ACTAGEOGRAPHICA GEOGRAFSKI ZBORNIK SLOVENICA 2019 59 2 ACTA GEOGRAPHICA SLOVENICA GEOGRAFSKI ZBORNIK 59-2 • 2019 Contents Drago PERKO, Rok CIGLIČ, Mauro HRVATIN The usefulness of unsupervised classification methods for landscape typification: The case of Slovenia 7 Vladimir M. CVETKOVIĆ, Kevin RONAN, Rajib SHAW, Marina FILIPOVIĆ, Rita MANO,Jasmina GAČIĆ, Vladimir JAKOVLJEVIĆ Household earthquake preparedness in Serbia: A study of selected municipalities 27 Iwona CIEŚLAK Spatial conflicts: Analyzing a burden created by differing land use 43 Ivan PAUNOVIĆ, Verka JOVANOVIĆ Sustainable mountain tourism in word and deed: A comparative analysis in the macroregions of the Alps and the Dinarides 59 Nikola Darko VUKSANOVIĆ, Dragan TEŠANOVIĆ, Bojana KALENJUK, Milijanko PORTIĆ Gender, age and education differences in food consumption within a region: Case studies of Belgradeand Novi Sad (Serbia) 71 Special issue – Franciscean cadaster as a source of studying landscape changes Matej GABROVEC, Ivan BIČÍK, Blaž KOMAC Land registers as a source of studying long-term land-use changes 83 Ivan BIČÍK, Matej GABROVEC, Lucie KUPKOVÁ Long-term land-use changes: A comparison between Czechia and Slovenia 91 Lucie KUPKOVÁ, Ivan BIČÍK, Zdeněk BOUDNÝ Long-term land-use / land-cover changes in Czech border regions 107 Drago KLADNIK, Matjaž GERŠIČ, Primož PIPAN, Manca VOLK BAHUN Land-use changes in Slovenian terraced landscapes 119 Daniela RIBEIRO, Mateja ŠMID HRIBAR Assessment of land-use changes and their impacts on ecosystem services in two Slovenianrural landscapes 143 Mojca FOŠKI, Alma ZAVODNIK LAMOVŠEK Monitoring land-use change using selected indices 161 ISSN 1581-6613 9 771581 661010 ACTA GEOGRAPHICA SLOVENICA 2019 ISSN: 1581-6613 COBISS: 124775936 UDC/UDK: 91© 2019, ZRC SAZU, Geografski inštitut Antona Melika Internationaleditorialboard/mednarodniuredniškiodbor: DavidBole(Slovenia),MichaelBründl(Switzerland),RokCiglič(Slovenia), Matej Gabrovec (Slovenia), Matjaž Geršič (Slovenia), Peter Jordan (Austria), Drago Kladnik (Slovenia), BlažKomac (Slovenia), Andrej Kranjc (Slovenia), Dénes Lóczy (Hungary), Simon McCharty (United Kingdom), SlobodanMarković (Serbia), Janez Nared (Slovenia), Drago Perko (Slovenia), Marjan Ravbar (Slovenia), Nika Razpotnik Visković(Slovenia), Aleš Smrekar (Slovenia), Annett Steinführer (Germany), Mimi Urbanc (Slovenia), Matija Zorn (Slovenia) Editor-in-Chief/glavni urednik: Blaž Komac; blaz@zrc-sazu.si Executive editor/odgovorni urednik: Drago Perko; drago@zrc-sazu.si Chief editor for physical geography/glavni urednik za fizično geografijo: Matija Zorn; matija.zorn@zrc-sazu.siChief editor for human geography/glavna urednica za humano geografijo: Mimi Urbanc; mimi@zrc-sazu.si Chief editor for regional geography/glavni urednik za regionalno geografijo: Drago Kladnik; drago.kladnik@zrc-sazu.si Chief editor for spatial planning/glavni urednik za regionalno planiranje: Janez Nared; janez.nared@zrc-sazu.si Chiefeditorforruralgeography/glavnaurednicazageografijopodeželja:NikaRazpotnikVisković;nika.razpotnik@zrc-sazu.si Chief editor for urban geography/glavni urednik za urbano geografijo: David Bole; david.bole@zrc-sazu.si Chief editor for geographic information systems/glavni urednik za geografske informacijske sisteme: Rok Ciglič; rok.ciglic@zrc-sazu.siChief editor for environmental protection/glavni urednik za varstvo okolja: Aleš Smrekar; ales.smrekar@zrc-sazu.si Editorial assistant/uredniški pomočnik: Matjaž Geršič; matjaz.gersic@zrc-sazu.si Issued by/izdajatelj: Geografski inštitut Antona Melika ZRC SAZUPublished by/založnik: Založba ZRC Co-published by/sozaložnik: Slovenska akademija znanosti in umetnosti Address/Naslov: Geografski inštitut Antona Melika ZRC SAZU, Gosposka ulica 13, SI – 1000 Ljubljana, Slovenija The papers are available on-line/prispevki so dostopni na medmrežju: http://ags.zrc-sazu.si (ISSN: 1581–8314) Ordering/naročanje: Založba ZRC, Novi trg 2, p. p. 306, SI – 1001 Ljubljana, Slovenija; zalozba@zrc-sazu.si Annual subscription/letna naročnina: 20 € for individuals/za posameznike, 28 € for institutions/za ustanove. Single issue/cena posamezne številke: 12,50 € for individuals/za posameznike, 16 € for institutions/za ustanove. Cartography/kartografija: Geografski inštitut Antona Melika ZRC SAZU Translations/prevodi: DEKS, d. o. o. DTP/prelom: SYNCOMP, d. o. o. Printed by/tiskarna: Tiskarna Present, d. o. o. Print run/naklada: 450 copies/izvodov The journal is subsidized by the Slovenian Research Agency and is issued in the framework of the Geography of Slovenia coreresearchprogramme(P6-0101)/revijaizhajaspodporoJavneagencijezaraziskovalnodejavnostRepublikeSlovenijein nastajav okviru raziskovalnega programa Geografija Slovenije (P6-0101). The journal is indexed also in/revija je vključena tudi v: SCIE – Science Citation Index Expanded, Scopus, JCR – Journal Citation Report/Science Edition, ERIH PLUS, GEOBASE Journals, Current geographical publications, EBSCOhost,Geoscience e-Journals, Georef, FRANCIS, SJR (SCImago Journal & Country Rank), OCLC WorldCat, Google scholar,and CrossRef. Oblikovanje/Design by: Matjaž Vipotnik Front cover photography: Exploration of the collapse dolines, such as the one at the Small Natural Bridge in RakovŠkocjan, has enabled a deeper understanding of karst processes in recent years (photograph: Matej Lipar).Fotografija na naslovnici: Raziskave udornice, kot je ta pri Malem Naravnem mostu v Rakovem Škocjanu, so v zadnjihletih omogočile globlje razumevanje kraških procesov (fotografija: Matej Lipar). THEUSEFULNESSOFUNSUPERVISED CLASSIFICATIONMETHODS FORLANDSCAPETYPIFICATION: THECASEOFSLOVENIA Drago Perko, Rok Ciglič, Mauro Hrvatin The landscape classification of countries with a high landscape diversity, such as Slovenia, using a computer and various classification methods is a difficult task. The picture shows three landscape types: fertile flysch Mediterranean hills with the Bay of Koper (foreground), Mediterranean karst plateaus (middle), and high Dinaric karst plateaus (background). DOI: https://doi.org/10.3986/AGS.7377 UDC: 911.5(497.4) COBISS: 1.01 The usefulness of unsupervised classification methods for landscape typification: The case of Slovenia ABSTRACT:Supervisedandunsupervisedclassificationmethodscanbeausefultoolindeterminingvarious geographicalspatialdivisions,especiallyregionalizationsandtypifications.BecauseSloveniaisgeographically verydiverse,itsdivisionsareaparticularlysignificantandinterestingresearchchallenge.Themainobjective of this article is to determine the effectiveness of unsupervised classification methods, and therefore we comparethewell-establishedlandscapetypologyofSloveniafrom1996withlandscapetypologiesthatwere modeled using various unsupervised classification methods. Our results show that landscape typologies modeledusing unsupervised classificationmethodsdeviatemore from the originallandscapetypology of Slovenia than landscapetypologies modeledusing randomand expert-supervised classification methods. KEY WORDS: geography, geographic information system, modeling, classification, landscape typology, Slovenia Uporabnost metod nenadzorovane klasifikacije za pokrajinsko tipizacijo na primeru Slovenije POVZETEK:Metodenadzorovaneinnenadzorovaneklasifikacijesolahkokoristnoorodjepridoločanju različnih geografskih prostorskih delitev, še posebej pri regionalizacijah in tipizacijah. Ker je Slovenija geografsko zelo raznolika, so njene delitve še posebej velik in zanimiv raziskovalni izziv. Glavni namen članka je ugotoviti učinkovitost metod nenadzorovane klasifikacije, zato primerjamo uveljavljeno pokrajinskotipizacijoSlovenijeizleta1996spokrajinskimitipizacijami,kismojihmodeliralizrazličnimi metodami nenadzorovane klasifikacije. Naši rezultati kažejo, da se pokrajinske tipizacije, modelirane z metodami nenadzorovane klasifikacije, izvirni pokrajinski tipizaciji Slovenije ne približajo tako dobro, kot pokrajinske tipizacije, modelirane z metodami naključne in ekspertne nadzorovane klasifikacije. KLJUČNE BESEDE: geografija, geografski informacijski sistem, modeliranje, klasifikacija, pokrajinska tipizacija, Slovenija Drago Perko, Rok Ciglič, Mauro Hrvatin Research Center of the Slovenian Academy of Sciences and Arts, Anton Melik Geographical Institute drago@zrc-sazu.si, rok.ciglic@zrc-sazu.si, mauro@zrc-sazu.si This article was submitted for publication on February 26th, 2018. Uredništvo je prejelo prispevek 26. februarja 2018 1 Introduction Regionalizations and typifications are among the