Modelling habitats in karst landscape by integrating remote sensing and topography data

Mateja Breg Valjavec 1 , Rok Ciglič 1 , Krištof Oštir 2 , and Daniela Ribeiro 1
  • 1 Anton Melik Geographical Institute, Research Centre of the Slovenian Academy of Sciences and Arts, Gosposka ulica 13, SI – 1000, Ljubljana, Slovenia
  • 2 Faculty of Civil and Geodetic Engineering, University of Ljubljana, Jamova cesta, 2 SI – 1000, Ljubljana, Slovenia
Mateja Breg Valjavec
  • Corresponding author
  • Anton Melik Geographical Institute, Research Centre of the Slovenian Academy of Sciences and Arts, Gosposka ulica 13, SI – 1000, Ljubljana, Slovenia
  • Email
  • Search for other articles:
  • degruyter.comGoogle Scholar
, Rok Ciglič
  • Anton Melik Geographical Institute, Research Centre of the Slovenian Academy of Sciences and Arts, Gosposka ulica 13, SI – 1000, Ljubljana, Slovenia
  • Search for other articles:
  • degruyter.comGoogle Scholar
, Krištof Oštir
  • Faculty of Civil and Geodetic Engineering, University of Ljubljana, Jamova cesta, 2 SI – 1000, Ljubljana, Slovenia
  • Search for other articles:
  • degruyter.comGoogle Scholar
and Daniela Ribeiro
  • Anton Melik Geographical Institute, Research Centre of the Slovenian Academy of Sciences and Arts, Gosposka ulica 13, SI – 1000, Ljubljana, Slovenia
  • Search for other articles:
  • degruyter.comGoogle Scholar

Abstract

Field mapping is an accurate but also time consuming method of detailed mapping of habitat types. Levels of habitat types are usually hierarchically nested at several levels. Our main research question therefore is: ‘How detailed can be modelling of habitat types with decision trees and digital data in karst landscape?’ Similar to studies in other (non-karst) environments we explored the basic properties of the habitats in Dinaric Karst study region (Classical Karst in Southwest Slovenia) and tested modelling of habitat types at three different levels of detail. To seek for the best set of predictor variables we used Rapid-Eye satellite images, airborne images and digital elevation model. We prepared more than 60 explanatory variables and divided habitat polygons into training and testing samples to validate the results. The results proved that modelling with decision trees in Dinaric Karst landscape does not result in high accuracy at high detailed levels. Due to the presence of mine fields in the large area of Dinaric Karst (e.g. in Croatia and Bosnia and Herzegovina) the field mapping in this area is difficult therefore the findings from this study can be used for further development of mapping through remote sensing.

1 Introduction

Typologies of the environment, including those presenting habitat types, have different purposes such as making inventory, assessing, monitoring, managing, planning, predicting, sampling, modelling, presenting, and analysing the environment [1,2,3,4,5,6,7]. Classifications help us to reduce the complexity of the natural world [8] and help us to understand it better. An overview of the environment and its characteristics is necessary [9].

One of the most extensive habitat monitoring scheme in Europe is arguably the Natura 2000 programme, where each European Union (EU) member state develops its own guidelines for field-based conservation status monitoring. These guidelines require detailed surveying of many ecologically relevant variables, most of these classically considered out of reach for remote sensing [10]. Therefore, detailed field mapping of habitats is still the most established and still the most accurate method for mapping, classifying and monitoring Natura 2000 habitats. This method is also in use for karst landscapes. Remote sensing is a well-established method in global land cover mapping [e.g. 3, 11,12,13]. Due to constant technological development, it must be constantly tested also for mapping of habitat types [e.g. 14,15,16]. Besides, remote sensing plays an increasing role in ecology, biodiversity, and conservation research [17].

Beginning with simple, often subjective, hand-drawn maps, habitat mapping now involves complex species distribution models using mapped predictor variables derived from remotely sensed data [18]. In addition, hyperspectral imaging enables an accurate mapping of vegetation classes up to a certain detail, but the availability and spatial resolution of even the most advanced airborne hyperspectral sensors limits the differentiation of certain fine scale vegetation characteristics, needed to assess the local conservation status of habitats in a spatially explicit manner [19]. Therefore, we can add additional datasets to improve differentiating. For example, remote sensing data and topography data are widely used in modelling of land cover [20] as well as in detailed mapping of habitats and species distribution [21]. Very-high spatial resolution airborne imaging spectrometry has shown its ability to map species composition with extremely high accuracy. This technology is still only available for dedicated, small-scale surveys, but wherever complemented with ground data it delivers maps of unprecedented thematic (and especially taxonomic) detail, highly relevant for habitat quality mapping [22].

Several authors have shown that modelling with high resolution multispectral satellite and airborne images in combination with topographic data can deliver up-to-date land-cover classifications for land resource planning, biodiversity studies, regional or global biogeochemical cycle, environmental change [23,24,25,26], and detailed habitat monitoring in remote tropical [18], desert [19] and in dangerous military or war landscapes [27]. The inclusion of topographic data for assessing vegetation diversity has already been performed in past researches [28].

An examination of above mentioned research cases shows that different modelling techniques may be used for classification of habitats and land cover types. A list of different methods for land cover classification was presented also by [29], nevertheless decision trees are one of the most frequently used methods for habitat modelling. The method has already been used for modelling of potential habitat for black poplar (Populus nigra L.) [30], modelling of heathland vegetation [19], different shrub and tree species [31, 32], forest cover [33], invasive or exotic plant species [34,35], grassland use intensities [36], habitat types [14], marine biotopes [37], plant distribution in marshes [15], abundance of taxonomic group of organisms [38], and wetland habitats [16]. Although decision trees are robust to variables (correlation, noise etc.) there are few researches that also used variable evaluation before the modelling [34].

It has been proven that the classification of satellite images is an efficient tool when determining the land cover in the Classical Karst region [11], which is a part of broader Dinaric Karst region. However, according to our knowledge there is no further research on more detailed typological level of habitat types when dealing with Dinaric Karst landscapes of South East Europe, beside that only Slovenia has established national system for habitat mapping and monitoring. Due to remaining mine fields, some parts of southern Dinaric landscapes, especially in Croatia and Bosnia and Herzegovina, are extremely dangerous for field research and economic activities [39]. Some projects of habitat mapping are already in progress and researchers have already faced with this threat (“Terrestrial Habitat Mapping of the Republic of Croatia, World bank project ID: P111205”). Therefore the use of remote sensing and geographic information system (GIS) methods is unavoidable. Modelling of habitats in Dinaric Karst and Mediterranean is in preliminary phase, few papers are focused on animal habitat modelling [40, 41] or land cover [12], but no comparable research on modelling mosaicked karst habitats has been published so far. Therefore, established methods reported above are not completely transferable to highly mosaicked and biodiversity rich karst landscape due to high land (karst landscape with dolines) and consequential habitat fragmentation, small land plots, land cover changes in short distance and underground hydrology.

Regarding this, the main objective of the paper is to analyse and compare habitat type modelling with decision trees, at different scales in Dinaric Karst landscape. The main research question is: »How detailed can modelling of habitat types be?«. Our aim is to explore the possibility and accuracy of mapping habitats in Dinaric Karst by using multispectral and topographic Earth observation data combined with field data. We used two established typology schemes: detailed Classification of Slovenian habitat types – HTS (si. habitatni tipi Slovenije) and general Land cover typology by Food and Agricultural Organisation of United Nations Land Cover Classification Scheme – FAOLCCS. We made a translation from more detailed habitat typology to more general land cover typology and created series of typologies at three hierarchical levels. Then we modelled typology at each level.

2 Study area and data

2.1 Dinaric Karst study area and habitat characteristics

Dinaric Karst landscapes are extending over two EU member countries (Slovenia and Croatia) and non-member countries of Western Balkans that were included in past war conflicts on Balkan in 1990s. Consequently, many unknown mine fields unable safe and professional field mapping of habitats that should be in the near future included in EU Natura 2000 protected area system. This is the largest continuous karst landscape in Europe [42].

We explored the basic properties of the existent karst habitats by studying Slovene Dinaric Karst (Figure 1). The study area is spreading over a 24 km2 of Slovenian region Kras. The area is characterized by its karst phenomena (dolines, caves, and underground hydrology) which contribute to the formation of rough stony terrain, high biodiversity, high land cover diversity and habitat fragmentation [11]. Karst habitats are characteristic for landscapes that are formed in carbonate rocks. The absence of surface water due to predominant groundwater discharge, the presence of specific concave karst landforms (dolines) and distribution of fertile soils in dolines caused traditional land use patterns where cultivated dolines and grasslands prevail.

Figure 1
Figure 1

Delimitation of the Dinaric Karst [43] and the location of the study area.

Citation: Open Geosciences 10, 1; 10.1515/geo-2018-0011

The study area is sparsely populated and overgrowing process is intensive, which makes it harder to distinguish between bushy and forested habitat types in different succession stages. In only 250 years an almost treeless stony grassland landscape was converted to a forest-dominated landscape [44].

In the past, the dry stony pastures with low sedge and rock knapweed (Carici humilis-Centaureetum rupestris) were formed on shallow soils due to grazing.

In slightly deeper soils with less pronounced stoniness and with more humus, dry and semi-dry grasslands have been formed through mowing and been used as meadows and pastures alternatively. Plant communities of the alpine oatgrass and villous viper’s grass (Danthonio- Scorzoneretum villosae) were formed in this area, and are considered as the most diverse grassland communities. These grasslands are being overgrown in many places, however they are still better preserved than pastures. Xeromesophile meadows where tall oatgrass (Arrhenatherum elatius) and hay meadows are common on sites with deeper soils.

