Professional Documents
Culture Documents
To cite this article: Jos M. C. Pereira & Lucien Duckstein (1993): A multiple criteria decision-making approach to GIS-based
land suitability evaluation, International Journal of Geographical Information Systems, 7:5, 407-424
To link to this article: http://dx.doi.org/10.1080/02693799308901971
I.
408
409
where /Ii (/,zO, Zbi= 1) are weights assigned to the criteria and p ranges from I to oo.
Varying p affects the relative contribution of individual deviations from the ideal
point, a greater emphasis being given to larger deviations as p tends toward m. In
practice, and if the natural scale scores are standardized to the range [0, 11, the largest
410
deviation (x:-xf) totally dominates the evaluation, when p becomes greater than a
value of approximately 10. In this case, the suitability of a database cell for a given
purpose is only as high as its lowest score on all criteria (map layers) and the evaluation
relies on a minimax rule, in which the minimum of the maximum weighted deviation
/Ii(x:-xf) is sought.
C P as usually applied includes a sensitivity analysis for the three strategic values
p = 1,2 and a nominal co (i.e. p > 10). When p= I (a 'city block'distance metric), total
compensation between criteria is assumed, meaning that a decrease of one unit of one
criterion can be totally compensated by an equivalent gain on any other criterion. This
is basically equivalent to a weighted linear additive model, of the kind available in some
commercial GIs, such as the Indexing function of SPANS. For p = 2 (a straight line,
shortest distance metric) there is only partial compensation and p = co represents a
totally non-compensatory situation (Zeleny 1982, pp. 322-325).
2.3. Relationships with ecological analysis and image classification techniques
Distance metrics are commonly used in multivariate analysis and classification
problems, namely in the fields of community ecology and the processing of remote
sensing images. Their application in MCDM methodologies, specifically in CP, is
related to these other applications. However, similarities are mainly algorithmic or
computational, while differences are of a more fundamental, conceptual nature. A
comparison of distance metrics used in ecological classification (e.g., Digby and
Kempton 1987, pp. 89-90), map classification ( e g , Burrough 1986, p. 141), image
classification (e.g., Schowengerdt 1983, pp. 50-51; Richards 1986, p. 191) and MCDM
(Goicoechea et a / . 1982, pp. 236-237; Zeleny 1982, pp. 316-325) clearly reveals that
distance metrics such as 'city block' and Euclidean are common to all these areas of
research, while others, such as Mahalanobis distance, are shared by community
ecology and image processing. Some other distance metrics have a more restricted
popularity, such as the Bray-Curtis distance, in ecology (Digby and Kempton 1987) or
the Bhattacharyya distance, in image processing (Haralick and Fu 1983).
There is, however, an essential difference between the use of distance metrics in an
MCDM technique, such as CP, and in fields like ecology and image processing. In
MCDM,distance is employed as a proxy for preferential similarity with an idealchoice,
in a value-laden, judgmental process. The use of distance as a proxy for ecological or
spectral similarities, on the contrary, is a value-neutral affair. This difference is
implemented through the inclusion of the Pi weights in the C P family of metrics, and
also in the transformation of measured physical magnitudes of evaluation criteria into
a standardized value scale, elicited from a decision-maker.
2.4. Data standardization
The nced for data standardization in CIS-based LSE often arises as a consequence
of the need to integrate into the evaluation process data measured not only in different
units but also in different scales of measurement, such as nominal, ordinal, interval and
ratio scales. The approach we propose, derived in the specific context of the application
presented below, is an implementation of a methodology for value measurement
suggested by von Winterfeldt and Edwards (1986, pp. 220-224).
The firststep of this methodology involves the definition of objects to which values
are to be attached, such as slopes and their inclination and orientation, o r stands
dominated by different tree species. Second, a natural scale is constructed, representing
some physical quantitative attribute of the objects being evaluated (e.g., percentage
41 1
slope, aspect in degrees from North, o r tree canopy density). The third step is the
construction o f a value function, with which the natural scale derived in the previous
step is converted to a value scale, through judgments of relative value of strength of
preference.
