Evaluation of Machine Learning Models to Predict the Probability of Forest Fires with Small Training Sample, Case of the Wilaya of Sidi Belabbes

The aim of this study is to evaluate machine learning models for predicting forest fire probability, based on 14 topographical, meteorological, anthropological and vegetation factors. A history of forest fires in the wilaya of Sidi Bel Abbes was used as a learning base, comprising 159 fire points. Eight models were tested, including Support Vector Machine (SVM), Random Forest (RF), Logistic Regression (LR), HistGradient Boosting Regressor (HGBR), Ada Boost Classifier (ABC), Gradient Boosting Classifier (GBC), Neural Network (NN), Gaussian Process Classifier (GPC). The Random Forest model posted the highest accuracy score, reaching 86.46%. The most influential factors in prediction were the median temperature and the Canadian Fire weather index (FWI).