Rapid Spectral Cloud Screening Onboard Aircraft and Spacecraft


Thompson, D R.; Green, R O.; Keymeulen, D; Lundeen, S K.; Mouradi, Y; Nunes, D Cahn; Castano, R; Chien, S A.

IEEE Transactions on Geoscience and Remote Sensing 52(11): 6779-6792

2014


IEEE
TRANSACTIONS
ON
GEOSCIENCE
AND
REMOTE
SENSING,
VOL.
52,
NO.
11,
NOVEMBER
2014
6779
Rapid
Spectral
Cloud
Screening
Onboard
Aircraft
and
Spacecraft
David
R.
Thompson,
Member,
IEEE,
Robert
0.
Green,
Dither
Keymeulen,
Member,
IEEE,
Sarah
K.
Lundeen,
Yasha
Mouradi,
Daniel
Calm
Nunes,
Rebecca
Castailo,
and
Steve
A.
Chien
Abstract—Next-generation
orbital
imaging
spectrometers
will
generate
unprecedented
data
volumes, demanding
new
methods
to
optimize
storage
and
communication
resources.
Here,
we
demonstrate
that
onboard
analysis
can
excise
cloud-contaminated
scenes,
reducing
data
volumes
while
preserving
science
return.
We
calculate
optimal
cloud-screening
parameters
in
advance,
exploiting
stable
radiometric
calibration
and
foreknowledge
of
illumination
and
viewing
geometry.
Channel
thresholds
expressed
in
raw
instrument
values
can
be
then
uploaded
to
the
sensor
where
they
execute
in
real
time
at
gigabit-per-second
(Gb/s)
data
rates.
We
present
a
decision
theoretic
method
for
setting
these
instrument
parameters
and
characterize
performance
using
a
continuous
three-year
image
archive
from
the
"classic"
Airborne
Visible/Infrared
Imaging
Spectrometer
(AVIRIS-C).
We
then
sim-
ulate
the
system
onboard
the
International
Space
Station,
where
it
provides
factor-of-two
improvements
in
data
volume
with
negligi-
ble
false
positives.
Finally,
we
describe
a
real-time
demonstration
onboard
the
AVIRIS
Next
Generation
(AVIRIS-NG)
flight
plat-
form
during
a
recent
science
campaign.
In
this
blind
test,
cloud
screening
is
performed
without
error
while
keeping
pace
with
instrument
data
rates.
Index
Terms
Cloud
screening,
imaging
spectroscopy,
lossy
compression,
pattern
recognition,
real-time
systems.
I.
INTRODUCTION
F
UTURE
Earth
science
missions
will
face
unprecedented
data
volumes.
Data
product
sizes
and
production
rates
have
steadily
increased
due
to
improvements
in
detector,
optics,
and
onboard
data
handling
technology.
High-resolution
spec-
trometers
such
as
NASA's
000-2
mission
will
yield
over
one
million
soundings
per
day
[1].
Proposed
imaging
spectrometers
such
as
HyspIRI
[2]
or
an
International
Space
Station
(ISS)
imaging
spectrometer
would
generate
data
rates
on
the
order
of
1
gigabit
per
second
(Gb/s).
These
rates
are
a
consequence
of
the
full
spectral
measurement
at
high
spatiotemporal
reso-
lution
required
for
a
range
of
unique
science
and
application
objectives
[3].
However,
the
large
data
volumes
affect
mis-
sion
requirements for
the
entire
data
handling
chain,
including
Manuscript
received
July
22,
2013;
revised
October
23,
2013
and
December
11,
2013;
accepted
January
18,
2014.
Date
of
publication
February
19,
2014.
The
Airborne
Visible/Infrared
Imaging
Spectrometer
is
operated
by
JPL
with
support
from
the
National
Aeronautics
and
Space
Administration.
The
MODIS
land-cover
data
were
obtained
through
the
online
Data
Pool
at
the
NASA
Land
Processes
Distributed
Active
Archive
Center
(LP
DAAC),
USGS/Earth
Resources
Observation
and
Science
(EROS)
Center,
Sioux
Falls,
SD.
The
authors
are
with
the
Jet
Propulsion
Laboratory,
California
Institute
of
Technology,
Pasadena,
CA
91109
USA.
Color
versions
of
one
or
more
of
the
figures
in
this
paper
are
available
online
at
http://ieeexplore.ieee.org
.
Digital
Object
Identifier
10.1109/TGRS.2014.2302587
onboard
digitization,
storage,
downlink,
ground
processing,
and
distribution
[4].
Bottlenecks
along
this
path
can
constrain
the
instrument
duty
cycle,
reducing
science
and
application
yield
[5].
In
particular,
bandwidth
constraints
have
motivated
new
advanced
lossless
compression
techniques
such
as
the
FL
algorithm
[6]—[9]
that
have
achieved
compression
rates
of
four
or
greater.
Efforts
to
optimize
lossless
methods
eventually
face
theoretical
limits,
but
data
rates
continue
to
increase.
The
challenge
has
driven
research
into
other
techniques
that
can
further
reduce
data
volumes
while
preserving
science
yield.
One
promising
approach
is
to
avoid
storing
or
transmitting
cloud-contaminated
data
[10],
[11].
Historically,
clouds
are
estimated
to
cover
54%
or
more
of
the
Earth's
land
area
and
68%
or
more
of
the
oceans
[12]—[14].
Many
algorithms
to
es-
timate
atmospheric
or
surface
properties
cannot
function
in
the
presence
of
condensed
water/ice
clouds.
This
makes
more
than
half
of
visible-to-short-wavelength
infrared
(VSWIR)
scenes
in
remote
sensing
archives
unusable
for
their
intended
science
and
applications
purpose
[15].
Excising
these
scenes
at
the
sensor
can
significantly
reduce
onboard
storage
and
bandwidth
requirements.
However,
the
community
lacks
a
practical
algo-
rithm
capable
of
real-time
execution
in
instrument
hardware.
This
paper
addresses
the
need.
We
present
a
real-time
cloud-
screening
method
that
executes
on
raw
sensor
data
for
use
on-
board
aircraft
and
spacecraft.
We
report
its
performance
both
in
simulations
and
in
a
deployment
on
Airborne
Visible/Infrared
Imaging
Spectrometer
-
Next
Generation
(AVIRIS-NG).
Our
approach
is
designed
for
the
unique
requirements
of
real-time
cloud
screening,
with
rapid
Gb/s
execution
rates
and
respon-
siveness
to
changing
terrain
and
illumination
conditions.
It
amounts
to
several
simple
channel
thresholds,
which
are
dy-
namically
adapted
to
account
for
predicted
brightness
of
clouds
and
terrain.
Some
inaccuracy
is
tolerable
since
any
missed
clouds
can
be
excised
later
on
the
ground.
It
is
operationally
very
simple
to
implement,
and
conservative
settings
ensure
that
good-quality
science
data
are
preserved
at
all
costs.
We
will
demonstrate
that
it
is
possible
to
achieve
data
volume
reduc-
tions
near
the
theoretical
maximum
without
any
significant
loss
of
science
data.
A.
Prior
Work
We
focus
on
the
VSWIR
electromagnetic
spectrum
from
0.4-2.5
,um.
Fig.
1
shows
an
example
scene
from
the
"classic"
Airborne
Visible/Infrared
Imaging
Spectrometer
(AVIRIS-C)
with
representative
spectra
of
different
materials
and
clouds.
There
are
many
studies
of
cloud
detection
in
these
wavelengths,
4.
"
"...•
a..
Z
-
Cloud
Bare
terrain
'
'Water
Snow
u.
14
4
6780
IEEE
TRANSACTIONS
ON
GEOSCIENCE
AND
REMOTE
SENSING,
VOL.
52,
NO.
11,
NOVEMBER
2014
x
10
4
4
3.5
3
2.5
2
0.5
0
0.5
1
1.5
2
2.5
Wavelength
(pm)
Fig.
1.
(Left)
AV1RIS-C
image
f100521t02p05,
a
challenging
scene
that
contains
both
clouds
and
snow.
(Right)
Spectra
from
bare
terrain,
snow,
open
water,
and
clouds,
in
units
of
dark-subtracted
instrument
digital
numbers
(DNs).
and
algorithms
vary
in
their
assumptions
and
complexity.
"Classical"
cloud
screening
applies
threshold
tests
to
spatial
and
spectral
properties
of
the
image
[16].
Pixels
whose
values
fall
outside
valid
ranges
are
marked
as
cloudy.
For
example,
the
MODIS
algorithm
compares
selected
visible
and
near-
infrared
(VNIR)
and
near-infrared
(NIR)
bands
to
predeter-
mined
thresholds
and
then
aggregates
the
result
in
different
combinations
depending
on
land
type
[17]—[19].
The
algorithm
uses
a
combination
of
14
wavelengths
and
over
40 tests.
This
underscores
the
intrinsic
difficulty
of
constructing
a
universal
and
complete
cloud-screening
procedure.
Even
more
complex
algorithms
are
possible.
Some
state-
of-the-art
cloud-screening
techniques
estimate
the
optical
path
from
absorption
features
such
as
the
oxygen
A
band,
as
in
Gomez-Chova
et
al.
[20]
or
Taylor
et
al.
[15].
Thermal
infrared
(IR)
channels
can
add
brightness
temperature
information.
Minnis
et
al.
predict
clear-sky
brightness
temperature
values
using
ambient
temperature
and
humidity
and
then
excise
pixels
outside
these
intervals
[21].
Texture
cues
can
be
also
used
to
recognize
clouds
by
their
high
spatial
heterogeneity
[22].
Martins
et
al.
demonstrate
that
a
simple
spatial
analysis,
i.e.,
the
standard
deviation
of
VNIR
isotropic
reflectances
in
a
3
x
3
pixel
window,
reliably
discriminates
clouds
from
aerosol
plumes
over
ocean
scenes
[23].
Murtagh
et
al.
represent
spatial
dependency
using
a
probabilistic
Markov
random
field
prior
[24].
Other
efforts
use
special
sensing
modalities
such
as
polarization
[25].
Of
direct
relevance
to
this
work,
onboard
cloud
screening
has
been
demonstrated
onboard
the
EO-1
spacecraft
[26].
EO-1
cloud
screening
uses
the
solar
zenith
angle
to
compute
the
apparent
top-of-atmosphere
(TOA)
reflectance.
Then,
it
applies
a
branching
sequence
of
threshold
tests
based
on
carefully
crafted
spectral
ratios
to
distinguish
clouds
and
bright
land-
forms
such
as
snow,
ice,
and
desert
sand.
The
EO-1
cloud
detection
also
acts
as
a
data
filtering
step
prior
to
onboard
flood
and
cryospheric
classification
[27],
[28].
To
our
knowledge,
it
is
the
only
previous
case
of
cloud
screening
performed
on
orbit.
Due
to
the
limitations
of
the
mission's
12-MHz
flight
computer,
screening
a
1024
x
256
image
requires
about
30
minutes
[29],
which
is
three
orders
of
magnitude
lower
than
our
desired
processing rate.
Nevertheless,
the
work
provides
an
important
proof
of
concept
and
a
foundation
for
our
study.
B.
Algorithm
Requirements
Previous
systems
try
to
screen
all
clouds
to
prevent
con-
tamination
of
later
retrieval
algorithms.
In
contrast,
we
aim
to
reduce
the
instrument
data
volume,
which
leads
to
distinct
requirements.
Completeness
is
not
critical
since
the
end
user
can
perform
more
precise
cloud
screening
later.
Our
algorithm
can
be
conservative,
abstaining
from
ambiguous
classifications
to
prevent
loss
of
science
data.
This
requires
some
way
to
represent
classification
certainty.
There
is
precedent;
for
exam-
ple,
Ishida
and
Nakajima
supplement
their
binary
decision
with
a
confidence
score
[30].
