Using citizen science in the photo-identification of adult individuals of an amphibian based on two facial skin features

Introduction

Animal biometrics is an emerging field that involves the identification of species or individuals based on their external phenotypic characteristics, including natural markings or color patterns (Kühl & Burghardt, 2013). It has been used as an effective technique for data collection in ecological procedures such as capture-mark-recapture (CMR), as photographic images of an individual’s unique markings can be cross-matched within a photo-database for detection of recapture events (Williams, Nichols & Conroy, 2002; Pebsworth & Lafleur, 2014). This process has been particularly useful for monitoring species that cannot be easily captured or artificially tagged for identification purposes (Frisch & Hobbs, 2007; Arandjelović & Zisserman, 2011; Hughes & Burghardt, 2015), and has been applied to a diverse number of taxa including mammals (Karanth & Nichols, 1998), large fish (Arandjelović & Zisserman, 2011; Hughes & Burghardt, 2015), crustaceans (Frisch & Hobbs, 2007), and herpetofauna (Gardiner et al., 2014). With the increased affordability and use of smartphone devices that are equipped with cameras, as well as the advent of camera trapping technology, individuals can now be photographed under field conditions and differentiated with very little cost, logistics or expertise required (Wagner et al., 2008; Haddock, Kim & Mukai, 2013; Pebsworth & Lafleur, 2014). Photo-identification is thus becoming an increasingly important ecological tool that is also providing greater opportunities for the use of citizen science in animal monitoring programs (Dickinson, Benjamin & David, 2010).

Nevertheless, some drawbacks have limited the application of photo-based CMR including difficulties in manually processing large image datasets ‘by eye’, particularly those in which markings are only subtly different between individuals, which becomes a time expensive process vulnerable to misidentifications (Katona & Beard, 1990; Bolger et al., 2012; Crunchant et al., 2017). Such obstacles can be overcome using computer-vision techniques that use pattern recognition algorithms, such as ‘hand-crafted’ feature descriptors or deep metric learning, which automatically detect, extract and compare feature information from images uploaded to a photo-database (Van Tienhoven et al., 2007; Takeki et al., 2016; Treilibs et al., 2016; Crunchant et al., 2017). These techniques have been shown to have high accuracy in animal identification and have led to significant labour savings (Morrison et al., 2011). Despite these benefits, most computer assisted systems are only partially automated and still require some degree of manual image processing (Burghardt, 2008). Instead of returning a definitive ‘match’ or ‘no match’ decision, a similarity score is calculated between each image pair, with the strongest matching candidates to the query subsequently needing to be visually inspected in order for a ‘true’ match to be confirmed. It is thus crucial that the feasibility of manual image matching is validated for species prior to the application of photo-based CMR, even for processes that are to become computer-assisted. This challenge of manual image matching, particularly for large databases, may be overcome through the use of citizen science (e.g., Willi et al., 2019), as the speed of matching can be increased by spreading the workload across a large group of people that can effectively be recruited online from any location.

Among amphibians, traditional techniques for identifying individuals include the placement of an artificial visual marker, the removal of toe pads or the insertion of dyes and microchip transponders (Turner, 1960; Brown, 1997; Simoncelli et al., 2005; Bainbridge et al., 2015). These are invasive processes that may influence animal survival, require expertise and are relatively expensive, thereby limiting their widespread use among citizen scientists (Reisser et al., 2008; Sacchi et al., 2010). In contrast, photo-identification is being increasingly used to differentiate amphibians at both the species and individual level (Bradfield, 2004; Church et al., 2007; Gamble, Ravela & Mcgarigal, 2008; Bendik et al., 2013; Sannolo et al., 2016; Konovalov, Jahangard & Schwarzkopf, 2018), and has been shown to have the capacity to outperform traditional marking techniques (Bendik et al., 2013). Nevertheless, the ability for skin features to be used for identification purposes requires investigation to determine if it is a viable alternative that can be performed by volunteers with little prior expertise.

In this study, we examined the feasibility of using citizen science in the visual identification of adult individuals of our model species, the sandpaper frog (Lechriodus fletcheri) based on two facial skin features (banding patterns and tubercles). We asked anonymous participants from the general public, recruited through social media, to visually examine small image datasets and correctly match images of individuals taken at different points in time under natural field conditions from a smartphone device. The main objectives of the study were to determine (i) the potential ease of obtaining sufficiently clear images of individuals from smartphones devices and (ii) the scalability of photo-matching as a technique that has the potential to be used in citizen scientist projects to assist in mark-recapture modelling of anuran populations. Additional objectives were to assess (i) the level of skin feature variability among adult L. fletcheri individuals, and (ii) the stability of skin features in adults over time, as well as to (iii) evaluate the efficacy of using our two facial skin features for accurate photo-identification of an anuran species.

