Classification and mapping of European fuels using a hierarchical-multipurpose fuel classification system

. Accurate and spatially explicit information on forest fuels becomes essential to designing an integrated fire risk management strategy, as fuel characteristics are critical for fire danger estimation, fire propagation and 15 emissions modelling, among other aspects. This paper presents the conceptual development of a new fuel classification system that can be adapted to different spatial scales and used for different purposes. The resulting fuel classification system encompasses a total of 85 fuel types, that can be grouped into six main fuel categories (forest, shrubland, grassland, cropland, wet and peat/semi-peat land and urban), plus a nonfuel category. For the forest cover, fuel types include two vertical strata, overstory and understory, to account for both surface and crown 20 fires. Based on this classification system, a European fuel map at 1 km resolution, was developed within the framework of the FirEUrisk project, which aims to create a European integrated strategy for fire danger assessment, reduction, and adaptation. Fuels were mapped using land cover and biogeographic datasets, as well as bioclimatic modelling, in a Geographic Information System environment. The first assessment of this map was performed by comparing it to high-resolution data, including LUCAS (Land Use and Coverage Area frame Survey) 25 data, Google Earth images, Google Street View images, and the GlobeLand30 map. This validation exercise provided an overall accuracy of 88 % for the main fuel types, and 81 % for all mapped fuel types. Finally, to facilitate the use of this fuel dataset in fire behaviour modelling, a first assignment of fuel parameters to each fuel type was performed by developing a crosswalk to the standard fuel models defined by Scott and Burgan (FBFM, Fire Behavior Fuel Models)


80
Although they have been developed for specific regions and conditions, they have been widely used to map fuel types in other regions (García et al., 2011;Palaiologou et al., 2013;Marino et al., 2016;Aragoneses and Chuvieco, 2021).
The only European-level fuel cartography is the 2000 EFFIS fuel map (European Forest Fire Information System (EFFIS), 2017), based on land cover and vegetation maps and using the NFFL system. Other works have mapped 95 FBFM fuel models (Scott and Burgan, 2005) for the European subcontinental scale, such as the Iberian Peninsula (Aragoneses and Chuvieco, 2021).
The lack of an adapted-to-Europe fuel classification strategy is limiting since fuel models are site-specific and should be applied to the region for which they were developed to obtain the most realistic fuel mapping and modelling (Arroyo et al., 2008). In this context, the ArcFuel project (Bonazountas et al., 2014) aimed in 2011-100 2013 to conceive a methodology to enable consistent fuel mapping production over Europe to support fire and emissions simulation scenarios, and the design of effective fire prevention and mitigation strategies. For this, it constructed a hierarchical vegetation fuel classification system adapted to Europe (Toukiloglou et al., 2013).
Nevertheless, fuel cartography was only created for southern European national (Portugal and Greece), and regional (Spain and Italy) scales and no European fuel map was generated (Bonazountas et al., 2014).

105
Considering the actual limitations of European fuel mapping, we aimed to three objectives. The first one was generating a European advanced fuel classification system to facilitate the integration of continental wildfire risk assessment, including both surface and canopy fuels. The new classification system should be hierarchical to facilitate scale integration, include both surface and canopy fuel types and be suitable for different purposes, from fire behaviour simulation to fire emissions or fire danger assessment. The second objective was to develop a 110 European fuel map at 1 km spatial resolution following the proposed fuel classification system. We aimed to develop a methodology that, combining expert knowledge, Geographic Information Systems (GIS), available datasets, and bioclimatic modelling, might be easily replicable and updated with low time and economic costs.
Finally, the third objective was to assign fuel parameters to the derived fuel types, by relating them to existing fuel models. We chose the FBFM standard fuel models (Scott and Burgan, 2005), as this system is widely used and 115 very flexible. These three objectives (development of a fuel classification system, generation of a European fuel map and fuel parameterization) serve to organise the structure of this paper around three sections (Fig. 1). This work is expected to lay the framework for an integrated and homogeneous fire management strategy across https://doi.org/10.5194/essd-2022-184 Preprint. Discussion started: 25 July 2022 c Author(s) 2022. CC BY 4.0 License.
European countries. The present study is part of the FirEUrisk project, which aims to create a European integrated strategy for fire danger assessment, reduction, and adaptation.

