1. Christian Pagé, CERFACS
Impact of climate change on France watersheds in 2050 : A comparison of dynamical and multivariate statistical methodologies
By :
Christian Pagé, CERFACS
Julien Boé, CERFACS
Laurent Terray, CERFACS
Florence Habets, UMR Sisyphe
Éric Martin, CNRM, Météo-France
Ouranos, 20 May 2008

2. Outline Uncertainties Comparisons against Quantile-Quantile
Problematic of Downscaling
Why use a statistical approach?
Methodology
Statistical Downscaling & Weather Types
Principles & Hypothesis
Validation (also Hydrology)
Application
Impact of climate change on France watersheds
Uncertainties
Comparisons against Quantile-Quantile
Summary & Future
Ouranos, 20 May 2008 2

3. Climate model Precipitations (mm/day) Downscaling Impact model
Perturbed climate meteorological fields ~ 250 km
Climate model
Precipitations (mm/day)
Problematic: Generalities
How can we evaluate impacts of climate change?
Meteorological forcings <10km
Impact model
Precipitations
(mm/day)
Downscaling
Ouranos, 20 May 2008 3

4. Better representation physical processes Much less CPU!
Problematic: Generalities
Downscaling
Two main methodologies
Better representation
physical processes
Statistical relationship:
Local fields & Large-scale forcings
Resolve dynamics and physics:
Numerical model
Much less CPU!
Statistical downscaling
Dynamical downscaling
Talk about pros and cons: Statistical: much less CPU; Dynamical: better physical processes representation
Can be used separately or in combination
Ouranos, 20 May 2008 4

5. Statistical Downscaling Dynamical Downscaling
Downscaling Methodologies
GHG, Aerosols
Statistical Downscaling
Dynamical Downscaling
MCGOA
CNRM-CM3
Boundary Conditions (also Oceanic)
Predictors
Predictors
Regional Model
ARPEGE-VR
F: Calibration
Validation
OBS.
Raw Forcings
Bias correction
Spatialisation
OBS.
Quelles sont les techniques de spatialisation?
Local Forcing variables
Impact Model: ISBA-MODCOU 5

6. Regional Climate Model
Weather Type: southerly winds
Arrows: 850 hPa Wind
Lines: MSLP anomalies
Dynamical downscaling
Precipitation anomalies (%)
Global Climate Model
280 km
Observations
8 km
60 km
Regional Climate Model 6

7. Several Methodologies (Déqué, 2007) ► Perturbation ► Quantile-Quantile
Dynamical downscaling 1)
Climate Change
Bias Correction
Several Methodologies (Déqué, 2007)
► Perturbation
► Quantile-Quantile Δ
Obs.
Scenario
Probability Density Functions
Ouranos, 20 May 2008 7

8. Several Methodologies (Déqué, 2007) ► Perturbation ► Quantile-Quantile
Dynamical downscaling 2)
Bias Correction
Several Methodologies (Déqué, 2007)
► Perturbation
► Quantile-Quantile
Model
Present
Model
Future
OBS.
Probability Density Functions
Ouranos, 20 May 2008 8 8

9. Several Methodologies (Déqué, 2007) ► Perturbation ► Quantile-Quantile
Dynamical downscaling 3)
Bias Correction
Several Methodologies (Déqué, 2007)
► Perturbation
► Quantile-Quantile
Corrected
Model
Present
Corrected
Model
Future
Probability Density Functions
Ouranos, 20 May 2008 9 9

10. Statistical downscaling: General methodology
Global Scale Climate Variable L (predictors) MSLP, geopotential, upper-level wind
Local Geographical Characteristics topography, land-use, turbulence
Local Scale Climate Variable R 10m wind, precipitation, temperature
R = F (L, β)
What is beta?
Local geo charac are included implicitely I assume?
β such that║R – F(L, β)║ ~ Min
F based on Weather Typing
Ouranos, 20 May 2008 10

11. Statistical downscaling: Current methodology
Based on:
NCEP re-analyses
Weather typing
Mean Sea-Level Pressure
Météo-France Mesoscale Meteorological Analysis (SAFRAN)
France Coverage
8 km spatial resolution from coherent climatic zones
7 parameters
Precipitation (liquid and solid)
Temperature
Wind Module
Infra-Red and Visible Radiation
Specific Humidity
SAFRAN 8-km resolution orography
Ouranos, 20 May 2008 11

12. For a given day j in which we know the Large-Scale Circulation
Statistical downscaling: Current methodology
For a given day j in which we know the Large-Scale Circulation
Find closest weather type (daily data)
Euclidian distance over first ten principal components
Select all Ri days of this type
MSLP and Temperature index
Reconstruct precipitation index: using regression of learning period and MSLP of climate model
Temperature index: mean of T over whole France
Regression is over 220 points, with 40 km radius of SAFRAN precip.
Learning data for regression is MSLP of NCEP-REA & SAFRAN precip averaged over 220 pts
Boe J., L. Terray, F. Habets and E. Martin, 2006: A simple statistical-dynamical downscaling scheme based on weather types and conditional resampling J. Geophys. Res., 111, D23106.
Ouranos, 20 May 2008 12

