Natural Resources Research Publications on Geochemical Anomaly and Mineral Potential Mapping, and Introduction to the Special Issue of Papers in These Fields

Abstract

In its 26 years of existence, the journal of Natural Resources Research (NRR) has published and continues to publish papers on geochemical anomaly and mineral potential mapping. This is consistent with its aims and scope to publish quantitative studies of natural (mainly but not limited to mineral) resources exploration, evaluation and exploitation, including environmental and risk-related aspects. Over the years, NRR has contributed significantly more to the publication of developments in mineral potential mapping and notably less to the publication of developments in geochemical anomaly mapping. In more detail, NRR has contributed significantly more to the publication of research on development of robust quantitative methods for analysis and synthesis of spatial evidence of mineral potential but notably less to the publication of research on development of geologically focused models of mineral potential. The editorship of NRR recognizes the latter as a challenge to promote further research on development of numerically robust as well as geologically focused mineral potential models, and this special issue is a major initiative in response to that challenge. The recent inclusion of Natural Resources Research for coverage by the Clarivate Analytics (formerly the Institute for Scientific Information) in the Science Citation Index Expanded™ and Journal Citation Reports^® (JCR) Science Edition will help make Natural Resources Research meet that challenge.

Introduction

The journal of Natural Resources Research (NRR) was founded in 1992 by Richard McCammon. Its original name, Nonrenewable Resources, was changed to the present one in 1999, but its objective remained the same: to promote quantitative approaches to mineral resource exploration, assessment, extraction and utilization. However, since its name was changed, the scope of the NRR has also broadened to publish quantitative studies of natural (mainly but not limited to mineral) resources exploration, evaluation and exploitation, including environmental and risk-related aspects. Thus, NRR covers a wide variety of resources including minerals, coal, hydrocarbon, geothermal, water and vegetation. Typical articles make use of geoscientific data or analyses to assess, test or compare resource-related aspects.

Finding new mineral deposits is becoming increasingly more and more difficult, as the most obvious ones have probably been all discovered already. This requires, among others, sustained research and development of methods to facilitate mineral deposit discovery. Significant developments in the geographic information system (GIS) technology during the last four to five decades or so have been useful in the development of quantitative methods of geoscientific data analysis and integration, particularly for geochemical anomaly and mineral potential mapping (Bonham-Carter 1994; Pan and Harris 2000; Carranza 2008). In this article, we review the general developments of GIS-based mineral potential mapping by looking into the yearly variations of published papers in this field. Then, we review the publications of NRR on mineral potential mapping, which is one of the major fields within the scope of this journal. Although NRR is not exclusively dedicated geochemical anomaly mapping, it has published a few articles on this field and these are also reviewed briefly here. Finally, we introduce the papers in this special issue on GIS-based geochemical anomaly and mineral potential mapping.

The field of geochemical anomaly mapping has existed since the early 1970s, and detailed explanations of the science of this field can be found in Levinson (1974), Rose et al. (1979), Govett (1983), Butt and Zeegers (1992), Kauranne et al. (1992), Hale and Plant (1994) and Hale (2000). In very simple terms, this field involves separation of background and anomalous geochemical samples and using the latter group to delineate targets for further exploration of undiscovered deposits. Traditionally, the task of delineating exploration targets based on geochemical anomalies was achieved by stacking on a light table same-scale maps of anomalies of single elements or suites of elements in order to define exploration targets based on intersections of anomalies. Since about three to four decades ago, this task was facilitated by the use of an “electronic light table” (i.e., a GIS). Then, in the late 1980s, the field of geochemical anomaly mapping intuitively evolved into the field of mineral potential mapping, as geochemical anomaly maps need to be integrated with geological datasets in order to delineate geologically meaningful exploration targets (Bonham-Carter et al. 1988, 1989; Agterberg et al. 1990).

Developments in Mineral Potential Mapping

Mineral potential mapping is concerned with quantifying and mapping the likelihood that mineral deposits are present in a study area. It is synonymous to mineral prospectivity mapping, which is concerned with quantifying and mapping the likelihood that mineral deposits may be found by exploration in a study area. Indeed, these two terms—mineral potential mapping and mineral prospectivity mapping—have been used interchangeably in the literature and are hereafter both denoted as MPM.

Thus, MPM involves the collection, analysis and integration of geochemical, geological and geophysical data from multi-sources to quantify spatial relationships between anomalies (i.e., indicators of mineralization) and existing occurrences of mineral deposits of the type sought and use the quantified spatial relationships to map mineral potential or prospectivity. The integration of anomalies, derived from a variety of geo-exploration data, has been done traditionally with the aid of a light table, upon which maps of the same sizes and scale are stacked on top of each to outline potential or prospective areas defined by intersections of anomalies. However, in the last three to four decades or so, MPM has been made more efficient with the aid of a geographic information system (GIS).

The procedure of GIS-based MPM starts with definition of a conceptual model of mineral potential (Fig. 1), which describes theoretical relationships among various factors or controls of how and where certain types of mineral deposits form. The conceptual model guides the choice of spatial geoscience datasets to be used in MPM. Analyses of spatial geoscience datasets to define predictive model parameters can provide feedback to the definition and fine-tuning of the conceptual model of mineral potential (cf. Carranza and Hale 2002b; Carranza 2009a) and eventually to the creation of predictor maps. The integration of predictor maps, using a certain method, yields a map of mineral potential or prospectivity, which needs to be validated in order to judge its predictive ability and, thus, its usefulness for decision-making to guide mineral exploration.

Figure 1

General flowchart of MPM (slightly modified from Carranza (2008))

In general, GIS-based MPM can be either data-driven or knowledge-driven. Methods of data-driven MPM, which involve quantitative analysis of spatial relationships between anomalies (i.e., indicators of mineralization) and existing occurrences of mineral deposits of the type sought, are suitable for “brownfields” or well-explored regions wherein the objective is to define additional targets for exploration. Methods of knowledge-driven MPM, which are based on expert judgment of spatial relationships between anomalies (i.e., indicators of mineralization) and existing occurrences of mineral deposits of the type sought, are suitable for “greenfields” or under-explored regions wherein the objective is to define new targets for exploration. The developments in MPM in the last three to four decades or so have involved testing and application of various methods, considering the assumptions as well as the strengths and limitations of each and every method. The different data- and knowledge-driven methods of GIS-based MPM are listed in Tables 1 and 2, respectively. The methods and references to these methods given in these tables are a representative sampling of the developments in GIS-based MPM as these are the most commonly used and are published mostly in peer-reviewed journals/books (see reference list).

Table 1 Methods commonly used for data-driven GIS-based MPM

Table 2 Methods commonly used for knowledge-driven GIS-based MPM

Developments of methods for data-driven MPM preceded those for knowledge-driven MPM by at least 10 years (Fig. 2). This is largely because MPM is mainly a form of deductive modeling, which involves analysis of patterns from observations (data) in order to derive a model (hypothesis/theory) of mineral potential or prospectivity (cf. Fig. 1). Weights of evidence (WofE) is the most widely used data-driven method of MPM, whereas fuzzy logic (FL) is the most widely used knowledge-driven method of MPM (Fig. 2). The development of WofE has been pioneered by Bonham-Carter et al. (1988, 1989) and Agterberg et al. (1990), whereas the development of FL has been pioneered by An et al. (1991). An analysis of annual publications regarding MPM shows that there have been generally 2–4 papers on WofE modeling of MPM since the development of WofE for MPM in 1988 up to the present, whereas there has been a significant increase in papers on FL modeling of MPM since the development of FL for MPM in 1991 up to the present (Fig. 3). In particular, the significant increase in papers on FL modeling of MPM took place in the last 5–10 years. There is an underlying reason for the growth in research on conceptual (or knowledge-driven) modeling of MPM relative to the decline in research on empirical (or data-driven) modeling of MPM, as discussed below.

