Content uploaded by Giacomo Fontanelli
Author content
All content in this area was uploaded by Giacomo Fontanelli on Jan 04, 2022
Content may be subject to copyright.
Please
cite
this
article
in
press
as:
Santi,
E.,
et
al.,
Application
of
artificial
neural
networks
for
the
soil
moisture
retrieval
from
active
and
passive
microwave
spaceborne
sensors.
Int.
J.
Appl.
Earth
Observ.
Geoinf.
(2015),
http://dx.doi.org/10.1016/j.jag.2015.08.002
ARTICLE IN PRESS
G Model
JAG-1133;
No.
of
Pages
13
International
Journal
of
Applied
Earth
Observation
and
Geoinformation
xxx
(2015)
xxx–xxx
Contents lists available at ScienceDirect
International
Journal
of
Applied
Earth
Observation
and
Geoinformation
journal homepage: www.elsevier.com/locate/jag
Application
of
artificial
neural
networks
for
the
soil
moisture
retrieval
from
active
and
passive
microwave
spaceborne
sensors
Emanuele
Santi∗,
Simonetta
Paloscia,
Simone
Pettinato,
Giacomo
Fontanelli
Institute
of
Applied
Physics—National
Research
Council
(IFAC-CNR),
Via
Madonna
del
Piano
10,
Firenze,
Italy
a
r
t
i
c
l
e
i
n
f
o
Article
history:
Received
24
March
2015
Received
in
revised
form
29
July
2015
Accepted
6
August
2015
Available
online
xxx
Keywords:
Artificial
neural
networks
Soil
moisture
content
Synthetic
Aperture
Radar
(SAR)
Scatterometer
Microwave
radiometers
a
b
s
t
r
a
c
t
Among
the
algorithms
used
for
the
retrieval
of
SMC
from
microwave
sensors
(both
active,
such
as
Syn-
thetic
Aperture
Radar-SAR,
and
passive,
radiometers),
the
artificial
neural
networks
(ANN)
represent
the
best
compromise
between
accuracy
and
computation
speed.
ANN
based
algorithms
have
been
developed
at
IFAC,
and
adapted
to
several
radar
and
radiometric
satellite
sensors,
in
order
to
generate
SMC
products
at
a
resolution
varying
from
hundreds
of
meters
to
tens
of
kilometers
according
to
the
spatial
scale
of
each
sensor.
These
algorithms,
which
are
based
on
the
ANN
techniques
for
inverting
theoretical
and
semi-empirical
models,
have
been
adapted
to
the
C-
to
Ka-
band
acquisitions
from
spaceborne
radiometers
(AMSR-
E/AMSR2),
SAR
(Envisat/ASAR,
Cosmo-SkyMed)
and
real
aperture
radar
(MetOP
ASCAT).
Large
datasets
of
co-located
satellite
acquisitions
and
direct
SMC
measurements
on
several
test
sites
worldwide
have
been
used
along
with
simulations
derived
from
forward
electromagnetic
models
for
setting
up,
training
and
validating
these
algorithms.
An
overall
quality
assessment
of
the
obtained
results
in
terms
of
accuracy
and
computational
cost
was
carried
out,
and
the
main
advantages
and
limitations
for
an
operational
use
of
these
algorithms
were
evaluated.
This
technique
allowed
the
retrieval
of
SMC
from
both
active
and
passive
satellite
systems,
with
accu-
racy
values
of
about
0.05
m3/m3of
SMC
or
better,
thus
making
these
applications
compliant
with
the
usual
accuracy
requirements
for
SMC
products
from
space.
©
2015
Elsevier
B.V.
All
rights
reserved.
1.
Introduction
The
amount
of
water
stored
in
the
soil
is
an
essential
variable
controlling
many
biophysical
processes
that
impact
water,
energy,
and
carbon
exchanges
at
the
land-atmosphere
interface.
In-situ
soil
moisture
(SMC)
measurements
are
labor
intensive
and
site-specific,
moreover,
frequent
and
spatially
distributed
soil
moisture
mea-
surements,
at
different
spatial
scales,
are
advisable
for
the
most
part
of
the
applications
related
to
the
environmental
disciplines,
such
as
climatology,
meteorology,
hydrology
and
agriculture.
The
possibility
of
observing
soil
moisture
and
its
temporal
evolution
from
space
is,
therefore,
regarded
as
being
extremely
attractive.
Among
the
instruments
operating
from
space
for
the
observa-
tion
of
the
Earth
surface,
the
sensors
operating
in
the
microwave
portion
of
the
spectrum
have
received
most
attention
because
this
frequency
range
has
the
unique
ability
to
return
information
on
media
(atmosphere,
vegetation,
soil)
that
are
opaque
to
the
much
shorter
visible/near-infrared
and
thermal
wavelength
and,
that
is
∗Corresponding
author.
Fax:
+39
055
5226434.
E-mail
address:
e.santi@ifac.cnr.it
(E.
Santi).
most
important,
because
microwave
scattering
and
emission
are
directly
related
to
the
water
content
of
the
observed
target.
In
par-
ticular,
remote
sensing
from
active
(Synthetic
Aperture
Radar—SAR
and
scatterometer)
and
passive
sensors
(radiometers)
have
demon-
strated
to
be
good
and
flexible
tools
for
observing
the
moisture
content
of
the
first
centimeter
layer
of
soil
and
for
detecting
its
spatial
and
temporal
variations
from
radar
(e.g.,
Barret
et
al.,
2009;
Notarnicola
et
al.,
2006,
2008;
Paloscia
et
al.,
2004;
Pathe
et
al.,
2009;
Pierdicca
et
al.,
2010;
Wagner
et
al.,
1999,
2007)
and
radio-
metric
microwave
sensors
(e.g.,
Jackson,
1993;
Jackson
et
al.,
2010;
Mladenova
et
al.,
2014;
Paloscia
et
al.,
2006).
These
two
types
of
sensors
(radiometers
and
SAR)
retrieve
infor-
mation
of
the
Earth
surface
at
different
scales.
SAR,
in
fact,
due
to
antenna
synthesis,
reaches
a
very
high
ground
resolution,
in
the
order
of
a
few
meters;
whereas,
radiometers
from
space
have
a
much
coarser
spatial
resolution,
in
the
order
of
tens
of
kilometers.
These
characteristics
allowed
the
use
of
these
sensors
for
differ-
ent
applications.
Fine
ground
resolutions
allowed
in
fact
a
more
detailed
investigation
of
the
surface
and
can
give
information
on
small
scale
phenomena,
useful
for
agricultural
applications
and
water
management
at
a
farm
or
basin
scale.
The
large
scale
of
space-
borne
radiometers
is
instead
very
useful
for
climatic
applications
http://dx.doi.org/10.1016/j.jag.2015.08.002
0303-2434/©
2015
Elsevier
B.V.
All
rights
reserved.
Please
cite
this
article
in
press
as:
Santi,
E.,
et
al.,
Application
of
artificial
neural
networks
for
the
soil
moisture
retrieval
from
active
and
passive
microwave
spaceborne
sensors.
Int.
J.
Appl.
Earth
Observ.
Geoinf.
(2015),
http://dx.doi.org/10.1016/j.jag.2015.08.002
ARTICLE IN PRESS
G Model
JAG-1133;
No.
of
Pages
13
2
E.
Santi
et
al.
/
International
Journal
of
Applied
Earth
Observation
and
Geoinformation
xxx
(2015)
xxx–xxx
and
for
detecting
the
trend
of
large
scale
phenomena,
as
global
changes,
e.g.,
deforestation
and
desertification
processes.
The
retrieval
of
soil
parameters
from
active
and/or
passive
microwave
measurements
is
nonetheless
not
trivial,
due
to
the
non-linearity
of
the
relationships
between
radar
or
radiometric
acquisitions
and
ground
parameters,
and
because,
in
general,
more
than
one
combination
of
soil
parameters
(soil
moisture,
rough-
ness,
vegetation
cover,
etc.)
has
the
same
electromagnetic
response.
Thus,
in
order
to
minimize
the
uncertainties
and
enhance
the
per-
formance
of
soil
parameter
retrieval
from
remote
sensing
data,
statistical
approaches
based
on
the
Bayes
theorem
and
learning
machines
are
widely
adopted
for
implementing
the
retrieval
algo-
rithms
(e.g.
Notarnicola,
2014;
Pasolli
et
al.,
2011;
Pierdicca
et
al.,
2014).
In
this
framework,
the
artificial
neural
networks
(ANNs)
rep-
resent
an
interesting
tool
for
implementing
accurate
and
flexible
retrieval
algorithms,
which
able
to
operate
with
radar
and
radio-
metric
satellite
measurements
and
to
easily
combine
information
coming
from
different
sources.
ANN
can
be
considered
a
statistical
minimum
variance
approach
for
addressing
the
retrieval
prob-
lem
and
they
can
be
trained
to
represent
arbitrary
input-output
relationships
(Hornik,
1989;
Linden
and
Kinderman,
1989
Linden
and
Kinderman,
1989).
In
the
training
phase,
training
patterns
are
sequentially
presented
to
the
network
and
the
interconnecting
weights
of
each
neuron
are
adjusted
according
to
a
learning
algo-
rithm.
The
trained
ANN
can
be
considered
as
a
type
of
non-linear
least
mean
square
interpolation
formula
for
the
discrete
set
of
data
points
in
the
training
set.
The
effectiveness
of
ANNs
in
solving
remote
sensing
problems
has
been
well
demonstrated,
since
they
can
easily
merge
data
com-
ing
from
different
sources
into
a
unique
retrieval
algorithm.
ANNs
have
been
successfully
applied
to
many
inverse
problems
in
the
remote
sensing
field,
and
in
particular
to
retrieve
soil
moisture
at
local
scale
from
SAR
(e.g.,
Del
Frate
et
al.,
2003;
Elshorbagy
and
Parasuraman
2008;
Paloscia
et
al.,
2008)
or
radiometric
(e.g.,
Santi
et
al.,
2012)
observations.
