arrow on maps points to padre island grid cell ownscaling ...mf > max (mh), as they may rely on...

1
REFERENCES: Dixon, K. W., et al., 2016: Evaluating the stationarity assumption in statistically downscaled climate projections: is past performance an indicator of future results? Climatic Change, 135, 395–408. Lanzante, J. R., et al., 2017: Some Pitfalls In Statistical Downscaling Of Future Climate. Bull. Amer. Meteor. Soc., in press. Li, H., J. Sheffield, and E. F. Wood, 2010: Bias correction of monthly precipitation and temperature fields from Intergovern- mental Panel on Climate Change AR4 models using equidistant quantile matching. J. Geophys. Res., 115, D10101. create data sets that serve as proxies for coarse resolution GCM output. A key difference between this PM experimental design and typical SD applications is the availability of "future observations" ( OF ) in the PM design, which allows us to evalute an SD method's future skill, thereby quantifying aspects of the stationarity assumption. Month: August. Location: near & inland of Padre Island, Texas. A 25km grid cell similar to many coastal points in that it is mostly land, but the overlying 200km grid values are more maritime. The Statistical Downscaling Challenge: Distribu- tions of daily August tasmax OH & OF values have larger variances, higher means, and different shapes (more positively skewed) than do the more maritime MH, MF distributions. Also, the climate change signal causes most late 21st century MF values to exceed the warmest historical MH value. Historical Period SD Performance: All three SD methods perform relatively well, yielding August monthly mean downscaling errors (DH - OH) of < 0.25K. Also, all SD methods generate DH distributions with shapes and variances that more closely resemble the OH distribution than the large-scale and more maritime MH tasmax data. Future Period SD Performance: The EDQM & BCQManom methods perform markedly better than BCQM does in the warmer future – the opposite of the results seen in the Carbondale, CO case. At the Padre Island 25km grid cell, BCQM introduces a mean bias of +4.6K for the late 21st century August case, and multi-degree warm biases occur at most U.S. southern coastal land points (upper map). Such SD-created warm biases are not found in the EDQM (lower map) & BCQManom cases. Methods like BCQM that do not adjust for future trends can break down when posed with MF > max(MH), as they may rely on ad hoc tail adjustments to apply the correction at max(MH) to all warmer MF values – a poor choice when the MH and OH distribution shapes differ. SD methods that incorporate adjustments for trends guard against this particular pitfall. Historical Period Downscaling Errors vs. MH values (BCQM results shown. Other SD methods are similar) Future Period Downscaling Errors vs. MF values (BCQM & EDQM shown. BCQManom similar to EDQM) arrow on maps points to Padre Island grid cell Examining the Performance of Statistical Downscaling Methods: Toward Matching Applications to Data Products Keith W. Dixon , John R. Lanzante , & Dennis Adams-Smith (1) NOAA Geophysical Fluid Dynamics Laboratory, Princeton, NJ (2) University Corporation for Atmospheric Research This work was funded in part by the U.S. Dept. of the Interior's South Central Climate Science Center Project Web Page: www.gfdl.noaa.gov/esd 1 1 2 PA43B-0321 Variable of Interest = Daily Maximum Temperature (tasmax); Future Scenario = Late 21st Century High Greenhouse Gas Emissions; Comparing 3 Univariate SD Methods

Upload: others

Post on 06-Oct-2020

0 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: arrow on maps points to Padre Island grid cell ownscaling ...MF > max (MH), as they may rely on ad hoc tail adjustments to apply the correction at max (MH) to all warmer MF values

RE

FE

RE

NC

ES

:D

ixon

, K. W

., et

al.,

201

6: E

valu

atin

g th

e st

atio

nari

ty a

ssum

ptio

n in

sta

tistic

ally

dow

nsca

led

clim

ate

proj

ectio

ns: i

s pa

st

per

form

ance

an

indi

cato

r of

fut

ure

resu

lts?

Cli

mat

ic C

hang

e, 1

35, 3

95–4

08.

Lan

zant

e, J

. R.,

et a

l., 2

017:

Som

e P

itfal

ls I

n S

tatis

tical

Dow

nsca

ling

Of

Fut

ure

Clim

ate.

Bul

l. A

mer

. Met

eor.

Soc.

, in

pres

s.L

i, H

., J

. She

ffie

ld, a

nd E

. F. W

ood,

201

0: B

ias

corr

ectio

n of

mon

thly

pre

cipi

tatio

n an

d te

mpe

ratu

re f

ield

s fr

om I

nter

gove

rn-

men

tal P

anel

on

Clim

ate

Cha

nge

AR

4 m

odel

s us

ing

equi

dist

ant q

uant

ile m

atch

ing.

J. G

eoph

ys. R

es.,

115,

D10

101.

crea

te d

ata

sets

that

ser

ve a

s pr

oxie

s fo

r co

arse

res

olut

ion

GC

M o

utpu

t.

A k

ey d

iffe

renc

e be

twee

n th

is P

M e

xper

imen

tal d

esig

n an

d ty

pica

l SD

ap

plic

atio

ns is

the

avai

labi

lity

of "

futu

re o

bser

vatio

ns"

(OF

) in

the

PM

de

sign

, whi

ch a

llow

s us

to e

valu

te a

n S

D m

etho

d's

futu

re s

kill,

ther

eby

quan

tifyi

ng a

spec

ts o

f th

e st

atio

nari

ty a

ssum

ptio

n.

