Professional Documents
Culture Documents
Daniel Swain Presentation
Daniel Swain Presentation
West Side of the San Joaquin Valley
Daniel
Swain
UC
Davis,
Atmospheric
Science
NASA/NSERC
SARP
2010
7/27/2010
Outline
• General
overview
of
land
surface
classifica(on
• Introduc(on
to
study
area
• Methods
• Results:
maps
and
features
of
interest
• Usefulness
to
other
SARP
projects
• Discussion
• Summary
and
Conclusions
Mature
almond
orchard,
Paramount
Farms,
6/29/2010
Classifica(on:
a
brief
overview
• Goal:
to
qualita(vely
and
quan(ta(vely
describe
the
land
surface
in
a
par(cular
region
• Use
sta(s(cal
analysis
to
assign
individual
elements
(pixels)
to
relevant,
pre‐defined
groups
containing
other
elements
with
similar
characteris(cs
Study
site
• Paramount
Farms,
located
on
the
west
side
of
the
Southern
San
Joaquin
Valley
• Terrain
is
very
flat,
and
climate
is
arid,
with
sparse
natural
vegeta(on
• Primary
land
use
is
agricultural,
mainly
orchards
• Two
tree
types:
almond
and
pistachio
• Numerous
roads
intersect
the
study
area;
most
are
unpaved
• There
are
no
natural
permanent
bodies
of
water,
but
two
major
irriga(on
channels
and
a
handful
of
holding
ponds
are
sca]ered
throughout
Study
site
(con(nued)
• DC‐8
overflights
occurred
on
6/29/2010
and
7/1/2010
(two
each
day)
DC‐8
Flight
Path
• MASTER
instrument
sampled
upward
radiance
from
an
al(tude
of
~13,000
feet
AGL
• Significant
cloud
cover
was
Study
Site
absent
at
all
levels
over
the
study
site
during
each
sampling
period,
N
which
increases
confidence
in
W E
accuracy
of
data
(Hook
et
al.,
S
2001)
Methods
• Spectral
data
from
MASTER
processed
using
ENVI
• Several
sta(s(cal
procedures
were
applied
to
the
data:
– Principal
Components
Analysis
(PCA)
– Minimum
Noise
Frac(on
(MNF)
– 3x3
pixel
smoothing
(median)
• Preliminary
qualita(ve
classifica(on
based
on
informal
ground‐truth
verifica(on
from
6/28‐6/29
fieldwork
Methods
(con(nued)
• Mul(ple
classifica(on
algorithms
were
used:
– Parallelepiped
– Minimum
Distance
– Maximum
Likelihood
– Mahalanobis
Distance
•
8
Regions
of
Interest
(ROIs)
were
defined,
and
representa(ve
“training”
areas
iden(fied
on
the
MASTER
imagery
based
on
qualita(ve
knowledge
of
the
study
area
Classes
used
• Almond
Trees
• Young
Almond
Trees
• Pistachio
Trees
• Field
Crops
• Bare
Earth
• Water
• Buildings
• Paved
Surface
Methods
(con(nued)
• ROIs
were
manually
outlined
in
ENVI
(pixel
by
pixel,
where
necessary)
•
Representa(ve
“training
polygons”
were
used
to
compile
sta(s(cs
characteris(c
of
each
land
surface
type
Methods
(con(nued)
• Very
high
resolu(on
photographs
from
Google
Earth
helped
to
verify
training
polygons
for
ROI
selec(on
in
the
study
area
Methods
(con(nued)
• Mahalanobis
Distance
produced
the
best
(least
noisy
and
most
accurate)
classifica(on
maps
– Takes
covariance
between
spectral
bands
into
account
in
probability‐weighted
distance
calcula(ons
(Kefauver,
2010)
– Works
well
with
mul(modal
distribu(ons
(in
this
case,
there
are
three
primary
tree
classes
with
dis(nct
spectral
signatures)
From
aiaccess.net
Results
• Successfully
able
to
differen(ate
between
the
three
primary
tree
classes
in
the
central
study
area
• Outside
of
the
boundaries
delineated
in
the
previous
slide,
uncertainty
increases
since