Development of the fuel classification system
We developed the FirEUrisk hierarchical fuel classification system with three main requirements: it should be adapted to the great variety of European environmental conditions, include both surface and canopy fuels, and be suitable to work at different spatial scales. The main driver of the classification system was fire behaviour modelling, but its use for fire risk assessment and fire emission estimations was considered as well. To  (UNESCO, 1973) and the FAO (Food and Agriculture Organization, 2000); and the European Environmental Agency's Corine Land Cover nomenclature (CLC) (Kosztra et al., 2019) were used. In addition to the mentioned sources, for the wet and peat/semi-peat land fuel types, the definitions provided by the International Peatland 135 Society (International Peatland Society, 2021) were also taken into account.
The FirEUrisk hierarchical fuel classification system used several criteria to discriminate fuel types. First, the main fuel cover, which differentiated six main fuel types: forest, shrubland, grassland, cropland, wet and peat/semi-peat land, and urban fuel types, plus a nonfuel category. For the forest fuel types, two vertical strata were identified: the first-level referred to the overstory (canopy) characteristics, and the second-level to the 140 understory characteristics. The former included three additional splitting criteria: leaf type (broadleaf/needleleaf), phenology (evergreen/deciduous), and fractional cover (open/closed). The latter included two aspects: understory type (grassland/shrubland/timber litter), and understory depth, that is, the height of the understory layer. For the https://doi.org/10.5194/essd-2022-184 Preprint. Discussion started: 25 July 2022 c Author(s) 2022. CC BY 4.0 License. rest of the main fuel types, only one vertical stratum (first-level) was identified. For shrubland and grassland fuel types, subcategories were created based on fuel depth (height of the vegetation layer). For cropland, the split was 145 based on cropland type (herbaceous/woody). For wet and peat/semi-peat land fuel subcategories, tree, shrubland and grassland formations were distinguished. For urban fuel types, the standard CLC division between continuous and discontinuous fabric was followed. For the nonfuel category, we distinguished water, snow, and ice; and bare soil and sparse vegetation, for high spatial resolutions.

The FirEUrisk hierarchical fuel classification system
The proposed hierarchical fuel classification system, FirEUrisk (Table A1 in Appendix A), encompassed a total of 85 fuel types for surface and canopy fuels, which were aggregated into six main fuel type categories, referred to the main fuel cover, which recall traditional land cover types, plus a nonfuel category. They were defined as follows:

155
• Forest: the tree cover is ≥ 15 % with a mean tree height ≥ 2 m. Understory type refers to the fuel type in which the surface fire will spread in the forest.
• Wet and peat/semi-peat land: it includes 1) Wetland: a permanent mixture of vegetation and water (salt, brackish, or fresh), including marshes; 2) Moorland/heathland: low and closed vegetation cover dominated by bushes, shrubs, dwarf shrubs and herbaceous plants, in a climax stage of development, including wet heath on humid or semi-peaty soils (peat depth < 30 cm), herbaceous vegetation, shrubs, 165 and trees of dwarf growth < 3 m; 3) Peatland and peat bog: terrestrial wetlands in which flooded conditions prevent vegetation material from fully decomposing, which results in accumulation of decomposed vegetation matter and moss (peat), including valley, raised, blanket and quacking (floating) bogs with > 30 cm of peat layer, and mosses and herbaceous or woody plants within natural or exploited peat bogs; and 4) Moss and lichen.
Fuel types subcategories were included to better estimate fuel models for each resulting fuel type category and would also lead to different fire behaviour. As previously indicated, two vertical strata were identified. The first-level identified the main vegetation cover, except for the forest fuel types, where it refers to the crown 175 characteristics. The second-level referred to the understory and only applied to forest fuels. For the nonfuel categories, water/snow/ice and bare soil/sparse vegetation were also discriminated for the second-level.
Discriminating all these subcategories may be quite challenging and should be adapted to the working scale of the fuel type product and, accordingly, to the quality of the input data available to produce it. The fuel type categories of the first-level (Table 1) should be more suitable for continental or global fuel products, while the second-level 180 should be better adapted to local or regional studies, where more detailed information can be available. In this paper, the European fuel type dataset was based on the first-level of the classification hierarchy. The area includes 33 countries and covers around 5 Mkm 2 of land. The spatial resolution for the target area is 1 km. This product https://doi.org/10.5194/essd-2022-184 Preprint.

Input data
The generation of the European fuel map with the targeted first-level fuel types (Table 1) was based on the combination of existing land cover and biogeographic regions datasets covering European territory and 205 bioclimatic models.
Due to the similarity between the fuel types of the FirEUrisk fuel classification system and the 2019 discrete Copernicus Global Land Cover map (Copernicus GLC map) legend (Buchhorn et al., 2020), this land cover dataset was used as base cartography for the generation of the European fuel map. The Copernicus GLC map has 100 m resolution and is based on PROBA-Vegetation (PROBA-V) sensor (Buchhorn et al., 2020) with 210 an overall accuracy of 79.9 % for continental land cover categories and 72.8 % for regional land cover categories over Europe . Whenever the land cover information of the Copernicus GLC map was insufficient to map a FirEUrisk fuel type, we used the three following input datasets to derive the required The Copernicus GLC map (Buchhorn et al., 2020) and the Built-up fraction cover Copernicus GLC map (Buchhorn et al., 2020) were downloaded in tiles for the study area and mosaicked. All input datasets were reprojected to ETRS89 Lambert Azimuthal Equal Area using the nearest neighbour method and with the same spatial resolution as the Copernicus GLC map. The input datasets were also clipped to the study area.

