You are on page 1of 13


ISSN 1330-0520 . UDK 581.526.42 57.04(497.5)

VOL. 9

No 1



March 31, 2000

original scientific paper / izvorni znanstveni rad


Rudjer Bo{kovi} Institute, Bijeni~ka 54, 10000 Zagreb, Croatia (e-mail:; Fax: (+385 01) 4680–084)


Croatian Natural History Museum, Demetrova 1, 10000 Zagreb, Croatia Oikon Ltd., Vlade Prekrata 20, 10000 Zagreb, Croatia

Andrija Mohorovi~i} Geophysical Institute, Department of Geophysics, Faculty of Science, University of Zagreb, Horvatovac bb, 10000 Zagreb, Croatia
Antoni}, O., Bukovec, D., Kri`an, J., Marki, A. & Hati}, D.: Spatial distribution of major forest types in Croatia as a function of macroclimate. Nat. Croat., Vol. 9, No. 1., 1–13, 2000, Zagreb. The model of spatial distribution of major forest types in Croatia was developed as a function of macroclimatic variables (monthly mean temperature, monthly precipitation, monthly mean global solar irradiation and monthly potential evapotranspiration) and variables derived from digital elevation model (terrain aspect and slope). Neural networks were used as modelling tool. The model was developed within the frame of a raster geographic information system with a spatial resolution of 300 × 300 m, and it was based on a forest vegetation map (in scale of 1 : 500000) and interpolation macroclimatic models. The agreement between modelled and mapped forest types was very good, which suggests a strong correlation between macroclimate and the main forest types in Croatia and high model reliability. The model was applied to the entire area of Croatia, aiming at the construction of the potential spatial distribution of major forest types. The model could be useful for reforestation planning and for prediction of vegetation succession under assumed climatic changes. Keywords: air temperature, DEM, GIS, neural networks, potential evapotranspiration, potential vegetation map, precipitation, solar irradiation Antoni}, O., Bukovec, D., Kri`an, J., Marki, A. & Hati}, D.: Prostorna razdioba glavnih tipova {uma u Hrvatskoj kao funkcija makroklime. Nat. Croat., Vol. 9, No. 1., 1–13, 2000, Zagreb. Model prostorne razdiobe glavnih tipova {uma u Hrvatskoj razvijen je kao funkcija makroklimatskih varijabli (srednja mjese~na temperatura, mjese~na oborina, srednje mjese~no Sun~evo ozra~enje i mjese~na potencijalna evapotranspiracija) te varijabli izvedenih iz digitalnog elevacijskog modela (orijentacija i nagib terena). Neuralne mre`e kori{tene su kao oru|e za modeliranje. Model je razvijen u okviru rasterskog geografskog informacijskog sustava, uz prostornu razlu~ivost od 300 x 300 m i
Croatian Natural History Museum, Demetrova 1, Zagreb, Croatia

