Sei sulla pagina 1di 14

C 2006)

Natural Resources Research, Vol. 15, No. 1, March 2006 (


DOI: 10.1007/s11053-006-9012-7

A Hybrid Fuzzy Weights-of-Evidence Model for Mineral


Potential Mapping
Alok Porwal,1,2 Emmanuel John M. Carranza,1,4 and Martin Hale1,3
Received 30 June 2005; accepted 9 August 2005
Published online: 4 November 2006

This paper describes a hybrid fuzzy weights-of-evidence (WofE) model for mineral potential
mapping that generates fuzzy predictor patterns based on (a) knowledge-based fuzzy membership values and (b) data-based conditional probabilities. The fuzzy membership values are
calculated using a knowledge-driven logistic membership function, which provides a framework for treating systemic uncertainty and also facilitates the use of multiclass predictor maps
in the modeling procedure. The fuzzy predictor patterns are combined using Bayes rule in
a log-linear form (under an assumption of conditional independence) to update the prior
probability of target deposit-type occurrence in every unique combination of predictor patterns. The hybrid fuzzy WofE model is applied to a regional-scale mapping of base-metal
deposit potential in the south-central part of the Aravalli metallogenic province (western
India). The output map of fuzzy posterior probabilities of base-metal deposit occurrence is
classified subsequently to delineate zones with high-favorability, moderate favorability, and
low-favorability for occurrence of base-metal deposits. An analysis of the favorability map
indicates (a) significant improvement of probability of base-metal deposit occurrence in the
high-favorability and moderate-favorability zones and (b) significant deterioration of probability of base-metal deposit occurrence in the low-favorability zones. The results demonstrate
usefulness of the hybrid fuzzy WofE model in representation and in integration of evidential
features to map relative potential for mineral deposit occurrence.
KEY WORDS: Fuzzy membership, conditional probability, base-metal deposits, GIS, Aravalli province.

INTRODUCTION

abilities of B, given the presence or absence of D


(Agterberg, Bonham-Carter, and Wright, 1990;
Bonham-Carter and Agterberg, 1990):

In the data-driven weights-of-evidence (WofE)


approach to mineral potential mapping, the spatial association of a predictor pattern B with target mineral deposits D is quantified in terms of a
pair of weights for that predictor pattern. The pair
of weights is calculated from the conditional prob-

P [ B| D]
W+ = loge P [ B | D]

and
|D]
P[B

W = loge P [ B | D ] .

From the preceding equations, it follows that the pair


of weights of a predictor pattern is determinable,
if, and only if, it contains at least one training deposit (i.e., loge of zero is indeterminable). Moreover, robustness of conditional probability of a predictor pattern estimated using the WofE method
decreases with a decrease in number of training deposits contained by the predictor pattern (Agterberg,

1 International

Institute for Geo-information Science and Earth


Observation (ITC), Enschede, The Netherlands.
2 Department of Mines and Geology, Govt. of Rajasthan,
Udaipur, India.
3 Faculty of Geoscience, Utrecht University, Utrecht, The
Netherlands.
4 To whom correspondence should be addressed at ITC,
Hengelosestraat 99, 7500 AA Enschede, The Netherlands;
e-mail: carranza@itc.nl.

1
C 2006 International Association for Mathematical Geology
1520-7439/06/0300-0001/0 

2
Bonham-Carter and Wright, 1990). The WofE modeling approach therefore, may be implemented using
binary predictor patterns, instead of multiclass predictor patterns, so that a predictor pattern is large
enough to contain a large proportion of training deposits. However, reclassification of spatial data into a
binary predictor map potentially results in some loss
of information in the predictor maps.
In fuzzy modeling approaches (for example,
An, Moon, and Rencz, 1991; Gettings and Bultman,
1993; Porwal and Sides, 2000; Knox-Robinson, 2000;
Carranza and Hale, 2001; Porwal, Carranza, and
Hale, 2003b), the spatial association of a predictor
pattern with a type of target mineral deposits is represented as a fuzzy membership value, which is estimated heuristically or algorithmically, based either
on expert knowledge or on empirical data, using appropriate fuzzy membership functions. Fuzzy modeling approaches, therefore, are implemented usually
using multiclass predictor maps.
In WofE modeling with multiclass predictor
maps, Cheng and Agterberg (1999) proposed a
fuzzy WofE approach that generalizes the ordinary
WofE approach by incorporating fuzzy mathematics.
The method involves (a) creating a fuzzy set of
favorable indicators of the target deposit-type for
each predictor map, (b) defining a knowledge-based
or data-based fuzzy membership function to calculate a fuzzy membership value for each pattern
in a predictor map represented as a fuzzy set, (c)
calculating a fuzzy WofE for each pattern in a predictor map, (d) combining the fuzzy WofE to
calculate, for each unique condition (Kemp,
Bonham-Carter, and Raines, 1999), a fuzzy posterior probability and the variance of the fuzzy
posterior probability because of missing patterns, to
miss-assigned patterns and to the fuzzy membership
functions, and (e) mapping of fuzzy posterior probabilities. The fuzzy posterior probability map can
be reclassified subsequently to create a binary or a
multiclass favorability map.
The fuzzy WofE approach (Cheng and
Agterberg, 1999) can be either (a) purely datadriven,
if a data-based function is used for calculating fuzzy
membership values or (b) hybrid knowledge-datadriven, if a knowledge-based function is used for
calculating fuzzy membership values. Cheng and
Agterberg (1999) demonstrated the fuzzy WofE
approach via a purely data-driven form to map gold
deposit potential in the Meguma Terrane, Nova
Scotia, Canada. In this paper, we explain and then
demonstrate a hybrid fuzzy WofE approach to

