Impact Statement
Wildfires are among the most destructive natural disasters, and due to climate change they are becoming more frequent, larger, and more intense globally. While wildfires can ignite naturally, most wildfires start from human activities, for example, burning of fields, sparks generated by heavy machines, or just carelessness. Meanwhile, the assessment of future wildfire risk tends to rely only on estimates of meteorologically driven fire weather from climate models, neglecting the role of humans and the land (fuel). In this paper, we evaluate different machine learning-based models for investigating wildfire risk at the end of the current century while attributing the risk to meteorology, land cover, topography, and population. We use the best trained model to uniquely explore the future risk of wildfires in Southern Europe under different climate and land use scenarios.
1. Introduction
Wildfires are one of the most adverse natural hazards in Southern Europe, where it has significant impacts on natural systems, humans, and economic activities (European Environment Agency, 2024). In the Mediterranean alone, the total burned area reached 0.5 million hectares in 2010; this exceeded 0.8 million hectares during the 2017 fire season. In total, the Mediterranean accounts for roughly 85% of Europe’s total burned area between 1999 and 2016 (de Rigo et al., Reference de Rigo, Libertà, Houston Durrant, Artes Vivancos and San-Miguel-Ayanz2017). While wildfire is a well-established risk due to the inherent vulnerability of the regional climates and landscapes to fire, recent observations clearly suggest that wildfire regimes in Southern Europe are becoming more extreme: fire seasons are becoming longer while wildfires are harder to control, resulting in an increase in the areas burned and losses suffered. For instance, the 2025 European fire season was historically extreme, with burned area exceeding 1.4 million hectares and several regionally unprecedented wildfires (Keeping et al., Reference Keeping, Zachariah, Haas, Grillakis, Barnes, Clay, Ekberzade, Jaupaj, Andreia, Trigo, Voulgarakis and Friederike2026). The Intergovernmental Panel on Climate Change (IPCC) has assessed with high confidence that extreme fire weather conditions will continue to increase with further warming, including in Europe (Intergovernmental Panel On Climate Change (IPCC), 2023). Meanwhile, the recent European Climate Risk Assessment underlines that the increased wildfire risk is a compound phenomenon driven not only by fire weather and amplified by climate change but also by factors such as land abandonment, fuel accumulation, forest homogenization, and other human-related factors (European Environment Agency, 2024). Thus, while extreme temperatures and fire weather conditions play important roles in creating fire-prone conditions, they are not enough on their own to explain the wildfire patterns observed in Southern Europe and the Mediterranean, where ignitions are predominantly caused by humans (Ganteaume et al., Reference Ganteaume, Camia, Jappiot, San-Miguel-Ayanz, Long-Fournel and Lampin2013). Rather, wildfire occurrence and the severity of the associated impacts result from the interaction of multiple drivers including the hydro-meteorological conditions, land cover characteristics, and human activities (Dorph et al., Reference Dorph, Marshall, Parkins and Penman2022; Mhawej et al., Reference Mhawej, Faour and Adjizian-Gerard2015; Oliveira et al., Reference Oliveira, Oehler, San-Miguel-Ayanz, Camia and José2012). Also, different types of land use exhibit different fire risks and are thereby associated with different fire ignition mechanisms (Bountzouklis et al., Reference Bountzouklis, Fox and Di Bernardino2023; Galizia et al., Reference Galizia, Curt, Barbero and Rodrigues2021).
Several studies have already investigated the response of European wildfire activity to future climate change (Garroussi et al., Reference Garroussi, Di Giuseppe, Barnard and Wetterhall2024; Zhang et al., Reference Zhang, Wang, Yang and Liu2024). However, even though the Shared Socioeconomic Pathways (SSPs) (O’Neill et al., Reference O’Neill, Tebaldi, van Vuuren, Eyring, Friedlingstein, Hurtt, Knutti, Kriegler, Lamarque, Lowe, Meehl, Moss, Riahi and Sanderson2016; Riahi et al., Reference Riahi, van Vuuren, Kriegler, Edmonds, O’Neill, Fujimori, Bauer, Calvin, Dellink, Fricko, Lutz, Popp, Crespo Cuaresma, KC, Leimbach, Jiang, Kram, Rao, Emmerling, Ebi, Hasegawa, Havlik, Humpenoder, Aleluia Da Silva, Smith, Stehfest, Bosetti, Eom, Gernaat, Masui, Rogelj, Strefler, Drouet, Krey, Luderer, Harmsen, Takahashi, Baumstark, Doelman, Kainuma, Klimont, Marangoni, Lotze-Campen, Obersteiner, Tabeau and Tavoni2017) used in climate research and by the IPCC (Intergovernmental Panel On Climate Change (IPCC), 2023) provide a structured framework for exploring plausible futures that combine representative pathways of greenhouse gas (GHG) emissions with socioeconomic development pathways, there are only few climate model simulations that also explicitly consider future changes in, for example, land use or population, and virtually all of these stem from coarse-resolution global climate models. As a result, existing projections of future wildfire risk generally rely on purely climatological indicators like the Fire Weather Index (Garroussi et al., Reference Garroussi, Di Giuseppe, Barnard and Wetterhall2024), neglecting the influence of potential socioeconomic development, including land use change (LUC), despite strong evidence that these factors jointly shape future fire regimes.
