The target region comprises a total area of 4444.25 km2, located in a borderland between Portugal and Spain, in the Transboundary Biosphere Reserve of Meseta Ibérica (Fig. 1). Topography varies widely, from mountains in the north (maximum elevation of ca. 2500 m) to flatter, undulated terrain in the south (minimum elevation of 0 m). Climate is characterized by the prevalence of Csb (dry and warm summer) climate, with a few areas with a Csa (dry and hot summer) climate under the Koppen-Geiger classification (Beck et al. 2018). At the end of the century, considering the RCP8.5 scenario, the Csa climate type will expand to the entire study area (Beck et al. 2018). Annual precipitation varies considerably over the territory, from around 1200 mm year− 1 to 700 mm year− 1 following a north-south decreasing gradient. Temperature varies along the same gradient, from an annual average of approximately 12°C in the northern mountainous areas to 17°C in the southernmost areas.
The region went through has suffered significant socioeconomic changes over the last several decades, leading to depopulation and abandonment of marginal agricultural land. This, together with mountain afforestation programs in the mid-20th century, led to an increase in the number of wildfires and the burned area from 1990 to the mid-2000′s, with a small decline in the last decade (ICNF, 2015). A joint effect of changes in land use and land cover, topography and vegetation type and structure were found to affect fire behaviour in the area (Azevedo et al. 2011; ; Silva et al. 2011; Magalhães et al. 2017). Depending on latitude, elevation and human agency, the region is dominated by four distinct woody vegetation types that are broadly representative of those occurring across the Mediterranean Basin, respectively pine forests, evergreen sclerophyllous oak woodlands and forests, deciduous forests, and a range of shrubland communities.
Monthly burned area distribution in the study area is concentrated mainly from July to September (Fig. 7). Another period of relevant wildfire activity can be detected in February and April, mainly due to agricultural stubble burning and for pastures renewal. Typically, the fires occurring in pasture and agricultural lands are smaller than the ones occurring in shrublands and forest areas, restricted mainly to the summer period.
Climate scenarios and data
The future climate was characterized under two distinct climate change scenarios, RCP4.5 and RCP8.5. The former describes a radiative forcing of ≈ 4.5 W m−2 (≈ 650 ppm CO2 eq.), implying an increase in global mean surface air temperature of 1.8°C (1.1–2.6°C) by the end of the century relative to the historical period of 1986–2005 (Collins et al. 2013). RCP8.5 describes a high-end emission scenario (van Vuuren et al. 2011) radiative forcing of ≈ 8.5 Wm−2 (≈ 1370 ppm CO2 eq.), with a projected increase in global mean surface air temperature of 3.7°C (2.6–4.8°C) in 2100 relative to the same historical period (Collins et al. 2013). The inclusion of more than one climate scenario is useful to assist in the decision-making process (Pedersen et al. 2020).
The following atmospheric variables were selected to compute the fire danger indices used herein: daily total precipitation, daily minimum, mean and maximum temperature, daily minimum, mean and maximum relative humidity and daily mean wind speed. Both observational and simulated data generated by climate models were retrieved within the target region. The ERA5-Land reanalysis dataset (Muñoz-Sabater et al. 2021) was used to characterize the historic period (1989–2005). The climate model simulations were considered for the historic (1989–2005) and future (2031–2050) periods, under RCP4.5 and RCP8.5. A 4-member ensemble of Global Climate Model (GCM) – Regional Climate Model (RCM) chains was provided by the EURO-CORDEX project (https://www.euro-cordex.net/). The present study used the variables Near-Surface Temperature, Near-Surface specific Humidity, Precipitation, Near-Surface Wind Speed, Eastward Near-Surface Wind, Northward Near-Surface Wind, Daily Maximum Near-Surface temperature and Daily Minimum Near-Surface temperature from the driving models: MPI-M-MPI-ESM-LR, IPSL-IPSL-CM5A-MR, ICHEC-EC-EARTH and CNRM-CERFACS-CNRM-CM5. Each driving model represents an institute and an RCM model (Table 1). For these projections, the time-frequency is daily, the ensemble is r1i1p1 and the experiments were carried out for the historical, RCP4.5 and RCP8.5 anthropogenic forcing.
