See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/248955306

Artificial Neural Networks: A New Method for Mineral Prospectivity Mapping

Article in Australian Journal of Earth Sciences · August 2000

DOI: 10.1046/j.1440-0952.2000.00807.x

CITATIONS READS

303 1,577

4 authors, including:

Warick Brown Tom Gedeon

Curtin University Australian National University
16 PUBLICATIONS 608 CITATIONS 471 PUBLICATIONS 9,433 CITATIONS

SEE PROFILE SEE PROFILE

D. I. Groves
University of Western Australia
479 PUBLICATIONS 33,328 CITATIONS

SEE PROFILE

All content following this page was uploaded by Tom Gedeon on 13 February 2019.

The user has requested enhancement of the downloaded file.

 The Use of a Multilayer Feedforward Neural Network for Mineral
Prospectivity Mapping
Warick Brown', Tamds Gedeon', and Robert Barnes'

' Centre for Strategic Mineral Deposits, Department of Geology and Geophysics, University of Western
Australia, Perth, WA 6907, Australia. E-mail: [email protected]
'School of Information Technology, Murdoch University, Perth, WA 6 150, Australia.
E-mail: tgedeon @rnurdoch.edu.au
'Geological Survey of NSW, NSW Department of Mineral Resources, PO Box 65, Armidale, NSW 2350,
Australia. E-mail: [email protected]

Abstract geophysical signatures of mineral systems in

order to predict the location of new deposits.
This paper reports a pilot study on the use of an
artificial neural network in conjunction with a
Geographic Information System (GIS) for the 2. Previous Work
integration large multi-source data sets used in Mineral prospectivity maps represent one way
regional mineral exploration and prediction of of combining the geological, geophysical,
mineral prospectivity. A multilayer feedforward geochemical maps used in regional-scale
neural network, trained with the error- exploration. Areas are ranked according to
backpropagation algorithm, is used to estimate the their potential to host mineral deposits of a
favourability for gold deposits from a raster GIs particular type. An empirical or data-driven
database for the Tenterfield 1:100,000 sheet area, approach to the prospectivity mapping
NSW. To validate and assess the effectiveness of involves the statistical analysis of the spatial
the neural-network method, mineral prospectivity relationships of map features (e.g., faults,
maps were also prepared using the empirical lithological units) to mineral deposit locations.
weights of evidence and the conceptual fuzzy logic The range of techniques used includes logistic
methods. The neural network produces a regression [ 11, Boolean algebra, the index-
geologically-plausible mineral prospectivity map overlay method [2], and weights of evidence
similar, but superior, to the fuzzy logic and weights [3]. Knowledge-driven or conceptual methods
of evidence maps. The results of this study feature in later work and employ a range of
indicate that neural networks have several artificial-intelligence computing techniques,
advantages over existing methods. such as fuzzy logic [4], evidential-belief theory
[5], and expert systems [6].
1. Introduction
Although neural networks have been applied to
The development of faster, more efficient and petroleum exploration [7,8], there are very few
objective Geographic Information System (GIs)- studies that describe the application of artificial
based methods for integration and analysis of neural networks to mineral exploration
regional exploration data sets has a potentially [9,10,11]. The use of neural networks for
important role in supporting the decision-making mineral prospectivity mapping has been
process for geologists in ground acquisition and described in only two previous studies [12,13].
selection of exploration targets.

Neural networks have the ability to extract 3. GIs Database

underlying patterns in a data set without pre- 3.1 GIs Thematic Layers
existing knowledge (e.g., a deposit model), and to
operate at acceptable accuracy with noisy data. The GIS database for the trial study
This suggests that they could be well suited to the corresponds to the area of the Tenterfield
integration of mineral exploration data and could 1:100,000 sheet (map 9339, Fig. l), and was
be trained to recognise the geological and compiled by the Geological Survey of New
South Wales [ 141. The database comprises the

0 1999 IEEE
0-7803-5871-6/99/$10.00 160
following thematic layers; solid geology, regional- most important hosts to gold mineralisation.
scale faults, airborne total magnetic intensity The mid-Triassic Stanthorpe Adamellite,
(classed image data), airborne gamma-ray appearing as a semi-circular body east of the
spectrometry (U, Th, K and total count, classed Demon Fault in the north of the study area,
image data) and deposit locations. The mineral hosts large, low-grade, primary, disseminated
deposit layer includes mineral occurrences, and and secondary alluvial deposits. Numerous
small and medium-sized deposits. Each cell in the alluvial gold occurrences are scattered across
grid data represents a 200 metre square on the the Bungulla Porphyritic Adamellite, which
ground. forms an arcuate body on the eastern margin of
the Stanthorpe Adamellite [141.
3.2 Geology
The area of the Tenterfield 1:100,000 sheet is Forty-seven secondary, thirteen primary, and
located in the southern portion of the New England five spatially-associated, granitoid-hosted gold
Orogen (Fig. 1). The highly fractionated Late deposits of unknown form were selected from
Permian to Early Triassic leucogranites are the the GIS deposit thematic layer for this study.