The
comparison
of
retrieval
algorithms
carried
out
in
Paloscia
et
al.
(2008)
demonstrated
that
ANNs,
with
respect
to
other
widely
adopted
statistical
approaches
based
on
Bayes
theorem
and
Nelder–Mead
minimization,
offer
the
best
com-
promise
between
retrieval
accuracy
and
computational
cost.
ANNs
in
some
cases
have
been
used
essentially
as
a
black
box,
without
further
effort
for
understanding
the
underlying
processes
and
the
physics
behind
them.
The
strategy
for
minimizing
these
problems
is
mainly
based
on
the
use
of
both
extensive
datasets
and
model
simulations
for
the
training
phase
of
ANN.
The
studies
pre-
sented
in
Paloscia
et
al.
(2010)
and
Santi
et
al.
(2013)
pointed
out
the
potential
of
the
ANN
technique
in
easily
and
effectively
ingest-
ing
information
extracted
from
different
sources
for
improving
the
retrieval
process,
such
as
NDVI
derived
from
optical
remote
sens-
ing
imagery
for
taking
into
account
the
presence
of
vegetation
on
the
ground.
In
Paloscia
et
al.
(2013)
nevertheless,
the
importance
of
a
robust
and
extensive
reference
dataset
for
the
training
was
con-
firmed,
in
order
to
obtain
a
retrieval
algorithm
able
to
work
at
large
and
global
scale
with
a
satisfactory
accuracy.
An
application
of
retrieval
algorithms
to
MetOp-A
ASCAT
backscatter
data
can
be
found
in
Gruber
et
al.
(2014)
where
the
ANN
techniques
are
compared
with
other
retrieval
algorithms,
in
their
ability
to
retrieve
soil
moisture
on
a
global
scale.
Correlation
and
triple
collocation
analysis
were
performed
using
in
situ
and
land
surface
model
data
as
a
reference,
pointing
out
again
the
effective-
ness
of
the
ANNs
compared
with
other
inversion
approaches.
In
this
paper,
an
overview
of
the
results
obtained
with
ANN
algorithms
applied
to
different
microwave
sensors
for
the
retrieval
of
soil
moisture
at
both
local
and
global
scales
is
presented.
We
have
summarized
and
homogenized
here
the
results
presented
in
other
papers
published
in
both
international
journals
and
confer-
ence
proceedings,
by
improving
and
tuning
the
algorithms
used
in
past
research,
and
validating
them
with
new
datasets.
In
some
cases,
more
accurate
results
have
been
obtained,
with
improved
accuracy
of
the
estimated
parameter.
In
this
work,
a
unique
proce-
dure
for
all
the
algorithms
was
thus
implemented,
by
choosing
the
same
strategy
for
training,
test,
and
validation.
In
Section
2
a
detailed
discussion
of
the
methods
and
in
partic-
ular
of
the
training
of
ANN
is
carried
out.
In
the
following
sections,
the
application
of
the
ANN
to
microwave
radiometers,
scatterom-
eters
and
SAR
sensors
is
presented
along
with
the
main
findings
of
the
methods.
Finally,
some
soil
moisture
maps
obtained
at
different
spatial
scales
by
using
the
data
from
these
sensors
are
shown
for
an
overall
validation
of
the
implemented
algorithms.
2.
Implementing
and
training
the
artificial
neural
networks
The
algorithms
presented
in
this
work
are
based
on
feed-
forward
multilayer
perceptron
(MLP)
ANNs,
with
a
certain
number
of
hidden
layers
of
neurons
between
the
input
and
the
output.
In
MLPs,
successive
layers
of
neurons
are
fully
interconnected,
with
trainable
connection
weights
that
control
the
strength
of
the
con-
nections.
The
ANN
training
was
based
on
the
back-propagation
learning
rule,
which
is
an
iterative
gradient
descent
algorithm
that
is
designed
to
minimize
the
mean
square
error
between
the
desired
target
vectors
and
the
actual
output
vectors.
It
should
be
noted
that
the
gradient-descent
method
sometimes
suffers
from
slow
conver-
gence,
due
to
the
presence
of
one
or
more
local
minima,
which
may
also
affect
the
final
result
of
the
training.
This
problem
can
be
solved
by
repeating
the
training
several
times,
with
a
resetting
of
the
initial
conditions
and
a
verification
that
each
training
process
led
to
the
same
convergence
results
in
terms
of
R2and
RMSE,
by
increasing
it
until
negligible
improvements
were
obtained.
In
order
to
define
the
optimal
ANN
architecture
in
terms
of
num-
ber
of
neurons
and
hidden
layers,
the
most
suitable
strategy
is
to
start
with
a
simple
ANN
architecture,
generally
with
one
hidden
layer
of
few
neurons.
This
ANN
is
trained
by
means
of
a
subset
of
the
available
data,
tested
on
the
rest
of
the
dataset
and
the
train-
ing
and
testing
errors
are
compared.
The
ANN
configuration
is
then
increased
by
adding
neurons
and
hidden
layers,
training
and
test-
ing
are
repeated
and
errors
compared
again,
until
a
further
increase
of
the
ANN
architecture
is
found
to
have
a
negligible
decrease
of
the
training
error
and
an
increase
in
the
test
error.
This
procedure
allows
defining
the
minimal
ANN
architecture
capable
of
provid-
ing
an
adequate
fit
for
the
training
data,
so
as
to
prevent
overfitting
problems.
Overfitting
is
related
to
the
oversizing
of
the
ANN,
and
may
cause
considerable
errors
when
testing
ANN
with
input
data
that
is
not
included
in
the
training
set
(Moody
et
al.,
1992;
Tetko
et
al.,
1995).
In
other
words,
the
ANN
is
able
to
reproduce
the
train-
ing
set
with
high
accuracy
but
fails
the
test
procedure.
Another
key
issue
for
defining
the
ANN
best
architecture
is
in
the
selection
of
the
most
appropriate
transfer
function:
in
gen-
eral
linear
transfer
functions
give
less
accurate
results
in
training
and
testing;
however,
they
are
less
prone
to
overfitting
and
are
more
robust
to
outliers,
i.e.,
input
data
out
the
range
of
the
input
parameters
included
in
the
training
set.
Logistic
sigmoid
(logsig)
and
tangental
sigmoid
(tansig)
transfer
functions
are
instead
char-
acterized
by
higher
accuracies
in
the
training
and
test;
however,
they
may
lead
to
large
errors
when
the
trained
ANN
is
applied
to
new
datasets.
Logsig
generates
outputs
between
0
and
1
as
the
neu-
ron’s
net
input
goes
from
negative
to
positive
infinity
and
describes
the
non-linearity,
g(a),
as
1/(1
+
e−a).
Alternatively,
multilayer
net-
works
can
use
the
tansig
function,
tanh(a)
=
(ea−
e−a)/(ea+
e−a).
Besides
these
problem,
however,
the
main
constraint
for
obtain-
ing
good
accuracies
with
the
ANN
approach,
as
it
has
been
demonstrated
in
Paloscia
et
al.
(2013)
is
represented
by
the
statis-
Please
cite
this
article
in
press
as:
Santi,
E.,
et
al.,
Application
of
artificial
neural
networks
for
the
soil
moisture
retrieval
from
active
and
passive
microwave
spaceborne
sensors.
Int.
J.
Appl.
Earth
Observ.
Geoinf.
(2015),
http://dx.doi.org/10.1016/j.jag.2015.08.002
ARTICLE IN PRESS
G Model
JAG-1133;
No.
of
Pages
13
E.
Santi
et
al.
/
International
Journal
of
Applied
Earth
Observation
and
Geoinformation
xxx
(2015)
xxx–xxx
3
Fig.
1.
Definition
of
training
and
validation
datasets
from
the
experimental
data
using
the
forward
e.m.
models.
tical
significance
of
the
training
set,
which
shall
be
representative
of
a
variety
of
surface
conditions
as
wide
as
possible,
in
order
to
make
the
algorithm
able
to
operate
at
a
global
scale.
Several
efforts
have
been
carried
out
in
the
last
years
to
create
large
databases
of
simultaneous
ground
SMC
measurements
and
satellite
acquisitions
(see
as
example
the
Climate
Change
Initiative
–
CCI
–
soil
moisture
activities
promoted
by
the
European
Space
Agency
http://www.
esa-cci.org/
or
the
SMEX
experiments
in
the
USA).
However,
the
datasets
derived
from
these
experimental
activities
are
not
large
enough
for
generating
the
“robust”
set
required
to
train
the
ANNs
for
global
monitoring
applications.
A
suitable
strategy
for
filling
in
any
gap
in
the
experimental
datasets
and
to
better
characterize
the
microwave
signal
depen-
dence
on
SMC
for
a
variety
of
surface
conditions
as
wide
as
possible
is
to
combine
the
experimental
measurements
with
simulated
data
obtained
from
electromagnetic
models
(e.m.).
The
consis-
tency
between
experimental
data
and
model
simulations
can
be
obtained
by
deriving
the
range
of
input
parameters
(namely
soil
moisture,
surface
roughness
and
vegetation
biomass)
from
the
experimental
measurements.
This
task
can
be
achieved
by
applying
the
Nelder–Mead
optimization
algorithm
(Nelder
and
Mead,
1965),
to
find
the
values
that
minimize
some
appropriate
cost
functions
between
microwave
measurements
and
corresponding
model
sim-
ulations.
After
defining
the
minimum
and
maximum
of
each
input
parameter,
the
input
vectors
for
the
e.m.
model
are
generated
by
using
a
pseudorandom
function,
rescaled
in
order
to
cover
the
range
of
each
parameter.
Thousands
of
inputs
vectors
for
running
the
model
simulations
can
be
generated
by
iterating
this
procedure,
thus
obtaining
datasets
of
surface
parameters
and
corresponding
simulated
microwave
data
for
training
and
testing
the
ANN.
The
flowchart
of
Fig.
1
represents
the
main
steps
for
generating
the
training
from
the
experimental
data.