most complicated fields of research in geography (Hammond1964;Dikau,BrabbandMark1991;Kladnik1996;Brabyn1998;McMahon,WikenandGauthier 2004; Gallant,Douglas andHoffer2005;IwahashiandPike2006; Ellison2010;Ciglič2014). Thisisespe­cially the case with the spatial divisions of Slovenia, which despite being a small country has some of the mostdiverselandscapesintheworld(Melik1935;Ilešič1956;1958;Gams1983;Natek1993;Gams,KladnikandOroženAdamič1995;GabrovecandHrvatin1998;Gams1998;Perko1998;Plut1999;Špeset al.2002; Kladnik,PerkoandUrbanc2009;PerkoandHrvatin2009;HrvatinandPerko2012;CigličandPerko2012; 2013). This is most likely why Slovenian geographers have only produced four true landscape typologies to date. The first one, which included eighteen landscape types, was produced by Anton Melik in 1946 (Melik1946),thesecondwasproducedbyDragoPerkoin1996andincludedninelandscapetypes(Kladnik1996;Perko1998;Perko2001;Perko2007),thethirdone,producedin2002byMetkaŠpeset al.,containedthirteen types (Špes et al. 2002), and the fourth was produced in 2014 by Drago Perko et al. and included twenty-four landscape types (Perko, Hrvatin and Ciglič 2015). Throughincreasingly more precise andaccessibledigital spatial data,technological development has also introduced changes to geographical classifications, including regionalization and typification, with variousmodelsandgeographicinformationsystemsbecomingwidelyusedandinfluentialresearchmeth­ods (Demeritt and Wainwright 2005). The term classification has several definitions (McGarigal, Cushman and Stafford 2000). Thus, for example, it can refer to any formal arrangement of data into a hierarchy of categories or distribution into classes(Whittow2000),orsystematicassignmenttoclassesorgroupsbasedonsharedcharacteristics(Clark 1998). Classifications can be roughly divided into supervised and unsupervised classifications. To date, Sloveniangeographershaveprimarilymodeledsupervisedclassificationsandcomparedthemagainstthe 1996landscapetypology of Slovenia, which comprisesninetypescombined intofourgroups andismost widely used and also included in Slovenian legislation. Landscape typology models were produced using various supervised classification methods based on the training samples of already classified or defined cells of individual landscape types. It was determined that the method selection had a strong impact on the results because modeled typologies matched the original ones by 51 to 75% (Ciglič 2014; Ciglič and Perko2015).Duringthelasttestingoftheoriginal1996typology,itsaccuracywasestimatedat94%(Ciglič et al.2017).Thecapacityandimpactofmethodswerealsodeterminedusingdistortedclassifications(Ciglič 2018), which allowed for additional evaluations of the original classification. However, for the first time in Slovenian research history, this article deals with models of Slovenia’s landscapetypologiesbasedonunsupervisedclassificationmethodsandcontainthesamenumberoftypes (groups) as the original 1996 landscape typology. Unsupervised models are compared against the origi­nal typology and supervised models, which contributes to evaluating supervised and unsupervised classificationmethodsaswellasassessingthesuitabilityoftheoriginaltypology’sconceptanddesign.With regard to unsupervised classification, it is important to know that the characteristics of individual types (i.e., groups) are not known in advance, but that unsupervised classification methods define or identify them by themselves. 2 The importance of landscape classification Thereisstillnoagreementonwhetherlandscapeclassificationentailssearchingforactualormerelyabstract units (Gams 1984; Udo de Haes and Klijn 1994; Bailey 1996), but, in any case, a classification is a minoror major abstraction of differences between features (Natek and Žiberna 2004). It is natural for people to seekorderandorganizationamongphenomena(Haggett2001),whichalsoappliestospatialorlandscape phenomena. Because landscapes constantly change, constant verification of landscape types is essential (Mücheretal.2003),facilitatingmoreeconomicaluseofnaturalresourcesandtheirreplenishment.Landscape changesmustbe accompanied byongoingadaptation of society’sorganizationandoperationtothe envi­ronment (Plut 2005), which makes it possible to implement principles of sustainability in the economy, society, and environment. Landscape types are important because they relatively homogenously respondto human impact (Špes et al. 2002) and demand similar landscape planning. Spatial classification following the natural characteristics of a landscape forms the basis for optimal spatial organization. Environmental issues also relate more to natural borders than administrative ones (Bailey 1996;Olsonet al. 2001), whichiswhyspatialclassifications based on natural factorsarebecoming increasinglycommonandarereplacingpoliticalclassifications(Bernertet al.1997).Forinstance,NUTS3 regions,whichinthe Mediterraneanoften include rural and urbancoastalareas (Hazeuet al. 2011),may vary greatly in terms of their natural and social characteristics. Various disciplines use classifications adapted to the content of their research and their needs (e.g., climate, vegetation, and soil classifications), but often their work would be made easier through uniform landscapeclassifications(Brabyn2018),especiallyatdifferentspatiallevels(McMahon,WikenandGauthier 2004).Alandscapeclassificationatthehighestspatiallevelcanbeusedtooutlinebordersforgeneralpur­posesandvariousdisciplines,whereasspatialunitsorlandscapetypesatalowerlevelcanserveasastarting pointformorespecificpurposes(Bailey1996)orclassificationwithinindividualdisciplines.Thisrequires understandingoftherelationsbetweenalandscapeasawholeanditscomponents(e.g.,landuseandbio-diversity),whichisvitalformanagingtheenvironment(landscape)anditsresources(Jongmanet al.2006). Landscape classification is alsoimportant forpreserving the naturalandcultural landscape;invento­rying,evaluating,andmonitoringthecurrentsituation;managing,planning,andconductingmeasurements; exploring scenarios; sampling; transferring models into the physical environment; presenting landscape diversity;analyzingenvironmentalpressures;andsoon(RunhaarandUdodeHaes1994;Bailey1996;Bunce et al. 1996; Bernert et al.1997; Bastian 2000; Mücher et al. 2003; Loveland and Merchant 2004; Romportl and Chuman 2012). Based onallthe above,itisnotsurprisingthatinsomeplaces classificationsevenhavea formalchar­acter,beingpartofofficialdocumentsorevenlegislationinspecificcountries.In1996,theEUintroduced thePan-EuropeanBiologicalandLandscapeDiversityStrategy(Pan-European…1996),andin2000itadopt­ed theEuropeanLandscapeConvention(TheEuropean…2018). Forexample,inSlovenia thelandscape typology of the country (Perko 1998) is used in defining land quality assessment criteria (Pravilnik o določanju…2008). 3 Classification methods Classificationentailscombiningsimilarunitsbasedonlogicalcriteria(Dodge2008).Aspartofgeographic information systems, units are defined with p data layers or, in other words, they have p dimensions. Classifications can be made based on a single criterion or factor (a monothetic approach) or several fac­tors (a polythetic approach). The former involves a one-dimensional and the latter a multidimensional data space (Loveland and Merchant 2004). Manyclassificationmethodsareused(McGarigal,CushmanandStafford2000;Rogerson2006;Abonyi and Feil 2007; Dodge 2008; Warner and Campagna 2009) and, in terms of results, each of them more or lessimposesaspecificstructureandleadstoaspecificsolution.Therefore,itisbesttocomparetheresults ofvariousmethods(Ferligoj1989;McGarigal,CushmanandStafford2000;TheodoridisandKoutroumbas 2006; Ciglič 2018). Classification methods are divided into soft and hard or relative and absolute, but most often a dis-tinctionismadebetweensupervisedandunsupervisedclassifications(WarnerandCampagna2009).With supervisedmethods, knownvaluesof training cellsare availablefor classification,whereasthat isnot the case with unsupervised ones (Theodoridis and Koutroumbas 2006); both are, however, at least partially shaped by an individual’s subjective judgment and knowledge (Warner and Campagna 2009). Aspartofsupervisedclassification,specificexamplesofunitsareselectedfromindividualgroups,which shouldhavethemosttypicalvaluespossible,andbasedontheseexamplesrulesaredesignedforassigning alltheremainingunitstothetypesdefined inadvance. Incontrast,aspartofunsupervisedclassification, unitscanbecategorizedbasedontheircharacteristicsorvaluesevenwithoutanypriorinformationabout