Artificial neural network for multifunctional areas

Francesco Riccioli & Toufic El Asmar &

Jean-Pierre El Asmar & Claudio Fagarazzi &Leonardo Casini

Received: 12 May 2014 /Accepted: 16 December 2015# Springer International Publishing Switzerland 2015

Abstract The issues related to the appropriate planningof the territory are particularly pronounced in highlyinhabited areas (urban areas), where in addition toprotecting the environment, it is important to consideran anthropogenic (urban) development placed in thecontext of sustainable growth. This work aims at math-ematically simulating the changes in the land use, byimplementing an artificial neural network (ANN) mod-el. More specifically, it will analyze how the increase ofurban areas will develop and whether this developmentwould impact on areas with particular socioeconomicand environmental value, defined as multifunctionalareas. The simulation is applied to the Chianti Area,located in the province of Florence, in Italy. Chianti is

an area with a unique landscape, and its territorial plan-ning requires a careful examination of the territory inwhich it is inserted.

Keywords Artificial neural network . GIS .

Land use change . Territorial planning

Introduction

The change brought about by new spatial technologiesis proposing the massive use of models able to modifythe way to tackle today’s complex dynamics of land useplanning focused on the interaction between humanactivities and the environment. The last century wasmarked by intense anthropogenic (urban) developmentresulting in the loss of natural resources: The issue ofmaking correct choices of spatial planning aimed atpreserving the environment on the one hand and at theachievement of anthropogenic development on the otheris inserted in this context. Different studies (i.e., Prieler2005;European Environment Agency 2006; Bernettiand Marinelli 2009) regarding this concern highlighthow the main evolutionary dynamics are oriented to-ward the reduction of the rural landscape in favor of twophenomena such as the abandonment and expansion(not always regulated) of urban areas (urban sprawl).The many complex variables involved in the land usechanges require the development of decision supporttools and forecasting models, which can simplify theplanning choices and involve more disciplines. Thegeomatics engineering appears to be among the most

Environ Monit Assess (2016) 188:67 DOI 10.1007/s10661-015-5072-7

F. Riccioli (*) : C. Fagarazzi : L. CasiniDepartment of Management of Agricultural, Food and ForestrySystems, University of Florence, Florence, Italye-mail: francesco.riccioli@unifi.it

C. Fagarazzie-mail: claudio.fagarazzi@unifi.it

L. Casinie-mail: leonardo.casini@unifi.it

T. El AsmarFood and Agriculture Organization of the United Nations(FAO) – Plant Production and Protection Division, Rome, Italye-mail: Toufic.Elasmar@fao.org

J.<P. El AsmarNotre Dame University – Louaize – Faculty of Architecture Artand Design, Zouk Mosbeh, Lebanone-mail: jasmar@ndu.edu.lb

appropriate, by focusing on the search for instrumentsthat give the greatest possible knowledge about thechanges of the territory, including the many aspects suchas employment, consumption, and conversion of nonur-ban land and the expansion of urban land. The recentterritorial environmental, energy, and landscape poli-cies, which often are bound by the Climate Changeglobal commitments, appear to be closely related tothe knowledge of land use. In recent years, many GISapplications (Malczewski 2004) have specialized in thisdirection: Among these applications are artificial intel-ligence models (AI) used to describe complex forecastscenarios through simulation of human reasoningreproduced by means of genetic algorithms, artificialneural networks, cellular automata, and fuzzy logictechniques. Used in various disciplines, ranging fromeconomic to medical or engineering (Pijanowski et al.2002), the present work is based on the application of amodel of artificial intelligence to predict land use chang-es: the artificial neural network methodology (ANN).This model is used to forecast the Bdelicate^ evolutionof the urban mosaic in particular contexts represented bymultifunctional areas (MF) or important areas fromsocial-economic and environmental point of view. Thearea of Chianti in the province of Florence, in Italy, isconsidered for this study; in this context, and withinprevious studies, the author (Riccioli 2007, 2009)highlighted multifunctional zones. The variables in-volved in the land use changes are first defined in orderto subsequently build a model of artificial neural net-work to predict the increase in urban areas and whetherthis increase may affect the multifunctional areas.

The paper is organized as follows: In Sect. 2, modelhas been illustrated; in Sect. 3, case study is introduced;in Sect. 4, the models have been applied to multifunc-tional areas; and finally, Sect. 5 is dedicated to conclu-sions and future recommendations.