The
same
procedure
allows
generating
the
independent
dataset
for
validating
the
ANN
after
training.
In
general,
the
available
data
are
divided
in
two
subsets
with
a
random
sampling,
the
first
subset
is
divided
again
in
60%
–
20%
–
20%
for
training,
test
and
validation
phases,
respectively
and
the
second
subset
is
reserved
for
an
independent
test
of
the
algorithm.
The
random
sampling
of
the
dataset
is
reiterated
5–6
times
and
the
training
is
repeated
each
time,
in
order
to
avoid
any
dependence
of
the
obtained
results
on
the
sampling
process.
Several
ANN
based
algorithms
have
been
developed
at
IFAC
following
this
strategy,
and
adapted
to
SAR/scatterometric
and
radiometric
satellite
sensors,
in
order
to
generate
SMC
products
at
both
local
and
global
scales
with
a
resolution
varying
from
hun-
dreds
of
meters
to
tens
of
kilometers.
3.
ANN
algorithm
for
spaceborne
multifrequency
radiometers
3.1.
HydroAlgo
algorithm
for
AMSR-E
and
AMSR2
The
“HydroAlgo”
algorithm
(Santi
et
al.,
2012)
applies
the
ANNs
for
estimating
the
surface
SMC
from
the
acquisitions
of
the
low
resolution
spaceborne
radiometers,
like
the
advanced
microwave
scanning
radiometer
for
the
Earth
observing
system
(AMSR-E)
(Lobl,
2001),
which
is
no
more
operating,
and
its
successor,
AMSR2
(Imaoka
et
al.,
2012).
The
main
characteristic
of
the
algorithm
is
the
exclusive
use
of
AMSR-E/2
data,
for
estimating
and
compen-
sating
the
effects
of
vegetation
conditions
and
surface
roughness
on
the
SMC
retrieval,
thanks
to
the
inclusion
of
data
acquired
at
the
higher
frequencies.
Brightness
temperature
data
collected
over
the
areas
of
interest
were
extracted
from
the
hierarchical
data
for-
mat
(HDF)
files
delivered
by
National
Snow
and
Ice
Data
Center
(NSIDC)
and
containing
the
calibrated
and
geocoded
acquisitions
of
AMSR-E
from
AQUA
satellite
(Level
2
data)
at
C,
X,
Ku
and
Ka
bands
in
both
polarizations
(H,
V).
A
check
of
data
for
possible
mis-
calibration
(Paloscia
et
al.,
2006)
and
for
the
presence
of
the
radio
frequency
interference
(RFI)
at
C
and
X
bands
was
carried
out,
by
using
a
simple
threshold
method
(Njoku
et
al.,
2005)
at
both
C
and
X
bands,
and
all
data
over
this
threshold
were
eliminated
from
the
dataset.
HydroAlgo
includes
a
disaggregation
procedure,
based
on
the
SFIM
filtering,
(Santi,
2010;
Parinussa
et
al.,
2013),
which
is
able
to
enhance
the
spatial
resolution
of
the
output
SMC
prod-
uct
up
to
the
nominal
sampling
of
AMSR-E
(∼10
×
10
km2),
thus,
minimizing
the
limit
of
these
sensors
for
the
monitoring
of
hetero-
geneous
landscapes,
due
to
their
coarse
resolution.
The
core
of
the
algorithm
is
composed
by
two
feedforward
multilayer
perceptron
(MLP)
ANNs,
trained
independently
for
the
ascending
and
descend-
ing
orbits
using
the
back-propagation
learning
rule.
Inputs
of
the
algorithm
are
the
brightness
temperatures
at
C
band
in
V
polariza-
tion,
the
polarization
indices
(PI)
at
10.65
GHz
and
18.7
GHz
(X
and
Ku
bands),
defined
as
PI
=
2*(TbV
−
TbH)/(TbV
+
TbH),
and
the
Tb
at
Ka
band
(36.5
GHz)
in
V
polarization.
The
C
band,
i.e.,
the
lowest
AMSR-E
frequency,
was
chosen
for
its
sensitivity
to
the
SMC,
which
is
slightly
influenced
by
sparse
vegetation.
PI
at
X
and
Ku
bands
were
taken
into
account
for
compensating
the
effect
of
vegetation
on
soil
emission
(Paloscia
and
Pampaloni,
1988),
and
for
flagging
out
the
densely
vegetated
targets,
where
SMC
cannot
be
retrieved
at
these
frequencies.
Tb
at
Ka
band,
in
V
polarization,
was
assumed
as
a
proxy
of
the
surface
physical
temperature,
to
account
for
the
effect
of
diurnal
and
seasonal
variations
of
the
surface
temperature
on
microwave
emission
(Njoku
and
Li,
1999).
HydroAlgo
was
developed
and
tested
using
a
set
of
several
thou-
sand
of
data
that
was
obtained
by
combining
the
experimental
data
collected
in
Mongolia
(Yang
et
al.,
2009)
and
the
Murrumbidgee
watershed,
in
Australia
(Smith
et
al.,
2012),
with
10.000
values
of
Tb
simulated
by
the
simple
implementation
of
the
radiative
transfer
theory
known
as
“tau-omega”
model
(Mo
et
al.,
1982).
These
experimental
datasets
were
kindly
provided
by
JAXA,
within
the
framework
of
the
JAXA
ADEOS-II/AMSR-E
and
GCOM/AMSR2
research
programs.
The
inputs
of
the
model
were
randomly
var-
ied
in
a
range
derived
from
experimental
data
as
follows:
SMC
(0.05–0.45
m3/m3),
surface
temperature
(275–320
K),
vegetation
optical
depth
()
at
C
band
ranging
between
0.16
and
1.1,
and
vegetation
single
scattering
albedo
(ω)
ranging
between
0.03
and
0.08.
At
the
higher
frequencies,
and
ω
values
were
derived
from
C
band
values
using
the
relationship
established
in
Santi
et
al.
(2012).
Model
simulations
were
iterated
for
each
input
vector
of
surface
parameters,
obtaining
a
training
set
of
10,000
brightness
temper-
atures
at
all
the
AMSR-E
bands
and
polarizations.
Fig.
2
shows
the
behavior
of
both,
experimental
and
simulated
Tb
at
C
and
Ka
bands
in
V.
pol.
as
a
function
of
volumetric
SMC
(m3/m3).
The
flowchart
of
the
whole
processing
chain
is
displayed
in
Fig.
3
along
with
a
detail
on
the
ANN
implemented
for
HydroAlgo.
The
ANN
independent
test
on
the
Australian
and
Mongolian
data,
which
were
not
used
for
the
training,
produced
the
diagram
of
Fig.
4,
where
the
soil
moisture
estimated
by
the
algorithm
(SMCest)
is
compared
with
the
soil
moisture
measured
on
ground
(SMCmeas).
The
corresponding
statistics
are:
determination
coefficient
R2=
0.8,
root
mean
square
error
RMSE
=
0.03
m3/m3,
and
BIAS
=
0.02
m3/m3.
Please
cite
this
article
in
press
as:
Santi,
E.,
et
al.,
Application
of
artificial
neural
networks
for
the
soil
moisture
retrieval
from
active
and
passive
microwave
spaceborne
sensors.
Int.
J.
Appl.
Earth
Observ.
Geoinf.
(2015),
http://dx.doi.org/10.1016/j.jag.2015.08.002
ARTICLE IN PRESS
G Model
JAG-1133;
No.
of
Pages
13
4
E.
Santi
et
al.
/
International
Journal
of
Applied
Earth
Observation
and
Geoinformation
xxx
(2015)
xxx–xxx
Fig.
2.
Experimental
and
simulated
Tb
values
at
C-
and
Ka
bands,
in
V-
polarization
as
a
function
of
SMC
derived
from
the
dataset
considered
for
training
the
ANN
algorithm.
Fig.
3.
HydroAlgo
flowchart
with
the
detail
of
the
ANN
architecture
This
results
was
obtained
thanks
to
the
contribution
of
the
higher
frequencies
in
accounting
for
the
vegetation
conditions,
trough
PI
at
X
and
Ku
bands,
and
in
estimating
the
surface
temperature,
trough
Tb
at
Ka
band,
thus,
reducing
the
uncertainties.
Fig.
4.
Test
of
the
ANN
algorithm
with
the
Australian
and
Mongolian
data
not
used
for
the
training
(Santi
et
al.,
2012).
A
further
and
new
validation
of
the
algorithm
was
carried
out
by
using
the
datasets
collected
on
the
four
agricultural
research
service
(ARS)
watershed
sites
in
the
US
from
2002
to
2009
(Jackson,
1993).
These
sites
represent
a
wide
range
of
ground
conditions
and
pre-
cipitation
regimes
(from
natural
to
agricultural
surfaces
and
from
desert
to
humid
regions)
and
provide
long-term
in-situ
data.
The
dimensions
of
each
watershed
are
compatible
with
the
AMSR-E
footprint,
close
to
10
×
10
km2at
all
frequencies
with
the
appli-
cation
of
the
SFIM
procedure.
These
sites
are
well-instrumented
and
characterized
and
have
been
the
focus
of
several
AMSR-E
validation
campaigns
(http://nsidc.org/data/amsr
validation/).
The
vegetation
characteristics
in
the
selected
watersheds
represent
typical
conditions
in
specific
climate
regions,
which
cover
a
range
of
semi-arid
to
humid.
All
watersheds
have
multiple
surface
soil
moisture
and
temperature
sensors
(5
cm
depth).
Characteristics
of
the
individual
watersheds
are:
Please
cite
this
article
in
press
as:
Santi,
E.,
et
al.,
Application
of
artificial
neural
networks
for
the
soil
moisture
retrieval
from
active
and
passive
microwave
spaceborne
sensors.
Int.
J.
Appl.
Earth
Observ.
Geoinf.
(2015),
http://dx.doi.org/10.1016/j.jag.2015.08.002
ARTICLE IN PRESS
G Model
JAG-1133;
No.
of
Pages
13
E.
