Water Resources Research,
Journal Year:
2024,
Volume and Issue:
60(7)
Published: June 27, 2024
Abstract
This
study
presents
a
new
approach
to
understand
the
causes
of
groundwater
drought
events
with
interpretable
deep
learning
(DL)
models.
As
prerequisites,
accurate
long
short‐term
memory
(LSTM)
models
for
simulating
are
built
16
regions
representing
three
types
spatial
scales
in
southeastern
United
States,
and
standardized
index
is
applied
identify
233
events.
Two
interpretation
methods,
expected
gradients
(EG)
additive
decomposition
(AD),
adopted
decipher
DL‐captured
patterns
inner
workings
LSTM
networks.
The
EG
results
show
that:
(a)
temperature‐related
features
were
primary
drivers
large‐scale
droughts,
their
importance
increasing
from
56.1%
63.1%
as
approached
6
months
15
days.
Conversely,
precipitation‐related
found
be
dominant
factors
formation
small‐scale
catchments,
overall
ranging
59.8%
53.3%;
(b)
Seasonal
variations
inversely
related
between
large
small
scales,
being
more
significant
summer
larger
winter
catchments;
(c)
exhibited
an
“trigger
effect”
on
causing
studying
areas.
AD
method
unveiled
how
network
behaved
differently
retaining
discarding
information
when
emulating
different
droughts.
In
summary,
this
provides
perspective
highlights
potential
prospect
DL
enhancing
our
understanding
hydrological
processes.
Hydrology and earth system sciences,
Journal Year:
2023,
Volume and Issue:
27(9), P. 1865 - 1889
Published: May 15, 2023
Abstract.
Hybrid
hydroclimatic
forecasting
systems
employ
data-driven
(statistical
or
machine
learning)
methods
to
harness
and
integrate
a
broad
variety
of
predictions
from
dynamical,
physics-based
models
–
such
as
numerical
weather
prediction,
climate,
land,
hydrology,
Earth
system
into
final
prediction
product.
They
are
recognized
promising
way
enhancing
the
skill
meteorological
variables
events,
including
rainfall,
temperature,
streamflow,
floods,
droughts,
tropical
cyclones,
atmospheric
rivers.
now
receiving
growing
attention
due
advances
in
climate
at
subseasonal
decadal
scales,
better
appreciation
strengths
AI,
expanding
access
computational
resources
methods.
Such
attractive
because
they
may
avoid
need
run
computationally
expensive
offline
land
model,
can
minimize
effect
biases
that
exist
within
dynamical
outputs,
benefit
learning,
learn
large
datasets,
while
combining
different
sources
predictability
with
varying
time
horizons.
Here
we
review
recent
developments
hybrid
outline
key
challenges
opportunities
for
further
research.
These
include
obtaining
physically
explainable
results,
assimilating
human
influences
novel
data
sources,
integrating
new
ensemble
techniques
improve
predictive
skill,
creating
seamless
schemes
merge
short
long
lead
times,
incorporating
initial
surface
ocean/ice
conditions,
acknowledging
spatial
variability
landscape
forcing,
increasing
operational
uptake
schemes.
Hydrology and earth system sciences,
Journal Year:
2023,
Volume and Issue:
27(12), P. 2357 - 2373
Published: June 30, 2023
Abstract.
As
a
genre
of
physics-informed
machine
learning,
differentiable
process-based
hydrologic
models
(abbreviated
as
δ
or
delta
models)
with
regionalized
deep-network-based
parameterization
pipelines
were
recently
shown
to
provide
daily
streamflow
prediction
performance
closely
approaching
that
state-of-the-art
long
short-term
memory
(LSTM)
deep
networks.
Meanwhile,
full
suite
diagnostic
physical
variables
and
guaranteed
mass
conservation.
Here,
we
ran
experiments
test
(1)
their
ability
extrapolate
regions
far
from
gauges
(2)
make
credible
predictions
long-term
(decadal-scale)
change
trends.
We
evaluated
the
based
on
hydrograph
metrics
(Nash–Sutcliffe
model
efficiency
coefficient,
etc.)
predicted
decadal
For
in
ungauged
basins
(PUB;
randomly
sampled
representing
spatial
interpolation),
either
approached
surpassed
LSTM
metrics,
depending
meteorological
forcing
data
used.
They
presented
comparable
trend
for
annual
mean
flow
high
but
worse
trends
low
flow.
(PUR;
regional
holdout
extrapolation
highly
data-sparse
scenario),
advantages
became
prominent.
In
addition,
an
untrained
variable,
evapotranspiration,
retained
good
seasonality
even
extrapolated
cases.
The
models'
pipeline
produced
parameter
fields
maintain
remarkably
stable
patterns
data-scarce
scenarios,
which
explains
robustness.
Combined
interpretability
assimilate
multi-source
observations,
are
strong
candidates
global-scale
simulations
climate
impact
assessment.
Water Resources Research,
Journal Year:
2024,
Volume and Issue:
60(1)
Published: Jan. 1, 2024
Abstract
Recently,
rainfall‐runoff
simulations
in
small
headwater
basins
have
been
improved
by
methodological
advances
such
as
deep
neural
networks
(NNs)
and
hybrid
physics‐NN
models—particularly,
a
genre
called
differentiable
modeling
that
intermingles
NNs
with
physics
to
learn
relationships
between
variables.
