Downscaling / Regionalization Techniques and methodologies AIACC Workshop, Apr 2002 Bruce Hewitson CSAG : University of Cape Town

Downscaling / RegionalizationTechniques and methodologies

AIACC Workshop, Apr 2002 Bruce Hewitson CSAG : University of Cape Town

AIACC Workshop, Apr 2002 Bruce Hewitson CSAG : University of Cape Town

Downscaling: Techniques and methodologies

General concepts and assumptions

Regional Climate Models : Overview of application

Empirical-Statistical downscaling : Overview and application issues

Decision approach to downscaling choices

AIACC Workshop, Apr 2002 CSAG : University of Cape Town

Simple

Complex

Reliable Dangerous

Downscaling: a valuable procedure of tremendous potential facing a minefield of choices

?

?

?

?

Do you need to downscale? What do you NEED rather than WANT

For the scientific question you are asking, can you do with a simple sensitivity study, use the native GCM data, apply interpolation, add the GCM anomaly to a baseline data set, or has some else already generated a suitable product???

?


Downscaling: A technique to take GCM atmospheric fields andderive climate information at a spatial/temporal scale finer than that of the GCM

Note: the downscaled predictand can only contain variance that exists in the cross scale relationship captured by f.

Anything else is/must be “made up”

“Local” Climate = f (larger scale atmospheric forcing)

R = f (L)

R: predictand - (a set of) regional scale variables

L: predictors - large scale variables from GCM

f: stochastic or quantitative transfer function conditioned by L, or a dynamical regional climate model.


Two options premised on the same assumptions

Regional Climate Models (RCMs) or Empirical cross scale functions

Assumptions: • The GCM is skillful (enough) with regard to the predictors used in the downscaling -- are they “adequately” simulated by the GCM

“Adequate” requires evaluating the GCM in terms of the predictor variables at the space and time scales of use!

e.g: For RCMs this could mean the full 3-dimensional fields of motion, temperature, and humidity, on a 6-12 hour time interval, over the domain of interest.

Note: Downscaling propagates the GCM error



Regional Climate Models or Empirical cross scale functions

Assumptions: • f is valid under altered climatic conditions - stationarity

Note: This applies to empirical downscaling and and RCMs. If the climate system is substantially non-stationary, then at the very least empirical downscaling becomes questionable, possibly much of RCM applications as well.

Local Response

ie: the bulk of future synoptic states are at least represented in present day records -- the future dominated by changes in frequency, intensity, and persistence.



Regional Climate Models or Empirical cross scale functions

Assumptions: • The chosen predictors represent / contain the climate change signal.

For example (empirical downscaling): if local temperature is well determined by synoptic scale sea level pressure (SLP), which shows minimal change into the future.

An effective empirical downscaling may be derived, but, what if atmospheric moisture content goes up?

The downscaled T from SLP may be ~0, yet a large T may actually exist from the moisture change.


Regional Climate Models

Computationally intensive, physically based (in part), likely the most viable/valid downscaling in the long term, still somewhat

developmental.

Conceptual approach:

Scale an AGCM to a finite domain, calibrate paramterizations for higher resolution, couple a land surface scheme, force at the boundaries with atmospheric fields from the GCM - simple?



Computationally intensive, physically based (in part), likely the most viable/valid downscaling in the long term, still somewhat

developmental.

Conceptual approach:

Scale an AGCM to a finite domain, calibrate paramterizations for higher resolution, couple a land surface scheme, force at the boundaries with atmospheric fields from the GCM - simple?

Conceptual issues:How to interface at boundaries Boundary field updatingInflow versus outflow Parameterization schemesLand surface scheme Number of levels

Spatial resolutionDomain sensitivityGCM versus RCM physics 1-way versus 2-way nesting



Practicalities:

• Complex procedure with many implementation decisions that can determine the result obtained.

• Need to understand why you get the results you see (right answer for wrong reason problem).

• Selection of domain, physics package, parameterization, and evaluation of performance is a time-consuming procedure, but essential!

Running a RCM, given suitable IT skills and resources, can be done in a matter of days.

Achieving understandable and justifiable results can be very lengthy.


