Enhanced forest fire susceptibility mapping by integrating feature selection genetic algorithm and bagging-based support vector machine with artificial neural networks

Forest fire is a natural disaster that threatens a large part of the world's forests. Considering the destructive effects of forest fires, the preparation of forest fire probability maps can be a very valuable step towards reducing such effects. This study proposes two novel wrapper feature selection-based ensemble models that combine the strengths of Support vector machine (SVM) and Artificial neural networks (ANN) with bagging (bootstrap aggregating) and Genetic Algorithm (GA) for forest fire susceptibility mapping in the Jerash and Ajloun provinces of Jordan. By integrating multiple learning algorithms through ensemble methods, we aim to increase predictive accuracy and enhance the robustness of our findings. GA was employed for feature selection utilizing data from 207 forest fire locations and fourteen predictor variables. 70% of the forest fire locations (145 locations) were used in the training phase, and the remaining 60% (62 locations) were employed to validate the models. The accuracy of the models was measured by using the area Under the Receiver Operating Characteristic (AUROC). The AUROC for single SVM, single ANN, GBSVM, and GBANN models was 69.3%, 66.9%, 70.9%, and 70.4% in the validation phase, respectively. The results showed that wrapper and bagging-based ensemble models did much better than single models. This shows that combining techniques can improve modeling performance for mapping the risk of forest fires.