Cities 30 (2013) 193–203

Contents lists available at SciVerse ScienceDirect

Cities
journal homepage: www.elsevier.com/locate/cities

Modeling urban evolution using neural networks, fuzzy logic and GIS: The case
of the Athens metropolitan area
George Grekousis ⇑, Panos Manetos 1, Yorgos N. Photis 1
Dept. of Planning and Regional Development, Adjunct Lecturer, University of Thessaly, Pedion Areos, Volos 38334, Greece

a r t i c l e i n f o a b s t r a c t

Article history: This paper presents an artificial intelligence approach integrated with geographical information systems
Received 20 October 2011 (GISs) for modeling urban evolution. Fuzzy logic and neural networks are used to provide a synthetic
Received in revised form 9 January 2012 spatiotemporal methodology for the analysis, prediction and interpretation of urban growth. The pro-
Accepted 29 March 2012
posed urban model takes into account the changes over time in population and building use patterns.
Available online 28 April 2012
A GIS is used for handling the spatial and temporal data, performing contingency analysis and mapping
the results. Spatial entities with similar characteristics are grouped together in clusters by the use of a
Keywords:
fuzzy c-means algorithm. Each cluster represents a specific level of urban growth and development. A
Urban growth
Fuzzy clustering
two-layer feed-forward multilayer perceptron artificial neural network is then used to predict urban
Neural networks growth. The model, applied to the prefecture of Attica, Greece, delineates the current and future evolution
Athens metropolitan area trends of the Athens metropolitan area, which are illustrated by maps of the urban growth dynamics. The
proposed methodology aims to assist planners and decision makers in gaining insight into the transition
from rural to urban.
Ó 2012 Elsevier Ltd. All rights reserved.

Introduction
Although the developed world and developing countries differ
in the percentage of people living in cities, as well as in the way
in which urbanization is occurring, there is a global trend of urban
population growth. The United Nations Population Fund has esti-
mated that by 2030 the urban population of the developing world
will have increased from 2.048 billion in 2000 to 3.991 billion,
while the urban population of the developed world is expected
to increase relatively little, from 870 million to 1.01 billion (United
Nations Population Fund, 2007). As a result, the spatial patterns of
urban areas, by means of their limits and borders, are changing in a
complex way. Urban areas continue to expand at the expense of
rural areas, intensifying among others, urban sprawl. Because of
consistent changes in their structure and shape, urban areas are
continuously at the epicenter of a wider scientific interest as the
problems arising are complex and exceed the pure urban aspect.
As contemporary cities are polycentric and avoid the classic
monocentric model of the past, understanding their evolution is
becoming constantly more complicated. The quantitative and spa-
tial demographic changes in urban regions are accompanied by
perpetual transformations that are manifested by urban sprawl
⇑ Corresponding author. Address: Pasxalias 15 Str, 15452 Psychico, Greece. Tel.:
+30 6939102064/2421074426.
E-mail addresses: geograik@gmail.com (G. Grekousis), pmanetos@prd.uth.gr
(P. Manetos), yphotis@prd.uth.gr (Y.N. Photis).
1
Tel.: +30 2421074426; fax: +30 2421074382.
and land use changes. Metropolitan areas have come under pres-
sure, and problems such as congestion and the loss of open space
are negatively affecting urban and transportation planning as well
as the quality of life and the environment (Waddell, 2000).
Modeling and simulation of complex dynamic systems such as
urban areas could greatly benefit from the synergy of the methods
and techniques of geographical information science with artificial
intelligence techniques such as fuzzy logic and neural networks.
For example, artificial intelligence algorithms are very capable of
capturing urban land use and land change patterns in a non-para-
metric approach and handle to a higher degree spatial heteroge-
neity well (McDonald & Urban, 2006; Wang & Mountrakis,
2011). Additionally, the main advantage of fuzzy clustering is that
objects can be associated with multiple clusters to different de-
grees, whereas in common clustering, each object corresponds
to one cluster only. Because administrative boundaries do not di-
vide people into totally different groups, fuzzy boundaries are
more appropriate than fixed boundaries to represent spatial clus-
ters (See & Openshaw, 2001). With fuzzy clustering, we delineate
a better profile for each spatial unit and avoid aggregating our re-
sults to the level of only one cluster (Grekousis & Hatzichristos,
2012). Additionally, neural networks do not depend on fixed func-
tional relationships and do not require any a priori knowledge of
the variable relationships (Cheng, 2003). This independence is a
major advantage of neural networks because they can be used
as powerful predictive tools when modeling complex nonlinear
problems (Olden & Jackson, 2001). Neural networks are robust
to noise and exhibit a high degree of automation (Openshaw,