Santi
et
al.
/
International
Journal
of
Applied
Earth
Observation
and
Geoinformation
xxx
(2015)
xxx–xxx
5
Fig.
5.
Comparison
between
the
HydroAlgo-derived
SMC
and
the
ARS
in
situ
data,
where
the
upper
[A]
and
the
lower
[B]
half
of
the
image
summarize
the
results
obtained
using
the
Ascending
(A)
and
Descending
(D)
overpasses.
Continuous
lines
represent
the
regression
equations.
Please
cite
this
article
in
press
as:
Santi,
E.,
et
al.,
Application
of
artificial
neural
networks
for
the
soil
moisture
retrieval
from
active
and
passive
microwave
spaceborne
sensors.
Int.
J.
Appl.
Earth
Observ.
Geoinf.
(2015),
http://dx.doi.org/10.1016/j.jag.2015.08.002
ARTICLE IN PRESS
G Model
JAG-1133;
No.
of
Pages
13
6
E.
Santi
et
al.
/
International
Journal
of
Applied
Earth
Observation
and
Geoinformation
xxx
(2015)
xxx–xxx
Table
1
Performances
of
HydroAlgo
soil
moisture
algorithms
terms
of
R2,
RMSE,
and
bias
for
both
ascending
and
descending
orbits.
Ascending
Descending
R2RMSE
(m3/m3)
Bias
(m3/m3)
R2RMSE
(m3/m3)
Bias
(m3/m3)
Little
Washita
0.374
0.046
−0.007
0.331
0.048
−0.009
Walnut
Gulch
0.297
0.019
−0.0003
0.255
0.02
0.0011
Little
River
0.279
0.043
0.017
0.358
0.039
0.0021
Reynolds
Creek 0.606 0.048 0.012 0.406 0.06
0.019
•Little
Washita
(OK):
20
soil
moisture
stations
in
a
region
dom-
inated
by
the
presence
of
rangeland
and
pasture
plus
some
agricultural
crops.
The
climate
is
sub-humid
with
about
750
mm
annual
precipitation.
The
size
of
the
area
is
610
km2.
•Little
River
(GA):
29
soil
moisture
stations
and
heavily
vegetated
(predominantly
forests,
croplands,
and
pasture).
The
climate
is
characterized
by
hot
humid
summers
and
short
mild
winters
(with
about
1200
mm
annual
precipitation).
The
size
of
the
area
is
334
km2.
•Walnut
Gulch
(AZ):
19
soil
moisture
stations,
with
brush
and
grass
cover
and
characterized
by
a
semiarid
climate
(with
about
324
mm
annual
precipitation).
The
size
of
the
area
is
148
km2.
•Reynolds
Creek
(ID):
19
soil
moisture
stations
in
a
rangeland
dominated
area,
with
snow
dominated
precipitation.
The
size
of
the
area
is
238
km2.
AMSR-E
data
collected
on
the
ARS
watersheds
from
2002
to
2009
were
used
for
this
comparison,
together
with
the
corresponding
ground
truth
data,
which
have
been
kindly
provided
by
Dr.
Tom
Jackson.
Due
to
the
diurnal
variations
in
surface
soil
moisture
that
occur
during
the
day
and
the
sensitivity
of
retrievals
to
assump-
tions
about
the
near
surface
temperature
and
moisture
profiles,
data
from
the
ascending
and
descending
orbits
were
analyzed
sep-
arately.
The
overall
performances
of
the
algorithm
are
summarized
in
Table
1,
in
which
the
correlation
coefficient
(R2),
RMSE
and
bias
obtained
for
the
different
sites
are
listed,
while,
the
SMC
estimated
for
each
test
area
using
HydroAlgo
is
plotted
against
the
SMC
mea-
sured
on
the
ground
in
Fig.
5.
The
overall
accuracy
(RMSE
≤0.06
m3/m3and
bias
<0.02
m3/m3)
is
compliant
with
the
AMSR-E
mission
SMC
accuracy
requirement.
The
spread
of
data
can
be
attributed
to
the
spatial
variability
within
Fig.
6.
ASCAT
measurements
in
forward
beam
as
a
function
of
the
measured
SMC
for
the
AMMA
(Spain)
and
MAQU
(Mongolia)
test
sites.
the
test
areas
and
the
rather
large
time
interval
covered
by
the
dataset
(8
years).
4.
ANN
algorithm
for
real
and
synthetic
aperture
radars
4.1.
ANN
algorithm
for
ASCAT/SCA
An
ANN
algorithm
for
SMC
retrieval
from
the
MetOp
ASCAT
scatterometer
(Figa-Salda˜
na
et
al.,
2002)
was
developed
in
the
framework
of
the
Round
Robin
exercise,
belonging
to
the
Climate
Change
Initiative
(CCI)
soil
moisture
activities
promoted
by
the
European
Space
Agency
(ESA).
Data
from
150
test
sites
of
the
Fig.
7.
(a)
Architecture
of
the
ANN
after
optimization.
(b)
ANN
independent
validation:
SMC
estimated
by
ANN
as
a
function
of
the
target
SMC
Please
cite
this
article
in
press
as:
Santi,
E.,
et
al.,
Application
of
artificial
neural
networks
for
the
soil
moisture
retrieval
from
active
and
passive
microwave
spaceborne
sensors.
Int.
J.
Appl.
Earth
Observ.
Geoinf.
(2015),
http://dx.doi.org/10.1016/j.jag.2015.08.002
ARTICLE IN PRESS
G Model
JAG-1133;
No.
of
Pages
13
E.
Santi
et
al.
/
International
Journal
of
Applied
Earth
Observation
and
Geoinformation
xxx
(2015)
xxx–xxx
7
Fig.
8.
Volumetric
SMC
(m3/m3)
estimated
by
the
ANN
algorithm
as
a
function
of
the
target
SMC
for
(a)
ASCAT
configuration
(VV
polarization
only)
and
(b)
SCA
configuration
(VV
+
VH).
International
Soil
Moisture
Network
(ISMN—Dorigo
et
al.,
2011),
containing
co-located
satellite
acquisition
and
ground
measure-
ments
of
SMC
were
provided
to
the
Round
Robin
participants
for
calibration
and
validation
of
the
algorithms.
The
calibration
dataset
contained
ASCAT
daily
acquisitions
over
75
test
sites
for
the
five
considered
years,
the
measurements
of
top
layer
SMC
from
the
ground
stations
and
the
Global
Land
Data
Assimilation
System
(GLDAS)
simulated
10
cm
soil
surface
temperature,
precipitation,
and
snow
water
equivalent,
while
the
validation
dataset
contained
only
the
ASCAT
acquisitions
on
the
remaining
75
test
sites.
A
preliminary
analysis
of
the
relationships
between
ASCAT
mea-
surements
and
SMC
measured
on
ground
confirmed
the
sensitivity
of
C-
band
radar
measurements
to
the
latter
parameter,
point-
ing
out
however,
that
the
geographic
locations
and
therefore,
the
different
climatic
and
vegetation
conditions
of
each
test
site
signifi-
cantly
affected
these
relationships.
This
is
evident
from
the
example
of
Fig.
6,
representing
the
ASCAT
backscattering
coefficient
(◦,
in
dB)
measured
in
forward
beam
as
a
function
of
the
SMC
(m3/m3)
measured
for
two
test
sites.
The
first
(ANMA
network)
is
in
Europe
(Spain)
and
is
characterized
by
seasonal
vegetation
cycles
while
the
second
(MAQU
network)
is
in
the
semi-arid
Mongolian
plateau.
The
diagram
well
demonstrates
that
the
presence
of
vegetation
affects
the
sensitivity
to
SMC,
as
indicated
by
the
two
different
slopes.
A
correlation
of
the
measured
signal
to
SMC
is
nonetheless
confirmed
in
both
cases,
as
it
is
pointed
out
by
the
determination
coefficients,
R2=
0.24
and
R2=
0.30,
respectively.
This
geographical
dependence
has
been
accounted
for
in
the
ANN
inversion
algorithm
by
adding
the
geographical
position
as
ancillary
input.
The
ANN
optimization
process
for
matching
the
ASCAT
RRDP
(Round
Robin
Data
Package)
dataset
resulted
in
an
architecture
with
three
hidden
layers
of
11
+
11
+
10
neurons
was
selected
(Fig.
7a).
A
subset
of
the
RRDP
calibration
dataset,
composed
by
about
the
25%
of
the
data
available
was
considered
for
generating
the
subset
for
training,
testing,
and
validating
the
ANN
(60%
–
20%
–
20%,
randomly
sampled),
and
the
remaining
75%
of
data
was
con-
sidered
for
the
independent
validation
of
the
algorithm.
ANN
inputs
were
the
ASCAT
backscatter
acquired
at
the
three
beams
and
the
corresponding
incidence
and
azimuth
angles.
Output
of
the
ANN
was
the
retrieved
SMC
(Fig.
7b).
The
overall
performance
of
the
algorithm
is
summarized
in
Table
2.
It
should
be
remarked
that,
in
this
case,
the
training
was
only
based
on
the
experimental
data,
since
the
CCI
data
con-
Table
2
Main
statistical
parameters
of
the
retrieval
obtained
with
the
ANN
algorithm.
RMSE
R2Slope
Intercept
0.042
0.67
0.67
0.072
sidered
for
this
application
were
representative
of
the
different
climatic
areas
of
the
world.
Moreover,
ancillary
information
on
soil
roughness
and
vegetation
cover
was
unavailable
for
most
part
of
the
test
sites,
making
challenging
the
definition
of
the
input
vec-
tors
for
running
the
model
simulations.
However,
by
adding
the
geographic
position
to
the
ANN
inputs
we
noticed
a
substantial
increase
of
the
accuracy
(R2from
0.49
to
0.67
and
RMSE
from
0.052
to
0.042
m3/m3),
with
respect
to
considering
backscattering
and
observation
angles
only.