Porwal, Carranza, and Hale


regional-scale mapping of base-metal deposit potential in the Aravalli metallogenic province, western
India. We used a knowledge-based logistic membership function (Porwal, Carranza, and Hale, 2003b)
for calculating fuzzy membership values, from which
we derive a fuzzy WofE model of mineral potential.
The hybrid fuzzy WofE approach we demonstrate
here is different from the data-driven fuzzy modeling
approach we demonstrated previously (Porwal,
Carranza, and Hale, 2003b), in that the latter made
use of estimated multiclass spatial contrast from
estimates of a pair of weights in order to estimate
fuzzy favorability. We implemented our hybrid fuzzy
WofE modeling in a GIS environment using Arc
View 3.2, Spatial Analyst 1.1 and Arc-SDM (Kemp
and others, 2001).
HYBRID FUZZY WEIGHTSOF-EVIDENCE MODEL
If Xi (i = 1 to n) is a set of multi-class predictor maps of potential for the occurrence of a target deposit-type, with each predictor map contain i (i = 1
ing xij (j = 1 to r) patterns, then fuzzy sets A
to n), each representing Xi , can be defined as follows
(Zimmerman, 1991 ):

Ai = {(xij , A )|xj Xi }

(1)

where A is the fuzzy favorability membership func i.


tion of Xi predictor map represented by fuzzy set A
In our hybrid fuzzy WofE modeling approach,
we used the following knowledge-based logistic
membership function to estimate fuzzy favorability membership values (Porwal, Carranza, and Hale,
2003b):

Axij =

1
,
1 + ea(xij b)

(2)

where b is the inflexion point, a is the slope of the


function and xij , (the class score of the jth pattern in
the ith predictor map) is calculated as:
xij = wj wi ,

(3)

where wj is the class weight of the jth pattern in the


ith predictor map and wi is the map weight of the ith
predictor map.
The map weight and the class weight are derived
as follows. Based on their subjectively of assessed favorability for indicating potential for the occurrence
of a target deposit- type, all patterns in a predictor
map are ranked in the scale of 1 to 10 in a reverse direction; That is, the most favorable pattern is ranked

A Hybrid Fuzzy Weights-of-Evidence Model for Mineral Potential Mapping


10, whereas the least favorable pattern is ranked 1. In
this scheme, the rank of a pattern is its class weight.
As much as possible, the patterns are ranked at equal
intervals. This method of allotting weights by ranking is simple, as most experts are likely to agree upon
the favorability rank of a predictor pattern but are
likely to disagree upon its favorability weight. Similarly, based on their importance with respect to pertinent recognition criteria, all predictor maps are assigned map weights according to the same procedure
for assigning class weights as described above. However, in the situation of assigning map weights, predictor maps are ranked in the scale of 5 to 10, instead
of 1 to 10. This is because we consider a rank below
5 to represent low or insignificant importance and a
map would not have been considered a predictor if
it is not important in indicating potential for mineral
occurrence.
Using the fuzzified predictor maps, we derive a
fuzzy WofE model as follows. Fuzzy loge posterior
odds of D, given Xi (i = 1 to n) predictor maps with
multi-class xij patterns, each predictor being repre i (i = 1 to n) fuzzy sets, can be estimated
sented by A
n 
r
as:

W
(4)
loge O[D|Xi ] = loge O[D] +
Aij ,
i=1 j =1

W
Ai (xij ) =



Ai (xij )P[Ai1 |D] + {1 
Ai (xij )}P[Ai2 |D]
loge
,


Ai (xij )P[Ai1 |D] + {1 
Ai (xij )}P[Ai2 |D]
(5)
where P[Ai1 |D] and P[Ai2 |D] are, respectively, the
conditional probabilities of two crisp sets Ai1 and
and
Ai2 , given the presence of D, whereas P[Ai1 |D]

P[Ai2 |D] are, respectively, the conditional probabilities of the two crisp sets Ai1 and Ai2 , given the absence of D. The two crisp sets Ai1 and Ai2 (i = 1 to
i (i = 1 to n) are defined here as
n) in the fuzzy set A
follows:
Ai1 = {xij |
Aij = MAX (
Aij )} and
(6)

Fuzzy posterior probability can be calculated from


Equation (4):
P[D|Xi ] =

eloge O[D|Xi ]
.
1 + eloge O[D|Xi ]

The following equations for estimating variance of


the fuzzy posterior probabilities because of missing
patterns, miss-assigned patterns, and fuzzy membership values are adapted from Cheng and Agterberg
(1999). The variance of the fuzzy posterior probabilities because of Xk missing patterns is estimated as:
k2 (P[D|Xi ]) = {P[D|Xk] P[D]}2 P[Xk]
+ {P[D|Xk] P[D]}2 P[Xk]. (8)
In Xl predictor map, the variance of the fuzzy posterior probabilities due to miss-assigned patterns from
Xl to Xl or vice versa is estimated as:
2
X
(P[D|Xi ]) = {P[D|Xk] P[D|Xk]}2 P[D],
l

and
2

X
(P[D|Xi ]) = {P[D|Xk] P[D|Xk]}2 P[D].

(9)

The variance of the fuzzy posterior probabilities due


to the fuzzy membership values, A i , is estimated
as:
2 
(P[D|Xi ]) =
Ai

2
Ai (1 
Ai )
p[Xi ]P[Xi ]
p[
]
Ai


2
2
(P[D|X
])
+

(P[D|X
])
.
X
i
i
Xl
l
(10)

where W
Aij is the fuzzy WofE of xij pattern in
Xi predictor map and is estimated as (Cheng and
Agterberg, 1999 ):

Ai2 = {xij |
Aij = MIN (
Aij )}.

(7)

APPLICATION TO BASE-METAL
DEPOSIT POTENTIAL MAPPING IN
ARAVALLI PROVINCE, WESTERN INDIA
The study area forms a part of the Aravalli metallogenic province in the state of Rajasthan, western
India (Fig. 1). It measures about 34,000 km2 and is
located between latitudes 23 30 N and 26 N and between longitudes 73 E and 75 E.
The geology of the province is characterized by two major fold belts, viz., the PalaeoMesoproterozoic Aravalli Fold Belt and the MesoNeoproterozoic Delhi Fold Belt, which are ingrained
in a reworked basement complex named the Banded
Gneissic Complex that contains incontrovertible
Archaean components. This three-fold tectonostratigraphic classification of the province, which was proposed first by Heron (1953), remains the basic framework of reference for all subsequent studies (Raja
Rao and others, 1971; Roy, 1988; Roy and others,
1993; Gupta and others, 1997). In an attempt to explain the evolution of the province in the framework
of Proterozoic Wilson cycles and plate tectonics,
Deb and Sarkar (1990) and Sugden, Deb, and

Porwal, Carranza, and Hale

Figure 1. Location map of study area in state of Rajasthan, India. Small black circles are locations of base-metal deposits.