In valleys with deepest and most fertile soils wet meadows with tall oatgrass can be found, which are regularly fertilized and mowed three times a year. This association is mainly formed on the surfaces that have been arable land in the past. Arable lands are nowadays declining. Around the villages there are still vegetable gardens and fields. In the villages, there are still common tall-tree orchards, while vineyards are less present. The development of human infrastructure has changed a large area of these cultural landscapes. The highway has fragmented the landscape. The habitats in its immediate vicinity have been changed, and are dominated by ruderal communities. Unrestrained expansion of craft zones around settlements [45] has destroyed the natural habitats in the bottom of the cultivated dolines by infilling dolines with waste, being this one of the major environmental issues within the study area.

2.2 Source data and derived variables

The study was based on several vector and raster data sources (Table 1). We used habitat type maps produced by field mapping during vegetation period in 2011. Colour aerial images in the scale of 1:2,000 and 1:4,500 were used as a base map for field mapping [45]. Mapping was based on the HTS classification which was adapted from “A classification of Palaearctic habitats” (catalogue of habitat types organized according to CORINE methodology, intending to increase knowledge of the diversity of Europe’s habitats [46]).

Table 1

List of source and derived derivative data layers.

Source data layer (resolution)Layer
Habitat types defined by Kaligarič et al. [45]HTS classification detailed level with 51 habitat types [45]
(polygon vector data)HTS classification level with 20 habitat types [45]
HTS classification/FAO comparative level with 7 land cover types
FAO LCCS level with 2 land cover types
Digital elevation model (5 m)Elevation (meters)
Difference between DEM and smoothed DEM (low filter)
Difference between DEM and smoothed DEM (focal, rectangle, 9)
Slope (degrees)
Aspect categories (8 categories)
Compound topographic index [48,49,50]
Heat load index [50, 51]
Integrated moisture index [50, 52]
Site exposure index [50, 53]
RapidEye images (6.5 m)RapidEye April, bands 1–5
RapidEye August, bands 1–5
RapidEye April NDVI
RapidEye April Red Edge NDVI
RapidEye August NDVI
RapidEye August Red Edge NDVI
NDVI Difference between August and April
Red Edge NDVI Difference between August and April
Aerial images (0.5 m)Infrared aerial image, RGB bands
Aerial image, RGB bands

We collected raster data with high spatial resolution: 6.5 m RapidEye satellite images (provided by Blackbridge, now Planet), 0.5 m aerial images and 5 m digital elevation model (both provided by The Surveying and Mapping Authority of the Republic of Slovenia). Digital elevation model vertical accuracy is between 1 (open landscape) and 3 m (forested area). They were used for the derivation of different explanatory variables – different vegetation and relief indexes, such as site exposure index (see Table 1 for complete list).

Aerial images and digital elevation model were already prepared for the analysis. However, RapidEye images needed additional processing (geometric and radiometric correction) before their usage. The study area was covered by two RapidEye images, taken in spring and summer of two consecutive years (2011-04-20 and 2012-08-18). The processing of satellite images has been performed with STORM–a fully automatic image processing chain developed by our group [47]. The images were first geometrically corrected by automatic extraction of ground control points (GCPs) and matching images onto reference roads, followed by sensor modelling with orthorectification. In the next step the images were radiometrically corrected – atmospheric corrections have been applied. We used ATCOR 2 (Atmospheric & Topographic Correction, version 2) to eliminate the effects of the atmosphere such as scattering and haze, and to define the extent of clouds and cloud shades. Several products have been calculated from the images once the pre-processing was completed. We computed the normalized differential vegetation index (NDVI)and the Red Edge NDVI (normalized difference of near infrared (NIR) and red edge bands) in order to get information on vegetation health and level of greenness. Near infrared band was also determined as very useful by previous research [26].

All of the explanatory variables were available in raster format with different resolutions varying from 0.5 to 6.5 m. Data on habitat types was the only vector layer (the dependent variable). Each patch was represented by individual polygon and encompassed at least few raster cells. Therefore we assigned information on mean and standard deviation value of specific part of each raster layer to each patch (polygon of habitat type). This resulted in a large database with 61 independent variables (30 different variables with 2 different statistics – mean and standard deviation). Aspect was the only variable with one statistical value (majority). In the case of aspect, the numeric values for mean and standard deviation would be misleading (1° is similar to 360°, but has completely different value). Therefore, we divided aspect into 8 categories (N, NW ...) and calculate the majority class for each polygon. In this way we obtained the information about the prevailing aspect direction for each unit. The polygon area was also a variable. Table 3 shows a complete list of variables.

3 Determination of classification levels and decision trees methodology

3.1 Translation from habitat to land cover classification scheme

While the conversion from land cover classification to habitat classification or vice-versa can be of great utility, differences in definitions and criteria have so far limited the establishment of a unified approach [54]. In this chapter we explain the translation from detailed HTS habitat typology to more general FAO-LCCS land cover typology for the study area. Land cover data at the European and global levels (e.g. Corine Land cover and FAO land cover maps) are the closest approximation of habitat type maps for Dinaric Karst region.

HTS habitat classification was adapted to the diversity of habitats existent in Slovenia [55]. Habitats mapped in the field were divided into 51 types and then grouped into 20 types.

FAO developed a Land Cover Classification System (LCCS) [56]. It is based on a set of diagnostic criteria for a parametric classification approach. The system has two stages: the dichotomous phase (where a dichotomous key identifies 8 major land cover categories) and the modular hierarchical phase (which is open ended and where a set of environmental classifiers allows the definition of more detailed land cover classes) [57]. FAO-LCCS was considered by Tomaselli et al. [54] as an appropriate classification scheme for long-term monitoring of the conservation status of habitats.

The translation from HTS habitat types to FAO-LCCS land cover classes generated comparative level with 7 land cover types. Both original level and comparative level are presented in Table 2A.

Table 2A

The translation from HTS habitat types to FAO-LCCS at the case study area.

DETAILED HABITAT LEVELCOMPARATIVE LEVEL (FROM HABITAT TO LAND COVER TYPES)GENERAL LAND COVER TYPOLOGY LEVEL
HTStranslation from HTS to FAO-LCCSFAO-LCCS
Habitat nameHabitatLand cover typologyComparativeLand cover codeLand cover type
codecode
Arable land82Agriculture1.2.1AllCultivated and managed terrestrial areas
Humid grassland and tall herb communities37Wetland1.1.2A12Natural and semi-natural vegetation
Water-fringe vegetation53
Steppes and dry calcareous grasslands34Natural dry grassland1.1.3
Temperate heath and scrub31Heathland1.1.31
Sclerophyllous scrub32
Mesophile grasslands38Artlficlal/harvested1.2.4AllCultivated and managed
Improved grasslands81grasslandsterrestrial areas
Urban parks and large gardens85Artificial trees1.2.5
Lines of trees, hedgerows, small anthropogenic84and grasslands
woodlands, mosaic landscapes
Orchards, olive groves, tree plantations83
Broad-leaved deciduous forests41Deciduous forest1.1.6A12Natural and semi-natural vegetation
Inland cliffs and exposed rocks62Natural bare land2.1.7B16Bare areas
Structures from stacked stones.84.S6
Unpaved roads and paths86.S72Nonvegetated2.2.7B15Artificial surfaces and
Fallow or recently abandoned arable land87artificial landassociated areas
Asphalt roads86.S712Urban2.2.8
Built-up areas (cities, villages, industrial areas)86
Villages, suburbs and individual buildings86.2
Surface standing waters22Water bodies2.9B27/B28Waterbodies

We limited our research to vegetated areas. Therefore, we selected suitable (vegetated) types at each level. All the categories/types used in the analysis are separately listed in Table 2B for each level (detailed habitat level, comparative level, and general land cover typology).

Table 2B

List of habitat types/land cover types with number of polygons and area covered in the study area.

LevelTypeArea (m2)Share (%)Number of polygonsShare (%)
Level B (HTS)Lines of trees, hedgerows, small anthropogenic woodlands, mosaic landscapes745364.68.782117.8
Improved grasslands94913.41.1631.4
Broad-leaved deciduous forests8901.00.190.2
Urban parks and large gardens66272.40.81443.1
Mesophile grasslands1116685.113.161913.5
Water-fringe vegetation6047.30.11<0.1
Humid grassland and tall herb communities3058.7<0.150.1
Steppes and dry calcareous grasslands1251858.414.677216.8
Arable land78439.00.9912.0
Temperate heath and scrub4985416.658.3178738.8
Orchards, olive groves, tree plantations191185.52.22896.3
Sclerophyllous scrub360.1<0.11<0.1
Total8548502.1100.04602100.0
Level C (HTS/FAOLCCS)Agriculture78439.00.9912.0
Artificial trees and grassland1002822.411.7125427.2
Artificial/harvested grassland1211598.514.268214.8
Deciduous forest8901.00.190.2
Heathland4985776.858.3178838.9
Natural dry grassland1251858.414.677216.8
Wetland (bogs/mires)9106.00.160.1
Total8548502.1100.04602100.0
Level D (FAOLCCS)Cultivated and managed terrestrial areas2292860.026.8202744.0
Natural and semi-natural vegetation6255642.173.2257556.0
Total8548502.1100.04602100.0

It is evident that the habitat typology of HTS arises from the need for evaluation and conservation, while in FAO-LCCS categorization of land cover the anthropogenic aspect is at the forefront. As an example of discrepancies we selected four habitat types on the level of HTS. Urban parks and large gardens (type 85), Lines of trees, hedgerows, small anthropogenic woodlands, mosaic landscapes (type 84), Orchards, olive groves, tree plantations (type 83), Improved grasslands (type 81) are included in the habitat division of the same class – agricultural and cultural landscape (types starting with 8). On the other hand, Mesophile grasslands (type 38) are placed in a completely different type – scrub and grasslands (types starting with 3). In terms of land cover the mentioned type 81 is much closer to the type 38. Hence it is very difficult to distinguish between these types using remote sensing.

