1. Professor A.K.M. Saiful Islam
DSMHT 403: Climate Modelling and Adaptation
Lecture-7: Statistical Downscaling Techniques
Professor A.K.M. Saiful Islam
Institute of Water and Flood Management (IWFM)
Bangladesh University of Engineering and Technology (BUET)
6 December 2016

2. Topics Approach of downscaling Techniques of downscaling
Strength and weakness
Statistical downscaling using SDSM

3. General Approach to Downscaling
Applicable to:
•Sub-grid scales (small islands, point processes)
•Complex/ heterogeneous environments
•Extreme events
•Exotic predictands
•Transient change/ ensembles

5. Types of downscaling
Dynamical downscaling using regional climate modelling
Statistical downscaling
Synoptic weather typing
Stochastic weather generation
Transfer-function approaches

6. Dynamic downscaling
Dynamical downscaling involves the nesting of a higher resolution Regional Climate Model (RCM) within a coarser resolution GCM.
The RCM uses the GCM to define time–varying atmospheric boundary conditions around a finite domain, within which the physical dynamics of the atmosphere are modelled using horizontal grid spacings of 20–50 km.

7. Limitations of RCM
The main limitation of RCMs is that they are as computationally demanding as GCMs (placing constraints on the feasible domain size, number of experiments and duration of simulations).
The scenarios produced by RCMs are also sensitive to the choice of boundary conditions (such as soil moisture) used to initiate experiments

8. Advantages of RCM
The main advantage of RCMs is that they can resolve smaller–scale atmospheric features such as orographic precipitation or low–level jets better than the host GCM.
Furthermore, RCMs can be used to explore the relative significance of different external forcings such as terrestrial–ecosystem or atmospheric chemistry changes.

9. Regional Climate Model
Limited area regional models require meteorological information at their edges (lateral boundaries)
These data provide the interface between the regional model’s domain and the rest of the world
The climate of a region is always strongly influenced by the global situation
These data are necessarily provided by global general circulation models (GCMs) or from observed datasets with global coverage
The nested regional climate modelling technique consists of using initial conditions, time-dependent lateral meteorological conditions and surface boundary conditions to drive high-resolution RCMs. The driving data is derived from GCMs (or analyses of observations) and can include GHG and aerosol forcing. A variation of this technique is to also force the large scale component of the RCM solution throughout the entire domain.
To date, this technique has been used only in one-way mode, i.e. with no feedback from the RCM simulation to the driving GCM. The basic strategy is thus to use the global model to simulate the response of the global circulation to large scale forcings and the RCM to a) account for sub-GCM grid scale forcings (e.g. complex topographical features and land cover inhomogeneity) in a physically-based way; and b) enhance the simulation of atmospheric circulations and climatic variables at fine spatial scales.
The boundary conditions for the PRECIS RCM are clearly an integral part of the system but as they comprise a very substantial amount of data (20-30 Gigabytes for a 30-year simulation) they have to be supplied separately. They are stored online at the Hadley Centre and will be made available on request through a web-based interface. The data will then be supplied on a storage medium specified by the user.

10. Statistical Downscaling
Transfer-function downscaling methods rely on empirical relationships between local scale predictands and regional scale predictor(s). Individual downscaling schemes differ according to the choice of mathematical transfer function, predictor variables or statistical fitting procedure.

13. Types of transfer functions
To date, linear and non–linear regression, artificial neural networks, canonical correlation and principal components analyses have all been used to derive predictor–predictand relationships.

14. Strength and weakness of transfer function
The main strength of transfer function downscaling is the relative ease of application, coupled with their use of observable trans–scale relationships.
The main weakness is that the models often explain only a fraction of the observed climate variability (especially in precipitation series).

15. SDSM Developed by Loughborogh university, UK www.sdsm.org.uk
Data can be downloaded from Canadian Climate Change Scenario network (CCSN)
SDSM is best described as a hybrid of the stochastic weather generator and transfer function methods.

16. SDSM- Statistical Downscaling Model

17. SDSM Algorithm
Optimisation Algorithm: SDSM 4.2 provides two means of optimising the model – Dual Simplex (as in earlier versions of SDSM) and Ordinary Least Squares. Although both approaches give comparable results, ordinary Least Squares is much faster.

18. The User can also select a Stepwise Regression model by ticking the appropriate box.
Stepwise regression: works by progressively adding all parameters into the model and selecting the model which models the predictand most strongly according to one of two criteria: either AIC(Akaike information criterion) or BIC(Bayesian information criterion).

19. The Akaike information criterion is a measure of the relative goodness of fit of a statistical model.
In statistics, the Bayesian information criterion (BIC) or Schwarz criterion (also SBC, SBIC) is a criterion for model selection among a finite set of models.
When fitting models, it is possible to increase the likelihood by adding parameters, but doing so may result in overfitting. The BIC resolves this problem by introducing a penalty term for the number of parameters in the model. The penalty term is larger in BIC than in AIC.

20. Patuakhali Tmin(1961-2001): Predictor variable Partial r
Ncepmslpas (mean sea level pressure)
0.763
ncepp500as (500 hpa geopotential height)
0.308
ncepp850as (850 hpa geopotential height)
0.61
ncepr850as (relative humidity at 850 hpa)
0.383
Mean E%
34.2
Mean SE
1.461