0264-2751/$ - see front matter Ó 2012 Elsevier Ltd. All rights reserved.
http://dx.doi.org/10.1016/j.cities.2012.03.006
194 G. Grekousis et al. / Cities 30 (2013) 193–203
1997). Furthermore, they do not make any assumptions regarding
the nature of the distribution of the data (Grekousis & Hatzichris-
tos, 2012).
In this paper, a prototype methodology is proposed for the
development of an urban model from a geodemographic perspec-
tive. Methods and techniques from artificial intelligence are used
for the spatiotemporal analysis, the prediction of future trends
and the development of urban regions. More specifically, we pro-
pose a method estimating future urban growth using neural nets,
fuzzy clustering and geographical information systems (GISs).
First, fuzzy clustering is used to group the spatial units into
clusters. Clustering is based on data from the national census, such
as population, number of buildings and building’s usage. With fuz-
zy clustering, each spatial unit will be assigned a factor of evolu-
tion and urbanism. Based on demographic time-series data, a
neural network is used to predict the future urbanism state of each
spatial unit. The proposed approach is applied to the Athens metro-
politan area of Greece.
The remainder of the paper is organized according to the fol-
lowing outline: Section 2 presents a literature review of urban
growth models and artificial intelligence, primarily focused on
neural networks; Section 3 describes the proposed methodology;
Section 4 presents the case study; and finally, Section 5 presents
the conclusions of this research.
Urban growth models and artificial intelligence
There are several multifaceted efforts investigating the question
of how urban space changes. Various models have been devised to
analyze urbanization along with the physical and socio-economic
factors impacting urban development (Liu, 2008). Urban growth
is a very complex process because it involves multiple actors with
differing patterns of behavior at various spatial and temporal
scales (Cheng, 2003). Furthermore, it is not easy to define urban.
For example, the United Nations Population Fund defines urban
as ‘‘settlements or localities defined as ‘urban’ by national statisti-
cal agencies’’ (United Nations Population Fund, 2007). As a result,
definitions vary based on the standards of national statistical agen-
cies, and there is not a global definition. Since the 1960s, several
theories have been developed for the analysis and simulation of
the evolution of cities, and the mathematical equations that have
resulted have led to the creation of several models simulating ur-
ban phenomena (Batty & Xie, 1994; Chapin & Weiss, 1968; Lau-
rence & Edward, 1981; Tobler, 1970). The liaison between the
theories and the models was often not very strong. Furthermore,
the initially limited computing capability did not always lead to
widely known and applicable tools. Urban growth modeling is an
interdisciplinary field as it involves numerous scientific areas, such
as GIS, remote sensing, urban geography and complexity theory
(Cheng, 2003). GIS and remote sensing have proven to be very
powerful tools for managing spatial information and extracting
valuable knowledge concerning urbanization and its dynamics
(Hasse, 2007; Kumar, Garg, & Khare, 2008; Masser, 2001). Agent-
based modeling (Benenson, 1998; Kerridge, Hine, & Wigan,
2001), fractal-based modeling (Batty & Longley, 1994; Shen,
2002) and decision trees (Triantakonstantis, Mountrakis, & Wang,
2011) have also been used in urban growth modeling.
Artificial intelligence tools including, but not limited to, neural
networks and cellular automata are gaining popularity in modeling
urban growth. Cellular automata are dynamic, discrete-space and
discrete-time systems that are widely used in urban modeling.
For example, Batty, Couclelis, and Eichen (1997), Batty (1998),
and Batty, Xie, and Sun (1999) studied urban systems and urban
dynamics with cellular automata, and Clarke, Hoppen, and Gaydos
(1997) studied the urban development of San Francisco based on
historical data. Almeida, Batty, and Monteiro (2003) designed an
urban model for estimating land-use transitions that used cellular
automata and elementary probabilistic methods.
Neural networks are widely used in geographical analysis
(Hewitson & Crane, 1994; Openshaw, 1997; Pradhan & Lee,
2010). Some applications in urban studies include telecommunica-
tion traffic flows (Fischer & Gopal, 1994), transport planning
(Shmueli, 1998), land cover classification (Foody, 2002; Lula,
2010; Mas & Flores, 2008) and spatial housing market structure
(Kauko, 2004). There are many studies in urban growth modeling
using neural networks. For example, Liu and Lathrop (2002) used
neural networks to detect newly urbanized areas in satellite sensor
images. Their results highlight the practical value of neural net-
works in detecting changes. Using GIS and neural networks, Pija-
nowski, Brown, Shellito, and Manik (2002) developed a land
transformation model based on social, political and environmental
factors to predict land use changes and urbanization. Lin, Lu, Espey,
and Allen (2005) explored the applicability of neural networks in
urban growth modeling. A regression model was used to compare
the results, and neural networks proved better in terms of predic-
tion accuracy.
Maithani, Jain, and Arora (2007) developed a three-layer back-
propagation neural network for simulating urban spatial growth.
Remote-sensing temporal data were used to provide empirical in-
puts related to urban growth and other spatial information. The
model results were evaluated using Moran’s spatial autocorrela-
tion index to predict the urban spatial pattern. In a later study,
Maithani (2009) improved his method using a faster version of
the back-propagation algorithm, the Levenberg–Marquardt algo-
rithm. Liu and Seto (2008) presented a spatiotemporal neural net-
work method to predict urban evolution based on transportation
routes, land use and topography. Pijanowski, Tayyebi, Delavar,
and Yazdanpanah (2009) adopted an urban expansion model
(UEM) based on GIS, neural networks and remote sensing. They
used a multilayer perceptron neural network, satellite imagery
from 1988 and 2000 and socio-economic and environmental vari-
ables to estimate urban expansion for the Tehran metropolitan
area in Iran for the year 2012.
Wang and Mountrakis (2011) developed the multi-network
urbanization (MuNU) model for studying urban growth. In this
very interesting work, the MuNU model integrates multiple neural
networks to capture spatial heterogeneity. The input space is split
into segments, and each input pattern is reallocated to the appro-
priate neural network. Their model is tested in a case study of Den-
ver, Colorado, based on satellite imagery datasets of land use and
land cover. The method was compared to two single-step algo-
rithms, stepwise logistic regression and a single neural network,
and several improvements were suggested by the authors.
In many applications, the combined use of cellular automata,
neural networks and fuzzy logic gives better results than heuristic,
probabilistic or statistical techniques (Kanungo, Arora, Sarkar, &
Gupta, 2006; Melchiorre, Matteucci, Azzoni, & Zanchi, 2008; Prad-
han & Pirasteh, 2010; Yeh & Li, 2002). For example, Guan and
Clarke (2005) used a constrained cellular automata model based
on a two-layer back-propagation neural network to simulate and
forecast urban growth in Beijing. The model estimates demand
for urban space in the future based on socio-economic data. Mant-
elas, Hatzichristos, and Prastacos (2007) presented a methodolog-
ical framework for modeling urban growth based on fuzzy systems
and cellular automata. First, a set of fuzzy systems processes the
data and calculates certain thematic indices regarding urban evo-
lution. Then, empirical rules are used in a cellular automata model
to predict urban growth. Their method was applied to Mesogeia,
which is part of the Athens metropolitan area, for the year 2004.
Liu (2008) used a fuzzy-constrained cellular automata model in a
GIS environment to simulate urban development in Sydney, Aus-
 G. Grekousis et al. / Cities 30 (2013) 193–203
tralia. Moghaddam and Samadzadegan (2009) used neural net-
works to calibrate a cellular automata model of urban growth,
using a historical raster dataset from 1975 to 1991. Ahmadi, Hep-
penstall, Hogg, and See (2009) combined genetic algorithms and
simulated annealing to calibrate an urban cellular automata model.
Then, fuzzy logic was used to formulate transition rules based on
social, physical and economic factors that influence urban growth.
Finally, Abide, Abidin, and Jamaluddin (2010) combined a back-
propagation neural network with cellular automata to build a
pixel-based urban growth model for predicting future patterns.
Temporal satellite imagery data were used, and future urban
growth was predicted based on the historical data.
These studies indicate that although artificial intelligence tech-
niques have limitations, in most cases the techniques provide
better performance than statistical models because they can better
handle the uncertainties of spatial data. All of the studies cited uti-
lized remote-sensing imagery and neural network algorithms or
cellular automata to simulate urban growth phenomena. In the
case of neural networks, historical satellite images and, in some
cases, socio-economic data were stored in a GIS environment,
and then the neural networks were used to predict the future
states of urban growth. There are two major differences between
our approach and these studies. First, we do not use raster data
or cells, as cellular automata models do, to model urban growth,
but instead we use basic geodemographic analysis based on such
variables as population and buildings. Fuzzy clustering gives a
description of the urban state of each spatial unit, and then a neu-
ral network predicts its future state of urbanization. Second, spatial
contingency is adopted in the method because spatial autocorrela-
tion, either positive or negative, usually exists. As a result, we do
