identity, in 92% of the videos.

Statistical analysis

We analyzed data in Microsoft Excel (v. 2016). Using Goswami et al. (2012)’s misidentification calculation and our aforementioned system for categorizing morphological traits, we determined the likelihood of the characterization process resulting in the misidentification of two different individual Asian elephants as the same elephant by calculating the maximum probability squared (p_max²). It is important to note that for the current study, we used this calculation to determine the probability of misidentification between easily visible elephants in our subset rather than between elephants identified in all videos (Goswami et al., 2012). Also, characteristics determined from multiple video clips of each individual were used for this calculation. In our study, to determine the maximum probability that the same traits exist and can be categorized in two different elephants using the camera trap footage, we first calculated the frequency of each trait state option per characteristic and used the most frequent in calculations. For example, the most common trait state for left ear lobe shape is a v-acute ear lobe shape which was observed in 63.9% of all adult elephant sightings (Table 2). Once the most common trait frequencies were calculated, they were ranked from the most to least commonly occurring morphological characteristics with those trait states. If there was more than one characteristic and trait state option that occurred the same number of times in the populations, the first occurring characteristic as listed in the identification protocol was put first into the ranking followed by the next on the list. The characteristic list order in the protocol was arranged for capturing information from the front of an elephant’s body to the back. However, characteristics that were seen from the whole elephant like sex and body condition were placed at the beginning of this order.

Exploratory statistical tests were used to determine whether characteristics were independent from each other. Independence in this case means that the traits of one characteristic cannot be predicted from the traits of another characteristic. Chi-square and Fisher exact tests were used to calculate whether the number of individuals with each combination of traits corresponded to the assumption of independence between those traits. As was the case in Goswami et al. (2012), many pairs of characteristics were not independent from one another. If traits are independent, then the probability of a combination of traits would be equal to the product of their individual probabilities. However, because of non-independence, a conditional probability calculation is more appropriate as it does not assume independence. Therefore, to estimate the probability that an individual possessed the most commonly occurring combination of traits (p_max), conditional probabilities were calculated by moving successively down the trait frequency ranking.

When computing p_max, we first calculated the probability of the most frequent trait state for presence of tusks/tushes. Next, we looked at the probability of back shape’s most frequent trait state occurring, when presence of tusks/tushes’ most frequent trait state occurred. Moving down the ranking, the next characteristic (L ear hole) and its most frequent trait state option contributed to the calculation for the probability of the L ear hole’s most frequent trait state occurring, given the presence of tusk/tushes most frequent trait and back shape’s most frequent trait. This process continued until the number of elephants with the combination of characteristics reached one (Table 2). The probability values were then squared to obtain the value for the probability of any two individuals showing the exact combination of morphological features (p_max²) (Goswami et al., 2012). p_max² was calculated separately for all adult elephants identified, adult males identified, and adult females identified.

Implications

This study supports the idea that the use of remote-sensing cameras to identify individual elephants in long-term studies of elephant behavior and ecology can be effective (Fernando et al., 2009, 2011; Goswami et al., 2012; de Silva et al., 2013; Vidya, Prasad & Ghosh, 2014) and provides a guide for identifying them using day and night videos. Given that the footage from these cameras can be as useful as photographs from research vehicles, our results hopefully can encourage researchers in other Asian elephant ranges to employ camera trap technology to systematically identify individuals in their local populations, especially in places where observing wildlife directly is challenging. As more scientists conduct research at the individual-, rather than just the population-level, this technology can better help monitor elephant movement and activity as well as characterize variation in behavior patterns between elephants (e.g., Rees, 2009; Horback et al., 2012; Sitompul, Griffin & Fuller, 2013; Horback et al., 2014).

