Predictive Modeling of Mineral Exploration Targets 17
subset of known occurrences of mineral deposits of the type correspond spatially
to the high prediction values resulting from the first subset? Suppose further that
we use the second subset to create a predictive model of mineral prospectivity.
How much of the first subset of known occurrences of mineral deposits of the type
correspond spatially to the high prediction values resulting from the first subset?
This question pertains to and is the essence of the so-called blind testing of data-driven
predictive models of mineral prospectivity (Fabbri and Chung, 2008; Chapter 8 of this
volume). This question also pertains to predictive models of mineral prospectivity
derived by one type or different types of data-driven techniques for predictive modeling
of mineral prospectivity (see Chapter 8). The best possible predictive model of mineral
prospectivity is, generally, the one which has high prediction values corresponding
spatially with the highest proportion or percentage of the known occurrences of mineral
deposits of the type sought.
Model validation thus aims at deriving the best possible predictive model of mineral
prospectivity. Deriving the best possible prediction model of mineral prospectivity
entails model calibration. Procedures for model calibration vary in every step of mineral
prospectivity modeling. Analysis of spatial distributions of mineral deposits of the type
sought (e.g., Carlson, 1991; Vearncombe and Vearncombe, 1999) and analysis of spatial
associations between mineral deposits of the type sought and certain geological features
(e.g., Bonham-Carter, 1985; Carranza and Hale, 2002b; Chapter 6 of this volume) can be
useful in testing and, if necessary, re-defining (thus, calibrating) a conceptual model of
mineral prospectivity and the prospectivity recognition criteria. Prior to the analysis of
predictive model parameters, training deposit-type locations to be used in data-driven
methods of creating predictor maps must be selected (thus, calibrated) systematically
instead of randomly (Stensgaard et al., 2006; Carranza et al., 2008b; Chapter 8 of this
volume). Every data-driven method of creating predictor maps has intrinsic ways of
analyzing and representing (thus, calibrating to reduce) parametric errors of uncertainties
in predictor maps, whereas knowledge-driven methods rely on expert opinion in judging
(thus, calibrating to reduce) parametric uncertainties in predictor maps. Finally, one must
quantify (thus, calibrate) fitting-rate and prediction-rate to characterise the performance
of a mineral prospectivity map (Agterberg and Bonham-Carter, 2005; Chung and Fabbri,
2005). The fitting-rate quantifies the goodness-of-fit between a predictive map of
mineral prospectivity and the training deposit-type locations. The prediction-rate
quantifies how well a predictive map of mineral prospectivity delineates the testing
deposit-type locations. The prediction-rate suggests the ability of a mineral prospectivity
map to direct further exploration activities toward undiscovered mineral deposits of the
type sought. The fitting-rate is pertinent only to data-driven mineral prospectivity maps,
whilst the prediction-rate is pertinent to either data- or knowledge-driven mineral
prospectivity maps. The fitting- and prediction-rates also quantify Type I (false-positive)
and Type II (false-negative) errors in a predictive model. These errors in a predictive
model, if not remedied, could render failure in mineral deposit discovery and thus
investment loss in the succeeding scales of target generation or phase of mineral
exploration. The various procedures for model calibration in every step of mineral