Thus, our purpose with scaling is not to produce a data set that performs better
(according to some figure of merit) under a distance metric transformation, as it is
sometimes done, for example, in community ecology work (Digby and Kempton 1987,
p. 12). The reasons for standardizing the data into a common value scale are: 1. to
ensure that all natural scales, some of which may include nominal or ordinal data, are
converted to a common value scale with interval properties, required for the legitimate
application of a distance metric evaluation; 2. T o account for the possibly non-linear o r
even non-monotonic character of the relationship between natural and value scales.
It may be worth clarifying that the data standardization is not designed to make the
multi-criteria distance scores independent from the absolute values of the criteria. In
fact, the distance functions are clearly not independent of a positive linear transformation of individual degrees of proximity to the ideal, and changing the scale of the
degrees of closeness has an effect similar to that of manipulating the /Ii weights (Zeleny
1982, p. 168).
2.5. Value and weights assessment
In addition to the reasons for data standardization mentioned in the previous
section, such a transformation also simplifies computation of the various distance
metrics, since when the natural scales are mapped onto a [O,l] standardized interval,
(x.-x:)
in (5) simplifies to (1 -x:). This does not require the ideal point to be at the
extreme of the range, o n the attribute's natural scale. It simply means that, with the
mapping of all potentially incommensurate attribute natural scales into a common
value scale, where 0 is least preferred and I most preferred, the ideal point must be at the
extreme of the value scale, by definition of value or preference. However, the maximum
value of 1 may be assigned to non-extreme scores of the natural scale. This would imply
the existence of a non-monotonic relationship between natural and value scales.
Assessment of the /Ii weights in (5)can also be performed in many different ways, but
when it is necessary to weight interval-scaled criteria against ordinal-scaled criteria the
range of options is somewhat reduced since it becomes impossible to use indifference
judgments (von Winterfeldt and Edwards 1986, Chapter 8). In indifference methods,
pairs of evaluation objects are varied, until a match is established in their respective
strengths of preference (von Winterfeldt and Edwards 1986). Specifically, with the
bisection method mentioned below, the respondent is first asked to find the two
extreme evaluation objects, spanning the entire range of values. Next, he/she is asked to
identify a n object halfway in value between the two extremes, and additional
subdivision of the scale refines the value function. In general, three carefully assessed
points of the value function provide enough information to produce a smooth curve
(von Winterfeldt and Edwards 1986, p. 236).
The case study we present includes both interval (continuous), and nominal and
ordinal (categorical) data, and requires therefore a combination of procedures for
standardization of single criteria evaluations, and the use of a weightingbrocedure that
allows comparisons between differently-scaled data. Derivation of a value scale for
criteria whose natural scale is a n interval scale, is accomplished using the bisection
method (von Winterfeldt and Edwards 1986, p. 235-236) as implemented in the MultiAttribute Tradeoff System (MATS) software package (Brown et al. 1986). Construction
4.12
of value functions for those criteria whose natural scales are of ordinal level, and
assessment of the pi weights were based on priority theory methods, using EXPERT
CHOICE(Forman et al. 1983) toimplement Saaty's Analytic Hierarchy Process(AHP)
(Saaty 1977, 1980). The relationships between A H P and value function assessment are
dealt with by Kamenetzky (1982) and the transformation he proposes (p. 708) is used to
rescale theAHP scores to a [0, I] value interval. These AHP-generated value functions
are not expressed on the methods' usual ratio scale, but can be given the strength of
preference interpretation (Kamenetzky 1982, pp. 708-709) required for use with a
distance-based MCDM technique.
413
3. Case study
The GIs-based LSE methodology outlined above was applied to the problem of
assessing quality of habitat for the endangered Mount Graham red squirrel,
Tamiasciurus hudsonicus grahamensis. Relevant autoecological information concerning this species and specifications of the digital cartographic database developed for
studying its habitat preferences are described in Pereira (1989) and Pereira et al. (1991)
and are only very briefly summarized here.