Gomez-Chova
et
al.
use
a
Gaussian
mixture
model
to
produce
posterior
probabilities
[20].
The
Bayesian
probabilistic
model
of
Merchant
et
al.
combines
observational
data
with
prior
predictions
from
atmospheric
forecasts,
leading
to
true
probabilistic
predictions
[31].
Rig-
orous
probabilistic
approaches
are
well
suited
to
an
onboard
algorithm
that
abstains
from
uncertain
classifications.
Onboard
cloud
screening
must
also
satisfy
strict
computa-
tional
constraints.
The
algorithm
must
process
all
data
collected
by
the
spectrometer
before
it
enters
the
flight
recorder.
In
many
cases,
this
requires
that
the
algorithm
run
in
instrument
hardware
such
as
a
field-programmable
gate
array
(FPGA),
entailing
additional
design
requirements
[10].
For
a
push
broom
instrument,
image
lines
arrive
sequentially.
Depending
on
the
buffering
strategy,
it
may
not
be
possible
to
pass
more
than
a
handful
of
values
from
one
line
to
the
next.
This
limits
the
use
of
spatial
context.
Moreover,
classifiers
instantiated
in
hardware
logic
typically
forgo
the
use
of
exponentials,
transcendental
functions,
and
even
floating-point
operations,
precluding
many
nonlinear
classifiers
and
naive
implementations
of
linear
clas-
sifiers.
Additionally,
cloud
screening
should
operate
with
Gb/s
throughput,
using
a
small
fixed
number
of
arithmetic
operations
on
locally
available
data,
and
have
a
deterministic
computa-
tional
path
without
recursion
or
iterative
loops.
This
excludes
many
classifiers
such
as
nearest
neighbor
or
decision
tree
algorithms.
Finally,
as
a
consequence
of
embedded
hardware
execution,
cloud
screening
must
operate
on
raw
instrument
data
values.
This
rules
out
most
classical cloud-screening
algorithms
since
it
is
not
realistic
to
reproduce
ground-side
processing,
which
could
provide
calibrated
reflectance
as
input.
This
paper
presents
a
technique
to
satisfy
these
requirements.
We
demonstrate
a
very
simple
cloud-screening
algorithm
that
operates
on
raw
instrument
data,
significantly
reducing
its
volume
while
achieving
a
higher
throughput
rate
than
any
previously
reported
cloud-screening
system.
As
with
EO-1,
the
screening
decision
is
a
sequence
of
threshold
tests
on
selected
wavelength
values.
However,
these
thresholds
are
recomputed
before
each
observation
using
foreknowledge
of
scene
pa-
rameters,
i.e.,
the
solar
irradiance
from
orbital
ephemeris
and
instrument
calibration
and
terrain
properties
from
historical
data.
These
define
distributions
of
raw
uncalibrated
instrument
values
for
cloud
and
terrain,
which
in
turn
prescribe
channel
THOMPSON
et
aL:
RAPID
SPECTRAL
CLOUD
SCREENING
ONBOARD
AIRCRAFT
AND
SPACECRAFT
6781
Threshold
cp
Excluded
region
2?
cover
effects,
and
illumination.
Fortunately,
we
can
predict
these
factors
to
first
order
using
historical
data
and
observation
geometry.
We
use
the
following
four
sequential
steps.
Terrain
1)
In
advance,
determine
the
channels
that
will
be
used.
U
pixels
Cloud
2)
Predict
pixel
brightness
by
extrapolation
from
historical
pixels
data.
Channel
value
y
1
3)
Optimize
channel
thresholds
to
reflect
data
reduction
and
false
alarm
requirements.
Fig.
2.
Thresholds
0
define
an
exclusion
region
to
classify
pixels
as
cloudy.
4)
In
real
time,
apply
these
thresholds
to
excise
cloudy
data.
thresholds.
Operators
update
thresholds
as
often
as
needed
to
track
changes
in
imaging
conditions
and
geometry.
This
partitions
the
cloud-screening
calculations
into
an
offline
part
that
benefits
from
powerful
computers
and
ancillary
meteoro-
logical
information
and
a
fast
real-time
part
suited
to
onboard
execution
and
FPGA
logic.
The
following
sections
describe
the
algorithm's
theoreti-
cal
assumptions.
We
present
a
formal
Bayesian
probabilistic
method
for
selecting
thresholds.
We
then
evaluate
performance
for
different
operation
scenarios
using
a
three-year
historical
image
archive
of
the
"classic"
Airborne
Visible/Infrared
Imag-
ing
Spectrometer
(AVIRIS-C)
[32].
A
case
study
quantifies
the
compression
benefits
using
orbital
parameters
of
the
ISS.
Finally,
we
report
the
results
of
a
field
deployment
onboard
the
AVIRIS
Next
Generation
(AVIRIS-NG)
airborne
imaging
spectrometer
[33].
A
cloud-screening
testbed
was
installed
in
parallel
with
the
regular
AVIRIS-NG
data
system
and
operated
without
error
during
a
recent
science
campaign.
II.
METHOD
Our
cloud-screening
approach
tests
specific
channels
with
user-defined
thresholds.
We
will
focus
on
VSWIR
imaging
spectrometer
measurements
and
will
refer
to
each
spectrum
as
a
pixel
(i.e.,
a
single
image
location
with
all
wavelengths).
In
mathematical
terms,
a
cloud-screening
algorithm
must
define
an
exclusion
region
l?.
C
Rd,
a
range
of
instrument
data
values
for
which
a
pixel
is
judged
to
be
cloudy.
The
observed
spectrum
of
instrument
data
forms
a
vector
y
with
multiple
spectral
channels
per
pixel.
The
cloud-screening
decision
maps
these
pixel
brightness
values
to
a
binary
classification
c
=
f
(y)
:
Rd
{ci,
c
2
},
where
c
1
represents
the
event
that
clear
sky
is
observed
and
c
2
represents
that
there
is
a
cloud
present.
The
corresponding
decision
rule
is
simply
c
i
,
if
y
E
f
(y)
=
'R.
c2,
if
Y
1Z.
(
1
)
Here,
we
define
'R.
with
a
set
of
channel
thresholds
0,
marking
any
pixel
that
exceeds
all
these
thresholds
as
cloudy.
Fig.
2
shows
the
decision
for
a
single
channel.
The
vertical
axis
in-
dicates
probability
density.
We
seek
thresholds
that
best
distin-
guish
terrain
classes
from
cloud
pixels.
Note
that
there
is
some
overlap
between
the
distributions;
hence,
in
this one
channel,
the
populations
are
inseparable.
Thus,
there
will
always
be
some
unavoidable
classification
error.
These
terrain
and
cloud
brightness
distributions
depend
on
scene-specific
factors
such
as
land
type,
seasonal
and
snow-
A.
Channel
Selection
Clouds
are
bright
across
the
ultraviolet,
visible
wavelengths,
and
1R.
However,
cloud-screening
algorithms
conventionally
use
only
a
small
subset
of
the
available
channels.
The
MODIS
cloud
mask
uses
the
0.659-pm
reflectance
channel.
The
0.865-pm
channel,
ratioed
with
0.659
pm,
can
identify
clouds
by
their
flat
spectra
[26].
Furthermore,
0.936-
and
0.940-pm
channels
discriminate
low
clouds
and
shadows,
respectively.
Additional
tests
on
NW
channels
at
1.24
and
1.65
pm
help
distinguish
snow
[34].
The
1.38-pm
band
indicates
cirrus
clouds
[35].
The
channel
lies
in
a
water
vapor
absorption
feature
that
is
typically
opaque
due
to
water
vapor
in
the
lower
troposphere;
thus,
large
reflectance
indicates
a
reflection
from
high-altitude
cirrus.
Postprocessing
can
often
correct
translucent
clouds
such
as
high
cirrus,
so
these
can
be
considered
be
considered
"good
data"
for
our
purposes.
We
will
focus
exclusively
on
low
opaque
clouds.
This
simplifies
the
problem
considerably
since
opaque
clouds
are
easiest
to
detect.
Our
approach
can
use
arbitrarily
many
frequencies,
but
the
following
experiments
use
just
two
channels
for
clarity.
A
blue
visible
channel
discrimi-
nates
clouds
from
land
and
ocean,
whereas
a
short-wavelength
infrared
(SWIR)
channel
excludes
snow
and
ice.
Section
III-B
demonstrates
that
this
pairing
has
the
highest
information
con-
tent
for
our
data
set.
B.
Estimation
of
Cloud
and
Surface
Appearances
This
section
describes
models
of
cloud
and
terrain
appear-
ance
that
are
used
on
the
ground
to
predict
pixel
brightness
dis-
tributions.
Our
method
is
similar
to
that
of
Merchant
et
al.
[31],
which
represents
explicit
distributions
of
cloud
appearances
under
different
imaging
conditions,
atmospheric
states,
and
terrain
types.
A
prior
P(ci
)
represents
the
known
probability
of
observing
clouds,
which
can
be
a
historical
average.
A
state
variable
x
represents
known
background
conditions
such
as
the
surface
type.
To
set
appropriate
thresholds,
we
must
ultimately
estimate
P(y
I
x,
c),
the
conditional
probabilities
of
pixel
values
for
clouds
and
terrain.
These
uncalibrated
instrument
values
are
sensitive
to
variations
in
solar
input
due
to
observation
ge-
ometry.
We
simplify
the
problem
by
estimating
the
related
distribution
P(z
I
x,
c),
i.e.,
a
normalized
representation
that
removes
the
solar
variability.
We
use
TOA
reflectance
values
z
that
have
been
adjusted
for
the
solar
zenith
angle
9
using
7rd
2
z
=
cos
(0)s
g
(y
b).
(2)
6782
IEEE
TRANSACTIONS
ON
GEOSCIENCE
AND
REMOTE
SENSING,
VOL.
52,
NO.
11,
NOVEMBER
2014
Here,
b
E
R
d
is
a
bias,
and
g
E
R
d
is
a
gain
that
translates
the
measurement
to
radiance
in
W/nm/sr/m
2
.
The
Earth—Sun
distance
d
is
a
function
of
the
Julian
day.
The
value
s
E
R
d
is
the
incident
solar
flux
per
channel,
typically
computed
by
convolving
a
solar
illumination
model
with
the
band
spectral
response.
The
resulting
z
are
solar
normalized
but
unitless
brightness
values
decoupled
from
observation
geometry.
We
accumulate
the
new
values
in
multidimensional
histograms
to
form
P(z
I
x,
c),
storing
a
different
histogram
for
every
distinct
surface
class
x.
Each
histogram
has
one
dimension
per
spectral
channel
used
in
the
test.
Operationally,
one
may
need
to
analyze
an
orbital
segment
spanning
a
range
of
solar
angles
and
surface
types.
In
this
case,
P(y
I
x,
c)
combines
each
surface
type
in
correct
proportion.
We
perform
a
weighted
summation
of
the
appropriate
pixel
brightness
distributions
at
each
segment
time
step
t
E
T,
simul-
taneously
accounting
for
solar
effects
and
transforming
the
so-
lar
normalized
representations
back
to
instrument
data
values.
For
a
histogram,
this
is
a
simple
operation,
with
normalized
bin
coordinates
z
related
to
new
timestep-specific
bin
coordinates
Yt
using
1
cos
Yt
xt,
=
E
[(z
xt,c)+
b
g7rd2
(
3
)
1
7
1
t
9t)s
Here,
x
t
and
O
t
refer
to
predictions
of
land
type
and
solar
zenith
angle
from
orbital
ephemeris.
In
the
special
case
of
models
having
two
spectral
channels,
inverting
solar
normalization
is
tantamount
to
a
simple
affine
transformation
of
a
2-D
image.
Fig.
3
illustrates
this
process
using
a
typical
AVIRIS-C
image
(f100521t02p05).
