Page 211 - Introduction to Mineral Exploration

194 C.J. MOON & M.K.G. WHATELEY

on the Internet (Arc-SDM 2001). The differing that it uses presence or absence of geological
complex methods can be divided into two main patterns, rather than a continuous variable.
approaches: (i) data driven, without a precon- The significance of the various geological con-
ceived model; (ii) model driven, dependent on trols is tested and a model built up.
the type of model used. Although the methods In the south Devon example (Fig. 9.14) using
differ, their aim is to map the probability of drainage catchments containing more than
finding mineralisation. 1000 ppb Au generated a probability p of:

Logit (p) = −3.81 + 1.47X5 − 1.74X11 +
Data-driven approaches
2.04X2X5 + 3.07X3X4
The main data driven techniques are (1)
weights of evidence and (2) logistic regression, where X3X4 is combined felsite and
both of which require the location of mineral- Dartmouth Group occurrence, X5 Meadfoot
isation in at least part of the area to be known, Group occurrence, X2X5 Meadfoot Group and
in order to use as a training set for use in the basic volcanics, and X11 faults. In this model,
remainder of the area. The reader can find the felsite and Dartmouth Group factor is the
further details in Bonham-Carter (1994) and at most important and faults are a less important
the Arc-SDM website (Arc-SDM 2001). control.
1 Weights of evidence was pioneered in min-
eral exploration use by Bonham-Carter et al.
(1988). This method uses Bayesian statistics Model-driven approaches
to compare the distribution of the known A number of model-driven approaches have
mineralisation with geological patterns that been used which develop on the index overlay
are suspected to be important controls. theme: (1) fuzzy logic, (2) Dempster Shafer
The first stage in using the method is to methods, and (3) neural networks.
calculate the probability of deposit occurrence 1 Fuzzy logic. This has probably been the
calculated without knowing anything of the most widely used technique. In the fuzzy logic
test pattern (prior probability). The distribution method variables are converted from raw data
of mineralisation in the training area is then to probability using a user-defined function.
compared with geological patterns in turn. The For example, if arsenic soil geochemistry is sus-
calculations for each pattern give two weights: pected as a pathfinder for gold mineralisation,
(i) a positive weight indicating how important then low As concentrations (say <100 ppm As)
the pattern is for indicating an occurrence; and will be given a probability of 0 and high con-
(ii) a negative value indicating how important centrations (say >250 ppm As) a probability of
the absence of the pattern is for indicating 1. Intermediate concentrations will be given a
no occurrence. The weights for the different probability of between 0 and 1 depending on
patterns are then examined to see if they are the function applied (Fig. 9.15). The variables
significant and combined to calculate an over- are then combined using algebra similar to
all posterior probability. This is then used to the Boolean functions but including sum and
calculate the probability of mineralisation product functions. The latter two functions
based on the geological patterns. are often combined together using a gamma-
In the example of the data from south Devon, function to generate an overall probability
the summary weights for different geological from individual layers (Bonham-Carter 1994,
controls are shown in Table 9.6. The most Knox-Robinson 2000). Knox-Robinson (2000)
important patterns using the known gold advocates the use of methods that allow the
occurrences are the occurrence of felsite and visualization of the degree of support of geo-
dolerite which can be seen to dominate the logical evidence. An example of the use of
final probability map (Fig. 9.13). these techniques on Pb-Zn deposits in Western
2 Logistic regression uses a generalized regres- Australia can be found in D’Ercole et al.
sion equation to predict the occurrence of min- (2000). In the south Devon example, a fuzzy
eralisation. Logistic regression is distinct from “or” has been used to combine layers which
the more familiar least squares regression in were estimated to indicate the occurrence of