230
Also, to account for fuel depth categories (low, medium, and high shrubland and grassland fuel types), we used datasets and bioclimatic models (Saglam et al., 2008;Smit et al., 2008;Fick and Hijmans, 2017;Bohlman et al., 2018;Zhang et al., 2018a) to relate environmental conditions with fuel depth.  (Zepner et al., 2020). The biogeographic regions whose climate graphs presented at least one dry summer month were assigned to the arid/semi-arid regime. A dry summer month is interpreted as a month whose sum of monthly precipitation (mm year -1 ) is less than twice the mean month temperature (ºC) (Zepner 240 et al., 2020). The biogeographic regions not meeting this condition were assigned to the sub-humid/humid regime.
The final general bioclimatic regimes were rasterized to 100 m and 1 km resolution using the maximum area method.

245
Methods to generate the European fuel map are summarised in Fig. 3.  Information on the leaf type, phenology, and fractional cover of forest fuels was obtained from the Copernicus GLC map (Buchhorn et al., 2020). This dataset defines all the first-level forest fuel types in the

B) Shrubland fuel types
The shrubland cover was extracted from the Copernicus GLC map (Buchhorn et al., 2020). To our knowledge, no global or European datasets on shrubland fuel depth, which is the height of the shrubland layer, are available. This variable is quite important, as shrubland depth is directly related to shrubland productivity (Radloff 265 and Mucina, 2007;Saglam et al., 2008;Ali et al., 2015), which is mainly determined by the Mean Annual Precipitation (MAP) (Shoshany and Karnibad, 2015;Paradis et al., 2016;Bohlman et al., 2018;Zhang et al., 2018b) through biomass accumulation (Keeley and Keeley, 1977;Schlesinger and Gill, 1980;Gray and Schlesinger, 1981;Bohlman et al., 2018). This is especially relevant in the arid/semi-arid regime, like the Mediterranean (Shoshany and Karnibad, 2011). Therefore, shrubland fuel depth was obtained from a bioclimatic 270 model adapted to arid/semi-arid conditions with three steps: first, mapping European MAP; second, estimating shrubland productivity from MAP; and third, estimating shrubland fuel depth from productivity.
Global 1970-2000 MAP at 1 km resolution was downloaded from WorldClim 2 dataset (Fick and Hijmans, 2017). The data were reprojected from WGS84 Geographic latitude/longitude to ETRS89 Lambert Azimuthal Equal Area using the bilinear method and clipped using the European shrubland mask.

275
The estimation of shrubland productivity was based on a linear model (Eq. 1) that related shrubland productivity and MAP for California (Bohlman et al., 2018). This model was derived from a literature review, and Californian bioclimatic conditions are similar to those of European arid/semi-arid regions, as can be checked in the ClimateCharts.net platform (Zepner et al., 2020). Therefore, it was used to calculate the mean potential shrubland productivity for each pixel.

280
Biomass (g m -2 ) = 9.6696 MAP (mm year -1 ) -1301.7 (1) Finally, we used a linear empirical model (Eq. 2) that related shrubland fuel depth and productivity for two study areas in Turkey (Saglam et al., 2008) that are similar to European conditions: 650 and 1200 mm year -1 285 mean precipitation. We applied this model to estimate shrubland fuel depth, constraining the outputs to the [0-6] m range. Last, each shrubland fuel depth pixel was assigned to its corresponding shrubland group of the FirEUrisk mainly the MAP: regions with more precipitation have higher grasslands with higher productivity (Smit et al., 2008;Nunez, 2019;Neal, 2021). The most productive grasslands are located in central Europe, while lower grasslands are located in the Mediterranean and Arctic regions (Smit et al., 2008). Information on the grassland fuel depth was obtained from a bioclimatic model with two steps: first, mapping grassland productivity, and 300 second, estimating grassland fuel depth from productivity.
First, European grassland productivity was derived from the consistent inventory of regional statistics (Smit et al., 2008) for the European environmental zones (Metzger et al., 2005), similar to the European biogeographic regions. The mean grassland productivity values were assigned to each polygon of the biogeographic regions' map and were subsequently rasterized using the maximum area method to 100 m 305 resolution, representing the European mean grassland productivity by biogeographic region. The map was then clipped by the grassland mask to obtain this information for the grassland pixels.
Second, to estimate European grassland fuel depth, we used a linear empirical model (Eq. 3) that relates grassland depth and biomass for China (Zhang et al., 2018a). We considered this model appropriate for Europe because Chinese grasslands are also generally temperate and the model was developed considering three study 310 areas that relate to European conditions: 1) 80-220, 2) 600, and 3) 850-1000 mm year -1 mean precipitation. With this model, we estimated grassland fuel depth for every pixel. Finally, each pixel was assigned to a FirEUrisk grassland group according to fuel depth. Outliers (pixels with < 0 m) were reclassified to 0 m.

