Wildfire Risk in the European Union

The regime

Two fire environments across one continent

The model

SWIM Uses Published European Fuel Maps

SWIM uses the FIRE-RES pan-European surface-fuel layer and the FirEUrisk fuel map, refined with ESA WorldCover and Copernicus tree-cover data. These inputs are mapped to the Scott and Burgan fuel classification used by the spread model and carried onto SWIM's 20-meter hexagons. Fire spread is calculated using the Rothermel and Albini equations across all forty published fuel models. European weather inputs come from ERA5 and CERRA reanalyses for the historical record and ECMWF and ICON-EU forecasts for live requests. Calibration uses the European historical fire record.

Fire-Size Estimates

Fire Size Is Represented as a Range of Possible Outcomes

Because final fire size is uncertain, SWIM estimates a range of possible outcomes rather than a single deterministic value. The model is developed using 156,080 European fire records and evaluated on geographically separate areas that were excluded from training.

For each request, SWIM uses the ignition location, time of year, surrounding fuel, terrain, and settlement, the day's weather, and the local temperature climate to estimate fire size and consequences. Results are reported in hectares and euros and are based on European data.

The test

Testing the 90th-Percentile Fire-Size Estimate

The 90th-percentile estimate represents a final fire size that about 90 percent of observed fires are not expected to exceed, while about 10 percent are expected to exceed it.

Across 15,360 observed European fires of at least 30 hectares, 90.0 percent of final fire sizes fell at or below the model's 90th-percentile estimate. Across the prediction range, coverage remained within 2.2 percentage points of the 90 percent target, for fires forecast at both 140 hectares and 2,700 hectares.

The model was evaluated using five-fold geographic cross-validation. The EU study area was divided into 379 blocks measuring 100 km by 100 km and assigned to five groups. In each round, one group was excluded from training and used only for evaluation. As a result, every historical fire was tested using a model that had not been trained on fires from the same geographic block. This separation reduces the risk that nearby fires with similar fuel, terrain, and weather make the model appear more accurate simply because it has already seen the same landscape.

Table 3 reports the performance of the 90th-percentile fire-size estimate across the full population. Figure 1 compares eight fires with their modeled ranges, and Table 4 presents their values.

Table 3. Performance of the 90th-percentile fire-size estimate on geographically held-out fires. Overall, 90.0 percent of observed fires fell at or below the modeled 90th percentile. Each fire was evaluated by a model that excluded its entire 100 km block from training.

Table 4. Modeled median and 90th-percentile fire sizes compared with observed final burned area for the eight examples in Figure 1. The start point identifies where each fire simulation began. Each estimate comes from a model that excluded the fire's 100 km geographic block from training. Observed burned area is the area recorded by EFFIS.

The eight fires come from eight member states, from the Portuguese interior and the mountains of León to Crete, Sicily and the southern Carpathians. Three were started where the fire agency recorded the ignition and five where satellites first saw them.

Historical Reconstructions

Historical Fire Reconstructions

Each fire in Table 6 was simulated using the same workflow as a live request and the weather observed on the day of ignition.

Each simulation begins at the earliest satellite detection so the model does not use information about the fire's eventual extent. This setup approximates the information available early in an incident: location, landscape conditions, date, and weather. Starting from a point inside the final burned area would introduce information about where the fire ultimately spread. Using the earliest satellite detection therefore keeps the reconstruction closer to a forward-looking test.

The primary spatial metric is the percentage of the observed burn contained within the modeled area. For planning applications, missed portions of the final burn fall outside the modeled planning area. Each fire is first evaluated using its 90th-percentile modeled size and, if the observed fire exceeds that range, also at the 95th percentile. These percentiles are reference points within a continuous modeled distribution, not maximum possible fire sizes. A request can be read through the 99.9th percentile. Table 5 presents the inputs, and Figure 2 shows four reconstructions.

Table 5. Inputs used for the historical fire reconstructions. Each simulation begins at the earliest satellite detection, taken as the center of that overpass rather than a single pixel, and uses observed weather for the day of the fire. The modeled area is drawn using the same rule applied to a live request.

Observed 19,112 ha95th-percentile modeled area 26,116 ha

Observed 5,806 ha90th-percentile modeled area 11,215 ha

Observed 7,914 ha90th-percentile modeled area 13,929 ha

Observed 8,124 ha95th-percentile modeled area 10,626 ha

The white line shows the observed fire perimeter, and the blue area shows the modeled area beginning from the earliest satellite detection. Imagery: Esri, Maxar, Earthstar Geographics. All four examples were generated using the same model and each fire's observed conditions, without event-specific tuning.