theunits(Ferligoj 1989; Oštir2006), whichisacertain advantagecomparedtosupervisedclassifications. The aim of unsupervised classification in groups is to achieve the maximum internal homogeneity and the minimum external isolation of groups (Ferligoj 1989) or to minimize variance within the group and maximize variance between groups (Rogerson 2006). Unsupervisedclassificationmethodshaveseveralweaknesses(McGarigal,CushmanandStafford2000), such as sensitivity to outliers, which these methods often assign to a separate class, and great dependency of results on the initial groups defined and their number. Theunsupervisedclassificationprocedurecanbedividedintofivebasicsteps(Ferligoj1989;Theodoridis and Koutroumbas 2006): • Step 1: selecting the units; • Step 2: selecting the variables; • Step 3: computing the similarities (and differences) between units; • Step 4: selecting and carrying out the classification method; and • Step 5: assessing the final classification, which can only be done by an expert in the relevant field. VariousrelativelydetaileddigitaldataonnaturalgeographicalfactorsareavailableforSlovenia(Ciglič et al. 2016). Among these, over forty variables or data layers were selected and then normalized to values from 0 to 100, and adjusted to a uniform resolution of 200m (the resolution of the least accurate layer), which means that Slovenia was divided into 506,450 units or cells. This completed the step 1 (selecting the units). Step2entailedanassessmentofalldatalayersintermsoftheirusefulnessforlandscapetypologymod­eling(Ciglič2012,2013,2014;CigličandPerko2017).Thefollowingthreecriteriawereapplied:correlation betweendatalayers,correlationbetween datalayersandavailable landscape typologies, andsuitabilityof data layers in terms of the classification level or scale (the scales were determined at which an individual data layer was still sufficiently diverse to be suitable for classification). Eliminating less important data layers can reduce the time and costs of implementation, while also simplifyingtheunderstandingofmodelingprocedures (Jianget al. 2008;TirelliandPassani2011). Based onthecriteriamentionedabove,thefollowingfourdatalayerswereselectedforlandscapetypologymod­eling: elevation, slope, rock permeability, and precipitation regime (the ratio between summer and fall precipitation). Steps 3 and 4 already dependon the individualmethods selected. Several unsupervisedclassification methods and their versions and settings in various software were tested for modeling the typologies of Slovenia.FourmethodsfromtheTerrSetsoftware(theformerIdrisi)wereselectedforpresentationinthis article:the histogrampeak analysis method, the iterativeself-organizingunsupervisedclassifiermethod, the k-means method, and the iterative self-organizing data analysis method. 3.1 Histogram peak analysis Thehistogrampeakanalysismethodisbasedonfrequencydistributionandclassifiescellsingroupsusing a multidimensional histogram (Richards and Jia 2006). It first looks for peaks in a multidimensional his-togram(i.e.,thevalueswiththehighestfrequency)andthenassignseverycelltoitsnearestpeak,forming groups. Theareasbetweenthepeaks(i.e.,valleys)havevalueswiththelowestfrequency,creatingbound­aries between groups (Eastman 2016). In the TerrSet program, this method is available as part of the CLUSTER module (TerrSet…2015), inwhichtheusercansetthefollowingparameters:numberofdatalayers,numberofclassesforeachdata layerorgraylevels,cutoffsforexcludingextremevaluesorthesaturationpercentage,generalizationlevel (broadorfine)foridentifyingpeaks,andclusteringrule,throughwhichtheusercandroplesssignificant clusters, set the maximum number of clusters, or retain all clusters. The following settings were selected: gray levels = 6, saturation percentage = 1%, generalization level = fine, and clustering rule = maximum nine groups. 3.2 The iterative self-organizing unsupervised classifier method InTerrSet,theiterativeself-organizingunsupervisedclassifiermethodisavailableaspartoftheISOCLUST module (TerrSet…2015), which in fact uses three other modules for this method. With the CLUSTER moduleitfirstderivestheinitialclusters(orseeds)usingamultidimensionalhistogram(likewiththehis­togram peak analysis; see Section 3.1), after which it employs the modules MAKESIG and MAXLIKE to automatically define the training sites and perform a supervised classification. It repeats the procedure several times using new training cells from individual clusters. Because of the efficiency of the seeding step, very few iterations are usually required to achieve a stable cluster. The user first selects the data layers and then specifies the number of iterations, the desired number of clusters, and the minimum sample size per class. The following settings were selected: number of iterations = 99, number of clusters desired = 9, and minimum sample size per class = 40. 3.3 K-means Classificationfollowingthek-meansmethodisbasedondistancesbetweencellsinmultidimensionalspace. The final classification strongly depends on the definition of the number of clusters (k) and the initial-izationofcentroids(centers) –thatis,centralcellsaroundwhichothercellsgather.First,theuserspecifies thenumberofclustersandthemethods(rules)forinitializingtheclustercentroids,afterwhichnew,more appropriate centroids are calculated. This procedure is repeated until the new centroids are the same as the centroids from the previous iteration (Ferligoj 1989; Richards and Jia 2006). Initial centroids can also be selected randomly or dispersed evenly between cells (Ferligoj 1989). In TerrSet, this method is available as part of the KMEANS module (TerrSet…2015), in which dif­ferent rules for initializing the centroids can be selected: • The random partition rule randomly assigns each cell to one of the k clusters and then determines the initial centroids; • The random seed rule randomly selects k cells as the initial centroids and then assigns each cell to one of the k clusters according to the minimum-distance rule; and • Thediagonalaxisrulesystematicallysortskcentroidsfromthen-dimensionalspacefromtheminimum to the maximum value of n data layers. In addition to specifying the rule for initializing the cluster centroids, the user selects the maximum numberofoutputclustersandtheoptiontomerge(overly)smallclusters,whichdonotexceedtheselect-ed percentage of the entire image cells, with larger ones. The user also specifies two stopping criteria to terminate the clustering process. With the first one, the process is terminated if during the last iteration the percentage of migrating cells is less than a specified percentage of all cells, and, with the second cri­terion, the process stops when a specified number of iterations has been reached. Thefollowingsettingswereselectedforthisstudy:maximumnumberofoutputclusters = 9,clustercen­troidinitializationrule =randomseed,mergeclusterswithproportionslessthanorequalto1%,andstopping criteria = percentage of migration cells (pixels) less than or equal to 1% and maximum iterations 999. 3.4 Iterative self-organizing data analysis Theiterativeself-organizingdataanalysis(ISODATA)isanimprovedk-meansmethod.Itcanmergeclus­ters,justlikethek-meanstechnique,butitcanalsosplitthem.Itdeterminestheinitialcentroidsandclusters the same way as the k-means method, after which it calculates the standard deviation within each cluster and the distances between cluster centroids. It splits a cluster into two if the standard deviation is higher than the one specified by the user, or it merges two clusters if the distance between two cluster centroids is smaller than the one specified by the user. It repeats the process with new clusters and new centroids. Theprocessisterminatedifthestandarddeviationanddistancesbetweencentroidsnolongermakeitpos­sible to merge or split clusters, if during the last iteration the percentage of migrating pixels is less than that specified by the user or if the process reaches the number of iterations specified. In TerrSet, this method is available as part of the ISODATA module (TerrSet…2015). The following settingswerespecifiedforthisstudy:initialnumberofclusters =9,maximumnumberofoutputclusters =9, cluster centroid initialization rule = random seed, stopping criteria = percentage of migration cells (pix­els)lessthanorequalto1%andmaximumiterations999,minimumclustersize = 100,standarddeviation withinaclusterforsplitting = 25,Euclideandistancebetweenclustersformerging = 12.5,andmaximum number of pairs to merge within an iteration = 2. Figure 1: CLUSTER, ISOCLUST, KMEANS, and ISODATA modules in the TerrSet program. 