Comparison of Different Algorithms and Vegetation Classes’ Importance Ranking in Wildfire Susceptibility Maps.

Comparison of Different Algorithms and Vegetation Classes’ Importance Ranking in Wildfire Susceptibility Maps. 
Wildfire Susceptibility Maps (WSM) and the analysis of the explanatory variables affecting the model’s predictions are innovative tools to support forest protection and management plans. Namely, WSM identify areas subject to wildfire, in terms of relative spatial likelihood, on the base of the observed past events, stored in spatio-temporal inventories, and on the local environmental and anthropogenic properties of an area. Approaches based on Machine Learning (ML) are particularly suited for WSM since they are capable to make predictions on data by modelling the hidden and non-linear relationships between a set of input variables and the output observations.
In the present work, Authors continue a research framework developed at local scale for Liguria Region, and lately improved at national scale (Italy), consisting in the implementation of a ML-approach, based on the algorithm Random Forest, allowing to assess the susceptibility to wildfires under the influence of different variables (e.g., land cover, vegetation classes, altitude and its derivatives, nearby infrastructures). In the present study the following improvements are introduced: (i) to evaluate which ML-algorithm performs better in terms of prediction capabilities we compared Random Forest (RF), Multi-layer Perceptron (MLP), and Support Vector Machine (SVM); (ii) to evaluate the impact of different classes of local and neighbouring vegetation on wildfires occurrence we used of a more accurate map of vegetation as input local explanatory variable; (iii) to consider both the spatial and the temporal variability of the burning seasons (summer and winter) we improved the selection of the testing dataset, based on a clustering approach. 
The output probabilistic predicted values resulting from the different ML-algorithms (RF, MLP, and SVM) allowed to elaborate the seasonal WSMs. Finally, the spatial distribution of the more susceptible areas will be presented. The performance of the three ML-algorithms was assessed by means of the AUC (Area Under the Curve) ROC (Receiver Operating Characteristics), evaluated over the testing dataset. In addition, the variable importance ranking was estimated as by-product of RF, which can handle both the typical numerical variables and native categorical variables (as for the classes of vegetation at pixel level). Vegetation resulted by far to be the most important explanatory variables; the marginal effect of each single class of vegetation was also assessed and the results will be discussed. 
Reference 
Trucchia, A.; Izadgoshasb, H.; Isnardi, S.; Fiorucci, P.; Tonini, M. Machine-Learning Applications in Geosciences: Comparison of Different Algorithms and Vegetation Classes’ Importance Ranking in Wildfire Susceptibility. Geosciences 2022, 12, 424. https://doi.org/10.3390/geosciences12110424