Algeria is one of the Maghreb countries most affected by wildfires. The economic, environmental, and societal consequences of these fires can last several years after the wildfire. Often, it is possible to avoid such disasters if the detection of the outbreak of fire is fast enough, reliable, and early. The lack of datasets has limited the methods used to predict wildfires in Algeria to the mapping risk areas, which is updated annually. This study is the result of the availability of a recent dataset relating the history of forest fires in the cities of Bejaia and Sidi Bel-Abbes during the year 2012. The dataset being small size, we used principal component analysis to reduce the number of variables to 6, while retaining 96.65% of the total variance. Moreover, we developed an artificial neural network (ANN) with two hidden layers to predict wildfires in these cities. Next, we trained and compared the performance of our classifier with those provided by the Logistic Regression, K Nearest Neighbors, Support Vector Machine, and Random Forest classifiers, using a 10-fold stratified cross-validation. The experiment shows a slight superiority of the ANN classifier compared to the others, in terms of accuracy, precision, and recall. Our classifier achieves an accuracy of 0.967±0.026 and F1-score of 0.971±0.023. The SHAP technique revealed the importance of the features (RH, DC, ISI) in the predictions of the ANN model.

Wildfires have changed in recent decades. The catastrophic wildfires make it necessary to have accurate predictive models on a country scale to organize firefighting resources. In Mediterranean countries, the number of wildfires is quite high but they are mainly concentrated around summer months. Because of seasonality, there are territories where the number of fires is zero in some months and is overdispersed in others. Zero-inflated negative binomial mixed models are adapted to this type of data because they can describe patterns that explain both number of fires and their non-occurrence and also provide useful prediction tools. In addition to model-based predictions, a parametric bootstrap method is applied for estimating mean squared errors and constructing prediction intervals. The statistical methodology and developed software are applied to model and to predict number of wildfires in Spain between 2002 and 2015 by provinces and months.