1. Downscaling and Uncertainty
Hayley Fowler, Newcastle University, UK
Linda O. Mearns, NCAR
ASP2014, NCAR, Boulder, CO
July 21 - August 6, 2014

2. Overview What is downscaling?
Different methods that are used – advantages/disadvantages
Comparisons
Uncertainties
How (not) to choose a downscaling method? Example applications
UKCP09 weather generator
Towards Climate Services

3. The Uncertainty Cascade or Pyramid
Wilby and Dessai (2010)
IPCC AR4 WG (modified after Jones, 2000, and "cascading pyramid of uncertainties" in Schneider, 1983)
not all layers of the pyramid are equally important, as the Wilby & Dessai illustration might suggest. The relative importance will depend on timescale, region, impact, relevant climate variables and other potential factors. For some impacts, such as health-related heat stress, the uncertainty might actually reduce when considering multiple variables.
The next figure attempts to visualise the cascade of uncertainty in projections of global mean surface temperature using the CMIP5 simulations. Each colour represents a different future emissions pathway (the Representative Concentration Pathways, RCPs) (top layer), with each model producing a different response to the same forcing (middle layer). The lowest layer of the pyramid illustrates the role of internal climate variability. This is seen as additional uncertainty for those models which have run multiple realisations of the same forcing pathway, but unfortunately many have not.
For the near-term ( ), the relative importance of the RCPs is far smaller than the uncertainty in the model response. However, at the end of the century, the RCP uncertainty tends to dominate more. If each simulation was then used to drive a regional climate model or an impact model then an additional layer could be added to represent the next step in the cascade.
This visualisation is potentially complementary to other approaches to describe the relative importance of difference sources of uncertainty in climate projections. Comments on whether this is useful, and suggestions on how it might be improved are very welcome! For example, it might be interesting to look at other climate variables such as precipitation, and at regional projections where variability will be a larger component.
Decision-Making
(Assessment of needs,
decision entry points,
institutional constraints,
politics etc.)

4. Downscaling There is a gap between climate
model resolution and that of
local-scale processes.
Problematic when assessing the impacts of climate change
e.g. hydrology, ecosystems, agriculture.
Downscaling refers to a range of techniques that
aim to bridge this gap.
Downscaling may be classified into a number of types; can classify into three main groups
(1) Dynamical downscaling, using nested atmospheric models at higher resolution than the GCM driving them. May be a regional climate model (50x50 km for example, run for Europe by the Hadley Centre) or a mesoscale model at some 10x10 km. Provides increased resolution of topography and orographic effects. Drawbacks - GCM errors are propagated down, rainfall fields are still too coarse in RCMs, cannot produce long (>100 years) simulations for flood frequency analysis. This method may ultimately provide the best solution, but not in the immediate future.
(2) Modification of observed rainfall by a factor derived from the change in mean rainfall from GCM output. Main drawback is that it retains the current observed variability, no change in extremes or sequences is possible. Also only change the rainday amounts, not the number of raindays.
(3) At WRSRL there is an ongoing programme developing and using a combination of statistical downscaling and stochastic rainfall models to try to avoid these problems.

5. image courtesy of Dr. Andrew Wood, NOAA/NWS NWRFC

6. Downscaling Types General Circulation Models (GCMs)
Statistical / Empirical Downscaling
General Circulation Models (GCMs)
e.g. HadAM3H, ECHAM4
Regression methods
Weather/circulation classification
Stochastic weather generators
Dynamical Downscaling
Regional Climate Models (RCMs)
e.g. HIRHAM, RCAO
Downscaled climate outputs
Change Factors

7. Simple downscaling methods: Analogues
Analogues make use of observed data
Spatial analogue
Select area with climate similar to that predicted
Simple but inflexible: limited by availability
Temporal analogue
Select time period with desired climate
Simple but inflexible: may not have period with predicted properties

8. Simple downscaling methods: Change Factors (Delta method)
Very widely used
Most commonly used method in UK water industry assessments (up to 2009!)
Take change factor between control and future simulations of climate models (GCM or RCM) and apply to observed climate series (e.g. monthly rainfall totals)
More sophisticated use of change factors is with stochastic methods such as weather generators – more later….