4 Landscape typology modeling using unsupervised classification methods Eachofthefourmethodspresentedwasusedtoproducetypologieswithvariousnumbersofclasses.This article describes in detail modeled landscape typologies with nine classes, which were compared against thebest-establishedlandscapetypologyofSloveniaof1996,whichcontainsninelandscapetypescombined into four landscape type groups. Theprocessofcreatingthe1996landscapetypologystartedin1995.Perko(1998)enteredthefollowing four data layers into the geographic information system: surface elevation, surface inclination, lithology, andvegetationtypes.Theinclinationandelevationdatawerebasedona100mdigitalelevationmodel,and the lithology andvegetation data were obtained through digitization of a1:250,000 lithological mapwith thirty-sevenbasicunits(Verbič1998)andavegetationmapwithsixty-twobasicunits(Zupančičetal.1998) converted to a 100m raster grid. All four layers were then generalized and simplified into seven classes. Perkooverlaid (intersected) allfourlayers. Altogether2,401 different combinationsweretheoretical­lypossible.Perkofilteredtheintersectedlayerthreetimesusingthemodusinsideofamoving11×11cell squarewindow,obtainingforty-eightlargerandspatiallyseparatehomogenouscoreswiththesamecom­binationofelevation,inclination,lithology,andvegetation. Heprintedthecoresona1:250,000mapand, withthehelpofexpertsforindividualpartsofSlovenia,manuallyplottedtheboundaries,mostlyinmor­phologicalboundariesandlargerwatercourses.Intheend,hecombinedtheseforty-eightmanuallydelineated landscape units into nine landscape types, which he merged further into four landscape type groups. Theninelandscapetypesare:Alpinemountains,Alpinehills,Alpineplains,Pannonianlowhills,Pannonian plains, Dinaric plateaus, Dinaric lowlands, Mediterranean low hills, and Mediterranean plateaus. ThefourlandscapetypegroupsareAlpinelandscapes,Pannonianlandscapes,Dinariclandscapes,and Mediterranean landscapes. This was the first partly computerized typology of Slovenia. Its research bases were first presented in 1998(Perko1998).IthasbeenpublishedinallmajorgeographicalworksonSloveniaissuedafterSlovenia’s independence. Since 2008, it has also been part of Slovenian tax legislation and has been used for rating agricultural land according to the Rules on Determining and Administering Land Rating. As already described in Chapter 3 the CLUSTER, ISOCLUST, KMEANS, and ISODATA modules in the TerrSet software were used for modeling. The modeled typologies were compared against the origi­nal1996typologyintermsofthecorrelationcoefficientbetweenthemodeledtypologiesandtheoriginal typology,theclusterdensityinthemodeledtypologiesbylandscapetypeoftheoriginaltypology,andthe ratiobetweenthe actualand theoretical clusterfrequencyinthemodeledtypologiesbylandscapetypeof the original typology. The first indicator was cluster density by landscape type of the original typology – that is, the num­berofcellsinanindividualclusterper100cellsofaspecificlandscapetype.Themaximumvalueofdensity possibleis100,whichiswhenallthecellsofanindividualclusterliewithinaspecificlandscapetype,and the minimum value is 0, when a specific landscape type does not contain even a single cell of this cluster. If the cells of all nine clusters were evenly distributed across the nine landscape types of the original typologyinthemodeledtypologies,thedensityofallclusterswouldbe11.Aclusterwithadensityofatleast twice as much (i.e., at least 22) was defined as a typical representative of this landscape type. If a modeled typologywascompletelythesameastheoriginaltypology,eachofthenineclusterswouldhaveadensityof 100 in only one landscape type, whereas the density in the remaining eight landscape types would be 0. If, forinstance,Cluster7ofamodeledtypologyhadthirty-threecellsperonehundredcellsofDinaricplateaus, Cluster7wouldbeagoodrepresentativeorapproximationoftheDinaricplateausintheoriginaltypology. Theindicator ratiobetweentheactualandtheoreticalclusterfrequency inthemodeledtypologies by landscape type of the original typology relies on contingency tables, in which the rows and columns representclustersandlandscapetypes,andthecellscontaintheactualfrequency(number)ofcellsinindi­vidualclustersbyindividuallandscapetype.Clustersthathadtheiractualfrequencyinaspecificlandscape typeatleast twicethe theoreticalfrequency weredefinedastypicalrepresentativesofthislandscapetype. Forexample,inTable1,attheintersectionofCluster9,whichwasspecifiedusingtheISODATAmethod, and the Alpine mountains landscape type the actual frequency – that is, the number of cells of Cluster 9 thatliewithintheAlpinemountains(theirtotalis27,975) –isprovided.Thetheoreticalfrequencyofthat table cell is 4,811 and equals the total of all cells of Cluster 9 in the last column (31,835) and all cells of the Alpinemountainsinthelastrow(76,533)dividedbythenumberofallcells(506,450).Becausetheratiobetween theactualandtheoreticalfrequency(5.82)isgreaterthan2,Cluster9isagoodrepresentativeoftheAlpine mountains. The correlation coefficient between the modeled typologies and the original typology also relies on contingencytables. Cramer’s V wasselectedfor the study; it haslower values than the Pearson correlation coefficient,butitdoesnotdependonthesizeofthetablesandthereforeallowscomparisonsbetweentables with different numbers of columns and rows or between typologies with different numbers of classes. Table 1: Example of arranging cells of a nine-cluster typology modeled using the ISODATA module by nine landscape types of the original typology. ISODATA Alpine Alpine Alpine Pannonian Pannonian Dinaric Dinaric Medi- Medi-Total module mountains hills plains low plains plateaus lowlands terranean terranean hills low hills plateaus Cluster 1 5,288 34,123 2,382 31,609 1,264 3,747 5,448 0 0 83,861 Cluster 2 250 6,543 2,435 30,360 25,360 0 0 0 0 64,948 Cluster 3 5,762 29,497 143 593 11 78 88 0 0 36,172 Cluster 4 8,365 1,395 224 0 0 33,590 9,724 5,654 13,885 72,837 Cluster 5 4,516 10,161 888 0 0 12,956 6,570 18,477 2,738 56,306 Cluster 6 14,932 8,997 115 5 0 11,493 366 663 183 36,754 Cluster 7 2,379 4,470 13,832 3,349 5,545 20,106 20,861 1,681 27 72,250 Cluster 8 7,066 20,840 447 8,803 156 9,845 4,330 0 0 51,487 Cluster 9 27,975 452 15 0 0 3,375 0 15 3 31,835 Total 76,533 116,478 20,481 74,719 32,336 95,190 47,387 26,490 16,836 506,450 4.1 Modeling using nine clusters A graphic presentation of cluster distribution according to the four selected unsupervised classification methods using nine clusters is provided in Figure 2. Theoriginaltypologywithninelandscapetypes and the original typology with four landscape type groups are added for comparison. Thecorrelationcoefficientsbetweentheoriginalclassificationandtheclassificationsproducedusing the four modules presented above are as follows: 0.4489 for the CLUSTER classification, 0.4159 for the ISOCLUST classification (the lowest among all four classifications), 0.4609 for the KMEANS classifica­tion (the highest among all classifications), and 0.4518 for the ISODATA classification. Thedensityofcellsinanindividualclusterbyninelandscapetypes(Table2)showshowwelltheclus­ters spatially match thelandscape types. Clusters withadensityof22ormorearegood representativesor approximations of a specific landscape type. IntheCLUSTERclassification,Cluster1isagoodapproximationoftheAlpinehillsandthePannonian low hills, Cluster 2 is a good approximation of the Alpine plains, Pannonian plains, Pannonian low hills, andDinariclowlands,Cluster3oftheAlpinemountainsandDinaricplateaus,Cluster5ofthePannonian plains,andClusters 6 and9oftheMediterraneanlowhillsandMediterraneanplateaus,whereasClusters 4, 7,and8arenotagoodapproximationofevenasinglelandscapetype. Cluster4demonstratesthehighest cell density (18) in the Alpine hills, Cluster 7 in the Alpine mountains (18), and Cluster 8 in the Dinaric plateaus(15).Thus,Cluster1hasadensityover22intwolandscapetypes,Cluster2infourtypes,Cluster 3 in two types, Cluster 5 in one type, and both Clusters 6 and 9 in two types. All landscape types are rep-resented:fivetypesinoneclusterandfourtypesintwoclusters,whichmeansthattheseclustersareagood spatial match with two landscape types, not just one. IntheISOCLUSTclassification,Cluster1isagoodapproximationoftheAlpinehills,Cluster2ofthe Dinaricplateaus,Cluster3oftheDinaricplateaus,Dinariclowlands,andMediterraneanplateaus,Cluster 4 of the Mediterranean low hills, Cluster 7 of the Pannonian