Fire risk modeling: an integrated and data-driven approach applied to Sicily

. Wildﬁres are key not only to landscape transformation and vegetation succession, but also to socio-ecological values loss. Fire risk mapping can help to manage the most vulnerable and relevant ecosystems impacted by wildﬁres. However, few studies provide accessible daily dynamic results at different spatio-temporal scales. We develop a ﬁre risk model for Sicily (Italy), an iconic case of the Mediterranean Basin, integrating a ﬁre hazard model with an exposure and vulnerability analysis under present and future conditions. The integrated model is data-driven but can run dynamically at a daily time step, providing spatially and temporally explicit results through the k.LAB (Knowledge Laboratory) software. This software provides an environment for input data integration, combining methods and data such as geographic information systems, remote sensing and Bayesian network algorithms. All data and models are semantically annotated, open and downloadable in agreement with the FAIR principles (ﬁndable, accessible, interoperable and reusable). The ﬁre risk analysis reveals that 45 % of vulnerable areas of Sicily have a high probability of ﬁre occurrence in 2050. The risk model outputs also include qualitative risk indexes, which can make the results more under-standable for non-technical stakeholders. We argue that this approach is well suited to aiding in landscape and ﬁre risk management, under both current and climate change conditions.


Supplement of
Fire risk modeling: an integrated and data-driven approach applied to Sicily Alba Marquez Torres et al.
Correspondence to: Alba Marquez Torres (alba.marquez@bc3research.org) The copyright of individual parts of the supplement might differ from the article licence.Ground fuels (cover >50%) Grass 2 Surface fuels (shrub cover >60%; tree cover <50%) Grassland, shrubland (smaller than 0.3-0.6mand with a high percentage of grassland), and clear-cuts, where slash was not removed.

4
Tall shrubs (shrub cover >60%; tree cover <50%) High shrubs (between 2.0 and 4.0 m) and young trees resulting from natural regeneration or forestation 5 Tree stands (>4m) with a clean ground surface (shrub cover <30%) The ground fuel was removed either by prescribed burning or by mechanical means.This situation may also occur in closed canopies in which the lack of sunlight inhibits the growth of surface vegetation 6 Tree stands (>4m) with medium surface fuels (shrubs cover >30%) The base of the canopies is well above the surface fuel layer (>0.5).The fuel consists essentially of small shrubs, grass, litter, and duff.

7
Tree stand (>4m) with heavy surface fuels (shrub cover >30%) Stands with a very dense surface fuel layer and with a very small vertical gap to the canopy base (<0.5m)

S3
We used the most appropriate discretization method, mostly according to the data distribution of each variable and by trial and error.However, factors to be considered include the shape and spread of the data, the purpose, and level of detail of the analysis, as well as the number and size of bins.The optimal number and size of bins depends on a trade-off between information loss and information gain.
In general, equal-width binning was applied to more uniformly distributed input data as for atmospheric temperature, maximum weekly atmospheric temperature, and solar radiation.For skewed distributions as for elevation, number of days without precipitation, slope, distance to protected area, distance to road, and distance to human settlement, we used Equal-frequency binning.The disadvantage of equal-frequency is that it can distort the distribution of the data and create irregular bin widths.That was the case with the "weekly precipitation" variable.After several tests, we realized that the equalfrequency produced a wrong data binning, this is the reason why we apply equal frequency in spite of its skewed distribution.

S5
Table S3.Area (km 2 ) of low, medium, or high ES (Ecosystem Services) value potentially exposed to fire and the percentage of change in area.

S1Figure
Figure S1.Global workflow of fire risk model.

Figure S3 .
Figure S3.Data distribution of atmospheric temperature, maximum weekly atmospheric temperature and solar radiation variables.

Figure S4 .
Figure S4.Data distribution of elevation, day count without precipitation, slope, distance protected area, distance to road and distance to human settlement variables.

Figure S5 .
Figure S5.Data distribution of weekly precipitation volume variables.

Figure S6 .
Figure S6.Uncertainty map of fire hazard model: standard deviation of the probability distributions simulated by the model ranges from 0 (blue) to 0.5 (red).