You are on page 1of 10

Spatial integration of geological data for predictive mineral mapping:

A case study from Taebaek-san area, Korea

Kwang-Hoon Chi1), No-Wook Park1), and Chang-Jo F. Chung2)

1) National Geoscience Information Center, Korea Institute of Geoscience and Mineral Resources,
30 Kajung-dong, Yusung-ku, Daejon 305-350, Korea
email : {khchi, nwpark}@rock25t.kigam.re.kr
2) Geological Survey of Canada, 601 Booth Street, Ottawa, Ontario K1A 0E8, Canada
email : chung@gsc.nrcan.gc.ca

Abstract

This study involves the integration of multiple geological data sets for mineral exploration
through spatial prediction models. To identify mineralized areas in Ogdong, Korea, we used
geological, geophysical, and geochemical data sets. As prediction models, joint conditional
probability function and likelihood ratio function were applied in order to obtain the potential
maps. Also, to evaluate the prediction results, cross-validation technique was carried out. From
the cross-validation analysis, most favourable potential areas of mineralization for further
investigation were delineated.

1. Introduction

Recently, the impact of GIS has been widely felt in many applications in the geosciences.
Especially, the integrated analysis of spatial data from multi-source has become increasingly
important to make optimized decisions.
Several schemes for mathematical representation and/or integration of spatial geological
information have been studied since late 1980s. Known methods include weight of evidence
(Bonham-Carter et al., 1988; Chung and Fabbri, 1998), the evidential belief function approach
(Moon, 1990; Chung and Fabbri, 1993), the probabilistic approach based on Bayesian
conditional probability theory (Chung and Moon, 1991; Chung and Fabbri, 1999), and fuzzy
logic approach (An et al., 1991). These approaches with their own mathematical backgrounds
have provided powerful schemes for decision-supporting information, through several case
studies using various geological and geophysical data sets for mineral exploration. However,
proper interpretation and quantitative evaluation of prediction result have not been fully
considered in real mineral exploration application, and is in the development stage (Chung,
2001; Chung and Keating, 2001).
In this study, prediction models for mineral exploration task using various geological,
geophysical, and geochemical data sets from Ogdong, Korea are investigated with real
applications. The objectives of this study include outlining of an area with the highest potential
and evaluating the prediction. Two data-driven prediction models are applied, and the prediction
results are evaluated with the help of cross-validation technique.

2. Study area and Data sets

For mineral potential mapping, multiple geological data sets covering the Ogdong, in
Taebaek-san area, Korea were used. Geologically, the major part of the study area is composed
of metasediments, and most of the polymetallic mines are located in the western and eastern
areas (Fig. 1). Ore deposits are mainly located within the thick limestone series of the Joseon
Supergroup and the vicinity of the granite; the rich elements of ore deposits are iron (Fe), lead
(Pb), and zinc (Zn).
The multiple geological data sets are listed in Table 1. After pre-processing, all data sets were
built into a cell-based database with 30 meter resolution and the whole study area consists of
451,870 pixels (730 by 619 pixels), covering approximately 407km2. In addition, seven known
mines in the study area were added to the database and used as prior evidences in the spatial
integration.

KOREA

Fig. 1. The location map and geological map of the study area. Black boxes denote the known mineral
deposits.
Table 1. Description of the data sets used in the study area.
Types Description
Geological
Lithology map Digitized and converted to raster map*
map
Airborne Gammaray, Potassium(K),
geophysical Thorium(Th), Digitized and converted to raster map**
data sets Residual magnetic intensity
Geochemical
Ag, Cd, Cu, Pb, Zn Interpolated and converted to raster map***
data sets
Data sources :
* Geological map of Ogdong, Geological survey of Korea, 1966
** Aerial gammaray and magnetic survey map at Jungseon, Samcheok, Yemi, Korean Institute of
Energey and Resources, 1988
*** Geochemical maps for Ogdong sheet in the Taebaegsan mineralized belt, Korean Institute of
Energey and Resources, 1984

