WMO course Statistics and Climatology Lecture VII

1
WMO course-Statistics and Climatology -
Lecture VII
Dr. Bertrand Timbal Acknowledgement
Dr. Oscar Alves Regional
Meteorological Training Centre, Tehran,
Iran December 2003
2
Statistics and climatology--- An example of
statistical predictions
Statistics and Climatology Lecture VII

• Content
• Regional climate change modelling
• Seasonal forecast a dynamical approach
• Hybrid Statistical downscaling and seasonal
  forecast

3
From global to local climate .
from a GCM grid to the point of interest.
4
A range of technique for Regional Climate Change

• Regional information based on Global Circulation
  Models (GCM)
• Increased resolution of GCM
• GCM with a variable resolution
• Regional Climate Modelling (RCM)
• Empirical (statistical) Downscaling

Courtesy of H. von Storch
5
CSIRO Coupled Model vs. Conformal Cubic forced
atmospheric model
Mark 3 CC4.8
Winter (JJA) Rainfall
Courtesy of J. McGregor
6
Regional Climate Models (RCM)
Courtesy of H. von Storch
Regional atmospheric modelling nesting into a
global state
7
Regional Climate Models (RCM)

• RCM are nested within a GCM
• Alternatively can be nested within analyses
• Ocean and Sea-ice components can be coupled
• Regional Climate System Models
• Usually SST forcing is used from observation
• Resolution is increased compare to the host
  model
• (around 50 km) and can be doubled nested (down to
  10 km)
• The RCM must exhibit improve climatology
• Some limitations arise from
• The lateral boundary impact and noise
• The parameterisations used in the GCM and RCM
• Unbalance initial state requires an integration
  to reach equilibrium
• The lack of double way interactions

8
Global model Regional model
Explained variance
Insufficiently resolved
Insufficiently resolved
Well resolved
Well resolved
Courtesy of H. von Storch
Finner spatial scales
Added value By RCM
Empirical Downscaling
9
The framework for empirical downscaling
In case of dynamical downscaling, the model F is
not fitted to data but based on first principles
and parameterizations of unresolved processes.
Courtesy of H. von Storch
10
Assumptions made for empirical downscaling

• Relies on large-scale predictors for which
  Climate System Models are most skilful
• Several grid lengths
• Tropospheric variables (away from the surface)
• Dynamic variables (geopotential, wind,
  temperature)
• The transfer function must remain valid in
  different climate conditions
• Hard to demonstrate
• Can be evaluated by comparison with other
  approaches
• The predictors must encompass the entire climate
  change signal
• Importance of testing several predictors
• Uncertainties related to the choice of predictors

11
Predictors

• Large scale fields which drive the local climate
• Variables well represented by Climate Models
• Statistical model validation
• Data quality
• The length of the dataset
• Stability in time
• Re-analyses NCEP, ERA
• Applying the SM Coupled GCMs Control and
  Transient simulations

12
The downscaling technique analogue
From a pool of analyses NCEP/NCAR
re-analyses (1958-2001)
Daily synoptic situation
Analogue Smallest Euclidean Distance
From NCEP reanalyses From coupled GCM runs
Observed surface parameters for the same period.
Associate surface parameter observations
13
An example of multi-comparison of downscaling
approaches

• Compare dynamical and several
• statistical downscaling

• Benefit in using detailed
• large-scale predictors
• Strengths and weaknesses
• of each type of approach

Analogue NHMM
14
Example Observed trends in the SW of WA

• Linear trend from in Winter (MJJ) 1958 to 1977,
  in mm/year-1
• Total rainfall has drop by 10 to 20
• Most of the reduction occurred in the 60s and
  70s
• Slight increase in Spring
• (1.51 mm/year-1)

15
Combined Predictors

• Validation on 79-98 period
• Most skilful combination includes
• A moisture field PWTR
• A dynamical field MSLP
• A index of the air flow Div1000
• Improved the skill
• Rain amounts
• Rain occurrences
• More skilful during Winter

? Summer Winter ?
16
Reproduction of the observed drying trends

• Over the 1958 to 1998 period, the drying trend is
  smaller
• The analogue technique reproduces the observed
  trend globally
• But reduce differences between local stations
• Trend in Katanning under estimated in both period

Mm/year-1
17
Reproduction of the observed drying trends

• Without MSLP, the trend is larger and often over
  estimated
• Without PWTR, the trend has the opposite sign for
  Southern stations
• Without DIV1000, the trend is mostly unchanged
• Small effect for the central stations (Cunderdin,
  Corrigin, Kellerberrin)

Mm/year-1
18
Improved rainfall series versus direct GCM
outputs

• Over estimation of rain days by GCMs is corrected
• Dry bias for total rain amount is also corrected

Total Rain Amount
Fraction of Rain Occurrence vs. Obs.
19
Reproduction of Wet spells
Observation Direct model outputs Downscaled GCMs
P (Wet gt days) / Total (Wet days)
20
Prediction of climate change impact on rainfall

• Net reduction in SWC of winter rain amounts
• Coherent scenarios 16 to 19
• GCMs indicate a larger spread
• a reduction of 12 to 36

• Rain occurrences in winter shows similar trend
• rainfall reduction is mostly driven by less rain
  days

21
Projections of future trends of Total rainfall
for each stations in the SW of WA
Future drying trend in percentage of total
rainfall with BMRC / CSIRO / LMD
22
Significant reduction of wet spells duration
Observation Control Climate Climate Change
scenarios
P (Wet gt days) / Total (Wet days)
23
Statistics and climatology--- An example of
statistical predictions