Figure 2

Years and numbers of publications of mostly journal papers on methods of (a) data-driven MPM and (b) knowledge-driven MPM. The publications included in these graphs are cited in Tables 1 and 2. DA = discriminant analysis. CA = characteristic analysis. LR = logistic regression. WofE = weights of evidence. FA = favorability analysis. LRA = likelihood ratio analysis. EBF = evidential belief functions. ANN = artificial neural network. SVM = support vector machine. RF = random forest. FL = fuzzy logic. MIO = multi-class index overlay. BL = Boolean logic. BIO = binary index overlay

Figure 3

Years and numbers of publications of mostly journal papers on the applications of weights of evidence (WofE) and fuzzy logic (FL) to MPM. The publications included in the graphs are cited in Tables 1 and 2

The development of GIS-based MPM has been influenced by the publication of three textbooks (Fig. 4), among which the textbook by Bonham-Carter (1994) has been the most influential as it has been cited at least 1500 times (according to Google Scholar). However, the later textbooks by Pan and Harris (2000) and Carranza (2008) certainly have also influenced the development of methods for MPM as, according to Google Scholar, these have been cited at least 100 times and at least 260 times, respectively. The influence of each of these three textbooks seems to be reflected by a decrease in number of journal/conference papers in the years after their respective publication year (Fig. 5), suggesting perhaps that researchers in the field of MPM were studying these textbooks after their publication and then published their findings later on. A similar pattern is associated with the publication of the first paper on WofE for MPM Bonham-Carter et al. (1988). However, it may be difficult to ascertain whether this supposition is true.

Figure 4

Front cover of textbooks relevant to MPM, authored by: (a) Bonham-Carter (1994); (b) Pan and Harris (2000); and (c) Carranza (2008)

Figure 5

Numbers and years of publication of mostly journal papers regarding MPM, as well as years of publication of the first paper on weights of evidence (WofE) modeling of MPM and those of three textbooks relevant to MPM (see text for details). The publications included in the graph are cited in Tables 1 and 2

Nevertheless, on the one hand, the first publication on mineral potential mapping that can be found by a year-to-year search of the literature using the search terms “mineral potential mapping” AND “GIS” in Google Scholar in the Internet is a book chapter by Bonham-Carter and Agterberg (1990) about the application of microcomputer-based GIS to mineral potential mapping. Thus, it can be said that the Canadians, as Bonham-Carter and Agterberg worked then for the Geological Survey of Canada, introduced the term mineral potential mapping. And, on the other hand, the first publication on mineral prospectivity mapping that can be found by a year-to-year search of the literature using the search terms “mineral prospectivity mapping” AND “GIS” in Google Scholar in the Internet is a journal article by Brown et al. (1999) about the use of a multilayer feed-forward neural network for mineral prospectivity mapping. Thus, it can be said that the Australians, as Brown et al. worked then for academic institutions in Australia, introduced the term mineral prospectivity mapping. Indeed, a compilation of papers on MPM in peer-reviewed geoscience journals shows that the Canadians have pioneered the development of GIS-based MPM for three decades or so since the late 1970s, although development in GIS-based MPM has grown to a global scale in the last two decades (Fig. 6).

Figure 6

Source countries of studies on MPM published as articles in peer-reviewed geoscience journals during the last four decades divided into four 10-year slices. Number of articles is denoted by n. The publications included in these charts are cited in Tables 1 and 2

Developments in GIS-based MPM in the last four decades have been published in various journals (Fig. 7); however, about 45% of journal papers on this topic have been published in the journals owned by International Association for Mathematical Geosciences (IAMG), namely NRR, Mathematical Geosciences (MG) and Computers & Geosciences (C&G). This reflects that developments in GIS-based MPM significantly involved the development of robust mathematical methods for analysis and synthesis of spatial evidence of mineral potential. It is remarkable that 30% of journal papers on GIS-based MPM have been published in NRR. It is remarkable as well that 18% of journal papers on this topic have been published in Ore Geology Reviewers (OGR), which is an economic geology journal. Intriguingly, the number of papers on GIS-based MPM published in NRR has decreased significantly during the last decade, whereas that in OGR has increased significantly during the last decade (Fig. 8). This reflects that developments in GIS-based MPM also significantly involved the development of geologically focused models of mineral potential. There is a common underlying reason for these observations and for the earlier observation regarding the growth in research on conceptual (or knowledge-driven) modeling of MPM relative to the decline in research on empirical (or data-driven) modeling of MPM. This common underlying reason is the adoption in GIS-based MPM of the mineral systems approach to exploration targeting.

Figure 7

Distributions of papers on MPM published in peer-reviewed geoscience journals in the last four decades (1977–2016). The journal papers included in this chart are among those cited in Tables 1 and 2

Figure 8

Numbers and years of publication of papers on GIS-based MPM published in Natural Resources Research (NRR) and Ore Geology Reviews (OGR). The journal papers included in this graph are among those cited in Tables 1 and 2

The concept of “mineral systems” for exploration targeting, proposed by Wyborn et al. (1994), describes “all geological factors that control the generation and preservation of mineral deposits, and stress the processes that are in involved in mobilising ore components from a source, transporting and accumulating them in more concentrated form and then preserving them throughout the subsequent geological history.” The mineral systems approach to exploration targeting is, according to Walshe et al. (2005), a paradigm that requires answers to five questions in order to guide mineral exploration, namely: (1) What is the architecture/size of the system?; (2) What is pressure, temperature and geodynamic history of the system?; (3) What is the nature of the fluids and fluid reservoirs in the system?; (4) What are the fluid pathways and drivers of fluid flux?; and (5) What are the transport and depositional mechanisms? Therefore, the mineral systems approach to exploration targeting focuses on three geological elements (or processes) that are critical to the formation of mineral deposits, namely: source of metals, fluid pathways and traps. For GIS-based MPM, these critical elements must be translated into mappable criteria (or spatial proxies) of mineral prospectivity (cf. McCuaig et al. 2010; Porwal and Kreuzer 2010; Porwal and Carranza 2015). Research on the adoption in GIS-based MPM of the mineral systems approach to exploration targeting has grown significantly in the last decade (2007–2016) (Fig. 9). Because the mineral systems approach to GIS-based MPM is intuitively a knowledge-driven approach and because FL is the mostly widely used (and perhaps the most efficient) method for knowledge-driven MPM, there is a statistically significant positive correlation (r = 0.6, p < 0.05) between the number of papers on FL modeling of MPM (Fig. 3) and the number of papers on mineral systems approach to MPM (Fig. 9) during the last decade (2007–2016).

