Wildfire susceptibility mapping via machine learning: the case study of Liguria Region, Italy

Risk and susceptibility analyses for  natural hazards are of great importance for the sake of  civil protection, land use planning  and risk reduction programs. Susceptibility maps are based on the assumption that future events are expected to occur under similar conditions as the observed ones. Each unit area is assessed in term of relative spatial likelihood, evaluating the potential to experience a particular hazard in the future based solely on the intrinsic local characteristics. These concept is well-consolidated in the research area related with the risk assessment, especially for landslides. Nevertheless, the need exist for developing new quantitative and robust methods allowing to elaborate susceptibility  maps and to apply this tool to the study of other natural hazards.  In  the presented work, such  task is pursued for the specific  case of wildfires in Italy. The  two main approaches for such studies are the adoption  of physically based models and the data driven methods. In  the presented work, the latter  approach is  pursued, using  Machine Learning techniques in order to learn  from and make prediction  on the available information (i.e. the observed burned area and the predisposing factors) . Italy is severely affected by wildfires due to the high topographic and vegetation heterogeneity of its territory  and  to  its   meteorological conditions. The present study has as its main objective the  elaboration of a wildfire susceptibility map for Liguria region (Italy) by making use of Random Forest, an ensemble ML algorithm based on decision trees. The quantitative evaluation of susceptibility is carried out considering two different aspects: the location of past  wildfire occurrences, in terms of burned area, and the related anthropogenic and geo-environmental  predisposing factors that may favor fire spread. Different implementation of the model  were performed and compared. In  particular,  the effect of  a pixel's  neighboring land cover (including the type of vegetation and no-burnable area) on the output susceptibility map is investigated. In order to assess the  performance  of the model, the spatial-cross validation has been carried  out, trying  out different  number of folders. Susceptibility maps for the two fire seasons (the  summer  and  the winter  one) were finally computed  and validated. The  resulting  maps show  higher susceptibility zones , developing closer to the coast in summer and along the interior part of  the region in winter. Such zones matched well with the testing burned area, thus  proving the  overall  good performance of the proposed method.

REFERENCE

Tonini M., D’Andrea M., Biondi G., Degli Esposti S.; Fiorucci P., A machine learning based approach for wildfire susceptibility mapping. The case study of Liguria region in Italy. Geosciences (2020, submitted)