powerpoint-präsentation - ecmwf€¦ · 2.) operational forecasts of the cosmo model (provided by...

Quantile based point-to-point validations

IACETHInstitute for Atmospheric and Climate Science

Johannes Jenkner (ETH Zurich Switzerland)Cornelia Schwierz (University of Leeds UK)

Silke Dierer (MeteoSwiss Switzerland)

3rd International Verification Methods Workshop Reading UK 01/02/2007

Pros and Cons of empirically bias corrected contingency tables

+ =Model

Observations

Matching

Outline

Motivation

Methodology and Proceeding

Present benefits and drawbacks

Application to QPF in Switzerland

Summary

IAC

ETH

Inst

itute

for A

tmos

pher

ic a

nd C

limat

e S

cien

ce

Motivation

x 2

Interpretation of (equitable) categorical scores ?

IAC

ETH

Inst

itute

for A

tmos

pher

ic a

nd C

limat

e S

cien

ce

Perfect matching possible at any gridpoint ?

8mm threshold

8mm threshold

OutlineIA

CET

HIn

stitu

te fo

r Atm

osph

eric

and

Clim

ate

Sci

ence

Motivation




Summary

Methodology and ProceedingIA

CET

HIn

stitu

te fo

r Atm

osph

eric

and

Clim

ate

Sci

ence

Equality

1.) Computation of local precipitation quantiles2.) Bias equal to quantile difference3.) Advanced contingency table4.) Determination of pattern overlap


CET

HIn

stitu

te fo

r Atm

osph

eric

and

Clim

ate

Sci

ence

hits misses / false alarms correct negatives

POD FAR


CET

HIn

stitu

te fo

r Atm

osph

eric

and

Clim

ate

Sci

ence

Hanssen-Kuipers discriminant equal to Heidke skill score:

Binomial distribution of misses:

Variance of HK proportional to 1/N:

→ multiple uncertainty for extreme

quantiles!!!

OutlineIA

CET

HIn

stitu

te fo

r Atm

osph

eric

and

Clim

ate

Sci

ence

Motivation




Summary

Present benefits and drawbacksIA

CET

HIn

stitu

te fo

r Atm

osph

eric

and

Clim

ate

Sci

ence

+ fair comparisons possible+ useful information about model behaviour

(model developers)+ only misses are counted+ accuracy limit can be implemented

(pixels with a negligible absolute difference are omitted)

- definition of quantiles:samples with non-rain data not easy to handle / cut-off for low intensities necessary

- quantiles less intuitive than fixed thresholds (end-users)- extreme quantiles (Q < 0.05 and Q > 0.95) suffer from small sample sizes

OutlineIA

CET

HIn

stitu

te fo

r Atm

osph

eric

and

Clim

ate

Sci

ence

Motivation




Summary

Application to QPF in SwitzerlandIA

CET

HIn

stitu

te fo

r Atm

osph

eric

and

Clim

ate

Sci

ence

Comparison of daily precipitation sums:1.) Observational analysis (provided by Christoph Frei)

- based on ∼ 450 pluviometers- gridded in connection with a monthly climatology

2.) Operational forecasts of the COSMO model (provided by MeteoSwiss)- nonhydrostatic and fully elastic dynamics- 7km grid spacing

→ period of 2001 - 2005


CET

HIn

stitu

te fo

r Atm

osph

eric

and

Clim

ate

Sci

ence

Relative quantile difference (Q90COSMO - Q90obs.) / Medianobs.

0

0

DJF MAM

JJA SON -250 %

-250 %

250 %

250 %

500 %

500 %


CET

HIn

stitu

te fo

r Atm

osph

eric

and

Clim

ate

Sci

ence

Hanssen-Kuipers score for 90%.

DJF MAM

SONJJA 0

0

0.5

0.5

1

1


CET

HIn

stitu

te fo

r Atm

osph

eric

and

Clim

ate

Sci

ence

Spatial means of intensity spectrum

OutlineIA

CET

HIn

stitu

te fo

r Atm

osph

eric

and

Clim

ate

Sci

ence

Motivation




Summary

SummaryIA

CET

HIn

stitu

te fo

r Atm

osph

eric

and

Clim

ate

Sci

ence

Simplification of contingency table

Symmetric setting of HK bene-ficial

Amplitude andmatching errors

independent from each other

Straightforward application3.)

1.) 2.)

4.)

Bias

HK

powerpoint-präsentation - ecmwf€¦ · 2.) operational forecasts of the cosmo model (provided by...

Documents