Beruflich Dokumente
Kultur Dokumente
167]
On: 05 April 2013, At: 18:00
Publisher: Taylor & Francis
Informa Ltd Registered in England and Wales Registered Number: 1072954 Registered
office: Mortimer House, 37-41 Mortimer Street, London W1T 3JH, UK
To cite this article: K. K. Benke & C. Pelizaro (2010): A spatial-statistical approach to the
visualisation of uncertainty in land suitability analysis, Journal of Spatial Science, 55:2, 257-272
To link to this article: http://dx.doi.org/10.1080/14498596.2010.521975
1. Introduction
Estimation of land suitability and its spatial
variability represents a continuing challenge
to resource planning and policy development. Optimisation of agricultural production is possible by matching crops to the
most appropriate regional areas, although
climate-based predictions can be subject to
many uncertainties (Dessai et al. 2007;
CAWCR 2008; Karoly 2008). Improvements are necessary in the supply and
accuracy of modelling and data sets relating
to land suitability in local regions, such as
the South West region in Victoria, where
data distribution is required to inform
farmers and policy-makers alike.
258
numerical
criterion from
numerical
predictors.
Alternatives to AHP-based approaches
to weight-assignment in MCDA are cited in
the literature (see, for example, Kujawski
2003; Lahdelma and Salminen 2006). Subject to the type of MCDA applied, the
variables can be expressed as degrees of
preferences, numerical scores or levels of
achievement for each criterion. In the AHP
approach, alternatives are compared sequentially by subject-matter experts with
respect to performance against a set of
decision criteria. A set of ratings (biophysical measurements) are used as inputs, and
a hierarchy of weightings is derived from
consideration of expert opinion gained
from regional workshops. A matrix of
judgements (given in numerical form) is
provided by decision-makers in a workshop
setting against a set of alternatives in a
process of repeated pair-wise comparisons
and prioritisation checked for consistency
by a process of eigenvalue decomposition
(Saaty 1994a). The procedure yields a set of
priorities or relative weights indicating the
decision-makers preferences for alternatives, and a decision tree is produced with
numerical values assigned to the criterion
weights (Saaty; 1994b; Hahn 2003). An
example of a decision tree producing a
prediction of land suitability is shown in
Figure 1, and applications in a GIS-based
context are provided by Hossain et al.
(2006) and Rahman and Saha (2008).
The AHP diers from a statistical
parametric model, based on regression
analysis, where one compares a set of model
predictions with a set of measurements in a
pre-existing training set (for error minimisation and parameter estimation). With
AHP, estimates reect predictions based
on using experts systematically weighting
the inputs using the AHP process described
in the previous paragraph. In some cases,
AHP may be the only analysis possible, if
biophysical models are not available or not
Figure 1. Example of a typical AHP decision tree for land suitability analysis. Experimental conditions for producing the model for ryegrass
pasture are described in Sposito et al. (2010).
260
Uncertainty analysis
Taxonomy of uncertainty
In the past, uncertainty has been attributed
to a variety of sources, such as the
consequences of statistical variation or
inherent randomness, measurement errors,
systematic errors in judgement, linguistic
imprecision or lack of agreement amongst
experts (Helton 1993; Burgman 2005). In
particular, Burgman (2005) has summarised
sources of uncertainty as being due to the
limitations of measurement devices, insucient data, extrapolations and interpolations, linguistic ambiguity and variability
261
Uncertainty in MCDA
262
263
xk 1
k1
xmean xmin
xmax xmin
4
xmax xmean
wn
xmean xmin
5
264
Figure 2.
265
the Latin Hypercube method, which executes stratied random sampling without
replacement across the full range of each
parameter (Vose 2000). The Latin Hypercube approach provides more uniform
sampling in the tails of the chosen distribution in comparison with traditional random
sampling and generally produces much
faster convergence for a specied number
of iterations (McKay et al. 1979; Iman and
Conover 1983).
Each Monte Carlo simulation experiment in the series was executed over 10,000
iterations. Skewness and kurtosis of the
output distributions approximated that of
the normal distribution in each case. Monte
Carlo simulation outputs can have skewed
distributions, although this has no eect on
the estimation of the condence interval in
this study as it is based on using percentile
dierences from the cumulative probability
distribution, a measure that is independent
of the shape of the distribution (see Benke
and Hamilton 2008). The simulation process is repeated (Figure 2), incrementing the
266
267
90%
CI
20th
perc
80th
perc
C.V.
Std.
Dev.
+0.22
+0.24
+0.26
0.59
0.68
0.77
0.81
0.92
1.00
0.18
0.16
0.13
0.13
0.15
0.16
Winter Wheat
Uncertainty Metrics
Prediction
0.70
0.80
0.90
1.00
90%
CI
20th
perc
80th
perc
C.V.
Std.
Dev.
+0.34
+0.39
+0.43
+0.48
0.52
0.59
0.66
0.73
0.88
1.00
1.00
1.00
0.30
0.29
0.29
0.28
0.21
0.24
0.27
0.29
268
K.K. Benke and C. Pelizaro
269
3
Figure 5. (a) Visualisation of uncertainty in ryegrass map, where uncertainty in land suitability is
represented by the standard deviation. This type of representation has been incorporated into a
modelling framework for optimisation of commodity production (see Sposito et al. 2010). (b) As in (a),
but for winter wheat land suitability predictions.
270
metric used. Normalised uncertainty estimates showed that that the CV ranged from
0.13 to 0.18 for ryegrass, and 0.28 to 0.30
for winter wheat. For high land suitability
values (i.e. LS 0.851.0), representing
close matches between crop type and
production location, more than 90 percent
of uncertainty (standard deviation) in prediction was due to the weights in the model
(representing expert opinion) rather than
the ratings (representing biophysical data).
In workshop conditions, ambiguity due
to linguistic responses may introduce further
epistemic uncertainty, which can be reduced
subject to improved experimental design.
The quantication of uncertainty in the
current case is conservative due to the range
imposed on weight values for the simulation.
More accurate estimates may be possible by
incorporating uncertainty analysis directly
into the AHP process itself, rather than as a
posterior analysis, but at the risk of greater
complexity in theory and operation.
References
Belton, V., & Steward, T. (2002) Multiple
Criteria Decision Analysis, An Integrated
Approach, Kluwer, Boston.
Benke, K.K., & Hamilton, A.J. (2008) Quantitative microbial risk assessment: uncertainty
and measures of central tendency for skewed
distributions. Stochastic Environmental Research and Risk Assessment, vol. 22, pp. 533
539.
Benke, K.K., Hamilton, A.J., & Lowell, K.E.
(2007) Uncertainty analysis and risk assessment in the management of environmental
resources. Australasian Journal of Environmental Management, vol. 14, pp. 1622.
Benke, K.K., Lowell, K.E., & Hamilton, A.J.
(2008) Parameter uncertainty, sensitivity
analysis and prediction error in a waterbalance hydrological model. Mathematical
and Computer Modelling, vol. 47, pp. 1134
1149.
Bomford, M., & Hart, Q. (1999) Risk assessment
for importing and keeping exotic vertebrates
in Australia. Australasian Journal of Environmental Management, vol. 6, pp. 149156.
271
272