To better understand the complex mechanisms governing the spatial and temporal occurrence of wildfires under current and future climatic conditions, this study projects future wildfire risk by combining high-resolution regional climate simulations, land cover, and population projections with data-driven modeling. Here, we focus on Southern Europe, but the same methodology can be applied in other regions, where comparable training data are available. We use a regional single-model/multi-scenario ensemble of the Canadian Regional Climate Model version 5 (CRCM5), covering the European CORDEX domain at a high spatial resolution (12 km) (Jacob et al., Reference Jacob, Teichmann, Sobolowski, Katragkou, Anders, Belda, Benestad, Boberg, Buonomo and Cardoso2020) that was produced within the second phase of the ClimEx project (Leduc et al., Reference Leduc, Mailhot, Frigon, Martel, Ludwig, Brietzke, Giguere, Brissette, Turcotte, Braun and Scinocca2019; Asselin et al., Reference Asselin, Leduc, Paquin, de Noblet-Ducoudré, Rechid and Ludwig2024). This ensemble combines climate projections for SSP1–2.6 and SSP3–7.0 with projections of land cover changes to explore the joint influence of climate and land use on climate extremes. For example, Asselin et al. (Reference Asselin, Leduc, Paquin, de Noblet-Ducoudré, Rechid and Ludwig2024) used these simulations to show how large reforestation under SSP1–2.6 can reduce summer heat stress in Central Europe. Finally, high-resolution, SSP-consistent population projections (Wang et al., Reference Wang, Meng and Long2022) provide the demographic information required to assess future human influences on wildfire.
Machine learning (ML) and other artificial intelligence (AI)-based models have emerged as important tools for understanding extreme weather and climate events (Reichstein et al., Reference Reichstein, Camps-Valls, Stevens, Jung, Denzler, Carvalhais and Prabhat2019; Camps-Valls et al., Reference Camps-Valls, Fernández-Torres, Cohrs, Höhl, Castelletti, Pacal, Robin, Martinuzzi, Papoutsis, Prapas, Pérez-Aracil, Weigel, Gonzalez-Calabuig, Reichstein, Rabel, Giuliani, Mahecha, Popescu, Pellicer-Valero, Ouala, Salcedo-Sanz, Sippel, Kondylatos, Happé and Williams2025), including wildfire risk assessment (Jain et al., Reference Jain, Coogan, Subramanian, Crowley, Taylor and Flannigan2020). Several recent studies have explored data-driven and deep learning models as a means to quantify risk under current climatic conditions, for example, to inform disaster risk management and planning (Ochoa et al., Reference Ochoa, Bar-Massada and Chuvieco2024; Zhang et al., Reference Zhang, Wang, Yang and Liu2024). However, to our knowledge, this study is the first to couple high-resolution, SSP-consistent regional climate projections that adjust both climate forcing and land cover with a data-driven modeling framework to assess future wildfire risk in Europe.
The paper is organized in three parts. First, we develop an ML model to estimate daily wildfire risk at a pixel level by combining detailed climate projections with land cover, population, topography, and other relevant predictors. Second, we apply the principles of explainable AI (Xu et al., Reference Xu, Uszkoreit, Du, Fan, Zhao and Zhu2019) to quantify the contribution of the different predictors to wildfire risk. Finally, we use the validated model to project end-of-century wildfire risk under the SSP1–2.6 and SSP3–7.0 scenarios at daily pixel resolution.
2. Methodology
Figure 1 provides an overview of the methodological framework with details provided in the sections below.
Overview of the modeling framework used to predict wildfire occurrence and project future risk under different climate and land use scenarios.
