Abstract
Estimating
the
size
of
animal
populations
plays
an
important
role
in
evidence‐based
conservation
and
management.
Some
methods
for
estimating
population
rely
on
animals
being
individually
identifiable.
Traditionally,
this
has
been
done
by
marking
physically
captured
animals,
but
increasingly,
with
distinctive
natural
markings
are
surveyed
noninvasively
using
cameras.
Animal
reidentification
from
photographs
is
usually
manually,
which
expensive,
laborious,
requires
considerable
skill.
An
alternative
to
develop
computer
vision
that
can
support
or
replace
manual
identification
task.
We
developed
automated
approach
deep
learning
identify
whether
a
pair
same
individual
not.
The
core
similarity
network
uses
paired
convolutional
neural
networks
triplet
loss
function
summarize
image
pairs
decide
they
individual.
Prior
main
matching
step,
two
additional
perform
segmentation,
cropping
object
within
image,
orientation
prediction,
deciding
side
was
photographed.
applied
four
species,
images
often
spanning
several
years:
systematic
surveys
bottlenose
dolphins
(
Tursiops
truncatus
,
2008–2019)
harbor
seals
Phoca
vitulina
2015–2019),
citizen
science
dataset
western
leopard
toads
Sclerophrys
pantherina
unknown
dates),
publicly
available
repository
humpback
whale
Megaptera
novaeangliae
dates).
For
these
our
best‐performing
models
were
able
different
individuals
95.8%,
94.6%,
88.2%,
83.8%
cases,
respectively.
found
functions
outperformed
binary
cross‐entropy
data
augmentation
curation
training
provided
small
consistent
improvements
performance.
These
results
demonstrate
potential
or,
more
likely,
facilitate
efforts.
Journal of Wildlife Management,
Год журнала:
2023,
Номер
87(3)
Опубликована: Фев. 8, 2023
Abstract
Non‐invasive
genetic
sampling
(NGS)
methods
are
becoming
a
mainstay
in
wildlife
monitoring
and
can
be
used
with
spatial
capture‐recapture
(SCR)
to
estimate
population
density.
Yet
SCR
based
on
NGS
remains
relatively
underused
for
ungulate
monitoring,
despite
the
importance
of
robust
density
estimates
this
ecologically
economically
important
group
species.
This
may
part
attributed
biological
characteristics
species
data
collection
that
lead
violations
model
assumptions.
We
conducted
simulation
study
evaluate
robustness
spatially
heterogeneous
(i.e.,
configuration
individuals
into
groups
variable
sizes
composition),
individual
heterogeneity
space‐use
patterns,
adaptive
variation
detectability
across
space
correlates
density).
evaluated
each
violation
separately
combination.
parameterized
our
simulations
published
information
preliminary
analyses
sets
3
species:
chamois
(
Rupicapra
rupicapra
),
red
deer
Cervus
elaphus
wild
boar
Sus
scrofa
).
While
were
grouping
sampling,
abundance
could
negatively
biased
(up
10%
simulations)
presence
unaccounted
use.
The
degree
which
underestimated
depended
mostly
amount
use
among
age
classes.
bias
was
also
accompanied
by
reduction
precision
coverage
probability
estimators.
discuss
implications
these
findings,
possible
approaches
identify
problematic
available
(goodness‐of‐fit
tests),
potential
further
developments
models
ensure
reliable
populations
from
data.
Global Ecology and Conservation,
Год журнала:
2024,
Номер
50, С. e02838 - e02838
Опубликована: Фев. 7, 2024
The
rapid
decline
in
global
biodiversity
underscores
the
critical
need
for
comprehensive
monitoring
of
wildlife
distribution
and
abundance.
This
study
explores
trends
applied
hierarchical
modeling,
which
is
an
important
tool
addressing
these
conservation
challenges.
By
analyzing
a
dataset
697
peer-reviewed
articles
published
between
2002
2022,
we
examine
taxonomic
focus,
detection
procedures,
designs,
modeling
choices
within
field
population
ecology.
Our
findings
revealed
that
most
studies
concentrated
on
single
groups,
particularly
mammals
birds.
Data
collection
methods
included
visual
surveys,
acoustic
camera
traps,
with
some
combining
multiple
techniques.
Notably,
United
States
dominated
geographical
accounting
46%
papers.
In
terms
approaches,
single-season
occupancy
was
prevalent,
followed
by
various
other
models,
including
multi-species
N-mixture
models.
While
has
gained
popularity,
citations
remained
relatively
modest,
only
few
achieving
over
100
citations.
Authorship
analysis
highly
collaborative
network
researchers,
key
authors
contributing
significantly
to
field's
development
dissemination.
Co-authorship
co-citation
networks
highlighted
importance
who
can
bridge
differing
scientific
groups
those
have
made
substantial
contributions
methods.
Despite
its
growth,
faces
challenges
related
standardization
reporting
practices.
efforts
address
issues
are
currently
underway,
cohesive
framework
ecology
still
emerging
stage.
