Download as pdf or txt
Download as pdf or txt
You are on page 1of 24

Environment and Planning B: Planning and Design 1998, volume 25, 103 -126

Simulation of land development through the integration


of cellular automata and multicriteria evaluation

F Wu, C J Webster
Department of City and Regional Planning, University of Wales Cardiff, PO Box 906, Cardiff
CF1 3YN, Wales; e-mail: WUF@cf.ac.uk; Webster@cardiff.ac.uk
Received 21 June 1996; in revised form 8 February 1997

Abstract. Cellular automata (CA) simulation has become a popular method of exploring the behaviour
of all kinds of self-organising systems. The city may clearly be viewed as such a system but one with a
particularly complex set of transition rules. Many natural processes such as the spread of fire or
vegetation can be modelled by a simple set of local rules. Insofar as the development of a piece of land
depends on the neighbourhood situation as well as on the characteristics of a site, urban evolution can
be treated in much the same way, with transition rules translating the evaluation of the location into a
land conversion outcome. If this modelling paradigm is to be used to gain insight into real-world
urban development processes, there is a need to discover ways of capturing the richness of land
conversion behaviour in the simplifying mechanisms of CA. Our paper contributes to this research
agenda by integrating multicriteria evaluation (MCE) into a CA simulation in order to define non-
deterministic, multidimensional, and multilevel transition rules. An analytical hierarchy process is
used to implement MCE-derived transition rules. The integrated MCE-CA model may be used in a
gaming mode to explore how urban form evolves under different development regimes caricatured by
the set of multicriteria weights. We use it to test loosely hypotheses about the nature of the regimes
that have governed the expansion of a fast-growing southern Chinese city.

1 Introduction
Abstractions about city growth processes are articulated at a variety of behavioural and
spatial scales. Macroeconomic models focus on whole urban economies; neoclassical
urban economics generalises about individual residents, firms, and governments; social
and cultural theories examine networks and the culture and history of localities; and
ecological models deal in habitats, competition, flows, filters, and barriers. The
knowledge domain of urban scientists is currently being enriched on many fronts at
a time when urban problems are gaining in public and political importance. Partly as a
result of developments in the natural sciences, particularly concerning open, complex,
and self-organising systems, t o p - d o w n totalistic modelling approaches are being com-
plemented by theories which emphasise the way in which decisions made locally give
rise to global patterns (Batty, 1995). Cellular automata (CA) are one of these new
approaches with wide appeal owing to their simplicity, their closeness to conventional
factor mapping, and to their flexibility and transparency as a modelling framework.
Governed by simple predefined rules of play, recursive local decisions that determine
how a cell changes from one state to another can be accumulated to produce very
complicated spatial forms. A literature that includes theoretical and empirical studies
is now well established (Batty and Xie, 1994; Couclelis, 1985, 1988; Deadman, et al,
1993; Itami, 1994; Phipps, 1989; Portugali et al, 1994; Portugali and Benenson, 1995;
White and Engelen, 1993).
Urban CA applications differ from the typical natural systems applications described
by Toffoli and Margolus (1987) and Wolfram (1984) in that they tend to have a very
limited total number of cells and a small number of temporal iterations (Batty and Xie,
1994). A more substantial difference, however, lies in the definition of transition rules.
Compared with fire spreading or the movement of a gas pixel, land development in a
104 F Wu, C J Webster

city follows a much more complicated decisionmaking process less easily mapped onto
the logic of a moving kernel. The evolution of land is a function of interdependencies
at many spatial scales and these may be characterised as local - global, neighbour-
hood-city, site - situation or micro-macro factors.
Early attempts at CA simulation of cities were typically more in the nature of
metaphors of urban growth, with little explicit relationship with underlying behavioural
theory (Batty and Xie, 1994; Couclelis, 1985). Various approaches to rule definition
have been tried in an attempt to enrich the theoretical basis of simulation. For
example, White and Engelen (1993) calculate the potential of development from a
predefined parameter matrix which is designed to describe general situations of land
development. Land conversions are simulated based on sorted potentials. Deadman
et al (1993) define rules such that some states change in a deterministic way whereas
others are subject to stochastic influences. Batty and Xie (1994) define the probability
of transition via nested neighbourhood space and a distance decay function from the
seeds of development. Portugali et al (1994) and Portugali and Benenson (1995) simulate
the transition through multiple steps. Each step is subject to a probability of change
which is defined through a mathematical formula with parameters adjusted to reflect its
socioeconomic context.
The objective of our paper is to present an alternative approach to capturing
behaviour when designing CA transition rules. We choose to model the decisionmaking
regimes that govern the generic form of development by making use of multicriteria
evaluation (MCE) techniques. This contrasts, on the one hand, with a deductive
theoretic approach that would attempt to govern simulation by the abstracted behav-
ioural assumptions of, for example, urban bid-rent theory and, on the other, with
nonbehavioural approaches based on physical metaphor or approaches based on
ad hoc rules. The behaviour represented by the MCE parameters can be that of any
actor in the development process, and actors can be assumed to be operating under
any particular market-intervention regime. Behaviour can be captured from the res-
ponses of real decisionmakers or from gaming exercises. MCE also provides an inter-
face for conducting what if simulation runs in a more sophisticated, transparent, and
reliable manner than in heuristic-driven CA.
The reader may feel that the use of an MCE front-end to the CA process results in
a loss of clarity when tracking relationships between individual rule weightings and the
ensuing simulated development patterns. We accept that this is the case when our
approach is compared with experiments that systematically adjust transition rules
and record changes in development patterns. However, our method has a somewhat
different purpose. It is an attempt to caricature the decisionmaking calculus of the
dominant development regime and to investigate the implications of different regimes
for generic growth patterns. The emphasis is therefore on providing a mechanism by
which the generic results of alternative development regimes may be explored. The
particular weighting of local and global factors and the manner in which these interact
in sequence as the simulation unfolds is not examined in detail in this paper. This kind
of systematic exploration of simulative behaviour would be appropriate after the MCE
has captured the decisionmaking behaviour of a number of types of developer or
regulative regime. In fact, the MCE-generated simulations would allow many hypotheses
of the general form to be tested: zm ^ zn, where z is a morphological or morphotemporal
pattern measurement and m, n denote the regime under which the measured pattern
was simulated.
In the remainder of this paper we present an experiment in which we have used
an analytical hierarchy process (AHP) for the multicriteria evaluation of transition
probability. In section 2 we discuss the methodology for integrating local, global, and
Simulation of land development 105

explicit sequencing information under MCE gaming scenarios. In section 3 we report


on how the simulation was implemented through links with geographic information
systems (GIS) software. In section 4 an empirical experiment is reported in which the
integrated technique is applied to Guangzhou, a fast-growing city in southern China.
It applies alternative MCE-derived transition rules that represent the very different
development processes operative in Guangzhou over the past two decades. Differences
in the dominant preferences represented in the alternative rules are shown to lead to
different spatial forms. Simulation results are compared with observed changes in the
urban form of Guangzhou and the issues of calibration and goodness of fit in the
context of our method are discussed. In the conclusion, we reflect on the significance
of our experiment and suggest further approaches to the definition and calibration of
behavioural transition rules.

