In general, statistical downscaling methods are conducted by two steps; 1) checking statistical relationship between a predictand and a predictor in training period, 2) applying the relationship to another period. There are some methods to calculate the statistical relationship such as linear regression, but we adopted Singular-Value Decomposition (SVD) method based on Chu and Yu (2010) in this study. Firstly, the relationship between 850hPa water vapor flux in the reanalysis dataset JRA-55 (Kobayashi et al., 2015, Ebita et al., 2011) as a predictor and observed heavy precipitation data in Kyushu island as a predictand was investigated for the training period of the end of 20th century. Kyushu island locates southwest in Japan and often experiences much rainfall events especially in the rainy season. Secondly, by applying the above relationship to CMIP5 dataset, the future heavy precipitation in Kyushu island in June was estimated for the period of the end of 21thcentury.
As for the selection of the predictor, we compared some atmospheric elements in JRA-55 such as equivalent potential temperature, specific humidity and water vapor flux, and we selected 850hPa water vapor flux as the most suitable predictor by evaluating cross-correlation coefficient, root mean squared error and variance ratio.
As a result, the frequency of extreme precipitation events and that of extreme dry events are both projected to increase in the future in many CMIP5 models. These future changes are thought to be related to the future change of water vapor flux.