Please cite the Published Version

Gounaridis, D, Chorianopoulos, I, Symeonakis, E and Koukoulas, S (2019) A Random Forest-

Cellular Automata modelling approach to explore future land use/cover change in Attica (Greece),
under different socio-economic realities and scales. Science of the Total Environment, 646. pp.
320-335. ISSN 0048-9697

DOI: https://doi.org/10.1016/j.scitotenv.2018.07.302
Publisher: Elsevier
Version: Accepted Version
Downloaded from: https://e-space.mmu.ac.uk/621512/

Usage rights: Creative Commons: Attribution-Noncommercial-No Deriva-

tive Works 4.0
Additional Information: This is an Author Accepted Manuscript of a paper accepted for publica-
tion in Science of the Total Environment, published by and copyright Elsevier.

Enquiries:
If you have questions about this document, contact openresearch@mmu.ac.uk. Please in-
clude the URL of the record in e-space. If you believe that your, or a third party’s rights have
been compromised through this document please see our Take Down policy (available from
https://www.mmu.ac.uk/library/using-the-library/policies-and-guidelines)
 A Random Forest-Cellular Automata modelling approach to explore

future land use/cover change in Attica (Greece), under different

socio-economic realities and scales

Dimitris Gounaridis, Ioannis Chorianopoulos, Elias Symeonakis, Sotirios Koukoulas

Science of the Total Environment, 646, DOI: 10.1016/j.scitotenv.2018.07.302

Accepted: July 2018

1. Introduction

Land use and land cover (LULC) changes are considered to be the most prominent

influence of humans on the environment. Technological and medical advancements

have brought about unprecedented increases in the human population and,

consequently, in the need for access to resources. This need has in turn caused

substantial and growing transformations to the Earth’s surface (Vitousek et al., 1997)

with often undesirable impacts and magnitudes that vary from local to global scales.

The dual role of humans to actively contribute to LULC changes and, at the same time,

be on the receiving end of experiencing the consequences of these changes, emphasizes

the need for a better understanding of the human-LULC change nexus.

A wide variety of LULC change models have been developed to meet the scientific

community needs for understanding how and why LULC evolves (Schrojenstein

Lantman et al. 2011). Generally, LULC models are widely used to analyze the complex

structure of linkages and feedbacks between drivers of change, determine their

relevance to particular changes and project how much land is used where and for what

purpose, under different predefined attributes and conditions. This type of information

is then adopted in a meaningful way in order to support policy decision making related
to land-use (Mallampalli et al. 2016). However, by definition, LULC models can not

exactly replicate the complex interactions and nonlinear relations which are apparent in

LULC systems. At a fundamental level, they are, rather, a process that provides a

platform that, allows computer experiments to be undertaken (Brown et al. 2013). When

the system in question is simple, the processes and interactions that characterize it can

be easily determined and the results are somehow expected, while projections and other

kinds of extrapolations are a straightforward task. When dealing, however, with

inherently complex systems, as is the case with LULC changes, the models are able to

represent and exemplify only a small fraction of the whole mechanism in order to

highlight important processes.

The recent methodological and technological advancements have paved the way for

more articulated LULC models which are able to answer more complex questions. Such

questions could be in regard to what the possible outcomes would be if alternative

pathways were followed, or which outcome is the most desirable from a list of

alternatives, as well as a diverse range of other ‘what-if’ scenarios. Increasingly,

scenario-based analysis is now being adopted by a range of disciplines pertaining to

LULC change, as fruitful experiments for exploring the possible future trajectories of

historical and current trends (Murray-Rust et al. 2013). Considering that the number of

potential futures is actually infinite (Greeuw et al. 2000), scenarios are not used to

predict the future in a precise manner, but to explore possible future directions and to

consider a range of alternative pathways. To do so, the scenario-based analysis fully

recognizes the infinity of potential futures and attempts to focus only to an

understandable and manageable set of alternatives. This is achieved by delineating

plausible, presumably coherent and internally consistent storylines of different socio-

economic development trajectories (Rounsevell and Metzger, 2010).

When modelling LULC, the scale, the spatial resolution and the extent of the study area

are important attributes of all spatially explicit models (Agarwal et al 2002). The term

scale refers to the spatial, temporal, quantitative, or analytic dimension used to measure

and study the processes that are modelled (Gibson et al., 2000). Scale also involves the

terms extent and resolution: extent refers to the magnitude of a dimension used in

measuring (e.g. study area boundaries on a map), whereas resolution refers to the

precision used in this measurement (e.g. pixel size) (Gibson et al. 2000). Moreover,

resolution refers not only to spatial resolution, but also to thematic, which is the level

of precision in LULC categories. In addition, the term temporal resolution is used to

refer to the time span and frequency of the analysis. Modelling LULC changes,

therefore, requires a range of scales to be defined since it is a phenomenon that involves

multiple processes that act over different scales. At each scale, different processes have

a dominant influence on the outcome (Meentemeyer, 1989; Van Delden et al. 2011).

Approaches that do not implement a multi-scale approach are prone to aggregation or

oversimplification errors and thus fail to reproduce cross-scale interactions. This is due

to the fact that features and processes that operate at local scales are not always

observable when dealing with larger areas and coarser spatial resolution data (Verburg

et al. 2004). On the other hand, studies that focus solely on the local level often fail to

incorporate information about the general context which can only be derived from

coarser spatial resolution data (Larondelle & Lauf, 2016). Given that all models are

driven by their input data, studies focusing on specific LULC processes, considering

only a single scale and using data that are particularly suitable only to a certain area,

are not representative, transferable or reproducible to different scales. Therefore, such

approaches are characterized by higher levels of uncertainty and depend on a number

of critical assumptions (Kok and Veldkamp 2001, Van Delden et al. 2011; Veldkamp
et al., 2001, Verburg et al. 2006). Moreover, it is a common assumption that the

modelling results are highly affected by the quality and the technical details, such as

the pixel size of inputs and the bias they entail (Kocabas & Dragicevic, 2006; Van

Delden et al. 2011).

Models designed to analyze LULC dynamics can be divided into categories according

to their perspective, their domain, the methodological framework they apply, their

spatial or non-spatial nature etc (literature reviews by Agarwal et al., 2002, Briassoulis

2000; Schrojenstein Lantman et al. 2011). However, LULC models that solely rely on

statistical approaches often suffer from limitations such as sensitivity to outliers and

noise, collinearity issues and factors compatibility (Dormann et al. 2013; Eastman et al.

2005). On the other hand, more recently, a variety of models pertaining to artificial

intelligence, such as agent-based models, have been successfully applied for addressing

the complex, non-linear behavior of human-nature interactions and decision making.

This type of models, however, are suitable to capture processes at the individual,

household or neighborhood levels and when it comes to agent behavior they can be

very complex and are often parametrized with qualitative social survey data and other

types of participatory approaches (Zagaria et al. 2017)

Cellular automata (CA) consist of a dynamic simulation framework where space is

represented as a grid of cells and time is considered as discrete unit. The basic principle

of this type of LULC modeling framework is that the state of a given pixel is determined

by taking into account its previous state, the spatial interactions with the surroundings

in a given neighborhood and a set of defined transition rules. These elements dictate

the possible change of a cell and can be expert-based or calculated from statistical

analysis of historical LULC changes (White and Engelen, 2000). A growing body of

the literature demonstrate that, although very simple, CA models have the strong ability
to represent rich LULC patterns and handle nonlinear, stochastic and spatially explicit

LULC processes (Sante et al 2010).

The biggest advantage of CA is that they are fully consistent with Geographic

Information Systems (GIS) and remote sensing. Additionally, CA can be coupled with

other types of models and thus they are flexible to allow the elaboration and extension

of the methodological procedures according to the needs of a case study (Aburas et al.

2016). For instance, CA have been previously combined with a plethora of modeling

frameworks such as Markov chains (Jokar et al. 2013), neural networks (Li and Yeh,

2002) support vector machines (Yang et al. 2008) and kernel-based methods (Liu et al.

2008) among others. More recently, CA have been successfully combined with

Random Forests (RF) (Kamusoko and Gamba, 2015; Gounaridis et al. 2018a).