Figure 1. Location and simplified geology of the Tenterfield 1:100,000 sheet area adapted from an unpublished
regional digital map supplied by the Geological Survey of New South Wales.

4. Methods Pre-Processing
4.1 Data processing All thematic layers in the GIS database were
converted to raster format prior to further
A Multilayer PercePhon ( m p ) I’Ieural network, processing. The original solid geology layer
trained using the error-back-propagation algorithm 41 rock units, many of which
[15], was chosen for this study because it performs zero or very few deposits.
very well for a large variety Of problem types and Consequently, units belonging to fie Same
has powerful function-approximation capabilites rock suite or stratigraphic group were
1161. combined, resulting in a simplified geology
layer consisting of 12 rock units. The GIs
database layer containing regional faults was

161
converted to a grid in which the cell values validation data set was used to process the
indicate the distance to the nearest fault. input vectors for the whole Tenterfield grid.

At each cell location in the Tenterfield 1:100,00 Post-Processing

sheet grid, values from each thematic layer were Output values produced by the trained network
combined to form a feature vector. Linear scaling for the complete Tenterfield data set were re-
was applied to both input feature vectors and the scaled and classed to produce a multiclass
target values. prospectivity map.
Input Coding, Network Topology and Network 4.2 Comparing Mineral Prospectivity Map
Parameters Quality
A one-of-n coding scheme was applied to the solid Prospectivity maps depict areas ranked
geology layer so that each rock type was assigned according to their potential to host mineral
a separate input unit. A single input unit was deposits of a particular type and therefore do
assigned to the magnetic, four gamma-ray, and not represent a classification of some extensive
fault-proximity layers. Binary coding was used for category, such as agricultural crop type, that
the single output unit, with one and zero can be readily checked. Instead, the following
representing the presence and absence of deposits, statistics were used to compare the quality of
respectively. The number of units in the hidden different mineral prospectivity maps:
layer of the network was determined
experimentally, resulting in an 18-2-1 network conditional probabilities of the occurrence
topology. The learning rate and momentum were of known mineral deposits for each
set to 0.5 and 0, respectively. Weights were prospectivity map class were calculated
randomly initialised in the range [-0.5,+0.5] and using the numbers of grid cells, assuming
were updated incrementally. that the deposits occupy a single cell each.
Training Data Set Selection and Training a high prospectivity map class should
Training data were randomly divided into training, define an area in which a high proportion
validation and test sets. (There were an insufficient of the known deposits are predicted in a
number of deposit and occurrence points available small proportion of the total map area.
to allow a third of the data to be reserved as a test This can be expressed as the following
data set in the statistically-based weights of ratio:
evidence method.) Approximately equal numbers
of deposit and non-deposit vectors were selected
for each of the data sets. Since there are only 63
gold deposits of the required type in the Tenterfield
area compared with 69,377 non-deposit cells, the
size of the data sets used for training were limited where n(D,) represents the number of
by the number of deposits available. The training, deposits in map class A, n(D,o,u,)is the total
validation and test data sets were each randomly number of deposits, n(A) is the area of map
assigned one third of the available deposit vectors class A, n(T) is the total area, P(D/A) is the
(21 each), and two non-deposit vectors from each conditional probability of a deposit, given
of the 12 main rock units, giving a total of 24 non- a particular prospectivity map class, and
deposit vectors in each training set. The training P(D) is the probability of a deposit for the
patterns were presented to the network in a map area as a whole. In the highest group
randomly permutated order. of prospectivity classes, the probability of
finding a deposit should be significantly
Ten networks, using a different random set of upgraded and the ratio should be greater
initial weights, were trained. Training was stopped than one.
at the number of epochs corresponding to the
minimum TSS error for the validation data set. the Chi-square statistic is used as a
The neural network giving the lowest error for the measure of the extent to which observed
and expected numbers of deposits differ.