Model je primjenjen na cijeli hrvatski teritorij s ciljem konstruiranja potencijalne prostorne razdiobe glavnih {umskih tipova. as well as raster modelling procedures within a framework of geographic information systems (GIS). due to the fact that Croatia is a country with large macroclimatic variability. DEM. MATERIAL AND METHODS Vegetation data The study area was the entire land area of the Republic of Croatia. from a warm and dry Mediterranean to a cold and wet mountainous climate.2 Antoni}. using the mentioned model. 1987). enables spatially explicit implementation of the correlation between vegetation and climate over larger areas (see e. 1909) and examined in numerous studies (see review in BERTOVI]. karta potencijalne vegetacije. THORNTON et al. LEATHWICK. 1997). at the level of major forest types. Podudaranje izme|u modeliranih i kartiranih tipova {uma vrlo je dobro. capable of predicting vegetation type as a function of macroclimatic variables. baziran je na karti {umske vegetacije (mjerila 1:500000) i interpolacijskih makroklimatskih modela. . Sun~evo ozra~enje INTRODUCTION Macroclimatic variables are usually major environmental factors which influence the large-scale spatial variability of vegetation (WOODWARD. potencijalna evapotranspiracija. Klju~ne rije~i: temperatura zraka.. Correlation between forest types and macroclimatic variables was recognized early in Croatia (BECK-MANNAGETTA. {to upu}uje na jaku korelaciju izme|u makroklime i glavnih tipova {uma u Hrvatskoj. 1975). which are useful for the management of forest resources. Thus. te na visoku pouzdanost modela. such as reforestation planning or the prediction of vegetation succession under assumed climatic changes. Model bi mogao biti koristan u planiranju po{umljavanja. BRZEZIECKI et al. 1995. oborina. The second aim was the construction of the spatial distribution of potential vegetation. including statistics of ecologically relevant macroclimatic variables recorded at the meteorological stations situated within the particular forest type. 1995. 1901. MITCHELL. kao i u predvi|anju vegetacijske sukcesije pod utjecajem pretpostavljenih klimatskih promjena. The development and testing of such a model for Croatia. It makes it possible to develop spatially explicit models. ADAMOVI]. BERTOVI]’s important study (1975) provides general knowledge about the relation between macroclimate and major forest types in Croatia (defined according to the Braun-Blanquet approach). This correlation is relatively strong.. 1998).: Spatial distribution of major forest types in Croatia . the vegetation pattern in Croatia is very dissected. LENNON & TURNER. GIS. were the basic aims of this research. some important tasks.. The recent development of methods for the spatial interpolation of macroclimatic variables (see e. O. neuralne mre`e. Primarily due to this fact.g.. which strongly affects local climate. The use of this knowledge in the management of the natural resources is limited by the fact that it is often hardly applicable in the real space.g. 1991. Approximately half of the area belongs to the Karst region with extremely rugged relief. et al. are solved arbitrarily.

All of these types contain edaphically conditioned subtypes not examined in this paper. from Quercetalia pubescentis Br..g. The chosen map originally contains 14 main forest type classes and 43 subclasses mainly based on phytocoenoses in the sense of Braun-Blanquet (compare also TRINAJSTI] et al. edaphically conditioned pine forests. 4) class 14 (relic. which are similar to the 'seed regions' of Croatia. 1992.-Bl. 1986 or ZIMMERMANN & KIENAST. 1975 or BERTOVI] & LOVRI]. (1992). 1.-Bl. because they do not present real. Thus. based on field surveys and generalised to a scale suitable for this research. Vol. 3) . The original map was digitised and rasterised in a spatial resolution of 300 x 300 m. only main classes were used. 1939) was merged with class 10 (altimontane neutrophilic mixed forests. 1963 and Listero-Piceetum abietis Ht. 1992) were inappropriate. Genisto elatae-Quercetum roboris Ht.. 1969). Orno-Ericion Ht. This data source was chosen because it is the only available representation of the real forest vegetation in Croatia. 1938) Borh. which are mainly related to the Mediterranean region. 2000 3 The forest vegetation pattern examined in this research was based on the 1 : 500000 map provided by TRINAJSTI] et al. The spatial distribution of these types in Croatia is shown in Fig. Populus sp. 1995). Fraxinus angustifolia Vahl and Alnus glutinosa (L.. 3) class 11 (middle European montane and altimontane acidophilic coniferous forests. the final vegetation data set contains the nine major forest vegetation types listed in Tab. see e. 1999): 1) monthly mean air temperature. are not discussed in this paper. Homogyno sylvestris-Fagetum sylvaticae (Ht. 1992 and RAU[ et al. Climate data Four variables were selected as macroclimatic estimators of great importance for the spatial distribution of vegetation (see e. presented in the work of BERTOVI] & LOVRI] (1992). 2) monthly precipitation. It has to be emphasised that the vegetation classification used in TRINAJSTI] et al.) Gärtn. LANDSBERG.). Croat. 1959. 1938 and Carpino betuli-Quercetum roboris (ANI]. non-climatic environmental influences (class number follows TRINAJSTI] et al. (1931) 1932).) was merged with class 5 (sub-Pannonian lowland forests of pedunculate oak..Nat. because the total area of class 13 in Croatia is relatively small in the context of this research. Lonicero borbasianae-Pinetum mugi (Ht. For the purpose of this research. 1963. Molinio-Quercetum pubescentis [ugar 1981) was merged with class 3 (Mediterranean thermophilic deciduous forest. (1992) has an alternative. 'Abieti-Fagetum' complex. 1992). 1969) was merged with class 12 (subalpine beech forest. derived by GRA^AN et al. 9(1). RAU[. 1. TRINAJSTI]. Vaccinio-Piceion Br. The differences between those two approaches.g. BERTOVI]. (1999) from the same vegetation data source. but potential vegetation which prejudices correlation between macroclimate and vegetation. 1998): 1) class 4 (middle European flood-plain and swampy forests of Salix sp. Other available sources (e. 1958) was omitted and 5) class 13 (subalpine coniferous forests. syntaxonomy follows VUKELI] & RAU[.g. 1938) Borh. 2) class 7 (pubescent oak forest on an impermeable flysch lithological substratum.. Some modifications were made to avoid vegetation variability caused by other.