D) Cropland fuel types
The herbaceous cropland cover was extracted from the Copernicus GLC map (Buchhorn et al., 2020), as this dataset only has information on this type of cropland cover. The CLC map (European Union Copernicus Land Monitoring Service, 2018) was overlapped with the previous map to extract the location of the woody cropland 320 pixels (CLC categories: 221, 222, 223).

E) Wet and peat/semi-peat land fuel types
The Copernicus GLC map (Buchhorn et al., 2020) was used to extract the location of the wetlandherbaceous cover, as this dataset only has information on this type of wetland cover; and the moss and lichens 325 cover. These categories were assigned to the grassland wet and peat/semi-peat land fuel type. Then, the CLC map (European Union Copernicus Land Monitoring Service, 2018) was used to extract the pixels of the peatland and moorland/heathland categories (CLC categories: 322, 412). These pixels were overlapped with the Copernicus GLC map to classify them into tree, shrubland or grassland wet and peat/semi-peat land fuel types, according to the cover type from the Copernicus GLC map they overlapped.

F) Urban fuel types and nonfuel types
The Built-up fraction cover Copernicus GLC map (Buchhorn et al., 2020) was used to extract the location of the pixels with ≥ 15 % and ≥ 80 % of urban cover. Pixels with ≥ 80 % of urban cover were assigned to urban continuous fabric and the rest of the identified urban pixels were assigned to urban discontinuous fabric. The permanent water bodies, open sea, snow and ice, and bare/sparse vegetation (< 10 %) categories from the Copernicus GLC map (Buchhorn et al., 2020) were reclassified to the nonfuel category.

Resampling to the target spatial resolution
After obtaining the first draft of the fuel type dataset at 100 m resolution, it was reprojected to the target 340 spatial resolution of 1 km. Before resampling, potential noise in the cross-tabulation process was minimised by using a majority filter. We performed filtering tests using 3 x 3, 5 x 5, and 7 x 7 moving windows and chose the most suitable according to a balance between information preservation and noise removal.
Resampling was carried out using a custom method that accounts for the spatial heterogeneity of European landscapes. First, the dominant categories within each 1 km 2 pixel were estimated by computing the 345 frequency of the fuel type categories within the 10 x 10 pixels contained in each 1 km 2 . The base resampling criterion was to choose the dominant (first-mode) category within the target pixel. However, to produce a more accurate resampled fuel map that considered complex landscape covers (e.g., mixed forest) and the most dangerous fuel type between two similar co-dominant fuel types; the combination of categories in Table 2 was performed whenever there were two co-dominant categories with the same frequency in a group of 10 x 10 pixels, or the 350 frequency of the co-dominant (second-mode) category was higher than half the frequency of the dominant category (first-mode). The combination of the categories in Table 2 was carried out regardless of which category was dominant and which co-dominant. In the case of a combination of co-dominant categories not included in Table 2, the resampling was performed by randomly choosing one of the co-dominant categories. After resampling, the number of first-mode categories within the 10 x 10 pixel groups was calculated to check the adequacy of the 355 smoothing and resampling method to the data. because these elements occupy a small fraction of a 1 km 2 pixel and are not identified in a fuel type product at this resolution, and photo on point. We selected only those LUCAS points with available photos, so our fuel types associated with fuel depth or multilayer structure could be estimated visually. Moreover, to avoid border effects and make LUCAS points more comparable to our target spatial resolution (1 km), they should be located within 375 large homogeneous areas. So, LUCAS points were buffered 200 m and only those points whose buffers met these three conditions were kept: 1) falling 88.5 % inside a polygon ≥ 4 km 2 of the 100 m vectorised fuel map, 2) falling completely inside a polygon of the 1 km 2 vectorised fuel map for the main fuel types, and 3) falling completely inside the study area. We used 88.5 % instead of 100 % to have enough pixels to perform validation for all main fuel types. Finally, we obtained 28,240 suitable LUCAS points, whose land cover categories were reclassified to 380 the most similar FirEUrisk main fuel types and used to generate 5,016 validation points by stratified random sampling. A confusion matrix was computed for quantitative analysis.
After validating the main fuel types from this automatic procedure, we performed a second validation exercise, aiming to assess all mapped fuel types, which required to obtain reference information on leaf type, phenology, fractional cover, fuel depth, and type. Since this required a visual interpretation, a 20 % subset of the 385 5,016 validation points was selected by stratified random sampling. Each point was assigned to a fuel category by visual interpretation of four information sources: 1) the latest Google Earth images to observe the 1 km 2 pixel, 2)  . 4).  The application of the tested smoothing window sizes (3 x 3, 5 x 5, 7 x 7) increased the percentage of 10 x 10 pixel groups with unimodal distributions after resampling, although in all cases the increase was marginal ( activities (e.g., spread and behaviour, emissions, post-fire, etcetera) because it allows mapping fuel models and their associated parameters.
The fuel types defined in this paper were matched to the Scott and Burgan Fire Behaviour Fuel Models

