Wildfire prediction using machine learning

Abstract

The use of supervised Machine Learning algorithms is widespread in the science of fires. The objective of this postgraduate thesis was to conduct three experiments utilizing only weather variables for the region of the Attica basin. More specifically, the prediction of the probability of fire occurrence (binary classification) for 12, 4 and 2 weather variables respectively, was implemented as first experiment, the prediction of the fire scale (multi-class classification: small fire, medium fire, large fire, wildfire) for 12 weather variables as second experiment and the prediction of the size of the burned area of forest fires for 12 and 4 weather variables as third experiment (regression task). Initially, a new dataset named “wildfire” was synthesized that included the prevailing weather conditions during the forest fires occurrences in the Attica basin. Based on this, an attempt was made to conduct the three experiments with the resulting predictions proving to be particularly impressive. The performance of the formed wildfire dataset was compared with the known prior art Montesinho dataset in order to evaluate which of the two functioned best in the application of supervised Machine Learning algorithms. The comparative results showed that for all 12 weather variables extracted by the wildfire dataset, a tuned Random Forest model (70%) outperformed other classification models regarding prediction accuracy of fire occurrence. In alternative embodiments for the best 4 and 2 selected weather features correspondingly the Extreme Gradient Boosting (XGBoost) prediction model achieved the best accuracy (67.4%) in terms of fire occurrence prediction and the Neural Networks performed marginally better (63.6%) than the Random Forest (63.3%). As for the problem of multi-class classification of fire scale prediction (small fire, medium fire, large fire, wildfire), it demonstrated that the model of the K- nearest neighbors implemented better (50%) than the other prediction models. The findings for forecasting of size of burned area of forest fires turned out that by using all the weather variables the K-nearest neighbors (r² score value 70%) outperformed other regression models while for 4 chosen weather features poor outcomes were provided by regression models with only the Linear Regression algorithm to carry out better than others (r² score value 2%). Finally, a comparison was made with the known prior art Montesinho dataset for 4 and 2 selected weather variables for the first experiment, as well as for 4 weather variables for the third experiment. The results showed that the newly created wildfire dataset functioned much better when applying the supervised Machine Learning algorithms.