Figure 9

Numbers and years of publication of journal papers yielded by a document search (as of March 11, 2017) via Google Scholar using the terms (a) “mineral systems” AND “mineral potential mapping” AND “GIS” (denoted as MSA-1) and (b) “mineral systems” AND “mineral prospectivity mapping” AND “GIS” (denoted as MSA-2)

Prior to the development of the mineral systems approach to GIS-based MPM (i.e., mainly prior to 1997; Fig. 9), conceptual models of mineral prospectivity for GIS-based MPM (Fig. 1) were based mostly on mineral deposit models (e.g., Cox and Singer, 1986; Roberts et al., 1988), which describe the typical spatial attributes (e.g., geological characteristics) of certain types of mineral deposits as well as their regional geological environments. Because the typical spatial attributes of certain types of mineral deposits may not be relevant to particular deposit types in particular study areas (i.e., mineral deposits are unique even though they are grouped into types according to their similarities), focusing on mineral systems (i.e., using spatial proxies of metal source, fluid pathways and traps) instead makes GIS-based MPM more process based and more geologically focused. Therefore, it can be said that there have been mainly two stages in the development of GIS-based MPM, namely: (1) an earlier stage devoted mostly to the development of robust quantitative methods for analysis and synthesis of spatial evidence of mineral potential and (2) a later stage devoted mostly to the development of geologically focused models of mineral potential. Whereas the initial stage was motivated by the Canadians, the second stage was stimulated by the Australians. The boundary between these two stages is fuzzy, and in fact these two stages strongly overlap each other because undoubtedly researchers who have contributed to the development of GIS-based MPM have endeavored and continue to endeavor to develop numerically robust as well as geologically focused mineral potential models; however, the initial stage was mainly during 1977–2006, and the second stage is mainly during 2007–present.

Accordingly, although NRR has significantly contributed to the publication of developments of GIS-based MPM in general (Fig. 7), NRR has contributed more significantly to the promotion of the initial stage and less significantly to the second stage (Fig. 8). The editorship of NRR recognizes the latter as a challenge to promote further the development of numerically robust as well as geologically focused mineral potential models. In the following sections, NRR articles relevant to GIS-based geochemical anomaly mapping and MPM are reviewed briefly in chronological order of publication.

NRR Publications on Geochemical Anomaly Mapping

Only a few articles relevant to geochemical anomaly mapping have been published in NRR, as it is not exclusively dedicated to the field of exploration geochemistry.

Cheng et al. (1996) used the concepts of fractal/multi-fractal dimensions and fractal measure to estimate prior and posterior probabilities of a small unit cell in a study area to contain at least one mineral deposit occurrence. This resulted in a new version of the WofE method for MPM, which can also be used to map geochemical anomalies.

Chen and Zhao (1998) subjected multi-element lithogeochemical data to factor analysis to understand zonation in primary halos around a gold deposit in China and to use the results of the analysis for mapping multi-element anomalies to support mineral exploration.

Costa and Koppe (1999) subjected pedogeochemical data to geostatistical analysis to derive probability maps, indicating uncertainty of geochemical anomalies, to select areas for further exploration.

Cheng et al. (2000) proposed a novel method for mapping geochemical anomalies by integration of spatial and spectrum analysis. The proposed method was demonstrated using soil geochemical data from an area in Sumatra (Indonesia).

Singer and Kouda (2001) discussed unsupervised and supervised learning methods for extracting useful information from exploration geochemical data.

Twarakavi et al. (2006) presented supervised learning of geochemical anomalies under condition of sparse data and showed that support vector machines and robust least-square support vector machines perform better than neural networks and kriging techniques.

El-Makky and Sediek (2012) subjected multi-element stream sediment geochemical data from Egypt to Q-mode cluster and R-mode factor analyses as well as to enrichment factor analysis to map anomalies associated with gold–sulfide deposits and their associated hydrothermal alteration zones. They found that R-mode factor analysis, compared to Q-mode cluster analysis, provided easily and reasonably interpretable results, whereas the enrichment factor analysis produced results indicative of a promising area for further detailed exploration.

He et al. (2013) demonstrated that, using multi-fractal and frequency distribution analyses, geochemical exploration data of most elements from secondary media (e.g., soils, stream sediments) should be modeled with nonlinear mathematical methods or should be transformed to linear distributions before modeling with linear mathematical methods.

Luz et al. (2014) subjected Cu- and Zn-soil geochemical data from an area in Portugal in order to map anomalies using the concentration–area fractal model, which has been proposed by Cheng et al. (1994).

None of the foregoing articles have addressed the closure problem inherent in the statistical analysis of compositional data, such as exploration geochemical data (Aitchison 1984). As this is a significant aspect of geochemical anomaly mapping, we have, since the publication of the last article above, ensured that no articles involving geochemical anomaly mapping gets published in NRR without “opening” of exploration geochemical data by log-ratio transformation prior to any kind of statistical analysis.

The above-reviewed nine articles represent only ~1% of all (=740) articles NRR has published since 1992. However, two of the above-reviewed articles are among the most cited/applied in this field (Table 3). This information suggests that NRR can potentially make a strong impact to research on geochemical anomaly mapping.

Table 3 Top 15 most cited articles on geochemical anomaly mapping obtained by a document search in Scopus (https://www.scopus.com) using the search terms “geochemical anomaly” AND “GIS” (as of May 24, 2017)

NRR Publications on Mineral Prospectivity Mapping

NRR Articles on Data-Driven MPM

Agterberg (1992) explained the WofE method, which was proposed by Bonham-Carter et al. (1988, 1989) and Agterberg et al. (1990), for integrating evidential layers for MPM with emphasis on undiscovered deposits’ effect on the calculation of weights and posterior probabilities. He distinguished between statistics that are sensitive to unit cell size and statistics that are more-or-less independent of it. He concluded that positive weights of evidence in the WofE method can be compared with coefficients of evidence in the logistic regression (LR) method.

Goossens (1993) used remote sensing (Landsat TM, airborne magnetic and radiometric) datasets to assess potential for granite-related mineralization in a 20-km × 20-km area in Spain. The nature of pixels surrounding a pixel classified as having potential was considered when deciding whether classification is correct. Weights were assigned to uncertainty of interpretation. Interpreted and weighed data were integrated in supervised classification to derive a probability map highlighting zones with potential for granite-related mineralization that were confirmed by all datasets.

Pan (1993b) demonstrated the indicator favorability theory in a case study for data-driven MPM in order to account for spatial correlations of geophysical, geochemical and geological datasets with assumption that mineral potential can be modeled by a combination (Θ) of a set of response variables derived from the datasets. The indicator favorability theory estimates mineral prospectivity of every location in two stages: (1) estimation of a linear combination of response variables by maximizing variance var(Θ) and (2) estimation of favorability function (F) by minimizing variance var[F − Θ]. The first stage is essentially a principal components analysis (PCA) of data at known deposit locations, and the coefficients of the target indicators are applied to the corresponding datasets covering the entire study area to map mineral potential.

Pan and Porterfield (1995) demonstrated a comprehensive methodology particularly for large-scale MPM in order to design optimal in-fill drilling (to check continuity of ore-grade and convert a part of geological resources into minable reserves) and select step-out drilling targets (for finding new orebodies around known ore deposits) in a gold-mining district. The central information synthesizer is canonical or indicator favorability analysis. The study resulted in delineation of several drilling targets.

Cheng et al. (1996) used the concepts of fractal/multi-fractal dimensions and fractal measure to estimate prior and posterior probabilities of a small unit cell in a study area to contain at least one mineral deposit occurrence. This resulted in a new version of the WofE method for MPM. The method was demonstrated in a case study to map potential for gold mineralization in the Iskut River area, northwestern British Columbia (Canada), where occurrences of gold deposits, which have fractal/multi-fractal properties, were integrated with several geochemical, geophysical and geological evidence layers.

Rostirolla et al. (1998) used WofE (Agterberg 1992) and PCA (Pan 1993b) in two case studies of MPM in Brazil. The results of both methods in either case study are fairly similar.