RF is a tree structured machine learning algorithm that generates a “forest” of

randomized independent to each other and identically distributed decision trees. Each

individual tree is composed with a random selection of the predictor variables and by

searching across a randomly selected subset, it predicts the target response, casting a

unit vote. This process is repeated until a user-defined number of trees has been built.

The outputs are determined from the majority of votes by each individual tree. For a

full detailed description of the RF algorithm, theory and applications, the reader is

referred to Breiman (2001). The independency of each individual decision tree and the

randomness in forming subsets of the input data makes RF insensitive to outliers, to

noise and to overfitting (Chan and Paelinckx, 2008). Additionally, normal distribution

of inputs is not a prerequisite and thus it can handle heterogenous data from various

sources, units and scales (Gounaridis et al., 2014; Gounaridis et al., 2016; Gounaridis

and Koukoulas, 2016). Another important advantage of RF is that it can handle large
datasets with thousands of imputs being accurate and at the same time computationally

faster (Rodriguez-Galiano et al. 2012).

The aim of this paper is, therefore, to explore potential future LULC dynamics in the

Attica region, using a CA modelling approach with scenarios that reflect different

economic performance realities and alternative planning options. The central premise

is to simulate all categories of LULC change at the regional level and to evaluate the

effects of different proximate and underlying causes. In order to spatially associate the

spatial determinants (proxies) with the observed historical changes, a set of factors

derived from multiple sources and expressed in different scales, units and resolutions

are incorporated in the modelling framework. A multi-resolution sensitivity analysis is

also carried out to assess the effect of spatial resolution of the input data to the model

outputs. The results will quantify the importance of various spatial determinants

(proxies) of change and shed light to the effect different economic performance realities

and land-use planning choices can have on the landscape.

2. Study area

The study area is the region of Attica in mainland Greece, an example of the rapid

socio-economic transformations that occurred in the country during the last decades,

including the demographic dynamics and population redistribution. The region includes

Athens, the capital city of Greece and the country’s major economic hub. According to

the latest census (2010), the region of Attica is inhabited by about 4 million people, or

35% of the country’s total population. In more recent decades, economic and population

growth triggered a persistent increase in housing demand and supply, and the

redistribution of middle-class Athenians to the outskirts of Athens (Mantouvalou et al,

1995; Leontidou et al. 2007). Additionally, socio-economic conditions favoured a

persistent amenity-driven trend for second homes along the coastal zone, albeit within
a commuting distance from the city-centre (Arapoglou and Sayas, 2009). As a

consequence, the landscape of peri-urban Athens has changed substantially. The urban

growth trend was indirectly emboldened by the weak presence of land use planning

checks and controls, which permitted the unhindered development at any

environmental, social or long-term economic cost (Pagonis, 2013). Moreover, after

successfully attracting national and foreign funds, and in preparation to host the

Olympic Games of 2004, the demand for construction sites to accommodate

commercial, industrial, transportation and recreational activities further increased the

built-up transformation of the urban periphery (Chorianopoulos et al. 2010). After the

phase of a rather stable economic growth, however, the area has recently been exposed

to the negative consequences of the sovereign debt crisis and the succeeding economic

recession (2010-2016). The decrease in purchasing power and a drastic drop in

consumer demand affected both the housing and the construction industries (Gounaridis

et al. 2018a).

Figure 1 about here

In terms of its topography, Attica also constitutes an interesting study case since it is

characterised by an undulated morphology (Figure 1). Mount Parnitha (elevation

1413m), Pateras (elevation 1132m), Penteli (elevation 1109m), Hymettus (elevation

1026m) and Egaleo (elevation 468m) are the main mountain ranges. These

geomorphological features separate the city of Athens from the adjacent flat districts of

Thriasio, Messoghia and Marathonas (Figure 1), which are the only available areas to

host residential and industrial settlements.

3. Material and methods

3.1 LULC Data

Five Landsat-based LULC maps spanning 25 years (1991, 1999, 2003, 2010, 2016) at

30m spatial resolution were used for the modelling. These maps were generated by

devising a semi-automated sampling extraction based on a context that combined the

no-change areas, spectral controlling, and prior knowledge of the area (Gounaridis et

al. 2018b). Overall accuracy for all maps is above 90%. Most importantly, the maps

come with a very high thematic resolution, achieved after disaggregating the urban-

related LULC categories (Gounaridis & Koukoulas, 2016). Specifically, the maps

depict eight land cover categories: i) continuous urban fabric, ii) discontinuous dense

urban fabric, iii) discontinuous medium density urban fabric, iv) discontinuous low

density urban fabric, v) industrial, commercial and transport units, vi) arable land and

permanent crops, vii) forests, scrubs and other natural areas and viii) other (includes

open spaces bare, mines and inland water bodies).

3.2 Transition probability modelling

Exploring future LULC patterns is a useful experiment for evaluating the causes and

identifying the impact of LULC changes. The scenario-based simulations have been

proven to be a useful way to sketch out how LULC patterns can evolve under different

pathways with a level of plausibility (Greeuw et al. 2000; Rounsevell and Metzger,

2010). Scenario-based analyses involves a certain degree of uncertainty originating

from the very nature of socioeconomic predictions that help define the scenarios. This

is due to the inability to foresee any unexpected circumstances and integrate any

emerging discontinuities or the data inputs for the models. Especially when dealing

with complex systems, such as LULC changes, assumptions are unavoidable. The level

of uncertainty can be minimized by combining an empirical analysis and sketching

different scenarios, attributes and conditions that deviate from historic trends in LULC

changes (Brown et al. 2013; Verburg et al. 2016)

3.2.1 Predictor variables

Taking into account previous LULC change modelling efforts (Gounaridis et al.

2018a), as well as data availability, a suite of 27 variables were chosen to best describe

the LULC change processes that took place throughout Attica in the study period (1991-

2016; Table 1). They are both categorical and continuous in nature and cover a broad

spectrum of potential LULC change factors. They can act as spatial determinants of the

changes that occurred during the last decades in Attica, and are derived from multiple

sources, with different scales and resolutions.

Table 1 about here

During the study period, changes related to artificial surfaces were dominant in Attica

and, therefore, the majority of the chosen variables represent factors that affect the

decision-making process when selecting locations for the construction of new housing

or infrastructure. Factors pertaining to social shifts, economic motives, inherent quality

and attractiveness of a given place and proximity to basic needs and amenities were

assumed to play a key role (Table 1). Variables related with the topography of the

terrain, such as elevation, slope and aspect influence the inherent quality of a certain

location and define the land suitability for built-up expansion. Proximity to the sea, to “blue-

flag” beaches (Foundation for Environmental Education- http://www.fee.global/), as well as to

natural reserves or urban green spaces are also perceived as added value in the pursuit for

a better quality of life and aesthetics for both primary or secondary homes. Proximity

to the city center of Athens or to nearest towns, to public transport, and the road density

are proxies that reflect the commuting distance to work. Additionally, distance to social
infrastructure including, among others, health provision, education and sports facilities,

together with the density of private enterprises (kernel density of geo-tagged newly

developed enterprises at 30m spatial resolution) serve as proxies to amenities.

Demographic and socio-economic variables such as changes in population density,

employment and unemployment rates provide insights on the shifts in the socio-

economic profile of the area estimated at local authority (municipal) level.

It is worth noting that, variables available at a higher administrative level, that of the

region, were not included since in Greece, implementation of local land use

management policies falls under the remit of local municipal authorities. Factors

expressed at the municipal level, therefore, were considered to represent an appropriate

spatial unit for our analysis (Panori et al. 2016; Gounaridis et al. 2018a). All spatial and

non-spatial datasets were collated in a GIS environment. Census data were mapped at

the municipal level while distances were computed using the Euclidean distance

function. The variables were then converted (resampled with bilinear interpolation) to

30m spatial resolution rasters to match the resolution of the Landsat-based land cover

classifications (Gounaridis et al. 2018b).

3.2.2 Leap-frog development index (LFDI)

To enhance the accuracy of the model, and to ensure the accurate detection and

representation of scattered development, the Leap-frog development index (LFDI),

originally proposed by Xu et al. (2007) was calculated and included in the modelling

scheme. Leap-frog development refers to the new urban patches that are formed

spontaneously and have no direct spatial connection and shared boundaries with the

existing urban patches. The index applies to artificial LULC types and has been proved

to effectively delineate any type of scattered development, classifying the historical

changes according to sharing boundaries properties (Gounaridis et al. 2018a).

