• The first one, composed by 1662 records of “stable green” cells at both the times, has
been excluded from NN processing.
• The second one, containing 1041 records for the cells urbanised at one of at both
times, has been used for training and prediction.
• The third one, concerning green cells with urbanised neighbourhood at 1994, has
been used only for prediction;
The split into three sets tries to improve the learning capability of the SANN avoiding
the simultaneous presence of records in which the same Input generates different
Outputs. In fact although one of the peculiarity of the SANNs is their ability to deal
with fuzzy behaviours, the process of inconsistent patterns should lead to
misinterpretations and errors.
The 1041 pattern selected for the experimentation have been randomly divided into two
sets (Set 1 and Set 2). Ten different architectures of SANNs have been trained with Set
1 and validated with Set 2. The same SANNs have been also trained with Set 2 and
validated with Set 1. In both cases the SANNs performances have been evaluated
through statistical functions.
In this way it was possible to evaluate the SANNs prediction capability on the whole
1041 records set.
At the end, the average of the 1041 prediction values of the 10 SANNs have been
calculated, and again the prediction capability have been evaluated through statistical
functions. In Fig. 5 is shown a flow diagram of the procedure.
Figure 5 - the training and validation procedure for the SANNs
11