. . see text for further explanation and Tab. The last two were modelled for each weather station as a function of relevant climatic variables observed at the respective station. White area is non-forested.4 Antoni}. see also BONAN.: Spatial distribution of major forest types in Croatia . Observed spatial distribution of major forest types used in analysis (simplified from TRINAJSTI] et al.. Spatial distributions of climatic variables were averaged for the period of 1956–1995. monthly mean global solar irradiation on a horizontal surface at ground level and 4) monthly potential evapotranspiration on a horizontal surface. 1. (1992). Fig. et al. O. using very accurate interpolation models presented in ANTONI] et al. The first two climatic variables were taken directly from weather station chronicles. 1 for legend). with a spatial resolution of 300 × 300 m. 1989) for potential evapotranspiration. using the NIKOLOV & ZELLER model (1992) for global solar irradiation and the PRIESTLY & TAYLOR model (1972. (in press).

1999). while eastness was adopted in this research as a general estimator of being leeward to the major cyclones (PAND@I]. which influence local precipitation (see PENZAR. The basic set of 48 independent estimators (4 climatic variables by 12 months) is reduced in this research to 5 composite estimators (non-linear analogues of principal components) using five-layered autoassociative NN (see BISHOP. respectively (GUISAN et al. The northness is general estimator of the local thermics. spatial resolution of 300 × 300 m) and were developed using neural networks (NN). the forest type prediction model was actually driven by all 48 independent macroclimatic estimators. Quercus ilex L. which were only filtered through the autoassociative NN. Description of major forest types used in the analysis. (1992) or VUKELI] & RAU[ (1998). 9(1). and Fagus sylvatica L. The terrain aspect is a circular variable and it is consequently transformed into a northness and eastness by the cosine and sine transformation. After the last two layers were cut. The logistic function was used as the activation function. Quercus pubescens Willd. this autoassociative NN was used as input for the development of forest type prediction model.79% of total macroclimatic variability was explained. (1992) and VUKELI] & RAU[ (1998). Fagus sylvatica L. to the coastal vegetation. To describe the basic topoclimatic variability. 1959).. type map class 1 2 3 4 5 6 7 8 9 4+5 6 8 10+11 12+13 9 3+7 2 1 general physiognomic and ecological description lowland deciduous forests including flood-plains and swamps colline and submontane deciduous forests montane deciduous mesophilic forests altimontane mixed (coniferous/deciduous) forests subalpine deciduous and coniferous forests montane deciduous thermophilic forests Mediterranean deciduous forests Mediterranean mixed (evergreen/deciduous) forests Mediterranean evergreen forests main tree species Quercus robur L. the same DEM is used for the calculation of terrain aspect and slope. over the mountainous region. which have 48 neurons in the first and last layer (48 basic estimators). For the syntaxonomic aspect of classes see TRINAJSTI] et al. (1992). A map class follows the numbers of main classes on the original map of TRINAJSTI] et al.g. Using this NN architecture. Pinus halepensis Mill. 15 neurons in the second and the fourth layer and 5 neurons in the central layer. Quercus petraea Liebl. 2000 5 Tab. 1998. and Pinus mugo Turra Fagus sylvatica L. 1989). which were used as additional independent estimators. Terrain slope was included as a general estimator of soil wetness. Abies alba L.Nat. 99. Fagus sylvatica L. Consequently. For species compositions see e. RAU[ et al. We . Croat. 1995). which is modified for this purpose (see text). The order of forest types implies a natural vegetation zoning from inland low areas. 1. These models are based on data from 127 weather stations and on elevation data from a digital elevation model (DEM. Vol.

