STARDEXSTARDEX

The lessons learned The lessons learned

http://www.cru.uea.ac.uk/projects/stardex/

……..so far…....so far…..

The STARDEX objectivesThe STARDEX objectives

• To rigorously & systematically inter-compare & evaluate statistical & dynamical downscaling methods for the reconstruction of observed extremes & the construction of scenarios of extremes for selected European regions.

• To identify the more robust downscaling techniques & to apply them to provide reliable & plausible future scenarios of temperature & precipitation-based extremes for selected European regions.

Assembling data sets is time consuming…

but STARDEX now has good data & software resources, publicly available wherever possible

Defining extremes so that everyone is happy is not easy…

We should have given more consideration to dissemination of

data deliverables in the proposal…

but our new DODS working group has come up with a solution…….

ETH Stardex Central Data Archive AbstractThis website provides links and useful information for the development of the STARDEX Central Data Archive.

This site is only for development and testing purposes. The entire content will eventually be moved to the STARDEX website.

This website contains work in progress.

DocumentsCurrent Draft of the Central Data Archive Description PDF html

DataTop level access to the Central Data Archive DODS NetCDF Station data example file (pre.al-fic.st.eth.obs.nc) DODS NetCDF CDL (ASCII) Indices example file (pind.al-fic.st.eth.obs.nc) DODS NetCDF CDL (ASCII)

LinksSTARDEX homepageNetCDF homepageClimate and Forecast (CF) conventions

History2004-08-04 Added CDL files 2004-07-16 Uploaded corrected FIC station files 2004-07-15 Initial version

Identification of methodologies for ensuring consistent and fair

comparisons requires a lot of thinking and planning…

Principles of verification for D12Principles of verification for D12

• Predictor dataset : NCEP reanalysis• Predictand datasets: “FIC dataset” and regional sets• Regions• Stations within regions• Core indices• Verification period: 1979-1993 (for compatibility with

ECMWF-driven regional models)• Training period: 1958-1978 & 1994-2000• Statistics: RMSE, SPEARMAN-RANK-CORR for each

station/index

Although STARDEX Although STARDEX is about downscaling is about downscaling

we have also had to upscale...we have also had to upscale...

D11 Study Regions

England (UEA)P: 13-27 per gpT: 8-30 per gp

German Rhine (USTUTT)P: ~500 per gpT: ~150 per gp

Greece (AUTH)P: 5-10 per gpT: 5-10 per gp

Emilia-Rom. (ARPA)P: 10-20 per gpT: 5-10 per gp

Europe (FIC)481 stations in total

Alps (ETH)P: ~500 per gp

Christoph Frei, ETH

D11 example: French part of Alpine Region

Winter (DJF)Summer (JJA)

Precipitation Indices

Juerg Schmidli, ETH

Spatially coherent Spatially coherent changes in extremes changes in extremes

have occurred over the have occurred over the last 40 years...last 40 years...

1958-2000 trend in frost days1958-2000 trend in frost days

Scale is days per year. Red is decreasing

Malcolm Haylock, UEA/STARDEX

1958-2000 trend in 1958-2000 trend in heavy summer (JJA) rain eventsheavy summer (JJA) rain events

Scale is days per year. Blue is increasing

Malcolm Haylock, UEA/STARDEX

Some of these Some of these changes/patterns are changes/patterns are

consistent with predictor consistent with predictor relationships...relationships...

Heavy winter rainfall and links with Heavy winter rainfall and links with North Atlantic Oscillation/SLPNorth Atlantic Oscillation/SLP

-3

-2

-1

0

1

2

3

4

1955 1965 1975 1985 1995

NAO -R90N PC2

CC1: Heavy rainfall (R90N) CC1: mean sea level pressure

Malcolm Haylock, UEA/STARDEX

In general, predictors are In general, predictors are well simulated by well simulated by

HadAM3P...HadAM3P...

Winter EOFs of winter Z500Winter EOFs of winter Z500HadAM3P (left) and NCEP (right) HadAM3P (left) and NCEP (right)

ARPA-SMR

But when identifying the But when identifying the best predictors, best predictors,

it is easier to make it is easier to make recommendations about recommendations about methodologies for doing methodologies for doing this than the predictors this than the predictors

themselves...themselves...

’’Traditional’ methods work Traditional’ methods work best, e.g., step-wise best, e.g., step-wise regression, correlation, regression, correlation, PCA/CCA.PCA/CCA.

