Improved Understanding of Urban Sprawl Using Neural Networks

Abstract

It is widely accepted that the spatial pattern of settlements is a crucial factor affecting quality of life and environmental sustainability, but few recent studies have attempted to examine the phenomenon of sprawl by modelling the process rather than adopting a descriptive approach. The issue was partly addressed by models of land use and transportation which were mainly developed in the UK and US in the 1970s and 1980s, but the major advances were made in the area of modelling transportation, while very little was achieved in the area of spatial and temporal land use. Models of land use and transportation are well-established tools, based on explicit, exogenouslyformulated rules within a theoretical framework. The new approaches of artificial intelligence, and in particular, systems involving parallel processing, (Neural Networks, Cellular Automata and Multi-Agent Systems) defined by the expression “Neurocomputing”, allow problems to be approached in the reverse, bottom-up, direction by discovering rules, relationships and scenarios from a database. In this article we examine the hypothesis that territorial micro-transformations occur according to a local logic, i.e. according to use, accessibility, the presence of services and conditions of centrality, periphericity or isolation of each territorial “cell” relative to its surroundings. The prediction capabilities of different architectures of supervised Neural networks are implemented to the south Metropolitan area of Milan at two different temporal thresholds and discussed. Starting from data on land use in 1980 and 1994 and by subdividing the area into square cells on an orthogonal grid, the model produces a spatial and functional map of urbanisation in 2008. An implementation of the SOM (Self Organizing Map) processing to the Data Base allows the typologies of transformation to be identified, i.e. the classes of area which are transformed in the same way and which give rise to territorial morphologies; this is an interesting by-product of the approach.