Abstract
This paper examines how thoroughly and effectively the self-organizing map (SOM) can analyze urban residential characteristics. As an artificial neural network, the SOM is able to extract characteristic patterns of multidimensional data by learning their features. Compared with the factor analyses used in previous research, the SOM reflects more information on residential characteristics in the research outcomes by directly classifying the characteristics. In addition, “feature maps” available through the SOM visualize their temporal changes. All districts in a dataset are classified with neurons on the SOM. Moreover, by calculating the statistics on all districts using each neuron, it is possible to visualize spatiotemporal transformations of the features on these maps, which are called “aggregated maps” in this paper.
To demonstrate the thoroughness and effectiveness of the SOM, this paper focuses on transformations of built-up areas in Kobe, Japan, by analyzing residential characteristics before and after the Great Hanshin-Awaji Earthquake in 1995. The earthquake hit the Hanshin area, severely damaging Kobe's inner-city area crowded with wooden houses. For this SOM analysis, a spatiotemporal dataset on the residential characteristics of Kobe was developed from small-area statistics from population censuses conducted in 1990, 1995, 2000, and 2005.