not examine each spatial unit by itself or pixel by pixel, as most
of the aforementioned studies do, but we take into account the
changes and the urban growth evolution through time of the
neighboring areas as well.
Theory and methodology
To model urban evolution, an innovative spatiotemporal ap-
proach is adopted based on artificial intelligence techniques and
GIS. The approach is based first on the cluster analysis of the study
area and then on the use of a neural network to predict urban
changes (Fig. 1).
Cluster analysis
The main objective of clustering is to detect common structures
in the data to create clusters. Fuzzy clustering has been proven
especially useful in spatial analysis (Openshaw, 1989). For exam-
ple, Feng and Flowerdew (1998) used fuzzy clustering for geode-
mographic analysis, and Grekousis and Photis (2011) used fuzzy
clustering in introducing a fuzzy index for detecting spatiotempo-
ral outliers. In fuzzy clustering, each cluster has a cluster center
that represents a typical object in the cluster. The advantage of fuz-
zy clustering is that objects can be associated with multiple clus-
ters to different degrees, whereas in common clustering, each
object corresponds to one cluster only. This degree is called the
membership value and reveals how close each object is to the cen-
ter of a cluster. If the membership value of an object in a cluster is
high, for example 95%, this means that it almost coincides with the
cluster’s center and is therefore classified in this cluster. Con-
versely, if the membership value is low (e.g., 5%), the object has lit-
tle similarity to the cluster.
In the proposed approach, spatial entities are grouped based on
common characteristics such as population, population density,
land uses, building uses, job sectors and income. The spatial refer-
195
ence unit is selected first, and then demographic and spatial data
are collected. Demographic data are temporal and should cover
at least four decades. The selection and segregation of the variables
that will be used for clustering is very important and depends on
data availability. Highly correlated variables should be excluded.
Fuzzy clustering using the fuzzy c-means algorithm (Bezdek,
1981) is used next to create clusters that express the level of urban-
ization. Each cluster is characterized by certain values (the cluster
center) and represents a specific level of urban growth and develop-
ment. The clusters identify the evolution of a region from a rural
area (e.g., a village) to a highly populated urban area. Each spatial
reference unit is assigned a membership value for every cluster in
every decade. Consequently, the state of urbanization of the spatial
unit is revealed by the cluster to which it belongs.
The changes in the membership values of a spatial unit that oc-
cur each decade may lead to the unit’s identification with a new
cluster. For example, a region may belong to cluster A (representing
a rural area) with a membership value of 58% in one decade and be-
long to cluster B (representing a semi-rural area) in the next decade
with a membership value of 52%. Changes also may be observed in
regions even if they do not change clusters. For example, a region
may belong to cluster A with a value of 88% in one decade, with a
value of 65% in the next decade and with a value of 52% in the final
decade. However, with crisp clustering, this region would belong
constantly to A without any indication of change. With fuzzy clus-
tering, the decreasing membership value for three decades reveals a
continuous change from cluster A to another cluster. As a result, we
have a better understanding of the region’s urban profile and can
easily interpret expected urban growth changes. This acknowledge-
ment involves the secession of entities in different clusters or the
changes in the same cluster over time indicating that, on one hand,
the stage of development is changing, and on the other hand, the
developmental prospects of each spatial unit are revealed. When
the process is complete, the final results are registered in a geodat-
abase and are visualized with the aid of GIS software. We can only
trace trends with cluster analysis, however; we cannot predict the
future urban growth. A neural network will be used in a second
phase to estimate the future urban evolution.
Neural network
A neural network with time-series data for each spatial unit
may be used to predict future urban evolution. In our approach,
we enhance the time-series data by taking spatial interaction into
account. Each spatial unit shares borders with other spatial units of
the same or of another evolution cluster. This contact facilitates the
bidirectional assimilation of characteristics and the interaction be-
tween the spatial units. As a result, we expect that the changes ob-
served in the neighboring area of a region will impact future
changes in that region. Consequently, we do not predict the future
urban state of a region based solely on its own time-series data, but
we incorporate the urban evolution of neighboring regions based
on the spatial autocorrelation primitive.
With the aid of GIS software, the total length of the borders of
each spatial unit is calculated as well as the lengths of the common
borders with each of the neighboring spatial units. In this way, new
percentage values are created for each spatial unit indicating the
percentages of the clusters that it borders. For example, a spatial
unit belongs to cluster A; of the neighboring areas, it borders with
32% of the municipalities belonging to cluster A, with 45% belong-
ing to cluster B and with 23% belonging to cluster C. Although the
border length is not the only parameter that can be used to study
spatial interaction, in this research we choose to use only this
parameter for simplicity.
The time-series data for each spatial unit and for the neighbor-
ing spatial units are inserted into the input table of the neural net-
196 G. Grekousis et al. / Cities 30 (2013) 193–203
Problem
Definition
Fuzzy Clustering
Data
Collection
Data Selection
Spatial Database
creation
Clustering
Output Matrices &
Maps
Fig. 1. Methodology.
work. In this way, we train the network to learn not only how each
spatial unit evolves over time but also how the neighboring areas
have affected these changes. This structure of the input table pro-
vides the history of a spatial unit and the history of its surround-
ings to enable the prediction of its future state.
A set of parameters must be regulated for each neural network.
These parameters include the following: the number of hidden lay-
ers, the number of neurons per layer, the learning rate, the
momentum, the epochs of iteration and the transfer functions.
The appropriate setting of these parameters is of crucial impor-
tance because the final precision and acceptance of the network
depends on them.
User experience and several trials are required to build an effi-
cient network. After the results are evaluated with the use of sta-
tistical indicators, such as the mean square error (MSE) and the
correlation coefficient r, and considered acceptable, the final pre-
diction maps are created. The correlation coefficient r is defined
in the range [ 1, 1], and values near 1 indicate good performance
by the neural network, i.e., that the output of the network is close
to the reference output.
The case of Athens metropolitan area
The proposed methodology is applied in the Athens Metropoli-
tan area which is characterized by substantial disparities in its
demographic and urban landscape. The Athens metropolitan area
contains small villages as well as big urban centers such as the city
of Athens, which is the capital of Greece. Because of its special
characteristics that include a high concentration of industrial and
productive activities, services and a specialized workforce, it con-
stituted a case study with particular interest. Metropolitan Athens
extends across 3808 sq km with a population of nearly 4 million
people divided into 108 municipalities.
Neural Network Model
Yes
Y
Validation & No
A t
Acceptance Prediction
of model
NN Training Output Matrices &
Maps
NN Model calibration Evaluation -
parameters Interpretation of results
Data
The spatial reference unit selected is the municipality because
data for minor spatial units were not available. Data were derived
from the NSSG (National Statistical Service of Greece) censuses
only concerning population and building uses from 1961 until
2001. We chose to exclude other demographic variables in this
study because our main motive was to test our method based on
the main variables that express urban growth, which are the pop-
ulation and the building use. However, a more in-depth analysis
using additional demographic data would add to this study and
is planned for future research. For a total of four decades, the fol-
lowing nine variables were finally selected: (1) population, (2)
growth population rate over the decade, (3) number of buildings,
(4) rate of new buildings over the decade, building use, such as
(5) residence use, (6) service use (shops, offices), (7) industrial
use, (8) public building use (schools, ministries, police depart-
ments, etc.), and (9) other use. With these variables based on the
living population, growth rate of the population over time and
building usages, one can make a profile of a municipality revealing
a lot in regards to the urban profile of an area. For example, high
values in office use along with high population and a high number
of buildings may reveal a city center. On the other hand, high pop-
ulation with high values in residential use may reveal a suburb.
Cluster analysis
A fuzzy c-means algorithm is implemented to create clusters.
Clustering is performed for each decade, and clusters defined the
degree of urbanization of each municipality. The optimal choice
for the number of clusters to be used is made using several valida-
tion criteria, such as the partition coefficient (PC) and classification
entropy (CE). The optimal number of clusters lay at the point,
where the graphs of the PC and CE indices for different choices of
 G. Grekousis et al. / Cities 30 (2013) 193–203 197