According to this and in order to harmonize with the FAO-LCCS we placed these two types, on the comparative level (HTS/FAO-LCCS), into one category Artificial/harvested grasslands (type 1.2.4), the remaining three types (83, 84 and 85) were placed in Artificial trees and dry grasslands (type 1.2.5). Similarly, we included the Humid grassland and tall herb communities (type 37) and the Water-fringe vegetation (type 53), which in HTS habitat typology are classified in two different types, in the same type – Wetland (type 1.1.2). For all the other categories the grouping in the transition from one to another categorization was less demanding.

The problematic category division resulting in the great variability of land use is one of the main characteristics when classifying satellite imagery of the Karst region [11]. The problem, which is a result of the fragmented land division, different land use types and rapid overgrowing of the Karst region, is especially evident, because meadows can be detected through thin bushes that create a similar spectral signature. Kokalj and Oštir [11] pointed out the problem in classification and distinguishing between field, bushes or meadows.

3.2 Supervised modelling with decision tree method

A decision tree is a machine learning method. The main purpose of modelling with decision tree is to clearly describe the relationship between dependent or objective variable (variable we predict) and one or more independent explanatory variables. Explanatory variables are selected at each modelling level according to their information value or impurity measure (e.g. Gini coefficient) and then used to define subsets. A decision tree consists of internal (parent) nodes – attributes, edges – corresponding to subsets of attribute values, and terminal nodes (child nodes, leaves) – class labels. We can classify certain sample simply by moving from the top of the tree to final terminal node according to rules. With this method we can explain existing data and predict future values [58, 59]. Decision trees do not have specific requirements regarding probability distribution of variables. They can handle nominal (e.g. aspect class) and numeric (e.g. slope) variables, and are not influenced by redundant variables or noise. They do not demand high computation resources and results are easy to interpret. High accuracy can be achieved by using different combination of trees [60].

In this research we used decision trees with algorithm CRT (Classification and Regression Trees; [61]) with Gini index as an impurity measure [59]. There are many possible settings (e.g. minimum number of units in nodes, number of levels, type of impurity measure, inclusion of pruning), thus we were forced into modelling with trial and error. 90% of polygons were used for modelling, and 10% were used for validation.

3.3 The selection of explanatory variables

Identifying the most useful variables for modelling landcover classes is a critical step [62]. If irrelevant data is removed from analysis, one can reduce costs and improve the implementation and understanding of models [63,64]. However the selection of variables usually receives less attention than the choice of methods [65]. Decision trees automatically select the most informative variable at given node and thus are not affected by unnecessary less informative variables. Despite that fact we would like to maximize modelling success and to remove variables that are not prominent. Thus all data layers were analysed according to their usefulness for modelling habitats/land cover types. We assessed them based on the:

  1. relationship between each variable and habitat types,
  2. correlation between variables.

To define a numeric objective relationship between each explanatory variable and habitat types classification we calculated information gain and gain ratio [58, 59, 66]. Information gain (InfoGain(A)) is defined as the amount of information, obtained from the explanatory variable, for determining the type of objective variable [59]. Gain-ratio is defined as information gain, normalized with the variable entropy [58, 59]. Both methods require nominal variables, thus our numeric variables were discretized (for details of discretization used in Weka software see [58, 67] and [68]).

Pearson’s and Spearman’s coefficients were used to calculate correlation between the most prominent explanatory variables.

3.4 Selection of polygons

We examined the size of the polygons. Mapped habitat types (field mapping data) sized between 100 m2 and 100,000 m2 were included in the modelling. Polygons larger than 100,000 m2 are mainly forested habitats, which have not been detailed mapped and therefore are highly heterogeneous. Polygons less than 100 m2 are mostly very small areas of habitat types, which encompass only few cells and thus cannot represent individual habitat type.

3.5 Complete methodology overview

After defining proper classification levels, selecting suitable modelling technique and evaluating appropriate explanatory variables, we modelled habitat types (objective variable) with several explanatory variables. The whole process is schematically presented in Figure 2.

Figure 2
Figure 2

Complete methodology of the habitat modelling. Different raster data layers were derived on the basis of DEM, Rapideye images, orthophotos, and infrared orthophotos. After the variables evaluation 9 data variables were selected, which were the explanatory (independent) variables. Original habitat types classification (with 51 types; level A) was hierarchically reclassified into more generalized types (levels B, C and D). Models of habitat types were produced for levels B, C, and D on the basis of training sets. Validation set was used to evaluate modelling results. At the end comparison between the levels was performed.

Citation: Open Geosciences 10, 1; 10.1515/geo-2018-0011

4 Results and discussion

4.1 Variable selection

We calculated information gain and gain ratio for each explanatory variable at each of three levels. Then we calculated indexed values of the information gain and gain ratio for each variable according to the average value of all variables at given level. Finally, we derived general average value for each explanatory variable (Table 3).

Table 3

Ranking of explanatory variables according to their information values.

VariableAverage valueVariableAverage value
mean RapidEye band 5 (April)3.294mean NDVI (August; bands 5, 3)0.804
mean NDVI (April; bands 5, 4)2.727mean RapidEye band 2 (April)0.772
mean difference between August and April2.503mean NDVI (August; bands 5, 4)0.673
NDVI (bands 5, 4)
mean difference between August and April2.471std. deviation of RapidEye band 3 (April)0.665
NDVI (bands 5, 3)
mean NDVI (April; bands 5, 3)2.211std. deviation of difference between original and smooth relief 20.651
mean infrared digital orthophoto band 11.861std. deviation of integrated moisture index0.645
mean infrared digital orthophoto band 31.768mean RapidEye band 1 (April)0.617
std. deviation of NDVI (April; bands 5, 4)1.613mean heat load index0.617
mean digital orthophoto band 11.514std. deviation of compound topographic index0.611
mean altitude1.509mean integrated moisture index0.607
std. deviaton of digital orthophoto band 21.503std. deviation of difference between August and April NDVI (bands 5, 3)0.582
std. deviaton of infrared digital orthophoto1.498std. deviation of RapidEye band 1 (April)0.575
band 1
mean digital orthophoto band 21.487mean difference between original and smooth relief 20.544
mean infrared digital orthophoto band 21.412mean site exposure index0.515
std. deviation of NDVI (April; bands 5, 3)1.410std. deviation of slope0.509
mean digital orthophoto band 31.379std. deviation of difference between August and April NDVI (bands 5, 4)0.484
mean RapidEye band 4 (August)1.365std. deviation of heat load index0.482
std. deviaton of infrared digital orthophoto1.338std. deviation of RapidEye band 1 (August)0.479
band 3
std. deviaton of digital orthophoto band 31.266std. deviation of RapidEye band 2 (April)0.452
mean RapidEye band 2 (August)1.258mean Rapideye band 5 (August)0.446
mean RapidEye band 1 (August)1.221std. deviation of RapidEye band 4 (April)0.442
mean RapidEye band 3 (August)1.162std. deviation of site exposure index0.436
std. deviaton of infrared digital orthophoto1.148std. deviation of RapidEye band 3 (August)0.390
band 2
mean slope1.102std. deviation of RapidEye band 2 (August)0.381
std.deviation of altitude1.077std. deviation of RapidEye band 4 (August)0.373
mean RapidEye band 4 (April)1.015mean difference between original and smooth relief 10.363
std. deviaton of digital orthophoto band 10.987std. deviation of difference between original and smooth relief 10.289
std. deviation of RapidEye band 5 (April)0.951aspect (majority class)0.285
mean compound topographic index0.922std. deviation of RapidEye band 5 (August)0.224
mean RapidEye band 3 (April)0.907std. deviation of NDVI (August; bands 5, 4)0.160
area0.901std. deviation of NDVI (August; bands 5, 3)0.148

On the basis of the highest values of information gain and gain ratio we can conclude that mean RapidEye band 5 (April), mean NDVI (April; bands 5, 4), and mean difference between August and April NDVI (bands 5, 4) are the most useful explanatory variables for classification of habitats and land cover. Relief indexes (e.g. heat load index and integrated moisture index) are less useful in our case. Among these the most prominent variable is mean altitude, and surprisingly aspect is one of the least important for habitat differentiation in karst landscape where a gently sloping hillside is prevailing.

In regard to the results of correlation analysis, both tests, Pearson and Spearman coefficient, gave similar results and we were able to drop large number of variables that were marked as redundant. After the complete evaluation of variables, we selected a group of variables with low correlation.

The assessment of variables resulted in the selection of 9 explanatory variables:

  • mean altitude,
  • mean slope,
  • mean Site exposure index,
  • mean RapidEye band 2 (green),
  • mean RapidEye band 5 (near infra-red),
  • mean Red Edge NDVI in August,
  • mean Red Edge NDVI difference (August - April),
  • mean aerial infrared color image (band 1), and
  • mean visual color aerial image band 1 (red).

There were 4 pairs with moderate correlation (>0.5 or <–0.5), however the rest of 36 pairs of variables had low correlation (Table 4). Nevertheless, mean RapidEye band 5 (April), mean NDVI (August; bands 5, 4), and mean difference between August and April NDVI (bands 5, 4) were kept in the database due to their high informative value. As some authors [34] use collinear variables when modelling with decision trees and the method can handle correlated variables we assume that the use of low and moderate correlated explanatory variables in our study is not problematic.

