Applied Sciences,
Journal Year:
2023,
Volume and Issue:
13(5), P. 3180 - 3180
Published: March 1, 2023
Rock
image
classification
is
a
fundamental
and
crucial
task
in
the
creation
of
geological
surveys.
Traditional
rock
methods
mainly
rely
on
manual
operation,
resulting
high
costs
unstable
accuracy.
While
existing
based
deep
learning
models
have
overcome
limitations
traditional
achieved
intelligent
classification,
they
still
suffer
from
low
accuracy
due
to
suboptimal
network
structures.
In
this
study,
model
EfficientNet
triplet
attention
mechanism
proposed
achieve
accurate
end-to-end
classification.
The
was
built
EfficientNet,
which
boasts
an
efficient
structure
thanks
NAS
technology
compound
scaling
method,
thus
achieving
for
Additionally,
introduced
address
shortcoming
feature
expression
enable
fully
capture
channel
spatial
information
images,
further
improving
During
training,
transfer
employed
by
loading
pre-trained
parameters
into
model,
accelerated
convergence
reduced
training
time.
results
show
that
with
92.6%
set
93.2%
Top-1
test
set,
outperforming
other
mainstream
demonstrating
strong
robustness
generalization
ability.
Frontiers in Earth Science,
Journal Year:
2023,
Volume and Issue:
10
Published: Jan. 16, 2023
Rock
image
classification
is
a
significant
part
of
geological
research.
Compared
with
traditional
methods,
rock
methods
based
on
deep
learning
models
have
the
great
advantage
in
terms
automatic
features
extraction.
However,
accuracies
existing
are
unsatisfied
due
to
weak
feature
extraction
ability
network
model.
In
this
study,
residual
neural
(ResNet)
model
transfer
method
proposed
establish
corresponding
for
seven
kinds
images.
ResNet34
introduces
structure
make
it
an
excellent
effect
field
classification,
which
extracts
high-quality
and
avoids
information
loss.
The
abstracts
from
shallow
features,
better
express
texture
case
fewer
To
improve
generalization
model,
total
3,82,536
images
were
generated
training
via
slicing
data
augmentation.
parameters
trained
Texture
Library
dataset
contains
47
types
reflect
characteristics
rocks
used
learning.
This
pre-trained
weight
loaded
when
dataset.
Then
fine-tuned
problem.
experimental
results
show
that
accuracy
without
reached
88.1%,
while
using
achieved
99.1%.
Aiming
at
engineering
investigation,
paper
studies
embedded
deployment
application
network.
transplanted
platform.
By
designing
system,
off-line
realized,
provides
new
solution
problem
survey.
can
automatically
classify
manually
extracting.
These
reduce
influence
subjective
factors
process
more
intelligent.
Swiss Journal of Geosciences,
Journal Year:
2024,
Volume and Issue:
117(1)
Published: May 7, 2024
Abstract
Deep
subsurface
exploration
is
important
for
mining,
oil
and
gas
industries,
as
well
in
the
assessment
of
geological
units
disposal
chemical
or
nuclear
waste,
viability
geothermal
energy
systems.
Typically,
detailed
examinations
formations
are
performed
on
cuttings
core
materials
extracted
during
drilling
campaigns,
geophysical
borehole
data,
which
provide
information
about
petrophysical
properties
rocks.
Depending
volume
rock
samples
analytical
program,
laboratory
analysis
diagnostics
can
be
very
time-consuming.
This
study
investigates
potential
utilizing
machine
learning,
specifically
convolutional
neural
networks
(CNN),
to
assess
lithology
mineral
content
solely
from
drill
images,
aiming
support
expedite
exploration.
The
paper
outlines
a
comprehensive
methodology,
encompassing
data
preprocessing,
learning
methods,
transfer
techniques.
outcome
reveals
remarkable
96.7%
accuracy
classification
segments
into
distinct
formation
classes.
Furthermore,
CNN
model
was
trained
evaluation
using
set
multidimensional
log
(silicate,
total
clay,
carbonate).
When
benchmarked
against
XRD
measurements
cores,
both
advanced
network
approach
developed
here
equally
good
performance.
work
demonstrates
that
deep
particularly
extracting
properties,
including
classification,
thus
offering
road
map
enhancing
performance
quality
image-based
cores.
Scientific Reports,
Journal Year:
2023,
Volume and Issue:
13(1)
Published: Aug. 24, 2023
Recent
advances
in
machine
learning
(ML)
have
transformed
the
landscape
of
energy
exploration,
including
hydrocarbon,
CO2
storage,
and
hydrogen.
However,
building
competent
ML
models
for
reservoir
characterization
necessitates
specific
in-depth
knowledge
order
to
fine-tune
achieve
best
predictions,
limiting
accessibility
geosciences.
To
mitigate
this
issue,
we
implemented
recently
emerged
automated
(AutoML)
approach
perform
an
algorithm
search
conducting
unconventional
with
a
more
optimized
accessible
workflow
than
traditional
approaches.
In
study,
over
1000
wells
from
Alberta's
Athabasca
Oil
Sands
were
analyzed
predict
various
key
properties
such
as
lithofacies,
porosity,
volume
shale,
bitumen
mass
percentage.
Our
proposed
consists
two
stages
AutoML
(1)
first
stage
focuses
on
predicting
shale
porosity
by
using
conventional
well
log
data,
(2)
second
combines
predicted
outputs
data
lithofacies
The
findings
show
that
out
ten
different
tested
(78%
accuracy),
(80.5%),
percentage
(67.3%),
classification
(98%),
distributed
random
forest,
gradient
boosting
models.
When
compared
manually
fine-tuned
algorithms,
AutoML-based
algorithms
provide
notable
improvement
property
higher
weighted
average
f1-scores
up
15-20%
problem
5-10%
adjusted-R2
score
regression
problems
blind
test
dataset,
it
is
achieved
only
after
~
400
s
training
testing
processes.
addition,
feature
ranking
extraction
technique,
there
good
agreement
domain
experts
regarding
most
significant
input
parameters
each
prediction.
Therefore,
evidence
has
proven
powerful
performing
advanced
petrophysical
analysis
minimal
time
human
intervention,
allowing
while
maintaining
model's
explainability.
Integration
subject
matter
could
advance
artificial
intelligence
technology
implementation
optimizing
data-driven