Cheng and Agterberg (1999) demonstrated a new approach of WofE modeling based on fuzzy sets and fuzzy probabilities in a MPM case study for gold deposits in Meguma Terrane, Nova Scotia (Canada). In this new hybrid method, fuzzy sets of subjective genetic elements are generated instead of converting data into binary or ternary evidence; fuzzy probabilities are then defined to derive posterior probability of a unit cell to contain mineral deposits based on fuzzy evidence. The hybrid method provides for objective or subjective definition of a fuzzy membership function of evidence as well as for objective definition of fuzzy or conditional probabilities. In a purely data-driven approach, derived posterior probabilities are completely dependent on existing data, but when the hybrid method is applied, they depend partly on expert knowledge.

Harris and Pan (1999) described a probabilistic neural network (PNN), which is a particular artificial neural network (ANN) architecture designed to compute the probability for membership in each of two or more classes (e.g., mineralized and barren), to classify mineralized and non-mineralized cells using geological, geochemical and geophysical variables. They have shown that the PNN outperforms two traditional multivariate supervised classification methodologies, namely LR and discriminant analysis (DA), for MPM.

Sahoo and Pandalai (1999) used LR to integrate indicator patterns for estimation of the probability of occurrence of gold deposits in a part of the auriferous Archaean Hutti–Maski schist belt. They used data consisting of categorical and continuous variables obtained from a coded lineament map and geochemical anomaly maps of the pathfinder elements of gold in soil and groundwater. Their study shows that LR is adequate for identifying mineralized areas with the type of data used.

Raines (1999) applied the WofE method to construct a prospectivity model for epithermal-Au deposits in the Great Basin (western USA). He concluded that the WofE MPM model is reasonable for the delineation permissive areas for epithermal deposits that are comparable to expert’s delineation and that have objective, reproducible and well-defined characteristics, and provided a quantitative measure of confidence.

Singer and Kouda (1999) compared the WofE method to PNN using data from Chisel Lake–Anderson Lake, Manitoba (Canada). Despite the sparse deposits in the study area, the results demonstrate the PNN’s ability to derive unbiased probability estimates and lower error rates compared to those derived by the WofE method. The WofE method resulted in strong bias, and errors are mostly known barren areas misclassified as prospective. Although the Chi-square test for independence indicated no significant correlations among the evidential layers, the test for expected number of deposits indicated that the results of the WofE method violated the assumption of conditional independence (CI). However, the PNN has no problem dealing with the CI assumption except that its performance strongly depends on having a completely representative training set.

Venkataraman et al. (2000) applied the WofE method and a FL algorithm to integrate information interpreted from remote sensing, geochemical, geological and ground-based datasets of Rajpura–Dariba, Rajasthan (India) to target potential base metal mineralized areas. Both their WofE and fuzzy MPM models showed four classes of potential zones of sulfide mineralization; however, their fuzzy model predicted more new potential areas.

Carranza and Hale (2000) applied the WofE method to integrate binary predictor patterns of geological features for prediction of gold potential using two sets—small-scale (n = 63) and large-scale (n = 19) occurrences—of gold deposits. The derived spatial associations between the binary predictor patterns and either set of gold occurrences, which indicate the geological features that are useful for mapping of prospectivity for gold in the Baguio District (Philippines), are consistent with the known geological controls on gold mineralization in the district. The resulting MPM maps based on the two sets of mineral occurrences data are similar, indicating that small-scale occurrences, which often outnumber large-scale occurrences, are important in MPM for economic deposits of the type sought.

Chen et al. (2001) introduced the concept of geo-anomaly unit (GU) as an area with distinct features that can be delineated by combining maps of ore-forming factors using computer techniques. The factor maps are binary (i.e., square cells coded with either 0 or 1 for absence or presence, respectively, of ore-forming features) and integrated by PCA to derive a score map of linear combination of ore-forming features. Cells with high scores are suggested as exploration targets.

Harris et al. (2001) used two data-driven methods (WofE, LR) and two knowledge-driven methods (weighted index and Boolean overlay) to produce gold prospectivity maps of the Swayze greenstone belt, Ontario (Canada), using geological, geochemical, geophysical and remotely sensed (Landsat) datasets assembled in a GIS. The derived gold prospectivity maps are more or less different as the evidential layers were purposely generated in different ways according to each modeling method. Some of the several areas classified by all modeling methods to have high gold potential coincide with known gold prospects. However, the data-driven prospectivity maps were better predictors of the known gold prospects.

Mihalasky and Bonham-Carter (2001) used the WofE method to quantify the spatial association of lithodiversity with metallic mineral sites in Nevada. They calculated lithodiversity by counting the number of unique geological map units within square-shaped sample neighborhoods of different sizes. They found that the spatial association between mineral sites and lithodiversity increased with increasing lithodiversity and that this relationship was consistent for (1) both basin-range and range-only regions, (2) four sizes of sample neighborhoods, (3) various mineral site subsets, (4) three scales of geological maps and (5) areas not covered by large-scale maps. They interpreted high lithodiversity to likely reflect the occurrence of complex structural, stratigraphic, and intrusive relationships that are thought to control, focus, localize, or expose mineralization. They proposed that lithodiversity measurements in areas that are not well explored may help delineate regional-scale exploration targets.

Carranza and Hale (2001a) used fuzzy sets of favorable distances to geological features and favorable lithologic formations, based upon qualitative and quantitative knowledge of spatial associations between known gold occurrences and geological features in the Baguio District (Philippines) and then combined the fuzzy predictor maps using FL as the inference engine. Their results, which are comparable to their previous work using WofE modeling (Carranza and Hale 2000), demonstrate the usefulness of fuzzy modeling of mineral potential.

Scott and Dimitrakopoulos (2001), in a case study in Australia, estimated mineral potential using the US Geological Survey three-part resource assessment process (Singer 1993) and data-driven GIS-based MPM by WofE. The results of their case study recognize that quantitative resource assessment and GIS-based MPM are complementary processes that support mineral resources development.

In a district-scale MPM case study for porphyry-Cu mineralization, Carranza and Hale (2002a) applied the WofE method to analyze the spatial association between known porphyry-Cu deposits and geologic features in Benguet, Philippines. They found that the porphyry-Cu occurrences are spatially associated with contacts of porphyry plutons, margins of batholithic plutons and strike-slip fault discontinuities and that the porphyry plutons are spatially associated with margins of batholithic plutons and strike-slip fault discontinuities. Based on these significant spatial associations, they further applied the WofE method to map zones favorable for porphyry-Cu mineralization and zones favorable for porphyry pluton emplacement in Benguet Province, Philippines. Validations of the predictive models show their usefulness for delineating targets for follow-up exploration.

Paganelli et al. (2002) applied the WofE method to determine the spatial associations of known kimberlite locations with variously trending lineaments at the Buffalo Head Hills area (Canada). They then used structural lineament maps, Bouguer gravity data, magnetic characteristics of the Buffalo High and Buffalo Utikuma terranes, and the boundary between those two terranes to derive a map of favorability for kimberlite emplacement. The results show highest favorability for kimberlite emplacement along the boundary between the Buffalo High and Utikuma terranes and correspondence with NNE-trending lineaments and their intersections with NE and ENE lineaments.

Raines and Mihalasky (2002) used a combination of WofE and weighted LR to generate pluton-related deposit tract maps and compared these maps with tract maps generated by US Geological Survey experts. They found that, in general, there is a very strong spatial correlation between the data- and knowledge-driven tract maps. This study suggests that, similar to the findings of Scott and Dimitrakopoulos (2001), data-driven MPM can be the first part of the 3-part quantitative mineral resource assessment proposed by Singer (1993).

