Page 233 - Geochemical Anomaly and Mineral Prospectivity Mapping in GIS

Knowledge-Driven Modeling of Mineral Prospectivity 235

modeling compared to the modeling techniques explained earlier is that one has to
consider not one but two evidential class scores (e.g., Bel and Unc) simultaneously.
However, with four complementary output maps, evidential belief modeling provides
better evaluation of predictive model performance especially in terms of determining
which input data are problematic; therefore, it provides more pieces of geo-information
that are potentially useful in guiding further exploration work.

Calibration of predictive modeling with multi-class evidential maps
There are three approaches by which predictive modeling with multi-class evidential
maps can be calibrated: (1) modification of evidential class scores; (2) modification of
evidential map weights; and (3) modification of inference networks for combining pieces
of spatial evidence.
The first approach to predictive model calibration is relevant to all three techniques
for modeling with multi-class evidential maps. The second approach to predictive model
calibration is relevant only to multi-class index overlay modeling and fuzzy logic
modeling, because evidential belief modeling does not provide ability to incorporate
evidential map weights in the modeling process. Porwal et al. (2003b) demonstrate
procedures for incorporating evidential map weights in fuzzy logic modeling. For each
evidential map, a map weight is assigned based on subjective judgment of relative
importance of pieces of spatial evidence. For each class in an evidential map, a class
weight is assigned and then the class score is obtained as the product of the map weight
and the class weight. The class scores are then transformed into the range [0,1] by
applying a fuzzy logistic membership function (see equation (7.21) further below). There
are certainly several possible meaningful evidential map weights and evidential class
scores that can be assigned and every modeler surely has different opinions about the
relative importance or weight of pieces of spatial evidence. Different sets of evidential
map weights and evidential class scores result in different mineral prospectivity models,
from which the best predictive has to be determined.
The third approach to predictive model calibration is relevant only to fuzzy logic
modeling and evidential belief modeling, because multi-class index overlay modeling
simply derives the average of weighted class scores. In fuzzy logic modeling and
evidential belief modeling, an inference network can be modified by changing operators
or by changing the combinations of evidential maps to be integrated by a certain
operator. Certainly, one must always evaluate the meaningfulness of an inference
network, but the ability to do so depends strongly on quality of available expert
knowledge. Different inference networks results in different predictive models, from
which the best predictive has to be determined.
Clearly, the generic approach to calibration of knowledge-driven predictive modeling
of mineral prospectivity is trial-and-error or comparative analysis to derive an optimum
predictive model. By ‘optimum’, it is meant that a knowledge-driven predictive model of
mineral prospectivity is geologically meaningful (i.e., consistent with the conceptual
model of mineral prospectivity) and has a high prediction-rate. This quality of an
optimum knowledge-driven predictive model of mineral prospectivity can be achieved