2 An integrated MCE-CA methodology


Land development bears some but not all of the fundamental characteristics of CA, as
defined by Wolfram (1984): space consists of a discrete lattice of sites; it evolves in
discrete time steps; each site takes on a finite set of possible states; the state of each site
evolves according to the same rules; and rules governing evolution of the system
depend only on a local neighbourhood of sites around it. The last two points demand
that the system has laws that are both uniform and local (Toffoli and Margolus, 1987).
In the context of land development, however, the idea of an exclusively local law is
problematic because it precludes any consideration of global (city-scale) factors, such
as site accessibility, attractiveness relative to other sites across a city, market conditions
of supply and demand, and so on (Couclelis, 1985). Fundamentally, urban development
is not governed by purely locally defined behaviour. Developers' own preferences and
the preferences of planners and urban governments both reflect a balance of macro-
scopic and microscopic considerations. Urban planning decisions clearly take account
of global criteria such as the demand for housing or industrial land in a city or the
need to locate large employers near public transport nodes but are also sensitive to
local factors, especially locaUsed positive and negative externality effects. In urban
simulations therefore it is important to encode global effects into transition rules in a
manner that reflects the decisionmaker's trade-off with local effects. Some global
factors have a compulsory effect, like zoning which in some planning regimes restricts
the possibility of development. In other regimes zoning will not be as deterministic and
may form the basis for negotiating community benefits. Nonrestrictive factors such as
accessibility to transport corridors and agglomeration attraction play their role by
reducing or increasing the probability of development according to the parameters
coded into transition rules. In the following simulations, the trade-off between global
and local effects is made through a simulated decisionmaking process using MCE.
Formally, the development of land can be thought of as being governed by a
location-specific profit function of developers [equation (1)]. Where there are many
types of urban land use, the development outcome will result from competition between
uses (and developers). In our simulation we assume a single urban use (m represents
urban or nonurban in the following expressions).

where
II f^ is the profit to developer derived from land-use m at location if;
Lfj is a vector of local neighbourhood positive and negative externalities;
Gy is a vector of global positive and negative externalities;
Ktj is a vector of site-specific and property-specific attributes;
106 F W u , C J Webster

X is a vector of private factor costs;


P is a vector of prices for properties developed;
Rijm is a vector of government regulation affecting use m at location ij;
Iijm is a vector of transfer payments to and from government (compensation and
betterment) arising from the use of location ij for land use m.
Both R and / are functions of the community's utility of use m at location ij, U^:
I = l(U^), R = R(U^), U^ = U^(/>, Cm, Bijm)
where CiJm and Bijm are the vectors of the social costs and benefits arising from the use
of location ij for use m. Developers are therefore assumed to have a certain preference
for land at location ij that relates to its profitability for a particular use. Depending on
how the community values the products of the developer and the related social costs
and benefits, it may impose regulations limiting or forbidding development, make
exactions, or pay the developer inducements. Alternatively, it may do none of these
and let the market take its course. It may also choose to influence the actions of
developers by investing in transport and other infrastructure in which case it can be
thought of as having an indirect effect on the developers' profit function by changing
the values in the L{j and Gtj vectors. We abstract from this model to specify a much
simpler set of transition rules [equation (2)]. If the simulation is being governed by a
developer, the weighting of factors in the rules can be assumed to reflect the behaviour
expressed more fully in equation (1). The rules determine the likelihood of a location's
transition to urban land use, an abstraction of the location's relative profitability.
Developers of different categories of urban land will have different profit functions
and will weight transition rules differently. In principle our M C E - C A framework
allows a multiuse gaming scenario in which different users evaluate a region and
compete against each other. The competition could be between uses after complete
simulations for each use have been run or, more interestingly, introduced iteratively
and interactively as a single simulation unfolds. If the simulation is being governed by
a community decisionmaker, the transition rules can be thought of as reflecting the
community's view of development suitability. In this case the transition rules are
abstractions of equation (1) with social costs and benefits, C and B, taken into account,
and R, I, G, and L adjusted to minimise social costs or maximise social benefits.
A community simulator could of course produce scenarios for different uses. The
community-run simulation could also be used in a positive rather than a normative
mode in which case the community simulator could weight transition rules in a way
that he or she thinks reflects the developer's behaviour. If the development factors were
sufficiently detailed and discriminating, this might show interesting differences in
perceived and actual developer behaviour—an important issue for the success of plan
implementation.

3 Definition of transition rules


The cellular automative land development simulation is achieved through iterative
looping of rules. The state of land use at t + 1 is determined by the state of land and
the development situation of its neighbourhood at time t in accordance with a set of
transition rules:
s++l = f ( « ; , ^ , r ) , (2)
where
Sjj+1 and Sy are the states of land use at location ij at the time t + 1 and /, respectively;
Qjj is the development situation in the neighbourhood space of the location ij;
T1 is a set of transition rules.
Simulation of land development 107

The equation indicates that, in a self-organising city, land development is an


historically dependent process during which development in the past affects the future
through interactions among land parcels. The interaction among developments is
captured in simulation through a moving kernel with a 3 x 3 pixel configuration that
defines the local neighbourhood. The kernel is applied to each cell and returns a value
indicating the proportion of the neighbouring eight cells that are developed in state S».
This local and dynamic information is summed with a variety of global variables to
produce an additive evaluation score which is then transformed into a probability that
determines the state of location ij at time t -f 1. With the locally derived information
added to the set of MCE criteria, the specification of the transition function becomes
similar to those of land evaluation (Chapin and Kaiser, 1979; Hopkins, 1977; Xiang
and Whitley, 1994). Some of the global variables are derived from GIS operations, for
example, an accessibility buffer around a highway that is exported as a constraining
development factor into the simulation process.
The probabilistic approach allows a flexible translation from context and local
dependencies into the rules of transition. As a nondeterministic CA, the state at time
t + 1 now becomes:
S^=f(pis,T'), (3)
where
pijs is the probability of transition to the state s at the location ij,
Pis = tfW,) = $[<*&}<, Wk)]; (4)
rijs rates the suitability of conversion s at location ij;
Fijk is the score of development factor k at location //, including the proportion of
neighbouring cells developed at state t;
Wk is the vector of weights that give relative importance to each development factor;
co is a combination function which sums the scores of the weighted development
factors;
<>
| is the function which translates the composite suitability score into the probability.
Where constraint factors take a binary value of 0 or 1, representing absolute
restrictions or not, the specification of co becomes:
/ m \ n
r F w F
» = \j2 «* *) n «*» ^
\k= 1 / k = m+l
where
Fijsk is the score of development factor k for a change s at location ij;
1 ^ k ^ m represents nonrestrictive factors of development;
m < k ^ n represents restrictive factors of development.
Among development factors are the state of land use, S^ and the neighbourhood
situation Qjj. Development factors can be grouped into static and dynamic factors in
terms of whether they are changing along with the simulation. Static factors are
constant throughout the simulation and these reflect the global configuration of the
cellular space, for example, the exogenously defined city centre, distance from it, and so
on. Dynamic factors, on the other hand, change during the simulation, reflecting the
consequence of transitions, for subsequent states. For example, the factor for local
intensity of development is updated whenever a neighbouring cell is developed as urban
land. A hierarchy can be constructed according to the role of various development
factors so as to mimic different approaches to decisionmaking, for example, a hierarchy
with the groups of static factors at its root or one starting with dynamic factors. In
behavioural terms static factors might be associated with greater certainty and dynamic
108 F W u , C J Webster