Windley (1990) divided the province into a number of tectonic domains based on lithogeochemical
and structural considerations. Our work on geophysical data also indicates that the province is comprised of subparallel linearly disposed belts (Fig.
2), each having distinct geophysical and tectonic
characteristics, coinciding broadly with the tectonic
domains proposed by Sugden, Deb, and Windley
(1990).
The Aravalli province holds substantial reserves
of base-metal sulfide deposits, particularly, lead and
zinc. The economically viable lead-zinc reserves in
the province stand at 130 million tonnes with 2.2%
Pb and 9.2% Zn with an additional 30 million tonnes
of possible resources in producing mines and deposits under detailed exploration (Haldar, 2001).
These form the entire economically viable lead-zinc
resource-base of the country.
A vast majority of the lead-zinc sulfide deposits
of the province are contained in the study area, which
is considered one of the prime exploration target
areas for base-metals in the country. The deposits

in the area are hosted by the supracrustal rocks


of the Bhilwara, Aravalli, and South Delhi belts
(Fig. 2). Major concentrations of mineralization in
the Bhilwara belt occur in the Rampura-Agucha deposit and in the Pur-Banera and Bethumni-DaribaBhinder mineralized zones. Rampura-Agucha is a
world-class Zn-Pb-(Ag) deposit with the highest
combined metal grade (about 15%) of all base-metal
deposits in India. In the Bethumni-Dariba-Bhinder
zone, ZnPb(Cu) deposits are located in a 17 kmlong belt extending from Bethumni in the north to
Dariba in the south, with a lean pyrite zone farther south nearby Bhinder. The Pur-Banera is a
lean polymetallic zone with several small deposits.
In the Aravalli belt, low-grade Cu, Au, and U mineralizations occur in basal sequences, whereas Zn
Pb deposits occur in upper sequences of the Zawar zone. The South Delhi belt hosts small deposits
of Cu(Zn) in the Basantgarh area. The details of
base-metal metallogenesis and nature of mineralization of occur in Porwal, Carranza, and Hale (2003a,
2003b).

A Hybrid Fuzzy Weights-of-Evidence Model for Mineral Potential Mapping

Figure 2. Generalized geological map of south-central Aravalli province showing study area for fuzzy WofE modeling
(demarcated in heavy black) and important mineralized zones (outlined in white). White circles are locations of basemetal deposits.

Recognition Criteria and Predictor Maps


Porwal, Carranza, and Hale (2003a, 2003b) identified the following regional-scale recognition criteria for base-metal deposit occurrence in the province:
(1) host rock lithology, (2) stratigraphic position, (3)
(paleo-)sedimentary environment, (4) association of
synsedimentary mafic volcanic rocks, (5) proximity

to regional tectonic discontinuities, and (6) proximity


to favorable structures. Each recognition criterion is
represented by at least one predictor map prepared
from available spatial data sets.
A regional-scale GIS was compiled by digitizing the lithostratigraphic map (Gupta and others,
1995a), the structural map (Gupta and others, 1995b)
and the map of total magnetic field intensity (GSI,

Porwal, Carranza, and Hale

1981). The locations of 54 known base-metal deposits were compiled from various sources. The digitized maps in a vector format were converted into
grid format for subsequent operations. Based on the
resolution of the original maps used for creating
the GIS database, a grid cell size of 250 m 250 m
was considered appropriate for representing the
smallest feature in the original analogue maps. The
grid maps were processed, interpreted, and reclassified, as described by Porwal, Carranza, and Hale
(2003b), to create predictor maps. Of the eight
predictor maps we prepared, seven are multiclass
(lithologies, stratigraphic groups, sedimentary environments, buffered regional lineaments, buffered
NE-trending lineaments, buffered NW-trending lineaments, and buffered fold axes) and one is binary
(mafic igneous rocks).
Pairwise Test of Conditional Independence. All
pairs of predictor maps were tested for conditional
independence (CI) with respect to the locations
of base-metal deposits using the 2 test (BonhamCarter and Agterberg, 1990). The test, although not
accurate in determining CI, because it is gives low estimates of correlations (see Singer and Kouda, 1999,
for a detailed discussion), can be used for identifying pairs of predictor maps that lack CI with respect
to target variable. The results indicate that predictor map of sedimentary environments and predictor
map of lithologies significantly lack CI with respect
to the locations of base-metal deposits (Table 1). The
predictor map of sedimentary environments was not

considered further in the hybrid fuzzy WofE modeling because the sedimentary environments were inferred mainly from available geoscience information,
whereas the lithologies were mapped in the field. After combining the predictor maps, we also performed
an overall test of CI (see next).

Estimation of Fuzzy WofE of Predictor Patterns


For each pattern in a predictor map, estimation of fuzzy WofE via Equation (5) requires estimates of (a) fuzzy membership values and (b) conditional probabilities (given the presence or absence of
a base-metal deposit) of map patterns with the highest and the lowest fuzzy membership values [Eq. (6)].
Fuzzy membership values were estimated
using the logistic membership function defined in
Equation (2). The values used for the variables b
and a in Equation (2) were 50 and 0.1, respectively,
which yield a curve that is symmetrical about the
inflexion point at (50, 0.5), where 50 is the class score
and 0.5 is the corresponding knowledge-driven fuzzy
membership value. The class scores were calculated
from class weights and map weights [Equation (3)],
which were assigned subjectively using the ranking procedure described earlier. The assignment
of subjective ranks to map patterns was based
on our experience of base-metal exploration in the
province, informal discussions with mining geologists
of M/s Hindustan Zinc Ltd. (a Government of India

Table 1. Pair-Wise 2 Test for Conditional Independence among Input Predictor Maps

Predictor maps
Stratigraphy
Lithologies
Sedimentary environments
Mafic igneous rocks
Buffered regional lineaments
Buffered NW lineaments
Buffered NE lineaments

Sedimentary Mafic Igneous


Lithologies environments
rocks
49.02 (56)
74.468

57.66 (48)
65.171
68.42 (42)
58.124

6.36 (8)
15.507
3.01 (7)
14.067
4.96 (6)
12.529

Buffered
regional
lineaments
0.74 (8)
15.507
0.85 (7)
14.067
0.11 (6)
12.529
0.46 (1)
3.84