The artificial neural network model

Costanza and Ruth (1998) consider that in buildingmental models, humans typically simplify systems inparticular ways. We base most of our mental modelingon qualitative rather than quantitative relationships, andwe linearize the relationships among system compo-nents, disregard temporal and spatial lag treat systemsas isolated from their surroundings, or limit our investi-gations to the system’s equilibrium domain. When

problems become more complex, and when quantitativerelationships, nonlinearities, and time and space lags areimportant, we encounter limits to our ability to properlyanticipate system change. In such cases, our mentalmodels need to be supplemented. We must thereforeresort to numerical methods, predictors that exploit theconsiderable potential of computers.

In literature, many works categorize and compare themodels used to analyze land use changes. Some re-searchers gather the models according to their finalpurpose or to the scale of the work (Baker 1989).Lambin (1997) proposes a classification of monitoringmethods of the Land Use Cover Change (LUCC) intropical areas: He analyzes the usefulness of the descrip-tive, empirical, statistical, and dynamic models relatedto the study of the phenomena of deforestation and soildegradation. Agarwal et al. (2002) select 19 models ofLUCC and analyze them according to their ability torepresent the spatial and temporal complexity of asystem.

The analysis in this study is based on artificial neuralnetworks (ANN) for the model’s remarkable ability toadapt to the observed data, especially in the presence ofdatabase characterized by incomplete information, witherrors.

ANN can be defined as nonlinear statistical datamodeling tools having a main purpose to reproducetypical activities of the human brain.

Lopez et al. (2001), Pijanowski et al. (2002), Engelen(2002), and Martinuzzi et al. (2007) used ANN inpredicting models for territorial planning: These modelshave in fact the ability to estimate any type of function,without taking account of its degree of nonlinearity andwithout a priori knowledge of its functional form. Onthe other hand, ANNs have a high degree of uncertaintyin choosing the most favorable network structure.Furthermore, the major limitation in implementingANN, as pointed out by Malczewski (2004), is theirblack-box style used to analyze spatial problems. Themeaning of Black box style is related to the difficulty ofexplaining the internal elaborations (computations) ofthe AI models: BThe ‘black box’ nature of the neuralnetwork methods is a limitation as far as real-worldapplications are concerned.^ It is unlikely that a solutionor a set of solutions obtained by AI-GIS techniques willbe acceptable to those who make decisions regardingland use and the public, if it is difficult or even impos-sible to clearly present and explain to them the internalworkings of the AI models. BOne needs a better answer

67 Page 2 of 11 Environ Monit Assess (2016) 188:67

then ‘because my AI model says so’ when faced withquestions regarding a recommended land-use plan^(O’Sullivan and Unwin 2003).

The ANN model implemented in the study is basedon the use of Multilayer Perceptron (MLP); it is basedon the neuron through which the structure of the humanmind tends to be simulated. Xia and Gar-On Yeh (2002)propose a simple structure of neural network (Fig. 1)consisting of three layers: an input layer (which in ourcase is represented by the variables involved in the landuse changes), a hidden layer, and an output layer (rep-resented by land use changes).

The first layer (input) is represented by ith neurons,each of which is associated to a variable x involved inthe land use changes. In turn, to each variable isassigned a weight w generating the signal that will besent to the neuron of the next layer (Eq. 1).

net j ¼X

j

xi⋅wi; j ð1Þ

where

netj signal sent from the ith neuron of the input layerto the jth neuron of the hidden layer

xi variable involved in the land use changes of theith neuron of the input layer

wi,j relative weight of the input layer and hidden layer

Subsequently, the signal (value) shown in Eq. 1 (netj)is sent to the jth neuron belonging to the hidden layer.This layer is activated if and only if it reaches a certainpredetermined threshold value (φ). Most common acti-vation functions can be linear or sigmoidal (Eq. 2, re-lating to a sigmoidal activation function).

φ j ¼1

1þ e−net jð2Þ

From the hidden layer, if activated, the signal istransferred to the next layer represented by the output:The output is format from the ith neuron, which values(pl) represent the probability of conversion from a givenland use to another (Eq. 3).

pl ¼X

j

w j;l⋅φ j ð3Þ

where

pl probability of conversion of the lth from theoutput layer

wj,l relative weight of the hidden layer and outputlayer

φj activation function of the jth neuron of the hiddenlayer

The algorithm used for the generation of the output isthe Bback-propagation^ which is a supervised learning,through which the output estimated by the network (pl,Eq. 3) is compared with a desired or known outputcalled out: Out represents the actual land use changesthat have occurred in the period examined.

