P1258900239jcgCB

1
Precipitation Downscaling with SDSM over Rio
de la Plata Basin Mario Bidegain and Madeleine
Renom
2

• Rio de la Plata Basin RPB and selected stations
• The Statistical DownScaling Model (SDSM)
• Downscaling precipitation process
• Preliminary results

3
RIO DE LA PLATA BASIN (RPB) and selected stations
4
The Statistical DownScaling Model SDSM
5
Downscaling of Precipitation Precipitation
occurrence process

• An nth order, two state process governs daily
  precipitation occurrence, with serially
  independent precipitation amounts on wet days

where ?t is the conditional probability of a wet
day, X is a K?1 vector of standard Gaussian
(i.e., normally distributed, with zero mean and
unit variance) explanatory variables, ? is the
coefficient matrix, and ? is random noise.
6
Precipitation occurrence (cont.)
The binary event of precipitation, Pt or no
precipitation is determined by
if ut??t
otherwise
where ut denotes uniform independent random
forcing for the occurrence process (probability
density fu 1, 0? u ?1).

• Notes
• Autocorrelation is incorporated implicitly by
  predictors
• Separate equations may be specified for each month

7
Conditional variables
Conditional variables, including nonzero
precipitation amounts rt are simulated by
where Z is a K?1 vector of standard Gaussian
(i.e., normally distributed, with zero mean and
unit variance) explanatory variables, ? is the
coefficient matrix, and ? is an error term which
is modelled stochastically (by assuming zero mean
and variance equal to model standard error).
8
Conditional variables (cont.)
Many conditional variables as precipitation
amounts are strongly skewed to the right.
Therefore, a range of transformations for rt are
available in SDSM (Version 2.3 and later),
including exponential, fourth root, and inverse
normal.
Illustration of the inverse normal transformation
9
Process to calibrate SDSM and generate downscaled
series
NCEP
GCMs
Global daily reanalysis 1948-2003
Global daily outputs 2000-2100
Predictand variables
Selection of predictand variables (webpage NCEP)
Station daily observed precipitation 1996-2001
Model Calibrated at location 1996-2001
Statistical downscaling 2000-2100
SDSM
SDSM
10
(No Transcript)
11
(No Transcript)
12
(No Transcript)
13
(No Transcript)
14
(No Transcript)
15
(No Transcript)
16
(No Transcript)
17
Monthly daily mean of precipitation observed vs.
generated
Monthly daily variance of precipitation
observed vs. generated
18
Preliminary results

• SDSM provides a technique of scenario
  construction that complements other methods
  (dynamic downscaling)
• Daily precipitation amount at individual stations
  is the most problematic variable to downscale
• SDSM should not be used uncritically as a black
  box (evaluate all relationships using
  independent data), the local knowledge is an
  invaluable source of information when determining
  sensible combinations of predictors
• The plausibility of all SDSM scenarios depends on
  the realism of the climate model forcing