480
(FBFM) (Scott and Burgan, 2005), which is a widely used fuel model classification system in Europe (Palaiologou et al., 2013;Aragoneses and Chuvieco, 2021;Alcasena et al., 2021). The FBFMs were based on the NFFL system (Anderson, 1982) and created to address fire behaviour predictions based on Rothermel's surface fire spread model (Rothermel, 1972) for the United States. They include 40 fuel models classified into 7 different groups according to the predominant fire-carrying surface fuel type: grass (GR), grass-shrub (GS), shrub (SH), timber-understory

485
(TU), timber-litter (TL), slash-blowdown (SB), and non-burnable (NB). Overall, the differences in fire behaviour among the surface fuel groups are mainly related to fuel load and its distribution among the particle size categories, Surface Area to Volume ratio, and fuel depth. Compared to NFFL models, the FBFM allows having a number of fuel models not fully cured or applicable in high-humidity areas. Regarding this point, to further improve the matching possibility and account for variations in fuel types and moisture conditions across Europe, we 490 distinguished arid/semi-arid and sub-humid/humid fuel types, as described in previous sections. Furthermore, FBFM data include more fuel models than the NFFL system for forest litter and litter with grass or shrub understory. Anyhow, a user can easily move from the proposed FBFMs to the NFFL system by using the crosswalk table between FBFM and NFFL fuel models (Scott and Burgan, 2005). In addition, our proposal of surface fuel mapping and characterisation for the European general conditions can be adjusted or adapted to specific study 495 areas or sites where more detailed information and measurements on fuels or custom data are available (Mutlu et al., 2008;Salis et al., 2016).
For the purpose of this study, we assigned to each fuel type a given FBFM and the related fuel parameters that most fitted the average conditions in the field, according to expert knowledge. As a general rule, we assigned grass models to fuel types related to grasslands and croplands and selected different sets of FBFM models 500 depending on the fuel depth and cropland type, as well as on bioclimatic conditions: arid/semi-arid versus subhumid/humid regimes ( Fig. B1 in Appendix B). Shrub models were indicated in shrubland areas, following the same considerations described for grass models. Moreover, we proposed the use of shrub models in conditions of open forests, where the fractional cover is low, and the high availability of sunlight can stimulate the presence of a shrubby understory. Timber understory and timber litter FBFMs were associated with closed forests: overall, we 505 assigned low fuel-load models to evergreen forests and higher load models to broadleaf forests. The FirEUrisk fuel types 51, 52 and 53 were associated with shrub or grass FBFM models, depending on the main surface fuels.
Finally, we proposed non-burnable (NB) conditions for urban continuous areas and other non-burnable zones (e.g., water, snow, ice, bare soils, sparse vegetation < 10 %), while shrub models were indicated for urban discontinuous areas, to account for the potential of a fire to spread in such environments.

