Logistic regression models for human-caused wildfire risk estimation: analysing the effect of the spatial accuracy in fire occurrence data

Abstract

About 90% of the wildland fires occurred in Southern Europe are caused by human activities. In spite of these figures, the human factor hardly ever appears in the definition of operational fire risk systems due to the difficulty of characterising it. This paper describes two spatially explicit models that predict the probability of fire occurrence due to human causes for their integration into a comprehensive fire risk–mapping methodology. A logistic regression technique at 1 × 1 km grid resolution has been used to obtain these models in the region of Madrid, a highly populated area in the centre of Spain. Socio-economic data were used as predictive variables to spatially represent anthropogenic factors related to fire risk. Historical fire occurrence from 2000 to 2005 was used as the response variable. In order to analyse the effects of the spatial accuracy of the response variable on the model performance (significant variables and classification accuracy), two different models were defined. In the first model, fire ignition points (x, y coordinates) were used as response variable. This model was compared with another one (Kernel model) where the response variable was the density of ignition points and was obtained through a kernel density interpolation technique from fire ignition points randomly located within a 10 × 10 km grid, which is the standard spatial reference unit established by the Spanish Ministry of Environment, Rural and Marine Affairs to report fire location in the national official statistics. Validation of both models was accomplished using an independent set of fire ignition points (years 2006–2007). For the validation, we used the area under the curve (AUC) obtained by a receiver-operating system. The first model performs slightly better with a value of AUC of 0.70 as opposed to 0.67 for the Kernel model. Wildland–urban interface was selected by both models with high relative importance.