Buffered NW
lineaments

Buffered NE
lineaments

Buffered
fold axes

0.96 (8)
15.507
6.44 (7)
14.067
2.53 (6)
12.529
0.01 (1)
3.84
0.002 (1)
3.84

0.89 (8)
15.507
7.28 (7)
14.067
7.09 (6)
12.529
0.07 (1)
3.84
0.03 (1)
3.84
1.53 (1)
3.84

1.91 (8)
15.507
1.19 (7)
14.067
1.49 (6)
12.529
0.03 (1)
3.84
0.38 (1)
3.84
3.83 (1)
3.84
0.32 (1)
3.84

Note. Values in bold and italics are calculated and tabulated 2 values (at 0.05% significance level), respectively. Figures in parentheses
are degrees of freedom ( = (v 1)(u 1), where v and u are number of classes in two predictor maps).
Null hypothesis of conditional independence is rejected at 95% confidence level.

A Hybrid Fuzzy Weights-of-Evidence Model for Mineral Potential Mapping


enterprise, which owns most of the large basemetal
deposits in the province), and recommendations of
various authors, especially Deb (1999) and Sarkar
(2000). The detailed rationale for ranking of the
predictor maps and individual map patterns are in
Porwal, Carranza, and Hale (2003b).
Conditional probabilities of individual patterns
in a predictor map, given the presence or absence of
a base-metal deposit, were estimated by using a randomly selected set of 23 training points (or known
deposits) out of the 54 base-metal deposits. The estimations were performed using the gridded predictor maps and taking 1 km2 as the unit cell size.
Thus, the estimated prior probability is (23/34000 )
0.0007. The remaining 31 base-metal deposits were
later used as validation points (or unknown deposits). The class weights, map weights, class scores,
knowledge-driven fuzzy membership values and conditional probabilities of the individual patterns in
each of the predictor maps are given in Table 2.
Based on the estimated fuzzy membership values and
conditional probabilities, Equations (5) and (6) were
used to estimate fuzzy WofE of every pattern in each
predictor map (Table 2).

Combining Predictor Maps


The predictor maps were combined by generating a unique conditions map (Bonham-Carter
and Agterberg, 1990; Kemp, Bonham-Carter, and
Raines, 1999). A unique condition is formed from the
unique combination of at least two patterns in different predictor maps (Kemp, Bonham-Carter, and
Raines, 1999). The attribute table (unique conditions
table) associated with a unique conditions grid has
one record (or row) per unique condition and one
or more fields (or columns) containing attribute data
(variables).
The multiclass predictor maps were combined
using digital overlay, which resulted in a map with
20258 unique conditions. The fuzzy posterior probability for each unique condition was derived from
estimates of the fuzzy WofE of predictor patterns
and the prior probability of base-metal deposits
[Eqs. (4) and (7)]. Figure 3 shows the spatial distribution of fuzzy posterior probabilities of basemetal occurrence in the study area. The variance of
the fuzzy posterior probabilities due to missing patterns, to miss-assigned patterns, and to fuzzy membership values were calculated using Equations (8),
(9), and (10), respectively. The spatial distribution

of total variance of the fuzzy posterior probabilities


are because of missing patterns and miss-assigned
patterns is shown in Figure 4A. The spatial distribution of variance of the fuzzy posterior probabilities because of fuzzy membership values is shown in
Figure 4B.
Overall Test of Conditional Independence. We
applied the new omnibus test (NOT) (Agterberg and
Cheng, 2002; Thiart, Bonham-Carter, and Agterberg,
2003) to determine violation or nonviolation of CI assumption in the final fuzzy posterior probability map.
The NOT statistic is ratio of the difference between
the predicted number of (training) deposits and the
observed number of (training) deposits divided by
the standard deviation of predicted number of (training) deposits. Values of NOT are assumed to approximate a standard normal distribution such that
probability that the difference between the predicted
number of (training) deposits and the observed number of (training) deposits is statistically greater than
zero can be estimated (from statistical tables) in order to judge whether to accept or reject the null hypothesis of CI.
We used total estimated variance of the fuzzy
posterior probabilities [i.e., sum of estimated variance using Equations (8), (9), and (10)] to estimate
standard deviation of the predicted number of (training) deposits. The NOT test showed that the final
fuzzy posterior probability map significantly violates
CI assumption. However, as suggested by Pan and
Harris (2000), violation of CI assumption can be ignored if, and only if, posterior probabilities are interpreted in a relative sense only and not in absolute sense (see further in the discussion section).
Therefore, we interpreted estimates of fuzzy posterior probabilities in a relative sense only in order to
classify zones in the study area with high-favorability,
moderate-favorability, and low-favorability for basemetal deposit occurrence.

Reclassification of Fuzzy Posterior Probability Map


It is cumbersome to interpret the fuzzy posterior probability map (Fig. 3) for selecting target
areas for base-metal exploration, because it shows
the fuzzy posterior probability of base-metal deposits in a continuous gray-scale from the least favorable (fuzzy posterior probability 0) to the most
favorable (fuzzy posterior probability 0.5). Moreover, a fuzzy posterior probability cannot be interpreted in the absolute sense of probability per

Porwal, Carranza, and Hale


Table 2. Map Weights, Class Weights, Fuzzy Membership Values, Conditional Probabilities, and Fuzzy WofE of Patterns
in Predictor Maps
Predictor Map/Pattern

Map
weight

Predictor map of lithotogies


Dolomite/dolomitic-marble
10
Calc-silicates
10
Graphitic meta-pelites
10
Magnetite quartzite
10
Meta-basites
10
Calc-schist/gneiss
10
Qzite-Arkose-Conglomerate
10
Migmatites; gneisses
10
Unrelated to base-metals
10
Predictor map of stratigraphies
Rajpura-Dariba group
9
Pur-Banera group
9
Debari group
9
Nathdwara group
9
Phulad Ophiolites group
9
Udaipur group
9
Jharol group
9
Sandmata Complex
9
Mangalwar Complex
9
Unrelated to base-metals
9
Predictor map of mafic Igneous rocks
Basic metavolcanic rocks
8
Unrelated to base-metals
8
Predictor map buffered regional magnetic lineaments
02 Km
8
24 Km
8
46 Km
8
68 Km
8
810 Km
8
>10 Km
8
Predictor map of buffered fold axes
00.5 Km
7
0.51 Km
7
11.5 Km
7
1.52 Km
7
22.5 Km
7
2.53 Km
7
33.5 Km
7
3.54 Km
7
44.5 Km
7
4.55 Km
7
>5 Km
7
Predictor map of NW-trending lineaments
01.5 Km
6
1.53 Km
6
34.5 Km
6
4.56 Km
6
67.5 Km
6
>7.5 Km
6
Predictor map of NE-trending lineaments
01.5 Km
6
1.53 Km
6
34.5 Km
6
4.56 Km
6
67.5 Km
6
>7.5 Km
6