The purpose of this comparison is to obtain an outputestimated as similar as possible to the desired output.The difference between the two outputs produces anerror (e) used to correct the weights (weights wereinitialized with random values at the beginning of thetraining). In our case, the error is quantified by thestandard deviation (Eq. 4). This training set is repeateduntil the error is less than a predetermined threshold.

el ¼ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiX

l

outl−plð Þ2s

ð4Þ

where

el relative error at the lth neuron of the output layeroutl known output of the lth neuron of the output layerpl estimated output of lth neuron of the output layer

In the Bback-propagation^ mechanism, the error ispropagated Bbackwards^ in the previous layers of themodel associating it with the weights. This mechanismfollows the Delta rule1 that is a learning rule based onthe decrease in the gradient δ (Eq. 5) to update theweights (Eq. 6).

δlt ¼elφl

φl

X

j

δ jtþ1 wjltþ1

8<

: ð5Þ

where

δlt gradient error of the lth neuron of the outputlayer at time t

el relative error at the lth neuron of the output layerφl activation function of the lth neuron of the

output layerδjt+1

1 The delta rule is based on the Bgradient descent,^ a non-linearoptimization algorithm, used to identify the local minimum of afunction.

Environ Monit Assess (2016) 188:67 Page 3 of 11 67

gradient error of the jth neuron of the hiddenlayer at time t +1

wjlt+1 relative weight to the hidden layer and to theoutput layer at time t +1

Δwji tþ1ð Þ ¼ ηδ ji xi þ αΔwji tð Þ ð6Þwhere

Δwji(t + 1) difference of weights between the hiddenlayer and the input layer after a number ofiterations t+1

η learning speed of the neuron or rate ofdescent toward the minimum of the error2

curveδ gradient of errorα momentum factor, constant of

proportionality which analyzes theprobability of oscillation of the weights3

Δwji(t) difference of weights between the hiddenlayer and the input layer after a number ofiterations t

The goal of the MLP is to minimize this gradientBadjusting^ the random weights, and bringing some

gradual and progressive changes to them. In otherwords, the value of the weights of the model variesthrough a number of iterations inducing thereby thevalue of the output to vary n times. When the gradientis sufficiently reduced, the training phase would haveproduced an estimated output very close to the desiredoutput. At the end of the training phase, the model willthen be able to recognize the unknown relationshipbetween the input variables and the output variables.In addition, this enables to create predictions in timewhere the output data are not known a priori. The finalaim of the supervised learning is a prediction of thevalue of output for each valid value of the input basedonly on a limited number of examples of correspon-dence (input-output pairs of values). To achieve this, thesystem uses two principles: mathematical distribution(that links the delta of input values to the output values)and likelihood function—once mathematical distribu-tion has been identified, the system chooses the param-eters that maximize the likelihood of the data and selectsthe correct likelihood function.

Case study: the multifunctional areas of Chianti

Chianti area includes five municipalities located inTuscany (center of Italy) in the province of Florence:Barberino val d’Elsa, Greve in Chianti, Impruneta, SanCasciano in val di Pesa, and Tavarnelle val di Pesa.

2 If the speed is too low, the training phasemay be too expensive interms of time and resources, if too high could lead to inaccurateresults.3 The momentum in practice analyzes the weights to determinedirection in which to search for the minimum error.

Fig. 1 ANN diagram

67 Page 4 of 11 Environ Monit Assess (2016) 188:67

The study area is located close to the south of thetown of Florence (Fig. 2) and has a total area of approx-imately 600 km2 and a total resident population of about77,000 inhabitants.

Chianti Area is characterized by a predominantlyhilly topography; the average annual temperature variesbetween 11.6 and 15 °C, while the rainfall conditions areestimated around 800 mm per year. The major land usesof the Chianti region are the forest and the vineyardsfrom which is produced the famous Chianti wine DOC(controlled origin) and DOCG (controlled and guaran-teed origin).

Using a spatial multicriteria analysis model (Riccioli2007, 2009), the author analyzes five functions, per-formed by agricultural activities in the area, which arethe socioeconomic, the aesthetic, the hydrological, theterritorial preservation, and the natural function. Thesefunctions have been quantified throughmultidimension-al indexes and aggregated through multicriteriaoperators.