clusters display the first sharp change in their gradients (Grekousis
& Hatzichristos, 2012).
Fig. 2 displays a change at point 5 of the PC and CE indices. Con-
sequently, the optimal number of clusters for this case study is five.
Each cluster is described by its cluster center values as shown in
Table 1. Each municipality belongs to some degree to a cluster in
each decade. Thus, for example, if a municipality belongs to cluster
A in the 1971–1981 decade, then the values of its demographic
variables are similar to the values of cluster center A. Conse-
quently, cluster center values refer to demographic variables over
a decade. The delineation of each cluster is based on its cluster cen-
ter values and is given below.
Cluster A: The growth rate of the population (54.6%) of cluster A
is the highest among all the clusters, due to the influx of new res-
idents. Simultaneously, the rate of construction of new buildings is
also high (96.5%), which is reasonable because an increase in the
population requires new buildings for residences. The residential
use of the buildings is the highest among the clusters (91.2%).
These residences are used extensively as secondary (i.e., vacation)
homes. The municipalities in this cluster have the lowest percent-
ages of use of industry (0.9%) and services (2.1%). Furthermore,
there is a low population density, an absence of boisterous activity
and a wide natural landscape. This cluster primarily consists of vil-
lages in the outskirts of the metropolitan area.
Cluster B: The growth rate of the population in this cluster is very
small (29.8%), but the rate of new building is very high. Residential
use presents the lowest percentage (76.2%) of all the clusters,
whereas industrial use and other uses have the highest percentages.
Cluster C: The population growth rate and the rate of construc-
tion of new buildings are very high in this cluster: 52.3% and 96.6%,
respectively. These municipalities are characterized by the second-
lowest percentage of residential use (86%) and the second-highest
percentages of service use (3.1%) and industrial use (2.3%). Munic-
ipalities in this cluster attract work activities such as the craft
industry, small agricultural units and the general service sector;
consequently, these municipalities are poles of attraction for the
population.
Cluster D: Low population growth and low building growth
rates are observed, revealing a tendency for slow evolution in this
cluster. Municipalities in this cluster have the third-highest per-
centage of residential use, only slightly less than the cluster E with
the second-highest residential use. The percentage of service use is
2.7%, demonstrating a concentration of services. Cluster D is char-
acterized by municipalities with a sufficient population and a suf-
ficient number of buildings, and the presence of the service
industry in this cluster is strong.
Cluster E: Municipalities in this cluster present the lowest rates
of both population growth and building construction: 22.9% and