Figure 1. Long description
The flowchart is divided into three numbered sections.
Panel 1. Datasets and processing. At the top is a map of Europe. Below the map, the data is split into two categories. The Predictand section includes Historical Fires E F F I S from June to October, 2008 to 2020. The Predictors section includes weather, land cover, topography, and population density, using a 50 percent fire and 50 percent no fire data split on a 12 by 12 km grid.
Panel 2. Model training and selection. The process uses a 75 percent training and 25 percent testing split. Training involves hyperparameter tuning with a 5-fold C V. Models listed are Random Forest, k N N, Logistic regression, and X G Boost. Performance metrics include Logarithmic loss, A U C, and F 1 score. X A I model explainability is assessed using S H A P.
Panel 3. Project future wildfire risk. This section describes running the M L model daily for each C R C M 5 simulation with 4 members. It compares three scenarios along a timeline. The Present scenario (1991 to 2010) uses present land cover. Two future scenarios (2081 to 2100) are S S P 1-2.6 and S S P 3-7.0, both evaluated with no L U C and including L U C. Footnotes specify that present land cover corresponds to 2015 and future land cover corresponds to 2100 including L U C.
2.1. Climate data and projections
Climate projections are obtained from a single-model/multi-scenario ensemble of CRCM5 regional climate simulations (Hersbach et al., Reference Hersbach, Bell, Berrisford, Hirahara, Horányi, Muñoz-Sabater, Nicolas, Peubey, Radu, Schepers, Simmons, Soci, Abdalla, Abellan, Balsamo, Bechtold, Biavati, Bidlot, Bonavita, De Chiara, Dahlgren, Dee, Diamantakis, Dragani, Flemming, Forbes, Fuentes, Geer, Haimberger, Healy, Hogan, Hólm, Janisková, Keeley, Laloyaux, Lopez, Lupu and Radnoti2020; Šeparović et al., Reference Šeparović, Alexandru, Laprise, Martynov, Sushama, Winger, Tete and Valin2013) that was produced within the second phase of the ClimEx project by dynamically downscaling the first four members (rXi1p1f1, X = 1–4) of the MPI-ESM-1.2 global ensemble from the Max Planck Institute for Meteorology (Mauritsen et al., Reference Mauritsen, Bader, Becker, Behrens, Bittner, Brokopf, Brovkin, Claussen, Crueger, Esch, Fast, Fiedler, Fläschner, Gayler, Giorgetta, Goll, Haak, Hagemann, Hedemann, Hohenegger, Ilyina, Jahns, Cuesta, Jungclaus, Kleinen, Kloster, Kracher, Kinne, Kleberg, Lasslop, Kornblueh, Marotzke, Matei, Meraner, Mikolajewicz, Modali, Möbis, Müller, Julia, Christine, Notz, Nyawira, Paulsen, Peters, Pincus, Pohlmann, Pongratz, Popp, Raddatz, Rast, Redler, Reick, Rohrschneider, Schemann, Schnur, Schulzweida, Six, Stein, Stemmler, Stevens, von Storch, Tian, Voigt, Vrese, Wieners, Wilkenskjeld, Winkler and Roeckner2019). CRCM5 is coupled to the Canadian Land Surface Scheme (Verseghy, Reference Verseghy1991; Verseghy et al., Reference Verseghy, McFarlane and Lazare1993), and the ensemble simulations were set up at a horizontal resolution of 0.11° (approximately 12 km) over the usual EURO-CORDEX domain for the periods 1979–2015 and 2064–2100 (Jacob et al., Reference Jacob, Teichmann, Sobolowski, Katragkou, Anders, Belda, Benestad, Boberg, Buonomo and Cardoso2020). For each member of the global climate model ensemble, three downscaled model configurations are available to examine the relative contributions of LUC and GHG emissions to the forcing at pixel level (i.e., present GHG and present LUC; future GHG and present LUC; future GHG and LUC). The MPI-ESM-1.2 ensemble members differ only by perturbed initial conditions, to allow sampling of the internal variability alongside the forced response. A reanalysis-driven CRCM5 model simulation using ERA5 (1979–2020) is used as the present-day reference for constructing ML predictors and for validation (Supplementary Figure A6 in the Supplementary Materials). For further information about the CRCM5 climate simulations, we kindly refer to (Asselin et al., Reference Asselin, Leduc, Paquin, de Noblet-Ducoudré, Rechid and Ludwig2024).