Material and Methods

Study species

Lechriodus fletcheri is a medium sized frog (4–5 cm) found in montane temperate forests along the east coast of Australia (Clulow & Swan, 2018). This species has a prolonged breeding season in the austral spring-summer (September–March) with adults congregating at ephemeral pools during periods of heavy rain to reproduce. Both sexes possess unique facial skin features that have the potential to be used for photo-identification (Fig. 1). A region of black banding that runs horizontally above the tympanum from the corner of the eye to the front leg is a strong candidate marker due to its relatively large size and high level of outline variability between individuals. The skin in this region also has many small epithelial projections (tubercles) that form a unique background pattern which might be an effective secondary feature for photo-analysis. An area of facial skin that included both features was selected as our region of interest (ROI) for visual analysis. All other skin features, particularly dorsal, ventral and leg patterns were not found to be distinct and sufficiently clear for photo-identification purposes.

Results

A total of 790 photographs were taken over the course of the three breeding seasons. Our database included 606 unique individuals, 15% of which were recaptured and photographed on more than one occasion. The number of days separating capture events varied from a few days to more than a year. Facial bandings used for photo identification were found to be highly polymorphic between individuals of both sexes, along with background arrays of tubercles (Fig. 2). Both features were also found among all individuals captured and clearly identifiable irrespective of sex or skin colouration. Both skin features were also found to be stable over time, with no apparent change after more than a year (Fig. 3).

The mean time difference between obtaining query and matching images of individuals in the field that were then used in the survey was mean ± SD = 15.57 ± 17.90 d. A total of 87 anonymous participants completed the survey; the majority were from Australia (75%) and accessed the survey from a mobile device (77%). The mean time required to complete the survey was mean  ± SD = 14.37 ± 25.03 min, with 90% of participants completing the survey within 20 min (less than one min per question). One participant had an irregularly long survey time (51 h) which was not included in these time estimates.

The average number of query images that were correctly matched ranged from 6/30 to 30/30, with a mean ± SD of 26 ± 5 (Fig. 4). The number of query images correctly matched by a participant was not related to the amount of time taken to complete the survey (GLMM, Z₈₃ =  − 0.003, P = 0.996). Some query frogs were matched correctly more often than others (Fig. 5). The lowest rate of successful matching for a frog was 65% of participants while the highest was 100%, with a mean ± SD of 86.02 ± 9.52%.

Discussion

We established that using photographs taken in the field to identify and differentiate between adult individuals of an anuran amphibian based solely on their natural skin marking features is feasible, including the ability of this process of photo-matching to be performed by citizen scientist with very little expertise or prior training. We also met essential criteria for the future application of this technique in capture-mark-recapture studies for our model species; namely demonstration of sufficient inter-individual variation in skin marking features so that individuals can be identified with a high degree of accuracy, and temporal stability of those features so that individuals can be re-identified across subsequent recapture events (Pennycuick, 1978; Marshall & Pierce, 2012).

We found a high level of stable inter-individual variation in facial skin features of both male and female L. fletcheri adults. While participants were asked to match frogs based on a combination of both feature types, the most discernible of these was the region of black banding that runs horizontally above the tympanum. Given its irregular shape, this feature is likely to be the most suitable for manual photo-identification for this species and potentially many other anuran amphibians given the widespread occurrence of this skin feature in this group (e.g., Australian species; Anstis, 2013; Clulow & Swan, 2018). This facial feature may also be a better candidate for inter-individual character discrimination than others previously examined in anurans amphibians such as dorsal patterns, which may be too complex for visual comparison or completely missing in a proportion of individuals of some species (Kenyon, Phillott & Alford, 2009). Differences in the number and positioning of tubercles was also apparent between individuals, though likely more difficult to discern than banding patterns. We suggest that the large number of uniform and repeated features that comprise each array of tubercles across the skin surface would result in a robust visual fingerprint for each individual that would be amendable for computer assisted techniques (Lowe, 1999).

