questions about empirical downscalingimsc.pacificclimate.org/proceedings/10imscpdfs/r... ·...

Questions about Empirical DownscalingBruce Hewitson1, Rob Wilby2, Rob Crane3

1 University of Cape Town, South Africa2 Environment Agency, United Kingdom

3 Penn State University & AESEDA, USA

University of Cape Townwww.csag.uct.ac.za

010203040506070

1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005

Year

Jour

nal p

ublic

atio

ns

"downscaling and climate" "statistical downscaling""dynamical downscaling" "downscaling and impact"

Peer reviewed journal publications listed on the Web of Science [accessed March 2006]

Wish I could be in Beijing!!

(Rob Wilby)

WWF


Downscaling: idealism and pragmatism

1 cm

12700km

A: The context for downscalinga) A complex environment; ranging from pessimism around uncertainty and

model limitations, to blind faith in perfect results

b) A user community misunderstanding projections versus forecasts

c) Yet, a pressing demand from stakeholders for any information that is better than tossing a coin

d) A tension between scientific idealism and stakeholder pragmatism

e) Downscaling is only one of a number of possible sources of information; e.g. AR4 CH 11 attributes for developing robust statements

- past trends, GCM envelope, downscaling studies, process understanding

f) Downscaling drivers come from GCMs that are not the real world, but a reduced dimensionality representation that is responsive to the same forcing parameters as the real world

g) Downscaling (as does all climate change science) operates in a realm of incomplete knowledge, imperfect tools, and a society whose policy and development measures will continue despite this


Future society

Emissionspathway

Climatemodel

Regionalscenario

Impactmodel

Impact


Downscaling is at the heart of the uncertainty cascade

B: Questions about empirical downscaling:a) What are we actually trying to achieve?

- The target lies somewhere between point-scale high temporal resolution and simple, broad, regional indications of the direction of change.

- The future under focus is dominantly the time horizon of policy and development plans

b) Just how robust are the tools? To what degree are the solutions sensitive to methodological choice and predictor suite (subject to minimum criteria), and to what degree do the differences matter?

c) As empirical downscaling bypasses model grid-cell parameterization and works from the skill-scale of the model, can one achieve better convergence?

d) Can downscaling tools be geographically transferable without custom case-by-case tuning?

e) What about stationarity and feedbacks

f) Do the limitations preclude usability? (e.g. “there are pressing philosophical issues to be answered first” & “downscaling cannot credibly succeed”)


C: Objectives of downscalinga) To provide an additional information resource targeted at

assessing regional climate responses that is:- Reflective of the first order response to large scale forcing- Consistent with physical process changes- At spatial and temporal scales of stakeholder relevance- Provides defensible information on projections (multiple

information)- Contextualized in term of uncertainty- Without trying to capture local feedback modulation

b) Serves to facilitate model diagnosis

c) Allow for rapid evaluation of regional attributes from many GCMs

d) Derive regional response for exotic variables (e.g. Storm surge) –even RCMs cannot directly achieve some of these

e) Aid understanding of process coupling across spatial scales


Examination: assess 2 techniques in 3 tough situations

Two downscaling methods in broad current usage- The two methods evolve from different starting points and are

implemented in different waysa. Begins with a weather generator and conditions this with the

atmosphere predictorsb. Starts with a cross-scale transfer function with the atmosphere

predictors and adds the local high frequency variance

Test against challenging situations- Target the downscaling of precipitation in different climate regimes of

Africa, using “dirty” training data in complex local climates

Apply the tools in a non-proprietary manner- i.e. apply without tuning the algorithm separately for each case



Case #1: Addis Ababa, Ethiopia


Case #2: Casablanca, Morocco


Case #3: Steenbras dam, South Africa

Downscaling conceptualization and goalsA procedure that derives a normative regional response to the large scale

forcing. Draws on the empirical information present in the observed data record, and in so doing:- Reflects the deterministic component of the large scale forcing- Includes the local (sub-GCM-grid scale) variance

Downscaling does not seek to reproduce the real world in observation for observation, but rather a realistic time evolution that:- at seasonal and inter-annual scales should match relative magnitude of

the temporal evolution of the forcing- at daily time scales should match the statistics of the daily events

(frequency of events, etc)

Downscaling should not seek to correct errors in the predictors; but predictor errors (such as too many low pressure systems) should be propagated

The methods lend themselves to ensemble downscaling and allows for an assessment of the envelope of response


• GCM boundary conditions are the main source of uncertainty affecting all downscaling methods

• Statistical and dynamical downscaling have similar skill

• Different downscaling methods yield different scenarios

• There are no universally “optimum” predictor(s)/domains, but there are guidelines to the baseline criteria

• Downscaling extreme events can be problematic

• Traditional skill measures for current climate may not be the best guide to the value of future scenarios of change

What do we know already?


Monthly totals (mm)

0

50

100

150

200

250

300

1 2 3 4 5 6 7 8 9 10 11 12

ObservedSeries2Series3Series4Series5


Case #1: NCEP predictors, daily precipitation, Addis Ababa

Continental environment, convective rainfall systems, tropical location.

Method A: three downscalings with different predictor setsMethod B: one downscaling, different predictor set to Method A

Predictors include parameters reflecting lower and mid troposphere circulation and humidity

365-day climatology of raindays per 30-day window

0

5

10

15

20

25

30


365-day climatology of 30-day total rain

0

50

100

150

200

250

300




1

10

100

1000

10000

5 10 15 20 25 30 35 40 45 50 55 60 65 70 75 80 85 90 95 100


Raindays > 10mm

0

1

2

3

4

5

6

7

8

9

10

1 2 3 4 5 6 7 8 9 10 11 12




Histogram of magnitude of rainfall events

Monthly time series (1979-1988)

- 50

0

50

100

150

200

250

300

350

400




Overall a credible representation of local climate in terms of both low frequency and high frequency response to the daily atmospheric predictors

One method over-predicts the frequency of low magnitude rain events not a major impact on totals, but relevant to, for example, soil moisture and landscape hydrology

Other method over predicts frequency of high magnitude events marginally gives too high totals in some months

No systematic evidence of one set of predictors outperforming another

Monthly totals (mm)

0

10

20

30

40

50

60

70

80

1 2 3 4 5 6 7 8 9 10 11 12



Case #2: Casablanca

Coastal environment without strong topographical forcing.

