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

unit vector, but also its neighbourhood according to a decreasing function (fig 1b).

SOM NN on square grids of 9,16, 25 nodes have been trained to group the data. The
more explaining output was obtained with a 4x4 nodes grid; this resulted the best in
order to show each group sufficiently different from each other.

c12                                    c13                                    c14

c21                                    c22                                    c23                                    c24

c33

CR80 Cell, residential, 1980

CP80 Cell, productive, 1980

CC80 Cell, commercial, 1980

RD Road distance

NR80 Neighbourhood, residential, 1980

NP80 Neighbourhood, productive, 1980

NC80 Neighbourhood, commercial, 1980

CR94 Cell, residential, 1994

CP94 Cell, productive, 1994

CC94 Cell, commercial, 1994

c43

I Codebooks range, values over the mean

Codebooks range, values under the mean

—— Codebook

Figure 2 - The codebooks

The spatial analysis carried out by SOM has been displayed by:

• cluster profiles and their codebook;

• charts with colour hatched plot of the zones, based on output units assignment.

In figure 2 all the codebooks, as prototypical profiles of each cluster, shows the most
relevant features in land use dynamics. On the x-axis are the variables, on the y-axis
their activation level. On the figure is charted the envelope of the records assigned to
each single cluster and, in yellow line, the codebook.

The colour map (Fig.3) shows the spatial organisation of the classes; it is crucial to
know if cells belonging to the same class are also spatially clustered, or if similar

<<   5   6   7   8   9   10   11   >>

More intriguing information