Debeljak, M., Dzeroski, S., Jerina, K., Kobler, A., and Adamic, M. (2001). Habitat suitability modelling for red deer (Cervus elaphus L.) in South-central Slovenia with classification trees. Ecological Modelling 138: 321-330. Keywords: 8SI/Cervus elaphus/classification/classification tree/deer/distribution/geographic information system/GIS/habitat/habitat model/habitat suitability/home range/Malme/management/ model/models/population dynamics/red deer/Slovenia/spatial/spatial distribution Abstract: We study and assess the potential habitats of a population of red deer in South-central Slovenia. Using existing data on the deer population spatial distribution and size, as well as data on the landscape and ecological properties (GIS) of the area inhabited by this population, we develop a habitat suitability model by automated data analysis using machine learning of classification trees. We assume that the recorded observations of deer approximate the actual spatial distribution of the deer population reasonably well. The habitat suitability models for individual animals have the form of classification trees. The induced trees are interpreted by domain experts and a generic model is proposed. The generic habitat suitability models can help determine potential unoccupied habitats for the red deer population and develop guidelines for managing the development of the red deer population and its influence on the environment. Ecological Modelling 138 (2001) 321–330 www.elsevier.com/locate/ecolmodel Habitat suitability modelling for red deer (Cer6us elaphus L.) in South-central Slovenia with classification trees Marko Debeljak a,*, Sasˇo Dzˇeroski b, Klemen Jerina a, Andrej Kobler c, Miha Adamicˇ a a Uni6ersity of Ljubljana, Biotechnical Faculty, Forestry Department, Vezˇna pot 83, 1000 Ljubljana, Slo6enia b Jozˇef Stefan Institute, Jamo6a 39, 1000 Ljubljana, Slo6enia c Slo6enian Forestry Institute, Vezˇna pot 2, 1000 Ljubljana, Slo6enia Abstract We study and assess the potential habitats of a population of red deer in South-central Slovenia. Using existing data on the deer population spatial distribution and size, as well as data on the landscape and ecological properties (GIS) of the area inhabited by this population, we develop a habitat suitability model by automated data analysis using machine learning of classification trees. We assume that the recorded observations of deer approximate the actual spatial distribution of the deer population reasonably well. The habitat suitability models for individual animals have the form of classification trees. The induced trees are interpreted by domain experts and a generic model is proposed. The generic habitat suitability models can help determine potential unoccupied habitats for the red deer population and develop guidelines for managing the development of the red deer population and its influence on the environment. © 2001 Elsevier Science B.V. All rights reserved. Keywords: Population dynamics; Habitat suitability; Models; Home range; Geographic information system; Classification trees; Machine learning 1. Introduction changed structure and composition of the herbal part of forest ecosystems and also in the non-nat- Human activity has mostly negative influence ural structure and density of wildlife, particularly on self-regulating processes in forest ecosystems. game animals. High quality timber production Co-operation between wildlife management and and a high number of high score antlers at the forest management in Slovenia could be better same place are illusory goals if the above men- despite the fact that both share the same place for tioned management branches do not start to work their activities. This discrepancy is reflected in the together. This is a well-known problem all over the world and Slovenia is not an exception. Problems related to the imbalance between the * Corresponding author. Tel.: +386-1-4231161, ext. 536; population size of ungulates and forest manage- fax: +386-1-2571169. E-mail address: [email protected] (M. Debeljak). ment escalated in Slovenia at the end of the 1960s. 0304-3800/01/$ - see front matter © 2001 Elsevier Science B.V. All rights reserved. PII: S0304-3800(00)00411-7 322 M. Debeljak et al. / Ecological Modelling 138 (2001) 321–330 The natural regeneration of forests in the South- Debeljak et al. (1999) extend the research of ern, most forested part of Slovenia, came close to Stankovski et al. (1998) by showing that severe a halt (Perko, 1979). The body weight and the climatic conditions during the summer and winter health condition of red deer (Cer6us elaphus L.) also have a negative effect on the body weight of and roe deer (Capreolus capreolus L.) were deteri- red deer. They show that the quantity of precipi- orating (Simonicˇ, 1982). Foresters and hunters tation during the summer has significant effect on were faced with a problem, which could only be the body weight of 1-year-old calves and that the solved with an interdisciplinary approach. The weather conditions