satellite - Stanford Universitycs231n.stanford.edu/reports/2017/posters/557.pdfMicrosoft PowerPoint...
Transcript of satellite - Stanford Universitycs231n.stanford.edu/reports/2017/posters/557.pdfMicrosoft PowerPoint...
Wide‐area
precipitatio
n estim
ation from
satellite im
agery
Paul M
. Aok
i
Motivation
Global precipitatio
n mod
els a
re crucial in
areas ra
nging from
clim
ate change re
search to
wire
less network planning
in
developing
cou
ntrie
s (my ow
n interest). Ra
ster data from
geostatio
nary sa
tellites is still the
only who
le‐globe
, near‐real‐
time, m
ulti‐de
cadal observatio
nal data record we have, but it is
not a
dire
ct m
easurement o
f rainfall presence (detectio
n) or
quantity (estim
ation).
WMO GPC
C an
nual ra
infall
estim
ate, 201
6
Prob
lem statem
ent
Mod
els for ra
infall de
tection and estim
ation from
infrared
satellite
imagery can be
con
structed
usin
g supe
rvise
d learning. R
ain/no
‐rain (R
/NR) labe
ls and rainfall values are not available glob
ally, but
quantitative precipita
tion estim
ate (QPE) d
ata is available for
select re
gion
s based
on weather ra
dar n
etworks.
The 20
17 state‐of‐the
‐art (SOTA) is b
ased
on pred
ictin
g rainfall at
the center pixel of e
very 15x15
patch (1
.2x 1.2) from each
hourly im
age of th
e stud
y region
and
I retain th
is approach here.
Data
Datasets
“Groun
d truth”: U
.S. N
WS QPE
grid
(hou
rly, 0.4)
Comparable op
erational system: U
CI PER
SIAN
N‐CCS
(ho
urly, 0.4)
Images: U
.S. N
OAA
GOES infrared
imagery (hou
rly, 0.4)
Stud
y region
U.S. G
reat Plains region, 30‐45N
, 105
‐90W (6
742,450,450,2)
948 M (3
0,30,1) p
atches = 6.8 TB @ 0.4
238 M (1
5,15,1) p
atches = 214
GB @ 0.8
Train/validation set: Winter/Summer 201
2, 4.8% ra
in pixels
Test se
t: Winter/Summer 201
3, 6.3% ra
in pixels
Metho
dIm
plem
entatio
nsin Keras/Ten
sorFlow:
•Keras“gene
rator” fo
r patches (
data augmentatio
n pipe
line)
•Several m
odels, includ
ing the on
es com
pared he
re:
Intuition
/motivation: The
SOTA
app
roach do
es not use any
convolutional layers, and
so m
ay m
iss out on any transla
tion‐
invaria
nt features th
at affe
ct probability of ra
in. It is a
lso re
latively
shallow, and
increased de
pth could im
prove hierarchical
representatio
n.Detectio
n measures : I app
ly standard m
easures –
binary
(sigmoid) cross‐entropy fo
r loss, binary accuracy and
critical sc
ore
index(CSI) for evaluation metrics.
(c.f.
(CSI is m
ore common
in m
eteo
rology;
is no
t used.)
Estim
ation measures : I app
ly m
ean squared error (MSE) as b
oth
loss and
evaluation metric.
Class im
balance: The
class distrib
ution is un
balanced
; to im
prove
CSI, balancing by oversam
pling of positive class and
by class
weightin
g have been exam
ined
.Re
ferences (see
repo
rt)
1.De
tection SO
TA. Tao
et a
l., J. Hydrometeorology, to appe
ar.
2.Estim
ation SO
TA: Tao
et a
l., IEEE CEC
2016.
3.NOAA
GOES. G
OES N Series D
ata Bo
ok (R
ev. D
), 2009.
4.NWS QPE. Lin & M
itche
ll, Proc. 200
5 AM
S Hy
drology Co
nf.
5.SD
AE. Vincen
t et a
l., JM
LR20
10.
6.Simplicity. Springenb
erget al., IC
LR 201
5. https://arxiv.org/abs/14
12.680
67.
VGG16
. Sim
onyan& Zisserman, ICLR 20
15. https://arxiv.org/abs/14
09.155
6
Selected
find
ings
Tuning
Sim
ple CN
NIn add
ition
to th
e usual tun
ing and sm
all‐n
overfitting tests, I
tune
d filter d
epth f(24, 32, 64, 96, 192
). The Simplicity
work used
f=96
for C
IFAR
‐10; here, f=
64 is used througho
ut.
Detectio
n compa
rison
It is im
possible to
determine Tao et al.’s
detectio
n accuracy with
certainty, but com
bining
statistics from tw
o of th
eir p
apers g
ives a
best gue
ss of 9
4.7%
. (the test se
t majority
‐class freq
uency).
Tao et al. repo
rt a te
st se
t CSI of 0
.306. So far, my accuracies are
substantially lower (e
.g., 89
.3% SOTA, 93.7%
Sim
ple CN
N) w
ith
class w
eightin
g/balancing that achieves c
omparable CSI.
Estim
ation compa
rison
SOTA
attaine
d an
MSE of 1
.32, far b
etter than the op
erational
PCCS
system
but larger th
an th
at of the
cou
nterfactual m
odel th
at
always e
stim
ates ze
ro ra
infall. Sim
ple CN
N app
ears to
do be
tter
than
eith
er but th
e freq
uency distrib
utions still needs to
be
checked against h
istorical prio
rs in
a rigorous way.
Conclusion
s / dire
ctions
While re
plication is no
t com
plete, it’s clear that the
SOTA
results
are plausib
le. Thu
s far, it a
ppears th
at SOTA
estim
ation results can
be
improved
upo
n; SOTA
detectio
n results re
main difficult to
improve past th
e majority
‐class freq
uency, and
add
ition
al layers
may be ne
eded
before large archite
ctural differen
ces can be
seen
.
NWS QPE
vs.
GOES IR
ban
d an
d WV ba
nd(U.S. G
reat Plains, 201
3‐02
‐21 16
00Z)
Cold (b
lue) im
agery at righ
t correspo
nds to intense rainfall at left.
SOTA
(SDA
E):
3‐layer fully‐con
nected
mod
elw/ S
DAE un
supervise
d pre‐training
•1 ba
nd mod
el•
2‐ba
nd ensem
ble (con
cat)
Simple CN
N:
Mod
ified
“Simplicity
” all‐C
NN m
odel
(+ dense to
p layers)
Tran
sfer CNN:
Pre‐trained VG
G16
mod
el
(+ dense to
p layers)
24 3264
96192
Filte
r depths >
64
overfit on this
dataset
All m
odels a
chieved
majority
‐class
validation an
d test
set a
ccuracy w/o
class w
eigh
ting.
Simple CN
N achieves
lower M
SE (left)
while re
taining a
reason
able frequency
distrib
ution (right)