In the following, we compare a present baseline (1991–2010) and a late-century window (2081–2100) to explore long-term, policy-relevant changes. We consider two future emission scenarios: the low-emission SSP1–2.6 scenario and the high-emission SSP3–7.0 scenario. These two scenarios provide contrasting end-of-century outcomes that bound likely climate variations and therefore help illustrate how wildfire risk responds to different mitigation pathways. For present land cover, a map derived from Reinhart et al. (Reference Reinhart, Hoffmann and Rechid2022) for the year 2015 is used, whereas for future scenarios, projected LUC maps for 2100 under SSP1–2.6 and SSP3–7.0 are obtained from the Land Use Harmonized Dataset version 2 (LUH2) (Hurtt et al., Reference Hurtt, Chini, Sahajpal, Frolking, Bodirsky, Calvin, Doelman, Fisk, Fujimori, Klein, Hasegawa, Havlik, Heinimann, Humpenöder, Jungclaus, Kaplan, Kennedy, Krisztin, Lawrence, Lawrence, Ma, Mertz, Pongratz, Popp, Poulter, Riahi, Shevliakova, Stehfest, Thornton, Tubiello, van Vuuren and Zhang2020). Future population data are obtained from (Wang et al., Reference Wang, Meng and Long2022), which provides projections for the SSP1, SSP3, and SSP5 narratives at 5-year intervals. We use the final available time step (2095). In the following, simulations driven by climate forcing only, with land cover fixed at present-day conditions, are referred to as no LUC, whereas simulations that include projected LUCs are denoted as incl. LUC for the SSP1–2.6 and SSP3–7.0 scenarios.
As indicated in Supplementary Figures A1 and A2 in the Supplementary Materials, the CRCM5 simulations exhibit small biases for temperature and precipitation compared to ERA5. Since we are comparing anomalies across different emission and land use scenarios and present and future climates using the same regional climate model, we did not adjust for these biases.
2.2. Assembling the dataset for the ML model
The predictand in the present study is a binary fire occurrence variable derived from the European Forest Fire Information System (EFFIS) (San-Miguel-Ayanz et al., Reference San-Miguel-Ayanz, Schulte, Schmuck, Camia, Strobl, Libertà, Giovando, Boca, Sedano, Kempeneers, McInerney, Withmore, Oliveira, Rodrigues, Durrant, Corti, Oehler, Vilar, Amatulli and Tiefenbacher2012). This analysis focuses on large wildfires, defined as events with a burned area greater than 10 ha occurring during the fire season (June–October). Following (Ochoa et al., Reference Ochoa, Bar-Massada and Chuvieco2024), a filtering was performed to remove fires associated with controlled agricultural practices by removing events for which more than 80% of the burned area overlaps agricultural land. We restrict our analysis to the period 2008–2020. The year 2008 marks the start of the EFFIS burned-area records, whereas 2020 is chosen to match the final year covered by the ERA5-driven CRCM5 simulations (see Section 2.1). For each grid cell and date, the target variable is set to 1 if any part of a filtered burned-area polygon intersects the cell on that day, and 0 otherwise. Finally, to ensure a balanced dataset for the supervised learning, an equal number of no-fire locations and dates were generated so that the training and testing dataset comprises 50% fire and 50% no-fire data.
The predictor selection includes a combination of hydro-meteorological variables representing fire weather conditions (temperature, precipitation, wind speed, relative humidity, and soil moisture), land cover, topography (elevation, slope, and aspect) as well as population density. This selection is similar to studies such as Oliveira et al. (Reference Oliveira, Oehler, San-Miguel-Ayanz, Camia and José2012) and Meuriot et al. (Reference Meuriot, Solo Martin, Zhang, Camara Pereira and Drews2025). A full list of the predictors is provided in Supplementary Table A1 in the Supplementary Materials. In the present study, only predictors for which projections under future climate scenarios are available are selected. The hydro-meteorological variables used to train the ML model are obtained from an ERA5-driven simulation with the CRCM5 (Hersbach et al., Reference Hersbach, Bell, Berrisford, Hirahara, Horányi, Muñoz-Sabater, Nicolas, Peubey, Radu, Schepers, Simmons, Soci, Abdalla, Abellan, Balsamo, Bechtold, Biavati, Bidlot, Bonavita, De Chiara, Dahlgren, Dee, Diamantakis, Dragani, Flemming, Forbes, Fuentes, Geer, Haimberger, Healy, Hogan, Hólm, Janisková, Keeley, Laloyaux, Lopez, Lupu and Radnoti2020; Šeparović et al., Reference Šeparović, Alexandru, Laprise, Martynov, Sushama, Winger, Tete and Valin2013), which poses as a reference dataset for present-day conditions. The ERA5-driven simulation is set up in the same way as the future climate simulations (see Section 2.1). Land cover categories include urban, grassland, forest (deciduous), forest (evergreen), and croplands. The historical population data are obtained from the WorldPop Global Population dataset (Tatem, Reference Tatem2017). To account for multicollinearity between the predictor candidates, pairwise Pearson correlations were calculated, and one variable from each highly correlated pair was removed. A threshold of 0.75 was applied, resulting in the removal of three predictors (pr monthly, N mean, and above30) as shown in the correlation matrix (see Supplementary Figure A7 in the Supplementary Materials).