low hills and Pannonian plains, Cluster 8 of the AlpineplainsandPannonian plains, and Cluster 9of theAlpine hillsandPannonian low hills, whereas Clusters5and6arenotagoodapproximationofevenasinglelandscapetype.Cluster5hasthehighestden­sity(17)intheAlpinehillsandCluster6in thePannonianlowhills(20). Thus,Cluster1hasadensityover 22 in one landscape type of the original typology, Cluster 2 the same, Cluster 3 in four types, and Clusters 7, 8 and 9 in twotypes. All landscape types arerepresented: sixin one cluster and three in two clusters. In the KMEANS classification, Cluster 1 is a good approximation of the Alpine hills and Pannonian low hills, Cluster 3 of the Alpine plains and Dinaric lowlands, Cluster 4 of the Alpine mountains, Cluster 6 of the Alpine hills, Cluster 7 of the Mediterranean low hills, Cluster 8 of the Pannonian low hills and Pannonian plains, and Cluster 9 of the Dinaric plateaus, Dinaric lowlands, and Mediterranean plateaus, 9 9 LANDSCAPE TYPE GROUPS LANDSCAPE TYPES Alpine landscapes Alpine mountains Pannonian landscapes Alpine hills Dinaric landscapes Alpine plains Mediterranean landscapes Pannonian low hills Pannonian plains Dinaric plateaus Dinaric lowlands Content by: Rok Ciglič,Mauro Hrvatin, Drago Perko Map by: Mauro Hrvatin Mediterranean low hills Mediterranean plateaus ©2018,ZRC SAZUAnton Melik Geographical Institute Figure 2: Nine-cluster typologies modeled using the unsupervised classification methods (in the key the clusters are ordered based on the percentage of cells in an individual cluster: the one with the highest percentage is at the top and the one with the lowest percentage is at the bottom). whereasClusters2and5arenotagoodapproximationofevenasinglelandscapetype. Cluster 2 displays the highest density (21) in the Alpine mountains and Cluster 5 in the Alpine hills (18). Thus, Cluster 1 hasadensityabove22intwolandscapetypesoftheoriginaltypology,asdoesCluster3,Clusters4,6,and 7 inonetype, Cluster 8 in twotypes, andCluster 9inthree types. All landscapetypesare represented:six in one cluster and three in two clusters. IntheISODATAclassification,Cluster1isagoodapproximationoftheAlpinehillsandPannonianlow hills,Cluster2ofthePannonianlowhillsandPannonianplains,Cluster3oftheAlpinehills,Cluster4ofthe DinaricplateausandMediterraneanplateaus,Cluster5oftheMediterraneanlowhills,Cluster7oftheAlpine plainsandDinariclowlands,andCluster9oftheAlpinemountains,whereasClusters6and8arenotagood approximation of even a single landscape type. Cluster 6 has the highest density (20) in the Alpine moun­tains and Cluster 8 in the Alpine hills (18). Thus, Cluster 1 has a density of over 22 in two landscape types, Cluster2thesame,Cluster3inonetype,Cluster4intwotypes,Cluster5inonetype,Cluster7intwotypes, andCluster9inone. Alllandscapetypesarerepresented:seventypesinoneclusterandtwointwoclusters. The indicator ratiobetweentheactualandtheoreticalclusterfrequency inthemodeledtypologies bylandscapetypeoftheoriginaltypology(Table3)showswhichclustersaregoodrepresentativesofaspe­cific landscape type. Good representatives are the clusters whose actual frequency is at least twice their theoretical frequency. IntheCLUSTERclassification,Cluster1isagoodapproximationoftheAlpinehills(aratioof2.37),Cluster 2 of the Alpine plains (3.93) and Dinaric lowlands (2.45), Cluster 3 of the Alpine mountains (3.08), Cluster 5ofthePannonianplains(asmuchas8.99),Cluster6oftheDinariclowlands(3.03),Mediterraneanlowhills (3.66),andMediterraneanplateaus(5.83),Cluster7oftheAlpinemountains(3.09),Cluster8oftheDinaricplateaus (3.02) and Dinaric lowlands (2.78), and Cluster 9 of the Mediterranean low hills (as much as 9.78) and Mediterranean plateaus (even 10.79), and Cluster 4 has a ratio below 2 with all the landscape types. Its ratioisthehighestwiththeAlpinehills(1.44).Threelandscapetypeswitharatioabove2appearinoneclus­ter,fourintwoclusters,theDinariclowlandseveninthreeclusters,andthePannonianlowhillsinnone.The ratioofthePannonianlowhillsisthehighestinCluster5(i.e.,1.79,whichisclosetothethresholdvalueof2). IntheISOCLUSTclassification,Cluster1isagoodapproximationoftheAlpinehills(aratioof3.47), Cluster2oftheAlpinemountains(2.08)andDinaricplateaus(2.40),Cluster3oftheDinaricplateaus(2.35) andMediterraneanplateaus(3.29),Cluster4oftheMediterraneanlowhills(6.01),Cluster5oftheAlpinehills (2.11), Cluster 6 of the Pannonian low hills (4.96), Cluster 7 of the Pannonian low hills (2.65) and Pannonian plains (2.73), Cluster 8 of the Alpine plains (4.88) and Pannonian plains (6.45), and Cluster 9 of the Pannonian low hills (2.94). Four landscape types with a ratio above 2 appear in one cluster, three in two clusters, and the Pannonian hills even in three clusters. In the KMEANS classification, Cluster 1 is a good approximation of the Pannonian low hills (a ratio of 2.58), Cluster 2 of the Alpine mountains (2.98), Cluster 3 of the Alpine plains (4.75) and Dinaric low­lands (3.13),Cluster4ofthe Alpinemountains(5.81),Cluster6oftheAlpinehills(3.55),Cluster7ofthe Mediterranean low hills (2.65), Cluster 8 of the Pannonian low hills (3.12) and Pannonian plains (6.21),and Cluster 9 of the Dinaric plateaus and Mediterranean plateaus, and Cluster 5 has a ratio below 2 with allthelandscapetypes.ItsratioisthehighestwiththeAlpinehills(1.69).Sevenlandscapetypeswitharatio above 2 appear in one cluster and two in two clusters. In the ISODATA classification, Cluster 1 is a good approximation of the Pannonian low hills (a ratio of 2.55), Cluster 2 of the Pannonian low hills (3.17) and Pannonian plains (6.12), Cluster 3 of the Alpine hills (3.55), Cluster 4 of the Dinaric plateaus (2.45) and Mediterranean plateaus (5.73), Cluster 5 of the Mediterranean low hills (6.27), Cluster 6 of the Alpine mountains (2.69), Cluster 7 of the Alpine plains (4.73) and Dinaric lowlands (3.09), and Cluster 9 of the Alpine mountains. Cluster 8 has a ratio below 2 withallthelandscapetypes.ItsratioisthehighestwiththeAlpinehills(1.76).Sevenlandscapetypeswith a ratio above 2 appear in one cluster and two in two clusters. Takingintoaccountallthreeindicatorspresented,thetypologymodeledusingtheiterativeself-orga­nizingdataanalysismethod(theISODATAmodule)andthetypologymodeledusingthek-meansmethod (the KMEANS module) are the closest to the original typology of nine landscape types. TheKMEANStypologyhasaslightlyhighercorrelationcoefficientthantheISODATAtypologybut, onthe other hand, the ISODATA typology displaysa density above the threshold value of 22 in twoclus­ters simultaneously only with two landscape types compared to three in the KMEANS typology. Thetypologymodeledusingtheiterativeself-organizingunsupervisedclassifiermethod(theISOCLUST module) matches the original typology the least. Table 2: Cell density of an individual cluster in the modeled typologies by nine landscape types of the original typology (green numbers indicate a density of 22 or more). Modules Alpine Alpine Alpine Pannonian Pannonian Dinaric Dinaric Medi- Medi- Total mountains hills plains low plains plateaus lowlands terranean terranean hills low hills plateaus CLUSTER 1 17.79 62.21 13.76 42.76 3.86 4.16 14.76 0.03 0.00 26.27 CLUSTER 2 1.25 6.12 63.91 26.86 28.21 13.01 39.76 2.60 0.27 16.26 CLUSTER 3 47.27 3.55 0.06 0.00 0.00 34.81 1.63 8.47 7.97 15.37 CLUSTER 4 10.25 18.15 4.85 16.62 0.58 13.17 8.11 17.97 1.75 12.64 CLUSTER 5 0.47 3.87 6.07 13.44 67.34 0.00 0.00 0.00 0.00 7.49 CLUSTER 6 1.11 0.58 0.00 0.00 0.00 10.01 19.91 24.05 38.33 6.58 CLUSTER 7 17.77 4.68 3.22 0.32 0.01 8.50 2.27 0.16 0.00 5.76 CLUSTER 8 2.02 0.82 8.12 0.00 0.00 14.74 13.56 0.13 0.31 4.88 CLUSTER 9 2.07 0.02 0.00 0.00 0.00 1.59 0.00 46.58 51.37 4.76 ISOCLUST 1 8.25 29.96 1.43 0.81 0.03 1.42 0.39 0.04 0.11 8.63 ISOCLUST 2 19.67 5.15 0.88 0.01 0.00 22.71 6.51 0.22 11.25 9.46 ISOCLUST 3 30.42 3.31 16.48 0.14 0.09 54.97 42.06 10.07 76.94 23.40 ISOCLUST 4 20.63 11.06 2.32 0.04 0.00 8.69 6.64 71.91 5.43 11.96 ISOCLUST 5 3.70 17.18 2.65 9.28 0.64 5.23 12.08 0.00 0.00 8.14 ISOCLUST 6 1.21 1.96 0.97 20.32 1.66 0.70 0.12 0.31 4.80 4.10 ISOCLUST 7 0.16 4.19 13.36 25.94 26.69 1.52 18.32 13.40 0.79 9.79 ISOCLUST 8 13.51 5.00 50.69 1.94 66.97 1.53 0.87 4.05 0.10 10.39 ISOCLUST 9 2.45 22.18 11.23 41.53 3.91 3.23 13.00 0.00 0.58 14.15 KMEANS 1 6.73 29.41 11.18 42.73 3.91 3.91 11.29 0.00 0.00 16.58 KMEANS 2 21.05 6.55 0.64 0.10 0.00 11.58 0.87 1.06 0.98 7.07 KMEANS 3 3.36 3.53 67.39 3.75 17.05 20.48 44.39 9.21 0.43 14.19 KMEANS 4 35.81 0.39 0.08 0.00 0.00 3.50 0.00 0.04 0.02 6.17 KMEANS 5 7.93 18.11 2.47 13.21 0.58 