Socioeconomic function has based on RuralDevelopment Plan guidelines and ISTAT census data-base (ISTAT 2001); it has been analyzed by some spe-cific indexes related to the farm and farmer characteris-tics (compared to total farm surface) such as number offarmers with a professional degree, number of farmswith high-quality wine production, and number of farmswith farmer under 60 years old: These indexes havebeen aggregated using ordered weighted average(OWA) operator (Malczewski 1999). Aesthetic functionhas been based on landscape values of land use. Anaesthetic value has been given to land use from pano-ramic viewpoint such as wine road and farmhouses(Riccioli 2004). Hydrological function has been ana-lyzed through Soil Conservation Service - CurveNumber (SCS, 1969) method; it has used to determinesurface flows in specific soils. Territorial preservationhas been analyzed by density of forest and rural road anddensity of cultural human rural construction (for exam-ple stone wall). Finally, natural function indexes havebeen based on Biopermeability Index (hectares of con-tinuous forest) and Shannon Index (degree of land usediversity): These indexes have been aggregated usingWeighted Linear Combination (WLC) operator(Malczewski 1999). The portions of territory showingthe simultaneous presence of the five features weretherefore considered multifunctional areas: In thisphase, an overlay (with AND operator) of previousfunctions has been used (see Riccioli 2007, 2009 for

more details). The purpose of the next phases is toevaluate, which of these areas will be affected by theprocesses of urbanization.

The ANN model applied to multifunctional areaspotentially involved in the process of urbanization

Creation of transition of land use rules

The preliminary phase of the investigation focused onthe analysis of land use changes that have emerged inthe decade between 1990 and 2000. This wasundertaken by using vector thematic maps of the casestudy land use. This analysis was influenced, as notedby Matheron (1978, 1989) by the set objectives, as wellas by the available data. Accordingly, the used sourcewas based on the Corine Land Cover (CLC) - EuropeanRegulation on Information Region (ENV 12657). TheCLC being the only source available allowing analyzingthe area through a multi-temporal reading.

I t is possible to observe how during theabovementioned period, seven typologies of land usewere changed: More particularly, the increase in woodedareas, in woody and agricultural crops, and urban areas isregistered on the one hand, and a decrease of heteroge-neous agricultural areas4 and arable land on the other.

The next phase is oriented toward the definition ofthe variable involved in the land use changes. Theliterature review (Pijanowski et al. 2002; Lombardo etal. 2005) shows that these variables are essentially re-lated to the morphology of the territory and the anthro-pogenic activity. Based on the data available, environ-mental variables such as slope and topography andanthropogenic variables such as human settlements(cities, towns, and small villages) and roads were select-ed. So, four layers were implemented through aGeographic Information System (GIS) that diversifythe land use changes according to:

1. Distance from roads2. Distance from inhabited centers3. Slope4. Altitude

4 Heterogeneous agricultural areas are considered temporary cropsassociated with permanent crops, cropping systems, and particlecomplex. Areas are predominantly occupied by agricultural fieldswith significant natural areas and areas of agricultural woods.

Environ Monit Assess (2016) 188:67 Page 5 of 11 67

A statistical analysis of spatial independence of theabove four layers was carried out. This was donethrough the application of tests based on the comparisonof pairs of layers (maps) in order to verify the reliabilityof the selected variables. As suggested by Bonham-Carter (1994), the Cramer’s V index was used for theanalysis.

A high Cramer’s V indicates that the potential ex-planatory value of the variable is good but does notguarantee a strong performance since it cannot accountfor the mathematical requirements of the modeling ap-proach used and the complexity of the relationship. Hisvalue varies between 0 (max independence) and 1 (maxdependence). The correlation of each variable with landuse changes was accordingly analyzed. The analysis

revealed a good relationship of dependency betweenthe data analyzed with values greater than 0.15, assuggested by Eastman (2006).

The four variables involved in the land use changeswere then entered as input data in the model of artificialneural network. This was done in order to Btrain^ theMLP by combining them with random weights. In thetraining phase, two constraints related to the maximumtolerable error between the estimated output and thedesired output (less than or equal to 0.0001) and thenumber of iterations (set at 5000) were fixed. The fol-lowing are the technical parameters used in the model.

& Speed of training (η) = 0.005& Momentum factor (α) = 0.5

Fig. 2 Case study map

67 Page 6 of 11 Environ Monit Assess (2016) 188:67

& Tolerance error value (e) = 0.0001& Number of iterations (t) =5000

Observing these constraints, the MLP has pro-duced estimated output evaluated from known rela-tionships between the input variables and the output(the desired output) , and generating Brules(probability) of transition.^ These probabilities wereused successively to predict future scenarios of theland use changes.