Table 4

Correlation between selected explanatory variables. The values show the Pearson’s correlation coeflcient. Asterisks refer to significance level: + - not significant.

mean altitudemean slopemean site exposure indexmean Rapideye band 2 (April)mean Rapideye band 5 (April)mean NDVI (August; bands 5,4)mean difference between August and April NDVI (bands 5, 4)mean infrared digital orthophoto band 1mean digital orthophoto band 1
mean altitude10,119***0,171***-0,165***-0,117***-0,035*0,205***-0,291***-0,211***
mean slope10,257***-0,023+-0,004+-0,002+0,159***-0,133***-0,156***
mean site exposure index10,129***0,261***0,123***-0,051**0,035*0,083***
mean Rapideye band 2 (April)10,034*-0,457***-0,070***0,278***0,492***
mean Rapideye band 5 (April)10,849***-0,665***0,459***0,240***
mean NDVI (August; bands 5, 4)1-0,582***0,268***-0,034*
mean difference between August and1-0,327***-0,645***
April NDVI (bands 5, 4)
mean infrared digital orthophoto band 110,377***

4.2 Accuracy assessment

4.2.1 Modelling land cover at general level (FAO-LCCS; level D)

At the general level of classification of land cover (FAO-LCCS level) we modelled two types: Natural and seminatural vegetation (A12) and Cultivated and managed terrestrial areas (A11).

In the differentiation between these two types producer’s and user’s accuracies were between 75 an almost 90% (Tables 5, 6, and 7). The agricultural landscape in the Dinaric Karst is being overgrown by natural vegetation. Therefore, it is difficult to distinguish between natural or anthropogenic grassland, which is overgrowing abandoned fields. This partly explains that the success of predictive models for Cultivated and managed terrestrial areas are always lower than the Natural and semi-natural vegetation.

Table 5

Producer’s and user’s accuracies for general land cover level (all polygons).

Land cover typeProducer’s accuracyUser’s accuracy
Cultivated and managed terrestrial areas80.9%83.5%
Natural and semi-natural vegetation87.4%85.3%

Table 6

Producer’s accuracy for general land cover level (training and test polygons separately).

Land cover typeTraining sampleTest sample
Cultivated and managed terrestrial areas81.8%73.3%
Natural and semi-natural vegetation88.1%81.2%

Table 7

User’s accuracy for general land cover level (training and test polygons separately).

Land cover typeTraining sampleTest sample
Cultivated and managed terrestrial areas84.4%75.5%
Natural and semi-natural vegetation86.0%79.3%

4.2.2 Modelling at comparative land cover level (level C)

At the comparative level we modelled the land cover in eight categories (Tables 8, 9, and 10). With extremely low achievement the categories Deciduous Forest and Wetland stand out. This is due to the reason that the two categories together are represented by only 15 polygons from a total of approximately 4,600 polygons, which are included in the modelling. From the remaining categories, the categories Agriculture and Natural dry grassland present the lowest success rate. The category Agriculture is often modelled as Artificial trees and grassland, Artificial/harvested grassland, Heathland, and Natural dry grassland. Natural dry grassland is frequently modelled as Heathland, Artificial trees and grassland, and Artificial/harvested grassland. The distinction between Artificial trees and grassland, and Artificial/harvested grassland is hard, especially, if the snapshot is made during the period when the heights of the grass are approximately the same. Moreover, the outstanding value of individual trees of the type Artificial trees and grassland loose in average a much larger representation from the surrounding pasture (grassland).

Table 8

Producer’s and user’s accuracies for comparative land cover level (all polygons).

Basic land cover level (FAO-LCCS, level D)Comparative land cover (level C)Producer’s accuracyUser’s accuracy
Cultivated and managed terrestrial areasAgriculture29.7%60.0%
Artificial trees and grassland71.3%73.6%
Artificial/harvested grassland67.7%67.2%
Natural and semi-natural vegetationDeciduous forest0.0%0.0%
Heathland82.9%70.8%
Natural dry grassland45.9%62.9%
Wetland (bogs/mires)0.0%0.0%

Table 9

Producer’s accuracy for Comparative land cover level (training and test polygons separately).

Basic land cover level (FAO-LCCS, level D)Comparative land cover (level C)Training sampleTest sample
Cultivated and managed terrestrial areasAgriculture31.3%12.5%
Artificial trees and grassland72.9%56.7%
Artificial/harvested grassland68.5%60.6%
Natural and semi-natural vegetationDeciduous forest0.0%0.0%
Heathland83.6%75.9%
Natural dry grassland47.7%27.5%
Wetland (bogs/mires)0.0%0.0%

Table 10

Use’s accuracy for Comparative land cover level (training and test polygons separately).

Basic land cover level (FAO-LCCS, level D)Comparative land cover (level C)Training sampleTest sample
Cultivated and managed terrestrial areasAgriculture60.5%50.0%
Artificial trees and grassland75.6%56.7%
Artificial/harvested grassland68.6%55.6%
Natural and semi-natural vegetationDeciduous forest0.0%0.0%
Heathland71.3%66.5%
Natural dry grassland64.7%42.2%
Wetland (bogs/mires)0.0%0.0%

4.2.3 Modelling at HTS habitats level (level B)

At this level we modelled 12 habitat types.

The classification reached its maximum performance with the type Temperate heath and scrub, where 86.7% from training polygons and 70.2% from the testing polygons were placed in the correct habitat type (Tables 11, 12 and 13). This habitat type includes mixed grasslands and shrubs, therefore it was expected that the maximum of incorrectly classified polygons were indexed in the similar types, which represent grasslands or trees or the mixture of both. A higher modelling performance was achieved in the classification of Mesophile grasslands (67.9% for training sample and 57.0% for the test sample), which refer to anthropogenic grasslands. Higher modelling performance was also obtained for the classification of Lines of trees,hedgerows, small anthropogenic woodlands, mosaic landscapes (68.1 % for the training sample and 70.5% for the testing sample). The modelling results show a mixture between the above mentioned land cover types.

Table 11

Producer’s and user’s accuracies for HTS habitat level (all polygons).

Producer’s accuracyUser’s accuracy
Lines of trees, hedgerows, small anthropogenic wood-lands, mosaic landscapes68.1%70.5%
Improved grasslands23.8%65.2%
Broad-leaved deciduous forests0.0%0.0%
Urban parks and large gardens33.3%55.2%
Mesophile grasslands67.9%57.0%
Water-fringe vegetation0.0%0.0%
Humid grassland and tall herb communities0.0%0.0%
Steppes and dry calcareous grasslands45.7%67.4%
Arable land24.2%43.1%
Temperate heath and scrub86.7%70.2%
Orchards, olive groves, tree plantations32.2%52.0%
Sclerophyllous scrub0.0%0.0%

Table 12

Producer’s accuracy for HTS habitat level (training and test polygons separately).

Training sampleTest sample
Lines of trees, hedgerows, small anthropogenic wood-lands, mosaic landscapes68.7%62.7%
Improved grasslands23.2%28.6%
Broad-leaved deciduous forests0.0%0.0%
Urban parks and large gardens35.8%0.0%
Mesophile grasslands69.0%52.4%
Water-fringe vegetation0.0%0.0%
Humid grassland and tall herb communities0.0%0.0%
Steppes and dry calcareous grasslands46.8%35.2%
Arable land24.1%25.0%
Temperate heath and scrub87.7%79.8%
Orchards, olive groves, tree plantations33.5%22.9%
Sclerophyllous scrub0.0%0.0%

Table 13

User’s accuracy for HTS habitat level (training and test polygons separately).

Training sampleTest sample
Lines of trees, hedgerows, small anthropogenic woodlands, mosaic landscapes70.8%67.5%
Improved grasslands68.4%50.0%
Broad-leaved deciduous forests0.0%0.0%
Urban parks and large gardens57.1%0.0%
Mesophile grasslands59.2%33.8%
Water-fringe vegetation0.0%0.0%
Humid grassland and tall herb communities0.0%0.0%
Steppes and dry calcareous grasslands69.1%51.0%
Arable land45.7%20.0%
Temperate heath and scrub70.3%69.7%
Orchards, olive groves, tree plantations53.5%40.0%
Sclerophyllous scrub0.0%0.0%

From the observation of tables 11,12, and 13 it is evident that some habitats, represented by a very small number of polygons (from 1 to 9 polygons), cannot be adequately modelled. These habitat types, such as wet habitat types, are rare in karst regions.

As we did not determine the importance of the land cover categories, those categories represented by a small number of patches are automatically deprived. Since there are several categories highly represented and these were at great extent correctly classified, the overall assessment of the classification has higher accuracy rate.

4.3 Comparison between levels B, C and D

Results from the comparison of the classification accuracies at different levels revealed that model accuracy decreases with the number of habitat type categories (Table 14). This is not surprising, and confirms previous findings [13]. The result was expected as the more detailed the classification is the more similar the categories are.

Table 14

Classification accuracies at all levels.

LevelNumber of habitat type categoriesOverall accuracy (%)Training accuracy (%)Test accuracy (%)
HTS (level B)2063.467.259.9
Comparative land cover (level C)769.271.159.3
FAO-LCCS (level D)282.685.377.7

The comparison of the classification success at different scales for each polygon (Figure 35) shows that there are 51.6% of polygons correctly classified at each level. Regarding to the overall classification accuracies at each level (between 63.4% and 82.6%) this percentage is relatively high. There is a relatively high number of polygons which are falsely classified at least at two levels (48.4%). 7.1% of polygons are falsely classified at all three levels.

Figure 3
Figure 3

Classification success of modelling land cover at general level – FAO-LCCS. Only patches included in the analysis are presented.

Citation: Open Geosciences 10, 1; 10.1515/geo-2018-0011

Figure 4
Figure 4

Classification success of modelling at comparative land cover level. Only patches included in the analysis are presented.

Citation: Open Geosciences 10, 1; 10.1515/geo-2018-0011

Figure 5
Figure 5

Classification success of modelling at HTS habitats level. Only patches included in the analysis are presented.

