seaturtle.org : MTN : ARCHIVES : Sign In
The attachment and use of flipper tags requires handling and sometimes capturing the individuals, which on occasion can be disruptive and injurious to turtles (Balazs 1982; Limpus 1992), affect their behavior, and reduce individual survivorship (Nichols et al. 1998). Furthermore, tags are not permanent and incorrect estimation of demographic parameters can occur if the rate at which tags are lost is not considered (Sibly et al. 2005).
The use of natural marks through photo-identification offers an alternative method to identify individuals (Richardson et al. 2000; Jean et al. 2010), which minimizes both spurious results caused by handling and/or injury, and potential animal welfare consequences. This technique is dependent on the assumption that markings are unique and do not change over time, and that the photo-identification method is robust, but when these assumptions are met, photo-id offers a cost effective and permanent method of individual recognition.
Figure 1. The facial scute ‘area of interest’ used by I3S Pattern for a hawksbill turtle. Three reference points, the beak edge, the edge of the outer eye area and the edge of the outer scale area are visible in blue. The outline of the area of interest is highlighted in green. Red circles show the 35 key-points annotated by the software.
Facial scutes have been used in individual turtle identification studies (e.g., Jean et al. 2010, Schofield et al. 2008) and have been shown to be stable for at least four years in adult hawksbill and green turtles, and at least 1,046 days in juveniles (Reisser et al. 2008).
I3S Pattern (Van Tienhoven et al. 2007; Den Hartog & Reijns 2014) is a semi-automated individual recognition software that was originally developed for green turtles (Chelonia mydas). It requires three reference points to be set by the user to orient the image and subsequently form an outline of the facial scute area (Fig. 1). A set number of key-points (Fig. 1) (Den Hartog & Reijns 2014) are automatically extracted and the size and location of each of the key-points are compared to those of each image in the database. After searching the database, the software will present a list of the 50 best matches in order of rank. The user then visually identifies the correct match (if present in the database) using the rank score as a guide.
Previous efficacy studies considered parameters such as the angle at which the picture was taken, clarity, contrast, the extent to which identification features were obscured to define the quality of a picture, and various ways to outline the area of interest (e.g., Den Hartog & Reijns 2012; Dunbar et al. 2014; Speed et al. 2007). I3S Pattern studies concluded that the software is suitable for individual recognition of green turtles. However, the software has not been tested on other turtle species, and the experience of the software user has not been considered in previous studies. The aim of this study is to investigate the parameters that potentially affect the usefulness of I3S Pattern on a nesting hawksbill turtle population, including the user’s ability to match individuals, how the user outlines the area of interest, and the quality of the photographs.
Figure 2. The geographical position of Seychelles within the Indian Ocean on the left with the study area of Mahé Island on the right. The hawksbill turtle nesting beaches used for monitoring are shown in orange.
The nesting population of hawksbill turtles on the island of Mahé, Seychelles (situated at 4°S 55°E) was the focus of this study. The Marine Conservation Society Seychelles (MCSS) monitoring program provided photographs taken during turtle encounters from 2011-2015. Six core nesting beaches were monitored in the south of the island, and eight additional nesting beaches in the south, west and east (Fig. 2), which are infrequently used by nesting turtles. Monitoring was stratified inside and outside of the peak hawksbill nesting season. Typically, patrols were carried out once a week between April and September, twice a week in March and October and three times a week between November and February for core nesting beaches. Additional beaches were monitored once weekly throughout the year. Pictures were mainly taken when a female started laying her eggs as she fell into a trance-like state (Diamond 1976). When possible, to obtain good quality photographs, the head was washed with seawater to remove sand, and photographs were taken when the neck was fully extended. In addition, opportunistic underwater sightings were obtained from divers and snorkelers.
