Page 303 - Geochemical Anomaly and Mineral Prospectivity Mapping in GIS

306 Chapter 8

mean that the map of discriminant scores based on training set BB (Fig. 8-24A) is better
than the map of discriminant scores based on training set AA (Fig. 8-23A).
Comparing and contrasting the performance of the maps of discriminant scores in
Figs. 8-23A and 8-24A with the maps of discriminant scores in Figs. 8-21A and 8-22A
indicate the following. If 40% of the case study area is considered prospective, then the
maps of discriminant scores in Figs. 8-23A and 8-24A have poorer fitting- and
prediction-rates (Figs. 8-23B and 8-24B) than those of the maps of discriminant scores in
Figs. 8-21A and 8-22A. If 20% of the case study area is considered prospective, then the
maps of discriminant scores in Figs. 8-23A and 8-24A have better fitting- and
prediction-rates (Figs. 8-23B and 8-24B) than those of the maps of discriminant scores in
Figs. 8-21A and 8-22A. Therefore, because mineral prospectivity mapping aims to
constrain the sizes of exploration targets in order to increase the chance of mineral
deposit discovery, the cross-validation results show that predictive modeling of mineral
prospectivity via application of LDA generally produces better predictive models by
using training sets consisting of all known non-deposit locations together with coherent
(proxy) deposit-type locations rather than by using training sets consisting of balanced
numbers of coherent (proxy) deposit-type locations and non-deposit locations. These
findings in the application of LDA are consistent with the findings of Skabar (2005) in
the application of artificial neural networks that using a training set of known deposit-
type and all known non-deposit locations results in better data-driven models of
prospectivity than when using a training set consisting of balanced numbers of deposit-
type locations and non-deposit locations.
Comparing and contrasting the performance of the maps of discriminant scores in
Figs. 8-23A and 8-24A with the maps of integrated Bel in Figs. 8-18A and 8-19A
indicate the following. If 40% of the case study area is considered prospective, then the
maps of discriminant scores in Figs. 8-23A and 8-24A have poorer fitting- and
prediction-rates (Figs. 8-23B and 8-24B) than those of the maps of integrated Bel in
Figs. 8-18A and 8-19A. If 20% of the case study area is considered prospective, then the
maps of discriminant scores in Figs. 8-23A and 8-24A have better fitting- and
prediction-rates (Figs. 8-23B and 8-24B) than those of the maps of discriminant scores in
Figs. 8-18A and 8-19A. These results illustrate that the application of LDA, using
training sets consisting of all known non-deposit locations together with (proxy) deposit-
type locations, generally produces better predictive models of mineral prospectivity than
the application of data-driven evidential belief modeling.
The more-or-less similar performances of the maps of discriminant scores in Figs. 8-
21 to 8-24 and the maps of integrated Bel in Figs. 8-16 to 8-19, depending on the
composition of training sets, is attributed to using the same sets of spatial evidence and
to the application of a scheme of spatial evidence representation (Fig. 8-20) in order to
adapt the spatial evidence layers used in data-driven evidential belief modeling in the
application of LDA. The scheme of spatial evidence representation for GIS-based
application of LDA (Fig. 8-20) was deemed necessary in performing controlled
experiments of bivariate and multivariate modeling of mineral prospectivity in the case