Modelling the daily probability of lightning-caused ignition in the Iberian Peninsula

Keywords: fire danger, forecast, fuel moisture, Iberian Peninsula, ignition probability, lightning strike, machine learning, wildfires.

Lightning is the most frequent source of natural fires worldwide, shaping fire regimes and ecosystems over millennia prior to the advent of human influence and use of fire (Bowman et al. 2009; Fernandes et al. 2022). In Europe, natural-caused fires represent a small percentage of the total number of fires, typically below 5–10% (San-Miguel-Ayanz et al. 2012). In contrast to anthropogenic fires, which occur in a spatial pattern close to the human footprint (e.g. road networks, agricultural lands or the wildland–urban interface), the spatial location of natural fires is often clustered in specific hotspots of intense lightning activity and fuel availability (Ganteaume et al. 2013; Nampak et al. 2021). These locations experience above-average burned area attributed to lightning-caused fires, which are particularly abundant in certain regions of the Iberian Peninsula compared with other Mediterranean regions (Dijkstra et al. 2022).

Lightning fires frequently start in regions distant from human habitation and trigger multiple simultaneous events, which adds an additional layer of difficulty to suppression (Rodrigues et al. 2019a) and increases their likelihood of growing into large fire events when fuel and relief conditions facilitate fire spread (Pineda and Rigo 2017). The stages conducive to wildfire from a lightning source – strike, ignition, survival and arrival – depend on a specific set of factors and conditions (Anderson 2002). In short, a flash with sufficient peak current and duration is needed to initiate the smouldering phase. Fuel moisture, availability and type modulate the chances of survival, along with fire-prone weather conditions (e.g. lack of or small rainfall episodes). Under such conditions, lightning-caused smouldering may begin to flame and initiate a wildfire event (Pineda and Rigo 2017).

Most climate predictions envisage a decrease in fuel moisture content while fuel accumulation due to land abandonment is creating the perfect conditions for larger and more extreme fires across the Western Mediterranean Basin (Moreira et al. 2020, p. 202). Although further research is needed, there is already evidence of an increase in the occurrence of natural fires related to the increased frequency of dry storms and heat waves (Coogan et al. 2020; Li et al. 2020), potentially increasing fire frequency, size and intensity (Turco et al. 2018; Dupuy et al. 2020; Barros et al. 2021). During the summer of 2022, unprecedented lightning-caused fire activity was observed in some Mediterranean-type regions like California or Spain while some studies warn about lightning-caused fires becoming larger (Dowdy and Mills 2012) and more frequent (Cattau et al. 2020) in Australia. Understanding the linkages and feedbacks between environmental and atmospheric conditions that ultimately relate to lightning-caused fires is relevant to better understand fire regimes (Barros et al. 2021; Cochrane and Bowman 2021) or develop fire risk assessments (Chuvieco et al. 2014).

Lightning-caused ignitions have attracted considerable attention worldwide, which has translated into a wide array of studies devoted to understanding lightning patterns and causes, their link with fire ignition or their future evolution under climate warming (Price and Rind 1994a, p. 199, 1994b). Without being exhaustive, we found examples in the USA (Cattau et al. 2020; Jiménez-Ruano et al. 2022), Canada (Wotton and Martell 2005; Wotton et al. 2010), China (Hu and Zhou 2014), Australia (Dowdy and Mills 2012) and Europe (Dijkstra et al. 2022). In the Iberian Peninsula, the region under study in the present work, lightning-caused fires have also been investigated historically (Vazquez and Moreno 1998; Castedo-Dorado et al. 2011; Nieto et al. 2012; Chuvieco et al. 2014; Couto et al. 2020). Numerous studies have focused on addressing the probability of fires caused by lightning, humans or both. Most studies dealing with natural fires are based on historical compilations of natural fires coupled with environmental and climatic variables (Nieto et al. 2012; Rodríguez-Pérez et al. 2020) and meteorological data to determine the atmospheric conditions conducive to lightning fires (Pérez-Invernón et al. 2021; Soler et al. 2021; Pineda et al. 2022) or the lightning features that foster fire events (Pineda et al. 2014, 2022; Couto et al. 2020; Soler et al. 2021). The interactions between atmospheric models and the rate of spread of wildfires have also been investigated by Couto et al. (2020). However, there is still a lack of homogeneity in the methods, variables and protocols to predict, and therefore anticipate, the occurrence of natural wildfires. Here, we advocate an approach that decouples the conditions that favour lightning strikes from the environments that favour ignition.