Very dirty data with missing days and some very suspect large values. Meta-data for the station is poor.

Raindays > 10mm

0

0.5

1

1.5

2

2.5

1 2 3 4 5 6 7 8 9 10 11 12

Series1Series2Series3Series4Series5

Dry spell duration

0

10

20

30

40

50

60

70

80

90

1 2 3 4 5 6 7 8 9 10 11 12


Case #2: Casablanca


Monthly time series of rainfall totals (1979-1988)

-50

0

50

100

150

200

250

300

350



Case #2: Casablanca

Problems of “over-prediction” in the summer – but how good is the station data?

Possible explanations:• The downscaling is infilling the missing days, hence increasing totals• The suspicious high values in the observed time series are possibly skewing the

downscaling function• Phase errors in the low quality data could be ascribing high precipitation to an

incorrect atmospheric state, which might have a high frequency of occurrence

Monthly totals (mm)

0

20

40

60

80

100

120

140

160

180

1 2 3 4 5 6 7 8 9 10 11 12



Case #3: Steenbras Dam

Coastal environment with strong topographical forcing, very sensitive to occurrence of orographiccloud, subject to winter rainfall from the mid-latitude westerly flow

On the face of it, high quality data, but some suspicion of a mid-series phase shift

365-day climatology of 30-day total rain

0

20

40

60

80

100

120

140

160

180

200




How much of a role could NCEP quality be playing?

Especially as the location is very sensitive to boundary layer moisture and boundary layer wind direction!


Monthly precipitation (1979-1990)

-50

0

50

100

150

200

250

300

350

1 8 15 22 29 36 43 50 57 64 71 78 85 92 99 106 113 120 127 134 141



Generally good, realistic, and captures the low and high frequency variance wellErroneously high precipitation in late winter, most apparent in one method which has greater sensitivity to the day-to-day phase matching with the predictors

Interim Conclusions

a) Both methods perform comparably; - Conditioned weather generator has a tendency to over do

frequency of low rainfall events - Transfer function with sampling of CDF has a tendency to over

estimate high magnitude events

b) Variation in predictor suite does not have a major influence; subject to minimum criteria of incorporating some representation of the base circulation and humidity attributes.

c) The results can be very credible; but have a vulnerability to data quality

d) Difficulty in separating out sources of error; station data, phase errors, NCEP realism.


Annual precipitation scenarios

-60

-40

-20

0

20

40

60

Tanger M ekness Casablanca Beni M ellal M arrakech Oujda M idelt Agadir Ouarzazate

% c

hang

e

UCT-CSIRO UCT-ECHAM4 UCT-HadAM3 SDSM-HadCM3

What about Climate change and the delta-issue?On the assumption that the GCMs are simplified representations of

reality, and proportionally sensitive to the real world anthropogenic forcing;

Given empirical downscaling propagates signal and error of the large scale atmospheric response;

And evidence that circulation-delta is largely consistent across GCMs


Downscaled annual precipitation scenarios for sites in Morocco by the 2080s under SRES A2 emissions. Source: World Bank (2007)

What about Climate change and the delta-issue?

For many locations, a strong multi-model agreement on the direction of change (better than using GCM grid cell data), but still a large inter-model range in magnitude.


Midelt rainfall totals (A2 emissions, 2080s)

-50

0

50

100

150

Jan

Feb

Mar

Apr

May Ju

n

Jul

Aug

Sep Oct

Nov

Dec

DJF

MA

M

JJA

SO

N

AN

N

% c

hang

e


Casablanca rainfall totals (A2 emissions, 2080s)

-100-80-60-40-20

02040

Jan

Feb

Mar

Apr

May Ju

n

Jul

Aug

Sep Oct

Nov

Dec

DJF

MA

M

JJA

SO

N

AN

N

% c

hang

e


High mountainsCoastal

Monthly anomaly for two locations

Change = drying; consistent with process understanding

Change = very uncertain! No confidence.

West Coast

Central Karoo

Eden

Overberg

Cape Winelands

#

City of Cape Town

West Coast

Central Karoo

Eden

Overberg

Cape Winelands

#

City of Cape Town

West Coast

Central Karoo

Eden

Overberg

Cape Winelands

#

City of Cape Town

10th percentile 90th percentile

Median

AR4 multi-model downscaling: multi-site example

Downscaled mean JJA rainfall (mm/month) response anomaly

XUniversity of Cape Town

www.csag.uct.ac.za

So what?Given:- The demand for regional information- The limitations and uncertainty of GCM grid cell data- The value of using multiple sources of information to assess change

Then, empirical downscaling is arguably:- Informative about the regional response to large scale forcing- Relatively insensitive to method subject to some baseline criteria- A fruitful avenue to support the impacts and adaptation community

But:- Whether traditional skill measures under the current climate are the

best guide to the skill of future scenarios of change needs more assessment

And:- There are fruitful avenues to be explored in using downscaling for

model diagnosis and understanding regional process


Source: BBC News

“Apathy is a sort of living oblivion.” Horace Greeley (1811-1872)

BBC news

questions about empirical downscalingimsc.pacificclimate.org/proceedings/10imscpdfs/r... ·...

Documents