3. Prediction models

In this case study, we applied two prediction models: joint conditional probability function
and likelihood ratio function (Chung and Fabbri, 1998; Chung and Fabbri, 1999). In both
approaches, the prediction models and estimation procedures are depending on the basic
quantitative relationships of geological data sets with respect to selected target features, e.g.
mineral occurrences. We assume that undiscovered Fe-Cu-Zn deposits exist in the study area
and geological data sets provide useful information to identify mineralized areas.
Consider a study area A where an exploration target for a specific mineral deposit type D is
sought, and a pixel p in A and a proposition:
Tp : p contains a deposit of type D
For each k-th layer Lk, the m layers of map data at every p in A are represented, in a
quantized form by:
{ ( vk(p), k=1,, m), A}
where, vk(p) is the quantized value for the k-th layer Lk at p.

Then, we define the favourability function fk of at each p for the k-th layer Lk.
fk(p) = function(Tp|vk(p))
A fk(p) value indicates that how each of the m pieces of evidence (vk(p), k=1,, m))
supports the sureness that the proposition is true at p.
Two models employed in this study are described briefly here. Details of conceptual
background are discussed in Chung and Fabbri (1993, 1998, 1999).

(1) Joint conditional probability function

In this prediction model, favourability function fk(p) is interpreted as the conditional


probability, denoted by Probk{Tp|vk(p) at p}. After Probk{Tp|vk(p) at p} is calculated for every k,
these conditional probability is combined by Bayesian combination rule. Under the conditional
independence assumption, the joint conditional probability can be expressed in terms of
bivariate conditional probabilities of the occurrences of the deposits given pixel values at each
layer separately. The conditional independence condition is not related to the actual spatial
pattern, but related to the conceptual understanding how data layers are related to the deposits.

(2) Likelihood ratio function

This model represents the ratio of the two distribution functions, those of the mineralized
and the non-mineralized subareas. So this model can highlight this difference. In this model,
we assume that the distribution functions of the mineralized and non-mineralized subareas
should be distinctly different. For each data layer, two empirical distribution functions for the
mineralized and non-mineralized subarea are computed, firstly. Then the likelihood ratio
function for the layer is computed. Using the Bayesian combination rule, the likelihood ratio
functions for all data layers are combined. The prediction maps are obtained from these joint
likelihood ratio functions.

4. Assessment of prediction model

Firstly, in order to carry out the spatial integration, the study area was divided into a number
of non-overlapping unique condition subareas by overlaying the input data layers specified in
the GIS database. An image consisting of the unique condition subareas was generated, where
the unique condition subareas mean that pixel values in a subarea have a unique identification
number, unique to that subarea. After generating the unique condition subareas, we computed
the bivariate conditional probabilities and the likelihood ratio of the known mineral deposit
occurrences. As for the prior evidence, the 25 pixles (5 x 5) surrounding each deposit are
considered as the mineralized area. So, 175 pixels are assigned to mineralized areas and the
remaining 451,695 pixels are assumed to be non-mineralized areas.
To visualize the prediction results, we first compute the score of either the joint conditional
probability or the joint likelihood ratio function for each pixel and then we sort all scores by
increasing order to determine the ranks of the scores. The pixel that has the smallest score (the
smallest prediction value) has rank one, and the pixel that has the maximum score has the
maximum rank, 451,870, the total number of pixels. Then the ranks are normalized so that the
maximum value is 1 or 100% and the normalized values are termed the favourability indices or
simply indices. The pixel with the index 100% had the largest score of the prediction function.
If the pixels have index, 99.5%, it means that the ranks of their function scores are within the
top 0.5% (99.5% - 100%) in the study area. These indices over the study area constitute a
potential prediction map. These indices allow us to compare the prediction results obtained by
different favourability functions.
The highest six classes between 70% and 100% with the 5% increment are shown in Fig. 2.
In both models, regional patterns are similar to each other in general. Overall, the resultant layer
fitted the real situation, showing in known mineral occurrences. To compare each prediction
pattern to the known seven occurrences quantitatively, we constructed success rate curve, in
terms of proportion of seven known mineralized areas with respect to prediction results (Fig. 3).
These success rates illustrate how well the estimators perform with respect to seven known
mineralized areas. In both models, all seven mineralized areas were included within the most
potential 12% (88% - 100% in the indices) area consisting of 48.8 km2, of the whole study area.
However, since all mineralized areas were used to construct the models, these results are not
prediction results, but rather assessment results.
The next question in prediction modeling is how successful these prediction maps would be
with respect to the locations of future discoveries of the same type of mineral deposits in the
study area. It leads to the next essential step of cross-validation.

