Mineral potential mapping in frontier regions: A Mongolian case study

Ore Geology Reviews 51 (2013) 15–26

Contents lists available at SciVerse ScienceDirect

Ore Geology Reviews journal homepage: www.elsevier.com/locate/oregeorev

Mineral potential mapping in frontier regions: A Mongolian case study Arianne Ford a,⁎, Craig J.R. Hart b a b

Centre for Exploration Targeting, University of Western Australia, Crawley WA 6009, Australia Mineral Deposit Research Unit, University of British Columbia, Vancouver, Canada BC V6T 1Z4

a r t i c l e

i n f o

Article history: Received 16 August 2012 Received in revised form 13 November 2012 Accepted 15 November 2012 Available online 23 November 2012 Keywords: Frontier regions Fuzzy logic Mineral potential mapping Mongolia Orogenic gold Weights of evidence

a b s t r a c t The quality of a mineral potential map is dependent on the quality of the input data used in the analysis. In frontier regions or those with limited or no exploration history, datasets are often of questionable quality, and are generally incomplete with data missing either due to incomplete mapping or data not being made available to the public. This study introduces a method for addressing these challenges in mineral potential mapping to derive exploration targets. Utilizing four established statistical measures, an iterative weights of evidence method is employed to assess the strength of the relationship between known deposits and a set of geological feature layers. This method acts as an indirect validation tool for assessing the quality of the data by allowing an expert user to determine whether the statistics conform to expected relationships. Taking data from Mongolia, this iterative weights of evidence method is used to produce a mineral potential map and to evaluate potential targets for orogenic gold mineralization. The success of the method is determined by the ability of the mineral potential map to predict the location of the known mineralization. © 2012 Elsevier B.V. All rights reserved.

1. Introduction Mineral potential mapping is used as a tool to delineate areas with a high potential to host mineral deposits. Utilizing a Geographic Information System (GIS) allows an expert user to rapidly evaluate spatial geoscience data for use in mineral potential mapping projects to identify exploration targeting opportunities. Quantitative methods for evaluating geoscientific data for use in mineral potential mapping include weights of evidence (Agterberg et al., 1993; Bonham-Carter, 1994; Bonham-Carter et al., 1989; Porwal et al., 2006), fuzzy logic (Bonham-Carter, 1994; Brown et al., 2003; Knox-Robinson, 2000), evidential belief (Carranza and Hale, 2002) and neural networks (Brown et al., 2000; Fung et al., 2005; Singer and Kouda, 1999). Previous studies have typically applied such techniques to wellstudied, well-explored, data-rich terranes (e.g. Carranza et al., 2005; Feltrin, 2009; Harris et al., 2008; Knox-Robinson, 2000; Mustard et al., 2004; Nykänen et al., 2008; Porwal et al., 2001; Raines, 1999). As a result of this critical dependence on data quality, brownfields-scale (near mine) approaches to exploration targeting have traditionally been the way in which mineral potential mapping has been utilized. However, because of the extensive exploration histories that are typical of brownfields districts, it is likely that the “big one” in these districts has already been discovered. Yet greenfields terranes have a greater potential for discovery of new mineral districts or world class deposits. By definition, greenfields districts lack extensive and intensive exploration histories, and as such may benefit greatly ⁎ Corresponding author. Tel.: +61 6488 5805; fax: +61 8 6488 1178. E-mail address: [email protected] (A. Ford). 0169-1368/$ – see front matter © 2012 Elsevier B.V. All rights reserved. http://dx.doi.org/10.1016/j.oregeorev.2012.11.002

from regional-scale mineral potential mapping exercises that may highlight regions of higher mineralization potential. Greenfields terranes may be in frontier regions of the world that have developing economies or have enduring geopolitical, economic and security issues that can hinder exploration (cf. Kreuzer et al., 2008; Penney et al., 2004; Singer and Kouda, 1999). These issues increase a nation's country risk, which can include risk factors such as its ability to cover national debt, its regulatory framework and restrictions on financial transactions (Trench and Packey, 2012). Consequently, previous exploration is often limited. As exploration facilitates an increasing knowledge-base and economic geology research, relevant academic literature in these areas is lacking which leads to poor data availability and quality. This paper shows how mineral potential mapping can be successfully applied to frontier terranes to successfully delineate valid exploration targets despite the data quality challenges posed by working in a frontier region. Mongolia was selected as a case study area for demonstrating the usefulness of model-based mineral potential mapping in a frontier country where reliable geoscientific data are not publically available. A modified weights of evidence model and a fuzzy logic model are applied to mineral potential mapping for orogenic gold in Mongolia. The results of the mineral potential mapping are verified by their ability to predict known mineralization within the study area. 2. Data challenges in frontier regions Successful targeting for potential mineralization in frontier regions using mineral potential mapping in a GIS poses many challenges in terms of the availability and quality of data. The adage of “garbage in,

16

A. Ford, C.J.R. Hart / Ore Geology Reviews 51 (2013) 15–26

garbage out” holds true for any mineral potential mapping project using digital scientific data. The quality of the final mineral potential map is most strongly influenced by the quality of the data used as inputs. Dealing with data availability and quality issues is by no means limited to frontier regions, however, these challenges are often compounded by having multiple issues within the same dataset, which are described below. 2.1. Data availability Frontier regions tend to have a lack of publically available data. Data may either (a) simply be non-existent for the complete (or parts of the) study area, or (b) exist in archives which are publically inaccessible, or (c) not be available in digital format. Dealing with non-existent data is largely a problem that cannot be overcome when dealing with a large study area. Such large scale data can usually only be acquired by national/state/provincial government surveys who have a long term view in trying to encourage investment in their part of the world and the monetary resources to acquire the data. However in some parts of the world, such government survey data are not always publicly available and may be considered secret in the interests of national security, etc. On a smaller scale, the data can often be proprietary and held by individual exploration and mining companies who will not provide the data to potential competitors. Many frontier parts of the world have undertaken some form of geological mapping in the past (e.g. West Africa and Central Asia). However, this type of data is not always available in digital format and can often come in the form of hand drawn maps or in image formats that lack spatial context such as projections or datums. This lack of digital data can sometimes be overcome by digitizing the hard copy data in a GIS. Problems can then arise due to the quality of digitizing and attributing geological maps. 2.2. Data quality The quality of the data can often be poor or questionable. This can be due to inaccurate data or having sporadic (heavily biased) data. Inaccurate data can be a result of digitizing of hard copy maps due to the poor quality of the original hard copy maps, lack of knowledge about the coordinate system used, distortion of images in image processing software as well as the general errors inherent in the digitizing process by humans. Georegistration of image files, while a simple task in a GIS, can lead to major discrepancies between the registered image and the actual location of the data if the original coordinate system is unknown or incorrect, or if the image itself has been distorted in some way through processing. Inaccurate data can be verified by comparing different geoscientific datasets against each other for logical inconsistencies. At some scales, this may require revision of geological maps through additional field mapping. However, at country-scales, this is infeasible due to the cost and time taken to undertake the geological mapping exercise. Data which are heavily biased towards certain regions can still be used in mineral potential mapping, however one must be careful how such data are used. Inaccuracy can also result as a consequence of incorrect coordinate systems being used to spatially represent the data. Trial and error can be used to test various alternative coordinate systems until the misrepresented data can be shown to fit an accurate dataset. However, this is not always successful if the data come in a userderived or local coordinate system that is not available or known under the standard systems within a GIS package. Incomplete data coverage can be accommodated by utilizing specific methods for mineral potential mapping. These methods are discussed below. However, this raises the question of whether the data coverage is truly incomplete or whether the geological features in question are simply not present in a given region. This can lead to subjective inputs

