Concurrent visual encounter sampling validates eDNA selectivity and sensitivity for the endangered wood turtle (Glyptemys insculpta).

Environmental DNA (eDNA) has been used to record the presence of many different organisms in several different aquatic and terrestrial environments. Although eDNA has been demonstrated as a useful tool for the detection of invasive and/or cryptic and declining species, this approach is subject to the same considerations that limit the interpretation of results from traditional survey techniques (e.g. imperfect detection). The wood turtle is a cryptic semi-aquatic species that is declining across its range and, like so many chelonian species, is in-need of a rapid and effective method for monitoring distribution and abundance. To meet this need, we used an eDNA approach to sample for wood turtle presence in northern Virginia streams. At the same time, we used repeat visual encounter surveys in an occupancy-modelling framework to validate our eDNA results and reveal the relationship of detection and occupancy for both methods. We sampled 37 stream reaches of varying size within and beyond the known distribution of the wood turtle across northern Virginia. Wood turtle occupancy probability was 0.54 (0.31, 0.76) and while detection probability for wood turtle occupancy was high (0.88; 0.58, 0.98), our detection of turtle abundance was markedly lower (0.28; 0.21, 0.37). We detected eDNA at 76% of sites confirmed occupied by VES and at an additional three sites where turtles were not detected but were known to occur. Environmental DNA occupancy probability was 0.55 (0.29, 0.78); directly comparable to the VES occupancy estimate. Higher probabilities of detecting wood turtle eDNA were associated with higher turtle densities, an increasing number of days since the last rainfall, lower water temperatures, and lower relative discharges. Our results suggest that eDNA technology holds promise for sampling aquatic chelonians in some systems, even when discharge is high and biomass is relatively low, when the approach is validated and sampling error is quantified.