Local feature importance of predisposing variables to explain the spatial heterogeneity of wildfires density in the Mediterranean area

In the southern European countries, the combination of climate change, substantial shifts in land use/land cover, and socio-economic factors acting over the last decades are anticipated to increase the frequency, scale, and intensity of wildfires unless enhanced prevention and control strategies are implemented. Statistical and data-driven approaches are widely used by researchers to evaluate the main variables controlling wildfires occurrences and spreading. Lately machine learning proved to be highly performant due to its flexible and non-linear nature, capable of capturing the complexity of the wildfire process. Nevertheless, conventional classification and regression methods like Support Vector Machine, pixel-based Neural Network, and Random Forest (RF), are global modelers, not calibrated to deal with the spatial heterogeneity of the investigated area. Thus, these algorithms turn out to be incapable of adequately addressing the spatially varying underlying relationship between wildfires pattern distribution and the predisposing variables.

While many studies seek to assess the importance of the predictor variables both at regional [1, 2] and at supranational level [3], up to now there is a lack of studies attempting to account for the spatial heterogeneity (i.e. non-stationarity) when modeling wildfires spatial patterns as function of geographical features.  To fill this gap, the present work explores the local feature importance of geographical independent predisposing variables on the spatial distribution of burned area density in the Mediterranean area. To this end, we have used the last development of Geographical Random Forest (GRF) [4], which integrates a parallelizable RF function, a procedure for the bandwidth optimization, and an option to spatially weight the local observations. As dependent variables we considered the percentage of burned pixels per map unit. Both geo-environmental features (i.e., variables providing information on the topography and land cover) and anthropogenic features (e.g., distances to urban areas and road network) have been select as predictors. The importance of these independent variables has been assessed by evaluating the Mean Decrease Accuracy (MDA) by using the Out of Bag samples available in RF: higher values mean that the model strongly benefits from the given variable when performing predictions. The spatial variation of each predisposing factor was illustrated by mapping the corresponding MDA values over the geographical space. Finally, the implemented model has been validated by using the root mean squared error computed over an independent testing dataset.   

[1] Trucchia A, Izadgoshasb H, Isnardi S, Fiorucci P, Tonini M, 2022. Machine-Learning Applications in Geosciences: Comparison of Different Algorithms and Vegetation Classes' Importance Ranking in Wildfire Susceptibility. Geosciences, 12 (11) p. 424. 

[2] Bustillo Sánchez M, Tonini M, Mapelli A, Fiorucci P, 2021. Spatial Assessment of Wildfires Susceptibility in Santa Cruz (Bolivia) Using Random Forest. Geosciences, 11 (5) p. 224. 

[3] Trucchia A, Meschi G, Fiorucci P, Provenzale A, Tonini M, Pernice U, 2023. Wildfire hazard mapping in the eastern Mediterranean landscape. International Journal of Wildland Fire. 32, 417-434. 

[4] Georganos S, Kalogirou S, 2022. A Forest of Forests: A Spatially Weighted and Computationally Efficient Formulation of Geographical Random Forests. ISPRS Int. J. Geo-Inf. 2022, 11, 471.