going into the mineral potential mapping process, which in turn raises an issue of data quality. 3. Methods for mineral potential mapping in frontier regions Mineral potential mapping is dominated by three main methods: data driven approaches such as weights of evidence and neural networks, and knowledge driven approaches such as fuzzy logic. Each of these methods have been shown to have strengths and weaknesses for mineral potential mapping (e.g. Agterberg and Bonham-Carter, 2005; Brown et al., 2003; Harris et al., 2003; Singer and Kouda, 1999). Given the potential challenges with data in frontier regions discussed previously, we chose to focus on using the weights of evidence and fuzzy logic methods. Neural networks offer a well established method for analyzing geoscience data (Bougrain et al., 2003; Feltrin, 2009; Singer and Kouda, 1999). However, this method was not considered in this study as it is a more “black box” method that does not allow for interpretation of the relative strengths of each evidential layer in terms of its data quality and relationship to known mineralization. The process for generating mineral potential maps using data from frontier regions is discussed below for the weights of evidence and fuzzy logic methods. 3.1. Weights of evidence Weights of evidence is a data driven method for mineral potential mapping. Subjective input is required from a geologist to develop a deposit or mineral systems model, and the data are then analyzed using statistical methods. This data driven method requires a mineral deposit dataset and a series of geological features in order to generate a mineral potential map. In a frontier region, the quality of both the mineral deposit data and evidential layers may be questionable. Selection of the mineral deposit data to be analyzed should attempt to filter the data for a specific style of mineralization, which will be analyzed using various statistical measures. However, with the lack of available research on existing deposits in frontier regions, the training data for a deposit style must be limited to those deposits or prospects for which data can be sourced. This may include peer reviewed scientific publications, government reports or company data such as NI43-101 reports from Canada. One of the benefits of using weights of evidence is that incomplete data can be used, as it is possible to specify a separate class for areas of missing data. Such data can be common in frontier regions, as structural mapping can be incomplete in remote mountainous regions where no geophysical data are available to complete the mapping. The strength of the spatial association between known mineral deposits and a geological feature can be measured by a contrast value in weights of evidence (Bonham-Carter, 1994). The contrast C can be calculated from  
C ¼ ln O B A − ln O B A where O(B| A) represents the odds of B (e.g. a mineral deposit) occurring given the presence of A (e.g., a specified geological feature) and O(B | Ā) represents the odds of B occurring given the absence of A. Further, statistics are calculated in order to determine which evidential layers are most appropriate for use in the final mineral potential map: • Confidence (C / σ) — studentized contrast values from the weights of evidence statistics where C is the previously defined contrast value and σ is its standard deviation (Bonham-Carter, 1994). • Deposit–Area statistic (d(d / a)) — measures the capture efficiency, where d is the percentage of the total number of deposits within a specified distance from a feature and a is the percentage of the total study area covered by that distance (Brown et al., 2003).

A. Ford, C.J.R. Hart / Ore Geology Reviews 51 (2013) 15–26

• Observed–Expected ratio ((O–E) / E) — the ratio between observed and expected numbers of deposits in a specified buffer distance. O is the observed number of deposits within the specified buffer distance and E is the number of deposits that would be expected within that buffer distance if the total number of deposits over the entire study area had been uniformly distributed (Brown et al., 2003). For vector data such as faults, initial buffer distances were placed around the geological features being analyzed for each evidential layer. The maximum buffer distances were chosen subjectively based on expert opinion on the maximum distances over which the features may be effective for ore genesis. Due to uncertainty in the quality of the data, not limiting the maximum buffer distance could lead to situations where the statistics are most promising at a distance from the feature that is meaningless (e.g., highest capture efficiency occurring at over 100 km from a granite when using intrusion related gold deposits as the mineral deposit dataset being analyzed). Buffer distances within the maximum buffer size can then be distributed, based for example on the cell size used for the study area. For raster data (e.g., Bouguer gravity), different classes can be assigned based for example on a normal or Poisson distribution. Each of the four statistics above can then be determined from the GIS software or from simple calculations performed in a spreadsheet for each buffer distance to (vector data) or for each class (raster data). The GIS software can calculate how many training data points fall within each buffer distance or class, and provide the contrast and confidence values. These data can then be exported to a spreadsheet and the deposit–area statistics and the observed–expected ratios can be calculated. Each of the four statistical variables can then be examined for each evidential layer and evaluated for its ability to detect the training data points. Any evidential layers which have a poor capture efficiency (i.e. ~b60% of total training data) within the maximum buffer distance could be eliminated. Buffer distances within the maximum buffer could then be redistributed to determine where critical distances of capture efficiency occur, and drops off. For raster data, classes with limited capture efficiency could be eliminated, or if all classes had poor capture efficiency, then the entire dataset could be eliminated. Evaluation of each of the four statistical variables allows an indirect verification of the quality of the data. For example, do the data indicate that more training data points can be seen closer to the feature? Or do the contrast values follow a logical pattern — is the relationship between the training data and the geological feature stronger closer to the feature? Do the evidential layers that should be critical for the genesis of the deposit style show the expected relationships? Any problems in these areas could result from inaccurate training data or inaccurate or incomplete data in the evidential layer. Since expert knowledge has already been used in the refining and selection of the training data in order to be confident in the deposit locations and homogeneity, such problems are more likely a result of an issue with the data in the evidential layer. Expert geological knowledge may be useful in correcting or refining the data for a given layer so that it may be reanalyzed. However, it is often the case that such geological knowledge does not exist for a frontier region and as such the evidential layer must be eliminated. Using re-buffered and reclassified evidential layers, the calculation of the four statistics is repeated. Evidential layers with weaker capture efficiency and lower observed–expected ratios are eliminated. Baseline measures for capture efficiency and observed–expected ratios were established at values of 100 and 1 respectively, values which are the equivalent of random chance. Buffer distances for the vector data are again redistributed to find the optimal buffer distance to best capture the training data. Classes for the raster data could again be eliminated based on their ability to capture the known mineral deposit data points. These refined evidential layers could then be reevaluated using the four statistical variables and through this iterative process of refinement and elimination, the ultimate result is the selection of

17

the evidential layers which best capture the known mineral deposits in the training dataset. The final evidential layers containing buffers need to be converted to binary maps with the threshold occurring at the critical buffer distance obtained from the statistical analysis. Weights of evidence methods then combine these selected evidential layers to generate a mineral potential map through calculation of posterior probability values. The posterior probability can be evaluated from the posterior logit: PostLogit ¼ PriorLogit þ w1 þ w2 þ w3 þ … þ wn and PriorLogit ¼ log ðPriorProb=ð1−PriorProbÞ where PriorProb is the probability of a deposit occurring at any given location within the study area ignoring the presence of any evidential layer. For example, if the study area was 100 km2 in 1 km× 1 km grid cells and there were five training data points each in a different 1 km× 1 km cell, then the prior probability is 5/100 = 0.05. The values of w1, w2, … wn represent either the positive weight w+ (ln O(B| A) − ln O(B)) or the negative weight w− (ln O(B |Ā)− ln O(B)) for each binary evidential layer depending on whether the geological feature (buffered or in a raster class) is present or not, respectively. The posterior logit must then be converted to posterior odds using the exponential function PostOdds ¼ expðPostLogit Þ and the posterior probability calculated as PostProb ¼ PostOdds=ð1 þ PostOddsÞ A unique conditions grid is generated from the unique combinations of the weights assigned to each evidential layer. The posterior probability is calculated for each unique condition. Higher posterior probability values represent locations where more of the evidential layers overlap and are therefore the areas with a higher potential for mineralization. In order to validate the mineral potential map, different combinations of the final evidential layers can be used as inputs (cf. feature selection in neural networks supervised classification). This can assist in determining which combinations of input layers allow the mineral potential map to best identify existing mineralization and whether the removal of certain input layers can improve this identification. One of the common criticisms of the weights of evidence method for mineral potential mapping is the problem of conditional independence (cf. Agterberg and Cheng, 2002; Zhang et al., 2008). Advances have been made in overcoming this problem and algorithms for evaluating the extent of the conditional independence have been developed and incorporated into commonly used mineral potential mapping software packages. 3.2. Fuzzy logic An alternative to using a statistical approach such as weights of evidence to generate a mineral potential map, fuzzy logic approaches that utilize the expert knowledge of the geologist to build models can be used to generate the mineral potential maps (e.g. Bonham-Carter, 1994; Feltrin, 2009; Nykänen et al., 2008). Whereas weights of evidence statistics select the evidential layers that show the best capture efficiency of known mineral deposits and uses statistically derived weights for combining these layers to generate the mineral potential maps, fuzzy logic allows the geologist to select the evidential layers they believe are the most critical for the particular style of mineralization being investigated, and additionally allows weights to be assigned to each of these layers based on their expert opinion. Mineral potential mapping using this approach is suited to new areas where data availability may be poor