Figure 3 shows the results of the application ofthese rules. It highlights which areas of arable land,and heterogeneous agricultural areas, will have inthe future high probability of being incorporatedinto the urban areas. The highest values (the areaswith the most intense chromatic scale) belong to theareas having the highest probability of becomingurban. In other words, the four variables involvedin land use changes are

1. The areas next to the roads2. The areas next to inhabited centers

3. The areas with minor slopes4. The areas located at minor altitude

Validation of the transition probability of land use

In the next phase, the accuracy of the data was verifiedthrough validation. The validation process involves thecomparison of a land usemap of a specific year developedby using the general transition rules of MLP (forecastingmap) and a land use map used as reference (referencemap). Based on the available cartography, year 2006 mapwas established as a reference (2006 is the most recentyear for which the Corine Land Cover is available): Thiswas considered the reference map. The forecastingmap at2006 was successively created. The literature highlightstheMarkovian approaches among the successful methodsused to develop hypothetic scenarios of land use changes(Aaviksoo 1995; Logofet and Lesnaya 2000).

By using the map of the potential transitions (Fig. 3),a land use map of year 2006 was therefore created. Thiswas done by applying Markov chain (Eastman and

Fig. 3 Potential areas of urbanization

Environ Monit Assess (2016) 188:67 Page 7 of 11 67

Toledano 2000) which is a stochastic process where thetransition probabilities (Eq. 3) have been used in amatrix (Pt) to obtain a projection to 2006 (Wt + 1) of thechanges occurring in the time interval from 1990 to2000 (Wt) as shown in Eq. 7.

Wtþ1 ¼ Wt⋅Pt ð7Þwhere

Wt + 1 land use at t+1Wt land use at tPt transition probability matrix or stochastic

matrix n×n of values pij

where n is the number of discrete states in the Markovchain and pij are the transition probabilities (between 0and 1) from the state j to the state i in the time intervalbetween t and t+1. As described by Coquillard and Hill(1997), the matrix obtained describes a system thatchanges through discrete increments of time, in whichthe value of each variable, at a given time, is the sum ofpercentages of the values of the variables in the previousinstant. The sum of the fractions along a row of thematrix is equal to one, and the diagonal contains thepercentages instead of pixels that do not change betweenthe start and end date.

Table 1 shows the transition matrix relative only tothe considered land use changes (arable land and het-erogeneous agricultural areas involved in urbanizationprocesses).

The prediction map at 2006 was compared with thereference map (CLC 2006) by the Cohen coefficient ofcorrelation (Cohen’s Kappa). Due to the use of rastermaps, the index has been calculated by comparing thespatial distribution and quantity of pixels for each cate-gory of land use. The statistical analysis has shown agood degree of agreement with a value of 0.7893 (assuggested by Landis and Koch 1977), statistically vali-dating the transition probability of land use obtained.

Multifunctional areas potentially involved in the processof urbanization

In order to highlight the probability that the multifunc-tional areas have to be involved in a process of urban-ization, an overlay of maps using the logical operator ofintersection AND was used. This was done by overlap-ping areas with probability of conversion to urban(Fig. 3), with the MF areas (shown in different shadesof orange in Fig. 4). The result is shown in Fig. 4 inwhich multifunctional areas potentially interested inurbanization (MFu) are highlighted in black.

Table 2 shows the hectares of areas, subdivided bymunicipalities, potentially interested by the urbanization.

The MFu areas stretch along approximately 110 haequivalent to 7.1 % of the total MF areas within thestudy area (1550 ha). The threatened areas are locatedexclusively in the municipalities of Greve in Chianti andSan Casciano representing respectively 7.4 and 8.5 % ofthe total MF areas of the municipality.

Conclusions and future recommendations

This paper is based on the study of the territory of asensitive area from the urban point of view in which theterritorial aspect must be preserved (the Chianti, despitebeing a rich areas with high environmental value, is veryclose to major industrial urban agglomerations). Basedon a previous work, the case study has been analyzedlooking at what the rural activities of man can offer fromthe economic, social, and environmental point of view,and highlighting multifunctional areas. By applying amodel of spatial multicriteria analysis, each of the threeaspects (economic, social, and environmental) has beenevaluated through multidimensional indexes. The in-dexes were appropriately aggregated with multicriteriarules and have allowed us to highlight the

Table 1 Transition matrix at year 2006

Land use at t+ 1 (2006)

Urban Arableland

Hetero. Agr.Area

Landuse at t(2000)

Urban 1.0000 0.0000 0.0000

Arable land 0.0853 0.9147 0.0000

Hetero. Agr. Area 0.1352 0.0027 0.8621

state i (t+ 1)

state j (t) p11 p12 p1np21 p22 p2npn1 pn2 pnn

67 Page 8 of 11 Environ Monit Assess (2016) 188:67

multifunctionality of the territory. The decision to focuson multifunctional areas was dictated by the increasingimportance that these areas play in the recent CommonAgricultural Policy (especially within the RuralDevelopment pillar).