The
top
panel
shows
the
original
populations
of
background
terrain
and
cloud
pixels
when
imaged
with
a
solar
zenith
angle
of
18°.
The
vertical
and
horizontal
axes
show
brightness
values
y
for
the
0.45-
and
1.25-µm
channels,
respectively.
As
an
example,
we
transform
these
data
to
predict
the
y
values
for
a
solar
zenith
angle
of
45°
(bottom).
Scene
dimming
is
most
obvious
in
the
lobe
corresponding
to
snow.
Combining
such
histograms
in
proportion
to
the
terrain
type
lets
an
analyst
predict
future
pixel
brightness
for
any
anticipated
mixture
of
terrain
types
and
observation
geometries.
C.
Threshold
Selection
Our
approach
cannot
predict
future
observations
exactly;
it
only
gives
probability
distributions
over
the
brightness
of
future
clouds
and
terrain.
We
account
for
this
uncertainty
in
our
thresholds
using
Bayesian
decision
theory
[36].
Recall
that
y
is
a
vector
of
brightness
values
across
several
spectral
channels
and
that
ci
and
c
2
are
clear
and
cloudy
cases,
respectively.
We
define
a
loss
function
with
a
penalty
ce
FN
for
false
negatives
(clouds
that
pass
the
filter)
and
a
separate
penalty
a
FP
for
false
positives
(clear
scenes
that
are
wrongly
excised).
The
total
expected
loss
E[G]
accounts
for
both
penalties
E[G]
=
f
a
F
pP(ci
I
y,
x)dy
f
aFNP(c2
I
y
,x)dy
Rd
\R.
(4)
30000
25000
,now
.7,
20000
a
FP
=
1000
15000
a
FP
=
10
`r
,
10000
5000
5000
10000 15000
20000 25000
1.251am
channel
intensity
30000
25000
7,
20
-
000
Sno
15000
a
FP
=1000
a
n,
=
10
10000
5000
errain
5000
10000 15000
20000 25000
1.25tim
channel
intensity
Fig.
3.
Brightness
distributions
in
0.45-
and
1.24-pm
channels
for
AVIRIS-C
image
f100521t02p05,
a
scene
that
contains
both
clouds
and
snow.
(Top)
Original
image
with
a
solar
zenith
angle
of
18°.
(Bottom)
Synthetic
distribution
after
transforming
the
image
to
a
solar
zenith
angle
of
45°.
Optimal
thresholds
are
shown
for
aggressive
cloud
screening
(a
FP
=
10)
and
conservative
cloud
screening
(a
FP
=
1000).
This
serves
as
a
figure
of
merit;
operators
simply
choose
the
channel
threshold
combination
with
the
lowest
expected
loss.
Following
Merchant
et
al.
[31],
we
rewrite
the
probability
of
the
cloud
case
ci
using
Bayes'
rule
P(y
I
x,ci)P(x
I
ci)P(ci)
Assuming
that
the
background
state
is
independent
of
the
cloud
probability,
we
have
P(x
I
ci)
=
P(x).
We
ignore
the
P(y
x)
term,
which
is
the
same
for
both
cloudy
and
clear
cases,
leaving
P(ci
I
y,x)
oc
P(y
I
x,
ci)P(ci).
(6)
The
two
possible
cases
are
clouds
ci
and
clear
sky
c
2
.
This
permits
the
following
decomposition:
E[G]
=
f
aFpP(y
I
x,
ci)P(ci)dy
f
aFNP(Y
I
x,
c2)P(c2)d3r.
(
7
)
Rdvre
P(ci
I
y,
x)
=
P(y
I
x)P(x)
(
5
)
6783
THOMPSON
et
aL:
RAPID
SPECTRAL
CLOUD
SCREENING
ONBOARD
AIRCRAFT
AND
SPACECRAFT
One
can
minimize
this
loss
using
any
nonlinear
optimization
method
appropriate
for
the
chosen
representation
of
P(y
I
x,
c).
The
proposed
multidimensional
histogram
representa-
tion
of
P(y
I
x,
c)
permits
a
direct
grid
search;
the
integrated
expected
loss
is
a
cumulative
sum,
computable
with
a
fast
recursive
operation.
More
generally,
gradient
descent
could
be
used
to
find
a
locally
optimal
threshold.
Using
11
to
denote
the
subspace
excluding
channel
v
and
O
n
the
set
of
points
on
the
decision
boundary
in
the
subspace
excluding
channel
v,
the
error
gradient
with
respect
to
a
specific
threshold
on
chan-
nel
v
is
=
f
aFpP(y
I
x,
ci)P(ci)dYn
d
f
aFNP(Y
I
x,
c2)P(c2)dYn•
(8)
This
permits
minimization
using
the
Newton
method
or
another
gradient-based
approach.
Fig.
3
shows
optimal
decision
boundaries
for
the
test
image
with
ce
FN
=
1.
The
two
thresholds
correspond
to
a
lenient
case
where
a
F
p
=
10
and
a
strict
case
where
a
F
p
=
1000.
The
scene
contains
both
clouds
and
snow;
lobe
of
the
background
distribution
corresponding
to
snow
features
has
high
brightness
in
the
0.45-µm
channel
but
low
brightness
in
the
1.25-µm
channel.
Consequently,
the
best
decision
boundary
carves
out
a
rectangular
exclusion
region
1Z.
The
optimal
thresholds
sig-
nificantly
vary
depending
on
geometry.
For
the
original
image
with
the
more
lenient
false
negative
penalty
aFN
=
10,
they
are
11
800
and
10
000
for
the
0.45-
and
1.25-µm
channels,
respectively.
For
the
stricter
case
of
ceFN
=
1000,
they
become
15
500
and
11
500,
focusing
on
the
fraction
of
cloud
pixels
that
are
completely
unambiguous.
In
the
dimmer
scene,
the
best
thresholds
are
9400
(0.45
µm)
and
8700
(1.25
µm)
at
ce
FN
=
10,
moving
to
12
200
and
9800
at
ce
FN
=
1000.
After
selection
of
optimal
threshold
values
for
a
new
obser-
vation,
the
flight
hardware
performs
these
tests
once
per
pixel,
designating
any
pixel
that
exceeds
all
thresholds
as
"cloudy."
D.
Spatial
Aggregation
The
pixel
classification
may
mislabel
isolated
bright
ter-
rain
pixels
such
as
anthropogenic
features,
sun
glint,
or
other
scene
clutter.
Such
localized
errors
can
be
addressed
by
spatial
smoothing,
with
methods
such
as
the
adjacency
tests
of
the
MODIS
approach
[17],
spatial
features
[20],
or
even
image
segmentation
[37].
Not
all
of
these
remedies
are
suited
for
real-
time
processing
since
instrument
buffers
can
only
store
a
small
portion
of
the
image
at
one
time.
Here,
we
evaluate
a
simple
spatial
aggregation
method
suited
for
real-time
execution
in
instrument
hardware.
It
operates
on
a
small
number
of
buffered
lines
simultaneously
and
makes
an
aggregate
decision
about
whether
to
keep
or
excise
the
block.
A
spatial
coverage
threshold
determines
the
number
of
cloudy
pixels
that
will
cause
a
vertical
block
to
be
excised.
This
spatial
aggregation
gives
additional
resilience
to
small
localized
bright
patches
or
single-pixel
artifacts.
We
use
a
spatial
coverage
procedure
SETTHRESHOLDS(aFp)
for
all
timestep
t
E
T}
do
calculate
surface
type
xt
calculate
solar zenith
angle
O
t
calculate
P
t
(y
el)
>
via
Equation
3
calculate
Pt(y
xt
0
2)
>
via
Equation
3
end
for
P(y
x,
c)
=
+
Et
Pt(Y
x
t
,
c)
ch
=
argmin
c
,E[G]
>
via
Equation
7
end
procedure
Fig.
4.
Algorithm
for
threshold
selection,
performed
offline.
Analysts
must
specify
the
false
positive
penalty
aFp
Pixel
classification
yi>
4
'?
Y2>
(1)
2
AND
Cloudy
Fig.
5.
Real-time
cloud
excision
algorithm
(onboard).
Here,
b
represents
the
minimum
number
of
cloud
pixels
that
triggers
an
excision.
Fig.
6.
AVIRIS-C
image
f100521t02p05,
which
contains
both
clouds
and
snow.
threshold
between
25%
and
50%.
Our
rationale
is
that
good-
quality
data
can
sometimes
be
recovered
from
images
with
less
cloud
cover
than
this
amount,
but
images
with
cloud
cover
greater
than
50%
would
almost
never
be
used.
Optionally,
each
block
can
be
subdivided
horizontally
into
two
or
more
subblocks,
with
separate
keep/reject
decisions
for
each.
This
finer
spatial
resolution
can
potentially
preserve
more
of
the
good-quality
data
near
the
clouds.
Fig.
4
shows
a
pseudocode
for
the
threshold-setting
proce-
dure.
Fig.
5
shows
the
real-time
portion
as
a
block
diagram
with
pixel-level
and
spatial
aggregation
thresholds.
Fig.
6
then
illustrates
the
result
for
the
image
associated
with
the
cloud
and
snow
distributions
previously
presented
in
Fig.
3.
This
d
E[G]
Spatial
threshold
(Cloudy)
>
b
?
yes
Excise
6784
IEEE
TRANSACTIONS
ON
GEOSCIENCE
AND
REMOTE
SENSING,
VOL.
52,
NO.
11,
NOVEMBER
2014
is
AVIRIS-C
image
f100521t02p00r05,
acquired
over
Senator
Beck
Basin,
Colorado,
on
May
21,
2010.
The
left
panel
shows
the
original
scene.
The
channel
threshold
operation
labels
cloudy
pixels,
shown
as
darkened
areas
in
the
middle
panel.
The
right
panel
shows
spatial
aggregation,
which
flags
a
segment
of
the
image
for
excision.
DI
EVALUATION
ON
HISTORICAL
AVIRIS-C
DATA
We
evaluated
the
method's
performance
using
the
AVIRIS-C
instrument's
2009-2011
data
archive
[32].
In
this
period,
AVIRIS-C
flew
throughout
North
America
on
campaigns
re-
lated
to
engineering
and
calibration,
mineralogy,
ecology,
and
disaster
response.
AVIRIS-C
operators
prefer
to
fly
on
clear
days,
which
reduces
the
total
fraction
of
cloud
cover.
However,
many
images
contained
clouds,
and
this
period
provides
a
rich
data
set
to
test
the
cloud-screening
approach.
The
AVIRIS-C
VSWIR
imaging
spectrometer
has
0.01-µm
resolution
in
224
bands
from
0.38
to
2.5
pm.
All
AVIRIS-C
images
have
680
samples
and
an
along-track
dimension
that
ranges
from
several
thousand
to
over
ten
thousand
image
lines.
A.
Data
Set
Much
previous
work
measures
cloud-screening
performance
by
correlating
observations
with
dedicated
cloud/aerosol
sen-
sors
[15],
[18]
or
with
authoritative
standards,
i.e.,
the
MODIS
cloud
mask
[38].
In
our
case,
the
AVIRIS-C
spatiotemporal
footprints
rarely
overlapped
with
other
cloud
sensors.
Instead,
we
evaluated
performance
by
labeling
every
cloud
pixel
in
the
archive
by
hand.
We
used
only
flights
above
10
000-m
altitude
for
maximum
fidelity
to
orbital
instruments.
We
inspected
this
entire
catalog
of
raw
uncalibrated
instrument
data,
marking
every
pixel
of
each
image
manually
as
either
"cloudy"
or
"clear"
to
form
a
ground
truth
cloud
classification.
We
would
typically
perform
this
labeling
in
an
image
editor
by
first
thresholding
each
image
and
then
cleaning
any
misclassified
points
using
a
manual
paint
tool.
We
only
labeled
clouds
that
were
opaque,
i.e.,
that
completely
obscured
the
terrain
color
and
texture
underneath.