11.81 11.00 0.00 0.00 10.70 KMEANS 6 7.26 23.70 0.41 0.79 0.03 0.00 0.00 0.00 0.00 6.68 KMEANS 7 6.99 10.95 5.06 0.00 0.00 7.20 8.59 71.25 6.52 9.88 KMEANS 8 0.35 5.50 11.94 39.43 78.43 0.00 0.00 0.00 0.00 12.62 KMEANS 9 10.52 1.86 0.82 0.00 0.00 41.52 23.86 18.44 92.05 16.11 ISODATA 1 6.91 29.30 11.63 42.30 3.91 3.94 11.50 0.00 0.00 16.56 ISODATA 2 0.33 5.62 11.89 40.63 78.43 0.00 0.00 0.00 0.00 12.82 ISODATA 3 7.53 25.32 0.70 0.79 0.03 0.08 0.19 0.00 0.00 7.14 ISODATA 4 10.93 1.20 1.09 0.00 0.00 35.29 20.52 21.34 82.47 14.38 ISODATA 5 5.90 8.72 4.34 0.00 0.00 13.61 13.86 69.75 16.26 11.12 ISODATA 6 19.51 7.72 0.56 0.01 0.00 12.07 0.77 2.50 1.09 7.26 ISODATA 7 3.11 3.84 67.54 4.48 17.15 21.12 44.02 6.35 0.16 14.27 ISODATA 8 9.23 17.89 2.18 11.78 0.48 10.34 9.14 0.00 0.00 10.17 ISODATA 9 36.55 0.39 0.07 0.00 0.00 3.55 0.00 0.06 0.02 6.29 Table 3: Ratio between the actual and theoretical cluster frequency in the modeled typologies by nine landscape types of the original typology (green numbers indicate ratios of 2 or more). Modules Alpine Alpine Alpine Pannonian Pannonian Dinaric Dinaric Medi-Medi-Total mountains hills plains low plains plateaus lowlands terranean terranean hills low hills plateaus CLUSTER 1 0.68 2.37 0.52 1.63 0.15 0.16 0.56 0.00 0.00 1.00 CLUSTER 2 0.08 0.38 3.93 1.65 1.74 0.80 2.45 0.16 0.02 1.00 CLUSTER 3 3.08 0.23 0.00 0.00 0.00 2.27 0.11 0.55 0.52 1.00 CLUSTER 4 0.81 1.44 0.38 1.31 0.05 1.04 0.64 1.42 0.14 1.00 CLUSTER 5 0.06 0.52 0.81 1.79 8.99 0.00 0.00 0.00 0.00 1.00 CLUSTER 6 0.17 0.09 0.00 0.00 0.00 1.52 3.03 3.66 5.83 1.00 CLUSTER 7 3.09 0.81 0.56 0.06 0.00 1.48 0.39 0.03 0.00 1.00 CLUSTER 8 0.41 0.17 1.67 0.00 0.00 3.02 2.78 0.03 0.06 1.00 CLUSTER 9 0.44 0.00 0.00 0.00 0.00 0.33 0.00 9.78 10.79 1.00 ISOCLUST 1 0.96 3.47 0.17 0.09 0.00 0.17 0.05 0.00 0.01 1.00 ISOCLUST 2 2.08 0.55 0.09 0.00 0.00 2.40 0.69 0.02 1.19 1.00 ISOCLUST 3 1.30 0.14 0.70 0.01 0.00 2.35 1.80 0.43 3.29 1.00 ISOCLUST 4 1.73 0.92 0.19 0.00 0.00 0.73 0.56 6.01 0.45 1.00 ISOCLUST 5 0.45 2.11 0.33 1.14 0.08 0.64 1.48 0.00 0.00 1.00 ISOCLUST 6 0.30 0.48 0.24 4.96 0.41 0.17 0.03 0.07 1.17 1.00 ISOCLUST 7 0.02 0.43 1.37 2.65 2.73 0.16 1.87 1.37 0.08 1.00 ISOCLUST 8 1.30 0.48 4.88 0.19 6.45 0.15 0.08 0.39 0.01 1.00 ISOCLUST 9 0.17 1.57 0.79 2.94 0.28 0.23 0.92 0.00 0.04 1.00 KMEANS 1 0.41 1.77 0.67 2.58 0.24 0.24 0.68 0.00 0.00 1.00 KMEANS 2 2.98 0.93 0.09 0.01 0.00 1.64 0.12 0.15 0.14 1.00 KMEANS 3 0.24 0.25 4.75 0.26 1.20 1.44 3.13 0.65 0.03 1.00 KMEANS 4 5.81 0.06 0.01 0.00 0.00 0.57 0.00 0.01 0.00 1.00 KMEANS 5 0.74 1.69 0.23 1.23 0.05 1.10 1.03 0.00 0.00 1.00 KMEANS 6 1.09 3.55 0.06 0.12 0.01 0.00 0.00 0.00 0.00 1.00 KMEANS 7 0.71 1.11 0.51 0.00 0.00 0.73 0.87 7.21 0.66 1.00 KMEANS 8 0.03 0.44 0.95 3.12 6.21 0.00 0.00 0.00 0.00 1.00 KMEANS 9 0.65 0.12 0.05 0.00 0.00 2.58 1.48 1.14 5.71 1.00 ISODATA 1 0.42 1.77 0.70 2.55 0.24 0.24 0.69 0.00 0.00 1.00 ISODATA 2 0.03 0.44 0.93 3.17 6.12 0.00 0.00 0.00 0.00 1.00 ISODATA 3 1.05 3.55 0.10 0.11 0.00 0.01 0.03 0.00 0.00 1.00 ISODATA 4 0.76 0.08 0.08 0.00 0.00 2.45 1.43 1.48 5.73 1.00 ISODATA 5 0.53 0.78 0.39 0.00 0.00 1.22 1.25 6.27 1.46 1.00 ISODATA 6 2.69 1.06 0.08 0.00 0.00 1.66 0.11 0.34 0.15 1.00 ISODATA 7 0.22 0.27 4.73 0.31 1.20 1.48 3.09 0.44 0.01 1.00 ISODATA 8 0.91 1.76 0.21 1.16 0.05 1.02 0.90 0.00 0.00 1.00 ISODATA 9 5.82 0.06 0.01 0.00 0.00 0.56 0.00 0.01 0.00 1.00 4.2 Modeling using four clusters A graphic presentation of cluster distribution according to the four selected unsupervised classification methodsusingfourclustersisprovidedinFigure3.Theoriginaltypologywithfourlandscapetypegroups and the original typology with nine landscape types are added for comparison. The four-cluster typologies modeled were also compared against the original typology. Taking into accountallthreeindicatorspresented,thetypologymodeledusingtheiterativeself-organizingdataanaly­sismethod(theISODATAmodule)istheclosesttotheoriginaltypologywithfourlandscapetypegroups and the original typology with nine landscape types, and the typology modeled using the iterative self-organizingunsupervisedclassifiermethod(theISOCLUSTmodule)matchtheoriginaltypologiestheleast. Itisinterestingthatthecorrelationcoefficientsbetweenthemodeledtypologiesandtheoriginaltypol­ ogy with nine landscape types are approximately one-third higher than with the original typology with four landscape type groups. 5 Quality of the typologies modeled Acomparisonbetweenthenine-clustertypologiesmodeledshowedthattheISODATAandKMEANStypolo­gies are the closest to the original typology with nine landscape types and can therefore be regarded as thebestapproximations.Thisappliestothemodeledtypologiesasawhole,butthequestionwaswhether this also applies to individual landscape types or the other two modules may prove to be more effective in assigning cells totheclusterthatisthebest approximationofaspecificlandscape type(e.g.,the Alpine mountains). The Herfindahl–Hirschman Index was used to calculate the concentration of nine landscape types of the original typology by nine clusters of the modeled typologies and vice versa (Tables 4 and 5). The value oftheindexranges from0 to 1. In the case presented, it hadvalue1 if all thecells ofa specificland-scape type were in only one cluster and it had value 0 if the cells of a specific landscape type were evenly distributed across all clusters. The higher the index value, the better a cluster of a specific module is an approximation of a specific landscape type. Every landscape type has the highest concentration index with one module (marked green in Table 4)andthelowestwithanother(markedredinTable4). IntheCLUSTERtypology,twolandscapetypes – the Alpine mountains and the Alpine hills – are the most concentrated across clusters (i.e., they have the highest concentration index), and as many as four types are concentrated the least: the Dinaric plateaus, Dinariclowlands,Mediterraneanlowhills,andMediterraneanplateaus.IntheISOCLUSTtypology,only the Dinaric plateaus are the most concentrated across clusters, whereas as many as five are concentrated the least: the Alpine mountains, Alpine hills, Alpine plains, Pannonian low hills, and Pannonian plains. In the KMEANS typology, the three types most concentrated across clusters are the Dinaric lowlands, Mediterranean low hills, and Mediterranean plateaus, whereas no type displayed a particularly low con­centration. Similarly, in the ISODATA typology, three types were most concentrated across clusters (i.e., the Alpine plains, Pannonian low hills, and Pannonian plains), and no type displayed a particularly low concentration. The values of the concentration indexes show that the CLUSTER module is the best for defining Alpine landscape types, the KMEANS module for determining Mediterranean landscape types, andtheISODATAmodulefordefiningPannonianlandscapetypes(Table4).Differencesbetweenthetypolo­giesmodeledarealsoevidentfromtheoppositeperspective –thatis,intermsofclusterconcentrationby landscape type (Table 5). Among the nine clusters used, the CLUSTER module proved to be the best for three clusters and the worst for three clusters, the ISOCLUST module was the best for one cluster, the KMEANS module was the best for two clusters and the worst for two clusters, and the ISODATA mod­ule was the best for three clusters and the worst for as many as four (Table 5). Thismeansthatindividuallandscapetypescanbedeterminedmoreeffectivelybyothermodulesthan the one used for a specific modeled typology; for example, the CLUSTER module for Alpine mountains and Alpine hills. The concentration index for Alpine mountains is 41% higher with the CLUSTER mod­ule than the KMEAN module, and even 93% higher for Alpine hills. The quality ofthe modeledtypologies isalso indicated bythe degree oftheir similarity with theorig­inal typology of 1996. According to the three indicators based on which the modeled typologies were LANDSCAPE TYPE GROUPS LANDSCAPE TYPES Alpine landscapes Alpine mountains Pannonian landscapes Alpine hills Dinaric landscapes Alpine plains Mediterranean landscapes Pannonian low hills Pannonian plains Dinaric plateaus Dinaric lowlands Content by: Rok Ciglič,Mauro Hrvatin, Drago Perko Map by: Mauro Hrvatin Mediterranean low hills Mediterranean plateaus ©2018, ZRC SAZUAnton Melik Geographical Institute Figure 3: Four-cluster typologies modeled using the unsupervised classification method (in the key the clusters are ordered based on the percentage of cells in an individual cluster: the one with the highest percentage