The focus has therefore shifted on the analysis of landuse changes in order to highlight what has altered overtime, how anthropogenic (urban) development mayevolve, and how it may affect the multifunctional areas.

The main purpose is to pursue a balanced socioeconom-ic and environmental urban development without arrest-ing the latter but by regulating its growth. Multitemporalanalysis of land uses was then carried out byimplementing an ANN model using the MultilayerPerceptron (a model of nonlinear analysis), in order tocreate a map of areas potentially affected by urbaniza-tion. The data were validated through a procedure thattook advantage ofMarkov chains to create a map of landuse forecasts to 2006 appropriately compared with areference map (Corine Land Cover 2006) by the statis-tical index of Cohen. Subsequently, the map of areaspotentially affected by urbanization has been used toidentify which of the multifunctional areas would beinvolved in this process.

In order to use and read the results of the proposedmodel, it is important to start from the assumption thatthe size of the multifunctional areas and the urbandevelopment are constant over time. This persistenceof the conditions is a limitation of the model as stressesTang et al. (2005) which can be overcome through theuse of new data such as the evolution of the road

Fig. 4 MF areas potentially involved in an urbanization process

Table 2 Statistics related to the MFu areas (data expressed inhectares and percentage over the total municipal MF areas)

Municipality MF MFu MFu on MFHa Ha %

Barberino 65.03 0 –

Greve 488.07 41.47 8.5

Impruneta 25.63 0 –

San Casciano 939.30 69.31 7.4

Tavarnelle 33.31 0 –

Totale 1551.34 110.78 7.1

Environ Monit Assess (2016) 188:67 Page 9 of 11 67

network, or a more detailed land use changes especiallyfrom the temporal point of view (Verburg et al. 2002).

This work can be classified then as an application ofan effective method of artificial intelligence based onartificial neural networks, applied in the environmentalfield to create predictive models of land use changes.Some current research studies have been conductedusing this methodology. Alsharif and Pradhan (2014)and Mahbood et al. (2015) analyze respectively devel-opment of urban areas in Tripoli and Pakistan usingremote sensing and landsat imageries, Mazzocchi et al.(2014) explore (through ANN) the evolution ofagricultural and natural areas near Milano fromenvironmental, cultural, and recreational point of view,and Grekousis et al. (2013) and Triantakonstantis andStathakis (2015) use ANNmethodology for the analysisof urban sprawl in Athens concluding that urban devel-opment depends on the available funds, accessibilityimprovement (railway and metro networks), landspeculation, and lack of land use control. Basse et al.(2014) combine ANN and cellular automata with theaim of identification of driving forces that are behindland use and land cover changes. Park et al. (2011)analyze various methods (also ANN method has beenexamined) to determine which best explained urbangrowth until the present for modeling future urbangrowth in Korea. All of these works aim to more accu-rate forecast of development of urban areas, but they donot analyze their relationship with the characteristics ofsurrounding territory that may be a crucial issue interritorial planning processes.

Starting with this consideration, by relating this ap-plication to the definition of multifunctional areas, weintend to provide the decision-maker with a powerfulplanning tool that can Bguide^ the urban developmentby controlling anthropogenic development, and the oth-er parts of the country deemed interesting from theeconomic, social, and environmental point of view.

The proposed methodology is a good compromisebetween adaptability of the model to input variablesselected or able to be selected, and the ability to under-stand the results. The results are able to be integratedand modified to further refine the research. For example,in order to expand the temporal range and the degree ofdetail of analysis, it may be useful to derive land usesfrom satellite photos. The extraction of rules for decisionmaking may include a greater number of variables in-volved in the land use changes, through perhaps, the useof discrete models (Choice Experiment) able to describe

human behaviors useful in understanding the evolution-ary dynamics. Furthermore, to forecast and developfuture scenarios, the analyzed data could be used toimplement a Cellular Automata (Basse et al. 2014) ableto consider the evolutionary dynamics by consideringthe so-called neighborhoods or areas adjacent to the areabeing analyzed.