Class
weight

Class
score

10
9
8
7
6
5
4
2
1

100
90
80
70
60
50
40
20
10

10
9
8
7
6
5
4
3
2
1

Fuzzy
membership

P[Xi |D]

P[Xi |D]

Fuzzy
WofE

0.99
0.98
0.95
0.88
0.73
0.5
0.27
0.05
0.02

0.1739
0.3043
0.3043
0.2174
0.0000
0.0000
0.0000
0.0000
0.0000

0.022
0.0118
0.0096
0.0002
0.0213
0.0467
0.0839
0.2092
0.5935

1.8265
1.6288
1.1837
0.5244
0.3280
1.2640
2.2360
4.1740
5.1200

90
81
72
63
54
45
36
27
18
9

0.98
0.96
0.09
0.79
0.60
0.38
0.20
0.09
0.04
0.02

0.2609
0.5217
0.0870
0.0000
0.0000
0.0870
0.0000
0.0435
0.0000
0.0000

0.0044
0.0117
0.0545
0.0036
0.0083
0.0961
0.1528
0.1005
0.1925
0.3738

3.0684
2.5593
1.7249
0.9093
0.0153
0.8700
1.7620
2.6880
3.5510
4.2650

10
1

80
8

0.95
0.01

0.7391
0.2609

0.1265
0.8731

1.4737
1.1810

10
8
6
4
2
1

80
64
48
32
16
8

0.9526
0.8022
0.4502
0.1419
0.0323
0.0148

0.2609
0.4348
0.2174
0.0435
0.0000
0.0435

0.1897
0.1633
0.1336
0.1066
0.083
0.3235

0.2455
0.0080
0.6220
1.4100
1.8430
1.9290

10
9
8
7
6
5
4
3
2
1
1

70
63
56
49
42
35
28
21
14
7
7

0.8808
0.7858
0.6457
0.4750
0.3100
0.1824
0.0998
0.0522
0.0266
0.0134
0.0134

0.6957
0.1739
0.0435
0.0000
0.0000
0.0000
0.0000
0.0000
0.0000
0.0000
0.087

0.1128
0.1026
0.0908
0.0787
0.0699
0.061
0.0523
0.0466
0.0415
0.0385
0.3051

1.7741
1.7303
1.6476
1.4992
1.2542
0.8911
0.4135
0.1540
0.7830
1.4440
1.4440

10
8
6
4
2
1

60
48
36
24
12
6

0.7311
0.4502
0.1978
0.0691
0.0219
0.0121

0.4783
0.2609
0.1739
0.0000
0.0435
0.0435

0.2430
0.1822
0.1169
0.0793
0.0574
0.3209

0.3142
0.1780
0.8580
1.4560
1.7950
1.8810

10
8
6
4
2
1

60
48
36
24
12
6

0.7311
0.4502
0.1978
0.0691
0.0219
0.0121

0.3913
0.2609
0.1304
0.087
0.0435
0.0870

0.1890
0.1678
0.1360
0.1056
0.0841
0.3171

0.3257
0.1470
0.6840
1.0490
1.2110
1.2470

A Hybrid Fuzzy Weights-of-Evidence Model for Mineral Potential Mapping

between high-favorability and moderate-favorability


zones, whereas the fuzzy posterior probability value
corresponding to the upper inflexion point (0.0012)
was used as a threshold value to distinguish between
moderate-favorability and low-favorability zones.
The threshold fuzzy posterior probability value of
0.0046 means that fuzzy posterior probabilities in
high-favorability zones range from about 6.6 to at
least 700 times greater than the estimated prior probability (0.0007). The threshold fuzzy posterior probability value of 0.0012 means that fuzzy posterior
probabilities in low-favorability zones range from
about zero to at most 1.7 times greater than the estimated prior probability. Fuzzy posterior probabilities in moderate-favorability zones, therefore, range
from about 1.7 to about 6.6 times greater than the estimated prior probability.

Validation of Favorability Map

Figure 3. Continuous grey-scale map of fuzzy posterior probabilities, which range from zero (white) to 0.484 (black).

se (see discussion section next; Singer and Kouda,


1999; Pan and Harris, 2000) in terms of making
decisions regarding selection of exploration targets.
We therefore reclassified the fuzzy posterior probability map (Fig. 3) into a ternary map of relative favorability for base-metal deposit occurrence
(Fig. 5). Threshold fuzzy posterior probability values to differentiate between high-favorability and
moderate-favorability zones and between moderatefavorability and low-favorability zones were determined using the graphical procedure described by
Porwal, Carranza, and Hale (2003a). The cumulative fuzzy posterior probability values (rearranged
in a descending order) were plotted against the percentage of cumulative area (Fig. 6). Two inflection
points were identified along the curve at which the
slope changes from steep to moderate and from moderate to almost flat. The fuzzy posterior probability value corresponding to the lower inflexion point
(0.0046) was used as a threshold value to distinguish

The favorability map was validated by overlaying the 23 deposit training points and 31 deposit
validation points on the favorability map (Fig. 5).
Table 3 shows the distribution of deposit validation points and deposit training points in the favorability map. In the favorability map, (a) the highfavorability zones occupy about 6% of the study area
and contain about 74% and about 96%, respectively,
of the deposit validation points and deposit training
points, (b) the moderate-favorability zones occupy
about 4% of the study area and contain about 13% of
the deposit validation points and 13% of the deposit
training points, and (c) the low-favorability zones occupy about 90% of the study area and contain about
13% and about 4%, respectively, of the deposit validation points and deposit training points.
Brown and others (2000) use the following probability ratio for expressing the quality of a favorability map:
n(DA)/n(Dtotal )
p(D|A)
=
,
p(D)
n(A)/n(T)