Model
simulations
have
been
instead
considered
for
under-
standing
the
potential
contribution
of
a
VH
polarized
channel
in
improving
the
SMC
retrieval
accuracy
from
scatterometric
acquisi-
tions.
The
EUMETSAT
polar
system
of
second
generation
is
indeed
under
study
to
replace
the
existing
satellite
system
and
provide
continuity
of
observations
in
the
2020
timeframe.
The
inclusion
of
a
cross-
polarized
(VH)
channel
on
the
mid-beam
antenna
in
the
EPS-SG
SCA
instrument,
heir
of
ASCAT,
is
at
present
under
study.
The
contribution
of
VH
for
correcting
the
effects
of
vegetation
cover
and
surface
roughness
on
soil
moisture
retrieval
was
then
evalu-
ated
through
the
implementation
of
another
ANN
algorithm.
Basing
on
the
ANN
architecture
implemented
for
ASCAT,
which
included
the
backscatter
acquired
at
the
three
beams
and
the
corresponding
incidence
and
azimuth
angles
at
VV-polarized
data
only,
a
further
input
was
added
for
the
VH
channel
that
will
be
acquired
in
mid-
beam
by
SCA
sensor.
In
this
case,
training
and
test
of
the
algorithm
were
carried
out
using
only
data
generated
by
a
simple
imple-
mentation
of
RTT,
based
on
the
coupling
of
Oh
model
(Oh
et
al.,
1992)
and
the
Vegetation
Water
Cloud
(WCM)
model
(proposed
by
Attema
and
Ulaby,
1978),
accounting
for
the
different
obser-
vation
geometries
of
the
3
SCA
beams.
The
latter
model
considers
the
volume
scattering
of
vegetation
and
the
attenuation
effect
on
the
surface
scattering
of
soil
under
vegetation.
By
coupling
these
models,
the
backscattering
sensitivity
to
soil
moisture
for
a
wide
range
of
vegetation
and
roughness
conditions
can
accurately
be
evaluated.
Please
cite
this
article
in
press
as:
Santi,
E.,
et
al.,
Application
of
artificial
neural
networks
for
the
soil
moisture
retrieval
from
active
and
passive
microwave
spaceborne
sensors.
Int.
J.
Appl.
Earth
Observ.
Geoinf.
(2015),
http://dx.doi.org/10.1016/j.jag.2015.08.002
ARTICLE IN PRESS
G Model
JAG-1133;
No.
of
Pages
13
8
E.
Santi
et
al.
/
International
Journal
of
Applied
Earth
Observation
and
Geoinformation
xxx
(2015)
xxx–xxx
Fig.
9.
ANN
SMC
algorithm
for
SMC
retrieval
from
C-
band
SAR.
The
blocks
in
the
circle
are
performed
offline
during
the
training
phase
of
the
ANNs.
A
dataset
of
about
60,000
backscattering
coefficients
was
gen-
erated
with
this
model
considering
volumetric
SMC
values
ranging
between
0.05
m3/m3and
0.45
m3/m3,
Plant
Water
Content
(PWC)
between
0
and
5
kg/m2and
roughness
(expressed
as
standard
devi-
ation
of
the
surface
heights,
Hstd)
between
0.5
and
2
cm.
Training
and
test
were
carried
out
by
dividing
the
available
data
in
two
subsets
of
30,000
samples
each.
The
first
subset
was
again
divided
randomly
in
60%
–
20%
–
20%,
for
training,
test,
and
validation
phase,
respectively,
while,
the
second
subset
was
retained
for
an
inde-
pendent
validation.
The
results
of
this
independent
validation
are
represented
in
Fig.
8a,
for
the
ASCAT
architecture
(VV
data
only)
and
Fig.
8b
for
the
SCA
architecture
(VV
+
VH).
This
comparison
demon-
strated
that
the
inclusion
of
the
VH
polarized
channel
is
able
to
improve
the
retrieval
accuracy,
with
an
increase
of
the
determina-
tion
coefficient
R2from
0.59
to
0.67
and
a
decrease
of
RMSE
from
0.048
to
0.042
m3/m3,
as
displayed
in
the
figures.
This
result
is
quite
interesting
since
the
simulations
confirmed
that
the
addition
of
a
cross
polarized
acquisition
to
the
future
SCA
sensor
may
help
in
improving
the
SMC
retrieval.
4.2.
ANN
Algorithm
for
C-
and
X-
band
SAR
Following
the
approach
proposed
in
Paloscia
et
al.
(2013)
an
ANN
algorithm
able
to
estimate
the
SMC
from
SAR
acquisitions
at
C-
and
X-
bands
in
different
acquisition
geometries
and
polariza-
tions
has
been
implemented,
tested
and
validated.
The
algorithm
flowchart
is
represented
in
Fig.
9.
Each
image
was
processed
by
applying
a
multilook
process,
in
order
to
average
the
intensity
in
range
and
azimuth
direction,
by
using
the
appropriated
window
size
according
the
considered
SAR
sensor.
The
radiometric
calibra-
tion
was
performed
considering
the
local
incidence
angle
based
on
the
orbital
parameters
and
the
DEM,
that
combined
with
satel-
lite
orbital
parameters
allowed
also
the
identification
of
layover
and
shadow
effects.
Layover
and
shadowing
were
almost
negligi-
ble
in
flat
areas.
The
geocoded
images
have
a
pixel
size
between
10
×
10
m2and
30
×
30
m2,
and
were
co-registered
in
order
to
be
comparable
‘pixel
by
pixel’
to
each
other
and
with
other
informa-
tion,
such
as
the
local
incidence
angle
(LIA).
The
core
of
the
algorithm
is
composed
by
6
+
6
ANNs,
trained
for
working
with
backscattering
in
VV
or
HH
polarization
with
and
without
the
ancillary
information
on
vegetation,
represented
by
co-located
NDVI
from
optical
sensor,
and
VV
+
VH
or
HH
+
HV
combinations,
at
C-
and
X-
band
respectively.
In
order
to
gener-
ate
an
algorithm
able
to
work
on
a
large/global
scale,
the
dataset
implemented
for
the
ANN
training
was
obtained
by
combining
experimental
satellite
measurements
of
backscattering
coefficients
(◦),
corresponding
ground
parameters,
and
data
simulated
using
e.m.
forward
models.
The
backscattering
of
the
bare
rough
surfaces
was
obtained,
in
this
case,
by
using
the
Advanced
Integral
Equa-
tion
model
(AIEM,
by
Wu
and
Chen,
2004)
for
co-polarized
SAR
Please
cite
this
article
in
press
as:
Santi,
E.,
et
al.,
Application
of
artificial
neural
networks
for
the
soil
moisture
retrieval
from
active
and
passive
microwave
spaceborne
sensors.
Int.
J.
Appl.
Earth
Observ.
Geoinf.
(2015),
http://dx.doi.org/10.1016/j.jag.2015.08.002
ARTICLE IN PRESS
G Model
JAG-1133;
No.
of
Pages
13
E.
Santi
et
al.
/
International
Journal
of
Applied
Earth
Observation
and
Geoinformation
xxx
(2015)
xxx–xxx
9
Fig.
10.
SMC
maps
of:
(a)
Sesto
test
area,
30
April
2011,
obtained
using
the
HH
polarization
only.
(b)
Sesto
test
area,
19
September
2011,
obtained
using
the
HH
polarization
and
NDVI.
(c)
Scrivia
test
area,
11
April
2010,
obtained
using
the
HH-HV
polarizations.
(d)
Scrivia
test
area,
28
September
2010,
obtained
using
the
HH-HV
polarizations.
signal,
and
the
Oh
model
(Oh
et
al.,
1992)
for
deriving
the
cross-
polarized
backscatter
from
the
co-polarized
simulated
by
AIEM.
The
contribution
of
light
vegetation
was
accounted
for
by
using
the
WCM,
deriving
the
information
on
vegetation
water
content
(VWC)
from
the
Normalized
Difference
Vegetation
Index
(NDVI),
mea-
sured
from
the
available
optical
sensors
(e.g.,
Landsat
and
MODIS),
trough
semi-empirical
relationships.
Minimum
and
maximum
values
of
the
soil
parameters
mea-
sured
during
the
experimental
campaigns
(SMC,
Hstd
and
correlation
length,
Lc)
were
considered
in
order
to
define
the
range
of
variability
of
each
soil
parameter.
Using
a
pseudoran-
dom
function
drawn
from
the
standard
uniform
distribution
on
the
open
interval
(0,1),
rescaled
in
order
to
cover
the
range
of
each
soil
parameter,
we
generated
input
vectors
for
the
coupled
AIEM
+
WCM
models,
in
order
to
simulate
the
backscattering
at
VV,
HH
and
HV/VH
polarizations.
More
in
detail,
the
input
parameters
were:
a)
Incidence
angle,
=
random
between
20
and
50◦.
b)
SMC
range
between
0.05
m3/m3and
0.45
m3/m3.
c)
Hstd
=
random
between
1
and
3
cm
for
C
band
and
0.5–2
cm
for
X
band.
d)
Lc
=
random
between
4
and
8
cm.
e)
Dielectric
constant
derived
from
random
values
of
SMC
between
5%
and
45%
using
the
Dobson
Model
(Dobson
et
al.,
1985).
f)
NDVI
=
random
between
0
and
0.8.
Since,
the
relationship
between
Hstd
and
Lc
is
rather
compli-
cated
and
reliable
measurements
of
the
Lc
parameter
are
barely
available,
we
decided
to
keep
these
two
quantities
independent,
associating
one
random
variable
with
each
of
these.
This
procedure
was
then
iterated
10,000
times,
thus
obtaining
a
set
of
backscatter-
ing
coefficients
for
each
input
vector
of
the
soil
parameters.
The
consistency
between
the
experimental
data
and
the
model
simula-
tion
was
verified
before
proceeding
to
the
training
phase.
The
ANN
training
was
carried
out
by
considering
the
simulated
◦at
the
var-
ious
polarizations
and
the
incidence
angle
as
input
of
the
ANN,
and
the
soil
parameters
as
outputs.