In conclusion, authors emphasize how the spatial-temporal simulation, integrated with socioeconomic in-formation, is the new frontier of territorial analysis. Asemphasized by Steyaert (1993), the evolutionary dy-namics in the real world typically take place in threedimensions, time-dependent, and are extremely com-plex. This complexity often includes nonlinear behaviorand stochastic components. The study of such behaviorgoes through the formulation of hypotheses and rules toexplain its functioning. The rules can, in turn, beexpressed by mathematical formulas or logical relation-ships, which often lead to a series of theoretical simpli-fications to reduce the number of equations used. Themathematical models are based on programming lan-guages that realistically simulate the evolution of spatialpatterns over time that are increasingly used for quanti-tative analysis, and no more only for qualitative analy-sis, of the complex issues at the local, regional, or globallevel. The goal is, ultimately, the realization of decisionsupport tools that are characterized by promptness, cost-efficiency, and ease of use, aiming at achieving a betterunderstanding and management of the territory.

References

Aaviksoo, K. (1995). Simulating vegetation dynamics and land-use in a mire landscape using a Markov model. Landscapeand Urban Planning, 31, 129–142.

Agarwal, C., Green, G. M., Grove, J. P., Evans, T. P., & Schweik,C. M. (2002). A review and assessment of land use changemodels (CIPES collaborative report series No.1).

Alsharif, A. A., & Pradhan, B. (2014). Urban sprawl analysis ofTripoli Metropolitan City (Libya) using remote sensing dataand multivariate logistic regression model. Journal of theIndian Society of Remote Sensing, 42(1), 149–163. Springer.

Baker, W. L. (1989). A review of models of landscape change.Landscape Ecology, 2(2), 111–133.

Basse, R., Omrani, H., Charif, O., Gerber, P., & Bòdis, K.(2014). Land use changes modelling using advancedmethods: cellular automata and artificial neural net-works. The spatial and explicit representation of landcover dynamics at the cross-border region scale.Applied Geography, 53, 160–171. Elsevier.

67 Page 10 of 11 Environ Monit Assess (2016) 188:67

Bernetti, I., & Marinelli, N. (2009). Cambiamenti in atto nelpaesaggio agrario toscano. Possibili scenari evolutivi. IGeorgofili, 4, 633–677.

Bonham-Carter, G.F. (1994). Geographic information systems forgeoscientists: modelling with GIS. Ontario, Canada:Pergamon.

Coquillard, P., & Hill, D. R. C. (1997).Modélisation et simulationd’écosystèmes. Paris: MASSON.

Costanza, R., & Ruth, M. (1998). Using dynamic modeling toscope environmental problems and built consensus.Environmental Management, 22(2), 183–195.

Eastman, J. R. (2006). IDRISI Andes, tutorial. MA, USA: ClarkLabs, Clark University.

Eastman, J. R., & Toledano, J. (2000). Markov chain and cellularautomata approaches to land cover change modeling. InWorkshop presentation (invited), 4th InternationalConference on Integrating GIS and EnvironmentalModeling, Banff, Alberta, Sept. 2–8, 2000.

EEA. (2006). Land accounts for Europe 1990–2000. EEA reportNo 11/2006.

Engelen, G., White, R., Van Der Meulen, M., & Hahn, B. (2002).Sustainable developments of islands: a policy support frame-work for the integrated assessment of socio-economic andenvironmental development. In H.-H.M.Hsiao, C.-H. Liu,&H.-M. Tsai (Eds.), Sustainable development for island soci-eties: Taiwan and the World (pp. 251–287). Taipei, Taiwan:Asia-Pacific Research Program, Academia Sinica andSARCS Secretariat.

Grekousis, G., Manetos, P., & Photis, Y. N. (2013). Modelingurban evolution using neural networks, fuzzy logic andGIS: the case of the Athens metropolitan area. Cities, 30,193–203. Elsevier.

ISTAT. (2001). V Censimento Generale dell’Agricoltura. Roma.http://censagr.istat.it/dati.htm. Accessed 24 Sept 2002.

Lambin, E. (1997). Modelling and monitoring land-cover changeprocesses in tropical regions. Progress in PhysicalGeography, 21(3), 375–393.

Landis, J. R., & Koch, G. G. (1977). The measurement of observeragreement for categorical data. Biometrics, 33(1), 159–174.

Logofet, D. O., & Lesnaya, E. V. (2000). The mathematics ofMarkov models: what Markov chains can really predict inforest successions. Ecological Modeling, 126, 285–298.

Lombardo, S., Pecori, S., & Petri, M. (2005). Investigating terri-torial dynamic using decision tree (Atti del convegnoCupum).