Automated methods Automated methods (neural networks, genetic (neural networks, genetic algorithm) are less suitablealgorithm) are less suitable

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

-0.09 -0.04 0.01 0.06 0.11 0.16

Moisture flux

Pro

ba

bil

ity

of

pre

cip

ita

tio

n

CP02

CP11

CP09

CP01

CP07

CP02 without moisture flux

CP11 without moisture flux

CP09 without moisture flux

Probability of precipitation at station 75103 conditioned to wet and dry CPs

Andras Bardossy, USTUTT-IWS

Handling many Handling many combinations of different combinations of different

methods (20+), regions (7), methods (20+), regions (7), indices (13) & seasons (4) indices (13) & seasons (4)

is difficultis difficult

Partners/regionsPartners/regionsIberia Greece Alps Germany UK Italy

UEA x x x x x x

KCL x x

ARPA-SMR x x

ADGB x

AUTH x x

USTUTT-IWS & FTS x x

ETH x x

FIC x x x x x x

DMI x x x x x x

UNIBE x

CNRS x x

Emilia Romagna, N ItalyEmilia Romagna, N Italy

ARPA-SMR

e.g, D12 – NCEP-based predictorse.g, D12 – NCEP-based predictors

UK – 90UK – 90thth percentile rainday amounts percentile rainday amounts

But results from more But results from more detailed regional analyses detailed regional analyses

will allow us to draw will allow us to draw clearer conclusions...clearer conclusions...

Iberia (16 stations) – Spearman correlations for each Iberia (16 stations) – Spearman correlations for each model and season averaged across 7 rainfall indicesmodel and season averaged across 7 rainfall indices

Iberia (16 stations) Iberia (16 stations) Averaged across all seasons, indices and stationsAveraged across all seasons, indices and stations

55thth & 95 & 95thth percentiles are also shown percentiles are also shown

Spearman correlation Rank of abs(bias)

Stakeholders, policy makers Stakeholders, policy makers and scientists are interested and scientists are interested

in what we are doingin what we are doing

e.g., State of Baden-WurttembergPLANAT Swiss Federal platform on natural disastersSENAMHI, Peru: Climate change scenarios 2004-2050 OURANOS, Canada: Regional climate consortium

So we are trying to publish So we are trying to publish papers and to provide a papers and to provide a range of material on the range of material on the

public web sitepublic web site suitable for suitable for different users.....different users.....

But the challenge now is to But the challenge now is to synthesise everything and synthesise everything and present it in a usable way...present it in a usable way...

D9: Observed trends

D16: Recommendationson robust methods

D18: Summary of changes in extremes

D19: Assessment ofuncertainties

D20: Final projectreport

Robustness criteria for statistical downscalingRobustness criteria for statistical downscaling

Criteria Relative ranking* Comments/Notes Strong predictor/predictand relationships Stable predictor/predictand relationships Physically meaningful predictor/predictand relationships Predictors well simulated by GCMs Coherent/strong change in predictors for 2070-2099 Range of predictors to address stationarity issues1 Stationarity of predictor/predictand relationships: Evidence from analysis of GCM output Theoretical evidence of physical processes

Low frequency variability/trends reproduced Uniform performance of downscaling method across: Stations Regions Seasons

Means and extremes are both well simulated

Application criteria forApplication criteria forstatistical and dynamical downscalingstatistical and dynamical downscaling

Method provides: Y/N Comments/Notes Station-scale information Y Grid-box information - European-wide information - Daily time series - Seasonal indices of extremes Y Temporally consistent temperature and precipitation Y Spatially consistent multi-site information Y Temporally consistent multi-site information - Information at sites with no observations - Method requirements : Relatively high/low Comments/Notes Computing resources High Volume of data inputs Low Availability of input data High

Performance criteria forPerformance criteria forstatistical and dynamical downscalingstatistical and dynamical downscaling

Relative Performance Confidence High Medium Low Temperature

Indices Txav, Tnav, Tav Txq90, Tnq10, Tnfd, Txhw90

Seasons Winter Spring, Summer Autumn Regions W Greece, E Greece

Precipitation Indices Pav Pxcdd, Pnl90 Pq90, Px5d, Pint, Pfl90

Seasons Winter Spring, Summer Autumn Regions W Greece E Greece x

Overall performance: Mean temperature

Temperature extremes Mean precipitation

Precipitation extremes

Good Average Good Poor

Optimal spatial scale: Higher resolution (<2.5o) Recommended impact applications:

What will we learn in the What will we learn in the next 10 months?next 10 months?

http://www.cru.uea.ac.uk/projects/stardex/

How will this feed into How will this feed into ENSEMBLES?ENSEMBLES?