is at the top and the one with the lowest percentage is at the bottom). compared against the original one, the best modeled typologies are those that have the highest correla­tioncoefficients,veryhighdensitiesbyindividualcellsofthecontingencytableandatthesametimevery low densities by other cells of the contingency table, and very high ratios by individual cells of the con­tingency table and at the same time very low ratios by other cells in that table. If the double density of cells in an individual cluster of modeled typologies by nine landscape types oftheoriginaltypology(Table2)istakenintoaccountasacriterionandthefrequenciesinthecellsofthe contingencytablethatmeetthisconditionareaddedup(valuesdensityof20ormorearetakenintoaccount), theISOCLUSTmodulehas66%ofallcells»properly«classified,thecorrespondingsharewiththeKMEANS and ISODATA modules is 65%, and with the CLUSTER module it is 59%. Based on the criterion of the doubleratiobetweentheactualandtheoreticalclusterfrequencyofmodeledtypologiesbyninelandscape typesoftheoriginaltypology(Table3),theISOCLUSTmodulehas56%ofcells»properly«classified,com­pared to 53% in the KMEANS module and 51% in the ISODATA module. The differences between the modules are very small. In terms of density, nearly two-thirds of the cells are »properly« classified, com­pared to only just over half in terms of the frequency ratio. Thedegreetowhichthetypologiesmodeledusingunsupervisedclassification methodsarelesscor­relatedwiththeoriginaltypologyof1996comparedtothetypologiesmodeledusingsupervisedclassification methods was assessed by comparing Cramer’s V (similar to how the modeled typologies were compared against the original one). The supervised classification models were designed using four supervised clas­sificationmethods:adecisiontree,k-nearestneighbors,maximumlikelihood,andminimumdistance.With allfourthesamefourdatalayerswereusedaswiththeunsupervisedclassifications.Themodelsweredesigned using two sets of training cells. For the first set, the cells were selected randomly and for the second set expert sampling was used, which means that the most representative areas were selected based on the researcher’sjudgment.Thus,withallmethods,randomsupervisedclassificationsweredistinguishedfrom theexpertones,whichyieldedeightmodeledtypologies.Typologiesusingsupervisedclassificationmeth­ods were modeled in 2013 (Ciglič 2014). Atfirstglance,itmayseemunusualthatthetypologiesmodeledusingarandomsamplegenerallyhad ahigher degree of correlation with the original typology(the correlation coefficient rangingfrom 0.5023 Table 4: Concentration indexes of landscape types of the original typology by clusters of modeled typologies (colors indicate the modeled typology or module where an individual landscape type is concentrated the most (green) or the least (red)). Modules Alpine Alpine Alpine Pannonian Pannonian Dinaric Dinaric Medi- Medi­ mountains hills plains low hills plains plateaus lowlands terranean terranean low hills plateaus CLUSTER 0.4589 0.5978 0.6092 0.4618 0.6902 0.3098 0.3885 0.4789 0.5871 ISOCLUST 0.3179 0.2962 0.4803 0.4488 0.6795 0.5353 0.3963 0.7001 0.7491 KMEANS 0.3263 0.3098 0.6479 0.5259 0.7755 0.3947 0.4439 0.7029 0.9128 ISODATA 0.3285 0.3168 0.6497 0.5292 0.7757 0.3431 0.4317 0.6920 0.8186 Average 0.3579 0.3801 0.5968 0.4914 0.7302 0.3957 0.4151 0.6435 0.7669 Table 5: Cluster concentration indexes of modeled typologies by landscape type of the original typology (colors indicate the modeled typology or module where an individual cluster is concentrated the most (green) or the least (red)). Modules Cluster 1 Cluster 2 Cluster 3 Cluster 4 Cluster 5 Cluster 6 Cluster 7 Cluster 8 Cluster 9 CLUSTER 0.5385 0.2778 0.5715 0.3326 0.5847 0.3768 0.4979 0.5700 0.5689 ISOCLUST 0.7859 0.4893 0.4308 0.3715 0.4642 0.7050 0.3698 0.4110 0.4951 KMEANS 0.4869 0.5102 0.3325 0.8686 0.3949 0.8086 0.3881 0.5512 0.4658 ISODATA 0.4834 0.5526 0.8075 0.4441 0.3433 0.4885 0.3351 0.4032 0.8698 Average 0.5737 0.4575 0.5356 0.5042 0.4468 0.5947 0.3977 0.4838 0.5999 withtheminimumdistancemethodto0.7261withthek-nearestneighborsmethod)thantheonesforwhich anexpertsamplewasused(0.5153withtheminimumdistancemethodand0.6020withthek-nearestneigh­bors method). The reason is that in expert-sample modeling some of the most typical areas of individual types (based on the researcher’s subjective judgment) were selected, and therefore the modeling did not covertheentirevariabilityofalandscapetypeoratleastnottothesameextentastherandom-samplemod-eling,where,giventhemoregeneralizedclassificationrules,thesamplewasmoreevenlydistributedacross a landscape type. Hence, with expert-sample modeling one can speak of over-fitting. This is also proven by the analysis of testing the success rate of modeling the training cells, where (precisely to the contrary) expert sampling achieved significantly higher scores than random sampling (Ciglič 2014). Thecorrelationcoefficientsbetweenthetypologiesmodeledusingunsupervisedclassificationmeth­odsandtheoriginaltypologyrangefrom0.4159withtheISOCLUSTERmoduleto0.4609withtheKMEANS module. Thismeans they areapproximately one-fifth lowerthanthe correlationcoefficientsbetweenthe originaltypologyandthetypologiesmodeledusingexpertsupervisedclassificationmethods,andathird lower than the correlation coefficients between the original typology and the typologies modeled using random supervised classification methods. 6 Conclusion The typologies modeled using the unsupervised classification methods presented only roughly approxi­matetheoriginaltypologyofSloveniaandtoalesserdegreethanthesupervisedclassificationmodels(Ciglič 2014).Thisisexpectedbecausetypologiesmodeledusingsupervisedclassificationmethodsarebasedon a referential or training classification or, in this case, the 1996 typology of Slovenia. On the other hand, the statistically significant correlation between the original typology of 1996 and the typologies modeled usingunsupervisedclassificationmethods(i.e.,completelyindependentlyofalreadyidentifiedlandscape types) confirms that the original landscape typology (at least in terms of the data layers that were taken into account) is fairly appropriate. However,because unsupervised classification methodshavecertainadvantages over supervised clas­sificationmethodsandbecausethenumberanddiversityofdatacoveringallofSlovenianterritoryisgrowing, itcanbe expectedthat thedifferences betweenthe qualityof supervisedand unsupervised classifications will become smaller and that modeled typologies will better approximate the actual situation, thus effec­tivelyreplacingtypologiesproducedusingtraditionalprocedures,whichareusuallymoretime-consuming and therefore more expensive. ACKNOWLEDGEMENT:TheauthorsacknowledgefinancialsupportfromtheSlovenianResearchAgency (program no. P6-0101: Geography of Slovenia and project no. L1-7542: Advancement of computationally intensive methods for efficient modern general-purpose statistical analysis and inference). 7 References Abonyi, J., Feil, B. 2007: Cluster analysis for data mining and system identification. Basel, Boston,Berlin, Birkhäuser. Bailey, R. G. 1996: Ecosystem geography. New York. Bastian, O. 2000: Landscape classification in Saxony (Germany) – a tool for holistic regional planning. Landscape and urban planning 50, 1–3. DOI: https://doi.org/10.1016/S0169-2046(00)00086-4 Bernert,J. A.,Eilers,J. M.,Freemark,K. E.,Ribic,C. 1997:Aquantitativemethodfordelineatingregions: AnexampleforthewesternCornBeltPlainsEcoregionoftheUSA. EnvironmentalManagement21-3. DOI: https://doi.org/10.1007/s002679900038 Brabyn, L. 1998: GIS analysis of macro landform. 