This document discusses assessing corrections to the MODIS C06 3km aerosol product over urban areas using high-resolution AERONET data. It finds biases in the 3km product over urban sites compared to AERONET measurements. The authors aim to improve MODIS' land surface correction algorithms, which were trained on non-urban surfaces, by combining AERONET and MODIS data over sufficiently clean days. They retrieve land surface spectral ratios using this approach and apply filters to ensure minimal aerosol contamination. The improved land surface models could then provide better aerosol retrievals over urban regions.
Assessing MODIS C006 urban corrections using the High Resolution Dragon AERONET Network
1. Assessing
MODIS
C06
Urban
Correc6ons
Using
the
High
Resolu6on
Dragon
AERONET
Network
Nabin
Malakar,
Adam
A/a,
Barry
Gross,
Fred
Moshary
Op#cal
Remote
Sensing
Lab,
CCNY
Min
Oo
CIMSS
/
UW-‐Madison
2. Mo6va6on
l Aerosol
Retrieval
over
land
is
greatly
affected
by
land
surface
albedo
(if
bright
enough).
l MODIS
land
surface
compensa6on
algorithms
for
global
applica6ons
were
trained
using
non
urban
land
surface
types
(mixtures)
such
as
vegeta6ons/
clays.
l As
urbaniza6on
con6nues
to
increase,
the
differences
in
land
surface
behavior
need
to
be
bePer
understood.
l These
issues
become
even
more
significant
as
higher
resolu6on
aerosol
products
such
as
C006
3km
Aerosol
Retrievals
become
available
Single
scaPering
Mul6ple
ScaPering
Photons
hit
land
surface
And
reflected
back
to
space
4. Approach
l We
previously
inves6gated
the
existence
of
high
bias
in
AOD
retrievals
in
C005
for
significantly
urbanized
areas
such
as
New
York
City
l By
combining
AERONET
with
MODIS
observa6ons
over
sufficiently
“clean”
days,
it
is
possible
to
improve
on
the
exis6ng
land
surface
model
needed
to
correct
for
land
reflec6on
l Applying
this
approach
over
a
region
is
complicated
by
the
fact
that
only
a
single
AERONET
sta6on
is
available
and
an
assump6on
that
the
AOD
and
phase
func6on
proper6es
are
homogeneous
on
a
regional
scale
are
clearly
an
issue
l Using
the
Dragon
Network
allows
for
the
poten6al
of
using
bePer
AERONET
informa6on
in
“tuning”
the
surface
while
also
providing
bePer
sta6s6cal
valida6on.
l We
also
inves6gate
the
neural
network
approach
to
correct
the
bias.
5. Opera6onal
satellites
retrieval
over
land
l MODIS
aerosol
retrieval
uses
three
wavelength
channels
(470,
660,
2120nm)
l Mul6
wavelength
measurements
help
separate
fine
/
coarse
components.
l However,
the
surface
reflec6on
contaminates
the
signals.
l To
es6mate
this,
MODIS
does
the
following
l Assumes
the
long
wavelength
channel
is
insensi6ve
to
the
atmosphere
so
the
signal
must
be
due
only
to
the
ground
reflec6on
(Rg_2120)
l Once
the
long
wavelength
reflec6on
is
es6mated,
use
semi-‐empirical
models
taking
into
account
how
vegeta6ve
the
surface
is
to
es6mate
the
VIS
to
SWIR
ra6os
(Rg_470)/
(Rg_2120),
(Rg_660)/
(Rg_2120)
l MODIS
uses
an
index
called
the
Modified
Vegeta6on
Index
(MVI),
which
combines
NIR
and
SWIR
to
es6mate
vegeta6on
class.
l We
demonstrate
that
these
ra6os
are
not
well
represented
in
opera6onal
algorithms
and
need
refinement
which
allows
bePer
aerosol
retrieval.