Agterberg and Cheng (2002) revisited the requirement of the WofE that updating the prior probability of mineral deposit occurrence with two or more evidential map layers is allowed only if the evidential layers exhibit CI with respect to known mineral deposits. They provided formal proof that CI of evidential layers implies that the sum of posterior probabilities weighted by unit cell area equals the number of known mineral deposits. They then proposed the “omnibus test” for CI, which is exact and simpler to use than existing tests of CI adapted from discrete multivariate statistics.

As the performance of ANNs for MPM is undermined by the paucity of deposit training patterns relative to barren training patterns, Brown et al. (2003a) proposed to overcome this problem by adding random noise to the original training patterns in order to create additional synthetic deposit training data. In the Kalgoorlie Terrane study area, the number of deposit training patterns increased from approximately 50–1000 by adding noise to the original deposit training data, which resulted in significant increase in both the classification performance of a trained multilayer perceptron (MLP) neural network, a feed-forward ANN, and the quality of the resultant prospectivity map.

Porwal et al. (2003a) described a GIS-based application of a radial basis functional link net (RBFLN), another type of ANN, to map potential for sedimentary exhalative (SEDEX) base metal deposits in an area in the Aravalli metallogenic province (western India). They trained a series of RBFLNs to determine the network architecture and estimate parameters that mapped the maximum number of validation vectors correctly to their respective targets. The trained RBFLN with the best performance with respect to the validation dataset was used for processing all feature vectors to generate a predictive map that was further reclassified into a prospectivity map showing zones with high, moderate and low prospectivity for SEDEX base metal deposits in the study area. The usefulness of RBFLN for MPM is indicated by the consistency of the spatial distribution of mapped high prospectivity zones with the conceptual models of base metal metallogeny in the study area.

Bougrain et al. (2003) used ANN to extract from a GIS database knowledge about factors relevant to the formation of precious and base metal deposits in the Andes. Results of the analysis indicate as much as 25 attributes as known or potential factors relevant to the formation of gold deposits in the Andes Cordillera. They used the trained ANN to distinguish potentially mineralized sites from non-mineralized sites. Their study demonstrates how ANN can be applied efficiently to assist mineral exploration, where general domain knowledge alone is inadequate to satisfactorily model the plausible controls on mineralization from a continent-scale database.

Brown et al. (2003b) proposed that the use of evidential layers represented as fuzzy membership variables is a useful method for integrating subjective knowledge with empirical data in an ANN approach to MPM. They used a MLP neural network to integrate up to 17 variables to derive prospectivity maps for orogenic gold deposits in the Archean Kalgoorlie Terrane of Western Australia. They used two types of fuzzy membership variables. For the first type, the spatial associations of data with known gold deposits were used to determine fuzzy membership values. The second type of fuzzy membership variables represents rheological contrast at lithologic boundaries, in which fuzzy membership values, although based on geological field data, are subjective. The methods described can be applied to various subjective data (e.g., favorability of tectonic environment, reactivation along major faults or host stratigraphy) used in regional exploration programs, but which normally would not be used as inputs in an ANN approach.

Harris et al. (2003) compared the performance of WofE, PNN, LR and DA for data-driven MPM in three case studies with contrasting scale and geologic information, and using randomly selected cells for training and validation in every case study. The deposit-scale Carlin study reveals that the performances of the various methods from lowest to highest are: PNN, DA, LR and WofE. The district-scale Alamos study shows that the performances of the various methods from lowest to highest are: DA, PNN and WofE. Unlike findings from the Alamos and Carlin studies, the regional-scale Nevada study DA, LR, PNN and WofE. Their study also demonstrated that the inferior performance of WofE is the result of the loss of information when data of the variables are discretized into binary maps to satisfy procedure requirement by WofE.

Carranza (2004) applied the WofE method to map potential for porphyry-Cu prospects in an area measuring ~920 km² with 12 known porphyry-Cu prospects. Uncertainty due to missing geochemical evidence is shown to have an influence on tests of assumption of CI among predictor maps with respect to prospects. However, for the final predictive model, the assumption of CI, which was initially rejected based on the omnibus test, was accepted based on a new omnibus test (Agterberg and Cheng 2002). Validation of the final predictive model demonstrates the plausibility of the WofE method for MPM in large areas with few mineral prospects.

Agterberg and Bonham-Carter (2005) contended that the random cell selection procedure followed by Harris et al. (2003) for training and validation of MPM necessarily results in better performance for the more flexible methods (i.e., DA, LR and PNN), but this did not necessarily indicate that these methods are better than WofE. They showed, by comparison with LR modeling, which does not need discretization of data, that the discretization of data into binary evidence in the WofE method is usually advantageous as it prevents occurrences of extremely high posterior probabilities. They concluded further that mineral occurrences must not be randomly sampled together with their surrounding environments in small cells but must be modeled as discoveries at points.

Carranza et al. (2005) demonstrated the application of data-driven evidential belief functions (EBFs), initially proposed by Carranza and Hale (2003), to map prospectivity for aquamarine-bearing pegmatites in the Lundazi District (eastern Zambia). Data-driven EBFs not only represent spatial association of target deposits with an evidential layer but also take into account spatial relationships among classes of evidences in an evidential layer. Spatial data that provide positive or negative evidence of prospectivity can be determined by data-driven EBFs. Data-driven EBFs of only positive evidence of prospectivity must be integrated for MPM. Validation of the results illustrated the usefulness of data-driven EBFs for MPM.

Skabar (2005) presented a new approach whereby MLP neural networks can be trained to yield output values that can be interpreted strictly as posterior probabilities. This approach uses all data in the generation of a model, thereby eliminating dependence on the choice of training data. The approach was followed to map prospectivity for gold in the Castlemaine region of Victoria (Australia). Comparison of the results with those of a method for estimating probability density functions showed that the MLP approach and the density estimation-based approach performed roughly equally per validation with the bootstrap “leave-one-out” method.

Chen et al. (2005) applied WofE modeling for MPM in large areas with small number of mineral prospects. They used predictor layers derived from a digital database that includes 1:200,000 scale geological, geochemical, and geophysical maps, and remote sensing images in study area. Their results show four main metal ore belts occupying 29% of their study area in China, which delineate 81% of the known porphyry-Cu occurrences.

Porwal et al. (2006a) developed a hybrid fuzzy WofE model for MPM that uses as inputs knowledge-based fuzzy membership values and generates outputs of data-based conditional probabilities. A knowledge-driven logistic membership function is used to derive fuzzy membership values, thereby allowing treatment of systemic uncertainty and generation of multi-class evidential layers. The fuzzy evidential layers are then integrated using the WofE method. The hybrid fuzzy WofE model was applied to regional-scale mapping of prospectivity for base metal mineralization in the south-central part of the Aravalli metallogenic province (western India). Validation of the results demonstrated the usefulness of the hybrid fuzzy WofE model for MPM.

De Quadros et al. (2006) compared the performances of the WofE and fuzzy methods for MPM based on a conceptual model for structurally controlled lode gold–quartz vein deposits in an area in Brazil. The found that, compared to the FL method, the WofE method delineated smaller highly favorable zones and that the WofE method produced higher biased probability within favorable zones.

Daneshfar et al. (2006) used the WofE and LR methods separately for MPM to delineate areas with potential for Zn–Pb Mississippi valley-type mineralization using evidential layers derived from Landsat TM data and regional geological data. Their two sets of results did not exhibit remarkable differences from each another, and the validation of the results demonstrated the usefulness of either the WofE or LR method for MPM.

Nykänen and Ojala (2007) validated predictive WofE and LR models of mineral prospectivity by the bootstrap “leave-one-out” method, actual field testing and follow-up analysis of previously drilled core within predicted highly prospective areas. All these methods of validation indicate that the WofE and LR mineral prospectivity models have successfully delineated both known and new areas of gold mineralization.