1995. 1994). Each tested architecture had an input layer with 8 independent estimators (5 composite macroclimatic estimators. 1996). The total data set used for model development was unbalanced. the second was used to check for overfitting (see e. are not suitable estimators for this research because they relate to a finer spatial scale. The first set was used for the finding of the NN parameters. The initial data set of 253447 pixels was split into three approximately even-sized sets (training. . the architecture of the final model illustrates the complexity of correlation between vegetation and climate in the real space. 2). 1998). performed in a separate control analysis. the model returns nine probabilities for each pixel. but a number of hidden layers (1–2) and its neurons (5–40) varied. O. are presented in Tab. If two or more forest types have equal frequency. a number of NN architectures was tested. verification and test set).g.g. with each pixel being finally classified into that forest type that is most frequent in ten independent classifications. A further increase in NN architecture complexity did not yield significant model improvement in a reasonable model training time. LAWRENCE. 1995 or PATTERSON. such as topographic solar irradiation (DUBAYAH & RICH. 1997). northness. The class with largest probability is used for the classification of the given pixel into the forest type. et al.e. or exposure to wind (ANTONI] & LEGOVI]. 2. Thus. All of NNs classify major forest types with a correctness between 78 % and 79 % in total. and the third contained fully independent data used for the testing of different NNs and for the evaluation of a final model.6 Antoni}.. The prediction model was derived using the feedforward NN with multilayer perceptrons (MLP) which is appropriate for classification problems (see e. soil moisture potential and snow accumulation potential (e. The NN architecture finally chosen had two hidden layers with 25 neurons. During the preliminary research. RESULTS AND DISCUSSION The NN model originally calculates the probability of incidence of each forest type for a given set of input values. Due to this fact. 1999). The use of a balanced data set. think that more precise DEM-based topoclimatic variables (see ANTONI].: Spatial distribution of major forest types in Croatia . Classification results for ten NNs. Data analysis Vegetation and climatic spatial distributions were overlaid and sampled for the final data matrix within the frame of a raster geographic information system (GIS). The logistic function was used as the activation function. did not significantly improve model reliability. ANTONI]. BROWN. the final model combines the results of these ten NNs. the selection between them is done randomly. 1996 for review). BISHOP. i. independently initialised and trained by thousand epochs. Classification correctness for particular vegetation classes varies for different NNs (Tab.. a particular forest type was represented with a different number of cases (pixels) according to its proportion in vegetation cover. However.g. eastness and terrain slope) and an output layer with 9 vegetation classes.

2 91.5 35. The scheme of spatial arrangement of adjacent forest types arising from the original map of TRINAJSTI] et al.7 7 94.1 64.0 93. Classification correctness for ten independent NNs.4 9 94.6 The final model has a total classification correctness of 79. Vol.3 46.4 92. aiming at the construction of a potential spatial distribution of nine major forest types (Fig.3 78. Types 3.7 91.2 62. forest type NN 1 NN 2 NN 3 NN 4 NN 5 NN 6 NN 7 NN 8 NN 9 NN 10 1 88.1 87.7 2 63.2 86.7 96. 1977.3 48.7 3 77.5 83.0 51.4 59.9 83.6 87.0 38. it could be assigned the value 'fair') between the observed and modelled pixels for type 5 (subalpine deciduous and coniferous forests. 1992) of 0. (1992).2 90.9 87. Moreover. Numbers of types follow Tab.2 50. All values are in percentages.7 89. Arrows indicate the existence of spatial contact between types.6 91.6 88. 1977). A lower agreement (according to LANDIS & KOCH.2 90.6 36.6 94.9 85.8 90.0 61. . 3).0 6 34.4 4 92.1 78.1 90.9 90. 3) could be explained by the hypothesis that the incidence of this type in Croatia.8 48.7 90. 2) is 96.7 24.6 89.8 61.8 % (see also Fig. 2000 7 Tab.6 37.3 64.5 78.1 90. 3 for detailed classification results).75 (see Tab.9 82. Tab.6 84. which frequently occupies the ridges of the Fig. 9(1). 5 and 6 belong to the complex of beech forests (monodominant or mixed).Nat.6 82.2 35.5 39.3 78. 2. 1.6 78. These results suggest: 1) a strong correlation between macroclimate and the spatial distribution of main forest types in Croatia and 2) high model accuracy. 4).2 30.7 30.6 90.3 62.4 35.3 81.8 90. However.9 89.9 51.5 87. 1. the spatial pattern of misclassified pixels is clearly non-random (Fig.1 39.2 65.3 91. Croat. Numbers of forest types follow Tab.4 8 52.5 27.5 77.7 79.7 55.6 78.6 82.2 77.9 41.7 58.5 80.6 78.5 94.9 92.6 90.4 54.3 25.0 27. 2. The final model was applied to the entire land area of the Republic of Croatia.3 93.1 91.8 87.2 92. 4.5 % and overall Kappa statistics (MONSERUD & LEEMANS.4 5 34.0 total 79. the total portion of correctly classified pixels together with pixels misclassified to the adjacent forest type (Fig.4 16.3 90. Consequently.7 21. which suggests the additional influence of topoclimatic factors. 3).1 40.1 91. the total agreement between observed and modelled forest types could be characterised as 'very good' (LANDIS & KOCH.9 61.