factors with greater risk. Where they sit in a hierarchical decision process will then
reflect the decisionmaker's attitude towards risk as well as his or her relative weighting
of factors. We do not explore this interesting area further in the present experiment.
Following most urban CA simulations we do not consider competing land uses and
therefore define the state of land as a binary case, where s^ = 0 if land is vacant (most
agricultural land), and stj = 1 if land is urban. Thus, there is only one probability of
land conversion, that is, from vacant land to urban land. The probability of develop-
ment p^ is shown in a matrix of land conversion as:

s<+1 = o s;+l = i

$} = 0 1 -ptj ptj
Sj} = l 0 1

For simplicity, a reversed order of development, that is, conversion from urban land to
vacant land is not allowed in this model. This can be relaxed in the future if it is
assumed that the probability of redevelopment p~ increases monotonically with the age
of the last development, that is,
p!j = f ( r „ * , ) , (6)
where
plj is the probability of redevelopment;
Xy is a development factor describing the time from last development {X(j is the
current time less the time of previously recorded development).
There are many ways of defining probability of urban transition in a nondeterministic
CA. Deterministic site-selection studies that choose the highest-suitability location are
only a special case in a complex location - allocation phenomenon. In such cases,
f 1 if ru — maximum (r») ,_>.
I 0 otherwise
Whether deterministic or nondeterministic, the site-selection type of transition rule
that we adopt here focuses only on a single land use and cannot be used to consider
competing land uses without a major extension of the algorithms. In a multiuse
problem, the best site for a particular use or a site with a high probability of conversion
to that use may not convert for reasons of competition rather than suitability; other
equally suitable uses may outbid the use in question. The location may even be less
suitable for the use with higher bidding power but nevertheless convert to that use
because of a limited set of alternative locations (perhaps itself a result of outbidding by
uses with even greater bidding power). We leave further discussion of the competing
land-use issue to the conclusion where we consider the possibilities of defining multiple
use probabilities within the additive factor score of any particular use as well as
alternative approaches such as iterating independent simulations for competing uses,
simulating competition outside of the MCE process.
Dependent on the context, there are many ways of translating an evaluation score
into a development probability. A predetermined lookup table may, for example, be
used to relate the chance of development to categories of suitability, such as highly
suitable, moderately suitable, marginally suitable, and not suitable. Similarly, probabil-
ity can be defined with reference to an ideal site or the best site (figure 1). The ideal site
is the site that has the highest possible scores over all criteria:
Pu=Hrv) = «^L, (8)
Simulation of land development 109

where
a is a dispersion parameter ranging from 0 to 1;
rtj is the composite evaluation score at location ij;
rldeal is the composite evaluation score at the ideal situation.
The ideal site where all criteria reach the highest score might not actually exist in
the geographical and temporal configuration of the automaton and, dependent on the
simulation at time t, the highest score, rmax, might be less than the ideal score, rldeal,
1 1 ea
(rl ^ r 1 ). The probability derived from the division of rtj by r would normally
be sufficiently small to render the site-selection process a stringent one. If rmax is used,
the range of probabilities is reduced and the probability surface becomes much less
discriminating: many more sites become candidates for transition and the simulation is
less influenced by the land evaluation criteria. This suggests that a nonlinear trans-
formation is required to depress the probability away from rmax in order to achieve
greater discrimination between cells in any one simulation, such as

v (9)
Pv = exp .max

where a is the dispersion parameter, a = 1,..., 10.


The value of the dispersion parameter governs the stringency of site selection, with
a higher value reflecting a more stringent selection process. A site with an evaluation
score equal to 80% of the best site, for example, has a development probability of 0.82
when a = 1; 0.37 when a = 5; and 0.14 when a = 10. The parameter value has a
strong influence on the overall pattern of land development since the development of
less desirable sites under a simulation with a low dispersion value will result in a more
arbitrary pattern, less influenced by the development factors.

Transformation based
on the ideal site score

a = 1

a< 1

Maximum rating score at time / Ideal site


Suitability, rtj
Figure 1. Alternative transformations of the evaluation score to development probabilities.

4 Definition of factor weights


After the transition rule in a multicriteria form is defined, the acquisition of factor
weights becomes the next step. In our experiment we have used an analytical hierarchy
process (AHP) originally developed by Saaty (1980). The AHP uses pairwise compar-
isons to acquire the preference of decisionmakers. It is argued that in real-world
situations decisions are made in a hierarchy descending from general criteria down to
subcriteria and that judgements can be made most effectively by comparing only a pair
110 F W u , C J Webster

of criteria at a time. The comparison uses a nine-point scale to measure the preference
in a pair of criteria. A matrix A is defined as:

A = («,) = ( 2 * ) , (10)

where n> is the weighting in vector W9 W = (H> ) r , and ranges from 1 to 9; and atj thus
ranges from 1/9 to 9.
It is obvious that atj > 0, atj = \/ajh and au = 1. Saaty (1980) observed that the
principal right eigenvector of A, when normalised, reflects the priorities of preference
of a decisionmaker:

AW = (—\wt)T = l^W. (11)

In a consistent situation where atj x ajk = aik, the principal eigenvalue of A is equal to
the dimension of A, that is, /lmax = n. To check the consistency Saaty (1980) proposed
a consistency index / con ,
;max n
/« = ^ _ ^ (12)
n— 1
where n is the scale of A, and Amax ^ n. Thus, the more consistent A is, the smaller the
z'con value. In a consistent environment, zcon equals 0. Following Saaty (1980) and Saaty
and Kearns (1985) we use these pairwise comparisons to recover a weight vector by
which suitability of the land can be computed. By incorporation of the local kernel-
derived neighbourhood configuration into the evaluation function and use of an
iterative simulation to update the evaluation over time, the land suitability becomes
dynamic. The evaluation of land at any locality changes at each iteration as a
consequence of neighbouring development decisions in the past and, in turn, affects
future decisions about neighbouring land development in subsequent iterations. In this
way, the AHP method is taken to depict the process of decisionmaking dominating the
simulated development.