This
was
a
strict
criterion
and
occasionally
left
ambiguous
translucent
pixels
around
the
edges
of
labeled
clouds.
We
prevented
edge
pixels
from
contaminating
the
"clear"
class
during
the
evaluation
by
disregarding
any
terrain
spectrum
within
10
pixels
of
a
cloud.
Our
rationale
was
that
either
classification
would
be
reasonable
for
these
ambiguous
cases;
hence,
they
should
not
count
for
either
credit
or
penalty.
Such
cases
constituted
a
tiny
fraction
of
the
data
set,
leaving
plenty
of
data
points
for
our
evaluation.
We
checked
our
data
set's
representativeness
using
typical
land-cover
taxonomies.
The
14-class
University
of
Maryland
(UMD)
system
[39]
is
attuned
to
biosphere
and
climate
re-
search,
but
its
categories
correlate
with
surface
reflectance
and
can
therefore
bear
on
the
observed
spectral
properties.
We
determined
UMD
categories
for
each
image
using
onboard
nav-
igation
telemetry
to
find
the
start
location
in
Global
Positioning
System
(GPS)
coordinates.
We
matched
this
location
to
land
types
recorded
in
the
MODIS
Collection
5
global
land
coverage
products
[40].
Table
I
shows
the
amount
of
data
assigned
TABLE
I
LAND
AND
OCEAN
COVER
REPRESENTED
IN
AVIRIS-C
DATA
SETS
Land
cover
Clear
pixels
Cloudy
pixels
Source
Water
6.7
x
10
8
2.2
x
10
7
UMD
Evergreen
needleleaf
forest
3.7
x
10
7
1.4
x
10
6
UMD
Evergreen
broadleaf
forest
5.8
x
10
6
5.8
x
10
5
UMD
Deciduous
needleleaf
forest
UMD
Deciduous
broadleaf
forest
2.1
x
10
8
4.4
x
10
6
UMD
Mixed
forest
1.3
x
10
8
3.7
x
10
6
UMD
Closed
shrublands
1.1
x
10
6
6.5
x
10
5
UMD
Open
shrublands
1.3
x
10
8
1.8
x
10
6
UMD
Woody
savannas
1.3
x
10
8
2.6
x
10
6
UMD
Savannas
UMD
Grasslands
8.9
x
10
7
3.1
x
10
6
UMD
Croplands
1.0
x
10
8
7.3
x
10
6
UMD
Urban
and
built-up
3.1
x
10
7
2.9
x
10
4
UMD
Snow
and
ice
1.3
x
10
8
4.4
x
10
6
Barren
9.7
x
10
7
1.7
x
10
2
Ocean
glint
3.6
x
10
8
1.5
x
10
7
to
each
category.
Urban
and
cropland
areas
were
particularly
well
represented.
A
large
fraction
of
AVIRIS-C
images
were
acquired
over
the
Gulf
Coast
ocean
due
to
extensive
operations
in
this
area
related
to
the
2010
Gulf
Oil
Spill.
The
land
types
differed
from
the
global
distributions
but
included
instances
from
nearly
all
of
the
UMD
land-cover
categories.
We
removed
some
specific
and
exceptional
image
features.
Except
where
noted,
we
excluded
images
containing
"sun
glint"
effects
since
they
would
not
be
used
for
most
applications.
We
also
excluded
a
set
of
images
of
the
White
Sands
National
Mon-
ument
in
Southern
New
Mexico,
a
feature
composed
of
nearly
pure
gypsum
that
is
highly
reflective
across
all
wavelengths.
This
feature
is
an
unusual,
if
not
globally
unique,
phenomenon
[41].
Finally,
we
removed
several
scenes
with
opaque
smoke
from
forest
fires
where
the
"correct"
answer
was
ambiguous.
This
left
a
data
set
of
507
images
included
in
the
study.
B.
Channel
Selection
The
AVIRIS-C
data
set
gave
insight
into
the
information
pro-
vided
to
the
cloud-screening
decision
by
different
combinations
of
channels.
As
apparent
in
Fig.
1,
snow
was
highly
reflective
in
visible
wavelengths
but
dark
in
the
SWIR.
Conversely,
bare
terrain
that
was
bright
at
SWIR
wavelengths
was
significantly
dimmer
than
clouds
in
VNIR
ranges.
This
favored
using
chan-
nels
in
both
regions.
Mutual
information
(MI)
indicates
the
information
provided
by
different
channel
combinations.
MI
is
a
quantity
from
in-
formation
theory
describing
the
information
value
of
an
obser-
vation
with
respect
to
another
unknown
variable
[42].
Specifi-
cally,
it
quantifies
how
knowledge
of
one
reduces
the
Shannon
entropy
in
the
other.
In
our
case,
MI
related
the
knowledge
of
frequency
channels
to
the
binary
cloud/clear
classification.
We
computed
the
MI
using
the
following
expression:
E
P(ci,
y)
log
(
P(ci)P(y)
yEY
+Ep(c2,y)
log
(
c2y)
p
p(
(c2)
,
(y)
)
(
9
)
yEY
0
Cloud
(0.45
pm)
Cloud
(1.65
ilm)
A
Tenuin
(0.45
pm)
V
Terrain
(1.65
tun)
-
Model
6785
THOMPSON
et
aL:
RAPID
SPECTRAL
CLOUD
SCREENING
ONBOARD
AIRCRAFT
AND
SPACECRAFT
TABLE
II
MI
OF
SELECTED
CHANNELS
FOR
CLOUD
SCREENING
OVER
LAND
SURFACES
(LARGER
IS
BETTER).
THE
FIRST
DATA
COLUMN
SHOWS
THE
INFORMATION
PROVIDED
BY
A
SINGLE
CHANNEL
TAKEN
ALONE.
OTHER
ENTRIES
SHOW
THE
VALUE
OF
TWO-CHANNEL
COMBINATIONS.
REFERENCES:
1
ACKERMAN
ETAL
[17];
2
GAO
ETAL
[35];
3
FRIEDL
ETAL.
[40];
AND
4
GOMEZ-CHOVA
ETAL
[20]
Channel
MI
M1
when
combined
with
isocro.
0.66
pm
0.86
pm
1.25
pm
1.38
pm
1.65
pm
Ref.
0.45
pm
0.58
0.66
pm
0.48
0.86
pm
0.39
0.61
0.59
0.54
0.63
0.55
0.47
0.59
0.51
0.46
0.63
0.55
0.49
4
1,3,4
1,3,4
I
;I:
10000.
1.25
pm
0.38
0.47
0.45
1,2,3
1.38
pm
0.20
0.41
1,2,3
5000.
1.65
pm
0.26
1,3
Ocean
TABLE
BI
MI
OF
SELECTED
CHANNELS
FOR
CLOUD
SCREENING
OVER
OCEAN
SURFACES
(LARGER
IS
BETTER).
REFERENCES
ARE
IN
TABLE
II
CAPTION
5000
10000 15000
20000
1.65nm
channel
intensity
30000.
25000.
20000
.
Fig.
7.
Channel
MI
MI
when
combined
with
0.66
pm
0.86
pm
1.25
pm
1.38
pm
1.65
pm
Ref.
0.45
pm
0.61
0.63
0.64
0.63 0.63
0.62
4
0.66
pm
0.60
0.62
0.62
0.62
0.60
1,3,4
0.86
pm
0.46
0.52
0.48
0.50
1,3,4
1.25
pm
0.48
0.48
0.49
12,3
1.38
pm
0.10
0.48
1,2,3
1.65
pm
0.43
1,3
4
TABLE
IV
MI
OF
SELECTED
CHANNELS
FOR
CLOUD
SCREENING
OVER
SCENES
CONTAINING
SNOW
AND
ICE
(LARGER
IS
BETTER).
REFERENCES
ARE
IN
THE
TABLE
II
CAPTION
Ml
when
combined
with
Channel
MI
0.66
pm
0.86
pm
1.25
pm
1.38
pm
1.65
pm
Ref.
2
0.45
pm
0.45
0.50 0.50
0.63
0.59
0.63
4
0.66
pm
0.44
0.51
0.62
0.59
0.62
1,3,4
0.86
pm
0.43
0.61
0.58
0.61
1,3,4
1.25
pm
0.54
0.60
0.58
1,2,3
1.38
pm
0.54
0.61
1,2,3
1.65
pm
0.40
1,3
Dark-subtracted
brightness
distributions
at
solar
zenith
angle.
x
10
4
2
1.8
l.6
1.4
1.2
0.8
0.6
0.4
0.2
20
30
40
50
60
70
Solar
zenith
0
where
y
was
the
domain
of
y.
Tables
II
and
111
show
scores
for
land
(368
flight
lines)
and
ocean
(139
flight
lines),
respectively.
For
these
data
sets,
the
0.45-µm
channel
was
the
strongest
single
indicator
of
clouds.
Table
IV
shows
scores
for
images
containing
snow
and
ice
(43
flight
lines).
The
cirrus
channel
at
1.38
urn
was
less
valuable
for
this
data
set
because
high-altitude
translucent
clouds
were
not
counted.
If
we
also
sought
to
screen
cirrus
clouds,
this
channel
would
have
been
more
important.
While
the
0.45-µm
channel
had
the
highest
overall
MI,
it
was
not
obvious
which
other
channel
made
its
best
partner.
Combinations
with
SWIR
wavelengths
gave
the
largest
net
im-
provement,
particularly
for
snow
and
ice
scenes.
The
1.25-µm
channel
had
been
used
previously
for
snow/ice
tests
in
produc-
tion
systems
such
as
the
EO-1
screening
algorithm
[26],
where
it
exploited
the
low
SWIR
absorption
of
snow
and
the
low
brightness
of
bare
terrain
in
this
spectral
region.
However,
for
our
system
that
considered
multiple
channels
simultaneously,
sensitivity
to
bare
terrain
provided
no
advantage.
The
0.45-µm
channel
already
discriminated
everything
but
snow,
for
which
either
SWIR
frequency
worked
well.
Consequently,
the
MI
of
the
[0.45
µm,
1.25
µm]
pairing
and
the
MI
of
the
[0.45
µm,
1.65
µm]
pairing
were
basically
indistinguishable.
We
return
to
this
question
in
the
following
section,
where
simulations
Fig.
8.
Cloud
and
terrain
brightness
as
a
function
of
solar
zenith
angle
for
all
images
over
land.
Instrument
DNs
have
been
corrected
for
dark
current
levels.
incorporating
spatial
information
show
a
slight
advantage
to
the
1.65-µm
channel.
Fig.
7
shows
the
combined
distributions
for
land,
ocean,
and
cloud
pixels,
with
values
given
as
dark-subtracted
data
numbers
translated
to
solar
zenith
angle.
The
terrain
distributions
had
axis-parallel
lobes
corresponding
to
barren
terrain
and
snow.
There
is
some
mixing
between
the
ocean
and
land
distributions
in
this
image.
We
labeled
each
scene
with
its
dominant
land
cover
class,
so
ocean
scenes
were
occasionally
contaminated
by
islands
and
shorelines.
There
was
also
some
natural
overlap
due
to
the
ambiguity
between
opaque
cloud,
which
was
labeled,
and
haze
or
thin
semitransparent
cloud,
which
was
not.
Fig.
8
compares
cloud
and
surface
brightness
to
predic-
tions
from
the
idealized
cos
(9)
illumination
falloff.
We
dark
subtracted
the
raw
instrument
values
and
corrected
them
for
Earth-Sun
distance.
We
then
binned
the
solar
zenith
angles
in
10°
increments
and
plotted
the
mean
of
the
resulting
image
pixels
in
each
zenith
bin.
Fig.
8
shows
the
result,
with
brightness
as
a
function
of
solar
zenith
angle,
with
a
solid
line
indicating
the
best
fitting
cosine-proportional
curves.
