2012 Volume 90 Issue 4 Pages 493-507
On the Korean peninsula, observation is being conducted so densely that the mean distance between stations of the extensive ground-based data network is only 12.7 km. Nevertheless, because of significant mountainous terrain and the fact that most observation sites are situated in low areas rather than mountain tops or ridges, the detailed topographical effect on temperature distribution is not reflected properly. A model using fine-scale grid spacing can represent such a topographical effect well, but due to systematic biases in the model, simulated temperature distribution will be different from the actual observation. This study therefore attempts to produce a detailed mean temperature distribution for South Korea through a method combining dynamical downscaling and statistical correction. For the dynamical downscaling, the Weather Research and Forecast (WRF) model developed by the U.S. National Center for Atmospheric Research (NCAR) is used. We applied a multi-nesting technique to obtain high-resolution climate information (3 km) with a focus on the Korean peninsula. The integration period was 10 years from January 1999 to December 2008. For the correction of systematic biases shown in downscaled temperature, a perturbation method divided into the mean and the perturbation part was used with a different correction method being applied to each part. The mean was corrected by a weighting function while the perturbation was corrected by the self-organizing maps method, which is one of the artificial neural networks method. The results with correction agree well with the observed pattern compared to those without correction, improving the spatial and temporal correlations as well as the root mean square error. In addition, they represented detailed spatial features of temperature including topographic signals, which cannot be expressed properly by gridded observation. Through comparison with in-situ observation with gridded values after objective analysis, it was found that the detailed structure correctly reflected topographically diverse signals that could not be derived from limited observation data.