The present work aimed to predict the probability of ignition of natural fires by translating lightning records into spatial-temporal patterns of lightning ignition potential on the Iberian Peninsula. For this purpose, we analysed a large lightning strike dataset (more than 17 million observations) to ascertain the conditions under which a lightning strike may trigger an ignition, i.e. the probability of ignition conditional on the occurrence of the lightning strike itself. The model was initially trained and validated using data covering all peninsular Spain plus the Balearic Islands, and we subsequently investigated its ability to anticipate ignitions in Portugal. Several variables related to fuel moisture content, vegetation structure and topography, as well as some factors related to polarity, intensity and density of impacts were used. The methodology is based on Random Forest modelling and other related techniques (bootstrapping, resampling and data stratification) to assess the accuracy of the probability obtained. Most of the existing approaches have focused on either the atmospheric conditions surrounding lighting and ignitions (Pérez-Invernón et al. 2021; Soler et al. 2021; Pineda et al. 2022) or on developing stationary likelihood maps (Nieto et al. 2012; Rodríguez-Pérez et al. 2020). Instead, here we delve into the spatial and landscape circumstances enabling natural wildfires. In this way, daily and moderate-resolution lightning ignition prediction for the whole Iberian Peninsula can be provided.

The core of our approach leverages a large and comprehensive dataset of cloud-to-ground lightning strikes recorded in the period 2009–2015. The dataset was compiled by the Spanish Meteorological Agency (AEMET) and provided by Meteogrid SL. AEMET’s lightning detection system for the Spanish territory is composed of 14 detectors on the Spanish mainland, 1 in the Balearic Islands and 5 in the Canary Islands (the latter were not used in this work); Meteo-France provides data from 10 of its detectors and the Portuguese Meteorological Agency adds 4 extra sensors. All data from these 34 detectors are consolidated in the calculation and tracking system at the AEMET’s climate service. The system is capable of detecting lightning strikes with a location accuracy of the order of 1 hm and a detection efficiency of 90% of total lightning strikes. The dataset used in this study consists of more than 17 million records in the period 2009–2020, reporting flash intensity, polarity and density. Lightning strikes were paired with actual lightning ignition data compiled by national agencies. Information about fire ignitions (location, date, cause and size) was retrieved from the Spanish (Estadística General de Incendios Forestales, EGIF; MAAyMA 2015) and Portuguese (Dispositivo Especial de Combate a Incêndios Florestais, DECIF; Pereira et al. 2011) databases. The Spanish dataset was used to identify those lightning strikes that started a fire while the Portuguese records were retained to evaluate the applicability of the model to the specific conditions of Portugal.

The response variable was constructed using historical ignitions in Spain. Approximately 2500 fires were matched to a lightning strike occurring within a radius of 1000 m and no more than 3 days after the ignition date (maximum holdover period according to Pineda et al. 2014). For each fire event, we retrieved all lightning strikes within a 3-day window. Then, we ranked lightning strikes by proximity to the fire’s ignition point. The closest lightning was associated with that ignition and further considered as lightning that triggered a fire when building the response variable. Only those flashes within a 1000-m distance from an ignition source were considered owing to the uncertainty in flash detection. From this information, we built a binary response variable combining strikes that triggered a fire (1 or presence) or were not related to a fire (0 or absence). The response variable comprised both lightning strikes on days where a wildfire occurred (fire days henceforth) and a balanced set of strikes from days without observed lightning-caused fires (non-fire days). We implemented a stratified sampling procedure to prevent undesired effects from the largely unbalanced response (more 0 s than 1 s) and minimise spatial autocorrelation.