After
training,
the
ANNs
were
tested
on
a
different
dataset
that
was
obtained
by
re-iterating
the
model
simulations
as
described
above.
The
use
of
a
pseudorandom
function
prevented
a
correlation
between
these
two
datasets:
this
fact
was
particularly
important
in
order
to
evaluate
the
capabilities
of
ANN
to
generalize
the
training
phase
and
to
prevent
the
overfitting
problem.
Incorrect
sizing
of
the
Please
cite
this
article
in
press
as:
Santi,
E.,
et
al.,
Application
of
artificial
neural
networks
for
the
soil
moisture
retrieval
from
active
and
passive
microwave
spaceborne
sensors.
Int.
J.
Appl.
Earth
Observ.
Geoinf.
(2015),
http://dx.doi.org/10.1016/j.jag.2015.08.002
ARTICLE IN PRESS
G Model
JAG-1133;
No.
of
Pages
13
10
E.
Santi
et
al.
/
International
Journal
of
Applied
Earth
Observation
and
Geoinformation
xxx
(2015)
xxx–xxx
Table
3
Statistical
parameters
(R2,
RMSE
and
bias)
of
the
SMC
retrieval
by
using
the
ANN
algorithm
at
C
and
X
bands.
C
band
X
band
ANN
R2RMSE
Bias
R2RMSE
Bias
VV
0.49
0.067
-0.014
0.33
0.056
0.012
VV
+
NDVI
0.65
0.052
0.009
0.61
0.046
0.003
HH
0.85
0.026
−0.003
0.57
0.040
0.005
HH
+
NDVI 0.88 0.023 0.010
0.77
0.034
−0.003
VV
+
VH
0.59
0.016
−0.009
0.71
0.041
0.008
HH
+
HV
0.87
0.028
−0.009
0.79
0.032
0.008
ANN
or
inadequate
training
could
cause
the
overfitting:
the
ANN
returns
outputs
outside
the
training
range
(outliers)
when
tested
with
input
data
that
are
not
included
in
the
training
set.
The
algorithm
was
finally
validated
by
considering
the
SAR
images
and
corresponding
ground
truth
in
several
test
areas
in
Italy
and
Australia,
for
a
total
of
about
700
field-averaged
values
of
◦
at
various
polarizations
and
corresponding
SMC
measurements
at
C
band
and
about
600
values
at
X
band.
The
results
of
the
overall
validation,
obtained
by
comparing
all
the
SMC
values
retrieved
by
the
algorithm
with
the
corresponding
ground
truth
pointed
out
that
the
worst
case
corresponds
to
the
VV
polarization
without
ancillary
information
of
vegetation
(co-
located
NDVI).
In
this
case
the
algorithm
is
able
to
provide
only
a
rough
estimate
of
SMC,
since
it
cannot
account
for
all
the
possi-
ble
surface
and
vegetation
conditions
that
affect
the
backscattering
sensitivity
to
SMC,
when
only
one
polarization
is
available.
The
HH
polarization
appears
more
related
to
the
SMC
than
the
VV,
and
the
NDVI
contributes
at
increasing
the
retrieval
accuracy.
The
combination
of
co
and
cross
polarizations,
either
VV
+
VH
or
HH
+
HV,
offers
the
best
performance
with
a
noticeable
accuracy
improvement
with
respect
to
the
other
combinations.
The
obtained
results
in
terms
of
R2,
RMSE
and
bias
are
summarized
in
Table
3
at
both
C
and
X
bands.
As
expected
the
best
results
are
achieved
at
C
band,
which
is
more
sensitive
to
SMC
and
less
influenced
by
the
vegetation
than
X
band.
At
the
latter
frequency
instead,
the
vegetation
effect
is
dominant,
although,
some
sensitivity
to
SMC
is
detectable
at
least
for
bare
and
scarcely
vegetated
surfaces.
5.
Generation
of
SMC
maps
at
local
to
global
scale
Looking
at
the
operational
application
of
these
algorithms,
the
generation
of
SMC
maps
in
real
or
near
real
time
at
different
resolutions,
depending
on
the
input
sensor
characteristics,
takes
advantage
of
the
reduced
computational
cost
of
the
ANN
tech-
niques
with
respect
to
other
statistical
methods.
In
Paloscia
et
al.
(2013)
the
ANN
retrieval
algorithm
was
demonstrated
to
be
able
to
processing
200,000
pixels/sec,
which
correspond
to
about
80
s
for
generating
a
SMC
map
at
25
×
25
m2resolution
from
an
input
SAR
image
of
100
×
100
km2.
Although,
the
SMC
maps
cannot
be
considered
a
real
valida-
tion
of
the
retrieval
algorithms,
since
adequate
ground
truth
for
comparing
the
algorithm
outputs
at
large
and
global
scale
is
barely
available,
these
maps
represent
an
effective
tool
for
verifying
qual-
itatively
the
validity
of
the
training
process.
Maps
characterized
by
extreme
SMC
variations
from
a
pixel
to
an
adjacent
one,
by
large
percentages
of
outliers,
and
maps
in
which
SMC
spatial
patterns
are
not
detectable
indicate
that
the
training
was
not
successful
or
that
the
training
set
was
not
well
representative
of
the
observed
target,
although
the
SMC
estimated
in
the
control
points
was
close
to
the
ground
truth.
In
these
cases,
the
ANN
should
be
retrained
and
it
should
be
verified
if
the
training
set
was
representative
of
the
whole
range
of
the
input
microwave
data
and
output
SMC.
As
an
example
of
the
operational
capabilities
of
these
algorithms,
four
SMC
maps
at
basin
scale
are
represented
in
Fig.
10.
These
maps
were
Fig.
11.
ANN
estimated
SMC
is
plotted
vs.
the
corresponding
ground
measured
SMC
(%).
The
regression
line
is
SMC
retrieved
=
0.78*SMC
measured
+
0.06,
with
R2=
0.79.
derived
from
COSMO-SKYMED
SAR
images,
which
were
collected
on
two
test
sites,
in
central
(Sesto)
and
northern
(Scrivia)
Italy.
Map
dimensions
are
40
×
40
km2,
and
white
and
blue
colors
represent
masking
for
urban
areas
and
water
bodies
respectively.
The
diagram
of
Fig.
11
was
inserted
here
as
a
partial
valida-
tion
of
the
previous
maps.
Estimated
values
of
SMC
at
X
band
in
HH,
VV,
and
HV
polarizations,
are
represented
as
a
function
of
SMC
measured
on
ground
on
selected
fields.
Since
all
the
COSMO-
SKYMED
configurations
have
been
taken
into
account,
we
can
note
that
the
resulting
determination
coefficient
and
slope
are
rather
high
(R2=
0.8,
s
=
0.78),
thus,
confirming
a
good
correlation
between
measured
and
estimated
soil
moisture.
The
RMSE
is
0.04
m3/m3,
pointing
also
out
that
the
role
played
at
this
frequency
by
surface
roughness
and
vegetation
is
significant.
Moving
from
basin
to
global
scale,
Fig.
12
represents
six
SMC
maps
of
a
portion
of
the
world
(Europe
and
Africa)
obtained
as
weekly
average
of
AMSR-E
acquisitions
collected
in
different
sea-
sons
between
December
2009
and
September
2010.
Masking
for
snow
cover
areas
(white),
dense
vegetation
(dark
green),
desert
(red),
and
water
(blue)
were
carried
out
in
the
maps.
A
quantitative
point-to-point
validation
of
each
image
is
obviously
problematic;
however,
the
algorithm
appears
to
be
able
to
reproduce
the
aver-
age
moisture
conditions
of
the
different
climatic
areas
of
the
Earth,
following
the
seasonal
variations
of
SMC.
This
is
particularly
evi-
dent
in
the
areas
covered
by
light
vegetation,
e.g.,
Sahel
region
and
Spain,
where
in
summer
there
is
a
general
decrease
of
SMC
(orange-
red
colors).
An
opposite
behavior
is
instead
visible
in
the
areas
in
the
austral
hemisphere
(e.g.,
South
Africa).
The
reliability
of
these
maps
can
be
also
evaluated
looking
at
the
presence
of
outliers,
i.e.,
estimated
SMC
values
outside
the
training
range,
which
indicate
Please
cite
this
article
in
press
as:
Santi,
E.,
et
al.,
Application
of
artificial
neural
networks
for
the
soil
moisture
retrieval
from
active
and
passive
microwave
spaceborne
sensors.
Int.
J.
Appl.
Earth
Observ.
Geoinf.
(2015),
http://dx.doi.org/10.1016/j.jag.2015.08.002
ARTICLE IN PRESS
G Model
JAG-1133;
No.
of
Pages
13
E.
Santi
et
al.
/
International
Journal
of
Applied
Earth
Observation
and
Geoinformation
xxx
(2015)
xxx–xxx
11
Fig.
12.
SMC
maps
of
Europe
and
Africa
obtained
as
weekly
average
of
AMSR-E
acquisitions
collected
in
different
seasons
between
December
2009
and
September
2010.
(For
interpretation
of
the
references
to
colour
in
this
figure
text,
the
reader
is
referred
to
the
web
version
of
this
article.)
an
inappropriate
training.
For
all
the
displayed
maps
indeed
the
percentage
of
outliers
was
really
negligible
(<0.1%).
6.
Conclusions
An
application
of
artificial
neural
networks
(ANN)
techniques
for
retrieving
the
SMC
from
active
and
passive
microwave
satel-
lite
acquisitions
was
presented
here.
ANN
based
algorithms
were
developed
and
tuned
for
working
with
active
and
passive
microwave
acquisitions
at
several
frequencies
varying
from
C
(5
GHz)
to
Ka
(37
GHz)
bands.
Large
datasets
of
co-located
satel-
lite
acquisitions
and
direct
SMC
measurements
on
several
test
sites
situated
worldwide
were
used
along
with
simulations
from
for-
ward
electromagnetic
models
for
setting
up,
training
and
validating
these
algorithms.
