Evaluation of artificial neural network techniques in predicting fire danger

Fire danger can be considered as the combination of a large number of factors related to both the probability of fire ignition (human factors, fuel moisture content, fuel flammability, etc.) and potential fire behaviour in space and time (terrain conditions, fuel characteristics and load, weather factors and wind field in particular). Considering the intrinsic complexity of relationships among large numbers of variables and fires, several works proposed methods, models, and tools that tried to integrate the main explanatory variables, providing predictions of fire activity at short and medium term, on different spatial and temporal scales. From a statistical point of view the relationship among variables can be analysed using different approaches: multiple linear regressions, logistic regressions, other non-linear statistical methods. The former approach can be followed using classical statistical techniques or alternative methods based on inductive processes where the prescription of a basic mathematical form is not required, since the relationships among the inputs and the expected outputs can be established by an interactive process called learning. One increasing way of modelling this kind of process is by using neural networks. The aim of this work is to develop a neural network model in order to study the interactions among environmental factors and human activities affecting the fire danger; spatial and temporal variations of fire danger are also estimated in order to describe the main fire danger features of the Sardinian fire season. The analysis highlights the capabilities of the neural network approach in estimating the level of fire danger, and suggests the large effect of the human factors in determining the fire danger during the fire season.