pollution is very dependent on the permeability of the lying geological layer, which determine percolation
time and effectivness of self purification of the soil.
Although the depth of the water table is extremely important in vulnerability assessment, it has not been
considered in this experiment, as its value can be assumed to be constant in the study area.
Applying the fuzzy classification framework, the multisource features involved in the classification
process have been qualified by linguistic labels.
These labels are quantified with standard quadratic membership functions, the parameters of which have
been elicited directly from the experts [Ref]. Figure 2 shows the membership function chosen to represent
the linguistic labels high , medium and /ow,qualifying the elevation feature.
The final decision class vulnerability is also modeled as a linguistic variable, assuming the values/ow,
medium and high.
The generation of classification rules proceeds by applying an empirical multistrategy learning
procedure to a training set 7j of 2283 examples. In the experiment described here a generic example is a
pixel represented by the vector u of values assumed by the observables listed in table 1, and jd is the
expert judgement (low. medium , high) expressing the degree of satisfaction with which the pattern u may
be assigned to the decision class vulnerability. Table 3 details the number of pixels off that the experts
have assigned to the vulnerability class with a low , medium and high degree of membership.
A draft set of 34 fuzzy production rules has been generated. Two examples of such rules, one for the case
of high vulnerability, and one for the case of medium vulnerability, are the following:
IF (soil_adj is typel) And (irrigjype is type3j And (veg main is type2) And (permeab. is high)
And (elevation is low) And (slope is low)
THEN (Vulnerability is High)
IF (soiljadj is typel) And (irrigjype is type!) And (vegjnain is type2j And (permeab. is high)
And (elevation is medium) And (slope is low)
THEN (Vulnerability is Medium)
Both antecedents of the two rules correspond to a fiat zone with the presence of rice on a highly
permeable soil; the only difference lies in the elevation of the land and in the associated conseguent: high
elevation and high vulnerability in the first rule, medium elevation and medium vulnerability in the
second. These results are in agreement with the assumption that depression zones are more vulnerable
than higher zones with the same territorial characteristics.
4.0 RESULTS
The generated know ledge base, interpreted by the fuzzy logic inference mechanism, produced a soft
map of ground water vulnerability of the study area. The values of the pixels represent different degrees
of vulnerability in a grey-scale ranging in color from dark (low vulnerability) to light (high vulnerability).
To simplify the vulnerability analysis a threshold mechanism can be introduced to obtain on fuzzy-
gradual vulnerability values a hardened map. Figure 3 shows a hard map in which the degrees of
vulnerability have been sliced into three classes: Low vulnerability( L). Medium Vulnerability (M). High
Vulnerability (H). Unclassified pixels (black coloured pixels inside the study area) correspond to
unproductive areas such as urban zones and water bodies that are not relevant for the study.