5 Implementation of the integrated simulation


The integrated simulation scheme is implemented by three modules: MCE, CA, and
GIS (figure 2). A strategy of 'tight' integration is adopted here, which means the first
two are built into the GIS so that AHP and CA are coded in the C programming
language and called from within the ARC/INFO GIS (ESRI Inc., Redlands, CA). An
alternative would have been to program the AHP and CA routines in a macro
language such as AML in ARC/INFO. The advantage of the latter strategy would be
the ability to use the cell-based functions of GRID, the raster module of ARC/INFO.
Set against this is the time inefficiency of implementing the highly iterative CA within
the cumbersome framework of a high-level macro programming language. For this
reason we decided on the external C module approach, accepting the loss of display
and data transfer functionality.
Within this implementation environment the GIS serves as a spatial database
which stores land use, constraints, infrastructure and related information for the city
that comes from satellite-derived digital data, plans, and topographic maps. The data
are processed by using a variety of GIS functions to generate indicators for each of the
global development factors and the processed data are exported into ASCII files which
can be read by the CA module. The results of the simulation are returned to the GIS
for storage and display or for the starting scenario in subsequent simulations. Breaking
down a simulation into a number of stages is a useful procedure, allowing the results of
Simulation of land development 111

Satellite data Planning and Socioeconomic


image processing topography maps context

MCE
(for example, AHP; written in C) CA (written in Q GIS (ARC/INRO GRID)

Identification of Projection Measurement of


development factors! of demand development factors

Construction of the Neighbourhood Acquisition of initial


analytical hierarchy configuration state of land use

Identification of
Pairwise comparison Transition rules development
constraints

Recovery of weightings, Probability of


measure of consistency land conversion

Monte Carlo
simulation

'Land \ Updating: », f+1


demand
, satisfied^ No

Graphic display
and comparison

Export into GIS


Refresh understanding for scenario library

Figure 2. Procedure for simulating land conversion by integrating cellular automata (CA) and
multicriteria evaluation (MCE).
one stage to be viewed and interpreted before submitting them to the CA for further
iteration. This also allows the decisionmakers running the simulation to restructure
preferences and recompute factor weights, leading to a modification of the transition
rules as the phased simulation proceeds. We do not explore this idea any further but
note that the process described is consistent with the idea of periodic (rather than
continuous) monitoring of development by strategic planning authorities and associated
adjustment of the balance between local - global, tactical-strategic, and state-market
influences in the development process.
112 F Wu, C J Webster

Functionally, the integration works as follows. The AHP procedure passes through
four stages: presenting the development factors; constructing the decisionmaking
hierarchy; pairwise comparison of development factors; and the recovery of weightings
and measurement of consistency. Pairwise comparison is assisted by a graphic user
interface (GUI) which reminds the user about the relative importance of a pair of
development factors. The recovered weights are then used in the calculation of transition
probability in the CA routine.
The cellular automaton begins with the arbitrary specification of land demand,
which is recorded as a simple urban cell count and governs the limit of simulated
growth. As there is no specific theory guiding the configuration of a neighbourhood,
a 3 x 3 neighbourhood is adopted giving any particular location eight adjacent cells.
An alternative configuration can easily be defined by users through a GUI which offers
an interactively defined neighbourhood template. Transition rules are defined, develop-
ment factor scores are processed by the GIS, and the probability of land conversion is
calculated from the AHP module. A Monte Carlo process randomly scans the cellular
space, constantly subjecting cells to a conversion examination. The calculated prob-
ability for a cell is compared with a random number that has the same range as the
probability in order to decide whether the transition is successful or not. If the random
number is smaller than the probability, the cell is converted; if not, another cell is
chosen randomly. The Monte Carlo algorithm may be taken to imply that developers
respond to land supply opportunities that follow a random pattern and that they
consider the local and global suitability of opportunities as they arise. When the
number of successful allocations exceeds a preset limit, the neighbourhood is updated
and the probabilities of development recalculated. During the process, the demand for
land is checked. If the demand has been satisfied, the process is terminated. The
iteration number of this simulation equals the number of neighbourhood updates,
that is, the total projected number of developments divided by the updating period
(in cells). When the neighbourhood is updated with every successful allocation, the
iteration number equals the projected demand. It should be noted that simulation
time has no necessary correspondence to real time (Cecchini, 1996). However, the
land demand parameter allows a certain degree of mapping into real time. The output
is either a final simulated surface indicating the extent of urban development or a series
of intermediate surfaces if the simulation has been partitioned to examine visually the
process of urban growth.

6 Application to land development in Guangzhou


Guangzhou, the largest city in southern China, is chosen for an experimental applica-
tion of our simulation methodology. Since land reform in 1987, the city has experienced
rapid growth driven by a potent mix of inward investment, indigenous manufacturing
growth, and leasehold development of government land. Encroachment of agricultural
land has become an alarming problem and prereform urban planning has arguably not
yet adjusted to the new market-led conditions of growth. Municipal authorities are
largely in league with these processes, following a strategy of rapid growth which aims
to transform the city into a global city in southern China. The huge volume of projected
land demand and the threat to peripheral agriculture means that there is a desperate
need to understand the processes of expansion. Our study selects an area of 225 km2
(15 km x 15 km) in the eastern suburbs for experimental C A - M C E simulation. Aerial
photography shows that the area was virtually all vacant (agricultural) land before
reform, with the exception of a few villages and institutions. In the 1980s a new
suburban centre was built in the newly established Tianhe district centred around a
new railway station which became the new hub of growth. The 1990s saw further urban
Simulation of land development 113

sprawl following, among other things, the early 'colonising' inwardly investing plants,
international hotels, and the government's zonal inducement policies. Two digital
satellite images (Landsat TM-5) for 1988 and 1993 were processed using PCI software
(PCI Corporation, Toronto, Canada) and a supervised classification made, based on
land-use information identified from 1:10 000 aerial photographs of the same period.
Because of the significant difference in their spectral signatures, urban and nonurban
land uses can be distinguished with a reasonable degree of confidence at a coarse level
of classification. Land uses were classified into three types: urban built-up area; forest
and agricultural land; and rivers and other water bodies. The classified land-use maps
are shown in figure 3 (gz88s represents the land use for 1988 and gz93s for 1993). This
gives us an observed starting point and an end point as reference distributions for our
simulations, the end point offering the possibility of calibration, although this proves
problematic, as is discussed later. In addition a virgin state of development is created
by reclassification to support simulations that start from the point of no existing
development (state_0 in figure 3).

