A first attempt to apply high speed spherical self-organizing map to huge climate datasets

We propose a use of high speed spherical selforganizing map (HSS-SOM) as a new visualization technique for huge climate datasets complementarily and alternatively to empirical orthogonal function (EOF) analysis. In order to reduce computational cost, we change the number and visible positions of neurons in our HSS-SOM method. This method allows us to use a large number of neurons and thus understand climate datasets better than conventional methods. We apply the HSS-SOM to a 4-times daily surface temperature dataset to check the accuracy. Then the HSS-SOM succeeds to map all data into 4 cluster regions in which each cluster region includes data of the same observational time. While 4 cluster regions are also classified by EOF analysis clearly, HSS-SOM locates several observational data points whose spatial patterns are similar to each other but observed at different time in a marginal area of different cluster regions. We also apply the HSSSOM to a huge climatology dataset which is not classified clearly in the EOF analysis. Then the HSS-SOM classifies observational data points of similar spatial patterns in the physical space closely, while those of different spatial patterns far away. These results show HSS-SOM extracts more detailed features than the EOF, and suggest that the HSS-SOM would be a useful tool to study future climate change and its assessment.