The functional approach to calculate the three indices of wildfire ignition
was performed through the use of a multi-layer perceptron
(MLP) that had been trained with the back-propagation algorithm
(Rumelhart and McClelland, 1986). All three networks had one hidden
layer with six hidden nodes for the FWI, four hidden nodes for the FHI
and eight hidden nodes for the FRI (Vasilakos et al., 2007). The logistic
function was used for the activation between inputs and hidden layer;
and the identity function was used for the activation of the output. A
learning rate of r = 0.1 was chosen, whereas the output neuron was
considered activated in case of an output value above 0.5. By taking
into consideration the mean square error in a training and a validation
dataset, best results were achieved after 100 epochs for the FWI, 1500
epochs for the FHI and 1000 epochs for the FRI (Vasilakos et al., 2007).
It should be noted also that the parameters have been chosen to reflect
the wildland fire ignition pattern based on the fire ignition causes of our
study area. Hence, the model itself is applicable to another study area
only if it will be trained with the fire history data of the new area. The
output Fire Ignition Index map (Fig. 4) portrays the geographical probability
of a new fire ignition, classified into 5 categories: low risk (0–40);
medium risk (41–60); high risk (61–80); very high risk (81–90); and
alarm risk status (91–100). The system is designed to function in an automated
mode daily at 10:00 am local time, in the same way used for
the creation of the weather maps. The SKIRON model provides all the
necessary forecasted weather prediction data to be used as inputs.
Fire ignition and weather prediction maps in Virtual Fire are produced
daily. However, the processing time needed for this creation
was considerably high, due to large input datasets. To confront this