Agreements for other forest types are 'good' (types 2. forest type 1 2 3 4 5 6 7 8 9 N correct (%) Kappa 1 11056 902 429 0 0 0 80 0 0 12467 88.4 0. 2). while on the Pannonian mountains. N is a number of pixels for each forest type in the test set (84468 pixels in total). Fig.6 0. without agreement expected by chance (see MONSERUD & LEEMANS. as on some patches in the Kvarner area. 7 and 9). is probably additionally conditioned by topography (especially by terrain exposure to wind).7 0. 4).1 0. Numbers of forest types follow Tab.. has a less defined self-specific macroclimate. while nature boundaries between classes are often blurred.9 0. This assumption could be supported by the classification results for type 4 (altimontane mixed forests): in Gorski kotar. Classification matrix for the final model (see text for further explanation).41 6 0 0 463 706 117 1224 677 0 0 3187 38. Kappa statistics indicates agreement between observed and modelled map for each forest type.0 0.49 7 0 0 395 130 0 444 14936 132 34 16071 92. it almost disappears from the modelled forest cover.61 3 150 1182 15664 1797 8 14 17 0 0 18832 83. 1992).85 Fair agreement for type 6 (montane deciduous thermophilic forests) suggest that this type.g. Besides.91 8 0 0 2 0 0 0 398 1656 721 2777 59. the model could be a macroclimatic frame for more detailed local models of the . Bold values indicate correctly classified pixels. where lowland pedunculate oak forest in the valley of the Mirna River is an enclave within the Mediterranean vegetation) and 4) model error (e. 'very good' (type 4) or 'excellent' (types 1. Total unexplained variability could be ascribed to: 1) vegetation mapping errors related to the scale used (1 : 500000).8 Antoni}. 2) the use of spatially discrete vegetation classes. it could be understood as a widely spread ecotone between adjacent forest types (compare Tab. 3. 1. Consequently.69 9 0 0 0 0 0 0 26 150 2825 3001 94. 3) remaining local influences (such as in Motovun forest. where this type mainly occupies ridges. et al.5 0. although relatively widespread in Croatia.. our model predicts its incidence well. Istria. Tab. where the model unrealistically predicts montane deciduous mesophilic forest.65 4 0 0 973 10972 23 10 1 0 0 11979 91. The presented model could be suitable for general reforestation planning and for the prediction of regional vegetation succession under assumed climatic changes. highest mountains.79 5 0 0 99 988 493 10 111 0 0 1701 29. O.2 0.: Spatial distribution of major forest types in Croatia . 1 and Fig.91 2 328 8304 5821 0 0 0 0 0 0 14453 57. 3 and 8).6 0. where this type is widely spread.

