Comparison and evaluation of statistical downscaling techniques for station-based precipitation in the Middle East

Several statistical downscaling techniques are intercompared and evaluated with respect to daily station-based precipitation in the eastern Mediterranean/Middle East region. The study introduces unconditioned and precipitation-conditioned SANDRA (Simulated ANnealing and Diversified RAndomization) cluster analysis (SCA) as new downscaling approaches and additionally uses the two widely used techniques of canonical correlation analysis (CCA) and multiple linear regression analysis (MR). For the precipitation-conditioned SANDRA cluster analysis different weights (percentages of contribution to the clustering) are evaluated. Furthermore, two different predictor combinations are used, a simple one only including mean sea level pressure (SLP), and a more complex one additionally including 500 hPa-geopotential heights, 500 hPa-vorticity and 1000 hPa-moisture flux. Analyses are carried out on a daily basis for the main rainy season from November to March for the period 1961–1990. It is shown that SLP, as single predictor, does not perform sufficiently well, but adding further predictors considerably improves model performance in terms of increased explained variance and model stability as well as reduced root mean square error (RMSE). From all selected techniques MR and CCA show the best performance for the SLP-based models, with comparable results for both techniques, whereas precipitation-conditioned SANDRA cluster analysis performs best when further predictors are included. Performance differences between all techniques are generally smaller than those for a particular technique using different predictor sets. Copyright © 2011 Royal Meteorological Society