(a) (b)

Fig. 2. Spatial integration results using (a) Joint conditional probability function, (b) Likelihood ratio
( (
function.
100

Success rate in percentage


90
80
70
60
50
40 Likelihood ratio function
30
20 Joint conditional probability
10 function
0
0 1 2 3 4 5 6 7 8 9 10 11 12

Area in percentage

Fig. 3. Success rate curves for joint conditional probability function and likelihood ratio function.

5. Cross-validation approach

The critical strategy in prediction model is the task of validating the prediction results with
the distribution of future discoveries.
Since known deposits are located in the western (four deposits) and north-eastern (three
deposits) in the spatially separated area, we subdivided seven deposits into two groups, western
part and northeastern part. Then, we have applied the following cross-validation procedure.
Assuming that four mineralized areas in western part selected have not yet been discovered,
we construct a prediction model using the remaining three mineralized areas in northeastern part.
To verify the prediction result, we construct prediction rate curve, in terms of proportion of
four mineralized areas in western part with respect to the prediction result from the northeastern
occurrences. By substituting western part for northeastern part, we repeated above procedure.

Fig. 4. Distribution of known deposits overlaid on Landsat TM imagery. These deposits are subdivided
into two groups expressed to red circle.
(a) (b)

Fig. 5. Predicted results using joint conditional probability function. (a) prediction maps using only
western mineralized zones, (b) prediction maps using only north-eastern mineralized zones.

(a) (b)

Fig. 6. Predicted results using likelihood ratio function. (a) prediction maps using only western
mineralized zones, (b) prediction maps using only north-eastern mineralized zones.

The cross-validation results are shown in Fig. 5 and Fig. 6. Both of two prediction models
show in general the same areas as having high potential. However, prediction rate curves for
two models show the difference of prediction power quantitatively (Fig. 7).
100 100
Prediction rate in percentage

Prediction rate in percentage


90 90
80 80
70 70
60 60
50 50
40 Likelihood ratio 40
Likelihood ratio
30 function 30 function
20 Joint conditional 20 Joint conditional
10 (a) probability function 10 (b) probability function
0 0
0 5 10 15 20 25 30 0 5 10 15 20 25 30

Area in percentage Area in percentage

(a) (b)

Fig. 7. Prediction rate curve (a) for north-eastern part deposits which are not used in computing Fig.
5(a) and Fig. 6(a), (b) for western part deposits which are not used in computing Fig. 5(b) an Fig. 6(b).

In the likelihood ratio function prediction result obtained using the northeastern occurrences,
all four mineralized areas were included within the most potential 25% (75% - 100% in the
indices) consisting of 101.7 km2. In addition, assuming that undiscovered deposits exit in this
study area, it is expected that all new discoveries will be located within the best 25% areas in
Fig. 6 (b). On the other hands, based on the likelihood ratio function prediction results obtained
using the western occurrences, all three mineralized areas were included within the most
potential 30% (70% - 100% in the indices) consisting of 152.6 km2. This result using the
western occurrences is much worse than that from the previous analysis using the northeastern
occurrences. Two prediction maps shown in Fig. 6 (a) and (b) appear somewhat different. For
any prediction models to generate reasonably good or significant results, the prediction results
should be robust and stable (Chung et. al, 2001).
To test the stability (Chung et. al, 2001) of the prediction results in Fig. 5 (a) and (b), and Fig.
6 (a) and (b) of the models, two prediction maps in each model were compared by first dividing
each map into two classes: the most favourable (the highest potential 10%) and the remaining
area. Then, the maps were overlain with the resulting four classes. We denote the most
favourable area in two maps to both favourable high potential area. When the most favourable
area appears in only one map, this area is denoted to one favourable high potential area.
Above procedures are displayed in Fig. 8. As can be observed in Fig. 8 (a) and (b), both of the
models are not stable. Consequently, it would be difficult to expect that the prediction maps
generate any significant results. Three out seven mineralized zones are located within both
favourable high potential area in joint conditional probability function prediction (Fig. 8 (a)).
Meanwhile, four out seven mineralized zones are located within both favourable high
potential area in likelihood ratio function prediction (Fig. 8(b)).
(a) (b)

