The tenth IMSC, Beijing, China, 2007 - International Meetings on ...
The tenth IMSC, Beijing, China, 2007 - International Meetings on ...
The tenth IMSC, Beijing, China, 2007 - International Meetings on ...
You also want an ePaper? Increase the reach of your titles
YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.
Answering questi<strong>on</strong>s about empirical downscaling methodologies<br />
<str<strong>on</strong>g>The</str<strong>on</strong>g> increasing demand for regi<strong>on</strong>al scale climate change projecti<strong>on</strong>s, and the increasing<br />
availability of empirical downscaling products, raise important questi<strong>on</strong>s about the role and<br />
influence of methodology choice <strong>on</strong> the derived projecti<strong>on</strong>s. <str<strong>on</strong>g>The</str<strong>on</strong>g>re are a wide range of applied<br />
downscaling methods in the literature, and which seek to accommodate the comp<strong>on</strong>ents of<br />
deterministic and stochastic variance in the local climate resp<strong>on</strong>se to large scale forcing. In<br />
general the methods fall into <strong>on</strong>e of two approaches; using some form of scale transfer<br />
functi<strong>on</strong> to capture the deterministic comp<strong>on</strong>ent and then adding the stochastic element, or<br />
alternatively using a technique such as a weather generator c<strong>on</strong>diti<strong>on</strong>ed in some form to<br />
accommodate the deterministic large scale forcing.<br />
We compare two such methodologies which have matured and been used to provide<br />
regi<strong>on</strong>al scenarios for a broad range of end user applicati<strong>on</strong>s. <str<strong>on</strong>g>The</str<strong>on</strong>g> two approaches are used to<br />
project climate change for four locati<strong>on</strong>s that present a challenge due to locati<strong>on</strong> and complex<br />
local features. <str<strong>on</strong>g>The</str<strong>on</strong>g> four locati<strong>on</strong>s span Africa and include tropical and higher latitudes, coastal<br />
and inland regi<strong>on</strong>s, and complex topography.<br />
Does n<strong>on</strong>linearity bring an improvement in statistical downscaling of daily temperature?<br />
Speaker: Radan Huth<br />
Radan Huth<br />
Institute of Atmospheric Physics, Prague, Czech Republic<br />
huth@ufa.cas.cz<br />
Stanislava Kliegrova<br />
Czech Hydrometeorological Institute, regi<strong>on</strong>al office, Hradec Kralove, Czech Republic<br />
Ladislav Metelka<br />
Czech Hydrometeorological Institute, regi<strong>on</strong>al office, Hradec Kralove, Czech Republic<br />
<str<strong>on</strong>g>The</str<strong>on</strong>g> aim of the c<strong>on</strong>tributi<strong>on</strong> is to evaluate the potential gain in the performance of<br />
statistical downscaling due to introducti<strong>on</strong> of n<strong>on</strong>linearity into the transfer functi<strong>on</strong>s. This issue<br />
has not been resolved so far. <str<strong>on</strong>g>The</str<strong>on</strong>g> n<strong>on</strong>linearity is introduced in two different ways: (i) by using<br />
artificial neural networks (ANNs), namely the multilayer perceptr<strong>on</strong> architecture, and (ii) by<br />
stratificati<strong>on</strong> of the dataset by a classificati<strong>on</strong> of circulati<strong>on</strong> patterns, the linear method being<br />
built in each class separately. <str<strong>on</strong>g>The</str<strong>on</strong>g> study c<strong>on</strong>cerns daily maximum and minimum temperatures<br />
in winter at eight European stati<strong>on</strong>s differing in their geographical and climatological settings,<br />
spread from Ireland to Russia and from northern Finland to Spain. <str<strong>on</strong>g>The</str<strong>on</strong>g> predictors are the 500<br />
hPa heights and 850 hPa temperature defined in a network covering whole Europe and a large<br />
part of the neighbouring Atlantic Ocean. <str<strong>on</strong>g>The</str<strong>on</strong>g> performance of the methods is quantified in terms<br />
of correlati<strong>on</strong>s between the downscaled and observed values. Other measures of<br />
corresp<strong>on</strong>dence, such as mean absolute error or root mean square error, lead to identical<br />
37