There
was
some
f
I
0.45
pm
and
1.65
pin
0.45
pm
and
1.25
pin
0.45,
1.25,
and
1.65
pm
6786
IEEE
TRANSACTIONS
ON
GEOSCIENCE
AND
REMOTE
SENSING,
VOL.
52,
NO.
11,
NOVEMBER
2014
natural
deviation
from
the
ideal
due
to
the
finite
data
set
size,
diversity
of
terrain types,
and
diversity
of
zenith
angles
within
each
bin.
However,
the
overall
result
corroborates
the
cosine
solar
zenith
angle
relationship
for
both
cloud
and
land
surfaces.
C.
Predictive
Thresholding
After
forming
2-D
histograms,
we
calculated
thresholds
for
pairings
of
the
0.45-µm
channel
with
either
the
1.25-
or
1.65-µm
channel.
We
separated
the
data
into
a
training
half
used
to
set
thresholds
and
a
test
half
used
for
evaluation.
We
then
transformed
the
training
half
to
a
TOA
reflectance
representation
to
construct
surface
and
cloud
brightness
distri-
butions.
We
calculated
bias
factors
using
an
automatic
closed-
shutter
pre-
and
postcalibration
segment
embedded
in
the
raw
image.
These
biases
were
estimated
once
on
a
single
calibration
image
and
then
uniformly
applied
across
the
whole
data
set,
relying
on
the
stability
of
the
instrument
as
would
be
the
case
during
extended
autonomous
operations.
We
tested
a
range
of
operating
regimes
with
different
false
alarm
costs.
Our
most
aggressive
setting
held
false
alarm
costs
equal
to
the
false
negative
cost.
This
setting
would
maximize
the
compression
rate;
it
would
be
appropriate
for
measurements
such
as
large-scale
atmospheric
profiling.
At
the
other
extreme,
we
set
false
alarm
penalties
to
be
a
factor
of
10
5
greater
resulting
in
very
conservative
cloud
screening
that
preserved
good
data
at
all
costs.
This
would
be
appropriate
for
scheduled
acquisitions
over
specific
high
value
targets.
For
each
false
alarm
cost,
we
used
an
exhaustive
grid
search
to
select
the
best
channel
thresholds.
This
gave
a
set
of
channel
threshold
pairs
representing
the
envelope
of
optimal
parameter
settings.
Initially,
we
calculated
all
thresholds
in
the
normalized
TOA
representation.
We
applied
them
to
each
new
test
scene
by
translating
back
to
the
appropriate
raw
instrument
values,
trans-
forming
brightness
according
to
the
test
scene's
solar
zenith
angle.
We
then
applied
cloud
screening
with
a
25%
spatial
cov-
erage
threshold.
We
evaluated
several
granularities
of
spatial
aggregation,
subdividing
each
along-track
segment
of
32
lines
into
one,
two,
or
four
cross-track
blocks.
We
also
sought
to
determine
the
potential
for
further
im-
provement
by
using
more
sophisticated
classifiers.
To
this
end,
we
evaluated
the
use
of
three
channels
in
different
combina-
tions
of
VNIR
and
SWIR
wavelengths.
The
three-channel
tests
used
the
same
solar
normalization
strategy
with
an
exhaustive
grid
search
to
compute
Bayes-optimal
thresholds.
Even
more
sophisticated
classifiers
are
possible,
e.g.,
a
linear
decision
boundary,
suitably
stripped
of
floating-point
arithmetic,
ar-
guably
meets
the
real-time
hardware
requirements.
This
classi-
fier
calculates
a
prediction
for
each
pixel
independently
by
first
applying
the
solar
zenith
correction
and
then
forming
a
linear
combination
of
all
channels.
As
before,
we
smooth
the
pixel-
level
classifications
using
a
spatial
aggregation
test.
We
trained
the
linear
decision
boundary
on
the
held-out
half
of
the
data,
randomly
sampling
from
cloud
and
background
distributions.
Training
used
a
stochastic
gradient
descent
algorithm
minimiz-
ing
a
sum
squared
error
objective
[43].
0.99
0.98
0
0.97
F
,
0.96
0
0
t
,
0.94
2
a
0.93
.2
0.92
0.91
0.9
0
Fig.
9.
Performance
on
AVIRIS-C
2009-2011
test
scenes.
We
evaluated
two
channel
combinations,
with
and
without
an
SZA.
D.
Evaluation
The
simulation
provided
two
performance
values
for
each
false
alarm
tolerance,
i.e.,
a
false
alarm
rate
giving
the
fraction
of
good
data
accidentally
deleted
and
a
true
positive
rate
giving
the
fraction
of
cloudy
image
blocks
that
were
successfully
removed.
We
defined
false
negatives
as
any
blocks
containing
greater
than
50%
cloud
that
were
not
excised.
An
excised
block
was
a
false
positive
if
it
did
not
contain
any
significant
clouds
(less
than
5%).
This
left
a
range
between
5%
and
50%
cloud
cover
where
neither
excision
nor
abstention
was
penalized.
The
free
range
was
necessary
since
cloud
edges
were
often
ambiguous
making
the
precise
fractional
coverage
of
small
image
areas
somewhat
subjective.
We
evaluated
performance
first
using
separate
thresholds
for
land
and
ocean
scenes
and
then
for
a
combined
data
set
which
ignored
surface
type.
Finally,
we
simulated
a
hypothetical
space
mission
to
evalu-
ate
data
compression
rates.
We
calculated
a
year's
orbits
of
the
ISS
at
10-min
time
step
intervals,
recording
at
each
time
step
the
solar
zenith
angle
and
the
terrain
type
under
observation.
Using
the
heuristic
of
Eastman
et
al.
[12],
which
was
the
lowest
of
recent
estimates,
we
conservatively
predicted
that
68%
of
water
scenes
and
54%
of
land
scenes
would
be
cloudy.
Given
that
instrument
operators
were
able
to
set
thresholds
with
fore-
knowledge
of
the
surface
type,
we
applied
the
appropriate
land
or
ocean
performance
values
in
proportion
to
their
appearance
in
each
orbit.
This
gave
the
expected
fraction
of
each
orbit
that
could
be
excised
for
a
desired
false
positive
level.
We
assumed
that
the
instrument
would
be
operating
whenever
the
local
solar
zenith
was
less
than
75°.
E.
Results
and
Discussion
Figs.
9-13
show
cloud-screening
performance
over
land
and
ocean,
as
well
as
for
the
combined
data
set
that
includes
sun
glint.
We
report
results
using
a
receiver
operating
characteristic
(ROC)
curve
plotting
false
alarm
and
true
positive
rates
[44].
0.95
2
3 4
5
False
positive
rate
x
10
-3
0.99
0.98
0.97
0.96
0
0.95
ro
›,
0.94
0.93
0.92
2
3
False
positive
rate
0.91
0.9
0
5
x10
-
3
1
Land
Ocean
-
Combined
•••
I
Two-channel
threshold
-
Linear
classifier
Screen
ing
e
ffic
iency
(
fr
ac
t
ion
o
f
c
lou
dy
bloc
ks
exc
ise
d)
0.99
0.97
0.96
0.95
0.94
0.93
0.91
0.98
0.92
0.9
0
2
3
False
positive
rate
4
5
x
10
-3
/
Normalized
- - -
Raw
0.99
O
0
0.98
0.97
0.96
0.95
›,
0.94
0.93
0.92
0.91
0.9
0.99
0.98
I
0.97
0.96
0.95
0.94
0.93
0.92
I
0.9
0
2
3
False
positive
rate
0.91
5
x10
-3
Full
swath
Half
swath
-
Quarter
swath
Fig.
13.
Performance
on
AVIRIS-C
2009-2011
test
scenes.
0.45-µm
channel
combination
at
0.1%
false
positive
ranges,
the
solar
correction
halved
the
false
positive
rate.
However,
per-
formance
was
surprisingly
good
even
without
solar
correction.
We
concluded
that
illumination
was
benign
for
most
AVIRIS-C
images
in
our
catalog,
and
effects
of
solar
variability
were
less
important
than
large
intrinsic
differences
in
surface
and
cloud
reflectance.
The
experiments
revealed
minor
differences
between
1.25-
and
1.65-µm
channels.
Many
cloud-screening
algorithms
favor
the
1.25-µm
channel
as
an
exclusion
test
for
snow
and
ice,
and
we
also
found
it
effective
for
this
purpose.
The
1.25-µm
channel
can
be
also
used
to
exclude
other
dark
terrain
types.
However,
when
combined
with
a
0.45-µm
channel,
this second
role
was
completely
redundant,
and
either
SWIR
channel
made
an
effective
pairing.
In
fact,
snow
absorption
was
even
stronger
at
1.65
µm
consistent
with
studies
by
Painter
et
al.
[45].
Consequently,
the
1.65-µm
channel
outperformed
1.25
µm
over
Sc
reen
ing
e
ffic
iency
(
frac
t
ion
o
f
c
lou
dy
bloc
ks
exc
ise
d)
THOMPSON
et
aL:
RAPID
SPECTRAL
CLOUD
SCREENING
ONBOARD
AIRCRAFT
AND
SPACECRAFT
6787
Fig.
10.
Performance
on
AVIRIS
-
C
2009
-
2011
test
scenes.
Fig.
12.
Performance
on
AVIRIS
-
C
2009
-
2011
test
scenes.
2
3
4
5
False
positive
rate
x
10
-3
Fig.
11.
Performance
on
AVIRIS-C
2009-2011
test
scenes.
The
ROC
curve
represents
the
envelope
of
performance
that
can
be
achieved;
designers
can
move
along
it
by
changing
the
channel
thresholds
to
be
more
or
less
aggressive.
Desirable
performance
lies
in
the
upper
left,
with
many
excised
clouds
and
few
false
positives.
We
show
performance
for
both
channel
combinations,
with
and
without
solar
zenith
adjustment
(SZA).
The
thick
gray
line
in
each
plot
marks
the
performance
of
our
reference
design,
i.e.,
a
two-channel
cloud-screening
algorithm
operating
over
land.
Performance
is
consistent
with
the
work
by
Williams
et
al.
on
a
threshold-based
FPGA
system
combining
VSWIR
and
thermal
channels
[10];
they
report
0.02%-0.09%
missed
clouds
at
a
0%-0.99%
false
detection
rate,
which
also
lies
on
this
ROC
curve.
Our
reference
design
uses
the
solar
zenith
correction.
We
found
that
the
correction
nearly
always
improved
performance,
particularly
in
the
case
of
the
combined
data
set,
including
diverse
land
cover,
ocean,
and
glint
scenes.
For
the
1.65-
and
I I I
I I I
I I I
I
I I I
I I I
I
I I I
I I I
I I I
I I
I I I
I I I
I I I
I I I
I I I
1 1 1
lc-5
lc-4
le-3
lc-2
Fraction
of
good
data
excised
6788
IEEE
TRANSACTIONS
ON
GEOSCIENCE
AND
REMOTE
SENSING,
VOL.
52,
NO.
11,
NOVEMBER
2014
0.65
0.6
7
,
0.55
0.5
0.45
Fig.
14.
Performance
on
ISS
orbits,
disregarding
scenes
with
significant
ocean
glint.
land.
All
methods
performed
equally
well
over
ocean.
We
also
considered
using
three
channels
simultaneously
and
found
that
the
combination
of
0.45,
1.25,
and
1.65
µm
provided
a
slight
additional
benefit.
Other
triples
failed
to
match
the
performance
of
the
0.45-
and
1.65-µm
pairings
and
are
excluded
from
the
plot.
Adding
a
fourth
channel
might
improve
performance
slightly
more,
although
the
histogram
representation
would
be
nearly
a
terabyte
in
size.
A
more
practical
approach
is
to
incorporate
more
channels
using
the
more
traditional
linear
classifier.
Fig.
12
compares
its
performance
to
our
reference
design.