A test database of 60 individuals (110 images) was created from a full database of 125 individuals (298 images) from 2011-2015 nesting seasons, using the best quality photographs available of both sides of the turtle’s head (when possible). The influence of four parameters on I3S Pattern recognition software’s performance was tested (Table 1). The matching success definition varied depending on the parameter being tested.
Table 1. Methods used to test the influence of four parameters on I3S Pattern recognition software’s performance. For parameters 2 - 4, in each test the samples were compared to the test database (60 individuals). For parameter 1, users compared the simulated encounters to the test database. The matching success definition varied depending on the parameter being tested.
The matching ability of six users with varying skill levels was examined using 25 simulated encounters (see Table 1) to determine if user experience influenced the overall performance of I3S Pattern. To determine if the way that a single user outlines the area of interest had any impact on I3S Pattern’s ability to select the correct matching individual from the database, the area of interest was outlined in six different ways (Fig. 3) for a subset of high quality images (Table 1). To assess the influence of image quality on matching success, the percentage of reflection and sand obscuring the area of interest (0-5%, 6-20%, 21-40%, and >40%) was estimated for a subset of photographs (Table 1). The environment that the image was captured in was also considered (beach or in-water). To determine I3S Pattern’s performance at identifying known resightings, one image from 50 known matched pairs of photographs in the test database was removed from the database and subsequently matched against the remaining database (Table 1). This means that 50 photographs from known matches (i.e., from 100 photographs, 50 from a first encounter and 50 from a second encounter) were deleted from the existing database. These were re-entered and it was noted whether the correct match was displayed by the software.
All statistical analyses were carried out using R version 3.1.2 (R Development Core Team 2010) in the R Studio environment (R Studio 2013). To investigate the influence of a parameter (Table 1) on the identification success where 1 = successful match and 0 = no identification (definitions of matching success vary depending on parameter tested), binomial GLMs with a complementary log – log (cloglog) link function were used to find the model that best fits the data. Model refinement was carried out using stepwise addition, and model selection was done using the lowest Akaike Information Criterion (AIC). Model significance was assessed using a one-way ANOVA with a χ2-test. Logistic regression coefficients were examined for all levels of variables within the best model to investigate their influence on the matching performance. Additional variables included as factors in the binomial GLMs were sand coverage, reflection coverage and the environment in which the image was captured (beach or in-water).
The previous software experience of the users in this study did not significantly affect the matching success scores. Matching success was significantly affected, however, by the image quality parameters sand and reflection obscuring the area of interest (final model χ2 3 = 17.79, p < 0.001, Table 2). The inexperienced user identified 72% of simulated encounters correctly, which was lower than users with other software experience or I3S Classic experience (82% and 84%, respectively) (Table 3). Users with previous experience identified all ‘new’ individuals not present in the test database, but failed to identify some resightings.
How the user outlined the area of interest did not significantly affect I3S Pattern matching success scores. Matching success was significantly affected, however, by the image quality parameters sand and reflection obscuring the area of interest (final model χ2 3 = 12.78, p < 0.01, Table 2).
Table 2. Results of the binomial GLMs for three investigated parameters: user experience, user outlining the area of interest, and image quality.
Table 3. Mean I3S recognition performance according to user experience. Correct values refer to the number and percentage successfully identified. Potentially correct values refer to the number and percentage displayed within the 50 displayed images by the software that the user could have recognized.
Table 4. I3S Pattern recognition performance. Ranking results per dataset of number of correctly identified images and percentage of correct identification per parameter. The two right-hand columns present the combined results where all images of sand coverage, reflection coverage and resightings were added and the percentage was calculated.
When testing image quality, the amount obscured by sand coverage significantly influenced the matching performance of I3S Pattern, whereas reflection coverage did not (final model χ2 1 = 36.91, p < 0.01, Table 2). From the calculated regression coefficients, it was evident that images with 21-40% and >40% of the area obscured by sand were less likely to be correctly identified when compared to images with 0-5% and 6-20% of the area obscured by sand. I3S performed well, ranking 80.95% (85 images) of the correct matches on position 1 (Table 4). Only five images corresponding to 4.76% of the data for sand and reflection were not in the top 50 matches displayed by the software (Table 4).