Case example from Africa of implementing an RCMs

Instituting an RCM in an environment where it has not been run before

16 scientists from around Africa, two week training workshop, full IT support, theory lectures, all software and scripts configured,email follow up with participants.

18 months later, 7 active participants, not all of whom achieved successful simulations at their home institution.


Running a RCM -- A brief exposure to typical activities

Select a model: preferably one that will run on available computational infrastructure, with an established user base, and make a (friendly) contact with an experienced user.

Develop appropriate skills: Unix literate, Fortran/C capable, data handling and visualization skills.

Implement appropriate infrastructure: Single PC can handle months to 1 year type simulations. Longer climate simulations require PC clusters or multiple-CPU workstations.



Select a model: preferably one that will run on available computational infrastructure, with an established user base, and make a (friendly) contact with an experienced user.

Develop appropriate skills: Unix literate, Fortran/C capable, data handling and visualization skills.

Implement appropriate infrastructure: Single PC can handle months to 1 year type simulations. Longer climate simulations require PC clusters or multiple-CPU workstations.

MM5 v3, land surface model, 110x100 grid points, 23 levels, 60km resolution

Simulation setup ResultsHardware Cost Speed # PCs Days Hours Min/ day ~Hrs/ day (1 PC)

Intel P4 $8,000 1500 MHz 6 120 35 17.5 1.75

AMD XP2000+ $12,000 1600 MHz 8 120 18 9 1.2

DEC ES40 $80,000 667 MHz "4" 120 22 11 N/A



Domain and resolution: If no one else has done it, establish domain sensitivity for region of interest. Select horizontal resolution, vertical levels, physics options. Undertake appropriate sensitivity studies.

Prepare boundary conditions: Establish a means of ingesting boundary field data into the RCM (and getting it out).

Develop reference climatology: Undertake a 10+ year simulation with reanalysis boundary conditions.



Domain and resolution: If no one else has done it, establish domain sensitivity for region of interest. Select horizontal resolution, vertical levels, physics options. Undertake appropriate sensitivity studies.

Prepare boundary conditions: Establish a means of ingesting boundary field data into the RCM (and getting it out).

Develop reference climatology: Undertake a 10+ year simulation with reanalysis boundary conditions.

Evaluate reference climatology: This is critical …. if the RCM is not appropriately simulating key processes, generating a future climate anomaly pattern has little meaning.

Note: “point and click” solutions are coming (and very welcome), BUT be wary of running an RCM over a new region without careful evaluation.



Apply GCM control simulation fields, run 10+ year nested simulation.This provides the reference climatology to which the future climate simulation will be compared. Evaluate the climatology, does the GCM/RCM combination generate an appropriate regional climate.

Apply GCM future climate simulation, run 10+ year simulation: Finally, the regionalized future climate!

“Signal to noise”: Ideally, repeat control and future climate nested simulations with other ensemble members from the GCM runs. Then repeat with another GCM!

Analyze your future climate, and the climate anomaly: Can you understand and explain why the future climate anomaly it the way it is.

Use the regionalized climate data: Either as direct results, or possibly by adding the regional anomaly to your baseline climatology.


Regional feedbacks

RCMs are powerful in allowing investigation of process response to feedbacks and forcings other than from GHG.

Example for southern Africa: Vegetation is almost certain to change from climate change forcing. What is the feedback to the atmosphere, and the consequent exacerbation or mitigation of climate change?

Average NPP 1901-1995

4

4.5

5

5.5

6

6.5

7

7.5

8

8.5

1901 1906 1911 1916 1921 1926 1931 1936 1941 1946 1951 1956 1961 1966 1971 1976 1981 1986 1991

Control

Dcrsd PPT

Incrsd RH

Dcrs RH

Dcrs TMP

Incrs PPT

Incrs TMP

NPP for 20% increase and decrease on the 1900-1999 record of T, RH, and ppt simulated by SDGVM.


Plant Functional TypesBare GroundC3 grassesC3 grassesEvergreen Broadleaf ForestDeciduous Needleleaf

Change in plant functional types modelled by the SDGVM for a 20% increase in precipitation. The cross-hatching shows areas of change


Regional feedbacks

Experiment design: MM5v3 RCM, domain over sub-equatorial Africa, albedo perturbed by 20% (within range of natural variability).

3 ensemble simulations for summer with and without perturbation.

