Frontiers in Earth Science,
Journal Year:
2023,
Volume and Issue:
11
Published: Jan. 17, 2023
The
prediction
performance
of
several
machine
learning
models
for
regional
flash
flood
susceptibility
is
characterized
by
variability
and
regionality.
Four
typical
models,
including
multilayer
perceptron
(MLP),
logistic
regression
(LR),
support
vector
(SVM),
random
forest
(RF),
are
proposed
to
carry
out
modeling
in
order
investigate
the
rules
different
predicting
susceptibility.
original
data
14
environmental
factors,
such
as
elevation,
slope,
aspect,
gully
density,
highway
chosen
input
variables
MLP,
LR,
SVM,
RF
estimate
map
distribution
index
Longnan
County,
Jiangxi
Province,
China.
Finally,
various
evaluated
using
ROC
curve
features.
findings
show
that:
1)
Machine
can
accurately
assess
region’s
vulnerability
floods.
all
predict
very
well.
2)
MLP
(AUC=0.973,
MV=0.1017,
SD=0.2627)
model
has
best
susceptibility,
followed
SVM
(AUC=0.964,
MV=0.1090,
SD=0.2561)
(AUC=0.975,
MV=0.2041,
SD=0.1943)
LR
(AUC=0.882,
MV=0.2613,
SD=0.2913)
model.
3)
To
a
large
extent,
factors
population
density
influence
Geoscience Frontiers,
Journal Year:
2023,
Volume and Issue:
14(3), P. 101541 - 101541
Published: Jan. 18, 2023
In
this
study,
the
future
landslide
population
amount
risk
(LPAR)
is
assessed
based
on
integrated
machine
learning
models
(MLMs)
and
scenario
simulation
techniques
in
Shuicheng
County,
China.
Firstly,
multiple
MLMs
were
selected
hyperparameters
optimized,
generated
11
cross-integrated
to
select
best
model
calculate
susceptibility;
by
calculating
precipitation
for
different
extreme
recurrence
periods
combining
susceptibility
results
assess
hazard.
Using
town
as
basic
unit,
exposure
vulnerability
of
under
Shared
Socioeconomic
Pathways
(SSPs)
scenarios
each
assessed,
then
combined
with
hazard
estimate
LPAR
2050.
The
showed
that
optimized
random
forest
combination
strategy
had
comprehensive
performance
assessment.
distribution
classes
similar
susceptibility,
an
increase
precipitation,
low-hazard
area
high-hazard
decrease
shift
medium-hazard
very
classes.
high-risk
areas
populations
County
are
mainly
concentrated
three
southwestern
towns
high
vulnerability,
whereas
northern
Baohua
Qinglin
at
lowest
class.
increased
intensity
precipitation.
differs
significantly
among
SSPs
scenarios,
"fossil-fueled
development
(SSP5)"
highest
"regional
rivalry
(SSP3)"
scenario.
summary,
proposed
study
has
a
predictive
capability.
assessment
can
provide
theoretical
guidance
relevant
departments
cope
socioeconomic
challenges
make
corresponding
disaster
prevention
mitigation
plans
prevent
risks
from
developmental
perspective.
Land,
Journal Year:
2023,
Volume and Issue:
12(6), P. 1125 - 1125
Published: May 25, 2023
Selecting
samples
with
non-landslide
attributes
significantly
impacts
the
deep-learning
modeling
of
landslide
susceptibility
mapping.
This
study
presents
a
method
information
value
analysis
in
order
to
optimize
selection
negative
used
for
machine
learning.
Recurrent
neural
network
(RNN)
has
memory
function,
so
when
using
an
RNN
mapping
purposes,
input
landslide-influencing
factors
affects
resulting
quality
model.
The
calculates
factors,
determines
data
based
on
importance
any
specific
factor
determining
susceptibility,
and
improves
prediction
potential
recurrent
networks.
simple
unit
(SRU),
newly
proposed
variant
network,
is
characterized
by
possessing
faster
processing
speed
currently
less
application
history
networks
optimized
Xinhui
District,
Jiangmen
City,
Guangdong
Province,
China.
Four
models
were
constructed:
model
sample
selection,
SRU
model,
results
show
that
best
performance
terms
AUC
(0.9280),
followed
(0.9057),
(0.7277),
(0.6355).
In
addition,
several
objective
measures
accuracy
(0.8598),
recall
(0.8302),
F1
score
(0.8544),
Matthews
correlation
coefficient
(0.7206),
receiver
operating
characteristic
also
performs
best.
Therefore,
can
be
sensitivity
improve
model’s
performance;
second,
weaker
than
performance.
