The urban sprawl dynamics: does a neural network understand the spatial logic better than a cellular automata?

dynamics. Given the complexity and variability of the location behaviours it appears
important to learn from the reality the true factors affecting the single location with
respect to the surrounding conditions.

Recent developments in the natural algorithms, and particularly in Neural Networks,
allow to reverse the approach by learning the rules and the behaviours directly from the
Data Base, following an inductive bottom-up process.

The aim of this paper is therefore to present an integrated approach on land use
dynamics where the transition rules of urban spatial evolution are learnt by Neural
Network. The proposed innovation concerns the heart of the CA itself: the growth rules
searching and identification.

In the paper the potentialities of NN are experimented with two different architectures:
SOM (Self Organizing Maps), (Kohonen 1995) and a set of Supervised NN (Semeion
1998).

SOM allow to investigate the different dynamic behaviors by showing the strengths of
the underpinning relationships with the environment. The classification produced by
SOM identifies the most relevant clusters of cells for transition rules in quantitative and
qualitative terms.

Then, for forecasting purposes, a set of Supervised NN is applied to learn the transition
rules and to produce a possible future scenario of urbanization.

The case study is the south metropolitan area of Milan, whose extension is
approximately 675 Km², which is a rich agricultural area with few historical small
centers. The area is under pressure for the spillover, in fragmented residential and
productive settlements, of Milan.

The paper is organized as follow:

the second section presents a short overview on NN and their potentialities in urban
analysis and forecast; the third paragraph sketches a brief description of the study area
and the GIS used. The methodology is explained in the following forth part which
describes the research path.

Section 5 is devoted to show the NN SOM implementation results. The implementation
of different architectures of Supervised NN is presented in the section 6-8:The input
data and the methodology in section 6; the learning and validation phase in section 7
and the results obtained in prediction in section 8.

Some final comments and perspectives on the adopted approach conclude the paper in
section 9.

<<   1   2   3   4   5   6   >>

More intriguing information