Results: indicate mean temperature change by up to 0.75 degrees. Response is from a of change in the dynamics of circulation, moisture transport, and cloud formation.

Future climates may perturb albedo by far more than 20%.


Empirical/Statistical downscaling

A plethora of competing and diverse algorithms of widely

different strengths and weaknesses

Region Technique Predictor Predictand Time Author (s)Africa

South Africa TF C P D Hewitson & Crane, 1996

AmericaUSA WT T Tmax, Tmin D Brown & Katz, 1995

USA WG C P D Zorita et al., 1995

USA WG, TF C, T, VOR P D Wilby & Wigley, 1997

USA TF C, Q P D Crane & Hewitson, 1998

USA WG, TF C, T, VOR T, P D Wilby et al., 1998a, b

USA WG, WT C T, P D Mearns et al., 1999

USA TF C, T, RH, W T D Sailor & Li, 1999

USA WG P D Bellone et al., 1999Mexico and USA TF C, TH, O P D Cavazos, 1997Mexico and USA TF, WT C, TH, Q P D Cavazos, 1999Central Argentina TF C, W T, Tmax, Tmin M Solman & Nuñez, 1999

AsiaJapanese coast TF C Sea level M Cui et al., 1995, 1996Chinese coast TF Sea level

variabilityM Cui and Zorita, 1998

OceaniaNew Zealand WT C Tmax, Tmin, P D Kidson & Watterson, 1995New Zealand TF C, TH,

VOR, WT, P D Kidson & Thompson, 1998

Australia TF C Tmax, Tmin D Schubert &Henderson-Sellers,1997

Australia TF C Tmax, Tmin D Schubert, 1998Australia WT C, T P Timbal & McAvaney, 1999Australia WT Schnur & Lettenmaier, 1999

EuropeEurope WG VOR, W Conoway et al., 1996Europe WG, TF C, P, Tmax,

Tmin, OT, P D Semenov & Barrow, 1996

Europe TF C, W, VOR,T, Q, O

T, P M Murphy, 1998a, b

Europe TF C T, P, vapourpressure

D Weichert & Bürger, 1998

Germany TF T Phenological event Maak &van Storch, 1997Germany TF C Storm surge M Von Storch & Reichardt, 1997Germany TF Salinity Heyen & Dippner, 1998

Germany WT Thunderstorms D Sept, 1998Germany TF Ecological

variablesKrönke et al., 1998

Iberian Peninsula WG C P D Cubash et al., 1996Iberian Peninsula TF C Tmax, Tmin D Trigo & Palutikof, 1998Iberian Peninsula TF P, NST Boren et al., 1999Iberian Peninsula TF P, NST Ribalaygua et al., 1999Spain (and USA) TF C Tmax, Tmin D Palutikof et al., 1997Spain (and USA) TF C Tmax, Tmin D Winkler et al., 1997Spain WT D Goodess & Palutikof, 1998Portugal TF C P M Corte-Real et al., 1995Portugal WT C D Corte-Real et al., 1999The Netherlands WT C, VOR, W T, P D,M Buishand & Brandsma, 1997Norway TF C, O T, P and others M Benestad, 1999a, bNorway (glaciers) TF C, O Local weather D Reichert et al., 1999Romania TF C P M Busuioc & von Storch, 1996Romania TF C P M Busuioc et al, 1999Switzerland TF P Buishand & Klein Tank, 1996Switzerland TF P Brandsma & Buishand, 1997Switzerland TF D Widmann & Schär, 1997Switzerland WG C Local Weather H Gyalistras et al., 1997Switzerland TF P Buishand & Brandsma, 1999Poland TF C T, sea level, wave

height, salinity,wind, run-off

D,M Mietus, 1999

Alps WT Fuentes & Heimann, 1996Alps TF C, T T, P M Fischlin & Gylistras, 1997Alps WT C Snow Martin et al., 1997Alps WT Fuentes et al., 1998Alps TF C, T T, P, Gyalistras et al., 1998Alps, TF C, T Snow cover Hantel et al., 1998Alps WT C, T Landslide activity Dehn, 1999a, bAlps WT T, P D Heimann and Sept, 1999Alps WT P D Fuentes & Heimann, 1999Alps TF, WG C, T Weather statistics M Riedo et al., 1999Alps TF C P M Burkhardt, 1999

