Data
Drought
in
the
Humid
Tropics:
How
to
Overcome
the
Cloud
Barrier
in
Greenhouse
Gas
Remote
Sensing
C.
Frankenberg
1,2
,
Y.
M.
Bar‐On
1
,
Y.
Yin
3
,
P.
O.
Wennberg
1,4
,
D.
J.
Jacob
5
,
and
A.
M.
Michalak
6,7
1
Division
of
Geological
and
Planetary
Sciences,
California
Institute
of
Technology,
Pasadena,
CA,
USA,
2
Jet
Propulsion
Laboratory,
California
Institute
of
Technology,
Pasadena,
CA,
USA,
3
Department
of
Environmental
Studies,
New
York
University,
New
York,
NY,
USA,
4
Division
of
Engineering
and
Applied
Science,
California
Institute
of
Technology,
Pasadena,
CA,
USA,
5
School
of
Engineering
and
Applied
Science,
Harvard
University,
Cambridge,
MA,
USA,
6
Carnegie
Institution
for
Science,
Stanford,
CA,
USA,
7
Department
of
Earth
System
Science,
Stanford
University,
Stanford,
CA,
USA
Abstract
Diagnosing
land‐atmosphere
fluxes
of
carbon‐dioxide
(CO
2
)
and
methane
(CH
4
)
is
essential
for
evaluating
carbon‐climate
feedbacks.
Greenhouse
gas
satellite
missions
aim
to
fill
data
gaps
in
regions
like
the
humid
tropics
but
obtain
very
few
valid
measurements
due
to
cloud
contamination.
We
examined
data
yields
from
the
Orbiting
Carbon
Observatory
alongside
Sentinel‐2
cloud
statistics.
We
find
that
the
main
contribution
to
low
data
yields
are
frequent
shallow
cumulus
clouds.
In
the
Amazon,
the
success
rate
in
obtaining
valid
measurements
vary
from
0.1%
to
1.0%.
By
far
the
lowest
yields
occur
in
the
wet
season,
consistent
with
Sentinel‐2
cloud
patterns.
We
find
that
increasing
the
spatial
resolution
of
observations
to
∼
200
m
would
increase
yields
by
2–3
orders
of
magnitude
and
allow
regular
measurements
in
the
wet
season.
Thus,
the
key
to
effective
tropical
greenhouse
gas
observations
lies
in
regularly
acquiring
high‐spatial
resolution
data.
Plain
Language
Summary
Our
research
looks
at
how
well
satellites
are
able
to
observe
greenhouse
gases
such
as
carbon
dioxide
and
methane
in
tropical
areas,
which
is
important
for
understanding
climate
change.
We
find
that
these
satellites
often
cannot
make
good
measurements
in
places
like
the
Amazon
rainforest
due
to
clouds.
By
using
space‐based
instruments
that
can
peek
in
between
clouds
(requiring
about
∼
200
m
spatial
resolution),
we
would
get
much
more
frequent
information,
even
during
the
rainy
season.
Our
study
shows
that
high‐spatial
resolution
is
needed
to
regularly
observe
greenhouse
gases
in
the
tropics.
1.
Introduction
While
in
situ
measurements
of
greenhouse
gases
provide
the
most
accurate
benchmark
(Andrews
et
al.,
2014
;
Komhyr
et
al.,
1985
),
they
cannot
provide
spatially
dense
global
coverage.
Remotely
sensed
observations
can't
match
the
accuracy
of
in
situ
measurements;
however
they
offer
the
potential
to
provide
dense
spatial
coverage,
especially
in
regions
where
in
situ
measurements
are
limited.
In
the
tropics,
space‐based
measurements
could
enable
substantial
knowledge
gains,
as
the
tropics
are
not
only
sparsely
sampled
by
in
situ
observations
but
also
essential
to
global
carbon
budgets.
The
tropics
are,
however,
much
more
cloudy,
and
these
clouds
obscure
the
view
from
space.
In
passive
optical
remote
sensing
of
Earth's
atmosphere
and
surface,
clouds
shield
the
lower
atmosphere
and
affect
photon
path‐
length
distributions,
greatly
complicating
the
retrieval
of
greenhouse
gas
concentrations.
This
issue
is
particu-
larly
challenging
due
to
the
stringent
accuracy
and
precision
requirements
for
greenhouse
gas
observations
(Merrelli
et
al.,
2015
;
Miller
et
al.,
2007
).
Consequently,
rigorous
cloud
filtering
is
necessary,
albeit
at
the
cost
of
reducing
the
fraction
of
useable
observations.
Understanding
the
trade‐off
between
cloud
filtering
and
data
us-
ability
is
vital
for
assessing
the
scientific
value
of
space‐borne
missions.
To
alleviate
the
impact
of
clouds,
the
Orbiting
Carbon
Observatory
(Crisp
et
al.,
2004
)
utilizes
a
pushbroom
technique
featuring
a
narrow
cross‐track
swath
width
of
10
km
and
a
spatial
resolution
of
1.29
km
cross‐track
and
2.25
km
along‐track,
which
is
the
finest
resolution
among
existing
missions
dedicated
for
accurate
measurements
of
atmospheric
greenhouse
gases.