2.3. ML framework for wildfire risk modeling
The predictand and predictors were aggregated to a common 12 × 12 Scikit-learn library: random forest (RF), k-nearest neighbors (kNN), logistic regression (LR), and XGBoost (XGB). We selected these models, as they represent linear and nonlinear approaches with different levels of interpretability and computational complexity. Moreover, good performance for wildfire prediction has previously been demonstrated.
We used 75% of the dataset for training and 25% for testing, with fivefold cross-validation applied to the training set for hyperparameter tuning. The predictors were standardized by subtracting the mean and dividing by the standard deviation of the training set. Model performance on the independent test set was evaluated using the Area Under the receiver operating characteristic Curve (AUC), logarithmic loss (logloss), and F1-scores (Murphy, Reference Murphy2012). The full set of tested hyperparameters, cross-validation results, and test-set performance metrics can be found in the Supplementary Materials (Supplementary Table A2). All models showed good performance in predicting wildfire occurrence. However, XGB achieved the best overall performance on the test set (AUC = 0.93, F1 = 0.85, logloss = 0.35), slightly outperforming RF (AUC = 0.92, F1 = 0.85, logloss = 0.36), and exceeding kNN and LR. In addition to its strong predictive performance, XGB was also more computationally efficient than RF, requiring 16.05 seconds of cumulative CPU time for training and hyperparameter optimization over the domain on our computers, compared to 51.45 seconds for RF. Given its combination of high predictive skill, ability to capture nonlinear interactions, and lower computational cost, we used the XGB model for further wildfire analyses. All results (e.g., Figure 2) refer to this model unless stated otherwise.
SHAP summary on the testing data from the XGB model. Panel (a) shows the mean absolute SHAP values of the top 10 predictors, indicating overall feature importance. Panel (b) shows the distribution of SHAP values for each predictor, illustrating their impact on model output; color encodes feature magnitude (orange = high, blue = low).
Figure 2. Long description
Panel a is a horizontal bar chart titled mean absolute S H A P value on the x-axis, ranging from 0.0 to 0.8. Ten features are listed on the y-axis in descending order of importance. Soil moisture has the highest value at approximately 0.7, followed by Slope less than 15 degrees at 0.55, 3-month precipitation at 0.5, and Relative humidity at 0.4. The remaining features—Grassland, 15 degrees less than slope less than 30 degrees, Max temperature, Population density, Broadleaf forest, and Weekly precipitation—all fall between 0.2 and 0.3.
Panel b is a beeswarm plot with the x-axis labeled S H A P value, ranging from minus 2 to 2. It shows the distribution of impacts for each feature. Soil moisture shows a dense cluster of blue dots on the positive side and a long tail of orange dots on the negative side. Slope less than 15 degrees and 3-month precipitation show long orange tails on the negative side. Relative humidity shows a blue tail on the positive side.
A vertical color bar on the right indicates Feature value, with blue at the bottom labeled Low and orange at the top labeled High.
To validate the performance of the XGB model, we compared the spatial distribution of mean predicted probabilities for present climate with the observed burned area from EFFIS (see Supplementary Figure A6 in the Supplementary Materials). Results suggest that the ML model rightly captures the main historical fire-prone regions. Likewise, the trained model seems to reproduce the interannual variability in fire occurrence, correctly identifying years with elevated fire activity such as 2012 and 2017. Although the predicted magnitudes differ from the observations in some years, the temporal agreement suggests that the model rightly identifies fire-prone conditions, despite occasional under or overestimation.
To interpret the model behavior, SHapley Additive exPlanations (SHAP) were applied to the testing data, providing insights into the contribution of individual predictors, that is, from an explainable AI perspective (Lundberg & Lee, Reference Lundberg and Lee2017).