However,
hydrologic
routing
simulations,
necessary
for
simulating
floods
stem
rivers
downstream
of
large
heterogeneous
basins,
had
not
yet
benefited
from
these
it
was
unclear
if
the
process
could
be
via
coupled
NNs.
We
present
novel
method
(
δ
MC‐Juniata‐hydroDL2)
mimics
classical
Muskingum‐Cunge
model
over
river
network
but
embeds
an
NN
infer
parameterizations
Manning's
roughness
n
)
channel
geometries
raw
reach‐scale
attributes
like
catchment
areas
sinuosity.
The
trained
solely
on
hydrographs.
Synthetic
experiments
show
while
geometry
parameter
unidentifiable,
can
identified
moderate
precision.
With
real‐world
data,
produced
more
accurate
long‐term
results
both
training
gage
untrained
inner
gages
larger
subbasins
(>2,000
km
2
than
either
machine
learning
assuming
homogeneity,
or
simply
using
sum
runoff
subbasins.
parameterization
short
periods
gave
high
performance
other
periods,
despite
significant
errors
inputs.
learned
pattern
consistent
literature
expectations,
demonstrating
framework's
potential
knowledge
discovery,
absolute
values
vary
depending
periods.
traditional
models
improve
national‐scale
flood
simulations.
Water Resources Research,
Journal Year:
2024,
Volume and Issue:
60(4)
Published: April 1, 2024
Abstract
While
deep
learning
(DL)
models
exhibit
superior
simulation
accuracy
over
traditional
distributed
hydrological
(DHMs),
their
main
limitations
lie
in
opacity
and
the
absence
of
underlying
physical
mechanisms.
The
pursuit
synergies
between
DL
DHMs
is
an
engaging
research
domain,
yet
a
definitive
roadmap
remains
elusive.
In
this
study,
novel
framework
that
seamlessly
integrates
process‐based
model
encoded
as
neural
network
(NN),
additional
NN
for
mapping
spatially
physically
meaningful
parameters
from
watershed
attributes,
NN‐based
replacement
representing
inadequately
understood
processes
developed.
Multi‐source
observations
are
used
training
data,
fully
differentiable,
enabling
fast
parameter
tuning
by
backpropagation.
A
hybrid
Amazon
Basin
(∼6
×
10
6
km
2
)
was
established
based
on
framework,
HydroPy,
global‐scale
DHM,
its
backbone.
Trained
simultaneously
with
streamflow
Gravity
Recovery
Climate
Experiment
satellite
yielded
median
Nash‐Sutcliffe
efficiencies
0.83
0.77
dynamic
simulations
total
water
storage,
respectively,
41%
35%
higher
than
those
original
HydroPy
model.
Replacing
Penman‒Monteith
formulation
produces
more
plausible
potential
evapotranspiration
(PET)
estimates,
unravels
spatial
pattern
PET
giant
basin.
parameterization
interpreted
to
identify
factors
controlling
variability
key
parameters.
Overall,
study
lays
out
feasible
technical
modeling
big
data
era.
Hydrology and earth system sciences,
Journal Year:
2024,
Volume and Issue:
28(3), P. 479 - 503
Published: Feb. 7, 2024
Abstract.
Deep
learning
(DL)
rainfall–runoff
models
outperform
conceptual,
process-based
in
a
range
of
applications.
However,
it
remains
unclear
whether
DL
can
produce
physically
plausible
projections
streamflow
under
climate
change.
We
investigate
this
question
through
sensitivity
analysis
modeled
responses
to
increases
temperature
and
potential
evapotranspiration
(PET),
with
other
meteorological
variables
left
unchanged.
Previous
research
has
shown
that
temperature-based
PET
methods
overestimate
evaporative
water
loss
warming
compared
energy
budget-based
methods.
therefore
assume
reliable
should
exhibit
less
when
forced
smaller,
energy-budget-based
PET.
conduct
assessment
using
three
models,
trained
tested
across
212
watersheds
the
Great
Lakes
basin.
The
include
Long
Short-Term
Memory
network
(LSTM),
mass-conserving
LSTM
(MC-LSTM),
novel
variant
MC-LSTM
also
respects
relationship
between
(MC-LSTM-PET).
After
validating
against
historical
actual
evapotranspiration,
we
force
all
scenarios
warming,
precipitation,
both
(Hamon)
(Priestley–Taylor)
PET,
compare
their
long-term
mean
daily
flow,
low
flows,
high
seasonal
timing.
explore
similar
national
fit
531
United
States
assess
how
inclusion
larger
more
diverse
set
basins
influences
signals
hydrological
response
warming.
main
results
study
are
as
follows:
substantially
estimation.
MC-LSTM-PET
matches
best
outperforms
estimating
evapotranspiration.
All
show
downward
shift
flows
but
median
shifts
considerably
(−17
%
−25
%)
than
(−6
−9
%).
model
exhibits
differences
different
forcings.
Conversely,
unrealistically
large
losses
Priestley–Taylor
(−20
%),
while
is
relatively
insensitive
method.
smaller
changes
timing
estimates
within
estimated
by
models.
Like
LSTM,
shows
(−25
stable
many
inputs
changed
better
aligns
for
flows.
Ultimately,
suggest
physical
considerations
regarding
architecture
input
may
be
necessary
promote
realism
deep-learning-based