Using multi variate data mining techniques for estimating fire susceptibility of Tyrolean forests

Abstract

Forest fires in the Alps might become more hazardous in the future due to their linkage to high temperatures and drought periods. To understand the driving forces and the actual distribution patterns of forest fires is crucial to adapt to those future challenges. We used the province Tyrol in Austria as a case study area. We tested two machine learning algorithms (MaxEnt and RandomForests) for their capability in use with historic fire data and environmental datasets. Presence-only data of 399 forest fires between 1993 and 2011 was compared with a range of topography, vegetation, climate and socio-economic datasets, in order to find out the main driving parameters of the spatial forest fire distribution and to delimit areas of particular fire danger. The results of both algorithms were compared for agreement and differences.Both algorithms delivered broadly similar results; climate (number of days above a certain fire weather index) and anthropogenic parameters (distance to buildings, population density) are the most important parameters in the current alpine fire regime. The models slightly disagreed on the role of forest type and of topography. Joined model results show a strong concentration of potential fire danger along the main valleys and in the drier Tyrolean Upland.