Specifically, the index is calculated after dividing the length of the common boundaries

between newly developed urban patches and already existing urban patches with the

perimeter of the newly developed urban patches Xu et al. (2007). When the resulting

value is higher than 0.5 the growth type is denoted as infilling. A resulting value lower

than 0.5 denotes the edge growth while when the result is 0, it denotes the absence of a

shared boundary, and the growth is identified as Leap-frog development. Therefore,

following the approach by Gounaridis et al (2018a) the maps of 1991 and 2016 were

converted to vector format and patches representing the four urban categories and

industrial commercial and transport units were assigned values denoting which patches

appeared in each date. Subsequently, using common functions in GIS, the length of

common boundaries, their perimeter and the index were calculated. The last step was

to convert the vector file to raster format at 30m spatial resolution.

3.2.3 Random Forests regressions

Following the approach adopted by Gounaridis et al. (2018a), the transition probability

surfaces were generated by employing the machine learning regression algorithm of

Random Forests (Breiman, 2001) using all variables, including the LFDI. Eighteen

possible transitions were identified (Table 2), under three assumptions: (a) it is

impossible for the urban fabric class to convert to any other land type as well as to

decrease in density; b) the industrial, commercial and transport units cannot convert to

any other land type, and c) the “other” category, that includes inland waters, bare land

and mines, cannot interact with any of the other 7 classes. To train each of the 18

models, 5000 random points were dispersed throughout the study area. Two possible

values were associated with these training points: 1 denotes change from any LULC

class to any other class, and 0 denotes no change. The RF regressions were then

implemented in R using the RandomForest package (Liaw & Wiener, 2002). To fine
tune the RF regressions, five predictor variables (equal to the square root of the total

number of 27 predictor variables) were used for each tree split and 700 trees for each

run. The modelling process generated 18 transition probability surfaces, each indicating

the degree of potential future LULC change.

Table 2 about here

RF also offer meaningful metrics about the importance of each predictor variable. To

quantify the importance and contribution for each of the 27 predictor variables, two

metrics, the Mean Decrease Gini and the Mean Decrease Accuracy were computed

(Gounaridis and Koukoulas, 2016). The mean decrease in Gini coefficient informs

about each variable’s contribution to the impurity of the resulting random forest model.

Variables with a high value in the decrease of Gini, tend to have nodes with high purity

which is a measure of model’s homogeneity. The mean decrease in accuracy provides

information about how much the accuracy would decrease if a variable were excluded

from the model. Therefore, the larger the value of mean decrease, the higher the

importance of a variable is.

3.3 LULC change scenarios

Figure 3 shows the LULC trends in Attica between 1991 and 2016, based on the

Landsat-based land cover maps. Three different phases of economic development and

performance can clearly be identified and based on these, we devised the following

three potential future scenarios:

Figure 3 about here

Low development scenario: this scenario reflects the 2010-2016 period, when urban

expansion rates curtailed significantly as a consequence of economic recession and a

drop in investment spending. Throughout this time, for instance, approximately

150,000 newly built houses in the region were left uninhabited (unsold), while over one
third of commercial facilities in the city of Athens closed down and remained shut

(Serraos et al., 2016). Under this scenario, economic growth, as well as the population

mobility would remain relatively stable.

Medium development scenario: this scenario reflects the period between 1991 and

1999, when the peri-urban areas of Athens conurbation, especially the uplands and the

Messoghia plain, experienced significant population gains. Increase in demand for new

houses boosted urban growth at the expense of other less profitable land uses, bringing

gradually major changes in the peri-urban landscape. In fact, during this time, peri-

urban Athens population had grown ten times faster than the Athens conurbation

population, which remained relative stagnant (Petrakos and Mardakis, 1999).

According to this scenario, peri-urban Athens experiences a steady population increase,

fueled by the relocation choices of Athenians questing residence in lower density areas.

High development scenario: this scenario reflects the sharp urban expansion rates noted

in the region in the 2000-2009 period, facilitated by stable economic growth and the

continuation of a rather “loose” approach to land use planning controls. The era is

chronologically framed by the effects of the 2008 global financial crisis, which were

felt locally, however, in late 2009, in the form of an excessive budgetary deficit and a

prolonged recession (Chorianopoulos and Tselepi, 2017). Under this scenario,

population decentralization from Athens conurbation continues apace, further enhanced

by labour migration from outside the country, fueling demand for new housing

constructions. Following the development trends of the reference period, the spotlight

of investment falls on the waterfront areas shifting further real estate dynamics towards

tourism-related facilities and secondary homes. Similarly, spatial planning continues to

play an important role in the “construction spree” by approving investment in

transportation infrastructure, enhancing peri-urban accessibility to the city of Athens.

Consequently, major infrastructure works, private enterprises and shopping centres will

keep on colonizing the Northern outskirts (Maranthon, Oropos, Messoghia and the

Thriasio plains forming a ‘suburban exploitation thesis’ case (Pacione, 2009).

All three scenarios draw from clear reference periods and assume that profound social

and political changes will not alter their traits. As far as the land use planning apparatus

is concerned, it is assumed that it will continue to be rather permissive to development,

following a political stance that approaches unregulated urban expansion as a “shortcut”

to economic growth.

3.4 Model calibration

The CA model was designed and implemented using the Dinamica EGO platform

(Soares-Filho et al. 2002). An important step, prior to the prediction phase, is model

calibration. To calibrate the model and evaluate the goodness of fit, a comparison of

simulated maps with reference maps is the most efficient way (Gounaridis et al. 2018a).

Any CA modelling framework involves four components: the probability maps, the

historical LULC maps, the transition rules and the neighborhood characteristics that

define the parameters of the simulation.

The CA model was trained based on the 1991-2010 period, and the observed changes

were used to predict the landscape structure and composition on 2016. To do so, the

annual rates of change per LULC category between 1991 and 2010 were calculated

generating a transition matrix. In order to replicate the actual structure and composition

of the area, three landscape metrics were computed: (i) the mean patch size, (ii) the

variance of patch size, and (iii) the patch isometry. In general, an increased patch size

results in less fragmented landscapes, while the patch size variance denotes the diversity

of newly developed patches. Isometry usually varies from 0 to 2 and thus, the greater
the isometry, the more isometric (i.e. equal) the newly developed patches are. The first

two metrics were computed for the input LULC map (2010) while the latter was

adjusted through a trial-and-error process. Finally, the 18 transition probabilities were

stacked together to drive the allocation of cells, based on the premise that the cells with

the highest likelihood values should change first. The model was then set to run and

predict LULC for 2016.

To evaluate the model's performance, the simulated LULC map of 2016 was compared

with the observed LULC map of 2016 (i.e. the outcome of the Landsat-based

classification; Gounaridis et al. 2018b) using the fuzzy similarity index at multiple

resolutions (Hagen, 2003). This index evaluates the accuracy of simulation results

considering the similarities of two maps (simulated versus observed) in a neighborhood

context and within increasing window sizes (Mas et al. 2012). This involves the

comparison of map fit and spatial agreement within a certain pixel vicinity allowing the

comparisons of maps not only in a strict pixel-by-pixel basis but also considering the

spatial similarity in multiple resolutions (Hagen, 2003). To gain insights about per class

agreement we also computed the error matrix between the simulated and the observed

maps of 2016. The sampling was based on 9399 samples holding LULC class values of

2016 (Gounaridis et al. 2018b). The samples come with relatively equal distribution

among the LULC classes that ensures equal representation.

After calibration, the simulation of LULC changes under the three scenarios was

implemented, taking 2016 as the initial year and 2040 as the final year, in a 5-year time

step. The parameters used to calibrate the model were kept constant and only the

quantity of LULC transitions per scenario were changed. A transition matrix was

constructed for each epoch, i.e. 1991-1999, 1999-2010 and 2010-2016, to reveal the

quantity of each possible transition per scenario (Table 2). Ideally, the predictor, and in
turn, the transition probability surfaces, would also change per scenario, to better reflect

the socio-economic conditions of each epoch. However, in our case, this option was not

feasible due to data availability and temporal mismatch issues.

