Page 304 - Geochemical Anomaly and Mineral Prospectivity Mapping in GIS

Data-Driven Modeling of Mineral Prospectivity 307

study area by using the same sets of spatial evidence layers represented in a highly
similar fashion (i.e., as numbers of unit cells) in order to demonstrate objectively the
utility of coherent (proxy) deposit-type locations. In fact the application of LDA is
advantageous compared to the application of data-driven EBFs because the former
method, like the other multivariate methods (Table 8-II), can take on raw values (i.e.,
unclassified data) of continuous fields as predictor variables whilst the latter method,
like the weights-of-evidence method (Table 8-I), requires classification of data of
continuous fields.
Classification of data of continuous fields, however, is useful in predictive modeling
of mineral prospectivity in terms of recognising spurious spatial evidence. For example,
Carranza et al. (2008a) used band4/band6 ratios of ASTER (Advanced Spaceborne
Thermal Emission Reflection Radiometer) data as a set representing hydrothermal
alteration intensity evidence for modeling prospectivity for epithermal Au deposits in the
Cabo de Gata area (Spain). They found that empirical spatial associations of epithermal
Au deposits with classes of high ASTER band4/band6 ratios are partly spurious because
these classes of spatial evidence coincide either with hydrothermally-altered volcanic
rocks or with greenhouses for which the Cabo de Gata area is (in)famous. Therefore,
proper classification of data and proper calibrations or adjustments of classified (i.e.,
categorical) spatial evidence data is useful not only in the application of bivariate
methods, like data-driven evidential belief modeling, but also in the application of
multivariate methods, like LDA, to predictive modeling of mineral prospectivity.
A disadvantage of LDA, like most multivariate methods (Table 8-II), is that the way
predictor variables are integrated (see, for example, equation (8.11)) is chiefly
mathematical but not necessarily logical representations of the inter-play of geological
processes involved in mineral deposit formation. In contrast, in data-driven evidential
belief modeling predictor maps are integrated not only mathematically but also logically
through the application of an inference network, which reflects inferences about the
inter-relationships of processes that control the occurrence of a geo-object (e.g., mineral
deposits) and spatial features that indicate the presence of that geo-object. Hybridisation
of some mathematical methods by incorporation of inference systems thus potentially
results in better predictive models of mineral prospectivity. For example, a hybrid neuro-
fuzzy model of mineral prospectivity (Porwal et al., 2004; Porwal, 2006) is better than a
neural network model of mineral prospectivity (Porwal et al., 2003a; Porwal, 2006).
Application of data-driven models/methods can be problematic in cases of (a) limited
number of occurrences of mineral deposits of the type sought and (b) incomplete
sampling or data over locations of known mineral deposit occurrences. In such cases,
approaches to mineral prospectivity mapping lead to so-called geographic expert systems
(GES), whereby elements of methods usually employed in knowledge-driven methods
(e.g., expert-based assignment of evidential weights, inference-based integration of
evidence) and elements of methods usually employed in data-driven methods (e.g.,
Bayesian ‘updating’ of evidence) are combined in so-called artificial intelligence (AI)
techniques (Duda et al., 1978; Campbell et al., 1982; McCammon, 1989, 1994; Katz,