Citation: Open Geosciences 10, 1; 10.1515/geo-2018-0011

The methods and data used in our research cannot assure highly accurate mapping of habitats in Dinaric Karst. It is possible to distinguish relatively well habitat types at very general level, where similar accuracies were achieved also in other environments with similar number (2 or 3) of categories [e.g. 19,31,32,33], but the success of correct classification at more detailed level does not give confident results.

The overgrowing of landscape is one of the major problems of the study area, resulting in the loss of the mosaic of habitats. Overgrowing also diminishes differences between certain types. Therefore, it is hard to define clear delineations even in the field.

Another reason for low accuracies is also the spatial resolution of data. Hofer [28] reported that topographic (relief) variability increases plant variability and Rapinel et al. [16] achieved better modelling results with lidar data. Therefore, we believe that higher resolution data should be used and tested also in karst areas.

Since this was one of the rare attempts of using remote sensing data in Dinaric Karst region, the research results can be presented as a basis for testing with other datasets (hyperspectral airborne images, lidar data) or in other countries within the region.

5 Conclusions

We can conclude that modelling of habitat and land cover types at several hierarchical levels for the study area on the basis of digital elevation model (5 m resolution), satellite images RapidEye (6.5 m resolution), and aerial images (0.5 resolution) did not achieve high accuracies. The highest accuracy (80%) was achieved at the general level with only two habitat types. At lower classification levels the accuracies were less than 70%. As expected, it is harder to model classifications with numerous types. Namely, the landscape is very diverse and larger number of habitat types is more difficult to model.

Producer’s and user’s accuracies for certain types rarely exceed 80%, as a result of similarity of habitat types. With the overgrowing of cultivated areas the differences between types are becoming even smaller. Therefore, the need to differentiate new types of habitats and land cover inside vegetated areas will arise.

In our case relief data with resolution of 5 m does not contribute to the modelling process and can be eliminated in further models of karst areas, especially when dealing with karst plateaus. Further improvements regarding relief data, could be achieved by using very high resolution laser scanning data, especially in highly karstified landscapes. Additionally, laser scanning would also provide data on vegetation height. With the increasing coverage of laser scanning data from national-scale surveys we expect lidar-based habitat quality assessment to allow a major step forward in regional-scale conservation.

Acknowledgement

The authors acknowledge the financial support from the Slovenian Research Agency (research core funding No. P6-0101, L1-7542, L6-6852, J2-6777, and P2-0227).

References

  • [1]

    Bailey R.G., Ecosystem geography. Springer, New York, 1996

  • [2]

    Bunce R.G.H, Barr C.J., Clarke R.T., Howard D.C., Lane A.M.J., Land classfiication for strategic ecological survey. J. Environ Manage., 1996, 47, 1, 37–60.

    • Crossref
    • Export Citation
  • [3]

    Varga K., Szabó S., Szabó G., Dévai G., Tóthmérész B., Improved land cover mapping using aerial photographs and satellite images. Open Geosci., 2014, 7, 1.

    • Crossref
    • Export Citation
  • [4]

    Bastian O., Landscape classification in Saxony (Germany) – a tool for holistic regional planning. Landsc. Urban Plan., 2000, 50, 1-3, 145–155,

    • Crossref
    • Export Citation
  • [5]

    Loveland T. R., Merchant, J. M., 2004, Ecoregions and Ecoregionalization: Geographical and Ecological Perspectives. Environ Manage., 2004, 34, S1-S13,

    • Crossref
    • PubMed
    • Export Citation
  • [6]

    Ribeiro D., Somodi I., Čarni A., Transferability of a predictive Robinia pseudacacia distribution model in northeast Slovenia. Acta geogr. Slov., 2016, 56,1, 25–43,

    • Crossref
    • Export Citation
  • [7]

    Marković V.N., Vasiljević D.A., Jovanović T., Lukić T., Vujićić M.D., Kovaćević M., et al. 2017, The effect of natural and human-induced habitat conditions on number of roe deer: case study of Vojvodina, Serbia. Acta geogr. Slov., 57, 2, 57–69,

    • Crossref
    • Export Citation
  • [8]

    Torres M., Qiu G., Automatic habitat classification using image analysis and random forest. Ecol. Inform., 2014, 23, 126-136

    • Crossref
    • Export Citation
  • [9]

    Mücher C.A., Bunce R.G.H., Jongman R.H.G., Klijn J.A., Koomen A.J.M., Metzger M.J., et al., Identification and Characterisation of Environments and Landscapes in Europe. Alterra rapport 832. Alterra, Wageningen, 2003

  • [10]

    Ichter J., Evans D., Richard D., Terrestrial Habitat Mapping in Europe: An Overview. European Environmental Agency, Luxembourg, 2014

  • [11]

    Kokalj Ž., Oštir K., Land cover mapping using Landsat satellite image classification in the Classical Karst – Kras region. Acta Carsologica, 2007, 36, 3, 433-440

  • [12]

    Rodriguez-Galiano V.F., Chica-Olmo M., Abarca-Hernandez F., Atkinson P.M., Jeganathan C., Random forest classification of Mediterranean land cover using multi-seasonal imagery and multi-seasonal texture. Remote Sens. Environ., 2012, 121, 93-107

    • Crossref
    • Export Citation
  • [13]

    Haest B., Thoonen G., Vanden Borre J., Spanhove T., Delalieux S., Bertels L., Kooistra, L., Mücher C.A., Scheunders P., An object-based approach to quantity and quality assessment of heathland habitats in the framework of Natura 2000 using hyperspectral airborne AHS images. ISPRS Archives, 2010, 38, 4/C7

  • [14]

    Erasmi S., Fuchs H., Westphal C., Kleinn C., Mapping habitat type with Terrasar-X, Radarsat-2 and Rapideye data. Proceedings of the 5. TerraSAR-X Science Team Meeting, Oberpfaffenhofen, 2013 European Commision (EC): Environment, 2015

  • [15]

    Hladik C., Schalles J., Alber M., Salt marsh elevation and habitat mapping using hyperspectral and LIDAR data. Remote Sens. Environ., 2013, 139, 318-330,

    • Crossref
    • Export Citation
  • [16]

    Rapinel S., Hubert-Moy L., Clément B., Combined use of LiDAR data and multispectral earth observation imagery for wetland habitat mapping. Int. J. Appl. Earth Obs. Geoinf., 2015, 37, 56-64,

    • Crossref
    • Export Citation
  • [17]

    Wang K., Franklin S.E., Guo X.L., Cattet M., Remote sensing of ecology, biodiversity and conservation: A review from the perspective of remote sensing specialists. Sensors, 2010, 10, 9647-9667

    • Crossref
    • Export Citation
  • [18]

    Ackers S.H., Davis R.J., Olsen K.A., Dugger K.M., The evolution of mapping habitat for northern spotted owls (Strix occidentalis caurina): A comparison of photo-interpreted, Landsat-based, and lidar-based habitat maps. Remote Sens. Environ., 2015, 35, 2, 277-283

  • [19]

    Delalieux S., Somers B., Haest B., Spanhove T., Vanden Borre J., Mücher C.A., Heathland conservation status mapping through integration of hyperspectral mixture analysis and decision tree classifiers. Remote Sens. Environ., 2012, 126, 222-231,

    • Crossref
    • Export Citation
  • [20]

    Zhai L., Sun J., Sang H., Yang G., Jia Y., Large area land cover classification with Landsat ETM+ images based on decision tree. ISPRS J. Photogramm. Remote Sens., 2012, 421-426

  • [21]

    Somodi, I., Čarni, A., Ribeiro, D., Podobnikar, T., Recognition of the invasive species Robinia pseudacacia from combined remote sensing and GIS sources, Biological Conservation, 2012, 150, 1, 59-67.

    • Crossref
    • Export Citation
  • [22]

    Zlinszky A., Schroiff A., Kania A., Deák B., Mücke W., Vári Á., et al., Categorizing Grassland Vegetation with Full-Waveform Airborne Laser Scanning: A Feasibility Study for Detecting Natura 2000 Habitat Types. Remote Sens., 2014, 6, 8056-8087

    • Crossref
    • Export Citation
  • [23]

    Mokarram M., Sathyamoorthy D., Relationship between landform classification and vegetation (case study: southwest of Fars province, Iran). Open Geosci., 2016, 8, 1, 302–309.

    • Crossref
    • Export Citation
  • [24]

    Friedl M.A., Sulla-Menashe D., Tan B., Schneider A., Ramankutty N. Sibley A., et al., MODIS Collection 5 global land cover: Algorithm refinements and characterization of new datasets. Remote Sens. Environ., 2006, 100, 223-236

  • [25]

    Chehbouni A., Escadafal R., Duchemin B., Boulet G., Simonneaux V., Dedieu G., et al., An integrated modelling and remote sensing approach for hydrological study in arid and semi-arid regions: the SUDMED Programme. Int. J. Remote Sens., 2008, 29, 5161-5181.