ground‐
truth
verifica(on
did
not
occur
in
these
areas
Young
almond
trees,
Paramount
Farms,
6/30/2010
MASTER
true‐color
imagery
(6/29/2010)
Mahalanobis
distance
classifica(on
• Almond
Trees
• Young
Almond
Trees
• Pistachio
Trees
• Field
Crops
• Bare
Earth
• Water
• Buildings
• Paved
Surface
Mahalanobis
distance
classifica(on
(with
3x3
median)
• Almond
Trees
• Young
Almond
Trees
• Pistachio
Trees
• Field
Crops
• Bare
Earth
• Water
• Buildings
• Paved
Surface
Mahalanobis
distance
classifica(on
(annotated)
Land
surface
area
by
class
40
35
30
Percent
of
total
area
(%)
25
20
15
10
5
0
Almond
Trees
Bare
Earth
Pistachio
Trees
Young
Almond
Field
Crops
Paved
Surface
Water
Buildings
Trees
Discussion
• Difficulty
in
separa(ng
young
almond
trees
from
pistachio
trees
– Both
regions
characterized
by
sparse
canopies
– Rela(vely
large
frac(on
of
bare
earth
in
these
two
orchard
types
led
to
high
spa(al
variability
• Mature/older
almond
orchards
were
also
characterized
by
increased
variability
(presumably
due
to
increased
tree
mortality
with
age),
but
separa(on
from
surrounding
younger
stands
not
a]empted
due
to
lack
of
ground
truth
in
these
areas
Discussion
(con(nued)
• Accuracy
was
assessed
by
comparing
each
classed
area
to
the
sta(s(cal
model
built
by
ENVI
for
each
ROI
• A
random
sample
of
points
within
each
ROI
was
selected
to
compile
a
reference
dataset
against
which
the
automa(cally‐classed
areas
were
compared
• The
degree
to
which
the
sta(s(cal
characteris(cs
of
both
the
defined
ROI
regions
and
the
automa(cally‐classed
regions
matched
the
reference
dataset
was
defined
as
the
accuracy
of
the
classifica(ons
Classifica:on
accuracy
by
group
100
90
80
Percent
Accuracy
(%)
Algorithm Accuracy
70 User Accuracy
60
50
40
Almond
Trees
Young
Almond
Pistachio
Field
Crops
Bare
Earth
Paved
Surface
Buildings
Water
Trees
Trees
Note:
overall
accuracy
=
92.4%
Applica(on
to
other
SARP
projects
• Evapotranspira(on
(ET)
mapping
– Use
of
proper
crop
coefficients
in
ET
model
• Vegeta(on
and
water
indices
– Correla(on
between
NDVI
and
crop
type
NDVI
(lep)
and
crop
type
(right).
Image
courtesy
of
Amanda
Rupiper.
Summary
and
Conclusions
• Successful
classifica(on
using
ENVI
for
the
study
area
• Mahalanobis
Distance
appears
to
be
the
best
op(on
for
classifica(on
in
this
seqng
• Different
tree
types
with
similar
canopy
extent
and
degree
of
open
space
can
be
difficult
to
separate,
even
if
different
species
• Demonstrates
the
importance
of
ground‐based
or
qualita(ve
verifica(on,
even
in
rela(vely
homogeneous
agricultural
seqngs
Sources
• ENVI
(Version
4.7,
ITT
2009)
used
for
all
calcula(ons
and
mapping
• Hook,
Simon
J.;
Myers,
Jeffrey
J.;
Thome,
Cur(s
J.;
Fitzgerald,
Michael;
Kahle,
Anne
B.
The
MODIS/
ASTER airborne simulator (MASTER) — a new
instrument for earth science studies.
Remote
Sensing
of
the
Environment.
76,
93‐102.
• Kefauver,
Shawn.
“Mul(spectral
Classifica(on
Techniques.”
Powerpoint
presenta(on,
June
2010.
Acknowledgements
• MASTER
data
provided
by
NASA,
with
pre‐
processing
by
the
Airborne
Science
Program
• Thanks
to
NSERC
for
sponsoring
SARP
2010
and
making
this
en(re
project
possible
• Thanks
also
to
the
DC‐8
flight
crew
• And
special
thanks
are
due
to
Shawn
Kefauver
and
the
rest
of
ET
group!
Ques(ons?