(11)

where n(DA ) is the number of deposits in favorability


class A, n(Dtotal ) is the total number of deposits, n(A)
is the area of favorability class A, n(T) is the total
area, p(D|A) is the posterior probability of a deposit
given favorability class A and p(D) is the prior probability of a deposit. Probability ratios higher than 1
indicate increase of probability of mineral deposit
occurrence, whereas probability ratios less than 1
indicate decrease of probability of mineral deposit

10

Porwal, Carranza, and Hale

Figure 4. Continuous gray-scale maps of variances of fuzzy posterior probabilities due (A) to missing patterns and to
miss-assigned patterns [variance ranges from zero (white) to 0.2148 (black)] and (B) to fuzzy membership values [variance
ranges from zero (white) to 0.0001 (black)].

occurrence. Probability ratios close to 1 indicate no


improvement of probability of mineral deposit
occurrence. An efficient predictive model should
delineate favorability zones with probability ratios
higher or lower than 1.
To validate the favorability map (Fig. 5) using
Equation (11), prior probability (n(Dtotal )/n(T)) is
re-estimated using all known base-metal deposit occurrences (i.e., n(Dtotal ) = 54), posterior probability
(n(DA )/n(A)) is re-calculated for each favorability
zone (A), and then the probability ratio in each favorability zone is calculated. The reason for using
all the 54 known base-metal deposit occurrences in
Equation (11) is to determine whether the favorability map, derived by using 23 training deposits, is
useful in guiding exploration towards zones with potential for undiscovered (i.e., validation) deposits.
The results of validation (Table 3) show that the (re)estimated prior probability (0.0016) of base-metal
deposit occurrence (a) increases to a posterior probability of 0.0226 in the high-favorability zones, (b)

increases to a posterior probability of 0.0027 in the


moderate-favorability zones, and (c) decreases to a
posterior probability of 0.0002 in the low-favorability
zones. Consequently, there is a distinct contrast in
magnitude of the probability ratios in the different
favorability zones. These imply that the favorability
map (Fig. 3) would be useful in guiding further basemetal exploration in the province.

DISCUSSION
The knowledge-based logistic function used in
the hybrid fuzzy WofE model provides a framework
for dealing with systemic uncertainties in a flexible
and consistent way. The function uses map weights
and class weights to derive fuzzy membership values of predictor patterns [Eqs. (2) and (3)]. The map
weight, which is similar in concept to the confidence
value of Knox-Robinson (2000), is assigned on the
basis of (a) the fidelity and precision of a predictor

A Hybrid Fuzzy Weights-of-Evidence Model for Mineral Potential Mapping

11

Figure 6. Plot of cumulative fuzzy posterior probability versus cumulative percent of study area.

Figure 5. Favorability map generated by reclassification of fuzzy


posterior probability map shown in Figure 3. Triangles are training base-metal deposits and circles are validation basemetal deposits.

map and (b) the relative importance of the recognition criteria represented by a predictor map. The
map weight therefore provides a method for treating systemic uncertainty, which generally arises from
(a) imprecision in mapping of predictor patterns, (b)
involvement of heuristics in generation of one or
more predictor patterns (for example, several predic-

tor patterns in the present study were based on interpretations of total magnetic field intensity data), and
(c) an unknown contribution of different genetic factors, and, hence, of predictor patterns which represent them, in spatial localization of mineral deposits.
The S-shaped logistic membership function
transforms linearly distributed class scores to logarithmically distributed fuzzy membership values,
so the differences in fuzzy membership values are
larger in the central part of the curve than along
its tails, as discussed by Porwal, Carranza, and Hale
(2003b). The function therefore separates unfavorable patterns from favorable patterns with less uncertainty, although among favorable (and unfavorable)

Table 3. Validation of Favorability Map


Validation based on size of study area, training deposits,
and validation deposits

Validation based on Equation (11) and all 54


known deposits

Zone

% of study area
34,000 km2 )

no. (and %) of
training deposits

no. (and %) of
validation deposits

Prior
probability

Posterior
probability

Probability
ratio

High-favorability
Moderate-favorability
Low-favorability

5.9
4.3
89.8

22 (95.6)
0 (0.0)
1 (4.4)

23 (74.2)
4 (12.9)
4 (12.9)

0.0016
0.0016
0.0016

0.0226
0.0027
0.0001

14.13
1.69
0.06

12
patterns, the separation is more uncertain (Porwal,
Carranza, and Hale, 2003a). In a spatial domain,
this results in a well-defined separation of highfavorability zones from low-favorability zones.
Cheng and Agterberg (1999) used a linear fuzzy
membership function based on an index of spatial association turned contrast (Bonham-Carter and
Agterberg, 1990) for estimating fuzzy membership
values in their data-driven fuzzy WofE model. Although the use of contrast for calculating fuzzy membership values incorporates a purely data-driven approach in the modeling procedure, it is suitable for
multiclass predictor maps if, and only if, each pattern contains at least one known deposit, because
the contrast value ( = W+ W ) of a predictor pattern is estimated from a pair of weights, which are
determinable if, and only if, the pattern contains at
least one known deposit. Therefore, the fuzzy WofE
of a predictor pattern in a data-driven fuzzy WofE
model may become indeterminable if the pattern
does not contain known deposits. On the other hand,
the knowledge-based logistic function [Equation (2)]
used in the present hybrid fuzzy WofE model can
be applied to derive fuzzy membership of a predictor pattern even if it does not contain a known deposit. The calculation of the fuzzy WofE of a predictor pattern in the present hybrid fuzzy WofE model
[Eqs. (5) and (6)] requires the conditional probabilities of only the patterns with the highest and the lowest fuzzy membership values, given the presence or
absence of a deposit. The fuzzy WofE of a predictor
pattern in the present hybrid fuzzy WofE model is,
however, indeterminable if, and only if, neither the
pattern with the highest fuzzy membership value nor
the pattern with the lowest fuzzy membership value
contains a known deposit, which can happen rarely.
At the least, the pattern with the highest fuzzy membership value always contains at least one known
deposit. Consequently, hybrid fuzzy WofE modeling
can be used more conveniently with multipass predictor maps, even if there are few known deposits available. For the same reason, the hybrid fuzzy WofE
model is applicable to mineral potential mapping in
poorly of explored provinces containing few known
mineral deposits.
The fuzzy WofE model [Eq. (4)] uses Bayes
rule under an assumption of conditional independence for combining fuzzy WofE to derive fuzzy posterior probabilities. The linear nature of the fuzzy
WofE model, however, entails that it is highly sensitive to violation of the assumption of CI among two
or more predictor maps with respect to the target

