Feasibility of using pattern-recognition software for photographic identification of tuatara (Sphenodon punctatus)

Abstract

The ability to distinguish individual animals is fundamental to ecological research, wildlife and livestock management. Individuals are recognised by using the animals’ natural marks that remain stable through time, by applying artificial temporary or permanent tags or marks to an animal. Recently, an increased interest in using non-invasive methods has arisen due to ethical concerns around the use of permanent tags or marks. Photographic identification (photo-ID) is a non-invasive identification technique that can be as reliable as other methods for species with distinctive patterns or marks. In this study, I aimed to determine the feasibility of photo-ID for adult and juvenile tuatara (Sphenodon punctatus) and to identify which of several body patterns is best to individually identify tuatara. This research also evaluated the comparative performance of three open-source photo-ID software packages (Wild-ID, Interactive Identification System (I3S) and StripeSpotter). First, I compared the identification performance of three different areas of the body: (i) the ventral area (further sub-divided into underside, throat and chest), (ii) head and body patterns including the spines on the right side of the body (“right side”) and (iii) the iris on the right side (“right eye”), among three photo-ID software packages. In total, I used photos from 196 tuatara, including captive and wild individuals. All tuatara were of Stephens Island (Takapourewa) origin. I found that the use of the chest pattern using Wild-ID was the best body-part- software combination to individually identify tuatara. In a database of 1,090 photos for the chest collected in 2013 and 2017, Wild-ID correctly identified 99% of the photos. In addition, from the corrected matches, 99.4% of the recaptures were ranked among the top five. With Wild-ID the right side (97% of correctly identified) showed a lower identification rate than chest pattern, but for a smaller sample size (172 individuals and 667 photos) and over a shorter elapsed time (up to 652 days). The right eye showed the lowest matching rate among all software packages. However, this result may have been influenced by the eye photos being of lower quality than chest and right-side photos. There was a significant difference in accuracy among software packages for all body parts, except when comparing right eye between Wild-ID and I3S. Wild-ID had the highest identification rate, and StripeSpotter the lowest for all body parts. These results present a reliable alternative to toe-clipping or PIT-tagging for individual identification of tuatara over a period of at least three years. This method is feasible from early juvenile tuatara. Until it is known whether chest patterns remain stable for longer than three years, I recommend that another marking technique should be applied in conjunction with photo-ID for at least most of a tuatara’s lifespan. My results encourage testing a computer-assisted identification in a field setting for tuatara and for other reptiles and amphibians in New Zealand where the assumptions of photo-ID are met.