BROWN. 1994. see Fig. local topoclimate. 1998. soil properties. 1998. 1998. 1994) and by other relevant environmental variables (lithological substratum. could have a crucial role in the inventory and mapping of data needed for optimal management of forest resources and protected areas in Croatia. 2 and text for further explanation). These local models..g. 3. 1998. Spatial distribution of correctly classified pixels (white) and misclassified pixels (grey – misclassifed into adjacent forest type. VAN RIJT. Non-forested area is also white. 9(1). FRANKLIN. BROWN. TAPPEINER. Croat. 1994 or MICHAELSEN et al. 1998. 1999. DE . black – misclassified out of adjacent forest type. especially supported by remote sensing (see e. These points will be the objects of future research. Vol. 1999) as well as the distribution of particular species (e. BRZEZIECKI.g. GOTTFRIED et al. LEATHWICK.Nat.g. 1995. ZIMMERMAN & KIENAST. 1996. 2000 9 Fig. ZIMMERMAN & KIENAST... spatial distribution of vegetation types (see e. human impact). 1999). GUISAN et al.

This research has been supported by the Croatian Ministry for Science and Technology and by OIKON Ltd. O. for access to weather station data. Fig. 1 for legend) for the entire territory of Republic of Croatia (modelled potential vegetation). Received February 15.10 Antoni}.: Spatial distribution of major forest types in Croatia .. 4. 2000 ... ACKNOWLEDGEMENTS The authors wish to express their gratitude to the Meteorological and Hydrological Service of Croatia (DHMZ). Spatial distribution of major forest types predicted by the model (see Tab. Zagreb. Zagreb. et al.

G. 405–419. & KOCH. – Cro. DUBAYAH. 733–748. California. O. J. 1995: Topographic solar radiation models for GIS. 34(1).. D. GUISAN.. Sci. 159–174. BERTOVI]. Rad.. – Ecol. BRZEZIECKI. Veg. G.. S. O. Veg. 31–40. BECK-MANNAGETTA. – Arctic and Alpine Research. – Ecol. 1992: Übersicht der Vegetation Kroatiens nach neueren Untersuchungen.. 1989: A computer model of the solar radiation. G.. 124. [umar. 2000 11 REFERENCES ADAMOVI]. Proceedings of the Fourteenth National Conference on Artificial Intelligence. GOTTFRIED.. – J.. @. S.. Met. – Ecol.. & LOVRI]. S. 1995: Neural Networks for Pattern Recognition. 1999: [umski sjemenski rajoni (jedinice) u Hrvatskoj. 143.. 33. GILES. BONAN. BISHOP. C.. BERTOVI]. 107–122. Geographical information systems. S. 5.. Modelling. 1995: Modeling potential impacts of climate change on the spatial distribution of zonal forest communities in Switzerland. – Acta Biol. M. A.. 257–268. J. 1986: Physiological ecology of forest production. 1997: Lessons in neural network training: overfitting may be harder than expected. 1909: – Die Vegetationsverhältnisse der Balkanländer.. 207–221. M. LAWRENCE. KIENAST. 9. Modelling 45. BROWN. PAULI. 1998: Predicting the distribution of shrub species in southern California from climate and terrain-derived variables. 9. J. 1994: Predicting vegetation types at treeline using topography and biophysical disturbance variables. S.. O. J. A.. 1977: The measurement of observer agreement for categorical data. 55–93. & GRABHERR.-P. T. C. G. 95–102.. Veg. Oxford. RAU[.. ANTONI]. Z. F. 641–656. Menlo Park. Vol. FRANKLIN. Leipzig. J. 1998: Modelling daily topographic solar radiation without site-specific hourly radiation data. D. 31. – Tuexenia 12. – J. A. 1996: Application of spatial modelling in the karst bioclimatology. – Int. 189–196. – J. C. O. Leipzig. J. ... Modelling. 1999: Estimating the direction of an unknown air pollution source using a digital elevation model and a sample of deposition. VII/2. 1998: Predicting of Vegetation Pattern at the limits of Plant Life: A New View of the Alpine-Nival Ecotone. AAAI Press.. \. 85–95.. 65–74. & WEISS. 540–545 pp. B. 29–48.. B. & TSOI. THEURILLAT. & RICH. 1998: Predicting the potential distribution of plant species in an alpine environment. Modelling. J. GUISAN. LANDSBERG. D. & WILDI. & BUKOVEC. L. LANDIS. Oxford University Press. Sci. 113... 6. J. A. MARKI. 30(3). soil moisture and soil thermal regimes in boreal forests. KRI@AN. & LEGOVI]. R. MATI]. 1975: Contribution to knowledge of relation between climate and vegetation in Croatia. J. – Ecol. 9... & SELETKOVI]. ANTONI]. 9(1). London.. P. Croat. ANTONI]. GRA^AN.. Veg. 1901: Die Vegetationsverhältnisse der illyrischen Länder. – Biometrics.Nat. L. G.. R. Sci. O. A. – J. 275–306. inst. A. & KIENAST. A. AAAI–97. – Plant Ecology. 1999: GLM versus CCA spatial modeling of plant species distribution. Sci. Academic press. H. ANTONI]. WEISS. (in press): Spatially-temporal interpolation of climatic variables over large region of complex terrain using neural networks. G. A.. KRSTINI]. F.