18

A. Ford, C.J.R. Hart / Ore Geology Reviews 51 (2013) 15–26

and there may be a lack of training sites, so expert knowledge is critical (cf. Brown et al., 2003). Fuzzy logic refers to the “weights” used in weights of evidence as fuzzy membership values (Bonham-Carter, 1994). Fuzzy membership values are dependent on how important the geologist believes a particular geological feature contributes to the mineralization process. The geologist can assign a value between 0 and 1 for each evidential layer they believe is critical for mineralization. Assigning a value of 0 means that the feature must not be present for mineralization to occur and assigning a value of 1 means the feature must be present for mineralization to occur. As the fuzzy membership values are multiplicative when generating the final mineral potential map, assigning a value of 0 will rule out a given area completely as multiplying any value by 0 gives a final value of 0. Due to uncertainties in the data quality in a frontier region, a small value of 0.001 can be assigned if the geologist believes that a given geological feature should not be present. This can account for incomplete, inconsistent, inaccurate or uncertain data for a given evidential layer which is a common feature in frontier regions. There are several ways to combine the evidential layers to generate a mineral potential map using fuzzy logic. The most common methods for mineral potential mapping are the fuzzy algebraic product, fuzzy algebraic sum and the gamma operation (cf. Bonham-Carter, 1994). These three methods allow for each of the fuzzy membership values for each evidential layer to have an effect on the result, which is unlike some of the alternative methods. The total combined fuzzy membership values are evaluated and each cell on the mineral potential map is assigned this total value. Areas with the highest totals represent areas where more of the evidential layers overlap and thus indicate a higher potential for mineralization. These mineral potential maps can be validated by using the same method previously discussed for verification of the maps generated using the weights of evidence method. 4. Application to a country-scale frontier region — Mongolia 4.1. Geology and orogenic gold in Mongolia Mongolia is composed of numerous crustal fragments of ancient continental margin, and Paleozoic island arcs, oceanic lithosphere and marine sedimentary origins that were variably amalgamated between the Siberian and North China cratons during Permian, Triassic and Jurassic times (Badarch et al., 2002). These crustal fragments are mostly recognized as terranes, which is appropriate since each has its own unique geological history and they are in fault contact with adjacent terranes. The terranes have been subsequently intruded by various suites of plutons, overlain by the products of various volcanic episodes, and overlain by successor basins of sedimentary rocks. Within many of these terranes, gold deposits and occurrences have been identified, and “ninja” (artisanal) mining is widespread. Historically, Mongolia has had a significant amount of placer gold production, particularly in the Zaamar, Bayanghol and Khentii regions, but until recently, has had only a few small hard rock gold mines. The Boroo gold deposit is an exception and is by far, the largest in Mongolia (Cluer et al., 2005) and has produced about 1.4 Moz since it started mining in 2004, and was the only operating gold mine in 2012. The Gatsuurt deposit is the largest unmined gold deposit with a reserve of about 1.6 Moz (Hendry et al., 2006). Both of these deposits are in the North Khentii area and are considered to be orogenic styles of mineralization. The only other significant gold mine is the Olon Ovoot deposit which operated intermittently in the past 10 years and is also an orogenic gold deposit. In addition, Mongolia's extensive placer goldfields are considered to be derived from orogenic gold deposits such as the Bumbat deposit in the Zaamar district. As a result, Mongolia is considered to be highly prospective for orogenic gold deposits and was the focus of considerable exploration effort which initiated this study.

4.2. Orogenic gold deposit model An orogenic gold model was constructed that was appropriate for this study based on the integration of several considerations; 1) accumulated expert knowledge obtained from field investigations on this deposit type in a range of geological settings in a global context; 2) compilation of features associated with this model from the scientific literature (i.e. Goldfarb et al., 2005; Groves et al., 2003); 3) accumulated knowledge specifically related to Mongolian gold deposits through expert knowledge accumulated during numerous field investigations on Mongolian gold deposits and from published scientific reports and unpublished exploration reports. The model had a bias towards a metamorphogenic fluid source, but fluid sources from intrusions were also incorporated. A total of 99 geological phenomena were considered as being potentially important contributors for the formation of an orogenic gold deposit. Because this experiment was at a country scale, and most data layers were at 1:1 M, many geological features that were considered were those that could be extracted from country scale data layers. For example, because the orogenic gold deposit model favors large crustal scale faults, there was a significant emphasis on defining the faulted margins of the tectonic elements (terranes) that comprise Mongolia. In order to do this, the gravity and magnetic data were utilized to define the extent of the crustal blocks, establish the locations of their edges, and features within them by evaluating the highs, lows, linear features, and gradients within the data. Most particularly, these data were used to define crustal scale features, terrane boundaries and to recognize many unmapped faults. Only a limited number of features of the mineral deposit model were specific to Mongolia due to the limited number of significant deposits, the lack of appropriate deposit descriptions in the scientific literature, and the overall bias of deposit characteristics of the North Khentii region which hosts the majority of known and significant orogenic gold occurrences. Reasonably good information was available for the Boroo (Cluer et al., 2005) and Gatsuurt (Hendry et al., 2006) deposits in the North Khentii region, as well as for a few other deposits. Other descriptions of Mongolian orogenic gold deposits may be available in industry reports or various scientific communications in foreign languages but there is little confidence about the deposit type of many of these deposits. Our decision-making of the mineral deposit model also fortunately benefited from “boots on the ground” expert knowledge on the Mongolian deposits. Each of the 99 features that were considered to characterize the orogenic gold model was divided into the following categories: gold source, non-source, tectonic, terrain-scale, camp-scale, rocks, metamorphic, structure, evidence of fluid flow and evidence of metal deposition. These categories were then broadly partitioned into the source, pathway, focus and trap mineral systems approach that has been suggested by Wyborn et al. (1994), Knox-Robinson and Wyborn (1997), Kreuzer et al. (2008) and McCuaig et al. (2010) to characterize the critical features and processes of ore formation (Table 1). In order to utilize any deposit model feature in a GIS prospectivity analysis, the feature, or a proxy for that feature, must be able to be extracted from a data layer in some way. Each of the layers identified in Table 1 provided between one and several derivative layers that provide a proxy for the geological features in the model. 4.3. Data A series of 10 primary datasets were sourced for use in this study. Each primary dataset is detailed below along with the data quality and availability issues and how these challenges were overcome. 1. Mineral deposit database — Database containing mineral occurrence/ deposit data for Mongolia. Deposits are characterized by location,

A. Ford, C.J.R. Hart / Ore Geology Reviews 51 (2013) 15–26 Table 1 Generic targeting criteria for orogenic gold mineralization in Mongolia separated into the source, pathway, focus and trap components of a mineral system. Source predictor layers Heat source: felsic intrusions

Considered to be potential heat source to drive hydrothermal system, but also a possible fluid source. Fluid source and sulfur source rocks: Considered to have high water contents that could generate fluids during shales, turbidites, and marine metamorphic devolatilization reactions clastics Metamorphic grade: greenschist Empirically orogenic deposits occur most frequently in greenschist grade metamorphic rocks and may represent fluid and sulfur-rich source rocks that are undergoing metamorphosis Pathway predictor layers Gravity gradients, ridges, troughs, highs, lows, and anomalies Major faults (at various scales) Density of fault intersections

Focus predictor layers Geological contacts: granite– sediment contacts

Fault intersections (various scales) Fault bends (various scales)

Trap predictor layers Brittle units: felsic intrusions, mafic intrusions, and quartzite Chemically reactive host rocks: ultramafic, basalts, and mafic intrusions Synclines

Features that represent crustal scale structures, most particularly terrane boundaries that can focus fluids Larger faults are recognized as being preferentially beneficial to fluid focusing Upper crustal regions with high density of faults and intersections can provide porosity pathways for fluids

Lithological contacts, particularly those juxtaposing rocks of differing rheologies are potential loci of permeability to focus fluids Fault intersections are potential loci of permeability to focus fluids Deviations in fault azimuths can provide permeability due to either extensional or compressional deformation, which can provide sites to focus fluids