• Content
• Regional climate change modelling
• Seasonal forecast a dynamical approach
• Hybrid Statistical downscaling and seasonal
  forecast

24
POAMA Predictive Ocean Atmosphere Model for
Australia

• Global coupled model GCM seasonal forecasting
  system
• Joint project between BMRC and CSIRO Marine
  Research
• Partly funded by the Climate Variability in
  Agriculture Program (CVAP)
• Run in real-time by BoM operational section since
  1st October 2002
• Operational products issued by the BoM National
  Climate Centre (NCC)
• Experimental products available on the POAMA web
  site
• www.bom.gov.au/bmrc/ocean/JAFOOS/POAMA

25
POAMA Web-site
www.bom.gov.au/bmrc/ocean/JAFOOS/POAMA
26
Experimental results pages Will also include
ocean analyses
27
Introduction- POAMA operational system
Observing network
Obs/data Assimilation
Model
Forecast/products
Daily NWP Atmos. IC
Real-time ocean assimilation latest ocean/ atmos
obs 9-month forecast once per day Ensemble
forecasts
Atmos. Model T47 BAM (unified)
Atmospheric observations
Coupler OASIS
Ocean observations
Ocean assimilation - Temp. OI every 3 days
current corrections
Ocean Model ACOM2 (MOM2)
28
POAMA Operational System
T-10 days
Today
Main ocean assimilation (GASP forcing, GTS T obs)
Catch-up ocean assimilation
Coupled forecast
GASP atmospheric/land IC
29
Skill of SST Predictions
Hind-casts one forecast per season, 1987-2001
(60 cases)
rms error (solid) Standard deviation (dashed)
Anomaly correlation
Green - model, red - anomaly persistence
30
2 months
Anomaly Correlation
4 months
6 months
31
Nino 3.4 of 97/98 El Nino
Forecasts starting 1st Dec 1996
32
Nino 3.4 of 97/98 El Nino
Forecasts starting 1st Dec 1997
33
Decay of 2002 El Nino POAMA Real-time forecasts
34
Decay of 2002 El Nino Forecasts available at BOM
Dec 2002 SCO
POAMA
ECMWF
NCEP
NASA
35
Statistics and climatology--- An example of
statistical predictions

• Content
• Regional climate change modelling
• Seasonal forecast a dynamical approach
• Hybrid Statistical downscaling and seasonal
  forecast

36
The Hybrid approaches
Current statistical scheme
1/ Direct GCM surface predictions
3/ Downscaling method, using atmospheric
large-scale response
2/ Bridging method using relationship between
SST anomalies and local climate
37
Bridging Pacific Tropical SST anomalies

• SST over the NINO4 box ranked into terciles
• T1 ? cold ? P1(Rgtmedian)
• T2 ? neutral ? P2(Rgtmedian)
• T3 ? warm ? P3(Rgtmedian)

1st step (Predictability) - If observed TX
P(Rgtmedian) PX(Rgtmedian)
2nd step (CGCM) If hindcast predict TX
PX aX P1 bX P2 cX P3
38
Skill score for Australian Rainfall
39
The downscaling technique analogue
From a pool of analyses NCEP/NCAR
re-analyses (1958-2000)
Daily synoptic situation
Analogue Smallest Euclidean Distance
From NCEP reanalyses From coupled GCM runs
Observed surface parameters for the same period.
Associate surface parameter observations
40
Surface Parameters rainfall, Tmin and Tmax
From the BOM high quality dataset
Two Extra-Tropical Australian domains SWC
and MDB
41
Statistical model parameters

• Predictors used
• MSLP, Z500
• PWTR
• T850

• Domain size to maximise signal/noise ratio

42
Statistical Model Interannual Variability

• Over the 43 years (1958-2000)
• Higher in the MDB than the SWC
• Higher for Temperature (Tmax or Tmin)
• Consistent across the seasons

Using seasonal mean
43
Interannual Variability
Tmax
At a station level (in Winter) over the 1958-2000
period.
Rainfall
Tmin
44
Predictability
Percent correct hindcast of seasonal anomalies
(for summer) using near perfect (reanalyses)
atmospheric predictors over the 1958-2000 period.
45
Dynamically modeled predictors

• Inter-annual correlation between
• seasonal means (Autumn) from
• the model hindcasts and the
• NCEP reanalysis (18 years).

46
Dynamically modeled predictors
Weak correlations if any (few are negative)
... Few seasons with several reasonnable
predictors
47
Hybrid scheme performances

• Percent of correct
• hindcasts over and
• below the median
• (in Autumn).

48
Hybrid scheme performances
Autumn
Rainfall

• At station level for the best
• cases, inter-annual variations
• are approximately reproduced.

Tmin
Tmax
49
Conclusions

• Bridging SST Pacific anomalies to forecast
  local seasonal anomalies
• Forecast useful and comparable to the
  operational scheme.
• Consistent up to 3 seasons ahead.
• Limited predictability need to search for
  Skillful SST predictions

• Downscaling atmospheric fields to forecast
  regional climate seasonal anomalies
• A lot of potential high predictability
  provide a lot of details on the
  
• season ahead suitable for probabilistic
  forecast.
• Despite a good behavior in the Tropics, the
  extra tropical performances
• of the coupled model limit this approach.