Both facial features in L. fletcheri were temporally stable over at least a 12 month period. This suggests that both are likely to be genetically determined, albeit influenced by environmental factors during development, and permanent (Murray, 1981; Arntzen & Wallis, 1999; Hoffman & Blouin, 2000; Wollenberg et al., 2008), which is critical for studies lasting multiple breeding seasons. Given the short lifespan of L. fletcheri adults (Gould, Clulow & Clulow, 2020), it is difficult to determine whether marking features are stable for longer periods than the recapture intervals obtained in the current study. It is also possible that the appearance of skin features may change if individuals are imaged between different life stages (e.g., sub-adult to adult) or if features are scratched or scarred between capture events (Yoshizaki et al., 2009), although there was no evidence of the latter occurring in our study. Features may also change in appearance due to external sources of variation, such as the method of photographing used. For example, as the skin in the ROI of this species is neither rigid nor flat, the manner in which individuals are held during image acquirement may lead to distortion of features, preventing observers from detecting a true match and leading to possible false rejections. However, this problem was readily mitigated by following a consistent imaging protocol between sampling periods in our study.

The capacity for participants from the public to consistently and correctly identify L. fletcheri individuals based on our target skin features was supported by the results of the online survey. Anonymous participants from the general public were able to correctly match capture (query) and recapture (answer) images of individuals when provided with small image pools, with a majority of participants correctly matching over 85% of queries and a majority of survey questions answered correctly more than 80% of the time. Such rates of successful image classification appear to be comparative to those achieved in the identification of individuals of other animal types (Schofield et al., 2008), and species from large camera surveys (Swanson et al., 2016). Given that participants were provided only minimal training prior to the survey, as well as the short time period taken to answer each question (less than one minute on average), these results suggest that L fletcheri adults can be easily and rapidly photo-identified, at least within small image pools, and that such citizen science programs may be useful in monitoring other amphibian populations.

There was, however, variability in matching score rates between query frogs, with some incurring more false matches than others. Although the reasons for this disparity were not analyzed as part of this study, it is likely that some matches are not as obvious as others, especially if the pool of answer images happen to be of frogs with less distinct features that increase the difficulty in discerning a true match. While 60% of participants correctly matched all individuals, there was also a level of variability in ability for true matches to be detected ‘by eye’. Misidentifications, even if they occur at a low rate, may result in inaccurate parameter estimates in mark-capture-recapture models (Stevick et al., 2001; Yoshizaki et al., 2009). The probability of incorrect matches could be reduced by providing additional training, particularly in the matching of the background array of tubercles that were more difficult to visually assess, by ensuring all images acquired are of sufficient quality before they are uploaded to the photo-database, or by having multiple observers score images and using consensus scores. Indeed, the impact of variable matching ability among participants was not found to be an issue as all query images were overwhelmingly matched correctly when taking into consideration the populous vote.

It is probable that the frequency of correctly photo-matching individuals would decrease as the pool of potential matches increases. Manual photo-matching becomes time consuming and may even become impractical with increasing database size (Katona & Beard, 1990; Speed, Meekan & Bradshaw, 2007; Bolger et al., 2012), a fact that has restricted the use of this technique to small populations (Karanth & Nichols, 1998; Langtimm et al., 2004). Under this scenario, automated approaches may be required. However, it is still important to validate the capacity for individuals from a population to be identified manually. This is because partially automated systems may, on some occasions, only be able to reduce the number of potential matching individuals down to a small pool of candidates with similar marking features, which an observer will then have to sort through to select a true match. This is likely to occur in most populations as some individuals would be expected to share some similar features (Ottensmeyer & Whitehead, 2003), even for those species with high inter-individual marking variability. Validating the ability for correct matches to be manually selected from small image pools is thus an important first step prior to the establishment of computer-assisted methods; a process we have shown could be performed by citizen scientists in future monitoring programs.

We showed that facial marking features in L. fletcheri adults are individually unique and temporally stable, making them effective for photo-identification for conducting non-invasive capture-mark-recapture studies. This process reduces handling times and circumvents the need for invasive tagging techniques with associated logistical, cost and ethical challenges (Reisser et al., 2008). It also has the added benefit of allowing members of the public to become involved in citizen science and conservation programs, either as individuals who are able to acquire images with just the use of a mobile device (under the strict guidance of scientists present), and those who can analyse the data online on their own at any given time. As an increasing number of species require monitoring for their effective management and funding and resources are becoming increasingly spread thin across conservation programs (Margules & Pressey, 2000; Butchart et al., 2010), reliable but less expensive methods of population monitoring are needed. Photographing individuals from mobile devices is likely to help. Although its use requires validation for new species, our results suggest that it is likely to be useful in many other anurans that display distinct facial skin features.

Supplemental Information