Development of a quantile-based approach to statistically downscale global climate models

Climate Change;Statistical Downscaling;Quantile Regression;Piecewise Regression;Extreme Temperature;Extreme Precipitation

Stoner, Annemarie K.

关键词: Climate Change;    Statistical Downscaling;    Quantile Regression;    Piecewise Regression;    Extreme Temperature;    Extreme Precipitation;   

Others  :  https://www.ideals.illinois.edu/bitstream/handle/2142/26339/Stoner_Annemarie.pdf?sequence=1&isAllowed=y

美国|英语

来源: The Illinois Digital Environment for Access to Learning and Scholarship

【 摘 要 】

Large-scale general circulation models give us an idea of how the climate may possibly develop over the future century. These models generally simulate the large-scale and global mean climate well; however, when applied to localized regions their output does not provide sufficient detail to perform local and regional assessments needed for evaluating necessary mitigation steps. To overcome this weakness I here introduce a novel method of statistical downscaling, which bridges the gap between the low-resolution output provided by climate models and the high-resolution data needed to perform local or regional climate assessments.The statistical downscaling method developed here, which is based on quantile regression, can downscale any variable simulated by AOGCMs and observed on a daily basis that has, or can be transformed into, a Gaussian-like or symmetrical distribution. One of the aspects of the quantile regression technique, along with our enhancements, is a high accuracy in projection of extremes, which often is the sole focus of impact studies when applying the downscaled output. Furthermore, the technique is applicable to both station-based as well as high-resolution gridded observations and can be applied to different types of climate anywhere in the world.The method is here evaluated for minimum and maximum temperature as well as precipitation for 20 stations in North America as well as for high-resolution gridded observations over the continental United States and Alaska.Station-based downscaling is evaluated based on seven different versions of the temperature model and eight versions for the precipitation model, each successive version having one added change or improvement to the downscaling process. Each version is evaluated in terms of three different quantities: the PDFs, giving a visual image of the skill each model; the coefficient of determination, R2, which is a measure of the portion of variance in observations that is reproduced by downscaling; and bias in nine quantiles distributed in order to evaluate both the central part of the distribution as well as the extremes.