13. Look for analogs (15 days) among all Ri days
Statistical downscaling: Current methodology
Look for analogs (15 days) among all Ri days
Closest in terms of precipitation and temperature index
Belonging to the same decile
Randomly choose one day
Use SAFRAN data for the chosen day
Apply temperature correction if Tindex - TNCEP > 2 C
Correct precipitation (solid/liquid) and IR radiation
Applicable if having long enough observed data time series
How do you explain synthetically the indexes?
Ouranos, 20 May 2008 13 13

14. Is Climate Model simulating correctly Weather Types ? YES
Statistical downscaling: Validation
Is Climate Model simulating correctly Weather Types ? YES
Precipitation mm/day
Period:
Downscaling:
MSLP ARPEGE
A1B Scenario Regional Simulation
SST from CNRM-CM3 model
Safran
Downscaling
DJF
0.6
0.6 7 7
What is TSO?
This downscaling is using MSLP, but also precip+temp indexes I assume.
It is not independent?
JJA
0.5 5
0.5 5 14

15. Strong link with regional climate Simulated correctly by model
Statistical downscaling: Validation: Hypothesis
3 Main Hypothesis
Predictors
Strong link with regional climate
Simulated correctly by model
Statistical relationship F still valid for perturbed climate.
Cannot be validated or invalidated formally. Also true for physical parameterisations and bias correction.
Predictors encompass completely the climate change signal.
Ouranos, 20 May 2008 15

16. Precipitation: 8 weather types
Hypothesis 1: predictors has strong link with regional climate
Precipitation: 8 weather types
Example for 2 winter type
WT1
WT2
MSLP Anomaly
NDJFM
MSLP Anomaly
NDJFM
-16
+16
+16
-16
Data courtesy of Météo-France
Variable de circulation de grande échelle: Pression (MSLP), provenant du projet EMULATE ( , journalier, 5°x5°), précipitations SQR (Météo-France)
Which period is shown here?
What is EMULATE?
Ratio
Pr(reg)/
Pr(moy)
Ratio
Pr(reg)/
Pr(moy)
+3.5
+3.5

17. Hypothesis 1: predictors simulated correctly by model
Winter types :
WT5 (MSLP, composite anomaly in hPa)
Spatial correlation > 0.96 for all weather types
NCEP Reanalyses
ARPEGE GCM-VR
Highlighted box is showing correlation between NCEP and ARPEGE-VR for each WT separately?

18. Perfect Model Validation
Hypothesis 2 & 3: Predictors encompass completely climate change signal
Statistical relationship still valid for perturbed climate
SPRING
Perfect Model Validation
Precipitation mean over France
Precipitation mm/day
Reconstructed
Precipitation amount
change in %
of current mean
(2100_2050) – (2000_1970)
A1B Scenario, Spring
What is difference between bottom pictures?
What is learning period?
Black curve=model, blue curve=downscaling?
What represent size of circles?
-0.35
+0.35 18

19. Precipitation Tendencies ΣPr 1951-2000 Observations vs Reconstruction
Statistical downscaling: Validation
Tendencies ΣPr Observations vs Reconstruction
Color: station latitude South North
Changes of weather type occurrence ► Precipitation Tendencies spatial structures (r=0.92)
Precipitation
What are units mm/dec ?? mm per decade
Precision again on northeastern stations (correlation not as good)
Because not totally explained by weather typing, but was improved because we use also the distance to the weather types. 19

20. RATIO Temperature Tendencies [Reconstructed] / [Observed]
Statistical downscaling: Validation
Temperature
RATIO Temperature Tendencies [Reconstructed] / [Observed]
Period
Data courtesy of Météo-France
Weather Type Occurrence changes cannot explain observed temperature tendencies
► Mandatory to take into account temperature as a predictor
Ouranos, 20 May 2008 20

21. Statistical downscaling: Validation: Hydrology
Flow Validation
150
800
1200
ARIEGE (Foix)
LOIRE(Blois)
SEINE (Poses)
Annual Cycle
OBS
NCEP
ARPEGE-VR
Jan to Dec
Jan to Dec
Jan to Dec
2500
250
LOIRE (Blois)
2500
SEINE (Poses)
ARIEGE (Foix)
CDF
OBS
NCEP
ARPEGE-VR
0 to 1
0 to 1
0 to 1
Cumulative Density Function
NCEP is downscaled using statistical downscaling, ARPEGE-VR also
Bottom: only obs + downscaled NCEP + SAFRAN obs because we are not showing a mean but a time series.
VIENNE (Ingrandes
500
Winter Mean
OBS
NCEP (0.85)
SAFRAN (0.97)
1960
2010