Brittle rock units are more likely to have high fracture density that can provide more favorable traps for gold deposition Chemically reactive rocks, largely those with high iron contents, may facilitate gold deposition through sulfidation reactions Fluid focusing in syncline hinges can result in fluid mixing and ore deposition

commodity, deposit type, size and resources (when applicable). Initially, all entries listing gold as a commodity were extracted. It was found that 47 different deposit types were specified for gold. These deposit types were reclassified into seven deposit types (orogenic, epithermal, porphyry, skarn, intrusion related, VMS and “other”) based on existing literature and expert knowledge. Further reclassification was necessary due to misclassification of “deposits” vs. “occurrences” as the world class Oyu Tolgoi Cu–Au deposit was initially classified as an “occurrence” and several of the “deposits” were simply resource estimates that had never been documented in scientific literature or company reports. This refined data was used to generate training datasets and validation datasets for each deposit style. Location information for the deposits in the training and validation datasets was verified for accuracy using a combination of expert knowledge, company reports and scientific literature. Based on the revised mineral deposit database, 100 orogenic gold deposits and occurrences were extracted for use in the analysis as training and validation data (Fig. 1). 2. Geology map — 1:1,000,000 scale digital geology map of Mongolia. This map contains information on the lithology and age for each polygon. However, the scale was deemed too coarse, and attempts were made to incorporate more detail. Rock types for each polygon were attributed from georeferenced 1:200,000 scale geology maps. This however resulted in multiple

19

rock types being attributed to each polygon. Investigation into the ages assigned to the lithology polygons indicated that many of the age attributes, particularly for the felsic intrusions, were incorrect. Without the support of a large-scale geochronology program, this is an issue that cannot be readily overcome. The original map also included units of surficial cover; these units were removed by querying out and removing all polygons attributed as Cretaceous or younger. Overlaying the major gold deposits (with verified locations) onto the 1:1,000,000 scale geology map of Mongolia indicates that many deposits are not properly located with respect to the geology. Further comparison of the geology map against Landsat images shows that the polygons on the digital geology map can be incorrectly positioned, locally and inexplicably near border regions, by up to 8 km. Testing of different coordinate systems revealed that the inaccuracies are not systematic and a result of incorrect registration, but that the discrepancies are different at different locations. It is impossible to determine the source of these inaccuracies, but is likely a mix of problems that range from poor documentation during initial field mapping to digitization errors. By increasing buffers around each of the critical rock types, the key issues with the inaccuracies (such as exclusion) can be reduced, but the spatial uncertainties are propagated through the process. Layers of information that were derived from the 1:1,000,000 scale geology map include: brittle rock units, chemically reactive rock units, source rock units, granite–sediment contacts, mafics, felsic intrusions, volcanics and reactivity and rheology contrasts. 3. Structural map — A 1:1,000,000 scale fault map was generated. As these faults were digitized from the 1:1,000,000 scale digital geology map of Mongolia, identical location uncertainties exist regarding the location of the deposits relative to the structures. This issue was overcome by increasing the buffer distances around the faults to account for the potential offset from their actual locations, although in most places it was likely less than 1 km. Additionally, the faults were digitized as a series of multiple line segments which made it impossible to accurately detect long faults, or bends, jogs and intersections in the faults. This issue was overcome by combining the relevant fault segments to form contiguous faults through visual inspection with the geology, and with the benefit of geophysical data. Attributes such as fault length and strike were added as attributes to the dataset for later analysis. A 1:200,000 scale structural map was generated by digitizing the faults from georeferenced scanned geology maps of Mongolia, however these higher resolution maps do not cover the entire country and as such data coverage is incomplete. To account for this incomplete coverage at country-scale, the 1:200,000 scale faults were only used in more focused regional-scale targeting exercises where coverage was complete over the requisite study areas. At the 1:200,000 scale, the faults do not show the same inaccuracies with respect to the location of the gold deposits as seen in the 1:1,000,000 scale mapping. Derivative layers from the fault maps at both scales include: different fault orientation datasets, different fault size datasets, fault bends/ jogs of different orientations, fault intersections between faults of varying orientations, density of fault intersections and magnetite destruction zones. 4. Aeromagnetic maps — Maps derived from aeromagnetic surveys included: magnetic susceptibility (200 m grid), reduced to pole (RTP; 100 m grid), RTP–analytic signal (200 m grid) and tilt derivative (100 m grid). There were issues with the registration of each of these maps, which, while minimized by correcting the coordinate system, could not be completely addressed. The available data covered approximately 60% of Mongolia. Large parts of the country have no magnetic survey data available, mainly in the mountainous regions of the Altai to the west.

20

A. Ford, C.J.R. Hart / Ore Geology Reviews 51 (2013) 15–26

Fig. 1. Location of orogenic gold training and validation data used in the iterative weights of evidence analysis.

Further issues were found with the leveling of the data in the magnetic maps between the various surveys. Attempts were made to reprocess the digital data, however as the original raw data were unavailable, this challenge proved to be insurmountable. The combination of these issues lead to the exclusion of the magnetic maps as a direct input into the country-scale mineral potential mapping. Magnetic highs and lows were extracted for areas with full magnetic survey coverage and could be used in regional-scale targeting. The magnetic data were used considerably to assist in improving other datasets, such as structure and terrane boundaries. 5. Gravity maps — Maps included: Bouger gravity (8 km grid), 1st and 2nd vertical derivatives (8 km grid), analytic signal (8 km grid), tilt derivative (8 km/200 m grid) and horizontal derivative (200 m grid). The 8 km Bouger gravity map was of good quality as a very high resolution georeferenced image, however the other maps had residual issues with registration as previously discussed. The lack of raw data precluded extracting derivative data layers. Further data layers were manually derived from the Bouger gravity map which included: blocks of gravity highs, blocks of gravity lows, gravity gradients, gravity ridges and gravity troughs. 6. Tectonic map — A digitized version of the Badarch et al. (2002) tectonic map was used, data attributes included terrane name and type. Comparison of the tectonic map to the high quality Bouger gravity map, the magnetic data and the 1:1,000,000 scale digital geology maps highlights many inconsistencies. As a result, the tectonic boundaries were shifted to be consistent with the other datasets. In doing so, a previously undefined tectonic unit was established in the North Khentii region. 7. Metamorphic map — A hand-drawn Russian metamorphic map of Mongolia including faults mapped at 1:1,500,000 scale was scanned and georeferenced. Initial problems with the georeferencing were

overcome by correcting the coordinate system used to register the image from a local projection to Latitude/Longitude (WGS84). In order to generate useable datasets, the metamorphic map and the faults on the map were digitized separately. The resulting metamorphic data layer was attributed with rock types, metamorphic grades and age date ranges. Some inconsistencies with rock types were found due to different naming conventions which were solved through expert knowledge of the system used to create the original metamorphic map. Some map units were attributed with multiple metamorphic grades (e.g. greenschist±amphibolite± granulite facies), likely because multiple rock units of potentially differing metamorphic grades occur within each polygon. Derivative layers from the metamorphic map include: a metamorphic grade dataset and contacts between units of different metamorphic grades. A digitized 1:1,500,000 scale fault map was attributed with fault length and strike for analysis. The initial map also indicated thrust faults and those relevant faults were so attributed. As with the 1:1,000,000 scale fault dataset, this dataset had issues with the location of the deposits relative to critical structures. This was again overcome by including a larger buffer around the faults in the analysis. Derivative layers from the fault include: different fault orientation datasets, different fault size datasets, fault bends/jogs of different orientations and fault intersections between faults of varying orientations. 8. Magmatic maps — Unregistered image files of the geochemistry of Mongolian granitoids with attributes such as metaluminous/ peraluminous, I/S/A classification, and K-rich were available. Problems with registering the poor scale images in a GIS could not be overcome, possibly due to distortion of the images during their creation. Additional issues existed with the data coverage which only covered approximately 60% of Mongolia and was

