possible scenarios was already experimented by the authors to investigate the complex
structure of urban sustainability in the Italian cities (Diappi, Buscema, Ottanà, 1998).
Supervised training implies a regimen in which the NN is supplied with a sequence of
examples (Xi, Yi), (X2, Y2)... (X, Yk).. of desirable or correct input/output pairs. As
each input Xk is entered into the NN, the “correct output” Yk also is supplied to the
network. In our study the input is given by the territory information at time t and the
“correct output” is the corresponding information at time t+1. Once the NN is trained
and has learned the rules of transition, it will be able to produce the “desired” land use
transformation of the present state of territorial system supplied as Input to the NN.
In self-organizing training, a network modifies itself in response to X Inputs. This
category of training is able to obtain a surprisingly number of information processing
capabilities: development of pattern categories based on clustering, estimation of
probability density functions, development of continuous topological mapping from
Euclidean space to curved manifolds (Hecht-Nielsen 1990). Self-organizing training
includes the Self-Organizing Map (SOM), presented in section 5.
SOM is able to develop a continuous topological mapping f : B ⊂ Rn→ C ⊂ Rm by
means of self-organization driven by Y examples in C, where B is a rectangular subset
of n-dimensional Euclidean space and C is a bounded sub set of m-dimensional
Euclidean space, upon which a probability density function ρ (Y) is defined. In the
paper their ability to classify has been used to distinguish the prototypical land use
dynamics in the case study area.
3. The study case, the Data and the GIS
The southern ring of metropolitan area of Milan presents large extensions of tilled land
and natural parks with rare urban centres historically grown on agricultural activities.
More recently, in the 70’ties the area has undergone a rapid urbanization process,
principally produced by spill-over effects from the city of Milan.
The scattered and dispersed form of both residential and industrial new settlements is
rapidly producing an high land consumption which is compromising the productivity of
one of the richer agricultural areas in Europe. The forecast of urban sprawl is therefore a
crucial issue which increases the scientific interest to test a new approach in urban
modelling.
The available GIS on the area concern the land use coverage only at two temporal
thresholds: 1980 and 1994. Even if this is an evident limit, it should be considered that