A neural network approach to the analysis of city systems

This study describes a method for analysing systems of cities and for assessing their sensitivity to change, It is based on the premise that the macroscopic appearance of a city is a result of a larger set of underlying processes which can be indicated by useful variables. Herein, a neural approach makes use of Kohonen's self-organizing maps (SOM) to create a phenomenological model of the (West) German city system, SOMs can display hidden patterns in input data as well as neighbourhood relations among the cities that make up the system. The 171 measurement vectors and 21 variables comprising the city system dataset can be reduced to just four dimensions that represent all relevant features of the system. The SOM technique permits classification of German cities into 24 groups that share common characteristics. By inputting a sequence of small changes to the data about a given city it is possible to observe whether and how it evolves towards the characteristics of another group. Some cities (e.g. Frankfurt, Stuttgart) are relatively insensitive to these data manipulations, whereas others respond quickly (e,g, Nurnberg), It is believed that the former are core representatives of discrete city types. With further refinement and broader application to global datasets, this technique may be useful for identifying cities that are susceptible to perturbations of human-nature interactions, including those that involve environmental hazards and disasters, (C) 1998 Elsevier Science Ltd, All rights reserved.