Fig. 8. (a) Favourable map of joint conditional probability function prediction. (b) Favourable map of
likelihood ratio function prediction. Black boxes denote the known mineral deposits.

6. Discussion and Conclusion

In this study, we applied two prediction models for predictive mineral potential mapping
using multiple geological data sets, and outlined the mineralized zones. Though the mineral
favourability maps generated in this study are not intended to be used directly for defining drill
targets but rather to define specific areas that warrant further detailed exploration, these results
can be used effectively for excluding areas that lack the general geological, geophysical, and
geochemical characteristics determined from the prediction model. With the help of cross-
validation technique, we can evaluate the prediction results quantitatively, compare models, and
outline confident favorable areas. Without this kind of the cross-validation technique,
prediction maps cannot be evaluated.
However, several aspects still need to be addressed further consideration should be given to
prediction. We had limited geological information in this study area in addition to the seven
limited number of mineralized areas.
The important tasks in spatial data integration are that one is the choice of prediction model
and the other is the selection of relevant data layers which contain useful information in causal
factors. We did not consider the influence of input causal factors in this study, related to latter
matter. We are currently analyzing the selection of causal factors, and evaluating other
prediction models such as fuzzy logic model and evidential approach.
7. References

An, P., W. M. Moon, and A. Rencz, 1991, Application of fuzzy set theory to integrated mineral
exploration, Canadian Journal of Exploration Geophysics, 27(1), p.1-11
Bonham-Carter, G. F., F. P. Agterberg, and D. F. Wright, 1988, Integration of geological data sets
for gold exploration in Nova Scotia, Photogrammetric Engineering & Remote Sensing,
54(11), p.1585-1592
Chung, C. F., and W. M. Moon, 1991, Combination rules of spatial geoscience data for mineral
exploration, Geoinformatics, 2, p.159-169
Chung, C. F., and A. G. Fabbri, 1993, The representation of geoscience information for data
integration, Nonrenewable Resources, 2(2), p.122-139
Chung, C.F., and Fabbri, A.G., 1998, Three Bayesian prediction models for landslide hazard. In,
A. Bucciantti, ed., Proceedings of International Association for Mathematical Geology
Annual Meeting (IAMG98), Ischia, Italy, October 3-7, 1998, p.204-211.
Chung, C. F., and A. G. Fabbri, 1999, Probabilistic prediction models for landslide hazard
mapping, Photogrammetric Engineering & Remote Sensing, 65(12), p.1389-1399
Chung, C.F., 2001, Use of airborne geophysical surveys for constructing mineral potential map,
Submitted to Economic Geology for publication
Chung, C.F., and Fabbri, A.G., 2001, Validation of spatial prediction models for landslide
hazard mapping, Submitted to Natural Hazard for publication.
Chung, C.F., and Keating, P., 2001, Mineral potential evaluation based on airborne geophysical
data, Submitted to Geophysical Exploration for publication.
Chung, C.F., Kojima, H., and Fabbri, A.G., 2001, Stability analysis of prediction models for
landslide hazard mapping, In, R.J. Allison, ed., Applied Geomorphology: theory and
practice, John Wiley & Sons, Ltd (in press)
Moon, W. M., 1990, Integration of geophysical and geological data using evidential belief
function, IEEE Trans. Geoscience and Remote Sensing, 28(4), p.711-720

You might also like