Prediction of Occurrence of Discrete Events - Geomathematics: Theoretical Foundations, Applications and Future Developments - page 176

that the three binary patterns used are probably conditionally independent of
occurrence of Au deposits in the training areas selected for these two experiments
as indicated by relatively low CI-test probabilities (0.676 and 0.569, respectively).
This CI-test result should be considered together with the fact that, with the
exception of the WLR coefficient for PC3 in Experiment 3, WLR coefficients for
non-binary variables in Experiments 2 and 3 are not significantly different
from zero.
In Experiments 4 and 5, the sums of WLR posterior probabilities exceed
numbers of deposits in the testing areas. In these two situations, WLR results are
probably worse than the corresponding WofE results. The main reason is that WLR
was performed on non-binary variables. Because of regional changes in the spatial
PC3 and aeromagnetic variability patterns, the testing areas include cells with
values that are outside the ranges of these variables in the training areas. As a
rule, posterior probabilities should not be used for prediction of mineral occurrence
unless in the special situation satisfied in Experiments 1 and 2 that the testing area is
the complement of a training area consisting of many small randomly selected cells
so that both training and testing area are representative of the entire study area.
The preceding conclusion also is reached when individual posterior probabilities
are considered. Smallest and largest WLR posterior probabilities are shown in the
bottom rows of Table 5.5 keeping training and testing areas separate for all five
experiments. A further separation is made by distinguishing between largest pos-
terior probabilities for cells with and without deposits, respectively. Smallest and
largest posterior probabilities for WofE in Table 5.4 were the same because there
were only eight unique conditions per experiment (and no deposits for the unique
condition with maximum posterior probability but relatively small area in each
experiment). In general, the largest posterior probability clearly exceeds the largest
probability for unique condition with one or more deposits. Worst-case scenario is
for WLR in Experiment 5 where the relatively unconstrained posterior probability
of 0.3217 in the testing area is nearly 20 times greater than its WofE counterpart
(

0.0162). It is mainly because of discretization that WofE results then are more
realistic than corresponding WLR results.

¼

5.2.5 Training Cells and Control Areas

The best strategy in mineral potential mapping is to base a mineral potential map on
all data available for a study area without geographical separation of training and
testing areas. In general, however, degree of knowledge about existence of mineral
deposits varies from place to place within the same study area according to patterns
that often cannot be quantified explicitly. Under-explored subareas become targets
of new exploration only if their probability index matches that in places where
mineral deposits are known to occur. Discretization, which is not necessarily
restricted to WofE, may help to stabilize extrapolation to relatively poorly explored
subareas by constraining ranges and magnitudes of the values of the variables.