3.5 Multi-resolution sensitivity analysis

After completing the model simulations at 30m spatial resolution, a sensitivity analysis

was also conducted at various spatial resolutions. It was hypothesized that when all

other parameters of the model are held constant and only the spatial resolution of inputs

changes, then the quantities, the spatial allocation and thus, the spatial patterns of

outputs, can differ. The central premise behind this step was that the spatial resolution

of the models’ inputs can have important and substantial effects on the output. Thus,

this parameter can limit or even enhance the ability of a model to project future

scenarios of LULC change. Sensitivity analysis is a process that examines the variation

in model outputs in response to variation in a set of model parameters, in this case the

spatial resolution of input data. To do so, the 1991 and 2016 Landsat-based

classifications were resampled (nearest neighbor) to 100m, 250m and 500m,

respectively and change detection was performed for each case. Next, the transition

probabilities were re-constructed through RF regression after resampling (bilinear

interpolation) all predictors for each case. The calibration followed the same steps as

aforementioned. The landscape metrics along with the transition quantities were re-

calculated and introduced to the models for each case. After calibration, each scenario

was simulated based on the transitions observed throughout each of the three epochs.

Finally, all maps generated from each run were overlapped using rule-based cross

classification in order to produce the final map per scenario. This step identified areas

of change that are common regardless the spatial resolution of the inputs. To explore

the influence of the spatial resolution on various consecutive steps of the modelling
process, we compared the transition probability surfaces produced at the native

resolution (30m) and at several coarser resolutions (100m, 250m, 500m). This was done

after sampling the transition probability surfaces at 1000 random points, and computing

the concordance correlation coefficient (Lin, 1989; 2000).

4. Results and discussion

4.1 Model calibration and performance

One common way to assess the level of model calibration and performance is to

compare the simulated map for a given year versus the observed map, which is often

derived from the classification of satellite data. Figure 4 depicts the resulting map of

2016 after calibration versus the reality (observed map of 2016). A visual comparison

of these maps shows the relatively high similarity. This suggests that the RF-CA model

was relatively accurate at allocating the LULC patterns of change in the study area.

Table 3 reveals the level of agreement per class between the simulated map of 2016 and

the observed map of 2016. Overall accuracy was acceptably high (88.36%) and the User

and Producer accuracies for all classes ranged from 83.4% to 96.5%. Regarding the

disagreements, confusion is evident between certain classes that are mostly spatially

adjacent. For instance, between “discontinuous medium density urban fabric” and

“discontinuous low density urban fabric”, as well as between “discontinuous low

density urban fabric” and “arable land and permanent crops”.

Figure 4 about here

Table 3 about here

Figure 5 illustrates the fuzzy similarity index computed based on the overlay of the two

maps. The accuracy assessment yielded a spatial fit of 85.18% within the 1x1 window

size radius which improved to 95.08 % when widened to a 15x15 window size. The

high scores in performance suggest that the suite of 27 predictor variables were used

efficiently and the RF algorithm performed well with an adequate fit.

Figure 5 about here

Figure 6 depicts the components of agreement and disagreement between the simulated

versus the observed maps. It reveals information about: (i) observed change simulated

correctly as change (i.e. hits); (ii) observed persistence (i.e. LULC that remained

unchanged) simulated correctly as persistence (i.e. null successes); (iii) observed

change simulated incorrectly as persistence (i.e. misses), and (iv) observed persistence

simulated incorrectly as change (i.e. false alarms). Most importantly, the model

predicted accurately the leap-frog development and this proves the added value of the

LFDI and the extensive training of the RF model.

Figure 6 about here

4.2 Variable importance

Figure 7 is the Mean Decrease in Gini coefficient which informs about each variable’s

contribution to the impurity of the resulting random forest model. Road density,

enterprises density and elevation contributed the most for changes related to dense

urban fabric. The same variables along with the distance to shoreline and education

centers are the most related to discontinuous dense and medium density urban fabric.

For the discontinuous low density urban fabric, which is a category broadly related to

second homes, distance to shoreline, to blue flag beaches, elevation, road density and

enterprises density were the most influential variables.

 Figure 7 about here

Figure 8 is the mean Decrease in Accuracy which informs about how much the accuracy

decreases if a variable would be excluded from the model. According to this, road

density, distance to natural reserves, to prefecture center and to shoreline, as well as

slope and elevation were the most influential variables for changes related to dense

urban fabric. The same variables along with the distance to beaches, to urban green

areas and to public buildings were the most influential to changes related to

discontinuous dense and medium density urban fabric. For the discontinuous low

density urban fabric, the elevation, slope, road density along with the distance to urban

green, to shoreline, to natural reserves and to prefecture center contributed the most into

the spatial changes description.

Figure 8 about here

4.3 Multi-resolution sensitivity analysis

The models yield similar patterns for each scenario but, as anticipated, as the resolution

increases, the patterns tend to become more aggregated and smaller patches of change

tend to be lost. Figure 9 depicts the concordance correlation coefficient (Lin, 1989,

2000) derived from transition probabilities for the continuous urban fabric class per

different spatial resolution. The higher concordance value can be observed between the

30m and 100m pixel size. Gradually, as the difference in spatial resolution increases so

is the distance of data's reduced major axis from the line of perfect concordance which

reflects the concordance between the transition probability surfaces.

Figure 9 about here

The multi-resolution sensitivity analysis results provide evidence that the technical

characteristics have substantial impact to the outputs of a model and thus to the
observed patterns and to the conclusions drawn. Even if a model is rigorously

calibrated, the predictability will decrease relative to the spatial resolution, and the

patterns revealed in the results will become less informative.

4.4 Model predictions for 2040

Figures 10-12 depict the LULC changes projection under the three scenarios while

Figure 13 provides a quantified insight to the final results.

Figure 10 about here

Under the medium economic development scenario (Figure 10), and with a pace of

urban growth equivalent to that of 1991–1999, artificial surfaces are expected to expand

predominantly at the expense of other, less profitable, land uses. Urban areas are

anticipated to reach 41% of the region’s surface, of which 17% will be discontinuous

low density urban fabric. Industrial areas are expected to occupy almost 8% of the total

area. At the same time agricultural land is expected to decline from 23.5% in 2016 to

10% in 2040 (Figure 13). Most changes will occur along the waterfront and in the

periphery of Athens conurbation, effecting notable changes in Messoghia, the

Thriassian plain, Marathonas, Oropos and Sounio. In these areas, pre-existing urban

and industrial clusters portray a tendency to become denser and to expand considerably,

ending up almost connected with Athens conurbation, especially in its northern parts.

The region’s coastline, especially in Messoghia, Marathonas, Oropos and Sounio, is

also expected to exhibit remarkable changes. Existing towns display a tendency to

become denser and to expand, transforming waterfront areas into a large and solid low

density urban patch. Leap-frog development is also expected to increase sharply around

road junctions of existing urban areas.

Figure 11 about here

Under the high economic development scenario (Figure 11), where the pace of urban

growth reflects the traits of the 1999-2010 period, artificial surfaces are expected to

increase remarkably. At the same time, they are expected to occupy more than half of

the total surface of Attica region (56.7%). In more detail, urban uses, are expected to

occupy an area of almost 48% in 2040, which can be translated to an increase of

approximately 21%. In this land use category, discontinuous low density urban fabric

will reach a high peak of almost 21% of the total area. The continuous dense,

discontinuous high density urban fabric and discontinuous medium density urban fabric

are expected to reach 9%, 10% and 8% respectively. At the same time agricultural areas

are expected to decrease by 18%, occupying only 5.2% of the total area (Figure 13).

All these accelerated landscape transformations are expected to occur throughout Attica

region, leading to a mosaic of mixed land uses. Pre-existing urban and industrial

clusters will become denser and expand considerably. In a similar fashion with the

medium growth scenario, most changes are observed along the coastline and to the peri-

urban zones of Athens conurbation. Changes are expected to be centered on the

northern suburbs of Athens, the Messoghia and the Thriassian plain, Marathonas,

Oropos and Sounio areas. Most notably, existing urban patches in the waterfront

(Marathonas, Messoghia, Sounio, western Attica and Oropos), are expected to be linked

with the conurbation forming an urban-rural continuum of low, and at places, medium

density. In the western part of Attica, the Thriassian plain is expected to experience a

considerable increase in industrial development and a notable increase in medium

density urban use. Last but not least, the density of urban areas will increase sharply,

especially in the northern and eastern suburbs of Athens.