A. Ford, C.J.R. Hart / Ore Geology Reviews 51 (2013) 15–26

critically missing data over what are currently the most highly mineralized regions. Both of these issues lead to the exclusion of the magmatic maps as an input for the mineral potential mapping. 9. Geochemistry databases — Grab samples, stream sediments and pan samples covering all of Mongolia existed from past industry activity. Each dataset was heavily biased towards the North Khentii region due to its more extensive exploration history. As a consequence of this location bias, the geochemistry was only used as an input layer when performing regional scale targeting in the North Khentii area. The grab sample geochemical database was comprehensive and contained sample locations and concentrations for 37 elements. The stream sediment database contained locations of Au, Ag, As, and Sb anomalies, however, no concentrations were listed. A database of pan samples contained the locations of Au anomalies but listed no concentrations. As no absolute values were available for anomaly thresholds in the stream sediment or pan samples databases, they could not be integrated with other data layers and were excluded from the analysis. Expert knowledge was used to extract relevant anomalous geochemical values from the grab samples database to create derivative datasets. 10. Alteration maps — 342 georeferenced ASTER scenes as pre-processed ECW files false colored with six alteration types (argillic, advanced argillic, iron-oxide, phyllic, propylitic and silicification) were available. Data coverage was incomplete and not contiguous and covered only approximately 60% of Mongolia. Critically, data coverage was missing over much of the highly mineralized North Khentii. In order to use the alteration data in the mineral potential mapping, the ASTER scenes were digitized and the alteration types mapped to polygons (15 m pixel size). Verification of the digitized data indicated that there were calibration problems within the original data as the alteration types did not always match up across adjoining scenes. As the raw data were unavailable, it was not possible to have the ASTER scenes reprocessed to improve the quality. Derivative datasets were generated for each of the different alteration types detailed in the original georeferenced ASTER scenes. Based on the above primary datasets, over 70 derivative datasets were generated and considered as proxies for the defined generic orogenic deposit model based on expert knowledge. It should be noted that not all of the features of the deposit model could be represented using the available data. Each derivative dataset was then classified as representing either a source, pathway, focus or trap (Wyborn et al., 1994). A mineral systems model (cf. the petroleum systems approach (e.g. Kreuzer et al., 2008)), requires that each of these four critical ore forming processes be present for the genesis of a hydrothermal mineral deposit (cf. Knox-Robinson and Wyborn, 1997; Kreuzer et al., 2008; Wyborn et al., 1994).

21

5. Application of mineral potential mapping techniques 5.1. Weights of evidence Three types of spatial relationship were tested: proximity, association and abundance. The proximity relationship looks at whether deposits tend to lie closer or further away from a given geological feature. This relationship is tested using a “buffer” analysis. An association relationship determines whether deposits lie within a particular polygon type on a map. This method is tested using an “overlay” analysis. Abundance relationships examine whether deposits are related to an abundance of a particular geological feature. This relationship is tested using “density” analysis. From the initial 99 geological features suggested for the orogenic gold deposit model, we were able to derive map proxies for 26 features from the 70 derivative datasets, some of which are detailed in Table 2. Maximum buffer distances were assigned for derived datasets intended for proximity analysis. These maximum distances were based on expert knowledge of the maximum distance that a given geological feature is effective in the mineralizing process, and prevents geologically meaningless associations. Within the maximum buffer size, a series of intermediate buffers were generated, initially at arbitrary distances for each dataset. Density grids were initially classified based on percentile ranges (e.g., densest 10% through to the least dense 10%). Each dataset was then processed to determine the area and number of deposits that fell within each intermediate buffer or density class. This allowed for calculation of the four statistics used to determine how good the dataset is at predicting known mineralization: contrast, confidence, deposit–area statistic and the observed–expected ratio as previously defined. Sample output from this first-pass statistical analysis is shown in Table 3. The shaded values highlight results for evidential layers that may be considered for further investigation. It is noted here that the evidential layer containing faults with lengths greater than 100 km failed to produce satisfactory statistics. Table 3 indicates that within 25 km of a 100 km fault, only 30% of the training data is captured meaning that it fails to capture 70% of the known mineralization. As such, it may be eliminated as a potential layer in the mineral potential analysis. This first pass statistical analysis resulted in decreasing the number of evidential layers to 18. The elimination of some layers presented some surprising results. For example, mineral deposit models consistently indicate a strong relationship between the proximity of large, crustal scale faults and orogenic gold deposits (Goldfarb et al., 2005; Groves et al., 2003). However, statistical analysis reveals that within a generous buffer of 25 km from such an intersection, only 6% of the training data were detected and poor values obtained for the contrast, confidence,

Table 2 Generic targeting criteria for orogenic gold mineralization in Mongolia for which proxies could be extracted from the available data. Density of fault intersections of any orientation and size fault (flat interpolation) from the 1:1,000,000 scale map Distance from fault intersections of any orientation and size fault from the 1:1,000,000 scale map Distance from gravity high (tilt derivative, 200 m grid) Distance from faults longer than 25 km from the 1:1,000,000 scale map Distance from intersections between geology polygons and faults from the 1:1,000,000 scale map Distance from chemically reactive rock units Distance from granite–sediment contacts Distance from contacts between different metamorphic grades Distance from gravity high blocks (Bouger gravity, 8 km grid) Distance from gravity gradients (Bouger gravity, 8 km grid) Distance from gravity troughs (Bouger gravity, 8 km grid) Distance from contact between geology polygons with high reactivity contrast

Density of fault intersections of any orientation and size fault (normal interpolation) from the 1:1,000,000 scale map Distance from fault intersections of 2 large faults of any orientation from the 1:1,000,000 scale map Density of faults longer than 25 km from the 1:1,000,000 scale map Distance from faults longer than 100 km from the 1:1,000,000 scale map Density of intersections between geology polygons and faults from the 1:1,000,000 scale map Distance from brittle rock units Distance from favorable host rock units Distance from greenschist grade metamorphism Distance from gravity low blocks (Bouger gravity, 8 km grid) Distance from gravity ridges (Bouger gravity, 8 km grid) Distance from magnetite destruction zones Distance from contact between geology polygons with high rheology contrast

22

A. Ford, C.J.R. Hart / Ore Geology Reviews 51 (2013) 15–26

Table 3 Example results from first pass analysis using four statistical measures.

Evidential Layer

Distance (m)

Chemically reactiverock units Chemically reactive rock units Chemically reactive rock units Chemically reactive rock units Chemically reactive rock units Chemically reactive rock units Chemically reactive rock units Chemically reactive rock units Chemically reactive rock units Chemically reactive rock units Chemically reactive rock units Distance from Faults>100km (1:1,000,000) Distance from Faults>100km (1:1,000,000) Distance from Faults>100km (1:1,000,000) Distance from Faults>100km (1:1,000,000) Distance from Faults>100km (1:1,000,000) Distance from Faults>25km (1:1,000,000) Distance from Faults>25km (1:1,000,000) Distance from Faults>25km (1:1,000,000) Distance from Faults>25km (1:1,000,000) Distance from Faults>25km (1:1,000,000) Distance from Geology-fault intersections Distance from Geology-fault intersections Distance from Geology-fault intersections Distance from Geology-fault intersections Distance from Geology-fault intersections

0−1000 1000−2000 2000−3000 3000−4000 4000−5000 5000−6000 6000−7000 7000−8000 8000−9000 9000−10000 >10000 0−1000 1000−5000 5000−10000 10000−25000 >25000 0−1000 1000−5000 5000−10000 10000−25000 >25000 0−1000 1000−5000 5000−10000 10000−25000 >25000

