Computer-assisted photo identification outperforms visible implant elastomers in an endangered salamander, Eurycea tonkawae

Abstract

Despite recognition that nearly one-third of the 6300 amphibian species are threatened with extinction, our understanding of the general ecology and population status of many amphibians is relatively poor. A widely-used method for monitoring amphibians involves injecting captured individuals with unique combinations of colored visible implant elastomer (VIE). We compared VIE identification to a less-invasive method - computer-assisted photographic identification (photoID) - in endangered Jollyville Plateau salamanders (Eurycea tonkawae), a species with a known range limited to eight stream drainages in central Texas. We based photoID on the unique pigmentation patterns on the dorsal head region of 1215 individual salamanders using identification software Wild-ID. We compared the performance of photoID methods to VIEs using both 'high-quality' and 'low-quality' images, which were taken using two different camera types and technologies. For high-quality images, the photoID method had a false rejection rate of 0.76% compared to 1.90% for VIEs. Using a comparable dataset of lower-quality images, the false rejection rate was much higher (15.9%). Photo matching scores were negatively correlated with time between captures, suggesting that evolving natural marks could increase misidentification rates in longer term capture-recapture studies. Our study demonstrates the utility of large-scale capture-recapture using photo identification methods for Eurycea and other species with stable natural marks that can be reliably photographed.

Figure 1. Head pattern recognition in Eurycea tonkawae.

Pair of images from (A) two different individuals and (B) the same individual one year apart. Lines connect matching SIFT features.

Figure 2. Diagram of the two-step image-matching process.

Image pairs above the score threshold of 0.1 were considered matches (criterion 1). The top 100 first-ranked image pairs below 0.1 were compared by eye.

Figure 3. Similarity scores and matching success.

Frequency of similarity scores for image pairs from different (black) and the same individuals (grey) from the high-quality dataset. Inset shows the lower range of similarity scores for top ranked image pairs only. The shaded region between dashed lines (2) indicates the range (similarity scores: 0.017–0.1) in which we visually (by-eye) compared 100 potential matching pairs. Similarity scores above this range (3) always involved photo pairs from the same individuals (according to VIE tags). Below this range (1), all but one photo pair came from non-matching individuals (according to VIE tags).

Figure 4. Effect of time on similarity score.

Similarity scores versus days between photo captures of the same individual for (A) low-quality (n = 896) and (B) high-quality photos (n = 1090). Tables.