Modeling spatial patterns of fire occurrence in Mediterranean Europe using Multiple Regression and Random Forest

Fire occurrence, which results from the presence of an ignition source and the conditions for a fire to spread, is an essential component of fire risk assessment. In this paper, we present and compare the results of the application of two different methods to identify the main structural factors that explain the likelihood of fire occurrence at European scale. Data on the number of fires for the countries of the European Mediterranean region during the main fire season (June-September) were obtained from the European Fire Database of the European Forest Fire Information System. Fire density (number of fires/km(2)) was estimated based on interpolation techniques and was used as the dependent variable in the model. As predictors, different physical, socio-economic and demographic variables were selected based on their potential influence in fire occurrence and on their availability at the European level. Two different methods were applied for the analysis: traditional Multiple Linear Regression and Random Forest, the latter being a non-parametric alternative based on an ensemble of classification and regression trees. The predictive ability of the two models, the variables selected by each method and their level of importance were compared and the potential implications to forest management and fire prevention were discussed. The Random Forest model showed a higher predictive ability than Multiple Linear Regression. Furthermore, the analysis of the residuals also indicated a better performance of the Random Forest model, showing that this method has potentiality to be applied in the assessment of fire-related phenomena at a broad scale. Some of the variables selected are common to both models; precipitation and soil moisture seem to influence fire occurrence to a large extent. Unemployment rate, livestock density and density of local roads were also found significant by both methods. Maps of the likelihood of fire occurrence were obtained from each method at 10 km resolution, based on the selected variables. Both models show that the spatial distribution of fire occurrence likelihood is highly variable in this region: highest fire likelihood is prevalent in the northwest region of the Iberian Peninsula and southern Italy, whereas it is low in northern France, northeast Italy and north of Greece. In the most fire-prone areas, preventive measures could be implemented, associated to the factors identified by both models. (C) 2012 Elsevier B.V. All rights reserved.