The FirEUrisk fuel classification system crosswalk to standard fuel models
The FirEUrisk fuel types crosswalk to the FBFM system (Scott and Burgan, 2005) is presented in Table 7, and the related FBFM map over Europe is provided in Fig. 6 and complemented with Table 8.   Figure 6. European fuel models based on the FBFM fuel models (Scott and Burgan, 2005) at 1 km resolution.
See Table C1 in Appendix C for the fuel descriptions and parameters.  The most relevant fuel model at the continental scale is GR6 (area covered about 1.6 Mkm 2 ), which refers to medium-high and moderate live-load grasslands of sub-humid/humid areas and is characterised by high moisture values. This fuel model is largely related to herbaceous croplands that cover the most productive agricultural flat areas of central and northern Europe. About 0.8 Mkm 2 of Europe is covered by TU2, which was associated with 530 closed needleleaf evergreen forests located in the sub-humid/humid regime. TU2 is related to timber understory characterised by moderate-load shrubs. TU3, which concerns timber understory with a combined presence of grasses and shrubs with moderate fuel load, is the third more common fuel model in Europe, covering 7.77 % of the area. We proposed TU3 in closed broadleaf deciduous forests of sub-humid/humid areas. For arid/semi-arid areas, GR4 is the dominant fuel model and occupies about 0.33 Mkm 2 (6.76 %) of land. This model represents 535 moderate load grasses of dry climates. We associated GR4 with herbaceous croplands of southern Europe. Among the fuel models that cover more than 5 % of the study area, we should also mention the GR9, which refers to tall and high live load grasslands of sub-humid/humid areas and is characterised by high moisture values; and the nonburnable fuels, which refer to urban continuous areas and other non-burnable areas including bare soil, water, and glaciers. The other FBFMs used in this work characterise approximately the remaining 24 % of the European 540 territory and range from 0.22 Mkm 2 of TL3 to 7,721 km 2 of GR7.
A description of the parameters of the FBFM fuel models used for the crosswalk is presented in Table C1 in Appendix C. As an example, we mapped the 1h dead fuel load and the surface fuel depth over Europe (Fig. 7). refers to the understory, not the crowns.

Discussion
The proposed FirEUrisk hierarchical fuel classification system was designed to be adapted to a wide and offers high versatility because it is expected to enable fuel mapping at various scales with different disaggregation of categories, depending on the detail and quality of the input data. Thus, whereas the fuel map developed at the European scale was based on existing European and global datasets integrated into a GIS framework, the same classification scheme could be applied to provide a more comprehensive fuel classification using a multi-sensor approach in a machine learning framework (García et al., 2011;Marino et al., 2016;Domingo 560 et al., 2020). Its structure has similarities (e.g., hierarchical scheme) with the ArcFuel classification (Toukiloglou et al., 2013), although this was only prepared for southern-European conditions. In addition, the involvement of expert knowledge in the development of the FirEUrisk hierarchical fuel classification system suggests high acceptance, and therefore usage, among the fire risk management community in the foreseeable future. It also allowed the development of a useful classification, intended to fill the actual gaps of the European fuel mapping, 565 towards a homogeneous and integrated fire risk prevention strategy. Nevertheless, it must be considered that the grouping of vegetation types into fuel types is a balance between generalisation of the landscape reality and loss of detailed information, which may not be the most suitable system for all study areas.
The predicted increase in fire intensity and occurrence of the so-called megafires (San-Miguel-Ayanz et al., 2013), which usually evolve from surface to crown fires, makes it necessary to improve our information on Regarding the European fuel mapping, the combination of existing land cover and biogeographic datasets,

575
and bioclimatic models, facilitated the generation of the fuel type dataset, being some of these data specifically adapted to European conditions (Europe's biogeographic regions map, CLC map). Nevertheless, the input datasets are a generalisation of the complex reality with their own uncertainties and errors, which are transferred to the final European fuel map. In fact, the errors of the final fuel type dataset are similar or even lower than those found in the main input land cover map used to obtain the fuel categories.

580
Estimating shrubland and grassland fuel depth was challenging. To our knowledge, there are no largescale reliable datasets in Europe on these variables. Although the models chosen to estimate fuel depth were not specifically developed for European areas, the biogeographical similarity to European conditions supported their usage for our purposes. For shrubland, 75 % of the shrubland fuels (in the 100 m resolution map) belong to the arid/semi-arid regime, which justifies the selection of a bioclimatic model developed for an arid/semi-arid area.

585
To avoid unrealistic estimations, we constrained the outputs to the range [0-6] m for the shrublands and to > 0 m for the grasslands, while no maximum cut-off threshold was applied to the grassland category as the obtained maximum value (1 m) was considered reasonable. In addition, the distribution of shrubland and grassland pixels led to considering the bioclimatic models adequate. The histogram for shrubland fuel depth showed the spatial continuity of the input variable (precipitation). The histogram for grassland fuel depth had an aggregated structure 590 due to the input productivity data by biogeographic region. Obviously, direct measurement of shrubland or grassland fuel depth is more desirable. Future works based on airborne or satellite LiDAR should provide a better estimation, but they are not yet available for the whole European territory (airborne) or need addition calibration efforts (satellite).
Concerning the final European fuel map (1 km 2 ), only 20 out of the 24 possible first-level fuel types were 595 mapped because for a fuel type to be mapped, it must occupy a continuous area large enough to be represented in 1 km 2 . The herbaceous cropland and the closed needleleaf evergreen forests are the most extended fuel types in Europe, related to the land use activities of the European society and the natural distribution of vegetation species due to bioclimatic conditions (García-Martín et al., 2001). Also, the large extension of forest fuel types constitutes an increasing potential risk in the light of the growing trends of land abandonment, particularly in remote areas:

600
forests with high surface fuel load can more easily turn into crown fires (Scott and Reinhardt, 2001;Weise and Wright, 2014), characterised by high intensity, emitting vast amounts of the stored carbon. Urban fuel types are the least represented in Europe, but they are the most dangerous from an economic, societal and human health point of view (Bowman et al., 2011).
Finally, the quantitative assessment of the mapped FirEUrisk main fuel types (forest, shrubland, 605 grassland, cropland, wet and peat/semi-peat land, and urban), plus the nonfuel category; obtained a high overall accuracy of 88.48 %: average commission errors of 34 % (highest for the nonfuel category and lowest for the forest fuel types) and average omission errors 30 % (highest for the grassland and lowest for the urban fuel types).
Although it is higher than the used base cartography, the Copernicus GLC map , and it surpassed the ideal 85 % minimum overall accuracy; not all fuel types presented the ideal ≥ 70 % accuracy 610 (Thomlinson et al., 1999). The overall accuracy was higher than the one for the 2019 Copernicus GLC map over Europe (79.9 %), probably due to the validation approach. The confusion matrix is aligned with the confusion matrix of the 2015 global Copernicus GLC maps over Europe , The quantitative assessment of all mapped FirEUrisk fuel types obtained a medium-high overall accuracy of 80.92 %: average commission errors of 41 % (highest for the high grasslands, and tree and shrubland wet and peat/semi-peat land fuel types; and lowest for the herbaceous cropland fuel type) and average omission errors of 50 % (highest for the medium grassland fuel type and lowest for the urban continuous fabric fuel type). These 635 results are higher than the used base cartography, the Copernicus GLC map , but do not surpass the ideal 85 % minimum overall accuracy, neither all fuel types with ≥ 70 % accuracy (Thomlinson et al., 1999). However, the visual assessment improved the validation method because it considered the entire 1 km 2 pixels and not only the area of the LUCAS points. This method could only be applied to a subset of the validation points because of its temporal and human cost compared to the previous validation method. The results are similar 640 to the confusion matrices of the FirEUrisk main fuel types and the Copernicus GLC map over Europe , although errors are higher and different due to the dissimilar validation methods and reference data, and that confusion appears between fuel types belonging to the same main fuel type. Most errors are due to pixels with a mixed cover of fuel types, and low quality of the reference data (unclear and blurred Google images and LUCAS photos; and pixels not meeting all ideal conditions for validation -that was needed to have a representative 645 sampling for every fuel type). Input and reference data temporal differences can also have affected the accuracy.
The obtained errors present the typical pattern for land cover and vegetation classifications with remote sensing (used to develop the input data), dependent on the separability of the spectral signatures of the land types. This explains why errors are dominant for fuel types belonging to the same main fuel type instead of fuel types from different main fuel types.

650
Forest fuel types have acceptable accuracy except for the closed mixed forest, highly confused with closed needleleaf evergreen forest. Many errors refer to the omission of open forest, assigned to the closed forest, as happens in the Copernicus GLC map over Europe . Shrubland and grassland fuel types' errors are significant, mostly between fuel depth categories. However, care must be taken for these results, as estimating fuel depth from photos is challenging, and fuel depth varies with time. These limitations specially affect 655 grassland due to its low depth, rapid growth, and that high grassland is frequently cut. Thus, grassland fuel depth is very changeable so we assume the European fuel map may only be accurate for some periods of the year. We The different levels of disaggregation of the proposed classification system, as well as the main fire behaviour characteristics of the diverse fuels, made the crosswalk challenging and did not allow to assign a specific standard model to each FirEUrisk fuel type. Moreover, the FBFM standard fuel models (Scott and Burgan, 2005) 670 were originally developed for the United States, so care must be taken when using the crosswalk in Europe (Santoni et al., 2011;Salis et al., 2016). From this point of view, our proposed approach can be improved in specific areas if customised information and data on given fuel types are available (Arca et al., 2007;Fernandes, 2009;Duguy Pedra et al., 2015;Kucuk et al., 2015;Ascoli et al., 2020). In other words, we propose a generic crosswalk scheme, but users are free to wisely choose or modify the best fitting standard fuel models according to their study area and 675 expertise, or to use different parameters from the standard ones if they have better information for given study areas. Plus, the main limitation of the crosswalk scheme relies on the reference to general bioclimatic regimes, which is not able to fully consider all inherent differences among European regions in terms of fuel characteristics, while moisture values can be spatially modified according to the specific status of each fuel type.
This work represents one of the first attempts to adopt a standardised fuel model mapping approach over