To predict the probability of ignition, we employed factors related to vegetation, topography and others directly linked to lightning features. Live fuel moisture content was characterised using the Drought Code (DC; Van Wagner 1987) as proxy, calculated following the standard procedure described in Wagner and Pickett (1985). DC was retrieved from the Copernicus Emergency System, which produces daily gridded (at 25 km resolution) historical data for a variety of meteorological fire danger indices, including the Canadian Fire Weather Index, to which DC belongs. We calculated the daily dead fuel moisture content (dFMC) as described in Nolan et al. (2016), previously used successfully on the Iberian Peninsula (Boer et al. 2017; Resco de Dios et al. 2022). The method is based on the exponential decline in dFMC with increasing vapour pressure deficit (VPD) and, in turn, it calculates VPD from MODIS Land Surface Temperature (MOD11A1 Collection 6). The dFMC measures 1–10 h fine fuels moisture while DC relates to seasonal drought conditions potentially affecting the moisture content of live vegetation (Viegas et al. 2001). The dFMC was calculated using surface temperature and relative humidity daily data from the ERA5-Land dataset at a 9 km spatial resolution (Muñoz Sabater 2019). Secondly, elevation and relief curvature (derived from elevation via topographic position index; Weiss 2001) were selected to depict the influence exerted by the terrain in the ignition of natural-cause fires. We retrieved elevation data from the NASADEM global digital elevation model at 30 m spatial resolution (NASA JPL 2020). The last landscape feature incorporated into the model was vegetation height, acting as a proxy for vegetation structure and cover. We used the global forest canopy height map using a novel dataset calibrated combining Global Ecosystem Dynamics Investigation (GEDI) footprints of canopy structure (Dubayah et al. 2020) and Landsat Operational Land Imager spectral imagery. All variables were regridded from their original resolution to 1 × 1 km. Topographic variables and vegetation height were aggregated by their median value into the 1 × 1 km destination cell while dFMC and DC were resampled using bilinear interpolation downscaling from 9 × 9 to 25 × 25 km. Our model also accounts for lightning-related features, such as strike intensity and polarity, and density of flashes, as provided by the lightning detection network.

The model was trained using Random Forest (Breiman 2001), a popular and effective algorithm for wildfire distribution modelling (Bar Massada et al. 2013; Rodrigues and de la Riva 2014; Vecín-Arias et al. 2016; Su et al. 2018). We trained a total of 100 models exploring different iterations of the response variable. In each realisation, we extracted a balanced sample of 1-presence (flashes igniting a fire; 1/3 of the final sample) and 0-absence (lightning strikes non-conducive to fire; 2/3 of the final sample distributed as 1/3 in fire days and another 1/3 in non-fire days). The process was stratified using a 100 × 100 km grid so that at each grid cell, we extracted a balanced sample of 1-presence and 0-absence to reduce the potential spatial autocorrelation (Wang et al. 2016). It must be noted that we only analysed the April-to-October period, coinciding with the brunt of the fire season.

Model calibration was conducted using the caret package (Kuhn 2008) in the R environment for statistical computing (R Core Team 2021). For each model realisation, we split data into a calibration (70% of data) and testing dataset (30% of data). During the calibration phase, we implemented a 5-fold repeated cross-validation (three repetitions), optimising model parameters: number of trees (ntree; from 100 to 2000 trees with steps of 100) and number of variables to use at each split (mtry; consecutive values from 1 to 6). The minimum node size (nodesize) parameter was kept constant at its default value for regression (5) because it had little influence on performance. Once the model was calibrated, we evaluated the importance and explanatory sense of each covariate using partial dependence plots (Greenwell 2017), and the overall performance of each model via area under the receiver operating characteristic curve (AUC) (Bradley 1997), using the test sample in the latter and the out-of-bag sample in the first. Out of the set of 100 models, we selected the one attaining average prediction performance (AUC = 0.828; mtry = 2, ntree = 1000 and nodesize = 5) as the candidate model to forecast and map the daily evolution of lightning-related ignition probability. The model was subsequently evaluated calculating the AUC from the Portuguese fire data, which consisted in 105 lightning fires (>1 ha) in the period 2009–2015. Like the procedure followed in the calibration stage, we created multiple random samples of 0-absence to account for the non-occurrence of fires in Portugal. In this case, we created a balanced sample of 105 non-fire locations on fire days, placing as many absences as fires were reported. This procedure was repeated 100 times to check the potential effect of the random placement of 0-absence. Finally, to perform daily spatial predictions over the entire Iberian Peninsula, we used daily information about dFMC and DC. Vegetation height and topography were kept constant. We set constant values of intensity, polarity and density of strikes equal to their respective average value over the study area and period (i.e. −10 kA and five discharges, respectively).