    • Crossref
    • Export Citation
  • [26]

    Schuster C., Förster M., Kleinschmit B., Testing the red edge channel for improving land-use classifications based on highresolution multi-spectral satellite data. Int. J. Remote Sens., 2012, 33, 17, 5583-5599

    • Crossref
    • Export Citation
  • [27]

    Neumann C., Weiss G., Schmidtlein S., Itzerott S., Lausch A., Doktor D., Brell M., Gradient-Based Assessment of Habitat Quality for Spectral Ecosystem Monitoring. Remote Sens., 2015, 7, 2871-2898

    • Crossref
    • Export Citation
  • [28]

    Hofer G., Bunce R.G.H., Edwards P.J., Szerencits E., Wagner H.H., Herzog F., Use of topographic variability for assessing plant diversity in agricultural landscapes. Agric. Ecosyst. Environ., 2011, 142, 144-148

    • Crossref
    • Export Citation
  • [29]

    Li C., Wang J., Wang L., Hu L., Gong P., Comparison of Classification Algorithms and Training Sample Sizes in Urban Land Classification with Landsat Thematic Mapper Imagery. Remote Sens., 2014, 6, 964-983,

    • Crossref
    • Export Citation
  • [30]

    Debeljak M., Ficko A., Brus R., The use of habitat and dispersal models in protecting European black poplar (Populus nigra L.) from genetic introgression in Slovenia. Biol. Conserv., 2015, 184, 310-319,

    • Crossref
    • Export Citation
  • [31]

    Coops N.C., Waring R.H., Beier C., Roy-Jauvin R., Wang T., Modeling the occurrence of 15 coniferous tree species throughout the Pacific Northwest of North America using a hybrid approach of a generic process-based growth model and decision tree analysis. Appl. Veg. Sci., 2011, 14, 402-414,

    • Crossref
    • Export Citation
  • [32]

    Hellesen T., Matikainen L., An object-based approach for mapping shrub and tree cover on grassland habitats by use of Li-DAR and CIR orthoimages. Remote Sens., 2013, 5, 558-583,

    • Crossref
    • Export Citation
  • [33]

    Hansen M.C., Roy D.P., Lindquist E., Adusei B., Justice C.O., Altstatt A., A method for integrating MODIS and Landsat data for systematic monitoring of forest cover and change in the Congo Basin. Remote Sens. Environ., 2008, 112, 5, 2495-2513,

    • Crossref
    • Export Citation
  • [34]

    Sadeghi R., Zarkami R., Sabetraftar K., Van Damme P., Application of genetic algorithm and greedy stepwise to select input variables in classification tree models for the prediction of habitat requirements of Azolla filiculoides (Lam.) in Anzali wetland, Iran. Ecol. Model., 2013, 251, 44-53,

    • Crossref
    • Export Citation
  • [35]

    Ghulam A., Porton I., Freeman K., Detecting subcanopy invasive plant species in tropical rainforest by integrating optical and microwave (InSAR/PollnSAR) remote sensing data, and a decision tree algorithm. ISPRS J. Photogramm. Remote Sens., 2014, 88, 174-192,

    • Crossref
    • Export Citation
  • [36]

    Asam S., Klein D., Dech S., Estimation of grassland use intensities based on high spatial resolution LAI time series. Int. Arch. Photogramm. Remote Sens. Spatial Inf. Sci., 2015, XL-7- W3, 285-291

    • Crossref
    • Export Citation
  • [37]

    Gonzales-Mirelis G., Lindegarth M., Predicting the distribution of out-of-reach biotopes with decision trees in a Swedish marine protected area. Ecol. Appl., 2012, 22, 8, 2248-2264,

    • Crossref
    • Export Citation
  • [38]

    Kocev D., Džeroski S., Habitat modeling with single- and multi-target trees and ensembles. Ecol. Inform., 2013, 18, 79-92,

    • Crossref
    • Export Citation
  • [39]

    Šiljković Ž., Čuka A., Pejdo A., Rural area transformation: From cropland to mine fields – Zemunik Donji municipality (Croatia) case study. Društvena istraživanja., 2011, 20, 4, 1163-1181,

    • Crossref
    • Export Citation
  • [40]

    Kobler A., Adamic M., Identifying brown bear habitat by a combined GIS and machine learning method. Ecol. Model., 2000, 135, 291–300,

    • Crossref
    • Export Citation
  • [41]

    Narce M., Meloni R., Beroud T., Pléney A., Ricci J.C., Landscape ecology and wild rabbit (Oryctolagus cuniculus) habitat modeling in the Mediterranean region. Anim. Biodivers. Conserv., 2012, 35, 2, 277-283

  • [42]

    Mihevc A., Prelovšek M., Geographical position and general overview. In: Mihevc A., Prelovšek M., Zupan Hajna N. (Eds.), Introduction to the Dinaric Karst, Inštitut za raziskovanje krasa ZRC SAZU, Postojna, 2010, 6-8

  • [43]

    Gams I., Kras v Sloveniji v prostoru in času. Založba ZRC, Ljubljana, 2004

  • [44]

    Kaligarič M., Ivajnšič D., Vanishing landscape of the “classic” Karst: changed landscape identity and projections for the future. Landsc. Urban Plan., 2014, 132, 148-158,

    • Crossref
    • Export Citation
  • [45]

    Kaligarič M., Škornik S., Šajna N., Otopal J., Pipenbaher N., Ivajnšič D., Kartiranje negozdnih habitatnih tipov Slovenije / Območje KRAS – Kozina in KRAS – Vremščica. Final Report (in Slovene), Faculty of natural sciences and mathematics, University Maribor, 2011

  • [46]

    Devillers P., Devillers-Terschuren J., A classification of Palaearctic habitats. Nature and environment, 1996, no. 78, Strasbourg, Council of Europe

  • [47]

    Pehani P., Čotar K., Marsetič A., Zaletelj J., Oštir K., Automatic Geometric Processing for Very High Resolution Optical Satellite Data Based on Vector Roads and Orthophotos. Remote Sens.- Basel, 2016, 8, 4, 1-26.

    • Crossref
    • Export Citation
  • [48]

    Moore I.D., Gessler P.E., Nielsen G.A., Petersen G.A., Terrain attributes: estimation methods and scale effects. In: Jakeman A.J., Beck M.B., McAleer M. (Eds.), Modeling Change in Environmental Systems. Wiley, London, 1993, 189-214

  • [49]

    Gessler P.E., Moore I.D., McKenzie N.J., Ryan P.J., Soil-landscape modeling and spatial prediction of soil attributes. Int. J. Geogr. Inf. Syst., 1995, 9, 4, 421-432

    • Crossref
    • Export Citation
  • [50]

    Evans J.S., Oakleaf J., Cushman S.A., Theobald D., An ArcGIS Toolbox for Surface Gradient and Geomorphometric Modeling, version 2.0-0, 2014, http://evansmurphy.wix.com/evansspatial

  • [51]

    McCune B., Keon D., Equations for potential annual direct incident radiation and heat load. J. Veg. Sci., 2002, 13, 603-606,

    • Crossref
    • Export Citation
  • [52]

    Iverson L.R., Dale M.E., Scott C.T., Prasad A., A GIS-derived integrated moisture index to predict forest composition and productivity in Ohio forests. Landsc. Ecol., 1997, 12, 331-348

    • Crossref
    • Export Citation
  • [53]

    Balice R.G., Miller J.D., Oswald B.P., Edminister C., Yool S.R., Forest surveys and wildfire assessment in the Los Alamos; 1998–1999. Los Alamos National Laboratory, Los Alamos, NewMexico, 2000

  • [54]

    Tomaselli V., Dimopoulos P., Marangi C., Kallimanis A. S., Adamo M., Tarantino C., et al., Translating land cover/land use classifications to habitat taxonomies for landscape monitoring: a Mediterranean assessment. Landsc. Ecol., 2013, 28, 5, 905-930

    • Crossref
    • Export Citation
  • [55]

    Jogan N., Kaligarič M., Leskovar I., Seliškar A., Dobravec J., Habitatni tipi Slovenije HTS 2004 – tipologija. Ministrstvo za okolje, prostor in energijo – Agencija RS za okolje, Ljubljana, 2004

  • [56]

    Di Gregorio A., Jansen L.J.M., Land Cover Classification System (LCCS): classification concepts and user manual. Food and Agriculture Organization of the United Nations, 2005, Rome

  • [57]

    Kobler A., Adamič M., Brown Bears in Slovenia: Identifying Locations for Construction of Wildlife Bridges Across Highways. Proceedings of the International conference on wildlife ecology & transportation, Missoula, Montana, 1999

  • [58]

    Witten I.H., Frank E., Data mining: practical machine learning tools and techniques. Morgan Kaufman Publishers, Amsterdam, 2005

  • [59]

    Kononenko I., Kukar M., Machine learning and data mining: Introduction to principles and algorithms. Hordwood Publishing, Chichester, 2007

  • [60]

    Dimitrovski I., Kocev D., Loskovska S., Džeroski S., Hierarchical classification of diatom images using ensembles of predictive clustering trees. Ecol. Inform., 2012, 7, 1, 19-29,

    • Crossref
    • Export Citation
  • [61]

    Lin N., Noe D., He X., Tree-based methods and their applications. V: Phoam H. (Ed.), Springer handbook of engineering statistics. Springer, London, 2006, 551-570

    • Crossref
    • Export Citation
  • [62]

    Lu D., Weng Q., A survey of image classification methods and techniques for improving classification performance. Int. J. Remote Sens., 2007, 28, 5, 823-870,

    • Crossref
    • Export Citation
  • [63]

    Jiang B., Ding X., Ma L., He Y., Wang T., Xie W., A hybrid feature selection algorithm: combination of symmetrical uncertainty and genetic algorithms. In: Proceedings of the second international symposium Optimization and systems biology, Lijiang, 2008, 152–157

  • [64]

    Tirelli T., Pessani D., Importance of feature selection in decisiontree and artificial neural- network ecological applications. Alburnus alburnus alborella: A practical example. Ecol. Inform., 2011, 6, 5, 309-315,

    • Crossref
    • Export Citation
  • [65]

    Williams K.J., Belbin L., Austin M.P., Stein J.L., Ferrier S., Which environmental variables should I use in my biodiversity model? Int. J. Geogr. Inf. Sci., 2012, 26, 2009-2047,

    • Crossref
    • Export Citation
  • [66]

    Ciglič R., Information values of absolute elevation and elevation difference for –illustration of thermal belt. Acta Geogr. Slo., 2010, 50, 2, 177-200,

    • Crossref
    • Export Citation
  • [67]

    Fayya, U.M., Irani, K.B., Multi-interval discretization of continuous-valued attributes for classification learning. In: Bajcsy R. (Ed.), Proceedings of the Thirteenth international joint conference on artificial intelligence 2, San Mateo, 1993, 1022-1027

  • [68]

    Hall M., Frank E., Holmes G., Pfahringer B., Reutemann P., Witten I.H., The WEKA data mining software: an update. SIGKDD Explor., 2009, 11, 1, 10-18

    • Crossref
    • Export Citation

Footnotes

***

P<0.001

***

P<0.001

***

P<0.001

***

P<0.001

***

P<0.001

***

P<0.001

***

P<0.001

***

P<0.001

***

P<0.001

***

P<0.001

***

P<0.001

***

P<0.001

***

P<0.001

***

P<0.001

**

P<0.01

*

P<0.05

***

P<0.001

*

P<0.05

***

P<0.001

***

P<0.001

***

P<0.001

***

P<0.001

***

P<0.001

***

P<0.001

***

P<0.001

***

P<0.001

***

P<0.001

***

P<0.001

***

P<0.001

***

P<0.001

***

P<0.001

If the inline PDF is not rendering correctly, you can download the PDF file here.