680
Europe, similar to the National Fire Danger Rating fuel models products available since the '90s for the continental United States (see for instance https://www.wfas.net/index.php/nfdrs-fuel-model-static-maps-44). Work is in progress to develop higher resolution products over Europe combining a set of remote sensing tools and data. This latter development at the European scale is highly complicated by the huge heterogeneity in the availability of high quality and resolution of ground and measured data, which vary a lot among and within the different regions.

685
The FirEUrisk fuel classification system can provide a number of insights and information for wildfire risk monitoring and assessment at the European scale. This is mostly related to the identified fuel categories crosswalk to the FBFM system (Scott and Burgan, 2005), which is specifically designed for the above purpose. In fact, the parameters included in each FBFM model allow the characterization of surface fuels and can serve as a baseline for surface wildfire spread and behaviour modelling. For instance, some existing fire spread models, such 690 as FlamMap (Finney, 2006) and FARSITE (Finney, 2004), could be used for this purpose, although canopy parameters should be additionally estimated. This should be subject of an extension of this project and could be based on airborne and satellite LiDAR systems. The fuel map is also expected to serve estimations of fire-caused carbon emissions and pollution, and estimations of biomass consumption. The Consume model (Prichard et al., 2006) could be used for this if a 695 crosswalk to FCCS fuels is previously made, including the necessary fuel parameters such as the combustion percentage. In addition, the European fuel map would be useful for regions that do not have fuel cartography.
Overall, we highlight that the main use of the map is providing a dataset able to rate fire danger and risk conditions across large geographic areas, while the application of wildfire spread models to very local scales or small areas may pose limitations in the quality of outputs due to low resolution (1 km 2 ) of the fuel input layer.

700
Finally, although it has been developed for European conditions, our methodology has the potential to be applied to other regions. The proposed fuel classification system could be used in other projects and applications apart from the FirEUrisk project, and adapted anywhere in the world, further extending the fuel subcategories wherever required. The classification of fuel types is dependent on existing land cover and biogeographic data, but it can also be directly estimated from satellite data, either coarse resolution for continental areas or higher 705 resolution for smaller territories. The fuel parameterization can also be based on standard models, such as the Scott and Burgan system (Scott and Burgan, 2005) used in this paper, but it can also rely on ground measurements or more detailed regional fuel characteristics. In any case, it is important to emphasise the need of estimating fuel parameters to use the fuel type products for quantitative estimations of fire risk, behaviour, and effects. This is a key aspect of the FirEUrisk project and a crucial point toward wildland fire prevention across the European Union.

Conclusions
This work, developed in the framework of the European FirEUrisk project , provided a hierarchical fuel classification system for surface and canopy fuels adapted to continental conditions. The final European fuel map 720 contains 20 fuel types, including both surface and canopy fuel categories. The estimated overall accuracy was 88 % for the main fuel types and 81 % for all mapped fuel types. A crosswalk between the proposed fuel types and commonly used standard fuel models, Fire Behaviour Fuel Models (FBFM) (Scott and Burgan, 2005), has been presented as well. Our approach, based on expert knowledge, Geographic Information Systems, existing land cover datasets, biogeographic data, and bioclimatic modelling, could be readily applied to other regions.

725
The results of this study constitute the first step toward a risk-wise landscape and fuel mapping development across Europe, which will help integrated, strategic, coherent, and comprehensive decision making for fire risk prevention, assessment, and evaluation.    Table B2. Percentage of 10 x 10 pixel groups with 1, 2 or > 2 first-mode categories for the 3 x 3, 5 x 5, and 7 x 7 smoothing moving windows, and without window applied.
Percentage (%) of 10 x 10 pixel groups with: Window size 1 first-mode category 2 first-mode categories > 2 first-mode categories      Disclaimer. This research reflects only the authors' view, and the European Commission is not responsible for any use that may be made of the information it contains.

Acknowledgements.
We would like to thank the members of the FirEUrisk consortium for their comments on the FirEUrisk fuel classification system. We would also like to thank María Clara Ochoa and Suresh Babu Kukkala 765 for their help with the product's validation. We are also grateful to anonymous reviewers for their helpful comments.
Financial support. This project has been granted funding from the European Union's Horizon 2020 research and innovation programme under Grant Agreement No. 101003890.