List of regional climate models (RCM) and respective institute and driving model, used in this study (EURO-CORDEX Data, 2020; Fraga et al. 2020)
Daily temperatures and precipitation were bias-corrected at the daily timescale using the quantile-mapping approach and the observation-based E-OBS gridded dataset version 22.0e (Cornes et al. 2018) as a baseline in the historic period (1989–2005). For each model, the same bias correction was applied to both the historic and future periods. The raw climate model data were defined over a 0.11° latitude × 0.11° longitude (~10 km) and were re-gridded, by bilinear interpolation, to the same grid as ERA5-Land and E-OBS, at a native spatial resolution of 9 km (0.10° latitude × 0.10° longitude), thus warranting the overlap of grid boxes from all datasets and variables.
Future changes in the climate variables were calculated by computing the anomalies for each location between future monthly values and the median observed value in the reference period. Fire danger and fire behaviour were calculated for the entire EURO-CORDEX grid within the study area, per climate model and scenario. For each day in the temporal series, we calculated the percentile 95 and the median of the study area to represent the fraction of the study area with the highest fire danger and potentially exhibiting a more dangerous wildfire behaviour and the median hazard in the region, respectively. We then calculated the monthly value of fire danger and fire behaviour for each climate model and scenario. We used the median monthly value to ensemble the different metrics.
Fire danger and wildfire behaviour characteristics
The Canadian Forest Fire Weather (FWI) Index System (van Wagner, 1987) was evaluated as the preferred method for fire danger rating in the Mediterranean region, particularly for the summer season (Viegas et al. 1999) and it has become the standard to assess fire danger at regional to global scales, including in climate change applications (Bedia et al. 2015; Abatzoglou et al. 2019). We evaluated the influence of climate change in the Fire Weather Index (FWI) and its two major sub-indices. These indices are presented and characterized below (van Wagner, 1987):
The Initial Spread Index (ISI), that represents the atmospheric effect in fire-spread rate.
The Buildup Index (BUI), that reflects the drought effect on the amount of fuel available for combustion.
The Fire Weather Index (FWI), a combined outcome of the ISI and the BUI that rates potential fire intensity and is often used by fire management agencies to assess fire danger conditions in the near future.
The FWI codes are calculated from noon readings of air temperature and relative humidity at a 2-m height, open wind speed at 10 m, and 24-hour rainfall. However, because of the study emphasis on potential fire behaviour rather than on fire danger conditions per se, calculation of the indices with an atmospheric component (FFMC, ISI, FWI) was based on daily maximum temperature and minimum relative humidity to capture the peak burning conditions associated to minimum fine dead fuel moisture content. The alternative would be to use daily means of input variables, which is not recommended as it induces a negative bias on FWI estimation (Herrera et al. 2013).
The FWI varies greatly between regions (or ecozones), with FWI values associated with the occurrence of large wildfires and distinct fire activity levels differing between locations (Amiro et al. 2004). Recently, Fernandes (2019) defined FWI thresholds associated with large wildfires in each administrative region (Distrito) in Portugal. For the study area, mainly located in the Bragança distrito (nearly 60% of the study area), wildfires that burn ≥500 ha are associated with an FWI of 79.9, whereas wildfires that burn ≥1000 ha are associated with an FWI of 90.1. Here, we used these FWI values to calculate the number of days above each of these thresholds for the historical and future climate (both under RCP4.5 and RCP8.5).
