bioRxiv (Cold Spring Harbor Laboratory),
Journal Year:
2024,
Volume and Issue:
unknown
Published: Aug. 26, 2024
ABSTRACT
This
study
addresses
the
challenges
in
pollinator
monitoring
by
proposing
an
effective
data
structure
for
automated
systems,
with
a
focus
on
use
of
machine
learning
to
handle
underrepresented
groups
small
datasets.
By
experimenting
grouping
top
three
pollinators
(bee,
butterfly,
hoverfly)
and
non-pollinators
datasets
fewer
than
300
samples,
research
aims
enhance
classification
detection
accuracy.
During
4-hour
filming
sessions,
181
images
insects
larger
1
cm
were
captured
classified
into
methods:
“Pollinator/
Non-pollinator”,
“Bee/Butterfly/Hoverfly/Ant”,
“Bumblebee/Honeybee/Butterfly/Hoverfly/
Ant”.
YOLO
V8
models
trained,
validated,
tested
these
based
different
class
methods.
The
found
that
“Pollinator/Non-pollinator”
YOLOv8
model
performed
best
across
all
metrics,
suggesting
it
is
more
reliable
categorizing
detecting
target
objects,
especially
smaller,
imbalanced
finding
aligns
trend
providing
training
opportunities
individual
classes
improves
Therefore,
using
broader
categorization
methods
can
reliability
accuracy
systems
when
insufficient.
Ecological Informatics,
Journal Year:
2023,
Volume and Issue:
77, P. 102278 - 102278
Published: Aug. 28, 2023
Cameras
and
computer
vision
are
revolutionising
the
study
of
insects,
creating
new
research
opportunities
within
agriculture,
epidemiology,
evolution,
ecology
monitoring
biodiversity.
However,
diversity
insects
close
resemblances
many
species
a
major
challenge
for
image-based
species-level
classification.
Here,
we
present
an
algorithm
to
hierarchically
classify
from
images,
leveraging
simple
taxonomy
(1)
specimens
across
multiple
taxonomic
ranks
simultaneously,
(2)
identify
lowest
rank
at
which
reliable
classification
can
be
reached.
Specifically,
propose
multitask
learning,
loss
function
incorporating
class
dependency
each
rank,
anomaly
detection
based
on
outlier
analysis
quantify
uncertainty.
First,
compile
dataset
41,731
images
combining
time-lapse
floral
scenes
with
Global
Biodiversity
Information
Facility
(GBIF).
Second,
adapt
state-of-the-art
convolutional
neural
networks,
ResNet
EfficientNet,
hierarchical
belonging
three
orders,
five
families
nine
species.
Third,
assess
model
generalization
11
unseen
by
trained
models.
is
used
predict
higher
were
not
in
training
set.
We
found
that
into
our
increased
accuracy
ranks.
As
expected,
correctly
classified
insect
ranks,
while
was
uncertain
lower
Anomaly
effectively
flag
novel
taxa
visually
distinct
data.
consistently
mistaken
similar
Above
all,
have
demonstrated
practical
approach
uncertainty
during
automated
situ
live
insects.
Our
method
versatile,
forming
valuable
step
towards
high-level
PLoS ONE,
Journal Year:
2024,
Volume and Issue:
19(4), P. e0295474 - e0295474
Published: April 3, 2024
Insect
monitoring
is
essential
to
design
effective
conservation
strategies,
which
are
indispensable
mitigate
worldwide
declines
and
biodiversity
loss.
For
this
purpose,
traditional
methods
widely
established
can
provide
data
with
a
high
taxonomic
resolution.
However,
processing
of
captured
insect
samples
often
time-consuming
expensive,
limits
the
number
potential
replicates.
Automated
facilitate
collection
at
higher
spatiotemporal
resolution
comparatively
lower
effort
cost.
Here,
we
present
Detect
DIY
(do-it-yourself)
camera
trap
for
non-invasive
automated
flower-visiting
insects,
based
on
low-cost
off-the-shelf
hardware
components
combined
open-source
software.
Custom
trained
deep
learning
models
detect
track
insects
landing
an
artificial
flower
platform
in
real
time
on-device
subsequently
classify
cropped
detections
local
computer.
Field
deployment
solar-powered
confirmed
its
resistance
temperatures
humidity,
enables
autonomous
during
whole
season.
On-device
detection
tracking
estimate
activity/abundance
after
metadata
post-processing.
Our
classification
model
achieved
top-1
accuracy
test
dataset
generalized
well
real-world
images.
The
software
highly
customizable
be
adapted
different
use
cases.
With
custom
models,
as
accessible
programming,
many
possible
applications
surpassing
our
proposed
method
realized.