Mon

th: A

ugus

t. L

ocat

ion

: ne

ar &

inla

nd o

f P

adre

Is

land

, Tex

as. A

25k

m g

rid

cell

sim

ilar

to m

any

coas

tal p

oint

s in

that

it is

mos

tly la

nd, b

ut th

e ov

erly

ing

200k

m g

rid

valu

es a

re m

ore

mar

itim

e.

Th

e S

tati

stic

al D

own

scal

ing

Ch

alle

nge

: D

istr

ibu-

tions

of

daily

Aug

ust t

asm

ax O

H &

OF

val

ues

have

larg

er v

aria

nces

, hig

her

mea

ns, a

nd d

iffe

rent

sh

apes

(m

ore

posi

tivel

y sk

ewed

) th

an d

o th

e m

ore

mar

itim

e M

H, M

F d

istr

ibut

ions

. A

lso,

the

clim

ate

chan

ge s

igna

l cau

ses

mos

t lat

e 21

st c

entu

ry M

F

valu

es to

exc

eed

the

war

mes

t his

tori

cal M

H v

alue

.

His

tori

cal P

erio

d S

D P

erfo

rman

ce:

All

thre

e S

D

met

hods

per

form

rel

ativ

ely

wel

l, yi

eldi

ng A

ugus

t m

onth

ly m

ean

dow

nsca

ling

err

ors

(DH

- O

H)

of

< 0

.25K

. A

lso,

all

SD

met

hods

gen

erat

e D

H

dist

ribu

tions

with

sha

pes

and

vari

ance

s th

at m

ore

clos

ely

rese

mbl

e th

e O

H d

istr

ibut

ion

than

the

larg

e-sc

ale

and

mor

e m

ariti

me

MH

tasm

ax d

ata.

Fu

ture

Per

iod

SD

Per

form

ance

: T

he E

DQ

M &

B

CQ

Man

om m

etho

ds p

erfo

rm m

arke

dly

bette

r th

an B

CQ

M d

oes

in th

e w

arm

er f

utur

e –

the

oppo

site

of

the

resu

lts s

een

in th

e C

arbo

ndal

e, C

Oca

se. A

t the

Pad

re I

slan

d 25

km g

rid

cell,

BC

QM

in

trod

uces

a m

ean

bias

of

+4.

6K f

or th

e la

te 2

1st

cent

ury

Aug

ust c

ase,

and

mul

ti-de

gree

war

m

bias

es o

ccur

at m

ost U

.S. s

outh

ern

coas

tal l

and

poin

ts (

uppe

r m

ap).

Suc

h S

D-c

reat

ed w

arm

bia

ses

are

not f

ound

in th

e E

DQ

M (

low

er m

ap)

&

BC

QM

anom

cas

es. M

etho

ds li

ke B

CQ

M th

at d

ono

t adj

ust f

or f

utur

e tr

ends

can

bre

ak d

own

whe

npo

sed

with

MF

> m

ax(M

H),

as

they

may

rel

y on

ad h

oc ta

il ad

just

men

ts to

app

ly th

e co

rrec

tion

at m

ax(M

H)

to a

ll w

arm

er M

F v

alue

s –

a po

orch

oice

whe

n th

e M

H a

nd O

H d

istr

ibut

ion

shap

esdi

ffer

. SD

met

hods

that

inco

rpor

ate

adju

stm

ents

fo

r tr

ends

gua

rd a

gain

st th

is p

artic

ular

pitf

all.

His

tori

cal P

erio

d D

ow

nsc

alin

g E

rro

rs v

s. M

H v

alu

es(B

CQ

M r

esul

ts s

how

n. O

ther

SD

met

hods

are

sim

ilar)

Fu

ture

Per

iod

Do

wn

sca

ling

Err

ors

vs.

MF

val

ues

(BC

QM

& E

DQ

M s

how

n. B

CQ

Man

om s

imila

r to

ED

QM

)

a

rrow

on

map

spo

ints

to P

adre

Is

land

grid

cel

l

Exa

min

ing

th

e P

erfo

rman

ce o

f S

tati

stic

al D

ow

nsc

alin

g M

eth

od

s:To

war

d M

atch

ing

Ap

plic

atio

ns

to D

ata

Pro

du

cts

Kei

th W

. Dix

on ,

John

R. L

anza

nte

, & D

enni

s A

dam

s-S

mith

(1

) N

OA

A G

eoph

ysic

al F

luid

Dyn

amic

s La

bora

tory

, Prin

ceto

n, N

J

(2)

Uni

vers

ity C

orpo

ratio

n fo

r Atm

osph

eric

Res

ear

ch

Th

is w

ork

 was

 fu

nd

ed i

n p

art 

by 

the 

U.S

. D

ept.

 of 

the 

Inte

rior'

sSou

th C

entr

al C

lim

ate 

Sci

ence

 Cen

ter

Pro

ject

Web

Pag

e: w

ww

.gfd

l.no

aa.g

ov/

esd

1

1

2

PA43B-0321

Var

iab

le o

f In

tere

st =

Dai

ly M

axim

um

Tem

per

atu

re (

tasm

ax);

F

utu

re S

cen

ario

= L

ate

21s

t C

entu

ry H

igh

Gre

enh

ou

se G

as E

mis

sio

ns;

C

om

par

ing

3 U

niv

aria

te S

D M

eth

od

s