3.1. Red squirrel autoecology and digital database
The Mount Graham red squirrel is an arboreal endemic subspecies of the higher
areas of Mount Graham, Graham County, Arizona, (max. elevation 3216111) whose
population is estimated at around 220 individuals. The red squirrel feeds primarily on
seeds from conifer cones, nests on large trees, in areas of dense canopy cover, and
searches for cool, shaded spots to store its winter food supply (U.S.Forest Service
1988). In the light of the ecological information and cartographic data available, a
raster G I s database was developed, for the top part of Mount Graham, covering a n
area of6460 ha. The database contains 12 920 squaregrid cells, each 0.5 ha at 70.7 m, on
a side. The layers elevation, slope, aspect, land cover types, canopy cover, tree diameter
at breast height (d.b.h.) and distance to clearings in the forest canopy were used for the
assessment of habitat (table 1).
3.2. Problem structuring and computational details
Figure 1 represents the habitat suitability analysis for the Mount Graham red
squirrel as a hierarchical structure, where the top level is the ultimate goal to be
achieved, i.e., habitat evaluation, the intermediate level lists and the relevant evaluation
criteria, and the bottom level contains the evaluation objects.
The weights for all criteria were assessed by having our guest expert perform
pairwise comparisons between the seven criteria, regarding their relative importance.
EXPERTCHOICE presents the respondent with a 7 by 7 matrix and a set ofquestions,
Map layer
Data type
Elevation
Slope
Aspect
Land cover
continuous
continuous
categorical
categorical
Units
Range
1956-3216
0-244
Categories
-
N, E, S, W, None
Spruce-fir
Mixed conifer
Ponderosa pine
Other
No canopy, l W 0
Canopy cover
categorical
D.b.h.
categorical
Distance to clearings
continuous
0-32
41-70,71-100
No trees, < 13
13-23, > 23
Note: Bold-faced categories and scores represent the ideal level for each individual criterion. The ideal
point is the actually unavailable multiple criteria alternative that offers the ideal level for each and every
criterion.
Figure I. Hierarchical structure olthe problem.The top level represents the ultimate goal ofthe
MCDM analysis process, which was evaluation of habitat suitability for the Mount
Graham red squirrel. The intermediate level lists the relevant evaluation criteria that were
compared pairwise to assess their relative weights. Some of these criteria refer to
categorical data, and branch o f to a bottom level representing the actual evaluation
objects or natural scale scores. Their values were also estimated by pairwise comparison.
The elevation, slope and distance criteria refer to continuous data. Their value was not
estimated by pairwise comparison, but by the indifference assessment techniquedescribed
in 9 3.2.
Value score
Spruce-fir
Mixed conifer
Ponderosa pine
Other
Value functions for interval-scaled. continuous criteria were assessed using MATS.
in interrogation mode. When the op;ion for value function construction is selected;
MATS asks for the name of the criterion, the full range of scores in the natural scale, and
Value
Value
b.353
0.4
0.2
0
2000 2200 2400 2600 2800 3000 3200
Elevation (m.)
Value
8-,353
0.8
0.8
0.4
0.2
X Canow cover
10 15 20 25
D i m e ( I ) cells)
30
35
Figure 2. Expert-derived weights and value functions for the Mount Graham red squirrel
habitat suitability evaluation. Weights are /3-values inside rectangles. Continuous
functions correspond to continuous data types and discrete functions to categorical data
types.
416
thc most and least preferred of these. Exemplifying with the elevation criterion, the
range is 2000-3200 m, where 3200 m is the most preferred score (value of I) and 2000 m
the least preferred (value of 0). A graph with a linear function linking the points (2000,O)
and (3200, 1) and passing through the value midpoint (2600, 0.5), is displayed,
accompanied by the question of which change would represent the greatest
improvement:
In order to calculate the distance between actual pixels and the ideal point, it is
necessary to subtract each value transformed map layer from the scalar 1, which was
accomplished with the SCALAR module of IDRISI. The output map layers were
exponentiated t o p = 1 , 2 and 10 and multiplied by the exponentiated piweights, using
the same IDRISI module. Then the 'addition' option of the IDRISI OVERLAY
module was used to add together all map layers, and the result was exponentiated to
Ilp. This set of operations was repeated three times, once for each p-value.