Mediterranean TF C, P T Palutikof & Wigley, 1995Mediterranean TF C P S Jacobeit, 1996North Atlantic TF C Pressure

tendenciesM Kaas et al., 1996

North Atlantic TF C Wave height M WASA, 1998North Sea TF Ecological

variablesDippner, 1997a, b

North Sea coast TF C Sea level M Langenberg et al., 1999Baltic Sea TF SLP Sea level M Heyen et al., 1996

Region not specifiedWT Frey-Buness et al., 1995WT C Matyasovszky & Bogardi, 1996WT Enke & Spekat, 1997TF C, VOR, W Kilsby et al., 1998TF Ecological

variablesHeyen et al., 1998


Significant lack of systematic evaluation ….

“More co-ordinated efforts are thus necessary to evaluate the different methodologies, inter-compare methods and

models”

IPCC, TAR 2001

Advantages:

• Computational efficiency

• Rapid application to multiple GCMs

• Tailoring to target variables (eg: storm surge)

• Applicability to broad range of temporal and spatial resolutions

• Accessibility beyond the modeling community

• Complementary to regional modeling


Two extremes to categories of downscaling:

• Transfer functions relating atmospheric forcing to target variable

• Stochastic functions and pure weather generators

For both: variance explained as a function of the large scale flow, residual variance can only be stochastically generated.

For future climate, only change due to the signal contained in the GCM scale forcing can be accounted for …..

Vari

ance

expla

ined

Synoptic scale Local scale

Synoptic ForcingSub-grid scale forcings


For climate change:

… what proportion of response will be due to sub-GCM grid scale structure -- independent of the large scale forcing?

…how stationary is the downscaling function -- applicable to both transfer functions and stochastic functions.

Q: for a given location, which is dominant: local or synoptic forcing?


High-res SST

1° SST

Zonal SST

Example of synoptic dominance: (From a RCM experiment)

Experiment: Precipitation as a function of three different SST fields with identical NCEP boundary forcing.

Precipitation (primarily convective) is temporally consistent independent of the SST fields. Implies dominance by synoptic state.

ie: The variance at sub-GCM grid cells is still conditioned by large scale flow, empirical downscaling of future change is strongly viable.


General categories of methodologies

Transfer Weather Stochastic conditioned StochasticFunction Typing on weather type

Typical downscaling modes:

Downscale from atmospheric instantaneous state to the local climate response (eg: daily precipitation)

Downscale secondary variable (eg: stream flow)

Relate atmospheric indices (eg: SOI, NAO) to climate statistics

Time downscaling -- downscale the diurunal cycle




Derives a quantitative relationship between predictor(s) and predictand(s)

eg: Station daily temperature = f (Sea Level Pressure & 500hPa gpm)

• f typically a regression style function, can / should be non-linear.

• Requires training data of adequate duration to span the range of events found in future climate.

• If predictands are patterns (eg: EOF) or indices (eg: NAO), one assumes stationarity of the pattern or index into the future.

• Derived function used with GCM field to downscale control and future climate.

• Can apply to any target scale, including stations.




- Method under-predicts peaks, over predicts minimum -- characteristic of a generalization function- Residuals represent variance not captured, either from inadequate predictors, or due to local forcing not represented in GCM fields

Example: ANN-based downscaling of daily rainfall

Effective at capturing temporal evolution consistent with atmosphere. Capture low frequency variability well (seasonal and interannual)Residuals (missing variance), can easily be added stochastically.


% variance of residuals is proportional to information in predictors

or

Skill of f is proportional to information in predictors

Downscaled station precipitation from 1° MRF assimilation data

Possible role in downscaling nested models to point resolution?

0

10

20

30

40

50

60

70

1 4 7 10 13 16 19 22 25 28 31 34 37 40 43 46 49 52 55 58 61 64 67 70 73 76 79

pre

cip

(m

m*1

0)

Observed

Downscaled




Derives a quantized relationship between predictor(s) and predictand(s)

eg: Station temperature = f (type of Sea Level Pressure pattern)

• Comes from the “synoptic climatology” discipline

• Weather patterns classified into N-different types

• Each type associated with a local climate response

• GCM weather patters matched to types, and assigned a local climate response




Stochastic / weather generators calibrated on observed data, conditioned on atmospheric state.