Behnia (2007) used the GIS-based RBFLN method for predictive mapping in Central Iran to delineate areas with potential for Proterozoic mineralization characterized mainly by several iron, apatite and uranium deposits. The results showed that a successful classification depends on the existence of spatially well-distributed deposits and non-deposits throughout the study area.

Raines et al. (2007) applied WofE modeling to compare the use of geological maps of different scales as inputs to GIS-based MPM to delineate permissive tracts for porphyry deposits in the USA, as the first step of a mineral resource assessment. The results indicate that porphyry tracts delineated using input from 1:2,500,000-scale geologic maps are similar to porphyry tracts delineated using input from either 1:10,000,000 or 1:35,000,000-scale geological maps. This finding demonstrates that conceptual context from small-scale maps is more appropriate for porphyry tract definition than from larger scale maps. The study also demonstrates the usefulness of the WofE method for analysis of strengths of spatial associations between mineral potential maps and known mineral occurrences.

Nykänen (2008) also used the RBFLN method to produce a series of prospectivity maps for the under-explored Paleoproterozoic Central Lapland Greenstone Belt, Northern Fennoscandian Shield, Finland, which is thought to be highly prospective for orogenic gold mineralization. They found that, when applied to the same evidential layers that are proxies for conceptual geological controls, the RBFLN method performed similarly as the LR method but outperforms the WofE and the FL methods. They also found that the performance of the RBFLN method improved when the training feature vectors were weighted according to the size of the known gold deposits.

For situations of MPM where mineral deposit occurrences are known, Fabbri and Chung (2008) discussed strategies for blind testing of MPM. They also described how to create a prediction rate graph, for validation of predictive models, whereby the X-axis of the graph represents proportion of a study area predicted to be prospective and the Y-axis represents proportion of “undiscovered” occurrences within the predicted prospective area. Such prediction rate graph is similar to the occurrence–area plots described by Agterberg and Bonham-Carter (2005), but the prediction rate graph makes use of occurrences for blind testing (i.e., occurrences not used for generation of a predictive prospectivity model).

Deng (2009) had shown that, when the CI assumption in the WofE method is violated, bias in posterior probability estimates has an intuitive and convenient interpretation, and then, he derived a formal expression for the bias. Then, using the correlation structure of the predictor patterns, he developed a modified WofE model to correct for the bias. For validation of the proposed modified WofE model in a case study, he proposed to use the receiver operator characteristic (ROC) curve analysis as it is often employed for binary response models (e.g., mineral occurrence is a binary variable). Results of validation using ROC curves demonstrated the usefulness of the proposed modified WofE model.

Oh and Lee (2010) performed a GIS-based ANN for mapping prospectivity for hydrothermal Au–Ag mineralization in the Taebaeksan District (Korea) using four different training datasets, derived by likelihood ratio and WofE methods, to analyze the effect of training. The results of their study showed that ANN performed best, followed by likelihood ratio and WofE.

Because the application of the WofE method to MPM often results in violation of the CI assumption, Agterberg (2011) proposed a modified WofE method whereby WofE is performed first and then WLR is applied to the weights. The results of the proposed modified WofE method are similar to those when the modified WofE method proposed by Deng (2009) is applied to correct for bias due to violation of the CI assumption.

Integration of several spatial data relevant to MPM is advantageous. However, for methods that involve estimating densities (e.g., PNNs) a high-dimensional input is disadvantageous due to the so-called curse of dimensionality. In view of this, Skabar (2011) described a substitute approach to the estimation of densities, using similarity-based learning, whereby he showed how (a) to estimate, using the concept of eigenvector graph centrality, the density of a set of deposit training examples from a graphical model of such examples and (b) to estimate from these data the likelihood of a test example without creating a new graph. Testing the proposed approach to a case study for gold deposits showed that, in terms of predictive capability, it is superior to conventional density estimation methods and is a bit better than MLP neural networks.

Mejía-Herrera et al. (2015) used curvature and proximity to certain geological features as predictor variables in LR to predict Cu potentials in the Kupferschiefer. Curvature is geometric attribute measurable in horizons of a surface structural model. Their LR models show positive correlation between curvature estimated on the surface depicting the mineralized layer and predicted probabilities of Cu potentials.

Carranza (2015) applied data-driven EBFs for MPM (cf. Carranza et al. 2005) to the same study area, measuring ~920 km² with 12 known porphyry-Cu prospects, where he applied the WofE method to map potential for porphyry-Cu prospects. The results indicate that method of data-driven EBFs is as efficient as the WofE method for MPM in areas where as few as 12 prospects are known and where evidential layers contain missing values.

Index overlay and Boolean logic are two methods traditionally used for knowledge-driven MPM. As the assignment of evidential weights based on expert opinion introduces bias that is difficult, if not impossible, to quantify, Yousefi and Carranza (2016) proposed a data-driven index overlay method for MPM, in which weights of evidential layers are estimated from data, and a data-driven Boolean logic method for MPM, in which thresholds for creating binary evidential layers are established based on data. The results of their MPM case study for porphyry-Cu deposits in an area in the Kerman Province in southeast Iran demonstrate the usefulness of the proposed methods.

Ford et al. (2016) examined the performances of the data-driven WofE and EBF methods and the knowledge-driven FL method of MPM to target for orogenic gold mineralization using incomplete data from the Carajás mineral province (Brazil). The results of their study demonstrate that the WofE method is the most effective, compared to the EBF and FL methods, for MPM with incomplete data. This paper by Ford et al. (2016) is the first to introduce in NRR the mineral systems approach to GIS-based MPM.

The random forests (RF) is a machine learning algorithm that has been shown recently as a viable method for data-driven MPM (Rodriguez-Galiano et al. 2014, 2015; Carranza and Laborte 2015a, b; Harris et al. 2015). Carranza and Laborte (2016) carried out a case study to map prospectivity for hydrothermal Au–Cu mineralization in Catanduanes Island (Philippines), where 17 prospects of hydrothermal Au–Cu deposits exist, in order to examine further the ability of RF modeling (a) for data-driven MPM in areas with few occurrences (i.e., <20) of mineral deposits and (b) for treatment of evidential layers with missing values. The results show that: (a) RF outperforms data-driven EBFs; (b) RF allows analysis of spatial associations between known prospects and each evidential layer, just as data-driven EBFs; and (c) and missing values in evidential layers can be treated in RF through RF-based imputation, whereas in data-driven EBFs, missing values are assigned maximum uncertainty. The case study demonstrates the usefulness of the RF algorithm for data-driven MPM in regions with few occurrences of the deposit type sought. This paper by Carranza and Laborte (2016) adapts a conceptual mineral systems model as framework for data-driven MPM.

In a MPM case study for gold deposits around the Huritz Group and Nueltin Suite, Nunavut (Canada), McKay and Harris (2016) compared the performance of the RF method to that of a knowledge-driven methodology involving weighted index overlay and FL methods. The results showed that the RF method outperformed knowledge-driven methodology and illustrated several advantages to the RF method, namely: (a) ability to take both categorical and/or continuous data as variable inputs, (b) ability to estimate importance of each input variable and (c) an unbiased internal estimation of mapping error, which makes cross-validation of final outputs to determine accuracy unnecessary.

Geranian et al. (2016) compared the performance of DA and support vector machine (SVM) for modeling subsurface gold mineralization by integrating surface soil geochemical anomalies and borehole data at the Sari Gunay gold deposit, NW Iran. The results of their study indicate that SVM outperforms DA in data-driven MPM.