Figure 12 about here

Under the low development scenario (Figure 12), an increase in artificial surfaces of

approximately 6% is noted in the region; a rate, however, that is significantly lower if

compared with the other two scenarios. Discontinuous low density urban fabric, for

instance, is expected to occupy 15% of the total area by 2040, an increase of only 3%

since 2016. Similarly, continuous dense and discontinuous high density urban fabric

are expected to reach 6.7% and 5.6% respectively (Figure 13). Following the traits of

the recession (2010-2016), urban expansion is observed throughout the region, yet at

relatively moderate rates and in rather compact form. Foreseen changes will mostly

occur around the road network and in the waterfront areas, particularly in the eastern

and northern parts of Attica. Already existing urban areas appear to increase in density,

rather than expanding, while leap-frog development is noted in areas of adequate

transportation infrastructure, guaranteeing ease of access to Athens. Regarding urban

density, slightly higher densities are expected in the northern suburbs of Athens.

Figure 13 about here

5 Discussion

5.1 RF-CA land use/cover modelling

The coupling of CA and RF proved to be a sound way to combine the advantages of

each approach. Implementing the RF algorithm for transition potential modelling,

allows the efficient combination of qualitative and quantitative data derived from

multiple sources and with different nature in terms of scale and origin. In addition, RF

proved insensitive to collinearity issues and normal distribution of data was not a

prerequisite. The predictors incorporated in the models proved capable to spatially

determine the phenomenon while the incorporation of the Leap-frog development index

at the regional level, assisted the models in LULC prediction. In this approach, a total
of 18 distinct transitions were identified and equal transition probability surfaces were

generated. Their combination in a CA modelling environment seemed challenging and

required intense training and calibration through trial and error. Currently, most LULC

models can only simulate limited possible transitions due to complexity in definitions,

attributes and transition rules (Liu et al. 2017). However, in reality, even in the same

location, different LULC dynamics occur simultaneously and affect each other. Thus,

a comprehensive outlook of these processes is much more effective in order to

realistically determine the future trajectories. The interactions and competition among

different types of LULC was explored by using a simple, yet effective competition

mechanism. This approach allows the incorporation of various LULC transition

probabilities as a single layer stack, containing all the probability surfaces. Each layer

represented one single possible transition, while each cell contained values denoting

the dominant LULC type and the likelihood to retain the current land type or transform

to another type. The reproduction of LULC patterns and the calibration procedure, as a

whole, improved considerably with the inclusion of the mean patch size, the variance

of patch size, and patch isometry. Introducing these metrics to the CA framework,

allows the models to take into account and to reproduce the actual parameters of the

study area. The adoption of the fuzzy similarity index (Hagen, 2003) for assessing the

model’s spatial fit was another advantage of the approach, as it performs comparisons

of simulated versus observed data within a neighborhood context, and not in a strict

per-pixel context.

5.2 LULC predictions for 2040

We employed three socioeconomic and associated urban growth scenarios to explore

potential LULC pathways to 2040. The ‘low development’ scenario draws from the

current economic austerity and recession reality, framing a long-term setting in which
economic downturn keeps on hindering urban expansion dynamics. Results obtained

from the medium and high economic development scenarios, however, are

multifaceted. Both scenarios shed light on the ways in which Attica would look like

when the current economic crisis is reversed. Against this backdrop, they point to the

critical role of land use planning in regulating urban expansion. Our results outline a

future landscape shaped by the unmediated prerogatives of rapid economic

development. They also underscore the significant socio-economic consequences such

as enhanced residential segregation, high infrastructure investment costs, central areas

underfunding. Moreover, the environmental impact such as increased car dependency

and usage, loss of agricultural land and natural habitats are also evident. In light of the

consequences, LULC changes that would occur locally are expected to create a

maladaptive and nonfunctional setting, liable to undermine future economic

development prospects (Chorianopoulos, et al. 2010).

Regarding the factors that contribute and the extent of this contribution to the different

types of LULC change, our study incorporated a total of 27 variables into the modelling.

By implementing 18 different models representing every permitted LULC transition,

the contribution of each factor was quantified using the Mean Decrease Gini and the

Mean Decrease Accuracy metrics. From the application of these models, three

messages emerge:

(a) Firstly, the results demonstrate that depending on the LULC type, different factors

play a key role in the spatial configuration of LULC change (Kizos et al. 2018). The

interrelationships of urban related classes, for example, can clearly be distinguished

according to their density, which translates to different residential use (e.g. secondary

homes). In densely built urban areas, spatial factors, such as road network density,

density of enterprises, proximity to social infrastructure (health services, educational

institutions) and accessibility to the municipal centres, were the dominant determinants

of change. In urban areas with lower density, distance to the shoreline and to “blue-

flag” beaches were among the most important. The results are in agreement with the

findings of other studies, especially with those related with the coastal zone of the

Mediterranean (Boavida-Portugal et al. 2016; Houet et al. 2016; Lagarias, 2012;

Marraccini et al. 2015; Petrov et al. 2009).

(b) Secondly, some factors that rank among the top determinants for a type of LULC

change, may have a strong positive or negative correlation coefficient with the

phenomenon. For instance, the slope and elevation variables, rank high in the urban

categories with strong negative correlation coefficients, mostly due to the topography

of Attica, limiting the majority of settlements within the plains.

(c) Finally, a possible important limitation that should be noted is that, all these patterns

and numbers are case-specific and the conclusions drawn from the quantitative insights

might not be transferable to other regions. This is mostly due to specificities present

only in Attica, for example, the physical constraints related to the topography, the

cultural choices for primary and secondary housing, or the presence or absence of a

regulatory planning mechanism. Future research directions should include cross-cases

comparison with areas that share common characteristics with Attica, e.g. coastal areas,

Mediterranean administrative regions that include a big metropolitan area and areas

with rapid socio-economic changes.

5.3 Multi-resolution sensitivity analysis

This paper demonstrated the importance of a multiple scales analysis, by incorporating

in the modelling framework data derived from multiple sources, expressed at various

scales and resolution. Given that, the data used as input in any model, affect the
outcomes, and in turn the usefulness and the accuracy of the model, studies that utilize

only data that concern a single scale or spatial resolution, fail to account for a wide

range of information. Moreover, their transferability is limited (Veldkamp et al. 2001).

Data expressed at coarse scales might hold information and patterns that are

undetectable at finer scales and vice versa (Brown et al. 2013; Van Delden et al. 2011;

Verburg et al. 2004). Furthermore, factors that determine a LULC change, might

operate at a distance from the area of focus. Thus, when dealing with a system that

involves multiple nonlinear relationships and various proximate and underlying factors,

it is necessary to consider all available information (Larondelle & Lauf, 2016). Here,

we exploited all possible resources and efficiently combined and integrated the

available multi-scale and multi-resolution data.

Additionally, the simulation results were subjected to a multiple resolution sensitivity

analysis. Since the modelling approaches generate outputs that are more or less driven

by the parameters and characteristics of input data (Kocabas & Dragicevic, 2006; Van

Delden et al. 2011), the results obtained by this approach are consistent to all pixel sizes

and thus insensitive to the effect of pixel size.

4. Conclusions

This paper demonstrated an integrated approach to explore potential future LULC

dynamics under different scenarios that reflect different economic performances and

policy options. Our integrated framework was able to sufficiently: i) take into account

socioeconomic, biophysical, legislative and land use factors spanning a broad spectrum

of LULC change spatial determinants (proxies); ii) provide insights into hidden patterns

by taking into account, not only the prominent changes between major LULC

categories, but also changes in density; iii) take into account the multiple scales
involved in LULC systems, and, v) provide results that are insensitive to the spatial

resolution of the inputs.

Acknowledgments

The authors wish to thank the anonymous reviewers for their constructive feedback and

suggestions on an earlier version of this manuscript.

References

Aburas, M. M., Ho, M. Y., Ramli, M. F., & Ashaari, Z. H. (2016). The simulation and

prediction of spatio-temporal urban growth trends using cellular automata

models: A review. International Journal of Applied Earth Observation and

Geoinformation, 52, 380–389.

