Page 294 - Geochemical Anomaly and Mineral Prospectivity Mapping in GIS

Data-Driven Modeling of Mineral Prospectivity 297

117 non-deposit locations of set 2 (Fig. 8-4) in order to avoid bias against the 86
randomly selected proxy deposit-type locations in training set A. The use of training sets
with equal numbers of deposit-type locations and non-deposit locations is adopted from
(a) the use of equal number of ‘zeros’ (e.g., non-deposit locations) and ‘ones’ (e.g.,
deposit-type locations) in logistic regression analyses when the latter are rare (Breslow
and Cain, 1988; Schill et al., 1993; King and Zeng, 2001) and (b) the suggestion of
Brown et al. (2000) and Porwal et al (2003a) that a gross imbalance between deposit-
type locations and non-deposit locations results in poor recognition of prospective zones
via application of artificial neural networks.
Of the properly calibrated classes of individual evidential data or predictor variables
(Table 8-IV), class ANOM5 (‘no data’) is excluded from the application of LDA
because it can result in multivariate outliers due to missing geochemical data. The
alternative of replacing missing data of a predictor variable with the mean of this
variable is also not considered because various parts of geochemical landscapes cannot
be appropriately represented by uniform mean uni-element concentrations or mean
multi-element scores. Thus, the training set A is left with 79 randomly-selected proxy
deposit-type locations and 81 non-deposit locations, whilst the training set B is left with
86 coherent proxy deposit-type locations and 81 non-deposit locations. The small (i.e., 8-
9%) difference between the numbers of deposit-type locations in training sets A and B is
not remedied because the results are a preliminary indication of the advantage of using
coherent rather than just (i.e., randomly-selected) proxy deposit-type locations in
modeling of mineral prospectivity. Based on the training sets of deposit-type and non-
deposit locations with data for all predictor variables, the predictive models of
epithermal Au prospectivity in the case study area derived via the application of LDA
are, as in the application of evidential belief modeling (Fig. 8-19), cross-validated
against the 13 known locations of epithermal Au deposits.
Table 8-V shows that the discriminant model based on training set B is slightly better
(i.e., lower Wilks’ lambda) than the discriminant model based on training set A. Both
discriminant models based on training sets A and B have common statistically
significant predictor variables. This is probably because most of the 79 randomly-
selected proxy deposit-type locations in training set A are the same as most the 86
coherent proxy deposit-type locations in training set B. However, except for the
standardised function coefficients of the ‘FI’ predictor variables (i.e., classes of
proximity to intersections of NNW- and NW-trending faults/fractures), most of the
standardised function coefficients of the predictor variables in the discriminant model
based on training set B are, to varying degrees, higher than the standardised function
coefficients of corresponding predictor variables in the discriminant model based on
training set A. In particular, the standardised function coefficients for the ‘NNW’
predictor variables (i.e., classes of proximity to NNW-trending faults/fractures) in the
discriminant model based on training set B are much higher than the standardised
function coefficients of the same predictor variables in the discriminant model based on
training set A. However, the standardised function coefficients for the ‘FI’ predictor