The
linear
classifier
performs
best
of
all
in
the
extreme
low
false
positive
regime,
excising
up
to
97%
of
clouds
with
negligible
loss
of
science
data.
This
suggests
that
still
better
performance
is
possible
if
the
instrument
hardware
can
support
the
required
multichannel
operations.
Finally,
Fig.
13
shows
the
performance
for
different
spatial
aggregation
strategies.
Subdividing
the
swath
horizontally
into
smaller
spatial blocks
harmed
classification
performance
since
small
isolated
bright
terrain
areas
triggered
false
alarms
more
easily.
As
some
compensation,
each
block
was
smaller
and
screening
decisions
were
made
at
a
fine
spatial
resolution.
Therefore,
an
excised
cloudy
block
was
less
likely
to
contain
clear
pixels.
We
used
this
result
to
estimate
compression
potential
for
future
Earth-orbiting
missions.
Fig.
14
shows
compression
performance
based
on
land
and
ocean
fractions
observed
during
a
year
of
ISS
orbits.
We
assumed
that
cloud
screening
used
the
0.45-
and
1.65-µm
channels.
The
horizontal
axis
shows
the
fraction
of
good
data
that
was
accidentally
removed,
which
depends
on
the
false
alarm
tolerance
chosen
by
the
designer.
We
considered
false
positive
rates
ranging
from
0.001%
(one
such
error
every
hundred
thousand
blocks,
essentially
negligible)
to
0.01%.
The
data
reduction
achieved
on
each
orbit
then
depends
on
the
proportion
of
land
and
ocean
encountered.
We
calcu-
lated
the
reduction
using
the
performance
estimates
from
our
AVIRIS-C
study,
presuming
that
the
ground
control
could
apply
land
or
ocean
threshold
sets
over
the
appropriate
surface
types.
The
vertical
axis
shows
the
fraction
of
pixels
removed.
Boxes
show
the
median
and
quartiles,
with
dashed
whiskers
indicating
the
extent
of
the
most
extreme
orbits.
At
a
false
alarm
rate
of
0.001%,
cloud
screening
reduced
data
rates
by
approximately
60%.
Performance
never
rose
very
far
above
this
level,
even
for
very
lenient
instrument
settings,
and
the
overall
span
of
system
performance
was
less
than
the
natural
cross-orbit
variability.
The
strictest
threshold
achieved
a
rate
reduction
better
than
90%
of
the
theoretical
optimum.
This
analysis
excluded
ocean
scenes
containing
glint;
hence,
performance
could
be
different
for
the
special
case
of
cloud
screening
during
glint-mode
atmospheric
retrievals.
Our
analysis
also
ignored
latitudinal
differences
in
cloud
cover
and
land
type
[14].
Northern
latitudes
pose
the
special
challenge
of
the
finest
grained
snow,
which
is
highly
reflective
[45],
[46]
and
somewhat
rare
in
this
test
data
set.
If
needed,
missions
could
mitigate
these
effects
by
defining
a
specific
new
land-cover
type
devoted
to
snow
which
would
nat-
urally
lead
to
more
conservative
thresholds
over
such
regions.
IV.
AIRBORNE
DEMONSTRATION
This
section
describes
a
deployment
onboard
the
AVIRIS-NG
airborne
imaging
spectrometer
[33].
AVIRIS-NG
is
a
next-
generation
push
broom
instrument
that
measures
the
0.38-
2.5-µm
region
with
480
bands
at
5-nm
spectral
resolution.
Its
640
cross-track
samples
provide
spatial
resolution
values
of
1
m
or
better,
depending
on
altitude.
Minor
differences
in
spectral
sampling
should
not
affect
cloud
screening
since
both
clouds
and
bright
terrain
features
are
spectrally
smooth
in
the
wavelengths
of
interest.
The
onboard
cloud
screening
system
used
a
commensal
data
path
that
ran
in
parallel
alongside
the
science
data
acquisition
and
synchronization
process.
The
primary
data
handling
path
used
a
Virtex-5
FPGA
connected
to
the
instrument
through
a
Camera
Link
interface
and
to
the
Inertial
Measurement
Unit
(IMU)/GPS
device
through
a
serial
and
digital
input
interface.
The
commensal
path
and
science
data
acquisition
and
synchro-
nization
were
housed
on
a
controller
board based
on
a
2.3-GHz
quad-core
Intel
i7
processor
from
National
Instruments.
This
processor
was
connected
through
a
64-Gb/s
PCI
Express
bus
to
a
1-TB
solid-state
drive
(SSD)
configured
as
RAIDO
to
achieve
a
read/write
bandwidth
of
6.4-Gb/s
throughput.
The
high
throughput
allowed
cloud
screening
to
access
the
SSD
simultaneously
with
the
processor's
own
acquisition
and
syn-
chronization
processes.
The
processor
board,
the
FPGA
board,
and
the
SSD
were
integrated
into
a
PM
chassis
from
National
Instruments.
A
GUI
was
designed
to
allow
the
user
to
control
the
data
acquisition
and
synchronization
process,
as
well
as
the
onboard
cloud
screening
system.
A.
Procedure
We
designed
the
system
to
implement
the
cloud
screening
algorithm,
computing
solar-normalized
cloud
thresholds
in
ad-
vance
using
the
archive
of
historical
AVIRIS-NG
test
flights.
We
used
these
data
to
calculate
optimal
thresholds
for
the
0.45-
and
1.65-µm
channels.
When
a
flight
began
the
real-time
system
monitored
the
file
system
for
new
data
and
immediately
THOMPSON
et
aL:
RAPID
SPECTRAL
CLOUD
SCREENING
ONBOARD
AIRCRAFT
AND
SPACECRAFT
6789
applied
cloud
screening.
It
started
by
first
reading
the
current
aircraft
location
from
the
IMU/GPS
information
synchronized
with
the
instrument
data
and
stored
with
the
images
on
the
SSD.
It
calculated
the
solar
ephemeris
and
adjusted
cloud
thresholds
to
the
new
observation
geometry.
The
system
then
analyzed
the
image
as
it
was
recorded,
producing
two
products,
i.e.,
a
real-time
operator
notification
displaying
the
average
cloud
fraction
of
the
scene
and
a
cloud
mask
image
recording
pixel-
level
classifications
for
later
analysis.
We
evaluated
the
system
during
a
12-day
science
campaign
over
Casper,
Wyoming,
USA,
in
June
2013.
This
campaign
collected
23
flight
segments
comprising
over
a
terabyte
of
raw
data.
The
images
showed
a
mixture
of
bare
terrain,
industrial
facilities,
and
vegetation.
While
the
terrain
content
was
diverse,
science
operations
required
clear
skies;
hence,
the
data
were
not
representative
of
global
cloud
and
clear-sky
distributions.
How-
ever,
in
one
flight
when
the
ground
observation
sortie
was
cut
short
by
clouds,
the
aircraft
intentionally
climbed
from
2290
m
above
mean
sea
level
(MSL)
to
4420
m
MSL
in
order
to
fly
above
the
rapidly
growing
cumuli.
B.
Results
Soon
after
reaching
the
desired
altitude,
the
software's
cloud-
fraction
display
abruptly
departed
from
0.0,
where
it
had
consistently
remained
during
the
earlier
portion
of
the
flight.
The
cloud-fraction
rapidly
climbed
to
1.0
while
flying
over
the
denser portions
of
the
clouds.
Post-analysis
revealed
that
the
system
had
successfully
identified
the
opaque
clouds
and,
when
compared
with
a
human
interpretation,
had
labeled
the
correct
segments
of
the
flight
line
for
excision.
Fig.
15
shows
the
result:
The
left
panel
is
the
original
image,
the
center
panel
shows
the
pixels
exceeding
both
thresholds,
and
the
right
panel
shows
the
excised
image
blocks.
Table
V
summarizes
the
entire
campaign.
Columns
show
the
time
and
the
number
of
images
in
each
batch,
the
total
number
of
image
lines,
the
fraction
of
pixels
that
were
flagged
as
cloudy
prior
to
spatial
aggregation,
and
the
final
cloud
screening
result,
i.e.,
whether
any
lines
were
excised
and
whether
clouds
were
actually
present.
In
cloudless
scenes,
isolated
structures
and
bright
objects
occasionally
exceeded
both
channel
thresholds.
This
occurred
for
a
fraction
of
pixels
from
0.02%
to
0.007%.
However,
these
small
pixel
areas
were
successfully
ignored
during
the
spatial
aggregation
step;
thus,
the
onboard
system
passed
all
clear-sky
images
without
modification
and
commit-
ted
no
false
excisions.
The
overall
result
was
that
cloud
screen-
ing
performed
without
error
and
kept
pace
with
the
sensor's
0.5-Gb/s
data
production
rate
throughout
the
campaign.
It
will
continue
to
be
deployed
on
future
flights.
V.
CONCLUSION
This
paper
has
described
a
novel
method
for
cloud
screening
onboard
spacecraft
at
Gb/s
data
rates.
We
perform
the
most
challenging
computations
on
the
ground,
exploiting
foreknowl-
edge
of
observation
geometry
and
surface
type
to
predict
the
brightness
of
terrain
pixels
and
cloud
pixels.
We
calculate
optimal
thresholds
in
uncalibrated
instrument
values
that
can
be
uploaded
for
real-time
execution
by
the
flight
system.
Analysis
Fig.
15.
AVIRIS-NG
image
ang20130625t174216.
The
images
at
center
and
right
show
the
result
of
onboard
real-time
cloud
screening.
on
a
three-year
archive
of
AVIRIS-C
images
demonstrates
that
this
simple
approach
can
reduce
instrument
data
volumes
by
a
factor
of
two
with
insignificant
loss
of
science
data.
A
deploy-
ment
onboard
the
AVIRIS-NG
platform
corroborates
this
per-
formance
in
blind
real-time
demonstrations.
The
testbed
system
will
continue
to
accumulate
additional
operational
experience
in
future
AVIRIS-NG
campaigns.
The
excised
data
are,
by
definition,
unrecoverable.
This
is
a
problem
for
investigations
that
are
wholly
intolerant
of
data
loss
or
that
study
the
clouds
themselves.
However,
having
implemented
the
screening
option,
designers
would
be
free
to
choose
the
threshold
level
most
appropriate
to
the
science
needs
of
each
observation.
This
contrasts
with
the
status
quo,
where
there
is
no
such
option
and
missions
cannot
escape
the
6790
IEEE
TRANSACTIONS
ON
GEOSCIENCE
AND
REMOTE
SENSING,
VOL.
52,
NO.
11,
NOVEMBER
2014
TABLE
V
AVIRIS-NG
FLIGHT
CAMPAIGN
RESULTS.
THE
lines
COLUMN
INDICATES
THE
NUMBER
OF
IMAGE
LINES
IN
EACH
GROUP.
THE
pixel
fraction
COLUMN
INDICATES
THE
PROPORTION
OF
INDEPENDENT
PIXELS
THAT
WERE
MARKED
AS
CLOUDY
DURING
THE
INITIAL
THRESHOLDING.