3. Results
3.1. ML model explainability
Figure 2 depicts the feature importance, calculated as the mean absolute SHAP values, for the top 10 predictors, as well as the contribution of each feature to the model output. Soil moisture is found to be the most influential feature followed by the slope fraction less than 15°, 3 monthly precipitation and relative humidity. The remaining six predictors are about equally important. Low soil moisture values indicate an increased wildfire risk (higher model output value), while low values of the 3 monthly precipitation (third most influential feature) similarly affirm the general importance of extended dry periods on wildfire risk. Low values of relative humidity and high values of maximum temperature (fourth and sixth most influential features, respectively) also increases the predicted risk as would be expected.
In addition to the hydro-meteorological variables, topography and land cover also influence the occurrence of wildfires. High fractions of slope below 15° (second most influential feature), which are indicative of flatter terrains, are associated with lower wildfire risk (decreases the model output value). Land cover variables, such as grassland and broadleaf forest, influence the wildfire predictions; however, their SHAP distributions suggest nonlinear and context-dependent effects, suggesting that vegetation might modulate wildfire risk through their interaction with climatic conditions. Finally, a high population density (eighth most influential feature) is found to be associated with an increased risk of wildfire, reflecting the role of human activity in fire ignition.
The results obtained from the XGB model agree well with previous wildfire studies, supporting a physical interpretation of the SHAP features, while providing a useful measure of relative importance. Soil moisture has often been shown to be a key control on fire occurrence, with dry soils reflecting prolonged moisture deficits that promote vegetation stress and fuel drying (Williams & Abatzoglou, Reference Williams and Abatzoglou2016). Similarly, the importance of multi-month precipitation aligns with studies demonstrating that antecedent drought conditions, rather than short-term weather alone, strongly influence wildfire risk by controlling fuel availability and moisture content (Littell et al., Reference Littell, McKenzie, Peterson and Westerling2009). In addition to these hydro-meteorological factors, topography and land cover are also widely recognized as key drivers of wildfire patterns. Fire spread is generally slower on gentle slopes than on steeper terrain, where uphill fire propagation is enhanced by heat transfer (Rothermel, Reference Rothermel1972; Bradstock et al., Reference Bradstock, Hammill, Collins and Price2010). Land cover further modulates wildfire risk through its interaction with climatic conditions, leading to nonlinear and context-dependent effects in which fuel availability and moisture can alternately limit the area burned depending on environmental conditions (Kampf et al., Reference Kampf, Stevens-Rumann, Nunes, Sequeira, Rego, Fernández, Hernández-Duarte, Mosso, Francois and Miranda2025). Other studies also explain why grasslands emerge as important predictors in the model, since these fuel types can promote fire spread under favorable climatic conditions and are among the land cover classes most prone to fire in the Mediterranean (Moreira et al., Reference Moreira, Viedma, Arianoutsou, Curt, Koutsias, Rigolot, Barbati, Corona, PedroVaz, Mouillot and Bilgili2011). Finally, the positive contribution of population density to wildfire risk is consistent with the established link between human activity and fire ignitions (Balch et al., Reference Balch, Bradley, Abatzoglou, Nagy, Fusco and Mahood2017). In general, the agreement between the relationships learned by the ML model and the established physical and socio-environmental drivers of the occurrence of wildfires increases confidence in both the predictive skill and in the interpretability of the approach.