Area % 1.764334421 1.129540698 1.223753966 1.386526484 1.432805569 1.549553633 1.626643003 1.672953917 1.689600383 1.780662599 84.74349801 1.473100839 5.611991567 6.726986499 19.32186817 66.86598926 8.52470934 28.35618854 25.37449774 36.22850282 1.51603790 3.57090567 35.35126908 32.78402118 32.15572053 0

deposit–area statistic and the observed–expected ratio. The most likely cause for such an unexpected result is that the model used in this analysis is not directly applicable to all geological settings. The orogenic model was created with an emphasis on Archean systems and a few large Phanerozoic systems that occur in forearc settings. However, many Phanerozoic gold districts with orogenic deposits such as the Victoria Goldfields, Otago, the Klondike and Meguma, lack direct associations with large faults. With the revised list of evidential layers, buffer distances and density classes were redistributed to best determine where critical distances and cut-offs may occur. This process also assists in determining the validity of the data. For example, are more deposits located close to the geological feature in question? Or do the contrast values follow a sensible pattern? It might reasonably be expected to see a stronger relationship (higher contrast value) between orogenic gold deposits and faults in areas proximal to the fault rather than further away, as such, contrast values should exhibit a direct and decreasing relationship with buffer distances. Using the new redistributed buffer distances and density classes, the statistical analysis was repeated. Sample output is shown in Table 4. Examination of the statistics allowed elimination of the evidential layers with weaker capture efficiency (d(d / a) values) and lower observed– expected ratios. This process resulted in removal of a further six evidential layers from the analysis, leaving 12. Buffer distances were again redistributed to best capture the known deposits and the process repeated. After several more iterations, a total of six evidential layers were chosen as inputs into a unique conditions map (mineral potential map) based on high contrast value, confidence in the contrast, capture efficiency and observed–expected ratio. Each of the layers was assessed for optimal buffer distances and density classes to best determine where critical distances and cut-offs may occur and converted to a binary format. The six evidential layers chosen for analysis were: 1:1,500,000 scale faults mapped from the original metamorphic map, density of intersections between faults and geology polygons from the corresponding 1:1,000,000 maps, gravity features from the Bouger gravity

Deps %

Contrast

Confidence

6 3 2 5 2 2 7 1 0 1 71 1 1 5 23 70 12 34 18 30 6 4 37 44 13 2

0.690655 0.366816 −0.163228 0.734369 −0.335958 −0.422775 0.980698 −1.237983 0 −1.30834 0 −0.29959 −1.777243 −0.241942 0.853916 0 0.43 0.34 −0.38 −0.22 1.47 0.178263 0.161704 0.56906 −1.07574 0

1.50654 0.601559 −0.222732 1.493768 −0.458429 −0.576897 2.259807 −1.216438 0 −1.285572 0 −0.294551 −1.747365 −0.493855 1.978168 0 1.41 1.57 −1.46 −0.98 3.49 0.349167 0.776011 2.801951 −3.612199 0

d(d/a) 20.40429499 7.967840397 3.268630878 18.03066893 2.791725609 2.581388547 30.12338903 0.597745096 0 0.561588703 59.48538966 0.678840154 0.178189862 3.716374338 27.378305 73.2809019 16.89 40.77 12.77 24.84 23.75 4.480656024 38.72562529 59.05315854 5.255674487 −

O-E/E 2.4007158 1.6559468 0.6343154 2.6061338 0.3958628 0.2906943 3.3033413 −0.4022549 −1 −0.4384113 −0.1621776 −0.3211598 −0.8218101 −0.2567251 0.1903611 0.04687 0.41 0.20 −0.29 −0.17 2.96 0.120164 0.0466385 0.3421172 −0.5957173 −

map (8 km grid), greenschist grade metamorphism, source rock units and granite–sediment contacts from the 1:1,000,000 scale geology map. These layers were then converted to binary (if not already in that format) based on critical threshold values for the contrast and combined to create a mineral potential map using weights of evidence. The statistics for the favorable regions for the six binary evidential layers are highlighted in Table 5. Fig. 2 shows the resulting mineral potential map colored by posterior probability with the known orogenic gold deposits. Verification of conditional independence between the input layers was performed using the new omnibus test, which is evaluated as the expected–observed number of deposits divided by the standard deviation of the expected number of deposits for each unique condition (Agterberg and Cheng, 2002). Based on the results of a simple chi-square test, two individual pairwise correlations were observed to not have conditional independence (1:1,500,000 scale faults vs. gravity features, and granite–sediment contacts vs. gravity features). However, when all six layers were combined, the conditional independence test was passed with a NOT value of 0.001. In order to generate targets for field testing, areas with the highest posterior probability were selected. A small buffer was then placed around each of these areas so that small individual areas would mesh with others to form potential trends. This buffering also expanded the spatial area which accounted for the location uncertainty that is inherent in some of the data layers. The number of known orogenic gold deposits, orogenic gold occurrences, geochemical anomalies or placer occurrences was then counted for each of these buffered areas. Areas containing geochemical anomalies or other indicators of mineralization have the highest level of confidence as they indicate that mineralizing processes were active in those areas, and also indicate that there was some previous exploration. As a result, these areas may be ranked higher. This iterative process can then be repeated for terrain-scale study areas that may have more specific deposit models. At a regional scale, the range of potential datasets available for use may also be increased. Datasets such as the magnetic maps and alteration that were excluded as direct inputs into the country-scale analysis due to incomplete coverage and insurmountable concerns over data quality could be considered

A. Ford, C.J.R. Hart / Ore Geology Reviews 51 (2013) 15–26

23

Table 4 Example statistics from revised targeting criteria.

Evidential layer

Class description

Area %

Deps %

Contrast

Confidence

Fault intersection density, all faults (1:1,000,000) Fault intersection density, all faults (1:1,000,000) Fault intersection density, all faults (1:1,000,000) Fault intersection density, all faults (1:1,000,000) Fault intersection density, all faults (1:1,000,000) Fault intersection density, all faults (1:1,000,000) Fault intersection density, all faults (1:1,000,000) Fault intersection density, all faults (1:1,000,000) Fault intersection density, all faults (1:1,000,000) Fault intersection density, all faults (1:1,000,000) Distance from 1:1,500,000 faults Distance from 1:1,500,000 faults Distance from 1:1,500,000 faults Distance from 1:1,500,000 faults Distance from 1:1,500,000 faults Distance from 1:1,500,000 faults Distance from 1:1,500,000 faults Distance from 1:1,500,000 faults Distance from 1:1,500,000 faults Distance from 1:1,500,000 faults Distance from 1:1,500,000 faults Distance from 1:1,500,000 faults Distance from 1:1,500,000 faults Distance from 1:1,500,000 faults

0% to 10% 10% to 20% 20% to 30% 30% to 40% 40% to 50% 50% to 60% 60% to 70%

10.44137643 10.44611892 10.46209698 10.44780585 10.45541293 10.44408188 10.45191177

2 4 5 10 19 10 25

−1.693642 −0.980374 −0.748454 0.000296 0.746733 0.000693 1.098523

−2.371091 −1.92112 −1.631204 0.000889 2.929362 0.002078 4.756593

0.383091255 1.531669333 2.389578307 9.571387662 34.52756981 9.574800462 59.79767281

−0.80845437 −0.61708267 −0.52208434 −0.04286123 0.817240516 −0.04251995 1.391906912

70% to 80% 80% to 90% 90% to 100% 0-1 km 1-2 km

10.4523892 10.45308943 10.45152982 7.476427586 6.860540173

5 13 7 13 17

−0.747422 0.296011 −0.389574 0.525372 0.935255

−1.628956 0.995475 −0.993977 1.764103 3.505653

2.391797658 16.16746906 4.688308874 22.60437864 42.12496286

−0.52164047 0.243651466 −0.33024159 0.738798357 1.477938992

2-3 km 3-4 km 4-5 km 5-6 km 6-7 km 7-8 km 8-9 km 9-10 km 10-15 km 15-20 km 20-25 km > 25 km

6.205248698 6.014084884 5.562402286 4.874199824 4.794054888 4.488943946 4.118361147 4.013262427 16.27063143 11.48886246 7.975089503 9.85769977

10 10 8 11 3 8 2 6 5 4 1 2

0.429938 0.463506 0.30131 0.793871 −0.575917 0.528346 −0.832051 0.336227 −1.410224 −1.233354 −2.242311 −1.773908

1.288318 1.388904 0.81669 2.480723 −0.98212 1.432055 −1.164619 0.797955 −3.071844 −2.415828 −2.230837 −2.482945

16.11538955 16.62763362 11.5058201 24.8245875 1.877325189 14.25725088 0.971260134 8.970258151 1.536510744 1.392653106 0.125390442 0.405774176