NONE
OF
THESE
RESULTED
IN
ANY
FALSE
POSITIVE
EXCISIONS
Date
Time
images
lines
pixel
fraction
excised
clouds
14
June
2013
14:43:44
1
960
0
N N
14
June
2013
22:52:28
4
186560
7.38
x
10
-5
N N
18
June
2013
20:25:42
2
4480
0
N N
18
June
2013
21:59:05
2
2880
0
N N
18
June
2013
22:59:41
3
70720
1.75
x
10
-3
N N
18
June
2013
23:18:56
1
32000
1.45
x
10
-3
N N
18
June
2013
23:26:57
2
36480
1.27
x
10
-3
N N
18
June
2013
23:36:12
1
24000
1.12
x
10
-3
N N
19
June
2013
15:56:23
1
4800
0
N N
19
June
2013
18:23:42
5
60480
1.55
x
10
-4
N N
19
June
2013
18:57:54
8
257920
6.58
x
10
-4
N N
20
June
2013
15:45:15
13
508160
3.01
x
10
-4
N N
20
June
2013
18:34:10
4
177280
3.36
x
10
-4
N N
21
June
2013
17:32:09
7
224000
4.96
x
10
-4
N N
21
June
2013
18:44:38
1
960
N N
21
June
2013
18:46:57
9
333440
3.14
x
10
-4
N N
23
June
2013
17:01:46
3
108800
1.76
x
10
-4
N N
23
June
2013
17:45:53
13
448320
2.09
x
10
-4
N N
24
June
2013
16:00:40
5
147840
3.00
x
10
-4
N N
24
June
2013
17:01:04
10
221120
2.06
x
10
-3
N N
24
June
2013
17:56:11
5
158080
1.41
x
10
-3
N N
25
June
2013
15:57:28
5
153600
3.94
x
10
-4
N N
25
June
2013
16:49:28
7
174720
6.32
x
10
-3
Y Y
Total
142
3847266
resource
cost
of
clouds.
Our
tests
suggest
that
high
rates
of
data
reduction
are
achievable
for
very
conservative
settings.
For
an
Earth-orbiting
spectrometer,
trivial
cases
such
as
open
ocean
yield
most
of
the
benefit,
and
very
strict
thresholds
can
excise
most
clouds
over
land
terrain.
We
anticipate
the
highest
risk
of
false
alarms
for
scenes
with
bright
snow
or
sun
glint.
As
compensation,
snow can
be
segregated
as
a
separate
terrain
type
and
tracked
over
time
using
snow-cover
products
from
other
missions.
As
an
extreme
measure,
one
could
simply
turn
cloud
screening
off
to
avoid
seasonal
areas
where
snow
could
appear.
Sun
glint
can
be
anticipated
from
imaging
geometry.
However,
many
investigations
consider
sun
glint
to
be
a
contaminant
similar
to
clouds.
The
specific
operational
concept
may
vary
depending
on
mission
needs.
Our
prototype
consists
of
a
real-time
component
requiring
streaming
Gb/s
processing
and
an
offline
component
having
effectively
no
restrictions
on
computation
or
communi-
cation.
While
this
is
an
idealized
view,
it
is
appropriate
for
real
space
operation
scenarios.
The
offline
computing
needs
would
not
be
a
significant
bottleneck
since
calculations
take
just
a
few
minutes
on
a
modern
laptop
computer.
High-performance
computing
resources
would
make
it
a
real-time
operation.
In
addition,
irregular
communication
is
not
an
obstacle
since
threshold
settings
could
be
computed
long
in
advance
based
on
the
known
observation
geometry.
Finally,
one
could
always
compromise
computing
or
communication
by
limiting
the
num-
ber
of
threshold
updates
and
applying
the
same
thresholds
over
longer
time
segments
spanning
multiple
terrain
types
and
solar
zenith
angles.
More
exotic
operational
concepts
are
possible.
If
the
spacecraft
or
instrument
can
be
pointed,
the
spectrometer
could
select
targets
based
on
the
cloud
screening
result,
perhaps
scanning
across
its
field
of
regard
until
a
suitable
clear
scene
is
found.
There
are
several
straightforward
ways
to
improve
accu-
racy.
One
could
use
three
or
more
channels
in
the
threshold
decision.
However,
it
is
likely
that
more
channels
will
of-
fer
diminishing
returns.
They
would
also
require
additional
histogram
dimensions,
which
quickly
become
intractable
as
the
number
of
channels
increases.
Future
work
could
seek
alternative
representations
that
scale
better
with
dimensionality.
A
more
promising
approach
would
be
to
incorporate
additional
domain
knowledge
into
the
state
vector.
One
could
condition
thresholds
on
very
specific
land
types
or
on
real-time
cloud
products
like
the
GOES
cloud
mask.
Preprocessing
and
feature
extraction
could
also
improve
performance.
For
example,
one
could
compute
spectral
derivatives;
sums,
differences,
or
ratios
of
channels;
or
continuum-relative
absorption
band
depths
[47].
Such
spectral
features
could
potentially
improve
results
at
a
low
computational
cost.
One
could
always
design
a
more
complex
cloud
classifier
to
disambiguate
the
most
difficult
pixels
and
consequently
achieve
a
slight
improvement
in
data
volume
reduction.
However,
our
simple
algorithm
already
achieves
better
than
90%
of
the
theoretical
maximum
making
it
sufficient
for
many
applica-
tions
and
a
useful
point
on
the
design
trade
space.
At
a
time
when
communications
and
storage
subsystems
struggle
under
increasing
data
rates,
the
potential
for
onboard
cloud
screening
has
remained
relatively
unstudied.
Mission
designers
should
bear
in
mind
that
a
few
simple
design
considerations,
i.e.,
the
introduction
of
channel
and
aggregation
thresholds,
can
enable
factor-of-two
reductions
in
data
volume.
ACKNOWLEDGMENT
The
authors
would
like
to
thank
C.
Sarture
and
the
AVIRIS
team
for
vital
assistance
in
performing
the
airborne
demon-
strations;
Dr.
P.
Mouralis,
Dr.
J.
Hyon,
and
the
Jet
Propulsion
Laboratory
(JPL)
Earth
Science
Directorate
for
their
support;
and
D.
Crichton,
R.
Doyle,
and
D.
Jones
for
their
counsel
and
insight
into
onboard
data
triage
concepts.
REFERENCES
[1]
D.
Crisp,
R.
Atlas,
F.-M.
Breon,
L.
Brown,
J.
Burrows,
P.
Ciais,
B.
Connor,
S.
Doney,
I.
Fung,
D.
Jacob,
C.
E.
Miller,
D.
O'Brien,
S.
Pawson,
J.
T.
Randerson,
P.
Rayner,
R.
J.
Salawitch,
S.
P.
Sander,
B.
Sen,
G.
L.
Stephens,
P.
P.
Tans,
G.
C.
Toon,
P.
0.
Wennberg,
S.
C.
Wofsy,
Y.
L.
Yung,
Z.
Kuang,
B.
Chudasama,
G.
Sprague,
B.
Weiss,
R.
Pollock,
D.
Kenyon,
and
S.
Schroll,
"The
Orbiting
Carbon
Observatory
(OCO)
Mission,"
Adv.
Space
Res.,
vol.
34,
no.
4,
pp.
700-709,
2004.
[2]
R.
0.
Green,
G.
P.
Asner,
S.
G.
Ungar,
and
R.
Knox,
"Results
of
the
decadal
survey
HyspIRI
imaging
spectrometer
concept
study:
A
high
signal-to-noise
ratio
and
high
uniformity
global
mission
to
measure
plant
physiology
and
functional
type,"
presented
at
the
IEEE
Geoscience
and
Remote
Sensing
Symposium
(IGARSS),
Boston,
MA,
USA,
2008,
Paper
TH1.109.3.
[3]
National
Research
Council,
Earth
Science
and
Applications
From
Space:
National
Imperatives
for
the
Next
Decade
and
Beyond,
Washington,
DC,
USA:
National
Academies
Press
2007.
[4]
L.
Mandrake,
C.
Frankenberg,
C.
W.
O'Dell,
G.
Osterman,
P.
Wennberg,
and
D.
Wunch,
"Semi-autonomous
sounding
selection
for
OCO-2,"
At-
mosp.
Meas.
Tech.
Discuss.,
vol.
6,
no.
3,
pp.
5881-5922,
Jun.
2013.
[5]
S.
Chien,
D.
Mclaren,
D.
Tran,
A.
G.
Davies,
J.
Doubleday,
and
D.
Mandl,
"Onboard
product
generation
on
earth
observing
one:
A
pathfinder
for
the
proposed
HyspIRI
mission
intelligent
payload
module,"
IEEE
JSTARS,
THOMPSON
et
al.:
RAPID
SPECTRAL
CLOUD
SCREENING
ONBOARD
AIRCRAFT
AND
SPACECRAFT
6791
vol.
6,
no.
1,
pp.
257-264,
Apr.
2013,
Special
Issue
on
the
Earth
Observing
One
(EO-1)
Satellite
Mission:
Over
a
Decade
in
Space.
[6]
M.
Klimesh,
"Low-complexity
lossless
compression
of
hyperspectral
im-
agery
via
adaptive
filtering,"
IPN
Pmgr.
Rep.,
vol.
42-163,
pp.
1-10,
Nov.
2005.
[7]
M.
Klimesh,
A.
Kiely,
and
P.
Yeh,
"Fast
lossless
compression
of
multi-
spectral
and
hyperspectral
imagery,"
in
Pmc.
2nd
Int.
Workshop
Onboard
Payload
Data
Compress.,
2010,
pp.
1-8.
[8]
N.
Aranki,
A.
Bakhshi,
D.
Keymeulen,
and
M.
Klimesh,
"Fast
and
adap-
tive
lossless
on-board
hyperspectral
data
compression
system
for
space
applications,"
in
Proc.
IEEE
Aerosp.
Conf
,
2009,
pp.
1-8.
[9]
D.
Keymeulen,
N.
Aranki,
B.
Hobson,
A.
Kiely,
M.
Klimesh,
and
K.
Benkrid,
"GPU
lossless
hyperspectral
data
compression
system
for
space
applications,"
in
Pmc.
IEEE
Aerosp.
Conf
,
2012,
pp.
1-9.
[10]
J.
A.
Williams,
A.
S.
Dawood,
and
S.
J.
Visser,
"FPGA-based
cloud
detection
for
real-time
onboard
remote
sensing,"
in
Pmc.
IEEE
Int.
Conf.
Field-Programm.
Technol.,
2002,
pp.
110-116.
[11]
E.
El-Araby,
M.
Taher,
T.
El-Ghazawi,
and
J.
Le
Moigne,
"Prototyping
automatic
cloud
cover
assessment
(ACCA)
algorithm
for
remote
sensing
on-board
processing
on
a
reconfigurable
computer,"
in
Proc.
IEEE
Int.
Conf.
Field-Pmgramm.
Technol.,
2005,
pp.
207-214.
[12]
R.
Eastman,
S.
Warren,
and
C.
Hahn,
"Variations
in
cloud
cover
and
cloud
types
over
the
ocean
from
surface
observations,
1954-2008,"
in
J.
Climate,
Nov.
2011,
vol.
24,
no.
22,
pp.
5914-5934.
[13]
M.
Mercury,
R.
Green,
S.
Hook,
B.
Oaida,
W.
Wu,
A.
Gunderson,
and
M.
Chodas,
"Global
cloud
cover
for
assessment
of
optical
satellite
ob-
servation
opportunities:
A
Hysp1RI
case
study,"
Remote
Sens.
Environ.,
vol.
126,
pp.
62-71,
Nov.
2012.
[14]
M.
King,
S.
Platnick,
W.
Menzel,
S.
Ackerman,
and
P.
Hubanks,
"Spatial
and
temporal
distribution
of
clouds
observed
by
MODIS
onboard
the
Terra
and
Aqua
satellites,"
IEEE
Trans.
Geosci.
Remote
Sens.,
vol.
51,
no.
7,
pp.
3826-3852,
Jul.
2013.
[15]
T.
Taylor,
C.
O'Dell,
D.
O'Brien,
N.
Kikuchi,
T.
Yokota,
T.
Nakajima,
H.
Ishida,
D.
Crisp,
and
T.
Nakajima,
"Comparison
of
cloud-screening
methods
applied
to
GOSAT
near-infrared
spectra,"
IEEE
Trans.
Geosci.
Remote
Sens.,
vol.
50,
no.
1,
pp.
295-309,
Jan.
2012.
[16]
M.
Wang
and
W.
Shi,
"Cloud
masking
for
ocean
color
data
processing
in
the
coastal
regions,"
IEEE
Trans.