417
418
would be to use the model with p = 2 at the class 4 threshold. This decision would
guarantee the protection of 70.3 per cent of the sites with squirrel activity, while
covering no more than 16 per cent if the study area, corresponding to a model based
gain in predictive power of 54.3 per cent. These results clearly confirm the visual
impression of a general good agreement between the classification of habitat suitability
obtained with CP, and the observed pattern of habitat used by squirrels.
5.
Conclusions
The alternative methodology for CIS-based LSE proposed in the present paper
contains three main elements that need to be assessed at this point. First, there is the
central issue of reducing the scaling and dependence problems of most conventional
studies of land suitability; second, it is necessary to consider whether the MCDM
techniques selected for implementation of the methodology performed as expected;
third, the question of dependence of the results on the ordinal definition of suitability
classes and the choice of threshold level (figure 6) should be investigated.
The scaling problem can be correctly solved with techniques of value, priority and
utility assessment whose theoretical soundness has been firmly established. The
diversity of techniques available facilitates inclusion of factors measured not only in
different units, but also in different scales in a single evaluation process, as was
demonstrated in thecase study. Independence is treated in the way proposed by Zeleny
(1982,pp. 161-165), with each alternative being considered as an inseparable 'bundle of
attributes'. for which is would be pointless to treat dependence as se~arableand assess
value and' preferential dependence. This approach'is justifiable given the natural
interdependence of factors in environmental analysis where, for example, occurrence of
a given elevation level conditions the range of temperature values or plant productivity
levels that can be achieved.
The choice of C P as the appropriate MCDM technique for the task seems to
racilitate sensitivity analysis of different strategies with respect to data aggregation
(p values), a very desirable feature in the absence of strong theoretical justification for a
specific procedure. The results thus obtained appear to be interesting and informative,
thus providing the stimulus for testing the potential of alternative MCDM techniques
for LSE. Finally, there is the issue of selecting the optimum threshold for binary habitat
classification based on the discrete ordinal data. Use of the original interval-scaled data
would have permitted a more precise determination of the optimum cutoff point and
probably produced slightly better classification results. However, it seems likely from
figure 5 (c)that the model where p = 2 would still produce the best binary classification,
at a value within or close to class 4. The other two models display flatter maxima, and
therefore the optimum threshold could suffer larger shifts, especially for p= co.
To summarize, the following points may be made:
1. MCDM seems to be applicable to CIS-based land suitability analysis.
2. C P makes it possible to investigate alternative strategies for data aggregation
and their conseauences.
3. The degree of accuracy of classification obtained from a subjective, expert-based
decision model matches that of inductive multivariate statistical models (Pereira
1989, 1991).
4. These results suggest the potential usefulness of CIS-based MCDM for other
problems of land suitability analysis, in areas such as urban and rural land-use
planning, facility location, and environmental impact assessment.
# cells
5000 1
suitability class
# of cells
5000
0=2
suitabilily class
"
0 1 2 3 4 5 6 7 8 9
suitability class
Figure 4. Histograms comparing the number of cells assigned to each suitability class, for
models with p = 1,2 and oo.
423
Acknowledgments
T h e U S D A Forest Service, C o r o n a d o National Forest, funded the M o u n t G r a h a m
Red Squirrel Geographic Information System Project. They also provided expert
information regarding M o u n t G r a h a m and the red squirrel, mainly through M a r k
Kaplan, Randy Smith and Jerry Connors, t o w h o m we a r e very thankful. Professor
Robert M. Itami directed the construction of th'e GIs database a n d contributed with
many helpful suggestions. W e also thank t w o anonymous reviewers for the constructive criticisms a n d recommendations.
References
BERRY,
J. K., 1986, Learning computer assisted map-analysis. Journal of Forestry, 84, 3943.
BROWN,C. A,, GRANT,R. W., and STINSON,D. P., 1986, Multi-Attribute Tradeoff System.
Personal computer version user's manual. Public and Social Evaluations Office, Division
of Planning and Technical Services, Engineering and Research Centre (Denver, CO:
Bureau of Reclamation, USDI).
BURROUGH,
P. A,, 1986, Principles of Geographical Information Systems for Land Resources
Assessment (Oxford: Clarendon Press).
DIGBY,P. G. N., and KEMPTON,
R. A.. 1987, Multivariate Analysis ofEcologica1 Communities
(London: Chapman and Hall).