3.2. End-of-century changes in the SSP1–2.6 and SSP3–7.0 scenarios
3.2.1. Land cover, weather, and population future changes
Figure 3 shows an overview of projected changes embedded in the SSP1–2.6 and SSP3–7.0 simulations, averaged over the domain area. Detailed maps are provided in the Supplementary Materials (Supplementary Figures A3 and A4). The projected changes in land use and population are in general consistent with the SSP narratives. In SSP1, the decrease in grassland fraction together with the increase in deciduous forest fraction reflects sustainability-oriented land management, including afforestation and ecosystem restoration. In SSP3, the strong reduction in crop fraction, accompanied by increases in grassland and deciduous forest fractions is consistent with agricultural decline and land abandonment under a fragmented, low-development future (O’Neill et al., Reference O’Neill, Tebaldi, van Vuuren, Eyring, Friedlingstein, Hurtt, Knutti, Kriegler, Lamarque, Lowe, Meehl, Moss, Riahi and Sanderson2016; Hurtt et al., Reference Hurtt, Chini, Sahajpal, Frolking, Bodirsky, Calvin, Doelman, Fisk, Fujimori, Klein, Hasegawa, Havlik, Heinimann, Humpenöder, Jungclaus, Kaplan, Kennedy, Krisztin, Lawrence, Lawrence, Ma, Mertz, Pongratz, Popp, Poulter, Riahi, Shevliakova, Stehfest, Thornton, Tubiello, van Vuuren and Zhang2020; Riahi et al., Reference Riahi, van Vuuren, Kriegler, Edmonds, O’Neill, Fujimori, Bauer, Calvin, Dellink, Fricko, Lutz, Popp, Crespo Cuaresma, KC, Leimbach, Jiang, Kram, Rao, Emmerling, Ebi, Hasegawa, Havlik, Humpenoder, Aleluia Da Silva, Smith, Stehfest, Bosetti, Eom, Gernaat, Masui, Rogelj, Strefler, Drouet, Krey, Luderer, Harmsen, Takahashi, Baumstark, Doelman, Kainuma, Klimont, Marangoni, Lotze-Campen, Obersteiner, Tabeau and Tavoni2017). The projected population decrease toward the end of the century in both SSP1 and SSP3, which aligns with SSP-based demographic projections; this is driven by low fertility rates in SSP1 and regional socioeconomic constraints in SSP3 (Samir & Lutz, Reference Samir and Lutz2017).
Domain-averaged changes in land cover, population, and key fire-season weather variables under present and future scenarios. (a) Percent change in land cover categories and population, (b) fire-season maximum temperature, and (c) fire-season precipitation sums. All panels show the mean values for the present period (1991–2010) and projected changes under SSP1–2.6 and SSP3–7.0 scenarios (2081–2100).
Figure 3. Long description
Panel a is a bar chart showing Change relative to present (%) for six categories across S S P 1 and S S P 3. The y-axis ranges from -60 to 20. In S S P 1, Crop is slightly positive, Grassland is -18, Forest (Evergreen) is +8, Forest (Deciduous) is +20, Urban is +10, and Population is -42. In S S P 3, Crop is -17, Grassland is +17, Forest (Evergreen) is +10, Forest (Deciduous) is -3, Urban is -10, and Population is -57.
Panel b is a bar chart of Max. Temperature in degrees Celsius. Present day is 19.3. S S P 1-2.6 shows 19.9 (+0.60) for no L U C and 19.8 (+0.50) for incl. L U C. S S P 3-7.0 shows 22.4 (+3.10) for no L U C and 22.5 (+3.20) for incl. L U C.
Panel c is a bar chart of Precipitation sum in millimeters. Present day is 2.66. S S P 1-2.6 shows 2.72 (+0.06) for no L U C and 2.76 (+0.10) for incl. L U C. S S P 3-7.0 shows 2.60 (-0.06) for no L U C and 2.58 (-0.08) for incl. L U C.
Legend indicates: light blue for Crop, medium blue for Grassland, dark green for Forest (Evergreen), light green for Forest (Deciduous), orange for Urban, and red for Population. Grey bars represent no L U C and orange-brown bars represent incl. L U C.
Considering the relative magnitudes of changes due to climate warming, the largest increase is, as expected, found in the SSP3–7.0 simulations (+3.1/+3.2 °C) compared to a much smaller increase in the SSP1–2.6 simulations (+0.6/+0.5 °C). For precipitation, the SSP1–2.6 simulation shows a slight increase (+0.06/+0.10 mm), whereas the SSP3–7.0 simulation shows a small decrease (−0.06/−0.08 mm) – none of these is significant.
More pronounced differences can be found when comparing the no LUC and incl. LUC simulations. In the SSP1–2.6 scenario, the no LUC simulation is seen to be warmer and dryer than in the incl. LUC simulation, where SSP1 land use is assumed. In contrast, in the SSP3–7.0 scenario, the no LUC simulation is on average cooler and wetter than the incl. LUC simulation, which implements a land use scenario corresponding to SSP3. Although these results are based on domain averages and therefore mask important spatial and temporal variability, they point to the potential influence of LUC on key hydro-meteorological variables. Thus, the relatively subtle differences at the aggregated scale may indicate emerging signals of LUC effects, consistent with Asselin et al. (Reference Asselin, Leduc, Paquin, de Noblet-Ducoudré, Rechid and Ludwig2024).
3.3. Future changes in wildfire risk
The XGB model is run daily for all members of the CRCM5 ensemble, using the end-of-century SSP1 and SSP3 population changes. A grid cell for a given day is considered at risk of wildfire if the classification probability derived from the XGB model is greater than 0.5.