during the first winter are very result of joint work was a more holistic wildlife important for the body size and health condition management system that became known as ‘the of the red deer in the second year of its life. It was control method’ (Simonicˇ, 1982). They established also shown that the climatic conditions have an a network of permanent plots for observing the important influence on the body size and health degree of browsing of new growth in the forest. condition of red deer until the age of 10 years. They also started with systematic measurements After including meteorological information in of several groups of parameters about the body the studies of red deer population dynamics, the size, health condition and location of shoot-off next logical step is to take into account more for every harvested animal in the South-central information on the area where the studied popula- part of Slovenia. Based on the first analysis of the tion of red deer lives. This leads us to consider collected data, they decided for a large reduction home ranges of red deer, as well as information of the red deer population size. They estimated from a Geographic Information System about the that they reduced the population size by half. study area in South-central Slovenia. The two are Unfortunately the degree of browsing didn’tfol- combined to derive habitat suitability models by low the trend of the red deer population size means of machine learning tools. This is the sub- reduction. ject of the present paper. The problem was just partly solved despite the Habitat models have become well-accepted fact that browsing effects of other ungulates tools for understanding the habitat characteristics (mostly roe deer) were eliminated from the obser- of different organisms, evaluating habitat quality vations of the degree of browsing. A reaction to this was a new approach to understanding of the and developing wildlife management strategies processes in forest ecosystems. Motivated by the (Verner et al., 1986). They become even more many positive results and experiences of applying useful if they are linked with GIS and expert artificial intelligence methods, and in particular systems. In the field of GIS, expert systems help machine learning methods, in ecological mod- to design maps, to extract geographic data for use elling (see e.g. Kompare et al., 1994; Kompare, in GIS, to create databases and to support deci- 1995; Dzˇeroski et al., 1998). Stankovski et al. sion making. Loh and Rykiel (1992) support an (1998), used the approach of machine learning of integrated approach to environmental manage- regression trees (implemented in the tool RETIS) ment that couples database management, GIS, to analyse the interaction between population dy- and expert systems. Donovan et al. (1987) build namics of red deer and forest development. models from GIS landscape data to evaluate nest- This work pointed out the important influence ing and brood-rearing habitats for wild turkeys. of meteorological parameters on the degree and Walker (1990) used GIS data to inductively de- location of browsing. It was shown that high velop models of the presence or absence of kanga- temperatures and high quantity of precipitation roos in relation to climate on a regional scale. during the summer, as well as high snow cover Fabricius and Coetzee (1992) have used GIS and during the winter have significant influence on the artificial intelligence for developing a prediction degree of browsing. It was thus shown that the model for the presence or absence of mountain meteorological parameters lead to non-linear in- reebuck in Karoo arid dwarf shrubland in South teractions between the state of the forest new Africa. Kobler et al. (1997) studied a brown bear growth and the size of the red deer population. ecological niche in the area of the Vrhnika–Pos- M. Debeljak et al. / Ecological Modelling 138 (2001) 321–330 323 tojna highway by a GIS analysis of radio-tracking bution and the home range from a random sam- data. Kobler and Adamicˇ (1999) have used a ple of locational observations made on an animal. combination of GIS and artificial intelligence for The kernel estimator can be summarised as fol- brown bear habitat modelling and identification lows. A probability density function, namely the of locations for construction of wildlife bridges kernel, is placed over each data point and the across highways in Slovenia. The model focused estimator is calculated. Where the points (loca- on the brown bear migratory routes from its core tions) are concentrated, the kernel estimate has a area in Southern Slovenia towards Western Slove- higher density than where there are only few nia and the Alps region. points. The smoothing parameter determines how smooth the resulting function will be. If a low value of the smoothing parameter is used, fine 2. The data detail of the data can be observed, while a high value obscures all but the most prominent fea- The study covers the forest area of High Karst tures.