Spatial Prediction of Fire Ignition Probabilities: Comparing Logistic Regression and Neural Networks

Abstract

The objective of this work was to develop and validate models to predict spatially distributed probabilities of ignition of wildland fires in central Portugal. The models were constructed by exploring relationships between ignition location/cause and values of geographical and environmental variables using logistic regression and neural networks. The conclusions are that (1) the spatial patterns of fire ignition identified can be used for prediction, (2) the spatial patterns are different for the different causes, (3) the logistic models and the neural networks both reveal acceptable levels of predictive ability but the neural networks present better accuracy and robustness, (4) the maps produced by the two methods are similar, and (5) the information contained in the spatial position of ignition events can be used to gain predictive capability over an important phenomenon that is difficult to characterize and, for that reason, has not been included in most of the currently used fire danger estimation systems.