Geosci.
Remote
Sens.,
vol.
44,
no.
11,
pp.
3105-3196,
Nov.
2006.
[17]
S.
A.
Ackerman,
K.
I.
Strabala,
W.
P.
Menzel,
R.
A.
Frey,
C.
C.
Moeller,
and
L.
E.
Gumley,
'Discriminating
clear
sky
from
clouds
with
MODIS,"
J.
Geophys.
Res.,
Atmosp.,
vol.
103,
no.
D24,
pp.
32
141-32
157,
Dec.
1998.
[18]
R.
A.
Frey,
S.
A.
Ackerman,
Y.
Liu,
K.
I.
Strabala,
H.
Zhang,
J.
R.
Key,
and
X.
Wang,
"Cloud
detection
with
MODIS.
Part
I:
Improvements
in
the
MODIS
cloud
mask
for
collection
5,"
J.
Atmosp.
Ocean.
Technol.,
vol.
25,
no.
7,
pp.
1057-1072,
Jul.
2008.
[19]
S.
Ackerman,
R.
Holz,
R.
Frey,
E.
Eloranta,
B.
Maddux,
and
M.
McGill,
"Cloud
detection
with
MODIS.
Part
II:
Validation,"
J.
Atmosp.
Ocean.
Technol.,
vol.
25,
no.
7,
pp.
1073-1086,
Jul.
2008.
[20]
L.
Gomez-Chova, G.
Camps-Valls,
J.
Calpe-Maravilla,
L.
Guanter,
and
J.
Moreno,
"Cloud-screening
algorithm
for
ENVISAT/MERIS
multispec-
tral
images,"
IEEE
Trans.
Geosci.
Remote
Sens.,
vol.
45,
no.
12,
pp.
4105-
4118,
Dec.
2007.
[21]
P.
Minnis,
Q.
Z.
Trepte,
S.
Sun-Mack,
Y.
Chen,
D.
R.
Doelling,
D.
F
Young,
D.
A.
Spangenberg,
W.
F
Miller,
B.
A.
Wielicki,
R. R.
Brown,
S.
C.
Gibson,
and
E.
B.
Geier,
"Cloud
detection
in
non-
polar
regions
for
CERES
using
TRMM
V1RS
and
terra
and
aqua
MODIS
data,"
IEEE
Trans.
Geosci.
Remote
Sens.,
vol.
46,
no.
11,
pp.
3857-3884,
Nov.
2008.
[22]
J.
Lee,
R.
Weger,
S.
Sengupta,
and
R.
Welch,
"A
neural
network
approach
to
cloud
classification,"
IEEE
Trans.
Geosci.
Remote
Sens.,
vol.
28,
no.
5,
pp.
846-855,
Sep.
1990.
[23]
J.
V.
Martins,
D.
Tatire,
L.
Remer,
Y.
Kaufman,
S.
Mattoo,
and
R.
Levy,
"MODIS
cloud
screening
for
remote
sensing
of
aerosols
over
oceans
using
spatial
variability,"
Geophys.
Res.
Lett.,
vol.
29,
no.
12,
pp.
MOD4-1-
MOD4-4,
Jun.
2002.
[24]
F
Murtagh,
D.
Barreto,
and
J.
Marcello,
"Decision
boundaries
using
Bayes
factors:
The
case
of
cloud
masks,"
IEEE
Trans.
Geosci.
Remote
Sens.,
vol.
41,
no.
12,
pp.
2952-2958,
Dec.
2003.
[25]
F-M.
Breon
and
S.
Colzy,
"Cloud
detection
from
the
spacebome
POLDER
instrument
and
validation
against
surface
synoptic
observa-
tions,"
in
J.
Appl.
Meteorol.,
Jun.
1999,
vol.
38,
no.
6,
pp.
777-785.
[26]
M.
K.
Griffin,
H.
K.
Burke,
D.
Mandl,
and
J.
Miller,
"Cloud
cover
detec-
tion
algorithm
for
EO-1
hyperion
imagery,"
in
Pmc.
17th
SPIE
AemSense
Conf.
Algorithms
Technol.
Multispectral,
Hyperspectral
Ultraspectral
Im-
ager),
IX,
Orlando,
FL,
USA,
2003.
[27]
F
Ip,
J.
Dohm,
V.
Baker,
T.
Doggett,
A.
Davies,
R.
Castao,
S.
Chien,
B.
Cichy,
R.
Greeley,
R.
Sherwood,
D.
Tran,
and
G.
Rabideau,
"Flood
detection
and
monitoring
with
the
autonomous
sciencecraft
experiment
onboard
EO-1,"
Remote
Sens.
Environ.,
vol.
101,
no.
4,
pp.
463-481,
Apr.
2006.
[28]
T.
Doggett,
R.
Greeley,
S.
Chien,
B.
Cichy,
A.
Davies,
G.
Rabideau,
R.
Sherwood,
D.
Tran,
V.
Baker,
J.
Dohm,
and
E
Ip,
"Autonomous
on-board
detection
of
cryospheric
change,"
in
Remote
Sens.
Envimn.,
Apr.
2006,
vol.
101,
no.
4,
pp.
447-462.
[29]
D.
Mandl,
M.
Griffin,
H.
Burke,
and
J.
Miller,
"Validation
of
on-board
cloud
cover
assessment
using
EO-1,"
in
Pmc.
ESTO
Technol.
Conf,
College
Park,
MD,
USA,
2003,
pp.
1-8.
[30]
H.
Ishida
and
T.
Y.
Nakajima,
"Development
of
an
unbiased
cloud
de-
tection
algorithm
for
a
spacebome
multispectral
imager,"
in
J.
Geophys.
Res.,
Atmosp.,
Apr.
2009,
vol.
114,
no.
D7,
pp.
D07206-1-D07206-16.
[31]
C.
Merchant,
A.
Harris,
E.
Maturi,
and
S.
MacCallum,
"Probabilistic
physically
based
cloud
screening
of
satellite
infrared
imagery
for
opera-
tional
sea
surface
temperature
retrieval,"
Q.
J.
R.
MeteomL
Soc.,
vol.
131,
no.
611,
pp.
2735-2755,
Oct.
2005.
[32]
R.
0.
Green,
M.
L.
Eastwood,
C.
M.
Sarture,
T.
G.
Chrien,
M.
Aronsson,
B.
J.
Chippendale,
J.
A.
Faust,
B.
E.
Pavri,
C.
J.
Chovit,
M.
Solis,
M.
R.
Olah,
and
0.
Williams,
"Imaging
spectroscopy
and
the
airborne
visible/infrared
imaging
spectrometer
(AVIRIS),"
Remote
Sens.
Envimn.,
vol.
65,
no.
3,
pp.
227-248,
Sep.
1998.
[33]
L.
Hamlin,
R.
Green,
P.
Mouroulis,
M.
Eastwood,
I.
McCubbin,
D.
Wilson,
D.
Randall,
M.
Dudik,
and
C.
Paine,
"Imaging
spectrometer
science
measurements
for
terrestrial
ecology:
AVIRIS
and
the
next
gen-
eration
AVIRIS
characteristics
and
development
status,"
in
Pmc.
NASA
ESTF,
2010,
pp.
1-44.
[34]
V. V.
Salomonson
and
I.
Appel,
"Estimating
fractional
snow
cover
from
MODIS
using
the
normalized
difference
snow
index,"
in
Remote
Sens.
Envimn.,
Feb.
2004,
vol.
89,
no.
3,
pp.
351-360.
[35]
B.
Gao,
P.
Yang,
W.
Han,
R.
Li,
and
W.
Wiscombe,
"An
algorithm
using
visible
and
1.38-pm
channels
to
retrieve
cirrus
cloud
reflectances
from
aircraft
and
satellite
data,"
IEEE
Trans.
Geosci.
Remote
Sens.,
vol.
40,
no.
8,
pp.
1659-1668,
Aug.
2002.
[36]
C.
M.
Bishop
and
N.
M.
Nasrabadi,
Pattern
Recognition
and
Machine
Learning.
New
York,
NY,
USA:
Springer-Verlag,
2006.
[37]
D.
Thompson,
B.
Bornstein,
S.
Chien,
S.
Schaffer,
D.
Tran,
B.
Bue,
R.
Castano,
D.
Gleeson,
and
A.
Noel,
"Autonomous
spectral
discovery
and
mapping
onboard
the
EO-1
spacecraft,"
IEEE
Trans.
Geosci.
Remote
Sens.,
vol.
51,
no.
6,
pp.
3567-3579,
Jun.
2013.
[38]
T.
Y.
Nakajima,
T.
Tsuchiya,
H.
Ishida,
T.
N.
Matsui,
and
H.
Shimoda,
"Cloud
detection
performance
of
spacebome
visible-to-
infrared
multispectral
imagers,"
Appl.
Opt.,
vol.
50,
no.
17,
pp.
2601-
2616,
Jun.
2011.
[39]
M.
C.
Hansen,
R.
S.
Defies,
J.
R.
G.
Townshend,
and
R.
Sohlberg,
"Global
land
cover
classification
at
1
km
spatial
resolution
using
a
clas-
sification
tree
approach,"
Int.
J.
Remote
Sens.,
vol.
21,
no.
6/7,
pp.
1331-
1364,
2000.
[40]
M.
A.
Friedl,
D.
Sulla-Menashe,
B.
Tan,
A.
Schneider,
N.
Ramankutty,
A.
Sibley,
and
X.
Huang,
"MODIS
Collection
5
global
land
cover:
Al-
gorithm
refinements
and
characterization
of
new
datasets,"
Remote
Sens.
Envimn.,
vol.
114,
no.
1,
pp.
168-182,
Jan.
2010.
[41]
J.
D.
Lindberg
and
M.
S.
Smith,
"Reflectance
spectra
of
gypsum
sanda
from
the
White
Sands
National
Monument
and
basalt
from
a
nearby
lava
flow,"
Amer.
Mineralogist,
vol.
58,
pp.
1062-1064,
1973.
[42]
D.
J.
MacKay,
Information
Theory,
Inference
and
Learning
Algorithms.
Cambridge,
U.K.:
Cambridge
Univ.
Press,
2003.
[43]
J.
Langford,
L.
Li,
and
A.
Strehl,
"Vowpal
wabbit
open
source
project,"
Yahoo!
Research,
Technical
report,
Yahoo!,
2007.
[44]
A.
P.
Bradley,
"The
use
of
the
area
under
the
ROC
curve
in
the
evaluation
of
machine
learning
algorithms,"
Pattern
Recog.,
vol.
30,
no.
7,
pp.
1145-
1159,
Jul.
1997.
[45]
T.
H.
Painter,
J.
Dozier,
D.
A.
Roberts,
R.
E.
Davis,
and
R.
0.
Green,
"Retrieval
of
subpixel
snow-covered
area
and
grain
size
from
imaging
spectrometer
data,"
Remote
Sens.
Envimn.,
vol.
85,
no.
1,
pp.
64-77,
2003.
[46]
T.
H.
Painter,
F
C.
Seidel,
A.
C.
Bryant,
S.
M.
Skiles,
and
K.
Rittger,
"Imaging
spectroscopy
of
albedo
and
radiative
forcing
by
light
absorbing
impurities
in
mountain
snow,"
J.
Geophys.
Res.,
Atmosp.,
vol.
118,
no.
17,
pp.
9511-9523,
Sep.
2013.
[47]
V.
Carrere
and
J.
E.
Conel,
"Recovery
of
atmospheric
water
vapor
to-
tal
column
abundance
from
imaging
spectrometer
data
around
940
nm-
sensitivity
analysis
and
application
to
Airborne
Visible/Infrared
Imaging
Spectrometer
(AVIRIS)
data,"
Remote
Sens.
Environ.,
vol.
44,
no.
2/3,
pp.
179-204,
May/Jun.
1993.