• Very effective at capturing high frequency variance, peaks and extremes

• Requires long term data sets to effectively define stochastic characteristics

• Question of stationarity


Note the underlying commonality:

All methods are, in effect, algorithms to implement an analog.

ie: each method simply draws a climate response from the historical record based on some atmopsheric state(s) from the same historical period

Thus: why not implement a true analog? Simply match a given GCM field to all possible comparable fields in the historical and take the closest match?


Evaluation of issues for effective empirical or statistical downscaling.

1) Choice of predictor variables

Most commonly used are circulation related variables




2) Predictor spatial representation

• Indices, EOFs, Synoptic classifications, Raw grid data

• Local versus remote (teleconnections)

• Surface versus upper air fields





3) Antecedent conditions

Local forcing as a function of antecedent events






4) Target (predictand) resolutions

Station scale, impacts scale (scale of user community), RCM scale?







5) Training data periods

Observational data that sufficiently spans the relationship for training downscaling function








6) Representing the climate change signal

Predictors explaining significant variance may not be predictors sensitive to the climate change signal









7) Stationarity of function / predictors

Is climate change primarily characterized by changes in frequency of existing events?

Are changes in local sub-grid-scale forcing small with respect to synoptic forcing?

Are residuals in downscaling from GCM-resolution due to low predictor resolution, or sub-grid scale forcing?









7) Stationarity of function / predictors

A given downscaling implementation needs to take cognizance of, and evaluate, the dependencies


Exploring the dependencies.

• Transfer function based methodology: gives dominance to synoptic forcing

• Challenging case: continental summer convective daily precipitation

• NCEP reanalysis 2.5 degree atmospheric predictors

• Station derived precipitation


Exploring the dependencies.

Transfer function based methodology (Neural nets): gives dominance to synoptic forcing• Problematic case: continental summer convective daily precipitation• NCEP derived predictors• Station derived precipitation

Regional context• steep topography

• elevated inversions

• strong interannual variability

Topography


1980-86 January mean SLP 1970-98 January mean precip

Dominance by semi-permanent high pressure systems with surface thermal trough

Strong spatial gradients of precipitation strongly dependant on moisture transport


Characteristic 7-day back trajectories into test region for downscaling

(shading by specific humidity).


Downscaling methodology:

Transfer function methodology - derives local response as function of synoptic forcing, excludes sub-grid scale local forcing (useful for evaluation of dependencies).- Artificial Neural Nets (analogous to non-linear multiple regression)- derives non-linear transfer functions between NCEP (2.5°) atmospheric variables and precipitation (0.25°)

20 years of training data (1980 - 1999)*

Focus not on optimizing results, but a sensitivity study

* Pre-1980 reanalysis data problematic for southern hemisphere


1: Evaluation of predictor variables (20 examples)

Surface 700hPa500hPatemperature temperature temperaturedivergence divergencedivergence

geopotential height geopotential heightvertical velocity vertical velocity vertical velocityrelative humidity specific humidity specific humidityu wind u wind u windv wind v wind v wind

Each predictor used independently to derive a transfer function to precipitation at 0.25°.

Predictor temporal resolution: 12 hourlyPredictor spatial resolution: 9 grid cells (7.5° by 7.5°) centered on

target location

48 hour antecedent predictor state included


Predictor variable R

Specific Humidity (500hPa) 0.56Vertical Velocity (500hPa) 0.55v wind (700hPa) 0.53Relative Humidity (Surface) 0.53Specific Humidity (700hPa) 0.49Divergence (700hPa) 0.49Temperature (Surface) 0.45Geopotential height (700hPa) 0.44v wind (500hPa) 0.44Divergence (Surface) 0.41Vertical Velocity (700hPa) 0.40Divergence (500hPa) 0.35u wind (Surface) 0.34u wind (500hPa) 0.34Vertical velocity (Surface) 0.34u wind (700hPa) 0.34v wind (Surface) 0.34Temperature (500hPa) 0.30Temperature (700hPa) 0.27Geopotential height (500hPa) 0.19

Results suggest:

Dominant relationship is with mid and upper troposphere humidity and predictors related to vertical motion.