W. VUKELI]. & TURNER. 81–92. P. D. PRIESTLY... Modelling 61. 1999: Predictive mapping of alpine grasslands in Switzerland: Species versus community approach. Zbornik II. J.... H. O. Zagreb. TRINAJSTI].. University of Zagreb. A. J.. J. ... & MEDVEDOVI]. – In: Forests of Croatia. J. & ZELLER. Veg. E.. Sci. Wirtschaftswald ein Vergleich der Floristischen Struktur. C. Singapore. Veg... rujan 1957. \. 469–482. 149–168. Veg. P. & TAPPEINER. F. R. & LEEMANS. T.: Spatial distribution of major forest types in Croatia . MONSERUD. F. D. 1992: Urwald. J. C. Zagreb. University of Zagreb and ’Croatian forests’ p. – J. – J. et al. PENZAR. C.. 1997: Generating surfaces of daily meteorological variables over large regions of complex terrain. & WHITE. & MEDVEDOVI]. 1998: [umarska fitocenologija i {umske zajednice u Hrvatskoj. J. – In: Forests of Croatia. RAU[. W. RAU[. I. DAVIS. Modelling 62. N. R. K. J. MICHAELSEN. & TAYLOR R. Weather Rev. & BLOM. & KIENAST.. – Mont. PATTERSON. Forestry Faculty. ZIMMERMANN. 275–293. F. Forestry Faculty. Modelling 113. 1991: The derivation of climate surfaces for New Zealand.. – Ecol. VUKELI]. J. R. J. M.12 Antoni}. K. 673–686. VUKELI]. M. TRINAJSTI]. Soc. 214–251. VAN DE RIJT. 1992: Vegetation Map of Forest Associations of Croatia (in Croatian with English summary). (in Croatian with English summary). L.. Sveu~ili{te u Zagrebu. W. 79–80 + 1 map. 100. TASSER. J. TRINAJSTI]. Ecol. 505–518. S. B. 1996: Vegetation zonation in a former tidal area: A vegetation-type response model based on DCA and logistic regression using GIS. 1995: Predicting the spatial distribution of climate: Temperature in Great Britain. Cambridge. J. – Ecol. – J. 1959: Razdioba godi{njih koli~ina oborine u Gorskom kotaru. & RAU[. B. 33–78. Z. 109–132.. C. Senj. A.. N. W... WOODWARD.. F. 190. G. Naturwald. I. Zagreb. – J. Veg. University of Zagreb and ’Croatian forests’ p. 1992: Comparing global vegetation maps with the Kappa statistics. J. R. 10. U. 1989: Factor analysis of the local climatic fields within global stream (PhD Dissertation). 1987: Climate and plant distribution.. Cambridge University Press. 13–24. 1972: On the assessment of surface heat flux and evaporation using large-scale parameters. – Sauteria 6.. D.. 1992: A solar radiation algorithm for ecosystem dynamic models. TAPPEINER. NIKOLOV. Hydrol. 1996: Artificial Neural Networks.. & DUBAYAH. Prentice Hall.. SCHIMEL. E. R.o. G. THORNTON. I. R..o. Sci. RUNNING. 719–732.. 225–237. Sci.. – N. J. PAND@I]. – J. 21.. 1998: Are New Zealand’s Nothofagus species in equilibrium with their evironment? – J. 64. \.. – Ecol. 5. I.. C.. Faculty of Natural Sciences. 1992: Plant Life of Croatian Forests. N. HAZELHOFF. Sci. savjetovanja meteorologa FNRJ za planinsku meteorologiju. Anim. E. FRIEDL.. 7. LENNON. LEATHWICK. 1994: Regression Tree Analysis of satellite and terrain data to guide vegetation sampling and surveys... HMZH Zagreb. A. \. 1998: Modelling vegetation patterns using natural and anthropogenic influence factors: preliminary experience with a GIS based model applied to an Alpine area. M. Croatia. 9. 370–392. S. and their application to the bioclimatic analysis of the distribution of Kauri (Agathis australis). MITCHELL.