  • [1]

    Bailey R.G., Ecosystem geography. Springer, New York, 1996

  • [2]

    Bunce R.G.H, Barr C.J., Clarke R.T., Howard D.C., Lane A.M.J., Land classfiication for strategic ecological survey. J. Environ Manage., 1996, 47, 1, 37–60.

    • Crossref
    • Export Citation
  • [3]

    Varga K., Szabó S., Szabó G., Dévai G., Tóthmérész B., Improved land cover mapping using aerial photographs and satellite images. Open Geosci., 2014, 7, 1.

    • Crossref
    • Export Citation
  • [4]

    Bastian O., Landscape classification in Saxony (Germany) – a tool for holistic regional planning. Landsc. Urban Plan., 2000, 50, 1-3, 145–155,

    • Crossref
    • Export Citation
  • [5]

    Loveland T. R., Merchant, J. M., 2004, Ecoregions and Ecoregionalization: Geographical and Ecological Perspectives. Environ Manage., 2004, 34, S1-S13,

    • Crossref
    • PubMed
    • Export Citation
  • [6]

    Ribeiro D., Somodi I., Čarni A., Transferability of a predictive Robinia pseudacacia distribution model in northeast Slovenia. Acta geogr. Slov., 2016, 56,1, 25–43,

    • Crossref
    • Export Citation
  • [7]

    Marković V.N., Vasiljević D.A., Jovanović T., Lukić T., Vujićić M.D., Kovaćević M., et al. 2017, The effect of natural and human-induced habitat conditions on number of roe deer: case study of Vojvodina, Serbia. Acta geogr. Slov., 57, 2, 57–69,

    • Crossref
    • Export Citation
  • [8]

    Torres M., Qiu G., Automatic habitat classification using image analysis and random forest. Ecol. Inform., 2014, 23, 126-136

    • Crossref
    • Export Citation
  • [9]

    Mücher C.A., Bunce R.G.H., Jongman R.H.G., Klijn J.A., Koomen A.J.M., Metzger M.J., et al., Identification and Characterisation of Environments and Landscapes in Europe. Alterra rapport 832. Alterra, Wageningen, 2003

  • [10]

    Ichter J., Evans D., Richard D., Terrestrial Habitat Mapping in Europe: An Overview. European Environmental Agency, Luxembourg, 2014

  • [11]

    Kokalj Ž., Oštir K., Land cover mapping using Landsat satellite image classification in the Classical Karst – Kras region. Acta Carsologica, 2007, 36, 3, 433-440

  • [12]

    Rodriguez-Galiano V.F., Chica-Olmo M., Abarca-Hernandez F., Atkinson P.M., Jeganathan C., Random forest classification of Mediterranean land cover using multi-seasonal imagery and multi-seasonal texture. Remote Sens. Environ., 2012, 121, 93-107

    • Crossref
    • Export Citation
  • [13]

    Haest B., Thoonen G., Vanden Borre J., Spanhove T., Delalieux S., Bertels L., Kooistra, L., Mücher C.A., Scheunders P., An object-based approach to quantity and quality assessment of heathland habitats in the framework of Natura 2000 using hyperspectral airborne AHS images. ISPRS Archives, 2010, 38, 4/C7

  • [14]

    Erasmi S., Fuchs H., Westphal C., Kleinn C., Mapping habitat type with Terrasar-X, Radarsat-2 and Rapideye data. Proceedings of the 5. TerraSAR-X Science Team Meeting, Oberpfaffenhofen, 2013 European Commision (EC): Environment, 2015

  • [15]

    Hladik C., Schalles J., Alber M., Salt marsh elevation and habitat mapping using hyperspectral and LIDAR data. Remote Sens. Environ., 2013, 139, 318-330,

    • Crossref
    • Export Citation
  • [16]

    Rapinel S., Hubert-Moy L., Clément B., Combined use of LiDAR data and multispectral earth observation imagery for wetland habitat mapping. Int. J. Appl. Earth Obs. Geoinf., 2015, 37, 56-64,

    • Crossref
    • Export Citation
  • [17]

    Wang K., Franklin S.E., Guo X.L., Cattet M., Remote sensing of ecology, biodiversity and conservation: A review from the perspective of remote sensing specialists. Sensors, 2010, 10, 9647-9667

    • Crossref
    • Export Citation
  • [18]

    Ackers S.H., Davis R.J., Olsen K.A., Dugger K.M., The evolution of mapping habitat for northern spotted owls (Strix occidentalis caurina): A comparison of photo-interpreted, Landsat-based, and lidar-based habitat maps. Remote Sens. Environ., 2015, 35, 2, 277-283

  • [19]

    Delalieux S., Somers B., Haest B., Spanhove T., Vanden Borre J., Mücher C.A., Heathland conservation status mapping through integration of hyperspectral mixture analysis and decision tree classifiers. Remote Sens. Environ., 2012, 126, 222-231,

    • Crossref
    • Export Citation
  • [20]

    Zhai L., Sun J., Sang H., Yang G., Jia Y., Large area land cover classification with Landsat ETM+ images based on decision tree. ISPRS J. Photogramm. Remote Sens., 2012, 421-426

  • [21]

    Somodi, I., Čarni, A., Ribeiro, D., Podobnikar, T., Recognition of the invasive species Robinia pseudacacia from combined remote sensing and GIS sources, Biological Conservation, 2012, 150, 1, 59-67.

    • Crossref
    • Export Citation
  • [22]

    Zlinszky A., Schroiff A., Kania A., Deák B., Mücke W., Vári Á., et al., Categorizing Grassland Vegetation with Full-Waveform Airborne Laser Scanning: A Feasibility Study for Detecting Natura 2000 Habitat Types. Remote Sens., 2014, 6, 8056-8087

    • Crossref
    • Export Citation
  • [23]

    Mokarram M., Sathyamoorthy D., Relationship between landform classification and vegetation (case study: southwest of Fars province, Iran). Open Geosci., 2016, 8, 1, 302–309.

    • Crossref
    • Export Citation
  • [24]

    Friedl M.A., Sulla-Menashe D., Tan B., Schneider A., Ramankutty N. Sibley A., et al., MODIS Collection 5 global land cover: Algorithm refinements and characterization of new datasets. Remote Sens. Environ., 2006, 100, 223-236

  • [25]

    Chehbouni A., Escadafal R., Duchemin B., Boulet G., Simonneaux V., Dedieu G., et al., An integrated modelling and remote sensing approach for hydrological study in arid and semi-arid regions: the SUDMED Programme. Int. J. Remote Sens., 2008, 29, 5161-5181.

    • Crossref
    • Export Citation
  • [26]

    Schuster C., Förster M., Kleinschmit B., Testing the red edge channel for improving land-use classifications based on highresolution multi-spectral satellite data. Int. J. Remote Sens., 2012, 33, 17, 5583-5599

    • Crossref
    • Export Citation
  • [27]

    Neumann C., Weiss G., Schmidtlein S., Itzerott S., Lausch A., Doktor D., Brell M., Gradient-Based Assessment of Habitat Quality for Spectral Ecosystem Monitoring. Remote Sens., 2015, 7, 2871-2898

    • Crossref
    • Export Citation
  • [28]

    Hofer G., Bunce R.G.H., Edwards P.J., Szerencits E., Wagner H.H., Herzog F., Use of topographic variability for assessing plant diversity in agricultural landscapes. Agric. Ecosyst. Environ., 2011, 142, 144-148

    • Crossref
    • Export Citation
  • [29]

    Li C., Wang J., Wang L., Hu L., Gong P., Comparison of Classification Algorithms and Training Sample Sizes in Urban Land Classification with Landsat Thematic Mapper Imagery. Remote Sens., 2014, 6, 964-983,

    • Crossref
    • Export Citation
  • [30]

    Debeljak M., Ficko A., Brus R., The use of habitat and dispersal models in protecting European black poplar (Populus nigra L.) from genetic introgression in Slovenia. Biol. Conserv., 2015, 184, 310-319,

    • Crossref
    • Export Citation
  • [31]

    Coops N.C., Waring R.H., Beier C., Roy-Jauvin R., Wang T., Modeling the occurrence of 15 coniferous tree species throughout the Pacific Northwest of North America using a hybrid approach of a generic process-based growth model and decision tree analysis. Appl. Veg. Sci., 2011, 14, 402-414,

    • Crossref
    • Export Citation
  • [32]

    Hellesen T., Matikainen L., An object-based approach for mapping shrub and tree cover on grassland habitats by use of Li-DAR and CIR orthoimages. Remote Sens., 2013, 5, 558-583,

    • Crossref
    • Export Citation
  • [33]

    Hansen M.C., Roy D.P., Lindquist E., Adusei B., Justice C.O., Altstatt A., A method for integrating MODIS and Landsat data for systematic monitoring of forest cover and change in the Congo Basin. Remote Sens. Environ., 2008, 112, 5, 2495-2513,