Fig. 1 illustrates the comparison of fire size and seasonal distribution of lightning-caused fires in the region. Lightning-caused fires were significantly larger than human-caused fires either due to arson or negligence (Fig. 1a, b). Note that accidental and negligence fires were grouped together owing to differences in the cause classification systems between Portuguese and Spanish agencies. Both the average fire size (burned area divided by number of fire events) and the 95th percentile of the size distribution show the potential for larger burned area size that lightning fires have. The average fire size of lightning fires was at least two times larger than human-related ignitions, even three times in the case of Portugal (Table 1). The interannual comparison of fire size revealed the prominent role lightning fires played in the waves of fires during the years 1994 (Spain), 2003 and 2017 (Portugal). Their intra-annual distribution is clearly concentrated in the summer months (June–September), though in Spain they peak in August and July, whereas in Portugal, they are more evenly distributed during these summer months. (Fig. 1c). In any case, they seldom occur outside that temporal window, being scarce in April and October, though the agencies still reported some fires in both countries in these 2 months (1% of total).

Fig. 1.  (a) Average fire size (ha) by cause of ignition, year, and country. (b) 95th percentile of fire size (ha) by cause of ignition, year, and country. (c) Monthly distribution of lightning-caused fires by country. Only fires larger than 1 ha and with a known source of ignition were considered. The period of the data was chosen based on those years with reliable data in terms of classification schemes and compilation completeness (Pereira et al. 2011; MAAyMA 2015).

Table 1.  Summary of fire activity by cause of ignition and country.

The Random Forest models denoted a high predictive capability (calculated test samples) in Spain (mean AUC of 0.82; Fig. 1a) and the feasibility of extrapolating them to similar regions, as exemplified in Portugal (AUC = 0.74; Fig. 1b). The sensitivity of the models’ performance and extrapolation potential to the sampling of the response variable (in the case of Spain) and to the random location of the background sample (in Portugal) was low. AUC values ranged between 0.80 and 0.85 in the first case and between 0.70 and 0.78 in the latter. The AUCs showed frequency distributions close to Gaussian, peaking at the means of the reported AUCs.

The major contributing factors (Fig. 2c, d), DC and dFMC, are related to fuel dryness and prolonged drought periods. DC and dFMC showed very clear profiles of association with predicted ignition probability. The former is related to increasing probability up to a DC value of 300 for which the predicted probability increases from 0.28 to 0.49, levelling off at 0.42 beyond the 600 mark. dFMC shows the reverse profile, decreasing rapidly from its maximum at 0.5 probability at 6% moisture content, decreasing to 0.34 towards 13%. After that threshold, the probability continues to show a general decline, stabilising at ~0.3 when fine dead fuels have more than 20% moisture content. Ignition probability was strongly linked to elevation and vegetation height, although to a lesser extent than for the former factors. The probability of ignition by lightning increased with elevation, from 27 at sea level to 0.4 at 1000 m elevation. Likewise, vegetation layers 10–15 m tall were more prone to ignition from lightning (0.45–0.48). Lastly, less important but still contributing was the density of sensor-detected bolts and their intensity and polarity. The probability of ignition increased almost linearly with respect to the density of detected flashes, with a probability greater than 0.4 when more than 10 flashes occurred. Polarity and intensity showed the most erratic profile. Negative flashes at −10 kA showed the highest probability. Most of the flashes that started a fire were observed around this range, as can be seen in the percentile lines on the x-axis in Fig. 2h. Relief curvature was not selected as a meaningful driver in any model.

Fig. 2.  Summary of variable and model performance. (a) AUC value of the original model for Spain. (b) AUC values of the validation in Portugal. (c–h) Partial dependence plots of the explanatory factors obtained for the average model. The brown solid line indicates the average value of the response across models. The dashed green line with grey shading shows the LOESS smoothed response and the 95 percent confidence interval range from the 100 model iterations, respectively. Rug lines along the x-axis indicate the 10-fold percentile position of the variable.

Fig. 3 exemplifies the predicted probability over the Iberian Peninsula on 2 days with contrasting meteorological conditions. In general, the spatial patterns reflect the conditions described above, i.e. a higher probability in mountain ranges with a taller canopy layer. However, comparison between the pairs of predictions highlights the leading role played by meteorological conditions and fuel moisture. The late May 2013 prediction exemplifies the situation under mild off-season conditions, while the early June 2015 prediction illustrates the effect of a heat wave episode that boosts the probability of ignition across the Peninsula. The temporal variability can be also seen in Fig. 4c–h, with increasingly hazardous conditions from May to October. Likewise, the predicted probability varies yearly (Fig. 4a, b), though an underlying spatial pattern emerges from the comparison of all temporal aggregations matching the spatial footprint of the stationary drivers, i.e. elevation and vegetation height.