Agarwal, C, Green, G M, Grove, J M, Evans, T P & Schweik, C M. (2002). A review

and assessment of land-use change models: dynamics of space, time, and human

choice. General Technical Report NE-297, Department of Agriculture, Forest

Service, Northeastern Research Station, Newtown Square, PA.

Arapoglou, V.P. and Sayas, J. (2009). New facets of urban segregation in southern

Europe. European Urban and Regional Studies, 16(4), 345-362.

Batty, M., Couclelis, H., & Eichen, M. (1997). Urban systems as cellular automata.

Environment and Planning B, 24(2), 159–164.

Boavida-Portugal, I., Rocha, J. & Ferreira, C.C. (2016). Exploring the impacts of future

tourism development on land use/cover changes. Applied Geography, 77, 82-

91.
Breiman, L. (2001). Random forests. Machine Learning, 45, 5–32.

Briassoulis, H. (2000). Analysis of Land Use Change: Theoretical and Modeling

Approaches. In the Web Book of Regional Science (Loveridge S. Ed.), West

Virginia University, Morgantown.

Brown, D.G., Goovaerts, P., Bumlckl, A. & Li, M-Y. (2002). Stochastic Simulation of

Land-Cover Change Using Geostatistics and Generalized Additive Models.

Photogrammetric Engineering & Remote Sensing, 68(10), 1051-1061.

Brown, D.G., Verburg, P.H., Pontius Jr, R.G. & Lange, M.D. (2013). Opportunities to

improve impact, integration, and evaluation of land change models. Current

Opinion in Environmental Sustainability, 5, 452–457.

Chan, J.C.W. & Paelinckx, D. (2008). Evaluation of Random Forest and Adaboost

treebased ensemble classification and spectral band selection for ecotope

mappingusing airborne hyperspectral imagery. Remote Sensing of Environment,

112 (6), 2999–3011.

Chorianopoulos, I., Pagonis, A., Koukoulas, S. and Drymoniti, S. (2010). Planning,

competitiveness and sprawl in the Mediterranean city: the case of Athens.

Cities, 27 (4), 249-259.

Chorianopoulos, I. and Tselepi, N. (2017) Austerity urbanism: Rescaling and

collaborative governance policies in Athens. European Urban and Regional

Studies, DOI: 10.1177/0969776417733309

Dormann, C. F., Elith, J., Bacher, S., Buchmann, C., Carl, G., Carré, G., et al. (2013).

Collinearity: a review of methods to deal with it 2012. Collinearity: a review of

 methods to deal with it and a simulation study evaluating their performance.

Ecography, 36, 27–46.

Dongjie, G., Weijun, G., Kazuyuki, W. & Hidetoshi, F. (2008). Land use change of

Kitakyushu based on landscape ecology and Markov model. Journal of

Geographical Science, 18, 455–468.

Eastman, J.R., Solorzano, L.A., & van Fossen, M.E. (2005). Transition potential

modeling for land-cover change. In GIS, Spatial Analysis, and Modeling;

Maguire, D.J., Batty, M., Goodchild, M.F., Eds.; ESRI Press: California, UK,

357–385.

Gibson, C.C., Ostrom, E. & Ahn, T.K. (2000). The concept of scale and the human

dimensions of global change: a survey. Ecological Economics, 32, 217–239.

Gounaridis, D., Zaimes, N.G. & Koukoulas, S. (2014). Quantifying spatio-temporal

patterns of forest fragmentation in Hymettus Mountain, Greece. Computers,

Environment and Urban Systems, 46, 35–44.

Gounaridis D., Apostolou A. & Koukoulas S. (2016). Land Cover of Greece, 2010: a

semi-automated classification using Random Forests. J Maps, 12(5), 1055-

1062.

Gounaridis, D. & Koukoulas, S. (2016). Urban land cover thematic disaggregation,

employing datasets from multiple sources and RandomForests modeling.

International Journal of Applied Earth Observation and Geoinformation, 51,

1–10.
Gounaridis, D., Chorianopoulos, I. & Koukoulas, S. (2018a). Exploring prospective

urban growth trends under different economic outlooks and land-use planning

scenarios: The case of Athens. Applied Geography 90, 134-144.

Gounaridis, D., Symeonakis, E., Chorianopoulos, I. & Koukoulas, S. (2018b).

Incorporating density in spatio-temporal land use/cover change patterns: the

case of Attica, Greece. Remote Sensing 10(7), 1034.

Greeuw, S., van Asselt, M., Grosskurth, J., Storms, C., Klomp, N., Rothman, D. et al.

(2000). Cloudy Crystal Balls: An Assessment of Recent European and Global

Scenario Studies and Models: Experts’ Corner Report. Office for Official

Publications of the European Communities.

Hagen, A. (2003). Fuzzy set approach to assessing similarity of categorical maps.

International Journal of Geographical Information Science, 17(3), 235–249.

He, Z. & Lo, C. (2007). Modeling urban growth in Atlanta using logistic regression.

Computers Environment and Urban Systems, 31 (6), 667-688.

Houet T., Marchadier C., Bretagne G., Moine M.P., Aguejdad R., Viguié V. et al.

(2016). Combining narratives and modelling approaches to simulate fine scale

and long-term urban growth scenarios for climate adaptation. Environmental

Modelling and Software journal, 86, 1-13

Jokar A, J., Helbich, M., Kainz, W. & Darvishi B, A. (2013). Integration of logistic

regression, Markov chain and cellular automata models to simulate urban

expansion. International Journal of Applied Earth Observations and

Geoinformation, 21, 265–275.

Kamusoko, C. & Gamba, J. (2015). Simulating urban growth using a random forest

cellular automata (RF-CA) model. ISPRS International Journal of Geo-

Information, 4, 447–470.

Kizos, T., Verburg, P.H., Bürgi, M., Gounaridis, D., Plieninger, T. Bieling, C. &

Balatsos, T. (2018). From concepts to practice: combining different approaches

to understand drivers of landscape change. Ecology and Society, 23, 25-36.

Kocabas, V. & Dragicevic, S. (2006). Assessing cellular automata model behaviour

using a sensitivity analysis approach. Computers Environment and Urban

Systems, 30, 921-953.

Kok, K. & Veldkamp, A. (2001). Evaluating impact of spatial scales on land use pattern

analysis in Central America. Agriculture. Ecosystems and Environment, 85,

205–221.

Lagarias, A. (2012). Urban sprawl simulation linking macro-scale processes to micro-

dynamics through cellular automata, an application in Thessaloniki, Greece.

Applied Geography, 34, 146-160.

Larondelle, N. & Lauf, S. (2016). Balancing demand and supply of multiple urban

ecosystem services on different spatial scales. Ecosystem Services, 22, 18–31.

Leontidou, L., Afouxenidis, A., Kourliouros, E. and Marmaras, E. (2007).

Infrastructure-Related Urban Sprawl: Mega-Events and Hybrid Peri-Urban

Landscapes in Southern Europe, in Urban Sprawl in Europe: Landscapes, Land-

Use Change & Policy (eds C. Couch, L. Leontidou and G. Petschel-Held),

Blackwell Publishing Ltd, Oxford, UK.

Li, X. & Yeh, A. (2002). Neural-network-based cellular automata for simulating

multiple land use changes using GIS. International Journal of Geographical

Information Science, 16, 323–343.

Liaw, A., & Wiener, M. (2002). Classification and regression by randomForest. R

News, 2(3), 18–22.

Lin, L. (1989). A concordance correlation coefficient to evaluate

reproducibility. Biometrics 45, 255 - 268.

Lin, L. (2000). A note on the concordance correlation coefficient.

Biometrics 56, 324 - 325.

Liu, X., Li, X., Shi, X., Wu, S. & Liu, T. (2008). Simulating complex urban

development using kernel-based non-linear cellular automata. Ecological

Modelling, 211, 169- 181.

Liu, X., Liang, X., Li, X., Xu, X., Oua, J., Chen, Y. et al. (2017). A future land use

simulation model (FLUS) for simulating multiple land use scenarios by

coupling human and natural effects. Landscape and Urban Planning, 168, 94-

116.

Ma, X. & Zhao, X. (2017). Land Use Allocation Based on a Multi-Objective Artificial

Immune Optimization Model: An Application in Anlu County, China.

Sustainability, 7, 15632-15651.

