urban sprawl in the area is a quite recent phenomenon whose interpretation and
description would be biased if based on a longer temporal series of data.
The model uses a regular square grid of 500 m with 2703 cells in total. The land uses
taken into consideration are: residential, commercial, industrial and “green” or unbuilt
land wich denotes rural areas.
In this study the information given to the NN has the same structure of a CA. The
following information for each cell are supplied to the Neural Network:
• Land use of the cell i at time t (1980)
• Land use of the neighbouring cells at time t (1980)
• Land use of the cell i at time t+1 (1994)
The state, of the cell or the neighbourhood, is described in term of share for each land
use with respect to the total surface of the unit. We have processed only the three
urbanized functions (residence, industry, commerce) because the unbuilt, green land use
share is redundant, being a linear combination of the other three.
4. The methodology
The initial idea was to test the approach in a “toy” example, based on a small scale
urbanisation process produced by a CA evolving for explicitly given rules.
Implementing an Associative NN on the system at different time steps our aim was to
test to what extent the NN are able to capture the imposed rules. The small toy was
implemented using with different neighbourhood, size and time lags. At the end the
experiment was successful: the NN was able to understand the CA rules, and relevant
information on the sensitivity of the NN to the Data were also available (Bolchi, Diappi,
Franzini, 2001).
But the same Associative NN, applied to the real Data Base of the south of Milan,
produced very poor results.
The scenario reconstructed in the querying phase, depicted a static situation where even
the estimated new residential cells were much lower than expected.
With an implementation of a different NN, the SOM (Self Organizing Map ) we tried to
investigate the fuzzy clusters of land use dynamics and to find out their prototypical
profiles. These profiles, called codebooks show the different activation levels of the
variables (nodes) allowing to investigate the underpinning relationships among
variables.
Finally, for forecasting purposes a set of supervised NN had been implemented. The