I3S Pattern matching success of known resightings was low with only 79.33% of images within the top 50 matches (Table 4). All unsuccessfully matched images were from the 2011/2012 nesting season, compared to 100% of resighting images from the 2014/2015 season being ranked in the top position. Resightings from 2011/2012 nesting season were generally poorer in quality during the monitoring program’s early use of this method.
Figure 3. Six different ways of outlining the area of interest in I3S Pattern individual recognition software. A. Normal: a general outline of the area of interest that does not follow every border of scales B. Outline scales: outline the area following the borders of scales. C. Outline scales 1: similar to B. but including the outer scale leading up the head as some individuals have one big scale instead of two small ones. D. Outline scales 2: similar to B. but including the scales from the lower jaw. E. Outline scales 3: a combination of C and D including the outer scale leading up the head as well as the scales from the lower jaw. F Outline scales 4: similar to B but instead of following the scale around the eye, the lower point is directly connected to the edge of the outer eye area.
For all parameters tested, matching success was shown to be affected significantly by image quality (reflection and/or sand). This indicates the need for a cut-off standard for low quality images to reduce matching errors. Previous studies graded photographs according to the extent to which distinctive features used for identification were hidden or obscured and avoided utilizing images with large hidden areas (Friday et al. 2000, Markowitz 2004). We suggest that photographs only with < 20% sand or reflection coverage in the area of interest should be used for matching and subsequent inclusion into a database. Likewise, appropriate image acquisition training should be given, sand should be washed off the area of interest with seawater, and the individual should be shielded from the sun to decrease any reflection.
It is widely accepted that the horizontal angle of photographs influences matching performance (Caci et al. 2013; Dunbar et al. 2014; Rocha et al. 2013; Speed et al. 2007), and although not tested in the present study, we noted that images taken at a high vertical or horizontal angle generally resulted in poorer match scores. Den Hartog & Reijns (2014) suggest that photographs are taken with an angle <30° from perpendicular as straight on at head level, which would provide best photographs for analysis, is not always possible.
Although the experience level of the software user did not affect matching success, all users still incorrectly identified some individuals. Missed matches for all users were primarily from the same encounters and consisted of low-quality images indicating that these would likely also be missed via manual matching. These images were however shown in the top 50 ranked images by I3S (96%, Table 3) highlighting that the software performs well regardless of the previous experience of users in this study. Appropriate training is essential when using any individual recognition software, along with employing quality control measures, such as matching by multiple experienced users, is advisable to minimize the likelihood of incorrect identifications.
The way in which the user outlined the area of interest was not found to affect the matching success of I3S Pattern, this stands in contrast to previous studies on the performance of similar software, I3S Classic (Den Hartog & Reijns 2012).
Flipper tag information was available for five resightings determined by using I3S Pattern software. In all but one case the flipper tag identification matched the photo-identification confirming the reliability of the software to correctly identify individuals. One resighting was solely identified by the flipper tag and failed to be identified by the software and user due to the low quality of the photograph. In this low size of resightings where both photographs and tag information were available, this number still represents 20% of matches confirmed by flipper tags where the software was not able to identify the individual. However, this photograph was of poor quality, and quality has increased immensely since then and with current photograph quality this number would likely be much lower.
For use as a semi-automated individual recognition tool, I3S Pattern performs well, with 95-96% of correct matches being shown in the top 50 ranked positions (for user experience and image quality datasets), from which trained users can manually determine the correct identification. This is similar to the only other evaluation of I3S Pattern. Rocha et al. (2013) reported 95% of iris patterns analyzed in lizards within the top 50 ranked images. I3S Pattern has the potential to generate accurate recognition of individual hawksbill turtles, when used in conjunction with good quality imagery, appropriate training and quality control measures.
