Machine-Learning Applications in Geosciences: Comparison of Different Algorithms and Vegetation Classes’ Importance Ranking in Wildfire Susceptibility

Abstract

Susceptibility mapping represents a modern tool to support forest protection plans and to address fuel management. With the present work, we continue with a research framework developed in a pioneristic study at the local scale for Liguria (Italy) and recently adapted to the national scale. In these previous works, a random-forest-based modeling workflow was developed to assess susceptibility to wildfires under the influence of a number of environmental predictors. The main novelties and contributions of the present study are: (i) we compared models based on random forest, multi-layer perceptron, and support vector machine, to estimate their prediction capabilities; (ii) we used a more accurate vegetation map as predictor, allowing us to evaluate the impacts of different types of local and neighboring vegetation on wildfires’ occurrence; (iii) we improved the selection of the testing dataset, in order to take into account the temporal variability of the burning seasons. Wildfire susceptibility maps were finally created based on the output probabilistic predicted values from the three machine-learning algorithms. As revealed with random forest, vegetation is so far the most important predictor variable; the marginal effect of each type of vegetation was then evaluated and discussed.