A 20m Pan-European Wildfire 12-class Fuel Map

Fuel map 12 classes methodology

Input susceptibility map:
A Pan-European susceptibility is generated by training a Machine Learning model based on the Random Forest algorithm using topographic, land cover, climate data, and fire presence/absence labels. The model infers the wildfire susceptibility for each geo-spatial pixel of the domain.

 

the native map resolution is at 100m and it has been resampled to 20m using bilinear interpolation.

Susceptibility map is categorized in 3 classes using as thresholds the 1st and 10th percentiles in the burned area distribution from 2008 to 2022, (retreived from the European Forest Fires Information System).

1 = Low
2 = Medium
3 = High

Input vegetation maps:

a) CLCplus backbone for European domain (10m resampled to 20 with nearest method) 
b) Global Copernicus land cover 100m for areas outside Eu (N Africa, middle east, Ukraine) - resample to 20m (nearest neighbor)
c) Global Copernicus land cover 10m for areas outside Eu (N Africa, middle east, Ukraine) - resample to 20m (nearest neighbor)

the 3 maps have been aggregated in a 4 type macro-veg classes:

0 = not burnable
1 = Grasslands and croplands 
2 = Broadleaves forest
3 = Shrublands
4 = Conifers forests

Fuel map generation:

the resulted Pan-EU map is built with the following logic:

for area outside Europe, the map at the 10m 'wins' always besides when tree forest is present.
In the forest class (2), the 100m map override with class 2 or 4. (Because the 100m map has broadleaves/conifer discrimination).
In the European domain, the CLCplus backbone always wins. 

the merged veg map is then combined with the categorized susceptibility map using the 12-class contingency matrix.

 FT 1 2 3 4
S
1   1 4 7 10
2   2 5 8 11
3   3 6 9 12

the dataset is present in HDF5 (.h5) format (lightweight) and a script (.py) for converting it in Geotiff with COG optimization is provided. The file size of the Geotiff can sensibly increase during the convertion.