Fig. 2. Partition coefficient (PC) and classification entropy (CE).

Table 1
Cluster centers.

Population Population rate Buildings Building rate Residence Public Service Industrial Other
A 5603 0.546 1527 0.965 0.912 0.013 0.021 0.009 0.045
B 6028 0.298 2127 0.969 0.762 0.019 0.027 0.024 0.170
C 9946 0.523 2640 0.966 0.860 0.013 0.031 0.023 0.074
D 30438 0.421 4879 0.672 0.898 0.010 0.027 0.018 0.046
E 66510 0.229 10707 0.429 0.900 0.008 0.034 0.019 0.039
198 G. Grekousis et al. / Cities 30 (2013) 193–203
42.9%, respectively. The municipalities are characterized as highly
urban areas, and they are in a state of very slow change due to the
over-concentration of population, the intense building activity of
past years and the lack of space to build. The percentage of residen-
tial use is 90%, whereas the public building percentage is the low-
est of all clusters. However, the percentage of service is the highest,
3.4%. Municipalities in cluster E are powerful urban centers, and
they constitute poles of attraction for both the population and
the installation of services.
For each decade, each municipality was assigned the value of
the membership function resulting from the fuzzy c-means classi-
fication algorithm. The membership values of the municipalities in
clusters A and E for the decades of 1981–1991 and 1991–2001 are
depicted in Fig. 3.
More can be understood about each spatial unit as its member-
ship values change. Whereas other clustering methods are unable
to trace differences for municipalities within the same cluster, fuz-
zy clustering provides a value expressing how typical a municipal-
ity is for a cluster. By this process, tendencies can be identified and
future evolution can be simulated. For example, in Fig. 3, the mem-
bership values of cluster A are depicted for the 1981–1991 and
1991–2001 decades. We can observe that in the 1991–2001 dec-
ade, more municipalities have low membership values (i.e., more
municipalities have less than a 20% membership in cluster A).
Fig. 3. Membership values for clusters A and E.
With fuzzy clustering, we can trace changes in a municipality
profile even if the municipality does not change clusters. For exam-
ple, the northern municipalities that still belong to cluster A during
the 1991–2001 decade have lower membership values than in the
previous decade, and consequently there is an increase in their
membership values in other clusters, such as cluster B (Fig. 4). This
reduction in the membership values of these municipalities in
cluster A suggests a high probability that they will change clusters
in the following decades. Cluster E is also depicted in Fig. 3 for the
1981–1991 and 1991–2001 decades. The municipalities belonging
to cluster E have higher membership values in 1991–2001. This
means that their values are moving closer to the cluster center,
i.e., toward greater urbanization. Nevertheless, although the
municipalities belong to cluster E, based on their membership val-
ues, some municipalities (the darker ones in the figure) are more
typical (i.e., more urban) than the others. To depict the results
more clearly, we defuzzified the results by assigning each munici-
pality to the cluster, where it had the maximum membership value
for each decade.
In the first decade, the majority of the municipalities outside
the Athens city complex were in clusters A and B, whereas the
municipalities primarily characterized by clusters D and E were
aggregated around the city of Athens (Fig. 4). Massive urbanization
had begun in Athens during the 1950s, and this urbanization con-
 G. Grekousis et al. / Cities 30 (2013) 193–203 199

Fig. 4. Clusters during 1961–2001.

tinued during the 1960s. A remarkable increase in the population Table 2

during the decade of 1971–1981 was observed in the majority of Membership values for the municipality by cluster for each decade.

municipalities, mainly around Athens. Furthermore, the population Decade A B C D E

increasingly chose to seek residence far from the urban center, 61–71 0.114858 0.046508 0.126751 0.640658 0.071224
mainly for a better quality of life. Northern suburbs and the munic- 71–81 0.110278 0.070643 0.12308 0.3293 0.366699
ipalities on the eastern coast became more attractive. During the 81–91 0.087252 0.059034 0.097683 0.251538 0.504493
decade of 1981–1991, changes were observed in the coastal 91–2001 0.067178 0.041567 0.077724 0.222563 0.590990

municipalities (where seasonal residences were common) and in

the southeastern municipalities (where the new national airport
was about to be built). Finally, during the decade of 1991–2001,
urbanization continued to increase, especially toward the south-
eastern municipalities. The majority of new residences in these
areas were now not for summer vacations but for use as perma-
nent homes. The commuting phenomenon increased.
For better comprehension of the algorithm results, we present
the output of the fuzzy clustering and the demographic data for
the municipality of N. Smyrni in Tables 2 and 3 and Fig. 5.
In the 1961–1971 decade, the municipality belongs to cluster D
with a 64% membership value (Table 2) and, as shown in Table 3,
its demographic data are close to the cluster centers of class D (Ta-
ble 1). During the 1971–1981 decade, the municipality transforms
gradually from the D cluster to the E cluster as the municipality’s
membership value in cluster E increases from 7.1% to 36.6%, versus
32.93% in cluster D. In the 1981–1991 decade, the municipality be-
longs to cluster E with a membership value of 50.4%. Finally, in the
1991–2001 decade, the municipality belongs to cluster E with a va-
lue of 59.09%.
GIS techniques were then applied to calculate the lengths of the
common borders of each municipality. For each municipality, a
new value was defined for the percentages of clusters that it bor-
dered. For the municipality of N. Smyrni, the results over the four
decades are depicted in Fig. 5 and Table 4. During the 1961–1971
decade, the municipality of N. Smyrni, which belonged to cluster
D with a membership value of 64% (Table 2), shared borders with
municipalities in cluster D (63.2%) and in cluster E (36.8%). During
the following decades, the municipality’s membership changes to
cluster E, as explained earlier in this section, with an increasing
membership value in this cluster. This can also be observed for
the neighboring areas because there was a consistent transition
200 G. Grekousis et al. / Cities 30 (2013) 193–203

Table 3
Demographic data for the municipality during four decades.

Decade Population Population rate Buildings Building rate Residence Schools Offices Industrial Other
61–71 42512 0.2938 3929 0.554803 0.929591 0.008685 0.016129 0.004032 0.041563
71–81 67408 0.5856 5425 0.380753 0.933481 0.008591 0.017428 0.003436 0.037064
81–91 69749 0.0342 6473 0.193174 0.936615 0.000205 0.018667 0.003692 0.032821
91–2001 73986 0.0607 6892 0.064730 0.938105 0.000195 0.019792 0.002022 0.030011

Fig. 5. Urbanization levels of the municipality and neighboring regions.