state_0 gz88s gz93s


Figure 3. Three initial simulation states: virgin land, 1988 (satellite derived), and 1993 (satellite
derived).
The choice of development factors for the simulation was informed by a recent
study of the determinants of land development in Guangzhou (Wu, 1995) in which the
relationship between development probability and various explanatory variables depic-
ting characteristics of a plot of land is calibrated through a series of logistic regressions.
Factors selected include cost of travelling the distance to the city centre (CENTRE); cost
of travelling the distance to the major industrial district (INDUSTRL); access to the
newly built railway station in the eastern subcentre (NEWRAILS); access of the highway
(HIGHWAY); development intensity in the neighbourhood (NEIGHBOR); and topographic
and regulative restrictions on land development (RESTRICT). The neighbourhood factor
captures in the simulation the effects of agglomeration such as positive externalities,
use-complimentary economies of scale in infrastructure supply, and so on. There is
much scope for experimenting with alternative neighbourhood definitions within the
CA to capture particular externality effects, but here we confine ourselves to a single,
generally conceived, agglomeration factor. The restriction factor is used to exclude
from conversion those areas that are nondevelopable owing to topographical barriers
or zoning as well as those areas that have been developed in previous iterations.
Redevelopment is not allowed in this simulation but this could in principle be relaxed
by defining the redevelopment probability as a function of the elapsing of simulation
time. These factors are measured in a raster format in GRID through GIS operations.
114 F Wu, C J Webster

The composite evaluation score is computed from the following expression:


rtj = (ft CENTRE + &INDUSTRL + fc NEWRAILS + fa HIGHWAY
+ j85NEIGHBOR)RESTRICT, (13)
where fa,..., fa are weighting parameters acquired from AHP; CENTRE, INDUSTRL,
NEWRAILS, HIGHWAY, NEIGHBOR, RESTRICT are development factors.
Although the CA program, using dynamic memory allocation, could handle a full-
size version of the cellular automaton based on the 28.5 m x 28.5 m pixel size of
Landsat TM, a reduced resolution (74 x 74 cells) cellular space was defined in order
to save computation time. Each cell represents a 200 m x 200 m area on the ground and
the images processed to the start and end points were resampled to this resolution.
The probability of transition is defined through the comparison of the composite
evaluation score and that of the most suitable cell at the simulation time following
equation (9) and, as discussed, the value represents the suitability of the site with
reference to some maximum suitability. The higher the value of the parameter a, the
more stringent the land development, with the value to some extent reflecting the
efficiency of the flow of information about land suitability. Large a values imply that
developers are more likely to search many sites until they find a satisfactory one.
Alternatively, it implies that planners have good information about suitable sites and
strong powers of development control. The most suitable site might not be selected
owing to the limited information and/or time and less desirable sites therefore still have
a chance of being developed. The more stringent the selection process is, the lower the
probability of development of these less suitable sites.
To test the effect of site selection, four simulations were run, with the parameter a
set equal to 1, 3, 5, and 10, respectively. Development factor weightings are kept the
same for these simulations. In the hierarchy of evaluation, the neighbourhood effect
shares equal importance with the first four factors and, within these four factors,
weights are divided equally again. The results of the simulations are shown in
figure 4. When a = 1, although some clusters can be seen, developments are generally
widely scattered. Along with the increase in the value of a, noise has been removed and
a general pattern has been formed. When a = 10, developments are concentrated on
the area with the highest composite score for development suitability: the area sur-
rounding the new railway station in the eastern subcentre. In each of the following
simulations, which test the impact of alternative decisionmaker preferences, the value
of a is set as 5.
Four schemes have been tested to simulate the expansion of urban built-up area in
Guangzhou under different assumptions. The first two start from a virgin state with an
estimated amount of land development ending in 1988. The third and fourth schemes
start from the land use classified from the 1988 digital image of Landsat TM-5. The
amount of land development to be simulated (the demand parameter) is obtained from
the comparison between 1988 and 1993 images.
The weights of preference in land development are obtained through AHP. Pairwise
comparisons lead to a series of matrices, A, for different simulation schemes (table 1).
A two-level hierarchy is adopted. At the top of the hierarchy the decisionmaker
controlling the simulation weights global factors in relation to local factors (how
important are citywide site attributes compared with the local agglomeration effect).
At the bottom, the first four development factors in expression (13) are compared and
weighted. This structure reflects an assumption that developers or development con-
trollers first make some sort of trade-off between localised and citywide locational
factors: how do they weigh the importance of existing neighbourhood development
Simulation of land development 115

Figure 4. Effect of dispersion parameter a on simulated land development (from a virgin start
point).
Table 1. Pairwise comparisons between development factors in four simulation schemes.

CENTRE INDUSTRL NEWRAILS HIGHWAY

(1) Centre-dominated growth (global: local weighting is 50: 50)


CENTRE 1 9 7 9
INDUSTRL 1/9 1 1/3 7
NEWRAILS 1/7 3 1 9
HIGHWAY 1/9 1/7 1/9 1
(2) Centre plus industrial district growth (global: local weighting is 60:40)
CENTRE 1 2 5 9
INDUSTRL 1/2 1 5 9
NEWRAILS 1/5 1/5 1 7
HIGHWAY 1/9 1/9 1/7 1
(3) Compact growth around the new suburban railway station (global: local weighting is 60:40)
CENTRE 1 1/3 1/5 3
INDUSTRL 3 1 1/3 3
NEWRAILS 5 3 1 5
HIGHWAY 1/3 1/3 1/5 1
(4) Highway-promoted growth (global: local weighting is 80:20)
CENTRE 1 1/3 5 1/9
INDUSTRL 3 1 7 1/9
NEWRAILS 1/5 1/7 1 1/9
HIGHWAY 9 9 9 1
116 F Wu, C J Webster

