1. Introduction
Reliable estimates of population size and demographic rates are central
to monitoring the status of threatened species. However, obtaining
individual-based demographic parameters requires long-term data,
gathered through intensive monitoring that is often costly and difficult
to conduct (Horswill, Humphreys & Robinson, 2018; Caughlan, 2001).
Identification of individuals from photographic records could provide an
inexpensive alternative, and open up the possibility of using camera
traps and citizen scientists to expand the spatial coverage of
monitoring (Seber, 1965; Marnewick et al., 2014; Wearn & Glover-Kapfer,
2019). This method can be used for species where individuals can be
identified from individual markings, including many threatened species
(Durant et al., 2014, Pierce & Norman, 2016).
Photographic records have already been used to estimate demographic
parameters in several endangered species. For example, long term
photographic data have been used to obtain survival and abundance
estimates of tigers Panthera tigris and cheetahs Acinonyx
jubatus (Karanth & Nichols, 2011; Kelly et al., 1998), and tourist
images have been used to estimate population sizes of whale sharksRhincodon typus (Davies et al., 2013). In addition, photographs
can provide data on individual movement, ranging behaviour, and social
structure (Randić et al., 2012; Armstrong et al., 2019). Many species
are photographed frequently as part of monitoring programs, and by
members of the public, including tourists. Such image catalogues
therefore represent a large, and potentially under-used, data resource
that inform conservation action.
Nevertheless, visually identifying all individuals in large image
databases is time-consuming. To partly automate this process, several
software packages are available to match images based on an individual’s
unique body markings (e.g. APHIS and WildID, Óscar et al. , 2015;
Bolger et al. , 2012). These image-matching software packages
assist the user by ranking potential image matches using a similarity
score. The algorithms underpinning the software packages find these
potential matches by comparing images on either a pixel-by-pixel or
feature basis. Pixel-based algorithms, such as APHIS, have been
successfully applied to numerous species, including horseshoe whip
snakes Hemorrhois hippocrepis and Balearic lizardsPodiarcis lilfordi (Óscar et al., 2015; Rotger, 2019). However,
they are susceptible to differences in camera angle, scale, and cropping
(Matthé et al., 2017), and are therefore unsuitable for animals that
cannot be caught and photographed using a standardised methodology. By
contrast, feature-based software packages, such as Wild-ID (Bolger,
2012), I3S-Pattern (Reijns, 2014) and Hotspotter
(Crall et al., 2013), match images based on unique features including
spots, stripes, blotches, or other marks. The algorithms that
feature-based packages use vary, but all have a higher tolerance to
differences in camera angle, scale, and lighting conditions than
pixel-based algorithms. However, researchers are still required to
select images that are suitable for identification, in that the
distinctive marks must face the camera and must be clearly visible.
Furthermore, when these suitable images are selected, the user has to
crop the region of interest from the image, which is laborious,
potentially preventing the application to large image catalogues (Miguel
et al., 2019).
The feature-based packages have been tested on a range of taxa (Table
1), and the reported proportion of true matches that the software
detects, i.e., accuracy rate, varies markedly, ranging between 36% and
100%. This variation can be attributed to differences in species
markings, image quality, size of database, how many potential matches
were inspected per image, and the image-matching software used (Nipko,
Holcombe & Kelly, 2020; Matthé et al., 2017; Crall et al., 2013).
Studies directly comparing the accuracy of different feature-based
packages are considerably more limited, even though the most accurate
software differs between species. For example, studies on jaguarsPanthera onca, ocelots Leopardus pardalis , and Saimaa
ringed seals Phoca hispida saimensis found that Hotspotter
outperformed Wild-ID (Nipko et al. 2020, Chehrsimin et al. 2018),
whilst studies on amphibian species found that Wild-ID outperformed
I3S-Pattern (Nipko et al ., 2020, Matthéet al. , 2017), and Hospotter (Morrison et al ., 2016). The
only study that directly compared all three software packages found that
Hotspotter was superior to I3S-pattern and Wild-ID for
identifying individual green toads Bufotes viridis (Burgstaller,
Gollmann & Landler, 2021). To date, studies comparing image-matching
accuracy across all three software packages for a mammal species are
lacking.
African wild dogs Lycaon pictus (hereafter ‘wild dogs’) have
unique coat markings, which vary between individuals (Figure 1, Maddock
& Mills, 1994). Wild dogs are classified as globally endangered, and a
lack of cost-effective large-scale monitoring has been highlighted as a
major limitation in developing effective conservation strategies
(Woodroffe & Sillero-Zubiri, 2020). Consequently, there is a pressing
need to devise new approaches for monitoring wild dogs. Demographic
processes of African wild dogs are typically studied by monitoring a
subset of individuals fitted with tracking collars (Woodroffe et
al. , 2019; Rabaiotti et al 2021; Jenkins et al. , 2015). Such
collar-based monitoring is labour-intensive and expensive, so upscaling
is difficult. However, many wild dog packs have already been
systematically photographed as part of monitoring programs, and many are
also regularly photographed by tourists. Therefore, photographic
identification of wild dogs potentially offers a non-invasive, cheaper
approach for monitoring, and could reduce uncertainties in demographic
rates and expand the spatial representation of monitoring (Maddock &
Mills, 1994; Marnewick et al., 2014).
Wild dog coat patterns contain tan, white, and black patches that can
vary considerably between populations. For example, wild dogs in eastern
African populations tend to have coats consisting of predominantly black
fur, whilst those in southern African populations have relatively more
white and tan blotches (Figure 1, McIntosh, Woodroffe & Rabaiotti,
2016). In amphibian species with contrasting colour patterns of similar
blotches, Wild-ID, Hotspotter, and I3S-Pattern have
been shown to effectively match images of the same individual, reaching
accuracy rates of up to 97% (Matthé et al., 2017; Burgstaller, Gollmann
& Landler, 2021). Therefore, it is likely that feature-based image
matching algorithms will effectively identify individual wild dogs from
image catalogues. However, variation in the degree of contrast in the
colour patterns amongst populations could affect the image-matching
accuracy, and the best performing software package could therefore also
vary between populations.
In this study, we develop a method to automatically isolate and crop
images from catalogues that are suitable for automated image-matching.
We then use these images to compare the efficacy of three feature-based
software packages with different underlying image-matching algorithms
(I3S-Pattern, Hotspotter, and Wild-ID; Reijns, 2014;
Bolger, 2012; Crall et al. , 2013). Finally, we compare whether
there is a difference in the accuracy of each software package between
two populations with differing coat patterns.
2. Methods