Lopez, E., Bocco, G., Mendoza, M., & Duhau, E. (2001).Predicting land cover and land use change in the urban fringe.A case in Morelia city, Mexico. Landscape and UrbanPlanning, 55, 271–282. Elsevier.

Mahboob, M. A., Atif, I., & Iqbal, J. (2015). Remote sensing andGIS applications for assessment of urban sprawl in Karachi,Pakistan. Science, Technology and Development, 34(3), 179–188. doi:10.3923/std.2015.179.188.

Malczewski, J. (1999). Gis and multicriteria decision analysis.Toronto, Ontario, Canada: John Winley & Sons, Inc.

Malczewski, J. (2004). Gis–based land-use suitability analysis: acritical overview. Volume 62, Issue 1, pp 3–65, Toronto:Elsevier.

Martinuzzi, S., Gould, W., & Gonzales, O. (2007). Land develop-ment, land use, and urban sprawl in Puerto Rico integrating

remote sensing and population census data. Landscape andUrban Planning, 79, 288–297. Elsevier.

Matheron, G. (1978). Estimer et choisir (Cahiers du centre demorphologie mathématique de Fontainebleau, Fasc.7, Ecolede Mines de Paris).

Matheron, G. (1989). Estimating and choosing – an essay onprobability in practice. Berlin: Springer.

Mazzocchi, C., Corsi, S., Sali, G. (2014). Agricultural land con-sumption in periurban areas: a methodological approach forrisk assessment using artificial neural networks and spatialcorrelation in Northern Italy. Applied Spatial Analysis. doi:10.1007/s12061-015-9168-9. Springer.

O’Sullivan, D., & Unwin, D. J. (2003). Geographic informationanalysis. Hoboken: Wiley.

Park, S., Jeon, S., Kim, S., & Choi, C. (2011). Prediction andcomparison of urban growth by land suitability index map-ping using GIS and RS in South Korea. Landscape andUrban Planning, 99, 104–114. Elsevier.

Pijanowski, B.B., Brown, D.G., Shellito, B.A., Manik, G.A.(2002). Using neural networks and GIS to forecast land usechanges: a Land Transformation Model. Computers,Environment and Urban Systems, 26(6), 553–575. Elsevier.

Prieler, S. (2005). Built-up and associated land area increases inEurope (WP 3. Environmental Evaluation, InternationalInstitute for Applied Systems Analysis (IIASA), Land UseChange and Agriculture Project).

Riccioli, F. (2004). Il paesaggio nelle valutazioni di impattoambientale. Atti del XXXIV Incontro di Studio CeSET BLegrandi infrastrutture: approcci di ordine giuridico, economicoed estimativo^ Firenze 15–16 ottobre 2004. University press:Florence, Italy

Riccioli, F. (2007). La funzione socio-economica delle attivitàvitivinicole letta attraverso il Piano di Sviluppo Ruraletoscano. In F. A. Milano (Ed.), Menghini S. (a cura di) BIlruolo del settore vitivinicolo nei processi di svilupposostenibile^ (pp. 52–92).

Riccioli, F. (2009).Una proposta metodologica per lo studio dellamultifunzionalità: l’analisi MultiCriteriale Geografica(Economia e diritto agroambientale n.3, pp. 99–118).Firenze: University Press.

Steyaert, L. T. (1993). A perspective on the state of environmentalsimulation modeling (Environmental modeling with GIS,Pag. 16–30). New York: Oxford University Press.

Tang, Z., Engel, B. A., Pijanowski, B. C., & Lim, K. J. (2005).Forecasting land use change and its environmental impact ata watershed scale. Elsevier, Journal of EnvironmentalManagement, 76, 35–45.

Triantakonstantis, D., & Stathakis, D. (2015). Urban growth pre-diction in Athens, Greece, using Artificial Neural Networks.International Journal of Civil, Environmental, Structural,Construction and Architectural Engineering, 9(3), 2015.

Verburg, P. H., Soepboer, W., Limpiada, R., Espaldon, M. V. O.,Sharifa, M. A., &Veldkamp, A. (2002).Modelling the spatialdynamics of regional land use: the CLUE-S model.Environmental Management, 30, 391–405.

Xia, L., & Gar-On Yeh, A. (2002). Neural network based cellularautomata for simulatingmultiple land use changes using GIS.International Journal Geographical Information Science,16(4), 323–343. Taylor & Francis.

Environ Monit Assess (2016) 188:67 Page 11 of 11 67