2017, no. 199-200, p. 167-176

Original scientific paper
UDK/DOI: 551.583:630

Development of Species Distribution Model Using Mechine Learning Methods

Lazar Pavlović 1, Dejan B. Stojanović 2, Milena Kresoja 3, Stefan Stjepanović 4, Saša Orlović 1,  Mirjana Bojović 5

1 University of Novi Sad, Faculty of Agriculture
1 University of Novi Sad, Institute of Lowland Forestry and Environment
3 University of Novi Sad, Faculty of Sciences
4 University of East Sarajevo, Faculty of Agriculture
5 University Educons, Faculty of Environment Protection

e-mail: lazar.pavlovic@polj.uns.ac.rs


Climate change that has been intensively occurring in the last few decades has a global effect on vegetation and forest cover, leading to major transformations in natural resources and the landscape structure. The impact of climate changes on species is often estimated using a species distribution models (SDMs). These models use environmental data and presence/absence of a species, determine their mutual relationship, in order to show on other locations whether environmental conditions are suitable or not for the existence of this species. Since models are easy to implement, they are now widely used to consider various issues in environmental research, as well as providing guidance for applied research. The aim of this paper is to develop and evaluate the Random Forest (RF) model based on current data on existence of European beech, ecological and climatic characteristics in the territory of Serbia. The model obtained will serve as the basis for building a model that will foresee the distribution of species in the future. The accuracy of the model was tested using adequate statistical methods. The True Skill Statistic (TSS) analysis indicates a high accuracy of the model (TSS = 0.87, specificity = 87.81, sensitivity = 99.44). The accuracy was confirmed by the analysis of the area under the ROC (Receiver Operating Characteristic) curve (AUC) (AUC = 0.97, specificity = 88.01, sensitivity = 99.27). Also, the results pointed to the need to include more environmentally relevant topographic variables when designing a SDM in relation to climate change, especially for species that are correlated with topography.

Keywords: Species distribution models-SDMs, European beech, machine learning, BIOMOD2