Table 4
umns contain the membership function values for each municipal-
Neighboring clusters for N. Smyrni.
ity for each decade. Columns 6–10 contain the membership values
61–71 (%) 71–81 (%) 81–91 (%) 91–2001 (%) of the clusters (e.g., from Table 4) of the surrounding municipalities
D 63.2 39.8 9.5 9.5 for the same decade. The last five columns contain the membership
E 36.8 60.2 90.5 90.5 values for the same municipalities during the next decade. The first
10 columns are used as the input data for the neural network, and
from cluster D (63.2%) in 1971 to cluster E (90.5%) in 2001. the next five columns are treated as the output reference data.
Although there appears to be no change between the decades of Thus, the objective of the neural network is to determine how
1981–1991 and 1991–2001, municipalities around N. Smyrni con- the first 10 columns are involved in producing the last five col-
sistently increase their membership values in cluster E. umns, which represents the future state of the urban evolution.
70% of this dataset is randomly selected and used as a training
Neural network dataset, and the other 30% is used as a testing dataset. More than
50 combinations of neural network settings are selected, and after
A multilayer perceptron neural network is selected, and a data trials and tests the best-performing neural network consisting of
set consisting of 15 columns is initially created. The first 10 col- two hidden layers connected with 550 neurons in a linear transfer
 G. Grekousis et al. / Cities 30 (2013) 193–203
function is selected (neural network C in Table 5). The learning-
rate momentum values for the two hidden layers are 0.1 and 0.9.
The evaluation produced promising results, represented by the
high average accuracies achieved. For example, the MSE of the
training error after 1000 epochs is 0.005 (Fig. 6), and r is 0.91.
Two additional neural networks (A and B) are presented in Table
5 to show the architectures (i.e., hidden layers, transfer functions,
step size and momentum) tested during the trial phase and the re-
sults after 1000 epochs.
The testing of the network using the test data is performed as
follows. The test dataset is not embedded during the training phase
and is used instead to compare the neural network results with the
actual data. The reference output and the actual network output
are plotted for a small part of the test data (Fig. 7). In this figure,
the actual membership values (identified by a straight line) of clus-
ters A–C for a set of 15 municipalities during the decade of 1981–
1991 are plotted with the network output (identified by a dashed
line). It can be observed from the plot that the network has recog-
nized the general pattern of the data and is able to predict the basic
trends. Although in some places there are deviations, the line trend
is nevertheless predicted with high accuracy. The final test MSE of
the test dataset is 0.0067.
Once the neural network learned to recognize patterns among
the data, it is ready for prediction. Data obtained since 2001 are
used to predict results for the year 2011. The final output is the
membership value of each cluster assigned to a municipality. The
results after defuzzification are depicted in Fig. 8.
Discussion
The concentration of the population in municipalities offering
many services and places of work attracts residents and the con-
struction of new buildings, thus propagating urban growth. The
Table 5
Neural network configuration during training.
Neural network Hidden layers Transfer function
A 1 Linear
B 2
1st Linear
2nd Linear
C 2
1st Linear
2nd Linear
0.6
0.5
0.4
MSE
0.3
0.2
0.1
0
1 100 199 298
Fig. 6. Mean square error versus epochs.
201
trends and pressures encountered by the municipalities can be dis-
cerned from the prediction map (Fig. 8) for the year 2011. We ob-
served a polarized state of the Athens metropolitan area between
municipalities with a high urban character and municipalities with
mainly residential (particularly vacation) use, almost causing the
municipalities to vanish from cluster C. Furthermore, the industrial
and agricultural sectors moved, to a large extent, out of the district.
During the decade of 1961–1971, 17.8% of the municipalities were
associated with cluster D, and progressively the percentage
reached 26.6% in 1991–2001 (Table 6). In the predicted decade of
2001–2011, this percentage rises to 32.4%. In this decade, 1 out
of 3 municipalities were associated with cluster D, whereas in
1961–1971 the correspondence was only 1–5 (Table 6). This result
was consistent with the general trend of urbanization.
In cluster A, a high concentration of municipalities was ob-
served; however, a slightly declining trend occurred. Thus, the con-
centration rose from 29.4% during the decade of 1961–1971 to 33%
in 1971–1981, but fell to 32% in 1981–1991. The prediction for
2001–2010 showed that 24.7% of the municipalities would be asso-
ciated with this cluster. In general, the municipalities that were
associated with cluster A were characterized by high stability over
time because out of a total of 30 municipalities that remained un-
changed over time, 20 (66%) were associated with cluster A.
In cluster B, which included the under-populated municipali-
ties, there was a continuously decreasing membership, demon-
strating the significant evolution of the prefecture of Attica. The
20% of municipalities associated with cluster B during the decade
of 1961–1971 fell to 14.6% in 1981–1991 and was predicted to
be only 2.7% in 2001–2011. Cluster C was characterized by munic-
ipalities that were in the middle of the evolution process. Only one
municipality was associated with this cluster over time, showing
the fluidity of this cluster. Cluster E followed an ascendant trend.
In the decade of 1961–1971, 8.5% of the municipalities (mainly
Neurons Epochs Momentum MSE r
252 1000 0.1 0.09 0.71
488 1000 0.1 0.07 0.87
0.5
0.5
550 1000 0.005 0.91
0.1
0.9
397 496 595 694 793 892 991
Epoch
202 G. Grekousis et al. / Cities 30 (2013) 193–203

0.7

0.6

0.5

Output
0.4 A 81-91
B81-81
0.3
C91-81
0.2
A 81-91 Output

0.1 B81-91 Output

C81-91 Output
0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
Exemplar

Fig. 7. Reference output and actual network output.

ral network predicts the future state of urbanization. Urban growth

estimates are based on population and building data, not on raster
datasets such as satellite imagery. Because only two datasets are
used, the applicability of the proposed methodology is increased.
Furthermore, spatial contingency is also taken into account, and
we calculate the changes and the urban growth of specific spatial
units together with the neighboring area. Nevertheless, the pro-
posed approach has some limitations. The spatial resolution is
determined by the administrative boundaries (e.g., the municipal-
ity or the postal code). Consequently, we can estimate the future
urban evolution for each spatial unit, but we cannot determine,
where it will take place inside the spatial unit.