10th colloquium of the Spatial Information Research Centre, University of Otago. Dunedin. Brabyn, L. 2018: NZ Landscape classification version II. Internet: https://www.waikato.ac.nz/eri/about/ people/larsb/New-Zealand-Landscape-Classification-v2.1.pdf (16.12.2018). Bunce,R.G.H,Barr,C.J.,Clarke,R.T.,Howard,D.C.,Lane,A.M.J.1996:Landclassificationforstrategic ecologicalsurvey.Journalofenvironmentalmanagement47-1.DOI:https://doi.org/10.1006/jema.1996.0034 Ciglič, R. 2012: Evaluation of digital data layers for establishing natural landscape types in Slovenia. Landscapes:Perception,knowledge,awareness,andaction.ProceedingsoftheFG-SHUInternational Symposium on Geography. New York. Ciglič, R. 2013: Uporaba geografskih informacijskih sistemov za določanje naravnih pokrajinskih tipov Slovenije. Ph.D. thesis, Univerza v Ljubljani, Filozofska fakulteta. Ljubljana. Ciglič,R.2014:AnalizanaravnihpokrajinskihtipovSlovenijezGIS-om.GeografijaSlovenije28.Ljubljana. Ciglič,R.2018:Assessingtheimpactofinputdataincongruityinselectedquantitativemethodsformodelling natural landscape typologies. Geografski vestnik 90-1. DOI: https://doi.org/10.3986/GV90107 Ciglič, R., Geršič, M., Perko, D., Zorn, M. (eds.) 2016: Digitalni podatki. GIS v Sloveniji 13. Ljubljana. Ciglič,R.,Perko,D.2012:SloveniaingeographicaltypificationsandregionalizationsofEurope.Geografski vestnik 84-1. Ciglič,R.,Perko,D. 2013:Europe’slandscapehotspots. ActageographicaSlovenica53-1. DOI:https://doi.org/ 10.3986/AGS53106 Ciglič, R., Perko, D. 2015: Modelling as a method for evaluating natural landscape typology: The case of Slovenia. Landscape Analysis andPlanning. Cham.DOI: https://doi.org/10.1007/978-3-319-13527-4 Ciglič,R.,Perko,D.2017:Amethodforevaluatingrasterdatalayersaccordingtolandscapeclassification scale. Ecological Informatics 39. DOI: https://doi.org/10.1016/j.ecoinf.2017.03.004 Ciglič,R.,Perko,D.,Hrvatin,M.,Štrumbelj,E.2017:Modelingandevaluatingolderlandscapeclassifications with modern quantitative methods. From pattern and process to people and action. Ghent. Clark, A. N. 1998: The Penguin dictionary of geography. London. Demeritt,D.,Wainwright,J.2005:Models,modellingandgeography.Questioninggeography:fundamental Debates. Malden. Dikau,R.,Brabb,E.E.,Mark,R.K.1991:LandformclassificationofNewMexicobycomputer.MenloPark. Dodge, Y. 2008: The concise encyclopedia of statistics. New York. Eastman, J. R. 2016: TerrSet. Geospatial Monitoring and Modeling System. Manual. Worcester. Ellison,A.M.2010:Repeatabilityandtransparencyinecologicalresearch. Ecology91-9.DOI:https://doi.org/ 10.1890/09-0032.1 Ferligoj, A. 1989: Razvrščanje v skupine. Teorija in uporaba v družboslovju. Ljubljana. Gabrovec, M., Hrvatin, M. 1998: Površje. Geografski atlas Slovenije. Ljubljana. Gallant, A. L., Douglas, D. B., Hoffer, R. M. 2005: Automated mapping of Hammond’s landforms. IEEE geoscience and remote sensing letters 2-4. Piscataway. Gams, I. 1983: Geografske značilnosti Slovenije. Ljubljana. Gams, I. 1984: Metodologija geografske razčlenitve ozemlja. Geografski vestnik 56. Gams, I. 1998: Pokrajinsko ekološka sestava Slovenije. Geografija Slovenije. Ljubljana. Gams,I.,Kladnik,D.,OroženAdamič,M. 1995:NaravnogeografskeregijeSlovenije. KrajevnileksikonSlovenije. Ljubljana. Haggett, P. 2001: Geography: a global synthesis. Harlow. Hammond, E. H. 1964: Analysis of properties in landform geography: An application to broadscale landformmapping.AnnalsofAssociationofAmericanGeographers54-1.DOI:https://doi.org/10.1111/ j.1467-8306.1964.tb00470.x Hazeu,G.W.,Metzger,M.J.,Mücher,C.A.,Perez-Soba,M.,Renetzeder,C.,Andersen,E.2011:European environmentalstratificationsandtypologies:Anoverview.Agriculture,ecosystemsandenvironment 142, 1–2. DOI: https://doi.org/10.1016/j.agee.2010.01.009 Hrvatin, M., Perko, D. 2012: Morphological typifications of Slovenia’s surface using global classification methods. Geografski vestnik 84-1. Ilešič, S. 1956: Slovenske pokrajine. Geografski obzornik 3-2. Ilešič, S. 1958: Problemi geografske rajonizacije ob primeru Slovenije. Geografski vestnik 29-30. Iwahashi,J.,Pike,R.J. 2006:AutomatedclassificationsoftopographyfromDEMsbyanunsupervisednested-meansalgorithmandathree-partgeometricsignature.Geomorphology86,3–4.DOI:https://doi.org/ 10.1016/j.geomorph.2006.09.012 Jiang,B.,Ding,X.,Ma,L.,He,Y.,Wang,T.,Xie,W.2008:Ahybridfeatureselectionalgorithm:combination of symmetrical uncertainty and genetic algorithms. Optimization and systems biology. Lijiang. Jongman,R. H. G.,Bunce,R. G. H.,Metzger,M. J.,Mücher,C. A.,Howard,D. C.,Mateus,V. L. 2006:Objectives andapplicationsofastatisticalenvironmentalstratificationofEurope.LandscapeEcology21-3.DOI: https://doi.org/10.1007/s10980-005-6428-0 Kladnik, D. 1996: Naravnogeografske členitve Slovenije. Geografski vestnik 68. Kladnik, D., Perko, D., Urbanc, M. 2009: Cultural landscapes in Slovenia from geographical perspective. Cultural landscape – Across disciplines. Kraków. Loveland, T. R., Merchant, J. M. 2004: Ecoregions and ecoregionalization: geographical and ecological perspectives. Environmental Management 34-S1. DOI: https://doi.org/10.1007/s00267-003-5181-x McGarigal,K.,Cushman,S.,Stafford,S.2000:Multivariatestatisticsforwildlifeandecologyresearch.New York. McMahon, G., Wiken, E. B., Gauthier, D. A. 2004: Toward a scientifically rigorous basis for developing mapped ecological regions. Environmental Management 34-1. DOI: https://doi.org/10.1007/s00267­004-0170-2 Melik, A. 1935: Slovenija – geografski opis. Ljubljana. Melik, A. 1946: Prirodnogospodarska sestava Slovenije. Geografski vestnik 18. Mücher,C.A.,Bunce,R.G.H.,Jongman,R.H.G.,Klijn,J.A.,Koomen,A.J.M.,Metzger,M.J.,Wascher, D. M. 2003:IdentificationandcharacterisationofenvironmentsandlandscapesinEurope. Alterrarapport 832. Wageningen. Natek, K. 1993: Tipi površja v Sloveniji. Geografski obzornik 40-4.Natek,K.,Žiberna, I. 2004: Naravnogeografske regionalizacijeSlovenije.Teorija in praksaregionalizacije Slovenije. Maribor. Olson, D. M., Dinerstein, E., Wikramanayake, E. D., Burgess, N. D., Powell, G. V. N., Underwood, E. C., D’Amico, J. A.,Itoua, I., Strand, H.E., Morrison, J. C.,Loucks,C.J., Allnut, T. F.,Ricketts, T. H., Kura, Y.,Lamoreux,J.F.,Wettengel,W.W.,Hedao,P.,Kassem,K.R.2001:TerrestrialecoregionsoftheWorld: a new map of lifeon earth: A newglobalmap ofterrestrial ecoregions provides an innovativetoolfor conserving biodiversity. BioScience 51-11. DOI: https://doi.org/10.1641/0006-3568(2001)051 [0933:TEOTWA]2.0.CO;2 Oštir, K. 2006: Daljinsko zaznavanje. Ljubljana. Pan-European Biological and Landscape diversity strategy, 1996. Internet: https://www.unibuc.ro/prof/ patru-stupariu_i_g/docs/2014/noi/03_11_52_38paneurop_strategie.pdf (12. 12. 2018). Perko, D. 1998: Regionalizacija Slovenije. Geografski zbornik 38. Perko,D.2001:AnalizapovršjaSlovenijesstometrskimdigitalnimmodelomreliefa.GeografijaSlovenije 3. Ljubljana. Perko, D. 2007: Landscapes. Slovenia in focus. Ljubljana. Perko, D., Hrvatin, M. 2009: Določanje morfoloških enot površja v Sloveniji s prirejeno Hammondovo metodo. Geografski vestnik 81-2. Perko,D.,HrvatinM.,Ciglič,R.2015:AmethodologyfornaturallandscapetypificationofSlovenia.Acta geographica Slovenica 55-2. DOI: https://doi.org/10.3986/AGS.1938 Plut, D. 1999: Regionalizacija Slovenije po sonaravnih kriterijih. Geografski vestnik 71. Plut, D. 2005: Teoretična in vsebinska zasnova trajnostno sonaravnega napredka. Dela 23. DOI: https://doi.org/10.4312/dela.23.59-113 Pravilnik o določanju in vodenju bonitete zemljišč. Uradni list Republike Slovenije 47/2008. Ljubljana. Richards,J.,Jia,X.2006:Remotesensingdigitalimageanalysis:anintroduction.Berlin,Heidelberg.DOI: https://doi.org/10.1007/978-3-642-30062-2 Rogerson, P. A. 2006: Statistical methods for geography: a student guide. London, Thousand Oaks, New Delhi. Romportl, D., Chuman, T. 2012: Proposal method of landscape typology in the Czech Republic. Journal of landscape ecology 5-3. DOI: https://doi.org/10.2478/v10285-012-0057-5 Runhaar,H.J.,UdodeHaes,H.A.1994:Theuseofsitefactorsasclassificationcharacteristiccsforecotopes. Ecosystem classification for environmental management. Dordrecht. DOI: https://doi.org/10.1007/ 978-94-017-1384-9_7 Špes,M.,Cigale,D.,Lampič,B.,Natek,K.,Plut,D.,Smrekar,A.2002:Študijaranljivostiokolja.Geographica Slovenica 35, 1–2. Ljubljana. TerrSetHelpSystem.ClarkLabs.Worcester,2015.Internet:https://clarklabssf.wpengine.com/wp-content/ uploads/2015/07/TerrSet-Help-System.zip (12. 12. 2018). Tirelli,T.,Pessani,D.2011:Importanceoffeatureselectionindecision-treeandartificial-neural-network ecological applications. Alburnus alburnus alborella: Apractical example. Ecological Informatics 6-5. DOI: https://doi.org/10.1016/j.ecoinf.2010.11.001 TheEuropeanLandscapeConvention.Internet:https://www.coe.int/en/web/landscape/home(10. 9. 2018). Theodoridis, S., Koutroumbas, K. 2006: Pattern recognition. San Diego. UdodeHaes,H. A.,Klijn,F. 1994:Environmentalpolicyandecosystemclassification. Ecosystemclassification for environmental management. Dordrecht. DOI: https://doi.org/10.1007/978-94-017-1384-9_1 Verbič, T. 1998: Kamnine. Geografski atlas Slovenije. Ljubljana. Warner, T. A., Campagna, D. J. 2009: Remote sensing with Idrisi® Taiga. A beginner’s guide. Hong Kong. Whittow, J. B. 2000: Dictionary of physical geography. London. Zupančič,M.,Marinček,L.,Puncer,I.,Seliškar,A.1998: Potencialnonaravnavegetacija. Geografskiatlas Slovenije. Ljubljana.