In
detail,
ANN
based
algorithms
for
the
SMC
retrieval
developed
and
validated
with
AMSR-E,
ASCAT,
ENVISAT
and
Cosmo-SkyMed
data
were
presented.
The
overview
of
the
retrieval
algorithms
presented
here,
demonstrated
that
ANN
are
a
very
powerful
tool
for
estimating
soil
moisture
at
both
local
and
global
scale,
provided
they
have
been
trained
with
consistent
datasets
made
up
by
both
experimental
and
theoretical
data.
The
performed
test
confirmed
the
flexibility
Please
cite
this
article
in
press
as:
Santi,
E.,
et
al.,
Application
of
artificial
neural
networks
for
the
soil
moisture
retrieval
from
active
and
passive
microwave
spaceborne
sensors.
Int.
J.
Appl.
Earth
Observ.
Geoinf.
(2015),
http://dx.doi.org/10.1016/j.jag.2015.08.002
ARTICLE IN PRESS
G Model
JAG-1133;
No.
of
Pages
13
12
E.
Santi
et
al.
/
International
Journal
of
Applied
Earth
Observation
and
Geoinformation
xxx
(2015)
xxx–xxx
of
this
method
and
the
possibility
of
using
it
for
both
active
and
passive
sensors
with
high
accuracy
and
computational
speed.
Test
of
the
algorithms
returned
accuracy
values
of
about
0.05
m3/m3
of
SMC
or
better,
making
these
applications
compliant
with
the
usual
accuracy
requirements
for
SMC
products
from
space.
More-
over,
the
possibility
of
repeating
the
training
with
new
datasets
enables
easily
the
improvement
of
the
retrieval
accuracy,
mak-
ing
this
technique
adaptable
to
new
data
and
sensors
and
flexible.
Another
advantage
of
these
algorithms
is
in
the
capability
of
merg-
ing
data
coming
from
different
sources,
as
other
sensors
or
ancillary
information,
into
a
unique
retrieval
approach.
It
was
the
case
of
the
algorithm
implemented
for
C-
and
X-
band
SAR
that
takes
advan-
tage
of
the
NDVI
information
from
optical
sensors
(Landsat/Modis)
when
available
for
improving
the
SMC
retrieval
accuracy.
The
main
constraint
for
accurate
retrievals
is
due
to
the
training
process:
the
retrieval
error
may
be
large
if
the
ANN
is
tested
with
data
not
correctly
represented
in
the
training.
Large
datasets
are
therefore
needed
for
properly
training
the
ANN,
in
order
to
cover
the
whole
range
of
the
microwave
data
and
corresponding
mois-
ture
condition
of
the
observed
surface.
It
should
be
noted
that
there
is
not
a
unique
way
for
defining
the
training
set.
Some
a
priori
knowledge
and
the
support
of
model
simulations
help
in
setting
the
range
of
each
surface
parameter,
in
order
to
make
the
training
set
as
representative
as
possible
of
the
observed
surface.
Testing
and
validation
on
independent
datasets
(i.e.,
scarcely
related
to
the
data
considered
for
training)
may
indicate
if
the
training
has
been
achieved
properly.
In
particular,
the
use
of
electromagnetic
mod-
els
for
generating
large
training
dataset
is
one
of
the
best
methods
for
avoiding
the
danger
of
‘black
box’
algorithm
and
to
make
sure
that
the
results
are
based
on
physical
assumptions.
Since
the
train-
ing
is
performed
off-line,
before
starting
the
data
processing,
the
computational
speed
of
ANN
is
not
hampered
by
this
procedure.
Next
steps
should
be
the
gathering
of
extended
datasets
for
a
more
accurate
validation
of
the
results,
especially
on
a
large
scale
where
the
possibility
of
validation
is
limited
by
the
scarcity
of
ground
truth
data.
Moreover,
this
method
should
be
applied
to
new
sensors,
whose
more
advanced
characteristics
will
allow
the
retrieval
of
more
accurate
results
in
terms
of
RMSE,
bias
and
other
statistical
parameters.
An
important
further
step,
even
in
view
of
the
new
SMAP
mission,
is
the
integration
of
active
and
passive
data
in
a
single
algorithm,
capable
of
retrieving
soil
moisture
at
different
scales
with
improved
accuracy
with
respect
to
that
one
of
separate
algorithms.
Acknowledgments
This
research
work
was
partially
supported
by
the
JAXA
ADEOS-
II/AMSR-E
and
GCOM/AMSR2
research
programs,
by
the
ESA/ESTEC
contract
no.
4000103855/11/NL/MP/fk
on
GMES
Sentinel-1
Soil
Moisture
algorithm
development,
by
the
EUMETSAT
contract
no.
EUM/C0/14/4600001368/JF
on
the
use
of
SCA
cross-pol
for
the
Soil
Moisture
retrieval
and
by
the
ASI
Hydrocosmo
proposal
no.
1720
on
the
retrieval
and
monitoring
of
Land
Hydrological
parameters
for
Risk
and
Water
Resources
Management.
The
authors
wish
to
thank
JAXA
for
providing
the
Mongolian
and
Australian
datasets,
Tom
Jackson
and
Iva
Mladenova
for
providing
the
ground
data
from
the
ARS
watersheds
and
Wolfgang
Wagner
for
the
involvement
in
the
ESA
CCI
Round
Robin
exercise.
References
Attema,
E.P.W.,
Ulaby,
F.T.,
1978.
Vegetation
modeled
as
a
water
cloud.
Radio
Sci.
13,
357–364.
Barret,
B.W.,
Dwyer,
E.,
Whelan,
P.,
2009.
Soil
moisture
retrieval
from
active
spaceborne
microwave
observations:
an
evaluation
of
current
techniques.
Remote
Sen.
1,
210–242.
Del
Frate,
F.,
Ferrazzoli,
P.,
Schiavon,
G.,
2003.
Retrieving
soil
moisture
and
agricultural
variables
by
microwave
radiometry
using
neural
networks.
Remote
Sens.
Environ.
84
(2),
174–183,
http://dx.doi.org/10.1016/S0034-
4257(02)00105-0.
Dobson,
M.C.,
Ulaby,
F.T.,
Hallikainen,
M.T.,
El-Rayes,
M.A.,
1985.
Microwave
dielectric
behavior
of
wet
soil—part
II:
dielectric
mixing
models.
IEEE
Trans.
Geosci.
Remote
Sens.
23,
35–46.
Dorigo,
W.A.,
Wagner,
W.,
Hohensinn,
R.,
Hahn,
S.,
Paulik,
C.,
Xaver,
A.,
Gruber,
A.,
Drusch,
M.,
Mecklenburg,
S.,
van
Oevelen,
P.,
Robock,
A.,
Jackson,
T.,
2011.
The
international
soil
moisture
network:
a
data
hosting
facility
for
global
in
situ
soil
moisture
measurements.
Hydrol.
Earth
Syst.
Sci.
15,
1675–1698.
Elshorbagy,
Parasuraman,
K.,
2008.
On
the
relevance
of
using
artificial
neural
networks
for
estimating
soil
moisture
content.
J.
Hydrol.
362,
1–18
http://
www.sciencedirect.com/science/article/pii/S0022169408004204
Figa-Salda˜
na,
J.,
Wilson,
J.J.W.,
Attema,
E.,
Gelsthorpe,
R.,
Drinkwater,
M.R.,
Stoffelen,
A.,
2002.
The
Advanced
scatterometer
(ASCAT)
on
the
meteorological
operational
(MetOp)
platform:
a
follow
on
for
European
wind
scatterometers.
Can.
J.
Remote
Sens.
28
(3),
404–412.
Gruber,
A.,
Paloscia,
E.,
Santi,
C.,
Notarnicola,
L.,
Pasolli,
T.,
Smolander,
J.,
Pulliainen,
H.,
Mittelbach,
W.,
Wagner,
W.,
2014.
Performance
inter-comparison
of
soil
moisture
retrieval
models
for
the
metop-A
ASCAT
instrument.
Proceedings
of
Geoscience
and
Remote
Sensing
Symposium
IGARSS
2014
Proceedings,
2455–2458,
ISBN:
978-1-4799-5774-3.
Hornik,
K.,
1989.
Multilayer
feed
forward
network
are
universal
approximators.
Neural
Net.
2
(5),
359–366.
Imaoka,
K.,
Takashi,
M.,
Misako,
K.,
Marehito,
K.,
Norimasa,
I.,
Keizo,
N.,
2012.
Status
of
AMSR2
instrument
on
GCOM-W1,
earth
observing
missions
and
sensors:
development,
implementation,
and
characterization
II.
Proc.
SPIE,
852815,
http://dx.doi.org/10.1117/12.977774.
Jackson,
T.J.,
1993.
Measuring
surface
soil
moisture
using
passive
microwave
remote
sensing.
Hydrol.
Process.
7
(2),
139–152,
Copyright©1993
John
Wiley
&
Sons,
Ltd.
Jackson,
T.J.,
Cosh,
M.H.,
Bindlish,
R.,
Starks,
P.J.,
Bosch,
D.D.,
Seyfried,
M.S.,
Goodrich,
D.C.,
Moran,
M.S.,
2010.
Validation
of
advanced
microwave
scanning
radiometer
soil
moisture
products.
IEEE
Trans.
Geosci.
Remote
Sens.
48
(12),
4256–4272.
Linden,
A.,
Kinderman,
J.,
1989.
Inversion
of
multi-layer
nets.
Proc.
Int.
Joint
Conf.
Neural
Net.
2,
425–443.
Lobl,
E.,
2001.
Joint
advanced
microwave
scanning
radiometer
(AMSR)
science
team
meeting.
Earth
Observer
13
(3),
3–9.
Mladenova,
I.E.,
Jackson,
T.J.,
Njoku,
E.,
Bindlish,
R.,
Chan,
S.,
Cosh,
M.H.,
Holmes,
T.R.H.,
de
Jeu,
R.A.M.,
Jones,
L.,
Kimball,
J.,
Paloscia,
S.,
Santi,
E.,
2014.
