Identifying brown bear habitat by a combined GIS and machine learning method

In this paper we attempt to identify brown bear (Ursus arctos) habitat in south-western part of Slovenia, a country lying on the north-western-most edge of the continuous Dinaric-Eastern Alps brown bear population. The knowledge base (in the form of a decision tree) for the expert system for identifying the suitable habitat, was induced by automated machine learning from recorded bear sightings, and then linked to the GIS thematic layers for subsequent habitat/non-habitat classification of the entire study area. The accuracy of the decision tree classifier was 87% (KHAT 73%). The decision tree mostly agreed with the existing domain knowledge. For the study area the main factors considered by the expert system to be important for brown bear habitat were the percentage of forest (positive), proximity to settlements (negative) and elevation above see (positive), however the decision tree did not account for habitat patch size. After filtering out habitat patches smaller than 5000 ha in GIS, the accuracy increased to 89% (KHAT 77%). Whereas 88% of the habitat was within forests, only 33% of all forests were considered suitable as habitat.