    • Crossref
    • Export Citation
  • [34]

    Sadeghi R., Zarkami R., Sabetraftar K., Van Damme P., Application of genetic algorithm and greedy stepwise to select input variables in classification tree models for the prediction of habitat requirements of Azolla filiculoides (Lam.) in Anzali wetland, Iran. Ecol. Model., 2013, 251, 44-53,

    • Crossref
    • Export Citation
  • [35]

    Ghulam A., Porton I., Freeman K., Detecting subcanopy invasive plant species in tropical rainforest by integrating optical and microwave (InSAR/PollnSAR) remote sensing data, and a decision tree algorithm. ISPRS J. Photogramm. Remote Sens., 2014, 88, 174-192,

    • Crossref
    • Export Citation
  • [36]

    Asam S., Klein D., Dech S., Estimation of grassland use intensities based on high spatial resolution LAI time series. Int. Arch. Photogramm. Remote Sens. Spatial Inf. Sci., 2015, XL-7- W3, 285-291

    • Crossref
    • Export Citation
  • [37]

    Gonzales-Mirelis G., Lindegarth M., Predicting the distribution of out-of-reach biotopes with decision trees in a Swedish marine protected area. Ecol. Appl., 2012, 22, 8, 2248-2264,

    • Crossref
    • Export Citation
  • [38]

    Kocev D., Džeroski S., Habitat modeling with single- and multi-target trees and ensembles. Ecol. Inform., 2013, 18, 79-92,

    • Crossref
    • Export Citation
  • [39]

    Šiljković Ž., Čuka A., Pejdo A., Rural area transformation: From cropland to mine fields – Zemunik Donji municipality (Croatia) case study. Društvena istraživanja., 2011, 20, 4, 1163-1181,

    • Crossref
    • Export Citation
  • [40]

    Kobler A., Adamic M., Identifying brown bear habitat by a combined GIS and machine learning method. Ecol. Model., 2000, 135, 291–300,

    • Crossref
    • Export Citation
  • [41]

    Narce M., Meloni R., Beroud T., Pléney A., Ricci J.C., Landscape ecology and wild rabbit (Oryctolagus cuniculus) habitat modeling in the Mediterranean region. Anim. Biodivers. Conserv., 2012, 35, 2, 277-283

  • [42]

    Mihevc A., Prelovšek M., Geographical position and general overview. In: Mihevc A., Prelovšek M., Zupan Hajna N. (Eds.), Introduction to the Dinaric Karst, Inštitut za raziskovanje krasa ZRC SAZU, Postojna, 2010, 6-8

  • [43]

    Gams I., Kras v Sloveniji v prostoru in času. Založba ZRC, Ljubljana, 2004

  • [44]

    Kaligarič M., Ivajnšič D., Vanishing landscape of the “classic” Karst: changed landscape identity and projections for the future. Landsc. Urban Plan., 2014, 132, 148-158,

    • Crossref
    • Export Citation
  • [45]

    Kaligarič M., Škornik S., Šajna N., Otopal J., Pipenbaher N., Ivajnšič D., Kartiranje negozdnih habitatnih tipov Slovenije / Območje KRAS – Kozina in KRAS – Vremščica. Final Report (in Slovene), Faculty of natural sciences and mathematics, University Maribor, 2011

  • [46]

    Devillers P., Devillers-Terschuren J., A classification of Palaearctic habitats. Nature and environment, 1996, no. 78, Strasbourg, Council of Europe

  • [47]

    Pehani P., Čotar K., Marsetič A., Zaletelj J., Oštir K., Automatic Geometric Processing for Very High Resolution Optical Satellite Data Based on Vector Roads and Orthophotos. Remote Sens.- Basel, 2016, 8, 4, 1-26.

    • Crossref
    • Export Citation
  • [48]

    Moore I.D., Gessler P.E., Nielsen G.A., Petersen G.A., Terrain attributes: estimation methods and scale effects. In: Jakeman A.J., Beck M.B., McAleer M. (Eds.), Modeling Change in Environmental Systems. Wiley, London, 1993, 189-214

  • [49]

    Gessler P.E., Moore I.D., McKenzie N.J., Ryan P.J., Soil-landscape modeling and spatial prediction of soil attributes. Int. J. Geogr. Inf. Syst., 1995, 9, 4, 421-432

    • Crossref
    • Export Citation
  • [50]

    Evans J.S., Oakleaf J., Cushman S.A., Theobald D., An ArcGIS Toolbox for Surface Gradient and Geomorphometric Modeling, version 2.0-0, 2014, http://evansmurphy.wix.com/evansspatial

  • [51]

    McCune B., Keon D., Equations for potential annual direct incident radiation and heat load. J. Veg. Sci., 2002, 13, 603-606,

    • Crossref
    • Export Citation
  • [52]

    Iverson L.R., Dale M.E., Scott C.T., Prasad A., A GIS-derived integrated moisture index to predict forest composition and productivity in Ohio forests. Landsc. Ecol., 1997, 12, 331-348

    • Crossref
    • Export Citation
  • [53]

    Balice R.G., Miller J.D., Oswald B.P., Edminister C., Yool S.R., Forest surveys and wildfire assessment in the Los Alamos; 1998–1999. Los Alamos National Laboratory, Los Alamos, NewMexico, 2000

  • [54]

    Tomaselli V., Dimopoulos P., Marangi C., Kallimanis A. S., Adamo M., Tarantino C., et al., Translating land cover/land use classifications to habitat taxonomies for landscape monitoring: a Mediterranean assessment. Landsc. Ecol., 2013, 28, 5, 905-930

    • Crossref
    • Export Citation
  • [55]

    Jogan N., Kaligarič M., Leskovar I., Seliškar A., Dobravec J., Habitatni tipi Slovenije HTS 2004 – tipologija. Ministrstvo za okolje, prostor in energijo – Agencija RS za okolje, Ljubljana, 2004

  • [56]

    Di Gregorio A., Jansen L.J.M., Land Cover Classification System (LCCS): classification concepts and user manual. Food and Agriculture Organization of the United Nations, 2005, Rome

  • [57]

    Kobler A., Adamič M., Brown Bears in Slovenia: Identifying Locations for Construction of Wildlife Bridges Across Highways. Proceedings of the International conference on wildlife ecology & transportation, Missoula, Montana, 1999

  • [58]

    Witten I.H., Frank E., Data mining: practical machine learning tools and techniques. Morgan Kaufman Publishers, Amsterdam, 2005

  • [59]

    Kononenko I., Kukar M., Machine learning and data mining: Introduction to principles and algorithms. Hordwood Publishing, Chichester, 2007

  • [60]

    Dimitrovski I., Kocev D., Loskovska S., Džeroski S., Hierarchical classification of diatom images using ensembles of predictive clustering trees. Ecol. Inform., 2012, 7, 1, 19-29,

    • Crossref
    • Export Citation
  • [61]

    Lin N., Noe D., He X., Tree-based methods and their applications. V: Phoam H. (Ed.), Springer handbook of engineering statistics. Springer, London, 2006, 551-570

    • Crossref
    • Export Citation
  • [62]

    Lu D., Weng Q., A survey of image classification methods and techniques for improving classification performance. Int. J. Remote Sens., 2007, 28, 5, 823-870,

    • Crossref
    • Export Citation
  • [63]

    Jiang B., Ding X., Ma L., He Y., Wang T., Xie W., A hybrid feature selection algorithm: combination of symmetrical uncertainty and genetic algorithms. In: Proceedings of the second international symposium Optimization and systems biology, Lijiang, 2008, 152–157

  • [64]

    Tirelli T., Pessani D., Importance of feature selection in decisiontree and artificial neural- network ecological applications. Alburnus alburnus alborella: A practical example. Ecol. Inform., 2011, 6, 5, 309-315,

    • Crossref
    • Export Citation
  • [65]

    Williams K.J., Belbin L., Austin M.P., Stein J.L., Ferrier S., Which environmental variables should I use in my biodiversity model? Int. J. Geogr. Inf. Sci., 2012, 26, 2009-2047,

    • Crossref
    • Export Citation
  • [66]

    Ciglič R., Information values of absolute elevation and elevation difference for –illustration of thermal belt. Acta Geogr. Slo., 2010, 50, 2, 177-200,

    • Crossref
    • Export Citation
  • [67]

    Fayya, U.M., Irani, K.B., Multi-interval discretization of continuous-valued attributes for classification learning. In: Bajcsy R. (Ed.), Proceedings of the Thirteenth international joint conference on artificial intelligence 2, San Mateo, 1993, 1022-1027

  • [68]

    Hall M., Frank E., Holmes G., Pfahringer B., Reutemann P., Witten I.H., The WEKA data mining software: an update. SIGKDD Explor., 2009, 11, 1, 10-18

    • Crossref
    • Export Citation
OPEN ACCESS

Journal + Issues

Search

  • View in gallery

    Delimitation of the Dinaric Karst [43] and the location of the study area.

  • View in gallery

    Complete methodology of the habitat modelling. Different raster data layers were derived on the basis of DEM, Rapideye images, orthophotos, and infrared orthophotos. After the variables evaluation 9 data variables were selected, which were the explanatory (independent) variables. Original habitat types classification (with 51 types; level A) was hierarchically reclassified into more generalized types (levels B, C and D). Models of habitat types were produced for levels B, C, and D on the basis of training sets. Validation set was used to evaluate modelling results. At the end comparison between the levels was performed.

  • View in gallery

    Classification success of modelling land cover at general level – FAO-LCCS. Only patches included in the analysis are presented.

  • View in gallery

    Classification success of modelling at comparative land cover level. Only patches included in the analysis are presented.

  • View in gallery

    Classification success of modelling at HTS habitats level. Only patches included in the analysis are presented.