(indicating the existence of established demand for that location and related positive
externality effects) against a citywide search for the best site irrespective of the level
of existing development in the vicinity of that site? A weighting strongly in favour of
global factors might therefore be taken to represent the preferences of a developer who
is relatively insensitive to the behaviour of other developers: large developments for
which accessibility to major industrial and commercial centres and transport corridors
is more important than whether a locality is already converted to urban use or not.
A simulation under the opposite weighting might conversely be taken to represent the
preferences of developers for whom the existence of other development in the neigh-
bourhood is relatively more important in decisionmaking than a systematic search
within the city for the most accessible site. The difference captured at this level of the
hierarchy therefore captures something of the developers' preference for systematic
search versus heuristic search, general accessibility versus specific accessibility, developed
neighbourhoods versus undeveloped neighbourhoods, as well as their attitude towards
market leading versus market following, risk, and their resources for site searching.
Behaviour in respect of risk and search time or resources is also captured in the value
of the dispersion parameter a, as already discussed.
The four simulation schemes are intended to illustrate how different considerations
in the land development process lead to different spatial patterns. They are abstractions
or caricatures of the development regimes operating in different periods of Guangzhou's
recent history, and MCE is used to capture the different blends of government and
private developer preferences that governed those regimes. A summary of the four
schemes is presented in table 2.
Table 2. Main features of the four simulation schemes for Guangzhou.
Simulation scheme
centre- centre plus compact growth highway-
dominated industrial district around the new promoted
growth growth suburban railway growth
station
Starting state virgin land virgin land 1988 land use 1988 land use
Projected land 6032 6032 1804 1804
development
(hectares)
CENTRE 0.325 0.294 0.084 0.080
(weight)
INDUSTRL 0.060 0.204 0.204 0.144
(weight)
NEWRAILS 0.100 0.078 0.324 0.032
(weight)
HIGHWAY 0.020 0.024 0.048 0.536
(weight)
NEIGHBOR 0.500 0.400 0.400 0.200
(weight)
RESTRICT 1 if vacant, 1 if vacant, 1 if vacant, 1 if vacant,
(value) 0 otherwise; 0 otherwise; 0 otherwise; 0 otherwise;
Consistency 0.212 0.106 0.068 0.189
index
Simulation result figure 5 figure 6 figure 7 figure 8
Simulation of land development 117

The first simulation—centre-dominated growth—emphasises the role of the city


centre. This represents the approach of municipal planners in a centrally planned
economy where investments in infrastructure have been constrained owing to the
ideological thinking that such expenditure is a nonproductive resource allocation
(in the period 1949-79). Development projects in such a regime have to rely on the
city centre for services and infrastructure provision and a radial public transport is the
dominant travel mode, which limits the commuting distance. As a result, the expansion
of the city is constrained. The preferences being captured in the MCE stage are
assumed to be principally those of the government which exerts strong control over
the location of urban developments.
The result of the simulation is shown in figure 5. As the iteration time has no
definite relation with the year, land uses are presented as a series of intermediate states.
In this case, seven states are shown. At state tl one seventh of total land demand has
been fulfilled, and at state t7 total demand has been satisfied. In this simulation, with
each successful development, the neighbourhood is updated. Thus, the number of
iteration equals the amount of successful land developments counted by the number
of cells.

Figure 5. The results of the first simulation scheme.


118 F Wu, C J Webster

Figure 6. The results of the second simulation scheme.


The second scheme is a refinement of the first, giving a higher preference to access
to the major industrial district on the southern bank of the Pearl River. Under the first
simulation, the city grows in all directions, a result that is inefficient in terms of
secondary infrastructure investment. In the period 1949 - 79 there was a belief that,
through industrial planning, agglomerative effects would be achieved which would
reduce the cost of infrastructure provision. The scheme therefore attempts to simulate
so-called 'socialist industrial allocation principles' which encouraged industries to be
located near industrial districts. The MCE preferences are, as in simulation 1, principally
those of the government which controls industrial as well as residential development.
The difference between the two schemes becomes obvious by the fourth stage
(figure 6). In the first simulation, development generally concentrated on the northern
bank of the Pearl River, whereas in the second scheme the growth has begun to shift to
the southern bank. This subtle change leads to a more realistic simulation of the actual
development depicted by the 1988 digital image. The centre-dominated growth over-
estimates the growth in the northern area of the city and underestimates the growth in
the southern area. Simulation 2 captures better the growth in the southern area. There
is, however, still a significant gap between the simulated scenario and the observed
Simulation of land development 119

development pattern. This would clearly be improved if more development factors were
introduced, such as distance to the port-oriented industrial district along the Pearl
River or the topographic constraints of the Baiyun Mountain in the northern area of
the city.
The third scheme simulates the situation that held in the 1980s. Local government
in Guangzhou tried to capture investments to build a subcentre in an area that was
suburban in the early 1980s. The developments included a great volume of municipally
managed housing construction and a new railway station. The fourth scheme tries
to reflect very recent changes in land development in the city. Foreign investment,
particularly from Hong Kong, has given rise to an intensive development of intercity
highways. Combined with a new land-leasing system (introduced in 1987) that enables
local governments to lease land to foreign investors, this has led to a very rapid phase
of urban expansion following the highways. The MCE weightings in this simulation
therefore principally represent the locational preferences of private developers.
The third and fourth simulations ascribe different levels of importance to access to
the new railway station and to highways, respectively (figures 7 and 8). An additional
difference is the global: local ratio which is weighted more highly in favour of global

Figure 7. The results of the third simulation scheme.


120 F Wu, C J Webster

t7
Figure 8. The results of the fourth simulation scheme.
factors in simulation 4. This is intended to reflect the fact that the big foreign investors
are strategic in their thinking and place great weight on car-borne accessibility, with
less importance given to the existing state of neighbouring land. These investors are
strategic in perspective, perhaps more so than the government under simulations 1 and 2
which was concerned mostly with infrastructure costs—hence the greater weighting
given to local factors. The private investors are demand shapers, creating new markets
in accordance with their view of how the liberalised land market will develop.
The difference in results for simulations 3 and 4 is subtle. The subcentre growth
produced a more compact form of growth around the new railway station as expected,
whereas the highway scheme promotes a more dispersed urban pattern. However, the
number of land developments is not large enough to produce significantly different
scenarios. Therefore, to examine the consequences of the two strategies more clearly,
simulations are rerun with a hypothetical projection of land development of 6000
hectares by using the same weights. The simulations produce two significantly different
scenarios (figure 9). This gives an indication of the difference between protracted
growth under the current regime of partnership between municipality and strategic
Simulation of land development 121

Figure 9. Extrapolation of urban growth (a) under simulation 3 and (b) under simulation 4.
foreign investors and the pattern that might have occurred under the alternative, public
transport - local government led, scenario.
However, it is entirely feasible that new rules of transition might be triggered off as
a kind of feedback to the current highway-dominated process, which may lead the
result away from the one shown in figure 9. The most likely scenario is that newly
formed subcentres will become dominant attractors once the settlement pattern has
reached a certain maturity threshold. The relatively undifferentiated attraction of the
highways to speculators will give rise to greater discernment as the centres of activity
take shape. From this point of view, the scenario shown in figure 8 represents a
colonising phase of urban development.