While
this
fine
resolution
was
chosen
to
provide
sufficient
data
even
in
the
tropics,
data
yield
predictions
for
OCO
were
based
on
cloud
climatologies
(Rayner
et
al.,
2002
)
based
on
AVHRR
RESEARCH
LETTER
10.1029/2024GL108791
C.
Frankenberg
and
Y.
M.
Bar‐On
contributed
equally
to
this
work.
Key
Points:
•
Data
yields
of
current
remotely
sensed
greenhouse
gas
(GHG)
missions
in
the
humid
tropics
are
often
below
1%
•
Shallow
cumulus
clouds
cause
most
of
the
low
data
yields,
esp.
in
the
wet
season
•
Finer
spatial
resolution
(
∼
200
m)
can
overcome
the
data
sparsity
in
the
tropics
Supporting
Information:
Supporting
Information
may
be
found
in
the
online
version
of
this
article.
Correspondence
to:
C.
Frankenberg
and
Y.
M.
Bar‐On,
cfranken@caltech.edu
;
ymbaron@caltech.edu
Citation:
Frankenberg,
C.,
Bar‐On,
Y.
M.,
Yin,
Y.,
Wennberg,
P.
O.,
Jacob,
D.
J.,
&
Michalak,
A.
M.
(2024).
Data
drought
in
the
humid
tropics:
How
to
overcome
the
cloud
barrier
in
greenhouse
gas
remote
sensing.
Geophysical
Research
Letters
,
51
,
e2024GL108791.
https://doi.org/10.1029/
2024GL108791
Received
12
FEB
2024
Accepted
9
APR
2024
Author
Contributions:
Conceptualization:
C.
Frankenberg,
Y.
M.
Bar‐On,
Y.
Yin,
D.
J.
Jacob,
A.
M.
Michalak
Formal
analysis:
C.
Frankenberg,
Y.
M.
Bar‐On
Investigation:
C.
Frankenberg,
Y.
M.
Bar‐On
Methodology:
C.
Frankenberg,
Y.
M.
Bar‐On
Project
administration:
C.
Frankenberg
Software:
C.
Frankenberg,
Y.
M.
Bar‐On,
Y.
Yin
Validation:
Y.
M.
Bar‐On
©
2024.
The
Authors.
This
is
an
open
access
article
under
the
terms
of
the
Creative
Commons
Attribution‐NonCommercial‐NoDerivs
License,
which
permits
use
and
distribution
in
any
medium,
provided
the
original
work
is
properly
cited,
the
use
is
non‐commercial
and
no
modifications
or
adaptations
are
made.
FRANKENBERG
ET
AL.
1
of
10
data
(James
&
Kalluri,
1994
)
aggregated
to
coarser
scales
(Stowe
et
al.,
1999
),
and
ignored
3D
effects
in
the
vicinity
of
clouds
(Massie
et
al.,
2017
,
2023
).
The
impact
of
small
clouds
ranging
from
tens
to
a
few
100
m
was
thus
not
fully
captured.
Although
several
point
source
imagers
for
greenhouse
gas
remote
sensing
do
have
much
finer
spatial
resolution
(see
Jacob
et
al.
(
2022
)
and
references
therein),
they
lack
the
capability
of
accurately
measuring
large‐scale
gradients
for
global
flux
inversions.
Here,
we
revisit
the
impact
of
clouds
on
GHG
remote
sensing
by
quantifying
long‐term
OCO‐2
data
yields.
These
findings
are
compared
against
cloud‐free
probabilities
computed
from
4
years
of
Sentinel‐2
cloud
data
at
10
m
resolution.
This
comparison
helps
us
explore
ways
to
improve
the
disappointing
data
collection
from
tropical
regions
in
current
satellite
missions
in
the
design
of
the
next
generation
of
satellites
focused
on
space‐based
observations
of
greenhouse
gases.
2.
Materials
and
Methods
To
assess
the
impact
of
clouds
on
greenhouse
gas
(GHG)
remote
sensing,
we
utilize
data
from
the
OCO‐2
and
OCO‐3
missions
(Taylor
et
al.,
2020
;
Wunch
et
al.,
2017
)
for
actual
GHG
measurements
and
Sentinel‐2
for
cloud
observations
(Tarrio
et
al.,
2020
).
OCO‐2
and
Sentinel‐2
are
on
sun‐synchronous
orbits
with
an
overpass
time
around
1:30
p.m.
and
10:30
a.m.,
respectively.
OCO‐3
is
hosted
on
the
ISS
with
a
precessing
orbit,
thus
overpass
times
vary,
enabling
measurements
from
early
morning
to
late
afternoon,
which
allow
an
analysis
of
time‐of‐day
dependence
of
data
yields
for
OCO‐3
(see
Text
S2
and
Figure
S2
in
Supporting
Information
S1
).