Place Arg Aus Bot Zam Bra Nin3 Cri Ban Mex Chi Atl Por Iow Ger SibLatitude -36 -34 -24 -16 -2 0 10 24 28 30 36 40 42 50 52Season W W W W W D D D D D W W D D D

Rank1 q7 rh7 rh7 q5 rh7 th q7 rh7 v0 q7 z8 v0 q7 v0 rh72 z5 q7 q5 z7 q5 q5 rh7 q7 rh7 v0 z7 z8 rh7 rh7 z53 rh7 q5 q7 z5 q7 z2 q5 z5 q5 rh7 z5 rh7 z8 z7 z74 z7 z5 z7 z8 u8 q8 z7 z7 q7 q5 z2 z5 z5 z5 z25 q5 v0 z5 rh7 d8 rh7 z5 q5 z5 z5 v0 z7 q5 z8 d86 z8 z7 z2 q7 d8 z8 v0 z8 z8 rh7 q7 z2 q57 v0 z2 z8 th d2 th vo0 v8 z7 q7 z2 z7 z28 z2 u8 z2 vo0 v7 z2 v7 q5 q5 d2 q79 th slp v0 u5 d2 q8 u5 u5 th

10 d8 v8 d2 u5 q8 th slp11 v7 u5 u8 d2 u8 slp vo012 u8 v7 v0 d8 d8 d2 v813 v8 d8 q8 d214 d2 u8 z2 u815 u5 u0 u016 vo0 v717 q8 slp

Similar examination of other locations supports the above results.

Suggests predictors should include mid-troposphere indicators of humidity and circulation dynamics.


Based on the above, a set of predictors may be chosen.

eg: Surface temperature, u and v winds700hPa specific humidity and geopotential

heights500hPa specific humidity and geopotential

heights

Trained function results: r = 0.7 predicted mean precipitation: 4.2mm/dayobserved mean precipitation: 3.8mm/day

0

50

100

150

200

250

300

1 6

11

16

21

26

31

36

41

46

51

56

61

66

71

76

81

86

91

96

Days

mm

* 1

0

Observed

Downscaled


Residuals

• From either:Lack of information in predictors (choice or predictor or resolution)Local sub-grid scale forcing unrelated to synoptic state

• May be stochastically modeled (stationarity issues)

0

50

100

150

200

250

300

1 6

11

16

21

26

31

36

41

46

51

56

61

66

71

76

81

86

91

96

Days

mm

* 1

0

Observed

Downscaled


Residuals

• Effect of stochastic addition of residuals to recover the higher frequency source of variance independent of the predictors

• Wmean: mean wet spell duration (number of days)


2: Predictor Spatial Resolution

Test relationship of target variable to atmospheric predictors progressively further away from region of interest.

ANN downscaling using mid-troposphere (700hPa) specific humidity and geopotential height

Predictors drawn from progressively larger regions:a) single NCEP grid cell co-located with targetb) 7.5° by 7.5° window centered on targetc) 15° by 15° window centered in targetd) 22.5° by 22.5° window centered on targete) 30° by 30° window centered on target



Test relationship of target variable to atmospheric predictors progressively further away from region of interest.

ANN downscaling using mid-troposphere (700hPa) specific humidity and geopotential height

Predictors drawn from progressively larger regions:a) single NCEP grid cell co-located with targetb) 7.5° by 7.5° window centered on targetc) 15° by 15° window centered in targetd) 22.5° by 22.5° window centered on targete) 30° by 30° window centered on target

Spatial resolution rSingle cell 0.397.5 x 7.5 0.5615 x 15 0.5722.5x22.5 0.5430x30 0.55

Increase in predictor window size, once large enough to represent spatial gradient, has minimal improvement.



Downscaling a function of information content in predictors -- a function of resolution.

eg: Station daily rainfall downscaled from MRF (1°) atmospheric fields:Average error: 0.5 mm/day

Observed and Predicted Rainfall

0

20

40

60

80

1 9 17 25 33 41 49 57 65 73 81 89 97 105

113

121

129

137

145

153

Day

Ra

infa

ll (

mm

/da

y)

Predicted Observed

Suggests variance from sub-grid scale forcing is minimal in this case


3: Predictor Antecedent State

Test relationship of target variable to inclusion of the antecedent state of atmospheric predictors.