Conclusions

Enhanced possibilities emerge by the use of fuzzy clustering

and neural networks in various spatial problems. The proposed
methodology is characterized by its simple and simultaneously dy-
namic structure. Fuzzy clustering proved to be particularly effec-
tive for clustering because each spatial unit had its own separate
profile for each decade. Thus, not only a complete picture of the
study area for each time interval was shaped, but by the diachronic
Fig. 8. Municipality urbanization levels in 2011. change in membership values, future trends for each spatial unit
were also revealed. With the use of neural network, the future
urbanization state of each municipality arose and as a result, a bet-
Table 6 ter profile for this area may be drawn in order to act at a planning
Municipalities by cluster. level. The proposed methodological approach delineated the future
61–71 (%) 71–81 (%) 81–91 (%) 91–2001 (%) 2001–2011 (%)
urbanism level more reliably in combination with a more tempo-
rally and thematically complete database, as well as a more exten-
A 29.4 33.0 32.0 25.6 24.7
B 20.0 14.6 13.3 2.9 2.7
sive spatial interaction model.
C 24.3 16.7 11.8 18.3 9.1 As far as the application was concerned, the applied approach
D 17.8 21.1 22.9 26.6 32.4 had a direct practical and explanatory value for Athens metropol-
E 8.5 14.6 20.0 26.6 31.1 itan area itself, in regards to the trends that prevailed in the munic-
ipalities for their demographic development and the building use
changes in the next decades. The output of the generated neural
the municipalities around the city of Athens) belonged to this clus-
network was judged satisfactory, taking into consideration the
ter. During the next decade, the percentage increased to 14.6%, and
amount of available data. Further development is under way,
during 1991–2001 it reached 26.6%. The percentage of the munic-
involving the enrichment of the methodology and the application
ipalities predicted to belong to this cluster in 2001–2011 was
for the year 2021 when required data from 2011 census become
31.1%.
available (estimated during late 2012).
In summary, in the period from 1961–1971, the municipalities
belonging to categories D and E (i.e., the developed urban centers)
comprised 25.8% of the total. In 1981–1991, the municipalities References
reached 42.9%, and during 1991–2001 the total was 53.2%. Finally,
Abide, M.Z., Abidin, S. Z. Z., Jamaluddin, M. F. (2010). Pixel based urban growth
the prediction for 2001–2011 was that 63.1%, or almost two out of model for predicting future pattern. In 6th International colloquium on signal
three municipalities in Attica, would belong to clusters D or E. processing & its applications (CSPAs) 21–23 May 2010. Malaysia.
In conclusion, we use fuzzy clustering and neural networks to Ahmadi, K., Heppenstall, A., Hogg, J., & See, L. (2009). A fuzzy cellular automata
urban growth model (FCAUGM) for the City of Riyadh, Saudi Arabia. Part 1:
analyze and predict urban growth. Fuzzy clustering provides a Model structure and validation. Applied Spatial Analysis and Policy, 2(1), 65–83.
description of the urban state of each spatial unit, and then a neu- http://dx.doi.org/10.1007/s12061-008-9020-6.
 G. Grekousis et al. / Cities 30 (2013) 193–203
Almeida, C. M., Batty, M., & Monteiro, A. M. V. (2003). Stochastic cellular automata
modeling of urban land use dynamics: Empirical development and estimation.
Computers, Environment and Urban Systems, 27(5), 481–509.
Batty, M., & Xie, Y. (1994). Urban analysis in a GIS environment. In S. Fotheringham
& P. Rogerson (Eds.), Spatial analysis and GIS (pp. 189–219). Bristol (PA): Taylor
& Francis.
Batty, M., & Longley, P. (1994). Fractal cities: A geometry of form and function. San
Diego: Academic Press.
Batty, M., Couclelis, H., & Eichen, M. (1997). Urban systems as cellular automata.
Environment and Planning B: Planning and Design, 24, 159–164.
Batty, M. (1998). Urban evolution on the desktop: Simulation with the use of
extended cellular automata. Environment and Planning A, 30, 1943–1967.
Batty, M., Xie, Y., & Sun, Z. (1999). Modeling urban dynamics through GIS-based
cellular automata. Computers, Environment and Urban Systems, 23, 205–233.
Benenson, I. (1998). Multi-agent simulations of residential dynamics in the city.
Computers, Environment and Urban Systems, 22(1), 25–42.
Bezdek, J. C. (1981). Pattern recognition with fuzzy objective function algorithms. New
York: Plenum Press.
Chapin, F. C., & Weiss, S. F. (1968). A probabilistic model for residential growth.
Transportation Research, 2, 375–390.
Cheng, J. (2003). Modeling spatial and temporal urban growth. PhD Dissertation,
Utrecht University.
Clarke, K. C., Hoppen, S., & Gaydos, L. (1997). A self-modifying cellular automaton
model of historical urbanization in the San Francisco Bay area. Environment and
Planning B: Planning and Design, 24, 247–261.
Feng, Z., & Flowerdew, R. (1998). Fuzzy geodemographics: A contribution from
fuzzy clustering methods. In Innovations in GIS: Innovations in GIS 5. Selected
papers from the fifth national conference on GIS research UK, Taylor & Francis, UK.
Fischer, M. M., & Gopal, S. (1994). Artificial neural networks: A new approach to
modeling interregional telecommunication flows. Journal of Regional Science, 34,
503–527.
Foody, G. M. (2002). Hard and soft classifications by a neural network with a
nonexhaustively defined set of classes. International Journal of Remote Sensing,
23(18), 3854–3864.
Grekousis, G., & Photis, N. Y. (2011). A fuzzy index for detecting spatiotemporal
outliers. Geoinformatica. http://dx.doi.org/10.1007/s10707-011-0145-4.
Grekousis, G., & Hatzichristos, T. (2012). Comparison of two fuzzy algorithms in
geodemographic segmentation analysis: The fuzzy C-means and
GustafsoneKessel methods. Applied Geography, 34, 125–136. http://dx.doi.org/
10.1016/j.apgeog.2011.11.004.
Guan, Q. L. W., & Clarke, K. C. (2005). An artificial-neural-network-based,
constrained CA model for Si-mulating urban growth. Cartography and
Geographic Information Science, 32(4), 369–380. http://dx.doi.org/10.1559/
152304005775194746.
Hasse, J. (2007). Using remote sensing and GIS integration to identify spatial
characteristics of sprawl at the building-unit level. In V. Mesev (Ed.), Integration
of GIS and remote sensing (pp. 117–143). New York: John Wiley & Sons.