In addition to the fire danger indices, we calculated fire behaviour characteristics, namely rate of spread and fireline intensity in common Mediterranean vegetation covers. This need comes from the fact that FWI does not consider important factors that affect fire ignition and spread (and consequently fire danger), such as vegetation cover. ROS and FLI are defined as follows:
Rate of Spread (ROS) is the linear rate of advance of the fire front. ROS determines the amount of resources needed to contain a wildfire and is important for shaping fireline intensity and consequently, suppression difficulty and firefighter safety (Sullivan and Gould, 2018)
Fireline Intensity (FLI) is the amount of heat released per unit time and unit length of the fire front (kW m-1). FLI is a major driver of aboveground fire effects and is often used as a proxy for fire suppression difficulty (Alexander and Cruz, 2018).
Fire behaviour modelling
We considered the four generic vegetation types that prevail in the region, respectively long-needle pine forest (Pinus pinaster, P. pinea), broadleaved deciduous forest (Quercus pyrenaica, Castanea sativa), broadleaved evergreen forest (Quercus suber, Q. ilex), and shrubland dominated by Erica, Cytisus, Pterospartium and Cistus species. The forest types were assigned fuel models from the Portuguese catalogue (Cruz and Fernandes, 2008; Fernandes, 2009) for the purpose of fire behaviour simulation, respectively M-PIN, M-CAD and M-ESC. These fuel models assume surface fuel complexes composed of litter and shrubs and their development included parameterization based on field observations of fire behaviour characteristics.
The forward (maximum) spread rate of a surface fire was simulated with the model of Rothermel (1972) for flat terrain, which depicts landscape fire spread better (Sullivan et al. 2014), and for baseline live fuel moisture contents for each forest type (Fernandes, 2009); because the effect of live fuel moisture content in the model is a mathematical artefact (Catchpole and Catchpole, 1991) the estimates were subsequently adjusted as per Rossa and Fernandes (2018). Live fuel moisture content was estimated from the Drought Code (DC) of the Canadian FWI System and an ensemble of representative species in the region (Viegas et al. 2001). The simulations also require “midflame” wind speed, i.e. ~2-m wind speed, and the moisture content of fine dead fuels, estimated respectively from 10-m open wind speed and the Fine Fuel Moisture Code (FFMC) of the Canadian FWI System (van Wagner, 1987), and then adjusted for vegetation type (Fernandes, 2009; Pinto and Fernandes, 2014; Andrews, 2018). Calculation of fire-spread rate for shrubland (SHRUBL) followed the same procedures but was based on the robust empirical model of Anderson et al. (2015). Surface FLI was calculated as the product of ROS, heat of combustion (assumed constant at 18,000 kJ kg−1) and fuel consumption (Byram, 1959; Alexander, 1982); the latter was estimated as an asymptotic function of the Duff Moisture Code (DMC) of the Canadian FWI (Forestry Canada Fire Danger Group, 1992; Fernandes and Loureiro, 2013). Assessment of surface fire transition to a crown fire and the type of crown fire (active or passive) followed Van Wagner (1977), but assuming that deciduous broadleaved forest does not support active crowning (Alexander and Cruz, 2006; de Groot et al. 2013). Crown fire rate of spread adopted the model of Cruz et al. (2005) for pine forest and assumed Anderson et al. (2015) in broadleaved evergreen forest. Surface and crown fire spread were weighted as a function of crown fraction burned (van Wagner, 1993) to calculate the overall fire-spread rate and FLI. Typical (median) values per vegetation type for the fuel and stand descriptors required by the calculations were computed from the national forest inventory plots located within the study region (Fernandes et al. 2019).
The influence of climate change in fire behaviour is presented by only considering the weather influence in ROS and FLI, and also by considering the potential effect that climate has over the biomass (i.e. the amount of fuel available to burn). To represent potential changes in fuel consumption during a wildfire due to climate change, we adjusted the Net Primary Production (NPP). NPP was calculated for each grid point, according to the global equations presented in Del Grosso et al. (2008) based on empirical non-linear relationships with temperature and precipitation. NPP was calculated differently for tree-dominated vegetation, i.e. forest, and for shrubland, respectively from precipitation and temperature and precipitation only.