Thus, NRR has several publications of research on novel methods for data-driven MPM as well as research on case-to-case innovation of existing methods for data-driven MPM.

NRR Articles on Knowledge-Driven MPM

McLaren (1992) described probably the first or one of the earliest applications of the weighted index overlay method for MPM. He classified mineral prospectivity of British Columbia (Canada) using three unique themes of field data (geological setting, geochemistry and mineral occurrences) and based on two factors (favorability and degree of confidence). He assessed prospectivity according to how well evidential data satisfy criteria defined from established mineral deposit models. He assessed degree of confidence depending on number of data themes determined to be favorable. Qualitative descriptions of each prospectivity class suggest the possibility for future exploration as an index of expected land use.

Reddy et al. (1992) proposed the use of GIS to develop an expert system for regional-scale MPM for volcanogenic massive sulfide (VMS) deposits in the Early Proterozoic Snow Lake greenstone belt of northwest Manitoba (Canada). Their expert system consisted of an inference network to model expert knowledge, to propagate evidence from input maps and to integrate information by FL and Bayesian updating to derive prospectivity maps depicting evaluation of hypotheses. A model’s sensitivity to changes in the parameters was evaluated by validating areas with predicted high prospectivity against the spatial distribution of known VMS deposits.

Chung and Fabbri (1993) used an artificially constructed dataset to demonstrate three representation methods—probability measures, Dempster-Shafer EBFs, and fuzzy membership functions—and their corresponding estimation procedures with analyses of the implications and of the assumptions that are required in each method to thematic mapping. They also discussed difficulties associated with the construction of probability measures, EBFs and fuzzy membership functions and proposed alternative procedures to overcome those difficulties.

An et al. (1994b) demonstrated EBFs for managing uncertainties in the integration of evidential data for MPM. The EBFs provide the capability to distinguish between negative information and lack of information, which is desirable when integrating diverse datasets with different spatial resolutions and spatial extents. Their study demonstrated that EBFs can provide a realistic quantitative model of mineral potential.

Rehder (1994) demonstrated an expert system for MPM, which constructs an inference network for assessing prospectivity for epigenetic gold deposits. He described three different approaches for constructing an inference network, discussed aspects of human reasoning and compared it with machine reasoning for assessing data as evidence of mineral prospectivity, and discussed some actual problems concerning uncertainty, contradictory evidence and prior probabilities. Included in the exploration model are technical aspects like costs and schedule.

An et al. (1994a) formulated an object-oriented modeling framework and respective reasoning processes using EBFs for a knowledge-based approach to integration of evidential layers. The mechanism developed for uncertainty propagation also worked well for MPM using real mineral exploration datasets from the Snow Lake area, northern Manitoba (Canada).

Cheng (1996) proposed fuzzy relations to derive weights for qualitative variables according to their partial order relations. They proposed two asymmetric measure indexes (incidence coefficient and probability difference) for quantifying asymmetric associations between evidential layers, from which the partial order relations can be generated. The combination of the fuzzy relations method with the asymmetric measure indexes leads to new methods for map overlay and data integration for MPM. Two types of models were demonstrated by an artificial example for MPM.

Carranza et al. (1999) demonstrated the application of a conceptual mineral exploration model and GIS to classify prospectivity for nickeliferous laterite in an area in northeastern Philippines. They modeled and integrated evidence maps in a way similar to that described by McLaren (1992) to furnish a nickeliferous laterite potential map. They compared this map with present land-use classification and policy in the area and found that the potential zones are in areas where mineral resources development is prohibited. Their study illustrated that MPM is a critical support to land-use policy-making to ensure that prospective land is considered in future mineral resource development.

Carranza and Hale (2001a) applied the theory of fuzzy sets to map prospectivity for gold in the Baguio District (Philippines). Maps of lithologic units and proximity to geological features, transformed into fuzzy evidence maps based upon empirical and expert knowledge of spatial associations of geological features with known gold occurrences in the district, were combined using FL as the inference engine. The results demonstrate the usefulness of the fuzzy set approach for MPM.

De Araújo and Macedo (2002) assigned expert-driven weights to geological, geochemical and airborne geophysical layers of evidence and integrated these for MPM in an area in the Ribeira Valley, São Paulo and Paraná States (Brazil) using weighted linear combination (WLC) and order weighted average (OWA) methods, which are variants of the weighted index overlay method. The evidence layers chosen and expert-driven weights were based on two mineralization models: (1) Perau type, sedimentary exhalative, and (2) Panelas type, vein-type carbonate hosted. The OWA method yielded the better results, with prospectivity maps depicting several classes of prospectivity occupying relatively minor areas. The WLC method yielded more coherent results but lacking details for minor areas. Both methods are inexpensive and viable for MPM in regions similar to the studied one.

Porwal et al. (2003b) proposed new fuzzy models for MPM, namely: (1) a knowledge-driven fuzzy model, which transforms input evidential values into fuzzy membership values by using a logistic function, and (2) a data-driven model, which derives fuzzy membership values of input evidential maps by using a piecewise linear function based on quantified spatial associations of evidential layers with known mineral deposits. They also described a graphical procedure for defuzzification and classification of output fuzzy prospectivity maps. The two new fuzzy models were demonstrated for mapping prospectivity for base metal deposits in an area in south-central Aravalli metallogenic province, Rajasthan (western India). Cross-comparison of the prospectivity maps derived by the proposed models indicates their strong similarity. Validation of both models again known deposits indicated their usefulness for delineating prospective zones to guide future exploration work.

Harris et al. (2008) performed MPM for four different mineral deposit types (i.e., SEDEX Zn–Pb–Ag, carbonate-hosted Zn–Pb–Ag, intrusion-related skarn and gemstones, and Carlin-type gold) within the Greater Nahanni Ecosystem. They followed an expert-driven approach to weighting of evidence layers similar to the WLC method used by De Araújo and Macedo (2002). Their validation of the individual mineral potential maps, using plots of number of mineral occurrences as a function of area, which depict a validation method that is essentially the same as the one proposed by Agterberg and Bonham-Carter (2005), showed strong goodness of fit with known occurrences per deposit type. Finally, they derived a composite mineral potential map by combining the four potential maps using a maximum operator.

Pazand et al. (2011) proposed the analytic hierarchy process (AHP) to evaluate weights of evidence layers for GIS-based knowledge-driven evaluation of potential for porphyry-Cu mineralization in an area in Iran using evidence layers derived from geological, geochemical, and geophysical, and remote sensing data. The results demonstrated acceptable outcomes for porphyry-Cu exploration in the study area.

Lusty et al. (2012) applied knowledge-driven FL modeling to map prospectivity for Caledonian-age turbidite-hosted orogenic gold mineralization in the Southern Uplands–Down–Longford Terrane, using geochemical and geophysical data from in conjunction with other spatial geoscience datasets. They emphasized that, since FL modeling depends on subjective judgment, it is important to understand well the key controls on the mineralization being sought and their relative importance, limited to the components of the model that can be mapped from available data.

Elliott et al. (2016) applied weighted index overlay and FL methods to determine new prospective sites for natural sand resources in the Central Texas Frac Sand District. The results of their study showed that the distribution of weights in the FL method clearly identifies prospective locations with less ambiguity compared to weighted index overlay method.

Asadi et al. (2016) integrated the AHP and the TOPSIS (Technique for Order of Preference by Similarity to Ideal Solution) algorithm, following Pazand and Hezarkhani (2015), to map prospectivity for porphyry-Cu in a district in central Iran with very few known porphyry-Cu occurrences. With mapped highly favorable to favorable areas covering roughly 2% of the study district and delineating most of the few known porphyry-Cu occurrences, several prospective areas with no known porphyry-Cu occurrences are now recognized as exploration targets.