I62
 Applying similar reasoning as above, the Fuzzy Logic Method
observed number of deposits in the map
The degree to which the value of a GIS
classes with the highest favourability should be
database parameter belongs to the fuzzy set
much higher than the number that be
“favourable for mineralisation” is expressed as
expected if the distribution of deposits with
a real number in the interval [0,1]. Fuzzy
respect to map classes were random.
membership values were assigned to each
parameter value stored in the GIS thematic
4) Spearman’s and Kendall’s rank correlation layers according to subjective judgement.
coefficients are used to assess the degree to
Fuzzy membership values from different layers
which the conditional probability of were combined using a variety of different
occurrence of known mineral deposits
operators and fuzzy-logic inference nets. The
increases with increasing map-prospectivity
prospectivity map that best accounted for the
class.
known deposit points while minimising the
4.3 Preparation of weights of evidence and area of the highest prospectivity classes was
chosen as the final fuzzy logic map. The
fuzzy logic prospectivity maps
gamma fuzzy operator with a value of 0.95
As a basis for validating and comparing the gave the best results. The fuzzy gamma
performance of the neural network method, operator is given by;
mineral prospectivity maps were prepared using
both an empirical method, based on statistics and
Bayesian probability, known as weights of
evidence, and a conceptual method, based on
fuzzy-logic.
where y = [0,1] and pjis the fuzzy membership
Weights of Evidence Method value for the ith data set. Detailed descriptions
The weights of evidence method is based on the of fuzzy operators are given by Zimmermann
concept of prior and posterior probabilities and [17], An et al. [4], and Bonham-Carter [2].
employs a log-odds formulation of Bayes rule [3]. Processing for the weights of evidence and
The method results in some loss of information in fuzzy logic method was performed using the
the original data, since multiclass maps must be GEODIPS GIS prospectivity analysis package
converted to binary maps prior to analysis and 1181.
requires an assumption that parameters in the
binary evidence maps are conditionally 5. Results
independent with respect to deposits. A full
description of the weights of evidence method is Mineral prospectivity maps are shown together
given by Bonham-Carter [2]. with known gold occurrence and deposit points
in Figure 2.

a) b) c)
Figure 2. Prospectivity maps. a) Weights of evidence method. b) Fuzzy logic method. c) Neural network
method. Crosses, dots and open squares represent known primary, alluvial and undifferentiated gold deposits
and occurrences, respectively.

163
Both the neural network and fuzzy logic The lower part of Table 2 shows the ratio of
favourability results are divided into nine equal conditional probability to probability for the
classes. Due to the necessity of combining study area as a whole. The nine-class fuzzy
several input maps to avoid problems of logic and neural network maps were
conditional dependence, the weights of reclassified into four-class maps to enable a
evidence map contains only three discrete direct comparison with the weights of evidence
probability levels. Classifying these values map.
results in a four-class prospectivity map in
which one of the classes does not correspond Table 2. Probability and probability ratio measures
to any cells on the map. of prospectivity map quality.

The values of the map quality indices Map Class Method

presented in Table 1 are all higher for the Weights of Fuzzy Neural
neural network map than those for the other Evidence Logic Network
two maps.
P(D/A) x io4
Table 1. Statistical measures of prospectivity map
quality. 1 2.52 2.85 2.60
2 0.00 1.65 2.15
3 8.67 6.89 11.66
Statistic Method 4 23.45 18.91 26.74
Weights of Fuzzy Neural P(D/A)/P(D)
Evidence Logic Network
1 0.28 0.31 0.29
Chi-square 66 63 83 2 0.00 0.18 0.24
Spearman’s p 0.80 0.50 0.93 3 0.96 0.76 1.29
0.67 0.53 0.83 4 2.58 2.08 2.95
Kendall’s 2

The value of the Chi-square statistic is Conclusions

significantly higher for the neural network map A neural network approach, using a simple
than the weights of evidence and fuzzy logic
implementation of the most commonly-
maps, indicating greater differences between
used neural network architecture is
the observed numbers of deposits and the successful in integrating geological and
numbers expected due to chance. Both the geophysical data sets representative of
Spearman’s and Kendall’s rank correlation
those used in a regional exploration
coefficients show the degree to which the
program to produce a mineral prospectivity
conditional probability of occurrence of known map.
deposits, P(D/A), increases with increasing
prospectivity map class. These statistics are
The similarity of the neural network map
also significantly higher for the neural network to the empirical weights of evidence and
map than for the weights of evidence and fuzzy conceptual fuzzy logic map indicates that
logic maps. the neural network map is consistent with
both the spatial relationships in the data,
The probability and probability ratio measures
and with geological knowledge about
of map quality (Table 2) indicate that the
factors important for gold mineralisation in
highest prospectivity map classes in the neural
the Tenterfield area, respectively.
network map are stronger predictors of
deposits than in the fuzzy logic map and
The neural network method has several
slightly stronger than in the weights of
advantages over existing methods,
evidence map.
including the ability to: a) respond to
critical combinations of parameters, rather
The upper part of Table 2 shows the
than automatically increasing the
conditional probability that a cell contains a
prospectivity due to all favourable
known deposit, given that the area of interest is
parameters, b) combine data sets without
restricted to a particular prospectivity class.