Mallampalli, V.R., Mavrommati, G., Thompson, J., Duveneck, M., Meyer, S.,

Ligmann-Zielinska, A., et al. (2016). Methods for translating narrative scenarios

into quantitative assessments of land use change. Environmental Modelling &

Software, 82, 7-20.

Manson, S.M. (2005) Agent-based modeling and genetic programming for modeling

land change in the Southern Yucatan Peninsular Region of Mexico. Agriculture,

Ecosystems and Environment, 111, 47–62.

Mantouvalou M., Mavridou M. and Vaiou, D. (1995). Processes of social integration

and urban development in Greece: Southern challenges to European unification.

European Planning Studies 3(2), 189-204.

Marraccini, E., Debolini, M., Moulery, M., Abrantes, P., Bouchier, A., Chery, J-P. et

al. (2015). Common features and different trajectories of land cover changes in

six Western Mediterranean urban regions. Applied Geography, 62, 347-356.

Mas, J-F., Perez-Vega, A. and Clarke, K.C. (2012). Assessing simulated land use/cover

maps using similarity and fragmentation indices. Ecological Complexity 11, 38–

45.

McBurney, P. (2012). What Are Models for? 9th European Workshop on Multi-Agent

Systems. In: Cossentino M. et al. (Eds) LNAI 7541. Springer, Maastricht, The

Netherlands, 175–188.

Meentemeyer, V. (1989). Geographical perspectives of space, time, and scale.

Landscape Ecology, 3, 163–173.

Millington J.D. (2005) Wildfire risk mapping: considering environmental change in

space and time. Journal of Mediterranean Ecology, 6, 33–42.

Pacione, M. (2009) Urban Geography: A Global Perspective. Oxon, Routledge.

Pagonis, A. (2013). The evolution of metropolitan planning policy in Athens over the

last three decades: Linking shifts in the planning discourse with institutional

changes and spatial transformation. Changing Cities: Spatial, morphological,

 formal and socioeconomic dimensions. University of Thessaly. Skiathos, 18-21

June.

Panori, A., Ballas, D. & Psycharis, Y. (2016). SimAthens: A spatial microsimulation

approach to the estimation and analysis of small area income distributions and

poverty rates in the city of Athens, Greece. Computers, Environment and Urban

Systems, 63, 15-25.

Petrakos G. and Mardakis P. (1999) 'Recent changes in the Greek system of urban

centers' In D. Economou and G. Petrakos (eds) The development of Greek cities:

Interdisciplinary approaches to urban analysis and policy. Volos: University of

Thessaly (pp. 45-64).

Petrov, L.O., Lavalle, C. & Kasanko, M. (2009). Urban land use scenarios for a tourist

region in Europe: Applying the MOLAND model to Algarve, Portugal.

Landscape and Urban Planning, 92(1), 10-23.

Poelmans, L. & Van Rompaey, A. (2010). Complexity and performance of urban

expansion models. Computer Environment and Urban Systems, 34, 17–27.

Robinson D.T., Murray-Rust D., Rieser V., Milicic V. & Rounsevell M. (2012)

Modelling the impacts of land system dynamics on human well-being: Using an

agent-based approach to cope with data limitations in Koper, Slovenia.

Computers, Environment and Urban Systems, 36, 164–176.

Rodriguez-Galiano, V.F., Chica-Olmo, M., Abarca-Hernandez, F., Atkinson, P.M.,

Jeganathan, C. (2012). Random Forest classification of Mediterranean land

cover using multi-seasonal imagery and multi-seasonal texture. Remote Sensing

of Environment 121, 93–107.

Rounsevell, M. & Metzger, M. (2010). Developing qualitative scenarios and storylines.

Wiley Interdisciplinary Reviews: Climate Change 1(4), 606-619.

Sante, I., Garcia, A. M., Miranda, D. & Crecente, R. (2010). Cellular automata models

for the simulation of real-world urban processes: A review and analysis.

Landscape and Urban Planning, 96(2), 108–122.

Schrojenstein Lantman, J., Verburg, P.H. Bregt, A. & Geertman. S. (2011). Core

principles and concepts in land-use modelling: A literature review. Chapter

3 in Land-Use Modelling in Planning Practice, ed. E. Koomen and J.

Borsboom-van Beurden. New York: Springer.

Serraos, K. Greve T. Asprogerakas, E. Balampanidis D. & Chani. A. (2016). Athens,

a capital in crisis. Tracing the spatial impacts. In Knieling J. & Othengrafen, F.

(Eds.) Cities in Crisis, Socio-spatial impacts of the economic crisis in Southern

European cities. Oxon, Routledge (pp. 116-138).

Soares-Filho, B., Pennachin, C. L. & Cerqueria, G. (2002). DINAMICA- a stochastic

cellular automata model designed to simulate the landscape dynamics in an

Amazonian colonization frontier. Ecological Modelling, 154(3), 217–235.

Van Delden, H., Van Vliet, J., Rutledge, D.T. & Kirkby, M.J. (2011). Comparison of

scale and scaling issues in integrated land-use models for policy support.

Agriculture, Ecosystems and Environment, 142, 18– 28.

Veldkamp A. & Fresco L.O. (1996) CLUE: a conceptual model to study the Conversion

of Land Use and its Effects. Ecological Modelling, 85, 253–270.

Veldkamp, A., Verburg, P.H., Kok, K., de Koning, G.H.J., Priess, J. & Bergsma, A.R.

(2001). The need for scale sensitive approaches in spatially explicit land use

change modeling. Environmental Modeling and Assessment, 6, 111–121.

Verburg P.H., Soepboer W., Limpiada R., Espaldon M.O., Sharifa M. & Veldkamp A.

(2002) Modeling the Spatial Dynamics of Regional Land Use: The CLUE-S

Model. Environmental Management, 30, 391–405.

Verburg, P.H., Schot, P.P., Dijst, M.J. & Veldkamp, A. (2004). Land use change

modelling: current practice and research priorities. GeoJournal, 61, 309–324.

Verburg, P.H., Schulp, C.J.E., Witte, N. & Veldkamp, A. (2006). Downscaling of land

use change scenarios to assess the dynamics of European landscapes.

Agriculture, Ecosystems and Environment, 116, 39–56.

Verburg, P.H., Dearing, J.A., Dyke, J.G., van der Leeuw, S., Seitzinger, S., Steffen, W.

& Syvitski, J. (2016). Methods and approaches to modelling the Anthropocene.

Global Environmental Change, 39, 328–340.

Vitousek, P.M., Mooney, H.A., Lubchenco, J. & Melillo, J.M. (1997) Human

domination of Earth’s ecosystems. Science, 277, 494–499.

White, R. & Engelen, G. (2000). High-resolution integrated modelling of the spatial

dynamics of urban and regional systems. Computers, Environment and Urban

Systems, 24(5), 383–400.

Xu, C., Liu, M., Zhang, C., An, S., Yu, W., & Chen, J. (2007). The spatiotemporal

dynamics of rapid urban growth in the Nanjing metropolitan region of China.

Landscape Ecology, 22(6), 925–937.

Yang, Q., Li, X. & Shi, X. (2008). Cellular automata for simulating land use changes

based on support vector machines. Computers & Geosciences, 34, 592–602.

Zagaria, C., Schulp, C.J.E., Kizos, T., Gounaridis, D. & Verburg, P.H. (2017). Cultural

landscapes and behavioral transformations: An agent-based model for the

simulation and discussion of alternative landscape futures in East Lesvos,

Greece. Land Use Policy, 65, 26–44.

List of Figures

Figure 1. Location and topography of Attica region

Figure 2. Flowchart of the presented methodology
Figure 3. Summary of statistics based on the classified maps and relative percentages
of artificial areas. Three different trends reflect three different levels of development
during the last 25 years
Figure 4. The simulated map of 2016 versus the observed of 2016.
Figure 5. Multi-resolution evaluation of model fitting using the fuzzy similarity index.
Figure 6. Result of cross classification between the simulated vs the observed map of
2016
Figure 7. Mean Decrease Gini (IncNodePurity) as assigned by the RF regression
algorithm
Figure 8. Mean Decrease Accuracy (%IncMSE) as assigned by the RF regression
algorithm
Figure 9. Concorance correlation coeficient between transition probabilities for
continuous urban fabric category. The values derived from 1000 random samples,
dispersed across the study area
Figure 10. LULC spatial configuration simulated for 2040, under the Medium
development scenario
Figure 11. LULC spatial configuration simulated for 2040, under the High development
scenario
Figure 12. LULC spatial configuration simulated for 2040, under the Low development
scenario
Figure 13. Rates of LULC simulated for 2040 in a 5-year step, under the three different
development level scenarios
Tables

Table 1. List of predictors used in the transition potential modelling process.