0.611538955 0.662763362 0.438227512 1.256780681 −0.37422494 0.782156359 −0.51436993 0.495043025 −0.69269785 −0.65183672 −0.87460956 −0.79711291

within more limited areas and may be included as potential evidential layers at that scale. 5.2. Fuzzy logic Based on expert opinion, a deposit model was generated for a countryscale orogenic gold deposit model. This model took into account the availability of data and the knowledge of the data quality gained through the iterative weights of evidence analysis previously performed. Table 6 details the evidential layers and fuzzy membership values defined from expert geological knowledge. Due to the potential for incomplete mapping on the metamorphic map, regions containing no data were assigned a value of 0.45, half the maximum (0.9 for greenschist). Where appropriate, buffers were placed around the geological features with the distance determined by expert knowledge of the orogenic gold deposit model and consideration for potential mapping errors (e.g. inaccuracies with the 1:1,000,000 scale geology and fault maps). The “missing data” value was set to 0.01 in order to avoid removing any area totally from the mineral potential map. The evidential layers were then combined using the fuzzy algebraic sum with the resulting mineral potential map shown in Fig. 3. This allows the weight assigned to each evidential layer to have an effect on the result (Bonham-Carter, 1994). The resulting values from the fuzzy algebraic sum represent the relative prospectivity at a given location. Higher values represent locations with higher potential for mineralization. Similar to the weights of evidence analysis, targets for field testing were selected based on areas with the highest fuzzy algebraic sum values (cf. highest posterior probability values). Small buffers were placed around each of these areas to link small areas and to identify potential trends and indicators of mineralization used to assist in ranking the targets as previously discussed. 6. Discussion Validation of the mineral potential maps is a critical part of the analysis. The ability to accurately predict the locations of known orogenic gold

d(d/a)

O-E/E

deposits is used to validate the mineral potential maps generated by the weights of evidence and fuzzy logic techniques employed in this study. Both the mineral potential map produced by weights of evidence and that produced by fuzzy logic are able to accurately predict the locations of the major orogenic gold deposits in Mongolia. The weights of evidence mineral potential map predicts 53% of the deposits within 13.2% of the study area while the fuzzy logic map predicts 74% of the deposits within 12.1% of the study area. The Boroo, Gatsuurt and Tsagaan Tolgoi deposits are directly detected from the weights of evidence mineral potential map as targets. The trends hosting the Bumbat, Nariin Gol and Tsagaan Tsahir Uul deposits are also clearly highlighted as areas of higher potential for mineralization. Fuzzy logic highlights the trends in which these six orogenic gold deposits sit, however none were directly selected as potential exploration targets for field testing based on the method herein outlined for target ranking due to the lack of indicators of hydrothermal activity (i.e. alteration, anomalous geochemistry, etc.) as previously discussed. The fuzzy logic success benefited from knowledge derived from the considerably more exhaustive weights of evidence iterative testing of evidential layers. The small number of actual orogenic gold deposits and mines in Mongolia presents a significant disadvantage in being able to verify the validity of the mineral potential maps generated. While expert geological knowledge was used to reclassify the gold deposit types from those originally assigned in the database, it is possible that some of the deposits assigned as orogenic may belong to another type. This may be one reason for the relatively poor prediction efficiency of both the weights of evidence and fuzzy logic mineral potential maps. Comparison of the mineral potential maps generated by weights of evidence (posterior probability value) and fuzzy logic (fuzzy algebraic sum) indicates an area-weighted Spearman's rank correlation coefficient of 0.4674 at 99% confidence (cf. Raines, 1999). Some level of dependence between the two mineral potential maps should be expected given that there is some degree of overlap in the evidential layers used in their generation. How this correlation might vary with the assignment of different expert driven fuzzy membership values or with the choice of a different fuzzy operator during the

24

A. Ford, C.J.R. Hart / Ore Geology Reviews 51 (2013) 15–26

Table 5 Statistics for the evidential layers for the orogenic gold weights of evidence model of Mongolia. Evidential layer

Buffer distances (km)

Area (km2)

# Deps

Contrast

Confidence

D2/A

O–E / E

“Brittle” rock unitsa 1.5 M scale faults Density geology–fault intersections Gravity features Greenschist metamorphism Granite–sediment contacts

0–3 0–2 60% to 100% intensity 0–15 0–5 0–2

379,826.5 225,220.5 656,753.5 676,054 493,752 302,735

39 30 74 64 70 40

0.4169 0.8471 1.4513 0.6594 1.1793 0.9739

1.9987 3.8646 6.3657 2.4716 4.7399 4.7549

62.9062 62.7746 130.9817 95.1762 155.8967 83.0246

0.6130 1.0925 0.7700 0.4871 1.2271 1.0756

a

“Brittle” rock units — felsic intrusions, mafic intrusions, and quartzite.

fuzzy logic analysis remains an open question (cf. Nykänen et al., 2008). A key feature in the success of these experiments was the combination of geological and GIS expertise. Decision making throughout the process requires critical evaluation of each situation, and working with known information allows the decision to be based on expert knowledge, and not simply an assumption. Similarly, understanding the processing parameters and significance of features such as the propagation of uncertainties is critical to decision-making and recognizing the value of the results. While this study was successfully able to predict the location of known orogenic gold mineralization in a frontier terrain using both weights of evidence and fuzzy logic, it also highlights the need for high quality, publicly available geoscience datasets to assist explorers in frontier regions. Access to such data would contribute to improvements in mineral potential mapping results, which in turn, serve to promote investment in the mineral resources sector in such areas. 7. Conclusions Quantitative evaluation of the spatial relationships between geological features and mineral deposits provides a simple yet effective method for mapping the potential for mineralization to have formed in a given

area. An iterative weights of evidence technique can be used to create a mineral potential map in a frontier terrain despite deficiencies in data quality and data availability. Expert driven fuzzy logic can also be used in regions where data availability presents the primary challenge. The validity of the resultant mineral potential maps is verified by their ability to predict known mineralization. The most thorough test of a map is its ability to predict currently unknown mineralization, something that can only be verified over time through exploration. This paper has discussed several drawbacks of this method: the elimination of critical datasets due to incompleteness, possible misclassification of training data used in the iterative weights of evidence process and difficulties in verifying the resultant mineral potential map due to lack of validation data. Yet, despite these challenges, application of the iterative weights of evidence technique to a countryscale analysis of orogenic gold potential in Mongolia indicates that this method can be successfully used to highlight areas of known mineralization and delineate targets for further investigation. The results of this study show how mineral potential mapping can be utilized as a tool in exploration in frontier terranes, and some of the methods utilized to overcome deficiencies in data quality and availability. Such analysis may be used as part of a broader exploration campaign to reduce country-scale targets to smaller regional-scale targets, however, field validation of targets is still critical.

Fig. 2. Weights of evidence mineral potential map derived using evidential layers outlined in Table 5. Orogenic gold data points are highlighted.

A. Ford, C.J.R. Hart / Ore Geology Reviews 51 (2013) 15–26 Table 6 Evidential layers and fuzzy membership values for Mongolia orogenic gold fuzzy logic model. Evidential layer

Fuzzy membership value

Distance from fault 1:1,500,000 faults from metamorphic map 1:200,000 faults Metamorphic grade Greenschist Amphibolite Sub-greenschist No metamorphic data Tectonic units Accretionary wedge Arc Backarc/forearc Ophiolite Cratonic Metamorphic Permian–Triassic volcanic Distance from terrain boundary Gravity gradients (Bouger gravity, 8 km grid) Favorable host rock units Granites Gold placers

0.8 0.7 0.9 0.2 0.2 0.45 0.9 0.7 0.7 0.6 0.4 0.1 0.2 0.75 0.6 0.7 0.8 0.9

The weights of evidence and fuzzy logic analyses described in this study were performed using the publicly available MapInfo Spatial Data Modeller package from Avantra Geosystems. Acknowledgments This study was undertaken as part of a collaborative research project between Troy Resources NL in Mongolia in 2008–09 and the Centre for Exploration Targeting at the University of Western Australia. The targeting products produced during the project have since been purchased by Meritus Minerals Ltd. The team of geologists from Troy Resources is thanked for providing expertise on the local geology and with compiling of the datasets, with special thanks to Tserenjav Tsevegdorj, Ganaa Ochirbal, Baigalmaa Lkhagvasuren and David

25

Otterman; and the Troy Resources management is commended for their forward-thinking in supporting this research. Steve Gardoll is thanked for discussions on targeting gold mineralization and Inna Mudrovska for assistance with translation of maps and documents. The authors would also like to thank the reviewers and editors, particularly Alok Porwal and John Carranza for their reviews which greatly improved the manuscript.