TOA
m
TOA
m
TOA
m
TOA
m
MVI
µµ
µµ
ρρ
ρρ
12.224.1
12.224.1
+
−
=
6. Retrieving
Land
Surface
Band
Spectral
Ra6os
l The
Collect
5/6
approach
allows
the
VIS-‐SWIR
ground
albedo
correla6on
coefficients
to
be
a
func6on
of
surface
type
(urban/vegeta6on
MVI)
and
observa6on
angles
(scaPering
angle).
l In
our
case,
we
ingest
AOD
from
Aeronet
to
atmospherically
correct
the
MODIS
images
l To
ensure
that
the
best
surface
retrieval
is
made,
the
following
filters
are
applied
– AOD
<
0.2,
– angstrom
exponent
>
1
to
assure
minimal
aerosol
contamina6on
at
2.1
um
– Homogeneous
condi6ons
(variability
of
AERONET
AOD
for
+/-‐
3
hours
<
20%)
which
helps
us
extrapolate
AOD
over
en6re
domain
– Mask
all
water
pixels
– For
Dragon
Network,
we
use
Aeronet
averages
when
possible
to
improve
quality
of
land
surface
reflec6on
and
remove
homogeneity
assump6on.
7. Obtaining
surface
albedos
using
combined
MODIS
–
Aeronet
Data
( )
( )
( ) albedosphericalicAtmospher
ontransmissitotaldownwardandUpward,
ereflectancpath,,,
,
=
=
=Δ
λ
θλ
φθθλρ
s
T ud
ivatm
g
udg
atmTOA
s
TT
ρ
ρ
ρρ
−
+=
1
Aeronet
Op6cal
Depth
+
MODIS
Aerosol
Phase
Func6on
consistent
with
AOD
Once
this
is
done,
we
can
Isolate
Lamber6an
albedo
)( atmTOAud
atmTOA
g
sTT ρρ
ρρ
ρ
−+
−
=⇒
Use
Aeronet
AOD
to
fix
the
MODIS
Aerosol
Phase
func6on
model
From
this,
we
can
get
all
relevant
atmospheric
scaPering
parameters
[⌧550
aer ]aeronet ! [Paer(⇥scat, ⌧550
aer , )]urban-nonabs
8. 80 90 100 110 120 130 140 150 160
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Rho0.66um/Rho2.12um Scattering angle
y = 1.2e-005*x + 0.77
data 1
linear
80 90 100 110 120 130 140 150 160
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Rho0.47um/Rho2.12um
Scattering angle
y = 0.00059*x + 0.45
data 1
linear
Band
Correla6on
with
ScaPering
Angle
( ) ( ) ( )
2:1
2120
=
Θ=
i
f gsiig ρλρ
Mean=0.5153
std=0.0858
Mean=
0.7734
std=0.0729
Rho
0.470/
Rho2.12
Rho
0.660/
Rho
2.120
Once
new
correla6ons
are
found,
we
can
replace
the
COO5
Correla6on
procedures
and
assess
retrieval
of
AOD
(for
all
cases)
Band
Correla6on
with
ScaPering
Angle
(water
mask
included)
shows
minimal
angular
dependence
valida/ng
lamber/an
assump/ons
9. General
Rela6onship
between
Surface
Type
and
the
VIS/SWIR
reflec6on
ra6os
in
urban
areas
Regional
surface
data
retrievals
(50km
x
50km)
around
different
ci6es
with
AERONET
at
center.
Note
that
VIS/SWIR
ra6os
decrease
with
MVI
index
in
contradic6on
to
the
MODIS
C005
opera6onal
models.
(Later,
we
see
that
C006
trend
is
improved
over
C005)
• When
MVI
is
low
(i.e
urban),
SRC’s
are
significantly
underes6mated
• The
C005
model
actually
shows
an
opposite
trend
indica6ve
of
the
differences
between
low
MVI
soils
and
urban
materials
• NYC
is
by
far
the
most
biased
region
over
other
urban
areas
in
comparisons
to
other
urban
centers.