9. Simple downscaling methods: Bias correction (local scaling)
raw model output
corrected model output
observed station data
raw model output

10. Simple downscaling methods: Bias correction (QQ correction)
Maraun, 2013

11. Statistical downscaling methods: Transfer functions

12. The UKCP09 Weather Generator
Observed rainfall data
(+ RCM change factors)
(1) Primary variable:
Precipitation (mm)
(2) Secondary variables:
Mean temperature (°C)
Daily temperature range (°C)
Vapour pressure (hPa)
Wind speed (ms-1)
Sunshine duration (hours)
NSRP
RAINFALL MODEL
Observed daily weather data
(+ RCM temperature change factors)
Multiple Simulated Rainfall Series
CRU
WEATHER GENERATOR
Multiple series of simulated weather variables + PET + direct and diffuse radiation

13. Change Factor Perturbation Method
Factors are multiplicative (except for mean temperature)
Hourly stats derived using observed regression relations (fixed for future)
Inter-variable relationships also fixed for future
So, no change information included at higher than daily resolution
Observed statistics X RCM change factors
Mean
Mean
Proportion Dry
Proportion Dry
Variance
Variance etc. X

14. Statistical downscaling
Advantages
Not computationally intensive
Applicable to GCM and RCM output
Provide station/point values
Disadvantages
Lack of long/reliable observed series
Affected by biases in the GCM/RCM
Not physically based e.g. climate feedbacks
Under-estimate variability and extremes
Assume stationary relationships in time

15. Comparison of downscaling methods
We know theoretical strengths and weaknesses of downscaling methods, where systematic inter-comparisons have been made, e.g. STARDEX, no single best downscaling method is identifiable
temperature can be downscaled with more skill than precipitation
winter climate can be downscaled with more skill than summer due to stronger relationships with large-scale circulation
wetter climates can be downscaled with more skill than drier climates
Direct comparison of skill of different methods difficult due to the range of climate statistics assessed in the literature, the large range of predictors used, and the different ways of assessing model performance

16. Largest uncertainties
Choice of downscaling method
Choice of predictor variables (statistical methods)
Lack of predictability (tropics, convective processes dominate)
Driving GCM boundary conditions (dynamical downscaling), parameterisations, structural assumptions, initial conditions etc.

17. How to choose a downscaling method?
Additional comparison studies are not needed
Little consideration given to the most appropriate downscaling method to use for a particular application
Need to define the climatic variables that it is necessary to accurately downscale for each different impact application
Different climates, different seasons and different climatic variables may be more accurately downscaled by using more appropriate downscaling methods

18. Flooding
High
Low
Medium
15minute, 2km
Hourly, 5km
UKCP09 sample applied to rainfall model and Urban Inundation Model
High res spatial rainfall model as shown
Used a more sophisticated method as we had to determine L, M, H projections from the rainfall series as too long to run ‘000s through flood model. Identfied appropriate statistical basis to do this.

19. Damaging winds Baseline ‘Low’ climate projection
‘High’ climate projection
‘Central estimate’ climate projection
35m/s threshold
Small increase in damaging wind events in relation to considerable uncertainties associated with climate model simulations of wind climate.

20. SDSM-DC

21. Use of SDSM
Wilby and Dawson, 2013

22. Towards Climate Services…

23. Towards Climate Services…
Central objective: to take the first step towards the realisation of a European Climate Service.
Researchers, in close cooperation with users, develop and demonstrate local climate services to support climate adaption policies.
Provides climate services for several climate-vulnerable regions in Europe, organized at a sectorial level: cities, water resources, coastal defence and energy production.
Will define, in conceptual terms, how a pan-European Climate Service could be developed in the future, based on experiences from local services and the involvement of a broader set of European decision makers and stakeholders.