The database of individual elephants created for the present study will be compiled into a guide and provided to Salakpra Wildlife Sanctuary rangers and local farmers as a reference for identifying resident elephants. Since frequent crop raiding has been observed around the sanctuary (van de Water & Matteson, 2018), individual identification may help provide insight into the behavior of particular elephants as it relates to their interactions with humans in the area (e.g., Cook, Henley & Parrini, 2015; Goswami et al., 2015; Ranjeewa et al., 2015; Plotnik & Jacobson, 2022). Identifying individuals from camera trap footage recorded at night may be particularly important in this context since many elephants are active in crop areas at night (e.g., Ranjeewa et al., 2015; Naha et al., 2020). A capacity for identifying elephants that frequently forage on crops could also aid in targeting HEC mitigation strategies at the individual level, a potentially more effective strategy that takes elephant behavior and personality into account (Mumby & Plotnik, 2018; Plotnik & Jacobson, 2022). More targeted strategies may help farmers manage their time as they direct their efforts towards specific individuals that they can identify within their own crop fields.

Identifying individuals is also crucially important for understanding individual variation in elephant behavior more generally. Remarkably, we know very little about such variation in wild Asian elephant behavior and how elephants adapt to rapid, human-generated environmental change (Plotnik & Jacobson, 2022). The current study helps form a foundation for future research in this area. Our own work aims to use the individual identification of wild Asian elephants to assess differences in personality and cognition, not only as a means to help in their conservation, but also as a method for understanding how flexibility in behavior facilitates adaptations to anthropogenic change (Mumby & Plotnik, 2018; Plotnik & Jacobson, 2022).

Limitations of camera traps for individual identification

The position and angle of the stationary camera traps sometimes limited our ability to collect comprehensive morphological data on the elephants. While a stationary camera provides an opportunity to capture some characteristics of an elephant in its field of view, what is captured is dependent on the elephant’s distance from the camera and movement across its view. Some videos only captured ears, backs, and tails, while other videos did not capture backs, tail length or ear top folds. Depending on the elephant’s approach and activity in front of the camera, sometimes only one side of the elephant was recorded. Camera traps deployed in the field were typically put up in a high place and were stationary for a long period of time. The only way to change the view would be to manually move the direction of the camera, and this was usually done infrequently due to their installation in remote areas. Overall, this limitation on the camera’s mobility increased the frequency of data points where the elephant could not be identified due to a lack of observable characteristics. Nonetheless, even though the position and immobility of camera traps at any given time limited the number of views of an elephant during a single observation, elephants were not included in our analysis unless we could confirm that they were the same individual in different videos based on a majority of the same traits. We suggest that future studies focusing on individual identification would benefit from installing multiple cameras together at different angles to ensure that both sides of the elephant are visible in an observation.

Collecting identification data from night videos was another challenge. During the night, when the infrared light was illuminated, characteristics like ear folds would sometimes blend in with the color of the ear, obscuring the shape and folds, making it difficult to identify the trait state. Even with some traits obscured, there were typically others visible that allowed for identification of the elephant. In the future, we would like to compare the utility of using either day or night video to identify specific traits, as well as compare how often elephants are not able to be identified depending on the lighting. This could also inform the differential installation of future cameras based on elephant movement patterns at different times of day.

While there are limitations for using camera traps to identify individual elephants, the installation of remote cameras allows for the capture of multiple screenshots from video at a much closer proximity than is usually possible when humans are present. Thus, the lack of a human presence during data collection, which could negatively impact the elephants’ behavior, may help offset the limitations camera trapping poses to individual identification. Videos also allow for determination of traits based on nuanced body movements. For example, analyzing frame-by-frame ear flapping can help identify side folds, while also making tears and holes more evident than they might be in a photographic snapshot. The use of video, when possible, could provide greater clarity than photographs during individual identification, especially when distinguishing between elephants with marked similarities in phenotype.

A limitation of identifying elephants using the characteristics discussed herein is that not all of them were temporally static. Our identification protocol included some characteristics that are temporally variable, although over different time scales (Goswami et al., 2012; Vidya, Prasad & Ghosh, 2014). For example, an individual’s body condition may change across seasons, degree of depigmentation may change across years, and new ear tears or tusk breaks could occur anytime. Researchers using this identification methodology must be aware of these potential changes and confirm the final individual identity using further analyses of the images. To maintain accurate identification, we suggest updating individual characterizations in a database as often as possible. Since the data included in the study only covered one year and no single characteristic determined a new individual, it is unlikely that any temporal changes affected the reliability of our dataset.