Wildlife Biology,
Год журнала:
2024,
Номер
2024(3)
Опубликована: Март 13, 2024
Wildlife
populations
can
be
unmarked,
meaning
individuals
lack
distinguishing
features
for
individual
identification.
Populations
may
also
exhibit
non‐independent
movements,
move
together.
For
of
either
unmarked
or
individuals,
models
based
on
spatial
capture–recapture
(SCR)
approaches
used
to
estimate
abundance,
density,
and
other
parameters
critical
monitoring,
management,
conservation.
However,
when
are
both
non‐independent,
few
model
options
available.
One
approach
has
been
apply
not
address
the
non‐independence
despite
unquantified
impacts
bias,
precision,
ability
make
robust
ecological
inferences.
We
conducted
a
simulation
study
quantify
impact
performance
count
(SC)
partial
identity
(SPIM)
–
two
SCR‐based
modeling
fully
marked
independent
SCR
as
reference.
varied
levels
(aggregation
cohesion),
detection
probability,
number
covariates
resolve
identities
in
SPIM
estimation.
expected
abundance
estimates
increasingly
biased
precise
aggregation
cohesion
increased.
Results
showed
that
indeed
became
less
increasing
non‐independence,
but
importantly
suggested
only
could
reliably
applied
under
low
sufficient
SC
yielded
consistently
with
poor
precision.
was
across
combinations
cohesion,
expected.
therefore
advise
against
use
estimating
population
known
caution
narrow
conditions,
encourage
continued
investigations
into
sampling
design
methods
development
individuals.
Wildlife Biology,
Год журнала:
2024,
Номер
2024(6)
Опубликована: Июнь 18, 2024
The
continuous
growth
of
the
global
human
population
results
in
increased
use
and
change
landscapes,
with
infrastructures
like
transportation
or
energy
facilities
being
a
particular
risk
to
large
carnivores.
Environmental
impact
assessments
were
established
identify
probable
environmental
consequences
any
new
proposed
project,
find
ways
reduce
impacts,
provide
evidence
inform
decision
making
mitigation.
Portugal
has
wolf
approximately
300
individuals,
designated
as
an
endangered
species
full
legal
protection.
They
occupy
northern
mountainous
areas
country
which
also
been
focus
over
last
20
years.
Consequently,
dozens
monitoring
programs
have
evaluate
status,
appropriate
mitigation
compensation
measures.
We
reviewed
Portuguese
answer
four
key
questions.
Do
examine
adequate
biological
parameters
meet
objectives?
Is
study
design
suitable
for
measuring
impacts?
Are
data
collection
methods
effort
sufficient
stated
inference
statistical
analyses
lead
robust
conclusions?
Overall,
we
found
mismatch
between
aims
reported,
often
neither
aligns
existing
national
guidelines.
Despite
vast
expended
diversity
used,
analysis
makes
almost
exclusive
relative
indices
summary
statistics,
little
consideration
potential
biases
that
arise
through
(imperfect)
observational
process.
This
comparisons
impacts
across
space
time
difficult
is
therefore
unlikely
contribute
general
understanding
responses
infrastructure‐related
disturbance.
recommend
development
standardized
protocols
advocate
account
imperfect
detection
guarantee
accuracy,
reproducibility,
efficacy
programs.
Abstract
Estimating
the
size
of
animal
populations
plays
an
important
role
in
evidence‐based
conservation
and
management.
Some
methods
for
estimating
population
rely
on
animals
being
individually
identifiable.
Traditionally,
this
has
been
done
by
marking
physically
captured
animals,
but
increasingly,
with
distinctive
natural
markings
are
surveyed
noninvasively
using
cameras.
Animal
reidentification
from
photographs
is
usually
manually,
which
expensive,
laborious,
requires
considerable
skill.
An
alternative
to
develop
computer
vision
that
can
support
or
replace
manual
identification
task.
We
developed
automated
approach
deep
learning
identify
whether
a
pair
same
individual
not.
The
core
similarity
network
uses
paired
convolutional
neural
networks
triplet
loss
function
summarize
image
pairs
decide
they
individual.
Prior
main
matching
step,
two
additional
perform
segmentation,
cropping
object
within
image,
orientation
prediction,
deciding
side
was
photographed.
applied
four
species,
images
often
spanning
several
years:
systematic
surveys
bottlenose
dolphins
(
Tursiops
truncatus
,
2008–2019)
harbor
seals
Phoca
vitulina
2015–2019),
citizen
science
dataset
western
leopard
toads
Sclerophrys
pantherina
unknown
dates),
publicly
available
repository
humpback
whale
Megaptera
novaeangliae
dates).
For
these
our
best‐performing
models
were
able
different
individuals
95.8%,
94.6%,
88.2%,
83.8%
cases,
respectively.
found
functions
outperformed
binary
cross‐entropy
data
augmentation
curation
training
provided
small
consistent
improvements
performance.
These
results
demonstrate
potential
or,
more
likely,
facilitate
efforts.