ANN multivariate analysis of factors that influence human-caused multiple fire starts

Abstract

Delays in the initial attack of new fire starts can happen locally when two or more fires burn simultaneously. The occurrence of multiple-fire-day situations may pose a real problem if suppression resources are limited, which almost always are. We analyzed multiple-fire-days in Galicia (Spain) from 2002 to 2005 with the goal of predicting these multiple fire situations by using Artificial Neural Networks. We carried out two types of analysis with our seasonally-structured data: to identify the relevant variables in the multiple versus single daily outcome (classification problem) and to predict the number of fires within the multiple-fire-days observations (prediction problem). The accuracy for the best Spring model was around 59-60\% which located multiple occurrences in higher altitudes and public forest properties, near roads and recreation areas, with lower temperatures, lower quantity of pastureland and higher FFMC. Best classifications for the Summer period were around 60-61\% and associated multiple fires to lower elevation areas, higher proportion of public and communal forests, near roads and higher drought indices. Predictions of actual number of fire occurrences in the Spring period reached 62\% accuracy with a similar variables selection as the Spring classification model. Predictions for the Summer period lacked accuracy (44-50\%) suggesting more complex patterns, probably due to mixed causes.