Correspondence author: T.A. de Lorm, tad20@ic.ac.uk
Affiliations:
1: Department of Life Sciences, Imperial College London, Silwood Park,
UK
2: Institute of Zoology, Zoological Society of London, London, UK
3: Division of Biosciences, Department of Genetics, Evolution and
Environment, Centre for Biodiversity and Environment Research,
University College London, London, UK
4: Department of Zoology, University of Cambridge, Downing St, Cambridge
CB2 3EJ, UK
5: African Wildlife Conservation Fund, Chishakwe Ranch, Zimbabwe
Running head: Automating African wild dog recognition
Abstract
Reliable estimates of population size and demographic rates are central
to assessing the status of threatened species. However, obtaining
individual-based demographic rates requires long-term data, which is
often costly and difficult to collect. Photographic data offer an
inexpensive, non-invasive method for individual-based monitoring of
species with unique markings, and could therefore increase available
demographic data for many species. However, selecting suitable images
and identifying individuals from photographic catalogues is
prohibitively time-consuming. Automated identification software can
significantly speed up this process. Nevertheless, automated methods for
selecting suitable images are lacking, as are studies comparing the
performance of the most prominent identification software packages.
In this study, we develop a framework that automatically selects images
suitable for individual identification, and compare the performance of
three commonly used identification software packages; Hotspotter,
I3S-Pattern, and WildID. As a case study, we consider
the African wild dog Lycaon pictus , a species whose conservation
is limited by a lack of cost-effective large-scale monitoring. To
evaluate intra-specific variation in the performance of software
packages, we compare identification accuracy between two populations (in
Kenya and Zimbabwe) that have markedly different coat colouration
patterns.
The process of selecting suitable images was automated using
Convolutional Neural Nets that crop individuals from images, filter out
unsuitable images, separate left and right flanks, and remove image
backgrounds. Hotspotter had the highest image-matching accuracy for both
populations. However, the accuracy was significantly lower for the
Kenyan population (62%), compared to the Zimbabwean population (88%).
Our automated image pre-processing has immediate application for
expanding monitoring based on image-matching. However, the difference in
accuracy between populations highlights that population-specific
detection rates are likely and may influence certainty in derived
statistics. For species such as the African wild dog, where monitoring
is both challenging and expensive, automated individual recognition
could greatly expand and expedite conservation efforts.
Key words: automated individual recognition, Hotspotter,
I3S-pattern, Lycaon pictus, photographic
identification, Wild-ID