Predictors used as:a) time coincident with targetb) time coincident with target and increasing lag in 12 increments to 96 hours lag.


3: Predictor Antecedent State

Test relationship of target variable to inclusion of the antecedent state of atmospheric predictors.

Predictors used as:a) time coincident with targetb) time coincident with target and increasing lag in 12 increments to 90 hours lag.

0.600.610.620.630.640.650.660.670.680.690.700.71

0 20 40 60 80 100 120

Lag (hours)

Co

rre

lati

on

Lags of at least 24 hours are very beneficial


5: Training data period

Test sensitivity of downscaled function to data used in training.

Case 1 Train on 1980s -- test with 1990sCase 2 Train on 1990s -- test with 1980sCase 3 Train on 1982/83 -- test with 1980s and 1990s


5: Training data period

Test sensitivity of downscaled function to data used in training.

Case 1 Train on 1980sCase 2 Train on 1990sCase 3 Train on 1982/83

For each case, test function on independent decades.

Case 1: Trained on 1980s, predicted 1980s: r = 0.66 mean ppt: +7%

Trained on 1980s, predicted 1990s: r = 0.59 mean ppt: -9%

Case 2: Trained on 1990s, predicted 1990s: r = 0.78 mean ppt: +6%

Trained on 1990s, predicted 1980s: r = 0.51 mean ppt: +18%

Case 3: Trained on 82/82, predicted 1980s: r = 0.33 mean ppt: -34%

Trained on 82/83, predicted 1990s: r = 0.11 mean ppt: -28%

Where training data spans the variability, performance good


6: Representing the climate change signal

Predictors that explain the most variance may not be the predictors that capture the climate change signal.

Test: for each predictor, determine the climate change signal

Train on the predictors, and predict from GCM control and future climate simulations


Predictor variable Future - present downscaled % change

Specific humidity (500hPa) 4.49Specific humidity (700hPa) 4.71Surface Temperature 2.43Surface u-wind -5.47Surface v-wind 1.06500hPa geopotential heights 0.26700hPa geopotential heights -1.63

Note: Choice of predictor may change sign of downscaled response


Predictors that explain the most variance may not be the predictors that capture the climate change signal.

Test: for each predictor, determine the climate change signal• Train on the predictors, and predict from GCM control and future climate simulations



Downscaling using: Future - control: +2.1%

Specific humidity (500hPa)Specific humidity (700hPa)Surface u-windSurface v-wind500hPa geopotential heights700hPa geopotential heights

Or excluding humidity: Future - control: -3.5%Surface u-windSurface v-wind500hPa geopotential heights700hPa geopotential heights


Spatial consequences

Downscaled summer precipitation anomaly (future - present)


7: Stationarity:

Predictors: Do future synoptic events have present day representation

Transfer function: Stability of relationship

Sub-grid scale forcing: % contribution to local variance, feedbacks to atmosphere

At a minimum, evaluate predictors ...


7: Stationarity

Consider distribution of 700hPa geopotential height fields in GCM control simulation


7: Stationarity

Frequency of occurrence of each mode may be determined, and change under future climate calculated

% change in frequency of occurrence from future-control simulations:

-53 -13 5 -7 7 -39 -48-16 22 35 7 45 10 -26-24 52 5 62 -28 30 124-39 -32 150 -42 -8 11 127-62 -28 -45 -30 -17 -36 48


7: Stationarity% change in frequency of occurrence from CSM future-control simulations:

-53 -13 5 -7 7 -39 -48-16 22 35 7 45 10 -26-24 52 5 62 -28 30 124-39 -32 150 -42 -8 11 127-62 -28 -45 -30 -17 -36 48

Similarity of future patterns to present day may be determined, and a measure of change in pattern calculated.

% change in pattern from CSM future-control simulations:

5 9 -4 4 13 20 -6-17 12 -1 1 2 3 -28 12 -6 3 11 22 73 1 -4 -14 8 5 931 4 3 4 3 -11 11

Where significant increases in frequency have occurred, variance of pattern modes has generally decreased. Hence: 700hPa geopotential height fields under a future climate are spanned by events in present day simulation


Some conclusions:

Empirical/statistical downscaling has pragmatic attractions.