Kona~na arhitektura ilustrira slo`enost korelacije izme|u {umske vegetacije i makroklime u realnom prostoru. Vol. Kao izvor podataka o makroklimi kori{teni su modeli visoke pouzdanosti. U radu su prikazani i komentirani rezultati klasifikacije modelom za svaki {umski tip posebno (Tab. Taj kona~ni model klasificira {umsku vegetaciju s ukupnom to~no{}u od 79. 2 i 3) iznosi 96. model bi mogao biti prikladan makroklimatski okvir za detaljnije. 3). Prostorna vegetacijska varijabilnost neobja{njena modelom mo`e se tuma~iti pogre{kom vegetacijskog kartiranja u sitnom mjerilu. 2). Model je razvijen u okviru rasterskog geografskog informacijskog sustava. te istovremeno na visoku pouzdanost modela. kona~ni model kombinira rezultate svih deset me|usobno nezavisnih mre`a.8 %. 1 i Fig.Nat. te antropogeni utjecaji. Svi ti rezultati upu}uju na jaku korelaciju izme|u makroklime i prostorne razdiobe {umske vegetacije u Hrvatskoj. A. Kri`an. lokalne modele prostorne razdiobe vegetacijskih tipova. Tijekom modeliranja testirane su razli~ite arhitekture neuralnih mre`a. udio to~no klasificiranih piksela zajedno s pikselima klasificiranim u susjedni tip {ume (Figs. srednje mjese~no Sun~evo ozra~enje i mjese~na potencijalna evapotranspiracija) i varijabli izvedenih iz digitalnog elevacijskog modela (orijentacija i nagib terena). zna~ajke tla. Croat. mjese~na oborina. Model bi mogao biti koristan u planiranju po{umljavanja. Ukupna Kappa statistika iznosi 0. Antoni}. u prostornoj razlu~ivosti 300 × 300. Deset nezavisnih inicijalizacija odabrane neuralne mre`e rezultiralo je ukupnom to~no{}u klasifikacije izme|u 78 % i 79 % (Tab. Upotrijebljene su vi{eslojne neuralne mre`e bez povratnih veza. . Kao izvor vegetacijskih podataka kori{tena je karta realne {umske vegetacije mjerila 1:500000. Testiranje modela provedeno je na nezavisnom uzorku.5 %.75 {to se ozna~ava kao ’vrlo dobro’. Kako je to~nost klasifikacije varirala za pojedine {umske tipove (Tab. Osim toga. koja je za potrebe ovog istra`ivanja svedena na devet glavnih tipova {uma (Tab. Hati} U radu je prikazan razvoj i testiranje modela prostorne razdiobe glavnih tipova {uma u Hrvatskoj. Bukovec. Nadalje. J. izra|eni u prostornoj razlu~ivosti 300 × 300 m. topoklimatski procjenitelji. 9(1). 4). lokalnim utjecajima i pogre{kom modela. kao i posebnih biljnih vrsta. 1). D. kori{tenju diskretnih tipova koje omogu}uju kartiranje dok su prirodne granice izme|u tipova ~esto nejasne. 2). kao i u predvi|anju vegetacijske sukcesije pod utjecajem pretpostavljenih klimatskih promjena. 2000 13 S A @ E TA K Prostorna razdioba glavnih tipova {uma u Hrvatskoj kao funkcija makroklime O. Model je razvijen kao funkcija makroklimatskih varijabli (srednja mjese~na temperatura. te uz kori{tenje neuralnih mre`a kao oru|a za modeliranje. Marki & D. uz kori{tenje dodatnih okoli{nih procjenitelja kao {to su litolo{ka podloga. Model je primijenjen na cijeli hrvatski teritorij s ciljem konstruiranja potencijalne prostorne razdiobe glavnih {umskih tipova (Fig.