Self-Organizing Maps (SOM): a clustering neural method for urban analysis
Paola Bolchi, Lidia Diappi, Lorena Franzini · Virtual Community of Pathological Anatomy (University of Castilla La Mancha) · 2001
The research focuses on the testing of a cognitive method for data processing based on Neural Networks. The approach followed is connectionistic and it is applied with the intention to investigate the relationships among urban indicators and understand the structural links of urban systems.The research analyzes the connections in a set of variables for the city of Milan subdivided into 144 statistical areas.The database of Milan is investigated starting from a specific definition of urban sustainability, meant as positive interaction between social, economic and environmental systems, and then collecting a set of indicators aimed to express the relationships between the three systems.During previous researches the same database has been investigated with a methodology of risk measurement based on a threshold approach and with the application of Self-Reflexive Neural Networks. This paper presents a new kind of application in which Self Organizing Maps (SOM) are experimented. SOM carry out a process of spatial organization in which the spatial units are classified on the basis of their own urban characteristics. Network splits up the input records into clusters, in which the main components that differ input data become prevalent.SOM are a typology of Networks that evolves through an unsupervised and self- organized learning process, by elaborating in competition the input units and referring to a winner unit concept. The application of this kind of Networks makes possible to construct maps of Milan by maintaining the spatial disaggregation in statistical areas.The outcomes of SOM also allow to obtain clustering maps where areas are identified both by their strong similarity and by their position in the output matrix (whose dimension has to be determined at the beginning of the neural process).