We
analyze
OCO‐2
(v11r)
and
OCO‐3
(v10r)
data
to
determine
the
number
of
high
quality
GHG
measurements,
applying
the
“xco2_quality_flag
=
0”
for
accuracy,
which
is
based
on
a
variety
of
filter
criteria
(See
Table
4.5
in
Payne
et
al.
(
2022
)).
We
calculate
the
total
number
of
measurement
counts
using
the
OCO‐2's
L1b
files
which
provide
a
total
count
of
the
number
of
observations
downlinked
from
the
spacecraft.
For
global
spatially
resolved
data
yields,
we
use
the
ratio
of
high
quality
(passing
the
quality
filter)
to
total
measurements.
Cloud
statistics
at
coarse
scales,
such
as
those
provided
by
MODIS,
are
insufficient
for
our
analysis.
Even
small
cloud
fractions
within
a
greenhouse
gas
measurement's
footprint
can
significantly
impact
data
quality
and
are
often
missed
by
cloud
climatologies.
One
reason
for
this
is
the
stark
surface
albedo
contrast
between
the
O
2
A‐
band
(about
0.4–0.5
at
760
nm)
and
the
GHG
bands
(as
low
as
0.05
at
1.6
or
2.3
μm).
Thus,
a
cloud
with
an
albedo
of
0.5
covering
only
1%
of
a
footprint
can
contribute
10%
of
the
signal
to
the
GHG
bands
but
only
1%
in
the
reference
oxygen
band.
If
this
cloud
shields
10%
of
the
column
(about
1
km
cloud
height),
it
can
cause
a
low
bias
of
1%
in
retrieved
gas
concentrations.
Thus,
requiring
a
<
ppm
bias
(Chevallier
et
al.,
2005
;
Miller
et
al.,
2007
)
might
require
screening
of
scenes
with
<
0.2%
fractional
cloud
cover.
This
drives
our
stringent
requirement
for
having
less
than
0.2%
of
a
footprint
affected
by
clouds
to
be
classified
as
cloud‐free
in
our
analysis.
A
sensitivity
analysis
with
stricter
and
looser
thresholds
settings
can
be
found
in
the
SOM,
showing
that
our
results
do
not
strongly
depend
on
it,
especially
at
finer
spatial
scales.
To
study
the
impact
of
clouds
on
OCO‐2,
we
thus
have
to
obtain
cloud
statistics
at
a
much
finer
resolution
than
OCO‐2's
footprint.
In
the
tropics,
frequent
shallow
cumulus
clouds
are
often
linked
to
forest
surface
fluxes
(Heiblum
et
al.,
2014
),
are
spatially
organized
and
have
cloud
gaps
that
are
smaller
than
1
km
in
scale.
Sentinel‐2,
with
its
frequent
revisits
and
10
m
resolution
(Drusch
et
al.,
2012
),
is
uniquely
suited
for
this
task.
It
enables
us
to
accurately
calculate
the
likelihood
of
obtaining
cloud‐free
measurements.
In
our
analysis,
a
“cloud‐free”
pixel
is
one
where
less
than
0.2%
of
the
footprint
area
is
covered
by
clouds
or
cloud
shadows.
We
employ
the
Cloud
Score
+
product
(Pasquarella
et
al.,
2023
)
based
on
Sentinel‐2
data,
using
a
threshold
of
0.65
to
identify
cloud‐free
pixels,
filtering
for
both
cloud
cover
as
well
as
cloud
shadows.
There
are
edge
effects
around
clouds
(Bell
et
al.,
2020
),
requiring
somewhat
larger
cloud‐free
areas
than
an
in-
dividual
satellite
footprint
size.
The
analysis
is
performed
using
Google
Earth
Engine,
focusing
on
specific
latitude
and
longitude
ranges
and
using
square‐sized
convolution
kernels
to
vary
spatial
resolutions.
Within
each
1°
×
1°
area,
we
compute
the
fraction
of
pixels
passing
our
cloud
filter
threshold
at
a
given
footprint
size
for
each
individual
Sentinel‐2
image,
which
are
acquired
every
5
days.
All
individual
cloud‐free
fractions
per
image
are
then
used
to
compute
probability
distributions
of
cloud‐fractions
within
a
certain
domain
and
time‐period.
Visualization:
C.
Frankenberg,
Y.
M.
Bar‐On
Writing
–
original
draft:
C.
Frankenberg,
Y.
M.
Bar‐On,
P.
O.
Wennberg,
D.
J.
Jacob,
A.
M.
Michalak
Writing
–
review
&
editing:
C.
Frankenberg,
Y.
M.
Bar‐On,
Y.
Yin,
P.
O.
Wennberg,
D.
J.
Jacob,
A.
M.
Michalak
Geophysical
Research
Letters
10.1029/2024GL108791
FRANKENBERG
ET
AL.
2
of
10
19448007, 2024, 8, Downloaded from https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2024GL108791 by California Inst of Technology, Wiley Online Library on [14/06/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License