10. Anomalies
in
Spectral
Ra6os
Tuned
Surface
Reflec6on
Ra6o
Strong
correla6on
between
urban
frac6on
and
regionally
tuned
surface
reflec6on
ra6o
Urban
Land
Cover
Deciduous
broadleaf
forest
12. Spectral
Ra6os
by
land
class
0 0.2 0.4 0.6 0.8 1
0
0.2
0.4
0.6
0.8
1
MVI Index
Ref660/Ref2120
cropland
regional
C006
0 0.2 0.4 0.6 0.8 1
0
0.2
0.4
0.6
0.8
1
MVI Index
Ref660/Ref2120
mixed forrest
regional
C006
0 0.2 0.4 0.6 0.8 1
0
0.2
0.4
0.6
0.8
1
MVI Index
Ref660/Ref2120
urban/built
regional
C006
0 0.2 0.4 0.6 0.8 1
0
0.2
0.4
0.6
0.8
1
MVI Index
Ref660/Ref2120
deciduous broadleaf
regional
C006
C006
generally
does
bePer
than
C005
(correct
trend)
but
urban
land
class
is
completely
underes6mated
at
low
MVI
13. Bias
Dependence
on
Different
Factors
l Small
but
posi6ve
bias/RMSE
dependence
on
%
urban
and
scaPering
angle
l Negligible
bias
on
C006
surface
reflec6on
ra6o
and
angstrom
Coefficient.
l Urban
classifica6on
should
be
ingested
into
high
resolu6on
algorithms
0 20 40 60 80
-0.2
0
0.2
0.4
0.6
Urban %
AODC006-AERONETAOD
100 120 140 160 180
-0.2
0
0.2
0.4
0.6
Single Scattering Angle%
AODC006-AERONETAOD
1 1.5 2 2.5
-0.2
-0.1
0
0.1
0.2
0.3
0.4
0.5
Angstrom Coefficient%
AODC006-AERONETAOD
0.4 0.45 0.5 0.55 0.6
-0.2
-0.1
0
0.1
0.2
0.3
0.4
660 /2120 Reflectance Ratio
AODC006-AERONETAOD
14. Case
Scenario
July
29
1740
UTC
l Strongest
correc6ons
occur
in
urban
zones
l Best
agreement
seen
when
correc6on
is
applied
l No
significant
correc6on
in
non
urban
area
(green
circle)
0 0.2 0.4 0.6 0.8 1
0
0.2
0.4
0.6
0.8
1
AERONET
MODIS3kmAOD
July 29 AQUA 1740 UTC
C006
Regional
17. Bias
Correc6on
using
Machine-‐
Learning
17
Target
Compare
Machine-‐Learning:
Neural
nets,
SVM,
RF,
GP
etc.
Input
18. Neural
network
18
yk =
0
@
nX
j=1
wkjxj
1
A
• Also
referred
to
as
mul6
layer
perceptron
method,
• Used
widely
for
classifica6on
or
func6on
approxima6on.
Where ø:
is
the
transfer
func6on
wkj: weight from unit j to unit k,
xj : n input variables
The
output
of
the
kth
neuron:
Inputs
Hidden layer
Outputs
22. Conclusions
l Assessment
of
3km
resolu6on
products
using
Dragon
Network
shows
somewhat
enhanced
bias
in
comparison
to
10km
l We
find
that
the
regionally
tuned
surface
spectral
ra6o
model
is
highly
correlated
to
several
dis6nguishing
land
classes
(Urban
/
deciduous
broadleaf
forest)
l The
current
MVI
parameter
used
to
get
the
VIS
channel
surface
albedo
es6mate
is
qualita6vely
and
quan6ta6vely
insufficient
to
separate
urban
land
areas
from
other
land
classes
(deciduous
broadleaf
forest)
l Significant
Improvement
can
be
seen
in
bias
reduc6on
using
regional
land
surface
model
with
negligible
differences
in
correla6on
l Adding
land
classifica6on
with
MVI
should
help
remove
anomalies
for
urban
retrievals.
l We
used
the
MODIS
3
km
AOD
products
from
AQUA
and
TERRA,
and
developed
a
machine-‐learning
framework
to
compare
and
correct
the
remote
sensing
product
with
respect
to
the
ground-‐based
AERONET
observa6ons.
l We
also
constructed
a
neural
network
es6mator
to
obtain
bias-‐corrected
AOD
product.
23. Future
Work
l Es6mate
PM2.5
from
the
bias-‐corrected
AOD
l Par6culates
with
a
diameter
of
2.5
microns
or
less
l Can
have
adverse
health
effects
l Once
in
the
body
may
lead
to
oxida6ve
inflamma6on
in
the
organs.
Ref:
hPp://www.airnow.gov