Fig. 3.  Spatial distribution of lightning ignition probability in the Iberian Peninsula: (a) prediction for milder conditions during 26 May 2013; (b) fire-prone conditions during a heat wave episode on 7 June 2015.

Fig. 4.  Spatial distribution of predicted probability of lightning ignitions in the Iberian Peninsula. Panels a and b display the annual average in 2013 (a) and 2012 (b). Panels c–h show the monthly averages in the period 2009–2015 from May (c) to October (h).

The Iberian Peninsula is the most fire-affected region in Europe (Camia et al. 2013). In this work, we produced a daily forecast of natural-caused wildfire ignition probability across the Iberian Peninsula, a region that despite being dominated by human-caused fires stands out as one of the main hotspots of lightning fires in the Western Mediterranean Basin. The benefits from managing natural ignitions are well documented, for example, in guiding fire reintroduction in heavily fire-excluded forests resilient or adapted to fire (Barros et al. 2021; Rodrigues et al. 2022).

Previous studies have set the basis for developing our methodological scheme to assess the probability of natural fire ignition over multiple temporal and spatial scales. Most of these were developed using binary classification approaches such as logit regression, assembling ignition locations along with environmental and climate-related drivers to derive the likelihood of ignition and identify the main driving forces (Castedo-Dorado et al. 2011; Nieto et al. 2012; Chuvieco et al. 2014). Castedo-Dorado et al. (2011) assessed the probability of ignition in the Spanish province (NUTS3, see https://ec.europa.eu/eurostat/web/nuts/background) of León, one of the major hotspots of lightning-caused fires. They employed binary logistic regression, analysing the stationary probability in the period 2002–2007, pointing to the confluence of thunderstorms over coniferous forests in high-altitude ranges as key factors, and reaching an AUC of 0.79. Using a similar approach but applied to two contrasting regions (the Autonomous Regions – NUTS2 – of Madrid and Aragón), Nieto et al. (2012) reached similar conclusions though with different levels of performance, the highest standing at AUC = 0.70. The method by Nieto et al. (2012) was subsequently extended over the entire Spanish mainland in Chuvieco et al. (2014). Their model was calibrated using categorical data on climate types, the number of lightning strikes and the Duff Moisture Code from the Canadian Fire Weather Index, giving an accuracy of 64.2%. Our proposal builds on these studies by delving into the temporal dimension.

We adopted a novel approach based on lightning strikes (Moris et al. 2020), in contrast to fire-based approaches (Amatulli et al. 2007; Nieto et al. 2012). That is, instead of building the response variable from the presence or absence of lightning-caused fires, we analysed the type of lightning and the landscape features conducive to fire, thus separating the ignition potential from the atmospheric conditions surrounding the occurrence of lightning and thunderstorms. In terms of risk assessment, we believe the strategy of breaking down ignition likelihood and the probability of lightning occurrence offers several advantages. Decoupling both phenomena allows us to better understand their relative importance while enabling embedding different models for lightning or thunderstorm forecast, or even real observations of lightning flashes. That is particularly important for short-term predictions. In turn, long-term modelling would also benefit from such an approach. Our forecast of ignition probability requires few weather-related factors (the same as are necessary for the calculation of the Canadian Fire Weather Index: rainfall, relative humidity, temperature and wind speed; Van Wagner 1987); thus, their prediction under climate scenarios (either Shared Socioeconomic Pathways or the former Representative Concentration Pathways) is already feasible. Along the same lines, several models are available to forecast lightning strikes (Woodard et al. 2014), hence giving the possibility to combine them with the ignition component. Furthermore, we translated model outcomes into spatial explicitly and scalable predictions as recommended in the literature (Vecín-Arias et al. 2016). Our modelling approach holds a high predictive accuracy (~0.82 AUC; Fig. 2a), which is noteworthy given the rare nature of thunderstorm-driven fires (Fernandes et al. 2021). Moreover, we demonstrated its suitability to perform forecasts in nearby fire-prone regions, evidenced by the successful validation in Portugal (Fig. 2b).