7 Goodness of fit
We conclude this section with a tentative discussion about goodness of fit. The
integration of MCE with CA allows simulations under various bundles of developer
preferences. Viewed as an exploratory tool—a method for exploring the generic form
of simulations governed by different processes—the issue of calibration does not
necessarily arise. The simulator may be viewed simply as an environment in which
alternative developer and planner preferences can be explored in relation to their
impact on the general physical evolution of the city. All parameters in the model may
be valued by the decisionmaker including the dispersion parameter (degree of search
effort), the global-local factor balance (importance of the strategic search for the best
site irrespective of other developers' actions), the choice of factors, and the weighting of
factors. Exploration may be carried out for broadly explanatory or predictive purposes.
It may, for example, be used to predict the generic pattern of market-led growth
(transition rules reflect the behaviour of private developers) or regulated growth under
different regulatory regimes if the rules reflect the preference of planners. As an
explanatory device, simulations may seek to discover the locational preferences that
best explain an observed pattern of growth. The four simulations reported above
illustrate both purposes. On the one hand, we have presented them to illustrate how
generic predictions can be made under a series of alternative models of locational
behaviour. On the other hand, we have loosely sought to test the hypothesis that
Guangzhou's recent development can be explained by the dominance of different
planning or development regimes in different periods. Explanatory applications like
this demand some attention to the issue of goodness of fit, the purpose of which is to
test the hypothesis that a particular decision model has dominated the evolution of city
form during the period in question. We do not explore this issue thoroughly but make a
number of observations and present the results from three possible calibration
approaches.
122 F Wu, C J Webster

It should first be noted that CA models are not well suited to the prediction of land
uses for individual sites. They assume the city to be, at least in part, a self-organising
system, in which case the use of particular sites is unpredictable: they predict the shape
of the total system under probabilistic local (and in our case jointly influential global)
transformation rules. Each simulated development can be viewed as a unique historical
process with a different microscopic spatial pattern compared with other simulations
run under the same assumptions. Nonetheless, some general features will emerge from
similar simulations which are related to the transition rules of the model. These
features can be treated as signatures of the simulation. If appropriate signatures are
compared between simulated and observed total system organisation we can say some-
thing about goodness of fit. To compare different land-use patterns, therefore, we need
to compare the signatures of land uses rather than the individual site-specific land uses
themselves. A good simulation is one that captures some important features of the
land-use pattern; some discriminating characteristics of urban form. This is essentially
a question of pattern recognition.
The measurement and description of urban form has long been the concern of
urban morphologists. Quantification of urban form by the use of various types of
signature has been approached from a number of directions including analysis of the
fractal dimension and, more generally, regularity of cities (Batty and Kim, 1992; Batty
and Longley, 1994) and pattern recognition in urban imagery (Webster, 1996). At one
level, a visual comparison between simulated and observed cities may offer some degree
of calibrative judgement. This is the approach taken in our discussion above. The human
brain is far more sophisticated in capturing the totality of similarities and differences in
two images than any contemporary computer vision algorithm. However, reductionist
approaches are necessary if we want to move towards automation and are desirable for
consistent and reliable methodology. The confusion matrix, used to evaluate classifica-
tion accuracy in remote sensing, is one method for comparing two land-use maps. We
present the results of a confusion matrix showing the accuracy of the four simulations
in predicting the observed urban form of 1988 (table 3). Goodness of fit ranges from
Table 3. Confusion matrix of the actual and simulated land uses.

Actual Simulated Sum


vacant urban

Land use of 1988 with the first scenario (t = 7)


Vacant 2791 795 3586
Urban 798 710 1508
Sum 3589 1505 5094
Land use of 1988 with the second scenario (t = 7)
Vacant 2843 743 3586
Urban 746 762 1508
Sum 3589 1505 5094
Land use of 1993 with the third scenario (/ = 7)
Vacant 2632 435 3067
Urban 508 1475 1983
Sum 3140 1910 5050
Land use of 1993 with the fourth scenario (t = 7)
Vacant 2549 518 3067
Urban 591 1392 1983
Sum 3140 1910 5050
Note: Percentages simulated correctly in the first, second, third, and fourth scenarios are
68.9%, 70.7%, 81%, and 78%, respectively.
Simulation of land development 123

69% to 81% correctly classified. The most accurate simulation in this sense is number 3
(socialist-industrial allocation) which gave 81%> of cells the same urban-nonurban
classification as the observed classification for 1988. This measure is flawed in an
important way, however. It is possible to imagine a sparsely populated area of non-
adjacent urban cells in which the geometric pattern of simulated development is
precisely the same as the observed, but dislocated by one cell. The confusion matrix
might at the extreme show a 0% classification accuracy which would be true in respect
of individual cells but misleading as a measure of accuracy in capturing pattern. This
implies a need for more structural measures that capture morphological regularity in
groups of cells not just information about isolated cells (albeit aggregated).
We experiment with two such signatures: the distribution of development across
concentric zones around the city centre and across concentric zones around the main
industrial district. The results are presented as profiles and indicate that simulations 1
and 2 both overestimate land development near to the city and underestimate a minor
industrial area around 12 km to the city [see figure 10(a)]. Looking at the concentric
zones around the industrial district [see figure 10(b)], we find that both simulations
show differences with the actual land-use pattern but simulation 2 performs better
(before and after the peak around 6 km from the district, the curve S2 is closer to
curve A). In more complex gaming applications of our methodology, morphological
signatures might be used to assess the differences between alternative generic forms

(a) Distance to city centre

0 2 4 6 8 10 12 14 16
Distance to industrial district
Figure 10. Cross-sectional density profile used as a morphological goodness-of-fit signature:
(a) profile measured from the city centre; (b) profile measured from the industrial district
[note: A, observed profile (from satellite imagery); SI, simulated profile (simulation 1);
S2, simulated profile (simulation 2)].
124 F Wu, C J Webster

produced under different decision rules and market-planning scenarios. Our interest
here has been only to indicate the kind of measures that may be used.

