INTRODUCTION
Marine ecosystems are dynamic, and conservation of key species that
inhabit these ecosystems requires long-term monitoring of populations
across a range of temporal and geographic scales Methods for long-term
monitoring of coastal species, including harbor seals, are often
invasive, costly, and time-consuming (Cunningham 2009), underscoring the
need for new techniques for systematic data collection and analysis. The
automation of these processes can be an effective technique for
monitoring population dynamics, as automation increases reproducibility
while decreasing cost and labor (.
Harbor seals are an ideal species for long-term monitoring as they are
highly mobile animals that inhabit a large geographic range and are
ecologically and economically important as top predators (Aarts et al.,
2019). In addition, harbor seals can be observed non-invasively as they
congregate at “haul-out” sites—essential areas where seals come out
of the water to rest on rocky islets, allowing them to thermoregulate
and avoid predation—which make them easily visible to researchers from
afar (. As top predators, seal populations effect ecosystem dynamics,
with healthy populations likely decreasing competition among species
such as flounder, sole and dab, and, in turn, influencing the balance of
both ecologically and economically critical fish populations . Increases
in seal populations along the Atlantic coast have also increased the
numbers of sharks that inhabit coastal waters, potentially affecting
tourism revenue in addition to local ecosystems .
Harbor seals are important indicators of ecosystem health because they
are susceptible to climate change and, given their extensive overlap
with human activities both in and out of the water, are particularly
vulnerable to increased anthropogenic activity (Allen et al., 1984).
Over the last century, the Atlantic coast populations of harbor seals in
northeastern North America have a history of heavy exploitation.
Following the Marine Mammal Protection Act of 1972, populations of
harbor seals off the Northeast coast of the U.S, successfully rebounded
to healthy population numbers, but the steep decline in abundance prior
to any legislation is evidence of the potential vulnerability of the
population to acute or chronic ecological challenges.
As key regulators and indicators of ecosystem health, monitoring harbor
seal population levels and movement patterns is essential. Tagging
methods have been widely used in the past, however these GPS-monitoring
devices are expensive, ranging from $1000 to $3000 for one device
(GPS and VHF Tracking Collars Used for Wildlife Monitoring ,
2017). In addition, the attachment of external devices may interfere
with behaviors such as swimming speed, oxygen consumption, and metabolic
rate, potentially corrupting the data collected or harming or disturbing
the individual . Aerial and visual observation methods limit
interference with seal behavior, but both techniques are time consuming
and expensive . Photo based identification techniques also have the
advantage of being non-invasive, but manual interpretation of
photographs is time-intensive and often limited to small-scale projects.
For seals, manual matching based on fur colors is difficult due to
changing coat colors as seals mature and during annual molting. However,
some promising progress has been made using analysis of pelage markings,
i.e. spots on the seal’s coat that can be reliably used as diagnostic
tools (Cunningham, 2009).
Here, we propose the use of automated facial recognition technology as a
system for identification of seal individuals for ecological and
population studies. We used deep learning methods and convolutional
neural networks to develop SealNet, a redesign of the PrimNet software
(Deb et al., 2018) developed for primates. CNN-based facial recognition
software achieves identification accuracies of 93.8% with lemurs ,
92.5% with chimpanzees , and 97.27% with pandas . Another software,
BearID, recently achieved close to 100% face chipping accuracy (number
of faces recognized in an unprocessed photo) despite an overall pipeline
identification accuracy of 82.4% SealNet contributes a new software
package to automate the process of seal identification for use by
researchers in the field.
In this paper, we outline the creation of a graphical user interface
(GUI), that allows the user to automatically identify, align and chip
seal faces to facilitate the processing of raw data. Then, using a deep
convolutional neural network (CNN) suitable for small datasets (e.g.,
100 seals with five photos per seal), we developed a seal face
recognition software. We trained and tested this software on a wild
population of Atlantic harbor seals in Casco Bay, Maine, U.S.A. We
compare the performance of SealNet with its predecessor PrimNet and show
that SealNet outperforms this software in the prediction of harbor seal
identities. In a time of rapid ecosystem changes, SealNet represents a
new tool non-invasive tracking of seals for use in ecological and
behavioral studies.