The observed relationships between ignition drivers and predicted likelihood are in line with their expected behaviour based on former modelling endeavours in the region. That is, low fuel moisture fosters ignition and higher chances of ignition are observed in mountain ranges. The important role of moisture content of the live/dead fuels in causing the ignition and propagation of new fire events has previously been documented in many works (Resco de Dios et al. 2021, 2022; Rodrigues et al. 2021; Baranovskiy and Kirienko 2022). Dry thunderstorms with a low precipitation rate preceded the largest lightning-caused fires recorded in the Mediterranean Basin (Pérez-Invernón et al. 2021). Indeed, the lack of rainfall is key during the survival phase of a fire (Pineda and Rigo 2017; Soler et al. 2021) and searching for these conditions has been at the core of many studies focused on the atmospheric and synoptic conditions surrounding lightning fires. But, according to our findings, it is the influence of fuel moisture and seasonal drought (DC) that seem more decisive, perhaps because they inherently involve reduced or complete lack of rainfall. It is nonetheless worth noting that the critical DC threshold was identified at a value of 300. According to the European Forest Fire Information System, this indicates low-to-moderate fire weather danger conditions, which suggests that lightning fires occur even during fairly short drought anomalies.

The type of vegetation was claimed as the most relevant factor for lightning ignition by Rodríguez-Pérez et al. (2020). We also found a clear association with vegetation, in terms of the height of the vegetation layer (though further research using fuel types would be useful), but it was outperformed by fuel moisture metrics and meteorological drought proxies (dFMC and DC). Topography also played a role, with most ignitions occurring at high elevations. Increased probability of ignition has been consistently found across the coastal and hinterland mountain ranges (the Spanish Plateau, Sierra da Estrela or the Iberian Mountains, among others). The altitude gradient in the Iberian Peninsula is related to the distribution and type of vegetation communities. But the arrangement of the relief, which runs parallel to the coast (east–west) in the Mediterranean and the Plateau, has been recognised to modulate ignition by influencing air humidity through Foehn effects (Rodrigues et al. 2019b) and is also linked to storm fronts and thunderstorm episodes (Soler et al. 2021).

Our findings were non-conclusive in terms of the role of the characteristics of the fire-causing flash. We did not find a strong signal in terms of intensity or polarity, though negative-current flashes seemed to be more strongly tied to fire ignition. Previous models in central Spain support increased ignition likelihood linked to the mean peak current of negative flashes (Vecín-Arias et al. 2016), while some studies conducted in Catalonia (northeast Spain) found no specific evidence of the role played by polarity (Pineda et al. 2014)

In this work, we produced a daily forecast of the probability of natural wildfires in the Iberian Peninsula. We analysed a large dataset of lightning strikes, consisting of more than 17 million observations retrieved from a comprehensive ground-based network of sensors. We combined lightning and wildfire data with environmental variables to train and test a binary Random Forest model.

We identified the critical thresholds promoting lightning-related ignitions across the Iberian Peninsula. Lightning fires tend to occur when dead fuels are below 10–13% moisture content, with moderate drought conditions (DC > 300). We observed a large temporal variability linked to these drivers, but also a structural spatial pattern related to topography and vegetation structure. Natural fires concentrate in mountain ranges above 1000 m with sufficient fuel load (vegetation height between 10 and 15 m tall). Our results revealed the major role played by fuel moisture content, the factor most sensitive to climate warming in the coming decades. This highlights the need for further research and consideration of natural ignitions in hazard mitigation plans, even though their contribution to fire activity is currently moderate.

Data about lightning strikes are managed and distributed by the Spanish Meteorological Agency (AEMET), and were acquired by METEOGRID SL to conduct this work. Data about lightning-related ignitions were provided by the Spanish and Portuguese agencies. Spanish data come from the EGIF database (Estadística General de Incendios Forestales; https://www.miteco.gob.es/es/biodiversidad/servicios/banco-datos-naturaleza/informacion-disponible/incendios-forestales.aspx), available on request to the Spanish Ministry for Ecological Transition and Demographic Challenge while Portuguese fire data can be accessed on the ICNF website (Instituto da Conservação da Natureza e das Florestas; http://www2.icnf.pt/portal/florestas/dfci/inc/estat-sgif). We retrieved elevation data from the NASADEM global digital elevation model (NASA JPL 2020). We used the global forest canopy height map by Dubayah et al. (2020).

The authors declare no conflicts of interest.

This work was financed by the projects FirEUrisk – Developing a Holistic, Risk-wise Strategy for European Wildfire Management, which received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement No. 101003890; and FIREPATHS (PID2020-116556RA-I00), funded by the Spanish Ministry of Science and Innovation.