Table 2. Transition probabilities of the eight LULC categories, allocated per scenario.
The numbers indicate transition rates per year in hectares.
Table 3. Error matrix - Simulated map (2016) versus the observed (2016). U.A: User’s
Accuracy; P.A: Producer’s Accuracy; O.A: Overall Accuracy. 1: Continuous urban
fabric. 2: Discontinuous dense urban fabric. 3: Discontinuous medium density urban
fabric. 4: Discontinuous low density urban fabric. 5: Industrial, commercial, and
transport units. 6: Arable land and permanent crops. 7: Forests, scrubs, and other natural
areas. 8: Other (open spaces, bare land, mines, inland water).
Figure 1
Figure 2
Figure 3
Figure 4
 1

4 Figure 5
5

10

11

12

13

14

15

16

17

1
 1

6 Figure 6
7

10

11

12

13

14

2
 1

4 Figure 7
5

10

11

12

13

14

15

3
 1

2 Figure 8
3

10

11

12

13

4
1

4 Figure 9
5

5
1

3 Figure 10
4

6
1

3 Figure 11
4

7
1

3 Figure 12
4

8
 1

8 Figure 13
9

10

11

12

13

9
 1 Table 1
Spatial
Variable Discription Source Time interval
resolution
Territorial variables
Elevation Elevation in m GLSDEM* (-) 30m
Slope Slope in degrees GLSDEM (-) 30m
Aspect Aspect in degrees GLSDEM (-) 30m
Climate Quality Climate quality index EEA* 1961-1990 1km
Visibility from residential areas at the parcel level GLSDEM and Urban
Viewshed (-) 30m
(centroids from UA) Atlas*
Euclidean distance from beaches signed with a blue Ministry of Environment
Distance from beaches 2010 30m
flag in m & Energy*
Distance from the sea Euclidean distance from the shoreline in m (-) 30m
Socio-economic variables
Distance from Education Euclidean distance from public education centers (all Ministry of Education &
2010 30m
centers levels) OSM*
Distance from public Society of Information*
Euclidean distance from public health centers (-) 30m
health centers & OSM
Euclidean distance from the center of the nearest
Distance from nearest
town (Markopoulo, Paiania, Koropi, Keratea, Artemida) OSM (-) 30m
town
in m
Distance from public Society of Information &
Euclidean distance from public buildings 30m
buildings OSM
Distance from public Euclidean distance from public hospitals and other
OSM (-) 30m
health public health care units in m
Distance from public Euclidean distance from public transport stops (bus,
OSM & opendata (-) 30m
transport metro, tram, suburban train) in m
Distance from road
Euclidean distance from road network in m OSM (-) 30m
network
Demographics Changes in population density at the municipality level ELSTAT* 1991-2011 30m
Total number of employed persons per total
Employment rate ELSTAT 1991-2011 30m
population at the municipality level
Total number of unemployed persons per total
Unemployment rate ELSTAT 1991-2011 30m
population at the municipality level
Landscape values
Landscape values quantifyed using Instagram data van Zanten et al. (2016) 2004-2015 1km
Instagram
Landscape values Flickr Landscape values quantifyed using Flickr data van Zanten et al. (2016) 2004-2015 1km
Landscape values
Landscape values quantifyed using Panoramio data van Zanten et al. (2016) 2004-2015 1km
Panoramio
Land use
Distance from green
Euclidean distance from green urban patches in m Urban Atlas 2006 30m
urban areas
Soil Sealing rate Average soil sealing per municipality EEA 2006-2012 30m
Average tree cover canopy percentage per
Tree cover USGS* 2010 30m
municipality
Cumulative total number of new houses built per
Built-up rate ELSTAT 1997-2016 30m
municipality
HeatMap of Enterprizes HeatMap of new enterprises registered to ACCI ACCI* 1991-2016 30m
Cumulative total number of new enterprises registered
Enterprises count ACCI 1991-2016 30m
to ACCI per municipality
Ministry of Environment
Distance from natural Euclidean distance from forested patches, areas of
& Energy & OSM & (-) 30m
reserves high nature value and protected areas in m
Natura 2000
2
a
3 Global Land Survey Digital Elevation Model (GLSDEM) http://glcf.umd.edu/data/glsdem/

10
 b
1 European Environmental Agency. https://www.eea.europa.eu/data-and-maps/data/indices-of-
2 climate-soil-and-vegetation-quality-1#tab-metadata
c
3 European Environmental Agency. Urban Atlas. GMES/Copernicus land monitoring services.
4 https://www.eea.europa.eu/data-and-maps/data/urban-atlas
d
5 Ministry of Environment & Energy. http://geodata.gov.gr/dataset/poioteta-udaton-akton-
6 kolumbeses-2013
e
7 Open Street Map. https://www.openstreetmap.org
f
8 Society of Information. http://geodata.gov.gr/dataset/demosia-kteria
g
9 Hellenic statistical authority. http://www.statistics.gr/
h
10 van Zanten et al. (2016). PNAS. http://geoplaza.vu.nl/data/dataset/continental-scale-quantification-
11 of-landscape-values-using-social-media-data
i
12 USGS. Global Tree Canopy Cover.
13 https://landcover.usgs.gov/glc/TreeCoverDescriptionAndDownloads.php
j
14 Athens chamber of commerce and industry
15 http://www.acci.gr/acci/catalogue/search.jsp?context=201
16

17

18

19

20

21

22

23

24

25

26

27

28

29

30

31

32

33

34

35

36

37

11
 1 Table 2
2

Medium High Low

From To
development development development

Discontinuous dense urban fabric Continuous urban fabric 0,319 0,392 0,051
Discontinuous medium density urban
Continuous urban fabric
fabric 0,029 0,040 0,005
Discontinuous medium density urban
Discontinuous dense urban fabric
fabric 0,356 0,384 0,070
Discontinuous low density urban fabric Continuous urban fabric 0,001 0,004 0,001
Discontinuous low density urban fabric Discontinuous dense urban fabric 0,044 0,049 0,008
Discontinuous medium density urban
Discontinuous low density urban fabric
fabric 0,383 0,436 0,022
Arable land and permanent crops Continuous urban fabric 0,001 0,002 0,000
Arable land and permanent crops Discontinuous dense urban fabric 0,010 0,019 0,001
Discontinuous medium density urban
Arable land and permanent crops
fabric 0,026 0,043 0,005
Discontinuous low density urban
Arable land and permanent crops
fabric 0,049 0,174 0,055
Industrial commercial and transport
Arable land and permanent crops
units 0,018 0,045 0,014
Forests Scrubs and other natural
Arable land and permanent crops
areas 0,090 0,099 0,083
Forests Scrubs and other natural areas Continuous urban fabric 0,000 0,000 0,000
Forests Scrubs and other natural areas Discontinuous dense urban fabric 0,001 0,002 0,000
Discontinuous medium density urban
Forests Scrubs and other natural areas
fabric 0,002 0,004 0,001
Discontinuous low density urban
Forests Scrubs and other natural areas
fabric 0,007 0,029 0,002
Industrial commercial and transport
Forests Scrubs and other natural areas
units 0,001 0,002 0,001
Forests Scrubs and other natural areas Arable land and permanent crops 0,060 0,064 0,056
3

10

11

12

12
 1

2 Table 3
3

Simulated
2016
Observed
2016 1 2 3 4 5 6 7 Totals P.A
1 1731 148 18 25 11 1933 89,55
2 112 1371 78 33 10 1604 85,47
3 36 77 1293 61 19 1 1487 86,95
4 3 23 99 1420 7 29 3 1584 89,65
5 17 21 11 7 529 12 597 88,61
6 3 9 121 14 957 32 1136 84,24
7 1 2 14 1 36 1004 1058 94,90
Totals 1899 1644 1510 1681 591 1034 1040 9399
U.A 91,2 83,4 85,6 84,5 89,5 92,6 96,5
O.A 88,36
4

13