Acknowledgments. We thank Peter Holden, Savinien Leblond, Michel Monthy, and Rachel Pool for participating in the parameter testing, and Vanessa Didon for monitoring support and guidance. The work conducted at MCSS was approved by the Government of Seychelles and supported under a component of UNDP-GEF Project ‘Strengthening Seychelles’ protected area system through NGO management modalities. The Maptool program (product of SEATURTLE.ORG <https://seaturtle.org>) was used for graphics in this paper.
ALLEN, Z.C., N.J. SHAH, A. GRANT, G-D. DERAND & D. BELL. 2010. Hawksbill turtle monitoring in Cousin Island Special Reserve, Seychelles: an eight-fold increase in annual nesting numbers. Endangered Species Research 11: 195-200.
BAILLIE, J., C. HILTON-TAYLOR & S.N. STUART. 2004. 2004 IUCN Red List of Threatened Species: A Global Species Assessment. IUCN, Gland, Switzerland. 191p.
BALAZS, G.H. 1982. Factors affecting the retention of metal tags on sea turtles. Marine Turtle Newsletter 20:11-14.
CACI, G., A.B. BISCACCIANTI, L. CISTRON, L. BOSSO, A.P. GARONNA & D. RUSSO. 2013. Spotting the right spot: computer-aided individual identification of the threatened cerambycid beetle Rosalia alpina. Journal of Insect Conservation 17: 787-795.
CAUGHLEY, G. 1994. Directions in conservation biology. Journal of Animal Ecology: 215-244.
DEN HARTOG, J. & R. REIJNS. 2012. Performance evaluation of I3S on whale shark data. Retrieved from <http://www.reijns.com/i3s>. Access date: 12/07/2018.
DEN HARTOG, J. & R. REIJNS. 2014. I3S Pattern Manual. Interactive Individual Identification System. Version 4.0.2 <http://www.reijns.com/i3s/download/I3S%20Pattern.pdf> Access date: 12/07/2018
DIAMOND, A.W. 1976. Breeding biology and conservation of hawksbill turtles, Eretmochelys imbricata L., on Cousin Island, Seychelles. Biological Conservation 9: 199-215.
DUNBAR, S.G., H.E. ITO, K. BAHJRI, S. DEHOM & L. SALINAS. 2014. Recognition of juvenile hawksbills Eretmochelys imbricata through face scale digitization and automated searching. Endangered Species Research 26: 137-146.
FRAZIER, J. 1971. Observations on sea turtles at Aldabra Atoll. Philosophical Transactions of the Royal Society B: Biological Sciences 260: 373-410.
FRIDAY, N., T.D. SMITH, P.T. STEVICK & J. ALLEN. 2000. Measurement of photographic quality and individual distinctiveness for the photographic identification of humpback whales, Megaptera novaeangliae. Marine Mammal Science 16: 355-374.
IUCN (2015). IUCN Red List of Threatened Species Version 2015.1 <http://www.iucnredlist.org>. Access date: 12/07/2018
JEAN, C., S. CICCIONE, E. TALMA, K. BALLORAIN & J. BOURJEA. 2010. Photo-identification method for green and hawksbill turtles - First results from Reunion. Indian Ocean Turtle Newsletter 11: 8-13.
LEBRETON, J.-D., K. P. BURNHAM, J. CLOBERT & D.R. ANDERSON. 1992. Modeling survival and testing biological hypotheses using marked animals: a unified approach with case studies. Ecological Monographs 62: 67-118.
LIMPUS, C. 1992. Estimation of tag loss in marine turtle research. Wildlife Research 19: 457-469.
MARKOWITZ, T.M. 2004. Social organization of the New Zealand dusky dolphin. PhD Thesis. Texas A&M University, College Station, Texas, USA. 255p.
