Wildfire Risk Assessment under present and future climate at national scale: a pan european approach

The authors present a framework designed to model wildfire risk and the future projection of wildfire risk patterns, also in view of climate change scenarios. The adopted modeling framework is inherently multi scale, giving results at national scale, after a data gathering process developed at regional / supranational scale. The risk assessment comprises the computation of susceptibility, hazard, exposures, and damage layers. Machine learning techniques are used to assess the wildfire susceptibility and hazard at regional level, analogously to [1, 2]. To this end, a two-models approach has been adopted. The first model, based on the Random Forest Classifier, is trained at pan-European level to capture the climate variability of the European continent and related fire regimes. Building upon the outcome of this model, a wildfire susceptibility map representative for the historical
conditions at pan-European level is produced and used in input of a second machine learning model, to provide results at national level. The strength of this model lies in using high-resolution downscaled climate data and annual temporal resolution, with the objective of computing a high resolution annual susceptibility map for the specific region. This approach facilitates the generation of annual outcomes for both historical and future conditions, using the climate projections available in the ISIMIP framework. The result of five different climate models and three climate change scenarios have been used to estimate the average annual losses due to wildfires. The wildfire hazard has been evaluated through empirical approaches, building a wildfire hazard classes map combining fuel type/severity maps with wildfire susceptibility. Then, a burning probability is estimated for each hazard class: a statistical analysis on historical wildfires at pan-European level has been performed in order to retrieve the annual relative burned area per hazard class. The method allows to estimate the average annual probability to be affected by a fire given a wildfire event. Several exposed elements were used to estimate the losses ranging from infrastructure to forest and roads: Global Earthquake Model [3] provides a dataset featuring economic values under both present and future conditions across five categories of infrastructures at European level. JRC, OpenStreeMap, and Copernicus provide information on the presence of roads and forests. Empirical vulnerability functions establish a link between severity maps, the presence of exposed elements, and their economic value, leading to the estimation of potential damage maps. The assessment of average annual losses involves coupling spatial information on average annual probability with potential damage maps. This approach allows for the evaluation of average values across various future timeframes associating a variance accounting for both the year to year and climate models’ variability. Results have been produced at national level for several countries characterized by different wildfire regimes, land cover and climate, such as Croatia, Romania and Bulgaria.