Using Crossborder Multisource Burned Area Datasets for Assessing Wildfire Susceptibility Using Machine Learning Techniques ^†

Susceptibility maps constitute a useful tool for wildfire management. They identify propensity for wildfire occurrence based on the intrinsic characteristics of the terrain.

Following a recent methodology, validated at the regional scale [1], the authors present maps of wildfire susceptibility for the winter and summer fire season over the whole MEDSTAR area. The predisposing factors that were considered belong to three categories: climatic data (mean temperatures and precipitation values), geographic data (CORINE land use, topography, slope and aspect) and anthropic data (proximity to roads and settlements, presence of protected areas). The adopted approach used a Machine Learning Technique (Random Forest) to link the predisposing factors with the history of wildfire data. The burned area dataset was obtained by merging data from the French regions of MEDSTAR (Corse and PACA) with data from 1973 to 2020, and the Italian region of Liguria (with data ranging from 1997 to 2019), and retrieving data from the coastal provinces of Tuscany and Sardinia from the official national wildfire dataset (burned area ranging from 2007 to 2018).

The robustness of the modelling framework allowed it to provide good results (in terms of AUC and RMSE indicators) by using open data as predisposing factors, as well as ground-retrieved burned polygons. The MEDSTAR example was developed in parallel with other cross-boundary implementations, using the same odelling framework and similar results, and thus confirming the goodness of the method in transboundary wildfire management. The method is currently being conveyed into a QGIS plugin, to allow for end users to replicate the produced map and/or perform their own experiments with other input data.