final_program_abstracts[1]

11 IMSC Session Program
Comparisons of extreme parameter estimation between
observed and modelled data, and implications for future
projections
Friday - Poster Session 7
Sarah E. Perkins
CSIRO Marine and Atmospheric Research, Melbourne, Victoria, Australia
Extreme value theory in terms of the Generalized Extreme Value (GEV) distribution
has been used comprehensively to estimate climatic extremes at the global and
continental scale using both global climate model (GCM) and observed daily datasets
(Wehner, 2004; Schaeffer et al., 2005; Kharin et al., 2007; Coelho et al., 2008;
Rusticucci and Tencer, 2008; Perkins et al., 2009). Fitting an extreme value
distribution such as the GEV to future projections by GCMs gives an indication as to
how the frequency and magnitude of extreme events may change in the future, which
is imperative to adaptation studies and policy. Evaluating GCMs on their ability to
simulate observed extreme events is therefore warranted to give some level of
confidence of their future reliability (Perkins and Pitman 2009, Perkins et al., 2009).
To explore this idea, Perkins et al. (2009) modified the evaluation metric introduced
by Perkins et al. (2007) to examine the difference in the daily observed and modeled
Probability Density Function (PDF) tail. This metric uses an arbitrary threshold of the
95 th (5 th ) percentile of the observed maximum (minimum) temperature PDF. This
metric therefore always includes at least observational data which is superfluous when
fitting the GEV, and does not quantify the difference in observed and modeled
extremes if the tails do not overlap in the prescribed range. A new question
consequently arose; how well do GCMs replicate the observed GEV parameters?
This study will compare by ratios each of the observed and modeled GEV parameters
(the location, scale and shape) estimated by the method of L-moments for minimum
temperature, maximum temperature and precipitation. Models evaluated include
CMIP3 models with daily data for the climate of the 20 th century simulation and are
compared to the Australian Bureau of Meteorology observational dataset for a number
of regions over Australia defined by Perkins et al. (2007). The use of these regions
will allow for an evaluation of the GCMs over various climatic types.
Although somewhat simplistic, calculating the observed to modeled ratio of each
GEV parameter quantitatively informs us whether the model is under or over
approximating the distribution and by how much (i.e. the model bias), and is not just
focused on the return levels which are the end product of fitting the GEV. Further, the
resulting ratios may be used in future scenarios to obtain extreme value projections
which may be more realistic than the actual GCM projections, assuming that the
relationship between the model and observations is consistent throughout time. This is
demonstrated by scaling the model-estimated GEV parameters by their respective
ratios for the SRES A2 scenario and comparing the consequential return values from
the scaled and non-scaled data.
Abstracts 321