Hewitson, B., & Crane, R. (1994). Neural nets: Application in geography. The
Netherlands: Kluwer Academic Publishers.
Kanungo, D. P., Arora, M. K., Sarkar, S., & Gupta, R. P. (2006). A comparative study of
conventional, ANN black box, fuzzy and combined neural and fuzzy weighting
procedures for landslide susceptibility zonation in Darjeeling Himalayas.
Engineering Geology, 85, 347–366.
Kauko, T. (2004). A comparative perspective on urban spatial housing market
structure: Some more evidence of local sub-markets based on a neural network
classification of Amsterdam. Urban Studies, 41(13), 2555–2579.
Kerridge, J., Hine, J., & Wigan, M. (2001). Agent-based modeling of pedestrian
movements: The questions that need to be asked and answered. Environment
and Planning B: Planning and Design, 28, 327–341.
Kumar, J. M., Garg, P. K., & Khare, D. (2008). Monitoring and modeling of urban
sprawl using remote sensing and GIS techniques. International Journal of Applied
Earth Observation and Geoinformation, 10(1), 26–43.
Laurence, J. C. M., & Edward, W. H. (1981). Urban development in modern China.
Boulder: Westview Press.
Lin, H., Lu, K. S., Espey, M., & Allen, J. (2005). Modeling urban sprawl and land use
change in coastal area – A neural network approach. In American agricultural
economics association annual meeting, providence, Rhode Island, July 24–27, 2005.
Liu, X., & Lathrop, J. R. G. (2002). Urban change detection based on an
artificial neural network. International Journal of Remote Sensing, 23(12),
2513–2518.
Liu, Y. (2008). Modeling urban development with geographic information systems and
cellular automata. Boca Raton Florida: CRC Press.
203
Liu, W., & Seto, K. C. (2008). Using the ART-MMAP neural network to model and
predict urban growth: A spatiotemporal data mining approach. Environment
and Planning B: Planning and Design, 35(2), 296–317.
Lula, V. (2010). Neural network classification of hyperspectral imagery for urban
environments: A case study. PhD Dissertation, Department of Earth and
Environmental Systems Indiana State University.
Maithani, S., Jain, R. K., & Arora, M. K. (2007). An artificial neural network approach
for modeling urban spatial growth. ITPI Journal, 4(2), 43–51.
Maithani, S. (2009). A neural network based urban growth model of an Indian city.
Journal of Indian Society and Remote Sensing, 37, 363–376.
Mantelas, L., Hatzichristos, T., & Prastacos, P. (2007). A fuzzy cellular automata
based shell for odeling urban growth – A pilot application in Mesogia area. In
10th AGILE international conference on geographic information science. Denmark:
Aalborg University.
Mas, J. F., & Flores, J. J. (2008). The application of artificial neural networks to the
analysis of remotely sensed data. International Journal of Remote Sensing, 29(3),
617–663.
Masser, I. (2001). Managing our urban future: The role of remote sensing and
geographic information systems. Habitat International, 25, 503–512.
McDonald, R. I., & Urban, D. L. (2006). Spatially varying rules of landscape change:
Lessons from a case study. Landscape and Urban Planning, 74(1), 7–20.
Melchiorre, C., Matteucci, M., Azzoni, A., & Zanchi, A. (2008). Artificial neural
networks and cluster analysis in landslide susceptibility zonation.
Geomorphology, 94, 379–400.
Moghaddam, K. H., 7 Samadzadegan, F. (2009). Urban simulation using neural
networks and cellular automata for land use planning. In: Proceedings REAL
CORP 2009 Tagungsband 22–25 April 2009. <http://www.corp.at/archive/
CORP2009_54.pdf>.
Olden, J. D., & Jackson, D. A. (2001). Fish–habitat relationships in lakes: Gaining
predictive and explanatory insight by using artificial neural networks.
Transactions of the American Fisheries Society, 130, 878–897.
Openshaw, S. (1989). Making geodemographics more sophisticated. Journal of the
Market Research Society, V31(1), 111–131.
Openshaw, S. (1997). Artificial intelligence in geography. London: John Wiley & Son
Ltd.
Pijanowski, B. C., Brown, D. G., Shellito, B. A., & Manik, G. A. (2002). Using neural
networks and GIS to forecast land use changes: A land transformation model.
Computers, Environment and Urban Systems, 26(6), 553–575. http://dx.doi.org/
10.1016/S0198-9715(01)00015-1.
Pijanowski, B. C., Tayyebi, A., Delavar, M. R., & Yazdanpanah, M. J. (2009). Urban
expansion simulation using geo-spatial information system and artificial neural
networks. International Journal of Environmental Research, 3(4), 493–502.
Pradhan, B., & Pirasteh, P. (2010). Comparison between prediction capabilities of
neural network and fuzzy logic techniques for landslide susceptibility mapping.
Disaster Advances, 3(2), 26–34.
Pradhan, B., & Lee, S. (2010). Landslide susceptibility assessment and factor effect
analysis: Backpropagation artificial neural networks and their comparison with
frequency ratio and bivariate logistic regression modeling. Environmental
Modeling and Software, 25(6), 747–759.
See, L., & Openshaw, S. (2001). Fuzzy geodemographics targeting. In Regional science
in business. Berlin: Springer.
Shen, G. (2002). Fractal dimension and fractal growth of urbanized areas.
International Journal of Geographical Information Science, 16(5), 419–437.
Shmueli, D. (1998). Applications of neural networks in transportation planning.
Progress in Planning, 50, 141–204.
Triantakonstantis, D., Mountrakis, G., & Wang, J. (2011). A spatially heterogeneous
expert based (SHEB) urban growth model using model regionalization. Journal
of Geographic Information System, 5, 195–210.
Tobler, W. R. (1970). A computer movie simulating urban growth in the Detroit
Region. Economic Geography, 46, 234–240.
United Nations Population Fund (2007). State of the world population 2007:
Unleashing the potential of urban growth. United Nations Population Fund,
United Nations Publications; 2007.
Waddell, P. (2000). Monitoring and Simulating Land Capacity at the Parcel Level. In
Monitoring land supply with geographic information systems: Theory, practice and
parcel-based approaches.
Wang, J., & Mountrakis, G. (2011). Developing a multi-network urbanization
(MuNU) model: A case study of urban growth in Denver, Colorado. International
Journal of Geographical Information Science, 25(2), 229–253.
Yeh, G.A., Li, X. (2002). Urban simulation using neural network and cellular
automata for land use planning. In Symposium on geospatial theory, processing
and applications, Ottawa; 2002.