DUCKSTEIN,
L., and OPRICOVIC,
S., 1980, Multi-objective optimization in river basin development. Water Resources Research, 16, 14-20.
EASTMAN,
J. R., 1990, IDRISI: A Grid-Based Geographic Information System. Version 3.2.
(Worcester, MA: Clark University Graduate School of Geography).
FORMAN,
E. H.,SAATY,T.L.,SELLY,
M. A.,and WALDRON,
R., 1983, Expert Choice(McLean, VA:
Decision Support Software Inc.).
GOICOECIIEA,
A,, HANSEN,
D. R., and DUCKSTEIN,
L., 1982, Mulfiobjectiue Decision Analysis with
Engineering and Business Applications (New York, NY: J. Wiley & Sons).
HARALICK,
R. M., and Fu, K.-S., 1983, Pattern recognition and classification. In Manual o j
Remote Sensing, vol. 1, edited by R. N. Colwell (Falls Church Va: American Society of
Photogrammetry), pp. 793-805.
HOPKINS,
L., 1977, Methods for generating land suitability maps: a comparative evaluation.
Journal o f t h e American Institute of Planning, October 1977, 386400.
JACQUET-LACR~ZE,
E., and SISKOS,J., 1983, Concepts et modeles en analyse multicritere. In
Mithode de Decision Multicritere, edited by E. Jacquet-Lagreze and J. Siskos (Paris: Ed.
Hommes et Techniques), pp. 7-38.
KAMBNETZKY,
R. D., 1982, The relationship between the analytic hierarchy process and the
additive value function. Decision Sciences, 13, 702-713.
NGOWI,J., and STOCKING,
M., 1989,Assessing land suitability and yield potential for coconuts in
Tanzania. Applied Geography, 9, 21-33.
PEREIRA,
J. M. C., 1989, A spatial approach to statistical habitat suitability modelling: the Mt.
Graham red squirrel case study. Unpublished Ph.D. dissertation, School of Renewable
Natural Resources, University of Arizona, Tuscon, AZ.
PEREIRA,
J. M. C., and ITAMI,R. M., 1991, GIs-based habitat modelling using logistic multiple
regression: a study of the Mt. Graham red squirrel. Photogrammetric Engineering and
Remote Sensing, 57, 1475-1486.
ROY,B., 1975, Vers une methodologie generale d'aide a la decision. Metro, 14, 459497.
RICHARDS,
J. A,, 1986, Remote Sensing Digital Image Analysis. An Introduction (Berlin: SpringerVerlag).
SAATY,T. L., 1977, A scaling method for priorities in hierarchical structures. Journal of
Mathematical Psychologg, 15, 234-281.
SAATY,
T. L., 1980, The Analytical Hierarchy Process (New York: McGraw-Hill).
SCHARLIG,
A., 1985, Decider sur Plusieurs CritPres (Lausanne: Presses Polytechniques
Romandes).
SCHOWENGERDT,
R. A,, 1983, Techniques for Image Processing and Classification in Remote
Sensing (Orlando, FL: Academic Press).
424
TOMI.IN,C. D., 1986, The IBM personal computer version of the Map Analysis Package.
GSD/IBM AcIS Project, Report No. LCGSA-85-16 Laboratory for Computer Graphics
and Spatial Analysis. Graduate School of Design, Harvard University.
U.S. FOR IS.^ SIXVICE
1988, Mount Graham red squirrel-an expanded biological assessment.
Coronado National Forest, Tucson, AZ.
V I N C KP.,~ 1986,
~
Analysis of muhicriteria decision aid in Europe. European Journal of0perarions
Research, 25, 160-1 68.
VON WINTERI:II~.~T,
D., and EDWARDS,
W., 1986, Decision Analysisand Behauioral Research (New
York: Cambridge University Press).
ZHLENY,
M., 1973, Compromise programming. In Multiple Criteria Decision Making, edited by
J . L. Cochrane and M . Zeleny (Columbia, SC: University of South Carolina Press).
ZEI.I:NY,M., 1982, Multiple Criteria Decision Making (New York: McGraw-Hill).