Philosophical Transactions of the Royal Society B Biological Sciences,
Journal Year:
2024,
Volume and Issue:
379(1904)
Published: May 5, 2024
Automated
sensors
have
potential
to
standardize
and
expand
the
monitoring
of
insects
across
globe.
As
one
most
scalable
fastest
developing
sensor
technologies,
we
describe
a
framework
for
automated,
image-based
nocturnal
insects—from
development
field
deployment
workflows
data
processing
publishing.
Sensors
comprise
light
attract
insects,
camera
collecting
images
computer
scheduling,
storage
processing.
Metadata
is
important
sampling
schedules
that
balance
capture
relevant
ecological
information
against
power
limitations.
Large
volumes
from
automated
systems
necessitate
effective
We
vision
approaches
detection,
tracking
classification
including
models
built
existing
aggregations
labelled
insect
images.
Data
account
inherent
biases.
advocate
explicitly
correct
bias
in
species
occurrence
or
abundance
estimates
resulting
imperfect
detection
individuals
present
during
occasions.
propose
ten
priorities
towards
step-change
vital
task
face
rapid
biodiversity
loss
global
threats.
This
article
part
theme
issue
‘Towards
toolkit
monitoring’.
Journal of Pollination Ecology,
Journal Year:
2025,
Volume and Issue:
37, P. 1 - 21
Published: Jan. 10, 2025
Monitoring
plant-pollinator
interactions
is
crucial
for
understanding
the
factors
influencing
these
relationships
across
space
and
time.
Traditional
methods
in
pollination
ecology
are
resource-intensive,
while
time-lapse
photography
offers
potential
non-destructive
automated
complementary
techniques.
However,
accurate
identification
of
pollinators
at
finer
taxonomic
levels
(i.e.,
genus
or
species)
requires
high
enough
image
quality.
This
study
assessed
feasibility
using
a
smartphone
setup
to
capture
images
arthropods
visiting
flowers
evaluated
whether
offered
sufficient
resolution
arthropod
by
taxonomists.
Smartphones
were
positioned
above
target
from
various
plant
species
urban
green
areas
around
Leipzig
Halle,
Germany.
We
present
proportions
identifications
(instances)
different
(order,
family,
genus,
based
on
visible
features
as
interpreted
document
limitations
stem
(e.g.,
fixed
positioning
preventing
distinguishing
despite
resolution)
low
Recommendations
provided
address
challenges.
Our
results
indicate
that
89.81%
all
Hymenoptera
instances
identified
family
level,
84.56%
pollinator
only
25.35%
level.
less
able
identify
Dipterans
levels,
with
nearly
50%
not
identifiable
26.18%
15.19%
levels.
was
due
their
small
size
more
challenging
needed
wing
veins).
Advancing
technology,
along
accessibility,
affordability,
user-friendliness,
promising
option
coarse-level
monitoring.
Scientific Reports,
Journal Year:
2023,
Volume and Issue:
13(1)
Published: June 16, 2023
Increased
global
production
of
sorghum
has
the
potential
to
meet
many
demands
a
growing
human
population.
Developing
automation
technologies
for
field
scouting
is
crucial
long-term
and
low-cost
production.
Since
2013,
sugarcane
aphid
(SCA)
Melanaphis
sacchari
(Zehntner)
become
an
important
economic
pest
causing
significant
yield
loss
across
region
in
United
States.
Adequate
management
SCA
depends
on
costly
determine
presence
threshold
levels
spray
insecticides.
However,
with
impact
insecticides
natural
enemies,
there
urgent
need
develop
automated-detection
their
conservation.
Natural
enemies
play
role
populations.
These
insects,
primary
coccinellids,
prey
help
reduce
unnecessary
insecticide
applications.
Although
these
insects
regulate
populations,
detection
classification
time-consuming
inefficient
lower
value
crops
like
during
scouting.
Advanced
deep
learning
software
provides
means
perform
laborious
automatic
agricultural
tasks,
including
insects.
models
coccinellids
have
not
been
developed.
Therefore,
our
objective
was
train
machine
detect
commonly
found
classify
them
according
genera,
species,
subfamily
level.
We
trained
two-stage
object
model,
specifically,
Faster
Region-based
Convolutional
Neural
Network
(Faster
R-CNN)
Feature
Pyramid
(FPN)
also
one-stage
YOLO
(You
Only
Look
Once)
family
(YOLOv5
YOLOv7)
seven
(i.e.,
Coccinella
septempunctata,
Coleomegilla
maculata,
Cycloneda
sanguinea,
Harmonia
axyridis,
Hippodamia
convergens,
Olla
v-nigrum,
Scymninae).
used
images
extracted
from
iNaturalist
project
training
evaluation
R-CNN-FPN
YOLOv5
YOLOv7
models.
imagery
web
server
publish
citizen's
observations
pertaining
living
organisms.
