Applying bootstrap and radial basis function neural networks developing a climate change statistical downscaling model

Abstract

This study applies bootstrap method to build a spatial statistical downscaling model. The Kaohsiung meteorological station is employed as a case study to test the performance of the built model. The stepwise regression procedure (SRP) and principal component analysis (PCA) are applied to select the best input variables from the monthly data collected from the Kaohsiung meteorological station and output data of three general circulation models (GCM), including CGMR, CSMK3 and GFCM2. The radial basis function neural network (RBFNN) is then used to as the building block of the models and genetic algorithms (GA) is used to optimize the parameters of RBFNN. Meanwhile, the bootstrap sampling method is used to estimate the uncertainty of the model. Simulated results show that SRP is better than PCA for the input variable selection of CGMR and GFCM2, but PCA is better than SRP for the input variable selection of CSMK3. In general SRP is better than PCA for the input variable selection. The projected average monthly rainfall of Kaohsiung meteorological station shows a trend of slightly decreasing in the future mid-term and long-term in summer compared with historical rainfall. As in winter rainfall of Kaohsiung meteorological station exhibits a slightly increasing trend in the future mid-term and long-term compared with the historical rainfall. It reveals that the overall future rainfall of B1 scenario is decreasing compared with the historical rainfall. Compared with the historical rainfall, simulated rainfall under A2 scenario in both summer and winter is the smallest one which are 1527.8% in winter in mid-term and -94.3% in summer in mid-term, and the most one which are 1769.3% in winter in long-term and -88.2% in summer in long-term, respectively.