MEYLAN, A.B. & M. DONNELLY. 1999. Status justification for listing the hawksbill turtle (Eretmochelys imbricata) as Critically Endangered on the 1996 IUCN Red List of Threatened Animals. Chelonian Conservation & Biology 3: 200-224.
MORTIMER, J.A. & R. BRESSON. 1994. The hawksbill nesting population at Cousin Island, Republic of Seychelles: 1971-72 to 1991-92. Schroeder B.A. & B.E. Witherington (Comps.) Proceedings of the 13th Annual Symposium on Sea Turtle Biology & Conservation. NOAA Tech Memo NMFS-SEFSC-341. pp. 115-116.
MORTIMER, J.A. & M. DONNELLY. 2008. Eretmochelys imbricata. IUCN Red List of Threatened Species. <http://www.iucnredlist.org>
MORTIMER, J.A. 1984. Marine Turtles in the Republic of the Seychelles: Status and Management: Report on Project 1809 (1981-1984), IUCN.
MORTIMER, J.A. & J. COLLIE. 1998. Status and conservation of sea turtles in the Republic of Seychelles. Epperly, S.E. & J.Braun (Comps.). Proceedings of the 17th Annual Sea Turtle Symposium. NOAA Tech Memo NMFS-SEFSC-415. pp. 74-76
MORTIMER, J.A., J. COLLIE & N. BONIFACE. 2011. Seasonality and status of nesting hawksbill (Eretmochelys imbricata) and green turtles (Chelonia mydas) at D’Arros Island, Amirantes Group, Seychelles. Chelonian Conservation & Biology 10: 26-33.
NICHOLS, W.J., J.A. SEMINOFF & A. RESINDIZ. 1998. Plastic ‘Rototags’ may be linked to sea turtle bycatch. Marine Turtle Newsletter 79:20-21.
R DEVELOPMENT CORE TEAM. 2010. R: A language and environment for statistical computing. Vienna, Austria, R Foundation for Statistical Computing. <http://www.R-project.org>
R STUDIO, I. 2013. R Studio Desktop 0.97.449 Integrated development environment (IDE) for R. <http://www.R-project.org>
REISSER, J.W., M.C. PROIETTI, P.G. KINAS & I. SAZIMA. 2008. Photographic identification of sea turtles: method description and validation, with an estimation of tag loss. Endangered Species Research 5: 73-82.
RICHARDSON, A., L.H. HERBST, P.A. BENNETT & U.K. BENNETT. 2000. Photo-identification of Hawaiian green sea turtles. In: Abreu-Grobois, F.A., R. Briseño-Dueñas, R. Márquez & L. Sarti (Comps.). Proceedings of the 18th Annual Sea Turtle Symposium. NOAA Tech Memo NMFS-SEFSC-436, p. 249.
ROCHA, R., T. CARRILHO & R. REBELO. 2013. Iris photo-identification: A new methodology for the individual recognition of Tarentola geckos. Amphibia-Reptilia 34: 590-596.
SCHOFIELD, G., K.A. KATSELIDIS, P. DIMOPOULOS & J.D. PANTIS. 2008. Investigating the viability of photo-identification as an objective tool to study endangered sea turtle populations. Journal of Experimental Marine Biology and Ecology 360: 103-108.
SIBLY, R.M., D. BARKER, M.C. DENHAM, J. HONE & M. PAGEL. 2005. On the regulation of populations of mammals, birds, fish, and insects. Science 309: 607-610.
SPEED, C.W., M.G. MEEKAN & C.J. BRADSHAW. 2007. Spot the match - wildlife photo-identification using information theory. Frontiers in Zoology 4: 2. DOI:10.1186/1742-9994-4-2
VAN TIENHOVEN, A., J.E. DEN HARTOG, R.A. REIJNS & V.M. PEDDEMORS. 2007. A computer-aided program for pattern-matching of natural marks on the spotted ragged-tooth shark Carcharias taurus. Journal of Applied Ecology 44: 273-280.