1. aBurón, León. The 95th-percentile modeled area captures 92.4 percent of the observed 19,112-hectare burn, including most of the fire's northern and southern extent in the Cantabrian Mountains.
2. bLa Teste-de-Buch, Gironde. The 90th-percentile modeled area captures 99.5 percent of the observed July 2022 burn in the pine forest south of Arcachon and reproduces the dominant north-to-south spread pattern.
3. cPirama, Chios. The 90th-percentile modeled area captures 99.4 percent of the observed burn and broadly reproduces its overall shape.
4. dSvilengrad, Haskovo. The 95th-percentile modeled area captures 92.8 percent of the observed burn, including most of the southern spread, while missing part of the western edge.

Table 6. Six fires reconstructed from their earliest satellite detection. First sighting identifies the simulation start point, modeled percentile identifies the level of the distribution used, modeled area is the area estimated at that level, and observed burn captured is the share of the mapped perimeter contained within the modeled area. Figure 2 shows four of the six fires. The Italian example, Custonaci, appears in the report published with the Italy wildfire risk map.

Two of the six fires are evaluated at the 90th percentile and four at the 95th. The simulations use only information available at or near the beginning of each fire: location, landscape conditions, date, and weather.

The risk map

Where the fuel and the people overlap

The European wildfire risk map contains 6,099,766 reporting cells. For each cell, SWIM estimates how often a fire is expected to start, how large it could become, and the economic consequences. Ignition frequency is estimated from the historical fire record and generalized across locations with similar conditions. On geographically held-out areas, the model assigns a higher burn probability to a burned cell than to an unburned cell 92 percent of the time. The continental model is calibrated to an average of 5,101 mapped fires and 572,374 hectares burned per year in the selected historical record. Fire size is estimated as it is for a live request and averaged over the months when fire occurs in each part of Europe.

Loss comprises three components, each valued using the same unit rates applied to a live request.

Structures and human safety. Structure loss uses each state's construction cost and rural dwelling size, multiplied by the share destroyed by a fire. That share is anchored to the 490 houses counted by the Pedrógão Grande inquiry across 28,914 hectares. Human-safety consequences use the EU value of a statistical life and a fatality rate derived from the JRC's 2023 season, which recorded at least 41 deaths among people living inside that year's mapped perimeters.

Land. Loss includes restoration costs and standing timber value in states that publish those figures. Suppression. Cost per hectare decreases as fire size increases. Together, the three components produce about €3.13 billion in modeled expected annual loss. Figure 3 shows its geographic distribution.

Expected loss in a reporting cellBelow the medianThe median or more5.9 times the median or more35 times the median or moreThe top 3%

Table 7 sets the twelve member states with the highest expected loss side by side, and Figure 4 shows how differently the member states rank on area, homes and loss.

Italy

Italy ranks third of 27 member states by expected loss

Italy accounts for about 14 percent of modeled EU expected annual wildfire loss, approximately €426 million, and 12 percent of expected annual burned area.

Italy has its own map and report. The Italy wildfire risk map presents the country cell by cell, and its report covers the 20 regions, the historical record, the two fire seasons, and six Italian fires reconstructed from their earliest satellite detection.

The calendar

Europe has no single fire season

The European Union does not have a single fire season. Mediterranean fire activity peaks in July and August. Temperate regions often experience a spring fire period before green-up and a second period during dry late-summer conditions. Boreal regions tend to peak in spring and early summer after snowmelt. A single continental seasonal curve therefore masks important regional differences. Figure 5 separates the historical record into three latitude bands.

The comparison

Fire Danger and Long-Term Wildfire Risk Answer Different Questions

EFFIS publishes the Canadian Fire Weather Index (FWI) across Europe each day at about 8 km resolution as an indicator of fire-weather severity. SWIM instead estimates long-term wildfire risk by combining fire likelihood, fire behavior, exposure, and consequence. Both identify the Mediterranean as an important area of wildfire concern, but they measure different quantities.

The Fire Weather Index describes how conducive current weather conditions are to fire. SWIM's continental risk map estimates where wildfire risk is concentrated over many years and the expected consequences of burning, while a live request estimates the potential behavior and consequences of a specific fire under current conditions. A location can therefore have a high fire-weather rating but relatively low long-term expected loss if exposure is limited, or the reverse.

Provenance

What the model is built from

Table 8 lists the published European sources used by the model.

Table 8. Model inputs, the organization publishing each source, and the applicable license.