8 Conclusion
The CA approach opens up a way of articulating global patterns through local processes
(Batty and Xie, 1994). Our study pursues the question of how to define microscopic
transition rules in a more realistic way, bearing in mind that the transition of nonurban
space into cities depends on global as well as local processes. We have chosen to
capture global processes and the relationships between global and local by using a
decision model rather than by systematic analysis of parameter changes. This has the
advantage of rendering the methodology suitable for gaming simulation but the dis-
advantage that the precise cause-effect relationships between individual weights and
outcomes are not necessarily clear. The latter need not matter if the object is primarily
to explore the urban pattern generated by operators in different development regimes.
The weightings generated by the MCE algorithms may be explored more systematically
if the interest is in understanding more precisely how factor weights contribute to
outcomes. In this sense, our methodology offers a higher level of analysis that may
precede a systematic analysis of weights and outcomes. We feel that an MCE with a
graphical interface for parameter definition coupled with a CA with a graphical kernel-
definition interface provides a powerful tool for capturing decisionmaking behaviour
to govern a simulation (Webster, 1994). Models are only as good as their representa-
tions, however, and more useful simulations require a deeper representation of the
knowledge domain. Models that deal only in urban - nonurban land use will remain
general in their insights and limited in their ability to inform explanation. We plan to
experiment with simulations that deal in competing land uses and that use transition
rules that reflect bidding power, macro market equilibrium, planning goals, and local
externality costs and benefits (in the models reported above, only the benefits due to
the coarseness of land-use classification are considered).
Other approaches to the development of gaming CA models may also be profitably
considered. The calibration of statistical relationships between the probability of
development and the explanatory variables is a potentially powerful approach to
transition rule definition. The approach is limited if dependencies are captured from
observed data. More interesting might be the use of stated preference models to
calibrate transition rules, a statistical approach directly analogous to our AHP
approach. A problem with all statistical approaches, however, is their huge demand
on the data, not only because of the very disaggregated spatial form of CA models but
also because of the need to capture intertemporal factors in the model. In addition,
there are problems of spatial interaction between the data which can be overcome with
explicitly spatial statistical models but might render model calibration excessively
difficult or costly owing to very large spatial association matrices for CA space.
Exploring behavioural rules from global patterns is another direction for research.
Contrary to forward-processing CA models in which the rules are given and the aim is
to investigate the stages in the development of a pattern, backward-processing CA
models seek to identify transition rules on the basis of known stages of development
of a spatial pattern (Portugali et al, 1994). Information about development patterns is
now generally much more available owing to remote sensing techniques. The develop-
ment of these backward-processing simulation models would be an innovative use of
remote sensing information and would be helpful in monitoring land development and
in the enforcement of development control. The satellite images of Guangzhou that we
have used indicate massive urban sprawl in the suburban areas. Urban planning there
has not been able or willing to deal with the problem of encroachment of agricultural
Simulation of land development 125

land. A backward-processing C A would be helpful in understanding the dynamics of the


problem. A C A method that explores the explicit sequence of development has potential
advantages over static equilibria-focused economic models. Backward-processing CA
linked to our M C E method of exploring the sequential effect of different but interacting
decision regimes seem to offer potential in this respect.
Finally, on the issue of the use of C A for generic prediction and the related issue of
calibration, it is unlikely of course that behaviour governing past phases of development
will hold exactly for future phases. More important are the attitudes of different actors
in the development process towards the future and the interactions between these. The
scientific endeavour underlying the development of methods to support this kind
of predictive activity involves preference measurement and combination. M C E is one
approach to capturing preferences to govern a CA simulation. Another, we have
suggested, is via stated preference models. CA simulations may even be used to
calibrate M C E or other decision models for more sophisticated simulations. For
example, a number of developers may each be asked to run a city simulation, inter-
actively adjusting factor weightings until their preferred development outcome arises:
a process of self-referential calibration. These calibrated MCEs might then be combined
in a m u l t i - M C E - C A simulation in which the city is grown under various market-
planning strategies.
Acknowledgement. Landsat TM-5 data were kindly provided by Professor Anthony Yeh, Director
of the GIS Research Centre, University of Hong Kong.
References
Batty M, 1995, "New ways of looking at cities" Nature 377 574
Batty M, Xie Y, 1994, "From cells to cities" Environment and Planning B: Planning and Design 21
531-548
Batty M, Kim K S, 1992, "Form follows function: reformulating urban population density functions"
Urban Studies 29 1043 -1070
Batty M, Longley P A, 1994 Fractal Cities (Academic Press, London)
Cecchini A, 1996, "Urban modelling by means of cellular automata: generalised urban automata
with the help on-line (AUGH) model" Environment and Planning B: Planning and Design 23
721-732
Chapin F S Jr, Kaiser E J, 1979 Urban Land Use Planning (University of Illinois Press, Urbana, IL)
Couclelis H, 1985, "Cellular worlds: a framework for modelling micro-macro dynamics"
Environment and Planning A 17 585 - 596
Couclelis H, 1988, "Of mice and men: what rodent populations can teach us about complex spatial
dynamics" Environment and Planning A 20 99 -109
Deadman P, Brown R D, Gimblett H R, 1993, "Modelling rural residential settlement patterns
with cellular automata" Journal of Environmental Management 37 147 -160
Hopkins L D, 1977, "Methods of generating land suitability maps: a comparative evaluation"
Journal of the American Institute of Planners 43 386-400
Itami R, 1994, "Simulating spatial dynamics: cellular automata theory" Landscape and Urban
Planning 30 27-47
Phipps M, 1989, "Dynamic behaviour of cellular automata under the constraint of neighbourhood
coherence" Geographical Analysis 21197-215
Portugali J, Benenson 1,1995, "Artificial planning experience by means of a heuristic cell-space
model: simulating international migration in the urban process" Environment and Planning A
271647-1665
Portugali J, Benenson I, Omer 1,1994, "Sociospatial residential dynamics: stability and instability
within a self-organising city" Geographical Analysis 25 321 - 340
Saaty T L, 1980 The Analytical Hierarchy Process: Planning, Priority Setting, Resource Allocation
(McGraw-Hill, New York)
Saaty T L, Kearns K P, 1985 Analytical Planning: The Organisation of Systems (Pergamon Press,
Oxford)
Toffoli T, Margolus N, 1987 Cellular Automata Machines (MIT Press, Cambridge, MA)
126 F Wu, C J Webster

Webster C J, 1994, "GIS and the scientific inputs to planning. Part 2: prediction and prescription"
Environment and Planning B: Planning and Design 21145 -157
Webster C J, 1996, "Urban morphological fingerprints" Environment and Planning B: Planning
and Design 23 279 -297
White R, Engelen G, 1993, "Cellular automata and fractal urban form: a cellular modelling approach
to the evolution of urban land-use patterns" Environment and Planning A 25 1175-1189
Wolfram S, 1984, "Universality and complexity in cellular automata" Physica 10D 1 - 35
Wu F L, 1995 Changes in the Urban Spatial Structure of a Chinese City in the Midst of Economic
Reforms—A Case Study of Guangzhou PhD thesis, Centre for Urban Planning and
Environmental Management, The University of Hong Kong, Hong Kong
Xiang X-N, Whitley D L, 1994, "Weighting land suitability factors by the PLUS method"
Environment and Planning B: Planning and Design 21 273 - 304

p © 1998 a Pion publication printed in Great Britain

You might also like