A Neural Network Approach for Forestal Fire Risk Estimation

This paper describes an intelligent system for the prediction of forest fire risk in Galicia, a region in north-west Spain. The system has been designed to calculate a risk fire index for each of the 360 squares of 10x10 kms into which the area map has been divided digitally. In our research, the problem was approached using a feedforward neural network. The information used to train the network was gathered at five meteorological stations on a daily basis from 1985 to 1999, and consisted of basically meteorological data, namely temperature, humidity and rainfall, in conjunction with previous fire records for the areas represented by squares. Network topologies were tested using 125,156 training data and validated over 13,906 test samples, and that achieving the best performance was the 6-9-1 topology. Finally, our results indicate that the system performs satisfactorily, with a sensitivity of 0.857 and a specificity of 0.768.