Thus, NRR has quite a good number of publications of research on novel methods for knowledge-driven MPM as well as research on case-to-case innovation of existing methods for knowledge-driven MPM.

Impact of NRR Articles on Development of MPM

The above-reviewed 51 and 16 articles on, respectively, data- and knowledge-driven MPM represent only ~7 and ~2%, respectively, of all (=751) articles NRR has published from volume 1 in 1992 to volume 26 in 2017. However, some of the above-reviewed articles are among the most cited/applied in this field (Tables 4, 5). This information indicates that NRR has made and continues to make strong impact to research on GIS-based MPM.

Table 4 Top 15 most cited articles on mineral potential mapping obtained by a document search in Scopus (https://www.scopus.com) using the search terms “mineral potential” AND “GIS” (as of May 24, 2017)

Table 5 Top 15 most cited articles on mineral prospectivity mapping obtained by a document search in Scopus (https://www.scopus.com) using the search terms “mineral prospectivity” AND “GIS” (as of May 24, 2017)

Organization of the Special Issue

There are 10 articles in this special issue. The first four articles are relevant to geochemical anomaly mapping and the last six to MPM.

Articles on Geochemical Anomaly Mapping

Carranza (2017) showed that anomalies of geochemical enrichment factors are no better than anomalies of log-ratio-transformed geochemical data. This study demonstrates further that exploration of geochemical data, being compositional data, should first be “opened” by log-ratio transformation in order to address the closure problem inherent in the statistical analysis of such data (cf. Aitchison 1984; Filzmoser et al. 2009a, b, Filzmoser et al. 2010).

Yousefi (2017) proposed and demonstrated a method of pixel-based modeling of geochemical landscape in order to avoid estimation errors resulting from interpolation of point lithogeochemical data. In this method, the square pixel covering each composite rock sample is considered its area of influence. By using this method together with the concept of multiplicative geochemical halos and the concentration–area fractal modeling, anomalies of multiplicative geochemical halos were delineated and then integrated to model metal zoning for vectoring into porphyry-Cu mineralization. Validation of the results against stockwork distribution and locations of Cu occurrences demonstrates that the proposed pixel-based method and the metal zoning concept is a powerful tool for targeting areas with potential for porphyry-Cu deposits.

Parsa et al. (2017) subjected log-ratio-transformed stream sediment geochemical data to robust principal components analysis to derive a multivariate geochemical signature of porphyry-Cu deposits in their area, and then, they subjected positively shifted values of multivariate mineralization-related geochemical signature to singularity mapping, as a filtering method, to enhance subtle but significant geochemical anomalies. The superiority of anomalies extracted from the filtered multivariate geochemical signature over anomalies extracted from the non-filtered geochemical signature was indicated by the higher success rate former compared to that of the latter.

Zuo (2017) reviewed the state of the art of applications of machine learning in identifying geochemical anomalies.

Articles on Mineral Potential Mapping

Hariharan et al. (2017) applied the synthetic minority over-sampling technique to modify the initial dataset and bring the deposit-to-non-deposit ratio closer to 50:50 in their RF-based gold prospectivity modeling of the Tanami Region, a greenfields terrain in Western Australia. They then objectively determined an optimal threshold prospectivity value by using statistical measures such as data sensitivity, specificity, kappa and percent correct classification. Their RF regression modeling with the modified dataset of close to 50:50 deposit-to non-deposit ratio delineated ~5% of the region as high prospectivity areas as compared to only ~1% by the original dataset, implying that the original “sparse” dataset led to underestimation of prospectivity.

Tessema (2017) discussed the chromite mineral system in the Bushveld Complex (BC) in South Africa as the framework for data-driven predictive mapping of prospectivity, using RBFLN, for chromite deposits in the Western Limb and the Nietverdiend layered mafic intrusion of the BC. The RBFLN model correctly classified 73% of the validation deposits into highly prospective areas, which cover 6.5% of the study area, and the RBFLN correctly classified all the non-deposit validation points into low prospectivity areas, which occupy 86.6% of the study area. Results of cross-validating the RBFLN model with a fuzzy WofE model show that the two models agree well on broad-scale targeting of high prospectivity areas for further exploration of chromite deposits.

Mutele et al. (2017) presented prospectivity mapping by FL modeling based on a mineral system model of critical processes responsible for the formation of polymetallic Sn–F–REE mineralization associated with the Bushveld granites of the Bushveld Igneous Complex, South Africa. They used spatial proxies of proximity to differentiated granites (representing heat and metal-rich fluid sources), Rb geochemical map (representing fluid-focusing mechanism), principal component maps (PC4 Y–Th and PC14 Sn–W, representing fluid pathways for high- and low-temperature mineralization, respectively) and proximity to roof rocks (representing traps). This study yielded encouraging results with delineation of 13 new exploration targets.

Duarte Campos et al. (2017) adopted the mineral systems approach to MPM in the Gurupi Orogenic Gold Belt, north–northeast Brazil. They used the knowledge-driven multi-class overlay method to weight and integrate spatial predictors of prospectivity derived from up-to-date regional-scale geological, geochemical and geophysical datasets compiled by the Geological Survey of Brazil. This study resulted in considerable reduction in the search area and delineation of new exploration targets.

Feizi et al. (2017) presented a novel hybrid AHP-Shannon Entropy approach for assigning weights to evidence for MPM in a case study for porphyry-Cu potential mapping in Iran. They evaluated their output mineral potential map by field checking and chemical analysis of rock samples. They found outcrops with evidence of a porphyry system in areas with high potential values, and they found good correlation between high potential values and Cu content of rock samples taken from the field.

Nykänen et al. (2017) combined a mineral deposit model and a mineral systems model into a conceptual model for MPM, which they tested in an active mineral exploration terrain within the Paleoproterozoic Peräpohja Belt (PB) in the Northern Fennoscandian Shield, Finland, where recent mineral exploration activities have indicated several gold-bearing mineral occurrences. They used the ROC spatial statistical technique to optimize rescaling of input datasets and integration of data using FL. Spatial coincidence of high prospectivity values with current exploration licenses and exploration drilling sites for gold indicates the validity of their conceptual mineral prospectivity model.

Concluding Remarks

The Natural Resources Research journal leads the promotion, through publication, of developments in research of methods for mineral potential mapping (Fig. 7). Since its foundation in 1992, the journal has contributed significantly to the publication of research on development of robust numerical methods for the analysis and integration of spatial datasets relevant to mineral potential mapping. It was only in recent years that the journal has started to publish research on development of geologically focused quantitative models of mineral potential. The reason for the latter is not that research on development of geologically focused models of mineral potential does not fall within the scope of the journal but because academics involved in this research field prefers to publish in peer-reviewed journals that are abstracted/indexed in the Science Citation Index Expanded™ (or SciSearch^®) and Journal Citation Reports^® (JCR) Science Edition. However, Natural Resources Research has just recently been selected for coverage in SciSearch and JCR as well as in Current Contents^®/Physical Chemical and Earth Sciences (see editorial of the author at http://springerlink.bibliotecabuap.elogim.com/article/10.1007/s11053-017-9328-5). This recognition from the Clarivate Analytics (formerly the Institute for Scientific Information) will make Natural Resources Research more attractive to researchers in the fields of geochemical anomaly and mineral potential mapping, and so we can foresee a raise in the number of submitted papers in these fields in the years to come.