Appropriate implementation can produce downscaled results consistent to changes in synoptic forcing.

Care is needed!!!

Decision process to implement downscaling

PreparationAre you looking for a sensitivity study, projection, or probabalistic prediction : What is needed versus wanted versus realistic?

Temporal: Resolution and duration. ie: hourly, daily, monthly seasonal, etc., and 1 year through to decadal etc.

Spatial: Resolution and domain. Ie: point or station scale, through regional scale / areal average.

Variable: Direct or derived (eg: temperature versus storm surge)?

If multivariate, is phase matching between variables important?

Do you need statistics or time series?

Source: Has/is an appropriate solution available elsewhere? Has it been evaluated?

Baseline data: What is available (and when)? Does it match all above requirements?

Are any shortcomings to the above critical? What are their implications?


GCM data

Which GCM(s), from where, and how/why are you selecting them? When will they be available?

What SRES or other forcing scenario(s) are used?

Are native temporal and spatial resolutions appropriate to the task? (Recognize skill level is typically > 7-9 grid cells)

Validation (Evaluation): (Essential for understanding what you get in the end!)

Has the GCM been evaluated at the spatial/temporal resolution of intended use?

If not, how will you evaluate it?

To what degree is the GCM future climate statistically stationary?

What skill level/margin of error is acceptable?


Resources

What computational hardware resources are available?

What are your own/team IT skills (programming, script writing, only point and click, system administration, data handling, etc)?


Choice

Do you need to downscale, is direct GCM output ok, is applying a GCM anomaly field to a baseline climatology ok, is interpolation ok? If so, do it!

Choose appropriate downscaling method based on answers above

RCMs or Empirical/statistical


RCM Regionalization

• Which model and why? How fast will it run under available resources? Will it even run on available resources?• Will areal averages (!) be what you need?• What spatial resolution and number of levels.• What map projection.• Domain selection and domain sensitivity.• Has a baseline climatology been run? What boundary conditions were used? What duration used to derive the climatology?• Land surface scheme -- what choice? • Model tuning, has it bben tuned, how and why? • Is stationarity of parameterization important?

Validation (evaluation): (Essential -- especially w.r.t feedbacks!)

•How was/will baseline evaluation done (see under GCM)?•Are the errors acceptable -- do they induce larger problems?•Nested control run climatology -- how long (long enough?)•What are the errors, are they acceptable? Be very careful here, paying attention to feedback processes.•What is the domain topography like -- is the model hydrostatic, does it need to be non-hydrostatic?•What is the driving : nested resolution ratio?•Is domain large enough to recapture sub-GCM grid-scale variance over domain of interest?•Feedbacks: are they recognized? What degree of consequence will ignoring them have on results? (eg: changing veg)•What is the synoptically forced versus locally forced variance ratio.•What is the signal to noise ration of the climate change anomaly?


Empirical downscaling

Note: Can ONLY generate predictand varinace that is inherent in the cross scale relationship with GCM-scale data. The rest is "made up"!!!

• Predictor-predictand relationship: how strong is it?• Are the required predictors available from the GCM?• Does the training/validation data adequately span the variance structure of the climate system?• Is the "synoptic" forced predictand variance enough for the application, do you need to recover locally forced variance?• Do the predictors carry the climate change signal?• Is the relationship strongly non-linear?• What domain size and temporal resolution/duration of the predictors?• Are the GCM predictors stationary, can the degree of non-stationarity be accepted?

• Is multivariate phasing important?• Method: Weather generator, transfer function, weather typing, true analogue, some combination?• Do computational and IT resources meet the methods requirements.• Predictor pre-processing -- yes, no, how, why? (eg, EOFs etc). • Are the pre-processed predictor forms stable -- eg: are the EOFs or climate indices of the training data valid under future climate?• Validation: how will you validate the training procedure? • Independent test versus training data -- where do they fall within data space?• What are the residuals, and biases after training?• Should/do the GCM predictor field need bias correction?• Downscaling function stationarity: can it be tested or evaluated?

Documents

Downscaling / Regionalization Techniques and methodologies AIACC Workshop, Apr 2002 Bruce Hewitson CSAG : University of Cape Town