Porwal, Carranza, and Hale


variable. As discussed by Singer and Kouda (1999),
the assumption of CI may be difficult to validate using a 2 test. Moreover, our previous work on the
application of the ordinary WofE method (Porwal,
Carranza, and Hale, 2003a) indicates that even if
Kolmogorov-Smirnov and 2 tests for goodness-offit (Agterberg, Bonham-Carter, and Wright, 1990;
Bonham-Carter and Agterberg, 1990) return a statistically significant goodness-of-fit between observed
and expected frequencies of deposits, the omnibus
test (Bonham-Carter, 1994) may indicate conditional
dependence between two or more input predictor
patterns (Porwal, Carranza, and Hale, 2003a; also
see Singer and Kouda, 1999). Here, we applied the
new omnibus test (Agterberg and Cheng, 2002) and
determined significant violation of CI assumption in
the final fuzzy posterior probability map. Therefore,
in practice, the possibility of conditional dependence
among two or more predictor patterns with respect
to locations of target deposit-type cannot be ruled
out completely. As a result, fuzzy posterior probabilities calculated using Bayes rule in general, are artificially inflated and, hence, cannot be interpreted in
an absolute sense for decision-making. Moreover, a
fuzzy posterior probability is an updated prior probability of a deposit-type, given the presence of a number of input predictor patterns, and can be accepted
in an absolute sense, if, and only if, it is assumed
that the input predictor patterns adequately represent all geologic processes that were responsible for
the spatial localization of the deposit-type. Such an
assumption is never justified in practice. In addition,
there is always an uncertainty (because of missing
patterns, miss-assigned patterns and fuzzy membership values) associated with fuzzy posterior probability values. However, the effect of over-estimation
of fuzzy posterior probability might be mitigated if
the exploration targets are selected on the basis of
relative favorability rather than absolute fuzzy posterior probabilities, as suggested by Pan and Harris
(2000) for the ordinary WofE approach. Accordingly, we interpret the fuzzy posterior probabilities
in relative terms only in order to classify zones with
high-favorability, moderate-favorability, and lowfavorability for base-metal deposit occurrence.
The validation of the hybrid fuzzy WofE model
of favorability for base-metal occurrence indicates
that, relative to estimated prior probability (using either the 23 training deposits or all the
54 known deposits), (a) the probability of basemetal deposit occurrence is significantly enhanced
in the high-favorability zones, (b) the probability of

A Hybrid Fuzzy Weights-of-Evidence Model for Mineral Potential Mapping


base-metal deposit occurrence is slightly enhanced in
the moderate-favorability zones, and (c) the probability of base-metal deposit occurrence is significantly
reduced in the low-favorability zones. A comparison
between total uncertainty due to missing and missassigned patterns and uncertainty due to fuzzy membership values (Fig. 4) shows that the former is lower
in all parts of the study area. This indicates that if
the missing pattern areas are assigned appropriate
fuzzy membership values then the total uncertainty
in fuzzy posterior probabilities can be reduced, as
suggested by Cheng and Agterberg (1999).
CONCLUSIONS
(1) The use of a knowledge-based fuzzy membership function in the hybrid fuzzy WofE
approach facilitates representation of spatial
evidence of mineral deposit occurrence as
multiclass predictor maps.
(2) In practice, the possibility of conditional dependence among two or more input predictor maps with respect to target variable cannot be ruled out completely. As a result,
estimates of fuzzy posterior probabilities
generally are artificially inflated. These estimates of posterior probabilities, therefore,
cannot be strictly applied in absolute terms.
However, they can be used to classify areas in
terms of relative favorability for occurrence
of target deposit-type.
(3) The hybrid fuzzy WofE approach is demonstrated to be useful in demarcating prospective ground for mineral exploration. In this
study, the hybrid fuzzy WofE approach was
used to demarcate high-favorability zones for
base-metal deposit occurrence in the southcentral Aravalli province based on regionalscale predictor maps. The same approach,
however, requires further testing in modeling of larger-scale predictor maps to demarcate specific target zones within predicted
prospective ground.
REFERENCES
Agterberg, F. P., and Cheng, Q., 2002, Conditional independence
test of weights-of-evidence modeling: Natural Resources Research, v. 11, no. 4, p. 249255.
Agterberg, F. P., Bonham-Carter, G. F., and Wright, D. F., 1990,
Statistical pattern integration for mineral exploration, in
G., and Merriam, D. F., eds., Computer Applications in
Gaal,