22. Strong link with regional climate Simulated correctly by model
Statistical downscaling: Validation: Summary
Predictors
Strong link with regional climate
Simulated correctly by model
Predictors encompass completely the climate change signal
Need to use Temperature as a predictor
Watersheds flows are correctly reproduced
Annual Cycle
Annual Variability
Cumulative Density Function
Ouranos, 20 May 2008 22

23. Quantifying Uncertainties
Application: Impact of climate change on France watersheds
Black-circled: at least 85% models has sign agreement
Multi-Model relative change of watershed
Flows (%), 2046/2065
Dispersion:
Spatial Mean
σ = 18%
Quantifying
Uncertainties
WINTER: DJF
The Simulated is direct model output or with a perturbation or quantile-quantile method?
Multi-Model relative change of
Downscaled Precip. (%), 2046/2065
Ouranos, 20 May 2008 23

24. Application: Impact of climate change on France watersheds
Relative change precipitation
2046/2065 vs 1970/1999 in Winter
-0.6
+0.6
Statistical downscaling
Dynamical
Quantile-Quantile downscaling
Ouranos, 20 May 2008 24 24

25. Statistical downscaling Quantile-Quantile downscaling
Application: Impact of climate change on France watersheds
Relative change watershed flows
2046/2065 vs 1970/1999 in Winter
-0.5
+0.5
Statistical downscaling
Dynamical
Quantile-Quantile downscaling
Ouranos, 20 May 2008 25 25

26. Summary - 1 Statistical downscaling methodology
Validation is very good
Hypothesis of stationarity (regression)
Weather Typing Approach
Low CPU demand
Evaluate uncertainties with many scenarios
Uncertainties of downscaling method are limited
Those of numerical models are, in general, greater
Ouranos, 20 May 2008 26 26

27. Summary - 2 Ensemble Mean of Watershed flows
Decreases moderately in Winter (except Alps and SE Coast)
2050 : important decrease in Summer & Autumn
Robust results, low uncertainty
Strong increase of Low Water days
Heavy flows decrease much less than overall mean
Ouranos, 20 May 2008 27 27

28. Down the Road… Whole Code Re-Engineering Modular approach
Implement several statistical methodologies
Configurable
End-user parameters
Core parameters
Web Portal
Climate-Change Spaghetti to Climate-Change Distribution
Probability Density Function
Re-sampled Ensemble Realisations
M. Dettinger, U.S. Geological Survey (2004)
Ouranos, 20 May 2008 28 28

29. Christian Pagé, CERFACS
Merci de votre attention! 
Christian Pagé, CERFACS
Julien Boé, CERFACS
Laurent Terray, CERFACS
Florence Habets, UMR Sisyphe
Éric Martin, CNRM, Météo-France
Ouranos, 20 May 2008 29

30. Future Present Application: France watersheds: Snow Cover5 30
Future
Present
Aug
Jul
Aug
Jul
250
500
Aug
Jul
Aug
Jul
Water Equivalent (mm) of Snow Cover
Pyrenees
2055
Grayed zones: min/max 30 30

31. + + - Application: France watersheds: Uncertainties Correlation
Atlantic Ridge
Blocking + ~0
20 days
NAO+
NAO- + -
-20 days
Models
Correlation
Weather Type Occurrence
Precipitation
-0.5
+0.5 31 31

32. Application: Impact of climate change on France watersheds
Relative change watershed flows
2046/2065 vs 1970/1999 Perturbation method
Winter
Corr 0.92
Spring
Corr 0.38
-0.5
+0.5
Summer
Corr 0.86
Autumn
Corr 0.72 32 32

33. Application: Impact of climate change on France watersheds
Relative change precipitation
2046/2065 vs 1970/1999 in Summer
-0.6
+0.6
Statistical downscaling
Dynamical
Quantile-Quantile downscaling
Ouranos, 20 May 2008 33 33

34. Statistical downscaling Quantile-Quantile downscaling
Application: Impact of climate change on France watersheds
Relative change watershed flows
2046/2065 vs 1970/1999 in Summer
-0.5
+0.5
Statistical downscaling
Dynamical
Quantile-Quantile downscaling
Ouranos, 20 May 2008 34 34

35. Domaine classification MSLP (D1)
Régimes de temps et hydrologie (H1)
Définition de régimes/types de temps discriminants pour les précipitations en France
Variable de circulation de grande échelle: Pression (MSLP), provenant du projet EMULATE ( , journalier, 5°x5°), précipitations SQR (Météo- France)
Classification multi-variée Précipitations & MSLP, pas de temps journalier, espace EOF. On conserve ensuite uniquement la partie MSLP pour définir les types de temps.
Domaine classification MSLP (D1)
* 310 stations pour les précipitations
Important: régimes discriminants, classification conjointe de la variable grande échelle avec la variable d’impact, échelle journalière, on peut représenter la variété des régimes de pluie avec 8 à 10 régimes.
8 à 10 régimes de temps !