Experimental
using
standard
metrics,
such
as
average
precision
(AP),
[email protected],
etc.,
shown
that
model
performs
best
coccinellid
[email protected]
high
97.3,
AP
74.6.
Our
research
contributes
automated
area
integrated
management,
making
it
easier
sorghum.
Agriculture,
Journal Year:
2023,
Volume and Issue:
13(12), P. 2287 - 2287
Published: Dec. 18, 2023
Using
neural
networks
on
low-power
mobile
systems
can
aid
in
controlling
pests
while
preserving
beneficial
species
for
crops.
However,
devices
require
simplified
networks,
which
may
lead
to
reduced
performance.
This
study
was
focused
developing
an
optimized
deep-learning
model
detecting
corn
pests.
We
propose
a
two-step
transfer
learning
approach
enhance
the
accuracy
of
two
versions
MobileNet
SSD
network.
Five
beetle
(Coleoptera),
including
four
harmful
crops
(belonging
genera
Anoxia,
Diabrotica,
Opatrum
and
Zabrus),
one
(Coccinella
sp.),
were
selected
preliminary
testing.
employed
datasets.
One
first
procedure
comprises
2605
images
with
general
dataset
classes
‘Beetle’
‘Ladybug’.
It
used
recalibrate
networks’
trainable
parameters
these
broader
classes.
Furthermore,
models
retrained
second
2648
five
species.
Performance
compared
baseline
terms
average
per
class
mean
precision
(mAP).
MobileNet-SSD-v2-Lite
achieved
mAP
0.8923,
ranking
but
close
highest
(0.908)
obtained
by
MobileNet-SSD-v1
outperforming
6.06%.
demonstrated
(0.9514)
Diabrotica
(0.8066).
Anoxia
it
reached
third-place
(0.9851),
top
value
0.9912.
Zabrus
position
(0.9053),
Coccinella
reliably
distinguished
from
all
other
species,
0.8939
zero
false
positives;
moreover,
no
pest
mistakenly
identified
as
Coccinella.
Analyzing
errors
revealed
good
overall
despite
size
training
set,
misclassification,
33
non-identifications,
7
double
identifications
1
positive
across
266
test
yielding
relative
error
rate
0.1579.
The
findings
validated
placed
place,
showing
high
potential
using
real-time
control
protecting
bioRxiv (Cold Spring Harbor Laboratory),
Journal Year:
2024,
Volume and Issue:
unknown
Published: April 15, 2024
Abstract
Arthropods,
including
insects,
represent
the
most
diverse
group
and
contribute
significantly
to
animal
biomass.
Automatic
monitoring
of
insects
other
arthropods
enables
quick
efficient
observation
management
ecologically
economically
important
targets
such
as
pollinators,
natural
enemies,
disease
vectors,
agricultural
pests.
The
integration
cameras
computer
vision
facilitates
innovative
approaches
for
agriculture,
ecology,
entomology,
evolution,
biodiversity.
However,
studying
their
interactions
with
flowers
vegetation
in
environments
remains
challenging,
even
automated
camera
monitoring.
This
paper
presents
a
comprehensive
methodology
monitor
abundance
diversity
wild
quantify
floral
cover
key
resource.
We
apply
methods
across
more
than
10
million
images
recorded
over
two
years
using
48
insect
traps
placed
three
main
habitat
types.
arthropods,
visits,
on
specific
mix
Sedum
plant
species
white,
yellow
red/pink
colored
flowers.
proposed
deep-learning
pipeline
estimates
flower
detects
classifies
arthropod
taxa
from
time-lapse
recordings.
serves
only
an
estimate
correlate
activity
flowering
plants.
Color
semantic
segmentation
DeepLabv3
are
combined
percent
different
colors.
Arthropod
detection
incorporates
motion-informed
enhanced
object
You-Only-Look-Once
(YOLO),
followed
by
filtering
stationary
objects
minimize
double
counting
non-moving
animals
erroneous
background
detections.
approach
has
been
demonstrated
decrease
incidence
false
positives,
since
occur
less
3%
captured
images.
final
step
involves
grouping
into
19
taxonomic
classes.
Seven
state-of-the-art
models
were
trained
validated,
achieving
F
1-scores
ranging
0.81
0.89
classification
arthropods.
Among
these,
selected
model,
EfficientNetB4,
achieved
80%
average
precision
randomly
samples
when
applied
complete
pipeline,
which
includes
detection,
filtering,
collected
2021.
As
expected
during
beginning
end
season,
reduced
correlates
noticeable
drop
method
offers
cost-effective