Remote
monitoring
of
soil
moisture
using
passive
microwave-based
techniques—theoretical
basis
and
overview
of
selected
algorithms
for
AMSR-E.
Remote
Sens.
Environ.
144,
197–213.
Mo,
T.,
Choudhury,
B.J.,
Schmugge,
T.J.,
Wang,
J.R.,
Jackson,
T.J.,
1982.
A
model
for
microwave
emission
from
vegetation
covered
fields.
J.
Geophys.
Res.
87,
11229–11237.
Moody,
J.E.,
Hanson,
S.J.,
Lippmann,
R.P.,
1992.
The
effective
number
of
parameters:
an
analysis
of
generalization
and
regularization
in
nonlinear
learning
systems.
Adv.
Neural
Inf.
Process.
Syst.
4,
847–854.
Nelder,
J.A.,
Mead,
R.,
1965.
A
simplex
method
for
function
minimization.
Comput.
J.
7,
308–313.
Njoku,
E.G.,
Li,
L.,
1999.
Retrieval
of
land
surface
parameters
using
passive
microwave
measurements
at
6–18
GHz.
Trans.
Geosci.
Remote
Sens.
37
(1),
79–93.
Notarnicola,
C.,
Angiulli,
M.,
Posa,
F.,
2006.
Use
of
radar
and
optical
remotely
sensed
data
for
soil
moisture
retrieval
on
vegetated
areas.
IEEE
Trans.
Geosci.
Remote
Sens.
44,
925–935.
Notarnicola,
C.,
Angiulli,
M.,
Posa,
F.,
2008.
Soil
moisture
retrieval
from
remotely
sensed
data:
neural
network
approach
versus
bayesian
method.
IEEE
Trans.
Geosci.
Remote
Sens.
46
(2),
547–557.
Notarnicola,
C.,
2014.
A
Bayesian
change
detection
approach
for
retrieval
of
soil
moisture
variations
under
different
roughness
conditions.
IEEE
Geosci.
Remote
Sens.
Lett.
11
(2),
414–418.
Oh,
Y.,
Sarabandi,
K.,
Ulaby,
F.T.,
1992.
An
empirical
model
and
an
inversion
technique
for
radar
scattering
from
bare
surfaces.
IEEE
Trans.
Geosci.
Remote
Sens.
30,
370–381.
Paloscia,
S.,
Pampaloni,
P.,
1988.
Microwave
polarization
index
for
monitoring
vegetation
growth.
IEEE
Trans.
Geosci.
Remote
Sens.
26
(5),
617–621.
Paloscia,
S.,
Macelloni,
G.,
Pampaloni,
P.,
Santi,
E.,
2004.
The
contribution
of
multi-temporal
SAR
data
in
assessing
hydrological
parameters.
IEEE
Trans.
Geosci.
Remote
Sens.
Lett.
1,
201–205.
Paloscia,
S.,
Macelloni,
G.,
Santi,
E.,
2006.
Soil
moisture
estimates
from
AMSR-E
brightness
temperatures
by
using
a
dual-frequency
algorithm.
IEEE
Trans.Geosci.
Remote
Sens.
44,
3135–3144.
Paloscia,
S.,
Pampaloni,
P.,
Pettinato,
S.,
Santi,
E.,
2008.
A
aomparison
of
algorithms
for
retrieving
soil
moisture
from
ENVISAT/ASAR
images.
IEEE
Trans.
Geosci.
Remote
Sens.
46
(10),
3274–3284.
Paloscia,
S.,
Pampaloni,
P.,
Pettinato,
S.,
Santi,
E.,
2010.
Generation
of
soil
moisture
maps
from
ENVISAT/ASAR
images
in
mountainous
areas:
a
case
study.
Int.
J.
Remote
Sens.
31
(9–10),
2265–2276.
Paloscia,
S.,
Pettinato,
S.,
Santi,
E.,
Notarnicola,
C.,
Pasolli,
L.,
Reppucci,
A.,
2013.
Soil
moisture
mapping
using
Sentinel-1
images:
algorithm
and
preliminary
validation.
Remote
Sens.
Environ.
134,
234–248,
http://dx.doi.org/10.1016/j.
rse.2013.02.027.
Please
cite
this
article
in
press
as:
Santi,
E.,
et
al.,
Application
of
artificial
neural
networks
for
the
soil
moisture
retrieval
from
active
and
passive
microwave
spaceborne
sensors.
Int.
J.
Appl.
Earth
Observ.
Geoinf.
(2015),
http://dx.doi.org/10.1016/j.jag.2015.08.002
ARTICLE IN PRESS
G Model
JAG-1133;
No.
of
Pages
13
E.
Santi
et
al.
/
International
Journal
of
Applied
Earth
Observation
and
Geoinformation
xxx
(2015)
xxx–xxx
13
Parinussa,
R.M.,
Yilmaz,
M.T.,
Anderson,
M.C.,
Hain,
C.R.,
de
Jeu,
R.A.M.,
2013.
An
intercomparison
of
remotely
sensed
soil
moisture
products
at
various
spatial
scales
over
the
Iberian
Peninsula.
Hydrol.
Process.
28
(18),
4865–4876,
http://
dx.doi.org/10.1002/hyp.9975.
Pasolli,
L.,
Notarnicola,
C.,
Bruzzone,
L.,
Bertoldi,
G.,
Della
Chiesa,
S.,
Hell,
V.,
Niedrist,
G.,
Tappeiner,
U.,
Zebisch,
M.,
Del
Frate,
F.,
Vaglio
Laurin,
G.,
2011.
Estimation
of
soil
moisture
in
an
alpine
catchment
with
RADARSAT2
images.
Appl.
Environ.
Soil
Sci.,
12,
http://dx.doi.org/10.1155/2011/175473,
Hindawi
Publishing
Corporation.
Article
ID
175473.
Pathe,
C.,
Wagner,
W.,
Sabel,
D.,
Doubkova,
M.,
Basara,
J.B.,
2009.
Using
ENVISAT
ASAR
Global
Mode
data
for
surface
soil
moisture
retrieval
over
Oklahoma,
USA.
IEEE
Trans.
Geosci.
Remote
Sens.
47,
468–480.
Pierdicca,
N.,
Pulvirenti,
L.,
Pace,
G.,
2014.
A
prototype
software
package
to
retrieve
soil
moisture
from
Sentinel-1
sata
by
using
a
bayesian
multitemporal
algorithm.
IEEE
J.
Selected
Topics
Appl.
Earth
Observ.
Remote
Sens.
7
(1),
153–163.
Pierdicca,
N.,
Pulvirenti,
L.,
Bignami,
C.,
2010.
Soil
moisture
estimation
over
vegetated
terrains
using
multitemporal
remote
sensing
data.
Remote
Sens.
Environ.
114,
440–448.
Santi,
E.,
2010.
An
application
of
SFIM
technique
to
enhance
the
spatial
resolution
of
microwave
radiometers.
Int.
J.
Remote
Sens.
31
(9–10),
2419–2428.
Santi,
E.,
Pettinato,
S.,
Paloscia,
S.,
Pampaloni,
P.,
Macelloni,
G.,
Brogioni,
M.,
2012.
An
algorithm
for
generating
soil
moisture
and
snow
depth
maps
from
microwave
spaceborne
radiometers:
HydroAlgo.
Hydrol.
Earth
Syst.
Sci.
16,
3659–3676,
http://dx.doi.org/10.5194/hess-16-3659-2012.
Santi,
E.,
Paloscia,
S.,
Pettinato,
S.,
Notarnicola,
C.,
Pasolli,
L.,
Pistocchi,
A.,
2013.
Comparison
between
SAR
soil
moisture
estimates
and
hydrological
model
simulations
over
the
Scrivia
test
site.
Remote
Sens.
5,
4961–4976,
http://dx.
doi.org/10.3390/rs5104961.
Smith,
A.B.,
Walker,
J.P.,
Western,
A.W.,
Young,
R.I.,
Ellett,
K.M.,
Pipunic,
R.C.,
Grayson,
R.B.,
Siriwardena,
L.,
Chiew,
F.H.S.,
Richter,
H.,
2012.
The
Murrumbidgee
soil
moisture
monitoring
network
data
set.
Water
Resour.
Res.
48,
W07701,
http://dx.doi.org/10.1029/2012WR011976.
Tetko,
I.V.,
Livingstone,
D.J.,
Luik,
A.I.,
1995.
Neural
network
studies.
1.
Comparison
of
overfitting
and
overtraining.
J.
Chem.
Inf.
Comput.
Sci.
35
(5),
826–833,
http://dx.doi.org/10.1021/ci00027a006.
Wagner,
W.,
Noll,
J.,
Borgeaud,
M.,
Rott,
H.,
1999.
Monitoring
soil
moisture
over
the
Canadian
prairies
with
the
ERS
scatterometer.
IEEE
Trans.
Geosci.
Remote
Sens.
37
(1),
206–216.
Wagner,
W.,
Blöschl,
G.,
Pampaloni,
P.,
Calvet,
J.C.,
Bizzarri,
B.,
Wigneron,
J.P.,
Kerr,
Y.,
2007.
Operational
readiness
of
microwave
remote
sensing
of
soil
moisture
for
hydrologic
applications.
Nord.
Hydrol.
38
(1),
1–20.
Wu,
T.D.,
Chen,
K.S.,
2004.
A
Reappraisal
of
the
validity
of
the
IEM
model
for
backscattering
from
rough
surfaces.
IEEE
Trans.
Geosci.
Remote
Sens.
42
(4),
743–753.
Yang,
K.,
Koike,
T.,
Kaihotsu,
I.,
Qin,
J.,
2009.
Validation
of
a
dual-pass
microwave
land
data
assimilation
system
for
estimating
surface
soil
moisture
in
semiarid
regions.
J.
Hydrometeorol.
10,
780–793.