References Agterberg, F.P., Bonham-Carter, G.F., 2005. Measuring the performance of mineralpotential maps. Nat. Resour. Res. 14, 1–17. Agterberg, F.P., Cheng, Q., 2002. Conditional independence test for weights-of-evidence modeling. Nat. Resour. Res. 11, 249–255. Agterberg, F.P., Bonham-Carter, G.F., Cheng, Q., Wright, D.F., 1993. Weights of evidence modeling and weighted logistic regression for mineral potential mapping. In: Davis, J., Herzfeld, U.C. (Eds.), Computers in Geology — 25 Years of Progress. Oxford Univ. Press, New York, pp. 13–32. Badarch, G., Cunningham, W.D., Windley, B.F., 2002. A new terrane subdivision for Mongolia: implications for the Phanerozoic crustal growth of Central Asia. J. Asian Earth Sci. 21, 87–110. Bonham-Carter, G., 1994. Geographic Information Systems for Geoscientists: Modelling with GIS. Pergamon Press, Oxford . (398 pp.). Bonham-Carter, G.F., Agterberg, F.P., Cheng, Q., 1989. Weights of evidence modelling: a new approach to mapping mineral potential. In: Agterberg, F.P., Bonham-Carter, G.F. (Eds.), Statistical Applications in the Earth Sciences. Geological Survey of Canada Paper 89–9, pp. 171–183. Bougrain, L., Gonzalez, M., Bouchot, V., Cassard, D., Lips, A.L.W., Alexandre, F., Stein, G., 2003. Knowledge recovery for continental-scale mineral exploration by neural networks. Nat. Resour. Res. 12, 173–181. Brown, W.M., Gedeon, T.D., Groves, D., Barnes, R.G., 2000. Artificial neural networks: a new method for mineral prospectivity mapping. Aust. J. Earth Sci. 47, 757–770. Brown, W., Groves, D., Gedeon, T., 2003. Use of fuzzy membership input layers to combine subjective geological knowledge and empirical data in a neural network method for mineral-potential mapping. Nat. Resour. Res. 12, 183–200. Carranza, E.J.M., Hale, M., 2002. Evidential belief functions for data-driven geologically constrained mapping of gold potential, Baguio district, Philippines. Ore Geol. Rev. 22, 117–132. Carranza, E.J.M., Woldai, T., Chikambwe, E.M., 2005. Application of data-driven evidential belief functions to prospectivity mapping for aquamarine-bearing pegmatites, Lundazi District, Zambia. Nat. Resour. Res. 14, 47–63. Cluer, J.K., Kotlyar, B., Gantsetseg, O., Togtokh, D., Wood, G., Ullrich, T., 2005. Geology of the Boroo gold deposit, northern Mongolia. In: Seltmann, R., Gerel, O., Kirwin, D.J. (Eds.), Geodynamics and Metallogeny of Mongolia with a Special Emphasis on Copper and Gold Deposits. IAGOD Guidebook Series, London, pp. 105–117.

Fig. 3. Fuzzy logic mineral potential map derived using evidential layers outlined in Table 6. Orogenic gold data points are highlighted.

26

A. Ford, C.J.R. Hart / Ore Geology Reviews 51 (2013) 15–26

Feltrin, L., 2009. Predictive modelling of prospectivity for Pb–Zn deposits in the Lawn Hill Region, Queensland, Australia. Comput. Geosci. 35, 108–133. Fung, C.C., Iyer, V., Brown, W., Wong, K.W., 2005. Comparing the performance of different neural networks architectures for the prediction of mineral prospectivity. Proceedings of the Fourth International Conference on Machine Learning and Cybernetics. Guangzhou, pp. 394–398. Goldfarb, R.J., Baker, T., Dubé, B., Groves, D.I., Hart, C.J.R., Gosselin, P., 2005. Distribution, character, and genesis of gold deposits in metamorphic terranes. In: Hedenquist, J.W., Thompson, J.F.H., Goldfarb, R.J., Richards, J.P. (Eds.), Economic Geology 100th Anniversary Volume. Society of Economic Geologists, Littleton, Colorado, pp. 407–450. Groves, D.I., Goldfarb, R.J., Robert, F., Hart, C.J.R., 2003. Gold deposits in metamorphic belts: overview of current understanding, outstanding problems, future research, and exploration significance. Econ. Geol. 98, 1–30. Harris, D., Zurcher, L., Stanley, M., Marlow, J., Pan, G., 2003. A comparative analysis of favorability mappings by weights of evidence, probabilistic neural networks, discriminant analysis, and logistic regression. Nat. Resour. Res. 12, 241–255. Harris, J.R., Lemkow, D., Jefferson, C., Wright, D., Falck, H., 2008. Mineral potential modelling for the Greater Nahanni ecosystem using GIS based analytical methods. Nat. Resour. Res. 17, 51–78. Hendry, J.W., Roscoe, W.E., Ross, D.A., 2006. Technical Report on the Gatsuurt Gold Project in Northern Mongolia. NI43-101 Report prepared for Centerra Gold Inc. Knox-Robinson, C.M., 2000. Vectorial fuzzy-logic: a novel technique for enhanced mineral prospectivity mapping, with reference to the orogenic gold mineralisation potential of the Kalgoorlie Terrane, Western Australia. Aust. J. Earth Sci. 47, 929–941. Knox-Robinson, C.M., Wyborn, L.A.I., 1997. Towards a holistic exploration strategy: using Geographic Information Systems as a tool to enhance exploration. Aust. J. Earth Sci. 44, 453–463. Kreuzer, O.P., Etheridge, M.A., Guj, P., McMahon, M.E., Holden, D.J., 2008. Linking mineral deposit models to quantitative risk analysis and decision-making in exploration. Econ. Geol. 103, 829–850. McCuaig, T.C., Beresford, S., Hronsky, J., 2010. Translating the mineral systems approach into an effective exploration targeting system. Ore Geol. Rev. 38, 128–138.

Mustard, R., Blenkinsop, T., McKeagney, C., Huddleston-Holmes, C., Partington, G., 2004. New perspectives on IOCG deposits, Mt Isa Eastern Succession, northwest Queensland. SEG 2004 Predictive Mineral Discovery Under Cover: Extended Abstracts. Centre for Global Metallogeny, University of Western Australia, Perth, pp. 281–284. Nykänen, V., Groves, D.I., Ojala, V.J., Eilu, P., Gardoll, S.J., 2008. Reconnaissance-scale conceptual fuzzy-logic prospectivity modelling for iron oxide copper–gold deposits in the northern Fennoscandian Shield, Finland. Aust. J. Earth Sci. 55, 25–38. Penney, S.R., Allen, R.M., Harrisson, S., Lees, T.C., Murphy, F.C., Norman, A.R., Roberts, P.A., 2004. A global-scale exploration risk analysis technique to determine the best mineral belts for exploration. Appl. Earth Sci. IMM Trans. B 113, 183–196. Porwal, A., Carranza, E.J.M., Hale, M., 2001. Extended weights-of-evidence modelling for predictive mapping of base metal deposit potential in Aravalli Province, western India. Explor. Min. Geol. 10, 273–287. Porwal, A., Carranza, E.J.M., Hale, M., 2006. Bayesian network classifiers for mineral potential mapping. Comput. Geosci. 32, 1–16. Raines, G.L., 1999. Evaluation of weights of evidence to predict epithermal-gold deposits in the Great Basin of the western United States. Nat. Resour. Res. 8, 257–276. Singer, D.A., Kouda, R., 1999. A comparison of the weights of evidence method and probabilistic neural networks. Nat. Resour. Res. 8, 287–298. Trench, A., Packey, D., 2012. Australia's Next Top Mining Shares: Understanding Risk and Value in Minerals Equities. Major Street Publishing, Australia . (340 pp.). Wyborn, L.A.I., Heinrich, C.A., Jaques, A.L., 1994. Australian Proterozoic mineral systems: essential ingredients and mappable criteria. In: Hallenstein, P.C. (Ed.), Australian Mining Looks North — The Challenges and Choices: Australian Institute of Mining and Metallurgy Publication Series, 5, pp. 109–115. Zhang, S., Cheng, Q., Chen, Z., 2008. Omnibus weights of evidence method implemented in GeoDAS GIS for information extraction and integration. J. China Univ. Geosci. 19, 404–409.