13

Resource Estimation Prediction and Assessment for Metals


and Petroleum: Pergamon Press, Oxford-New York, p. 121.
An, P., Moon, W. M., and Rencz, A., 1991, Application of fuzzy set
theory for integration of geological, geophysical and remote
sensing data: Can. Jour. Exploration Geophysics, v. 27, no. 1,
p. 111.
Bonham-Carter, G. F., 1994, Geographic Information Systems for
geoscientists: modeling with GIS: Pergamon Press, Ontario,
398 p.
Bonham-Carter, G. F., and Agterberg, F. P., 1990, Application
of a microcomputer-based geographic information system to
mineral-potential mapping, in Hanley, J. T., and Merriam,
D. F., eds., Microcomputer-based Applications in Geology,
II: Pergamon Press, New York, p. 4974.
Brown, W. M., Gedeon, T. D., Groves, D. I., and Barnes, R. G.,
2000, Artificial neural networks: a new method for mineral
prospectivity mapping: Australian Jour. Earth Sciences, v. 47,
no. 4, p. 757770.
Carranza, E. J. M., and Hale, M., 2001, Geologically constrained
fuzzy mapping of gold mineralisation potential, Baguio district, Philippines: Natural Resources Research, v. 10, no. 2, p.
125136.
Cheng, Q., and Agterberg, F. P., 1999, Fuzzy weights of evidence
and its application in mineral potential mapping: Natural Resources Research, v. 8, no. 1, p. 2735.
Deb, M., 1999, Metallic mineral deposits of Rajasthan, in Kataria,
P., ed., Proc. Seminar on Geology of Rajasthan Status and
Perspective (Udaipur, India), p. 213237.
Deb, M., and Sarkar, S. C., 1990, Proterozoic tectonic evolution
and metallogenesis in the Aravalli-Delhi orogenic complex,
NW India: Precambrian Research, v. 46, no. 12, p. 115137.
Gettings, M. E., and Bultman, M. W., 1993, Quantifying favorableness for occurrence of a mineral deposit type an example
from Arizona: U. S. Geol. Survey, Open-File Rep. 93392,
23 p.
GSI, 1981, Total intensity aeromagnetic map and map showing the
magnetic zones of the Aravalli region, southern Rajasthan
and northwestern Gujarat, India: Geol. Survey India, Hyderabad, India, map, scale: 1: 253,440.
Gupta, S. N., Arora, Y. K., Mathur, R. K., Iqballuddin, Prasad,
B., Sahai, T. N., and Sharma, S. B., 1995a, Lithostratigraphic
map of Aravalli region (2nd edn.): Geological Survey of India, Calcutta, India, map, scale: 1: 250,000.
Gupta, S. N., Arora, Y. K., Mathur, R. K., Iqballuddin, Prasad,
B., Sahai, T. N., and Sharma, S. B., 1995b, Structural map
of the Precambrian of Aravalli region (2nd edn.): Geological
Survey of India, Calcutta, India, map, scale: 1:250,000.
Gupta, S. N., Arora, Y. K., Mathur, R. K., Iqballuddin, Prasad,
B., Sahai, T. N., and Sharma, S. B., 1997, The Precambrian
geology of the Aravalli Region: Geol. Survey India, Mem. v.
123, GSI, Hyderabad, India, 262 p.
Haldar, S. K., 2001, Grade-tonnage model for lead-zinc deposits
of Rajasthan, India: Proc. Intern. Workshop on Sedimenthosted Lead-Zinc Deposits in the Northwestern Indian
Shield (New Delhi and Udaipur, India), p. 153160.
Heron, A. M., 1953, The geology of central Rajputana. Geol. Survey India Mem., v. 79, no. 1, 389 p.
Kemp, L. D., Bonham-Carter, G. F., and Raines, G. L., 1999, ArcWofE: ArcView extension for weights of evidence mapping.
http://gis.nrcan.gc.ca/software/arcview/wofe.
Kemp, L. D., Bonham-Carter, G. F., Raines, G. L., and
Looney, C. G., 2001, Arc-SDM: ArcView extension for

14
spatial data modelling using weights of evidence, logistic regression, fuzzy logic and neural network analysis:
http://ntserv.gis.nrcan.gc.ca/sdm/.
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: Australian Jour.
Earth Sciences, v. 47, no. 5, p. 929942.
Pan, G. C., and Harris, D. P., 2000, Information synthesis for mineral exploration: Oxford Univ. Press, Inc., New York, 461 p.
Porwal, A., and Sides, E. J., 2000, A predictive model for
base-metal exploration in a GIS environment: Intern.
Archives Photogrammetry and Remote Sensing, v. 33, pt. B7
(Amsterdam), p. 11781184.
Porwal, A., Carranza, E. J. M., and Hale, M., 2003a, Extended
weights-of-evidence modelling for predictive mapping of
basemetal deposit potential in Aravalli Province, Western India: Exploration and Mining Geology, v. 10, no. 4, p. 155
163.
Porwal, A., Carranza, E. J. M., and Hale, M., 2003b, Knowledgedriven and data-driven fuzzy models for predictive mineral
potential mapping: Natural Resources Research, v. 12, no. 1,
p. 125.
Raja Rao, C. S., Poddar, B. C., Basu, K. K., and Dutta, A. K.,
1971, Precambrian stratigraphy of Rajasthan: a review: Geol.
Survey India, Records v. 101, no. 2, p. 6079.
Roy, A. B., 1988, Stratigraphic and tectonic frame work of the Aravalli Mountain Range, in Roy, A. B., ed., Precambrian of

Porwal, Carranza, and Hale


the Aravalli Mountain Rajasthan, India: Geol. Society India
Mem., v. 7, p. 331.
Roy, A. B., Sharma, B. L., Paliwal, B. S., Chauhan, N. K., Nagori,
D. K., Golani, P. R., Bejarniya, B. R., Bhu, H., and Ali
Sabah, M., 1993, Lithostratigraphic and tectonic evolution
of the Aravalli Supergroup: a protogeosynclinal sequence, in
Cassyap, A. M., ed., Rift Basins and Aulacogens: Gyanodaya
Prakashan, Nainital, p. 7390.
Sarkar, S. C., 2000, Geological setting, characteristics, origin
and evolution of sedimenthosted sulfide ore deposits of
Rajasthan: a critique with comments on their implications
for future exploration, in Deb, M., ed., Crustal Evolution and
Metallogeny in the Northwestern Indian Shield: Narosa Publ.
House, New Delhi, p. 240292.
Singer, D. A., and Kouda, R., 1999, A comparison of the weights of
evidence method and probabilistic neural networks: Natural
Resources Research, v. 8, no. 4, p. 287298.
Sugden, T. J., Deb, M., and Windley, B. F., 1990, The tectonic
setting of mineralisation in the Proterozoic Aravalli-Delhi
orogenic belt, NW India, in Naqvi, S. M., ed., Precambrian
Continental Crust and its Economic Resources: Elsevier,
Amsterdam, p. 367390.
Thiart, C., Bonham-Carter, G. F., and Agterberg, F. P., 2003,
Conditional independence in weights-of-evidence: application of an improved test: Proc. IAMG 2003 Ann. Conf.
(Portsmouth), England, CD-ROM.
Zimmerman, H.-J., 1991, Fuzzy set theory and its applications
(2nd edn.): Kluwer Acad. Publ., Dordrecht, 399 p.

Potrebbero piacerti anche