Reducing identification errors of African carnivores from photographs through computer-assisted workflow

Abstract

Precise and accurate estimates of population size are fundamental to the study and conservation of wildlife. Identification of individual animals is often required to obtain such estimates, yet manual classifications by human observers induce bias, which can propagate across long-term datasets or large spatial scales. Pattern recognition algorithms have been developed to aid identification efforts and here, we demonstrate the efficacy of this technology to reduce misidentifications of African large carnivores in historic data. We used a 7-year camera-trapping dataset from north-central Namibia and revised cheetah and leopard individuals identified by human observers through a pattern recognition algorithm, HotSpotter, implemented in a web-based and open-source application, the African Carnivore Wildbook (ACW). Verification of individuals with ACW resulted in a reduction from 43 to 40 cheetah individuals and from 59 to 46 leopard individuals. This is equivalent to a difference of 7% and 22% of individuals identified for cheetahs and leopards, respectively. Additionally, this revision increased the proportion of individuals that were detected over multiple years and at multiple locations. Our findings may have implications for population and trend estimates of these and other species, given that current estimates often rely on manual identification that could overestimate population size.