I 64
 the loss of information inherent in the Katz, S . S., “Emulating the Prospector expert
weights of evidence method, c) overcome system with a raster GIS,” Computers and
Geosciences, vol. 17, no. 7, pp. 1033-1050, 1991.
problems related to conditional Osborne, D. A., “Neural networks provide more
independence in the weights of evidence accurate reservoir permeability,” Oil and Gas
method, and d) produce results that are Journal, vol. 90, no. 39, pp. 80-83, 1992.
relatively unaffected by redundant data, Wong, P. M., Taggart, I. J. & Gedeon, T. D., “Use
of neural network methods to predict porosity and
spurious data and data containing multiple permiability of a petroleum reservoir,” AI
populations. Applications, vol. 9, no. 2, pp. 27-37, 1995.
Singer, D. A. & Kouda, R., “Classification of
4) Statistical measures used to compare map mineral deposits into types using mineralogy with a
quality indicate that the neural network probabilistic neural network,” Nonrenewable
Resources, vol. 6, no. 1, pp. 27-32, 1997.
method performs as well, or better, than An, P., Chung, C. J. F. & Rencz, A. N., “A
existing methods, while using preliminary study of alteration mapping from
approximately a third less data than the airborne geophysical and remote sensing data using
weights of evidence method. feed-forward neural networks,” In: Society of
Exploration Geophysicists 64th Annual
(Approximately one third of the available International Meeting; Technical Program
data was reserved as a test data set in the Abstracts, vol. 64, pp 837-840, 1994.
neural network method and was not used Gong, P., “Integrated analysis of spatial data from
in training.) multiple sources: using evidential reasoning and
artificial neural network techniques for geological
mapping,” Photogrammetric Engineering and
7. Acknowledgments Remote Sensing, vol. 62, no. 5, pp. 513-523, 1996.
Singer, D. A. & Kouda, R., “Application of a
The authors gratefully acknowledge the feedforward neural network in the search for
Geological Survey of New South Wales, Kuroko deposits in the Hokuroku District, Japan,”
Department of Mineral Resources for Mathematical Geology, vol. 28, no. 8, pp. 1017-
1023,1996.
providing the GIS data used in this study. Clare, A., Bolt, G. & Daneel, G., ‘The use of neural
Weights of evidence and fuzzy logic mineral analysis as an exploration tool,” In: Proceedings of
prospectivity maps were produced using the the 12th International Conference on Applied
GEODIPS GIS software package written by Geologic Sensing, Environmental Research
Institute of Michigan (ERIM), Ann Arbor, vol 2,
Geoffrey Taylor and Jusmady of the pp. 199-205, 1997.
Department of Applied Geology, University of Barnes, R. G., Henley, H. F. & Henley, J. E.,
New South Wales. Joe Leach is thanked for “Exploration data package for the Tenterfield and
contributing the initial idea for this study. Coaldale 1:100,000 sheet areas,” Geological Survey
Report No. GS199Y004. Geological Survey of
New South Wales, 1995.
8. References [15] Rummelhart, D. E., Hinton, G. E. & Williams, R.
J., “Learning representations by back-propagating
Reddy, R. K. T., Agterberg, F. P. & Bonham- errors,” Nature, vol. 323, 9 October, pp. 533-536,
Carter, G. F., “Application of GIS-based logistic 1986.
models to base-metal potential mapping in Snow [ 161 Hornik, K., “Approximation capabilities of
Lake area,” In: Proceedings of Canadian multilayer feedforward networks,” Neural
Conference on GIs, Ottawa, pp. 607-618, 1991. Networks, vol. 4, no. 4, pp. 251-257, 1991.
Bonham-Carter, G. F., “Geographic Information [17] Zimmermann, H . J., “Fuzzy Set Theory and its
Systems for Geoscientists: Modelling with GIS,” applications,” Kluwer Academic Publishing,
Pergamon Press, New York, 1994. Boston, 1985.
Bonham-Carter, G. F., Agterberg, F. P. & Wright, [18] Jusmady & Taylor, G.R.T., “Geological Data
D. F., “Integration of geological datasets for gold Integration and Processing System (GEODIPS)
exploration in Nova Scotia,” Photogrammetric software package, Version 1.03,” Department of
Engineering and Remote Sensing, vol. 54, no. 11, Applied Geology, University of New South Wales,
pp. 1585-1592, 1988. 1997.
An, P., Moon, W. M. & Rencz, A., “Application of
fuzzy set theory to integrated mineral exploration,”
Canadian Journal of Exploration Geophysics, vol.
27, no. 1, pp. 1-11, 1991.
Moon, W, M., “Integration of geophysical and
geological data using evidential belief function,”
IEEE Transactions, Geoscience and Remote
Sensing, vol. 28, no. 4, pp. 711-720, 1990.

165

View publication stats