Panel (a) from Figure 4 shows the spatial average wildfire risk expressed as the percentage of time steps for both the present-day conditions (1991–2010) and projected end-of-century conditions (2081–2100) under SSP1–2.6 and SSP3–7.0, with and without LUC, with values in parentheses indicating changes relative to the present. Under SSP3–7.0, the model projects an increase in wildfire risk of more than 7%, whereas under SSP1–2.6 there is a reduction of about 3% when LUC is included and around 1% when it is not. These reductions are consistent with the relatively small increase in temperature and the slight rise in precipitation in the SSP1–2.6 scenario, which together lead the model to predict less favorable conditions for wildfire occurrences. Although the meteorological changes are modest, their combined effect with changes in land cover (see Figure 3) helps explain why the changes in forest fire risk in the model are more pronounced than those of the meteorological variables alone.
Spatial distribution and spatial mean of wildfire risk in Southern Europe, expressed as the percentage of time steps with wildfire risk during the fire season (June–October). (a) Domain-averaged wildfire risk for the present period (1991–2010) and projected end-of-century conditions (2081–2100) under SSP1–2.6 and SSP3–7.0, with and without LUC; values in parentheses indicate changes relative to the present. (b) Present-day wildfire risk (1991–2010), averaged across the four CRCM5 ensemble members. (c, d) Projected changes in wildfire risk under SSP1–2.6 (2081–2100), with (c) including LUC and (d) excluding LUC. (e, f) Projected changes under SSP3–7.0, with (e) including LUC and (f) excluding LUC. Positive (negative) values indicate an increase (decrease) in the percentage of time steps at risk relative to the present period.
Figure 4. Long description
The figure consists of six panels labeled a through f.
Panel a is a bar chart titled Spatial mean wildfire risk. The y-axis represents Mean percent time step at risk, ranging from 10 to 30.
- Present (1991-2010) is a blue bar at 19.68.
- S S P 1-2.6 (2081-2100) shows two bars: incl. L U C (orange) at 16.35 (minus 3.34) and no L U C (grey) at 18.53 (minus 1.15).
- S S P 3-7.0 (2081-2100) shows two bars: incl. L U C (orange) at 27.19 (plus 7.50) and no L U C (grey) at 26.69 (plus 7.01).
Panel b is a map of the Mediterranean showing Present (1991-2010) risk. A color scale from white (0) to black (100) indicates the percent time step at risk during fire season. High risk is concentrated in the Iberian Peninsula, Southern Italy, and the Balkans.
Panels c and d show the difference between S S P 1 and Present for incl. L U C and no L U C respectively. A diverging color scale from dark blue (minus 36) to dark red (plus 36) is used. Panel c shows significant blue areas (decreased risk) in the Balkans and Turkey, while panel d shows more neutral or light blue tones.
Panels e and f show the difference between S S P 3 and Present for incl. L U C and no L U C respectively. Both maps are dominated by red and dark red pixels across the entire Mediterranean basin, indicating a widespread increase in wildfire risk, with the highest increases in the Iberian Peninsula, Italy, and the Balkans.
Panel (b) shows the percentage of time between June and October (1991–2010) where an area is at risk of wildfire, averaged over four ensemble simulations. Areas that are particularly prone to wildfires include the north of Portugal as well as Western Spain, Italy, and the Balkan countries. Similar regions are identified in the historical risk maps produced by Oliveira et al. (Reference Oliveira, Oehler, San-Miguel-Ayanz, Camia and José2012) and Ochoa et al. (Reference Ochoa, Bar-Massada and Chuvieco2024). Panels (c–f) show changes in the percentage of time with a risk of wildfire for the SSP1–2.6 and SSP3–7.0 simulations (with panels d and f corresponding to no LUC simulations and panels c and e corresponding to incl. LUC simulations).
Considering the SSP3–7.0 scenario, both the no LUC and incl. LUC simulations show an increase in the percentage of time with a risk of wildfire by the end of the century. The strongest increases are found in regions where the percentage of time with a risk of wildfire is already high in the present-day simulation, indicating that some regions are at high risk of wildfire for much of the fire season. This increase is more pronounced in the incl. LUC simulation than in the no LUC simulation, suggesting that LUC amplifies the effect of the warmer and drier future climate conditions, together with the increase in grassland fraction shown in Figure 3. In contrast, under SSP1–2.6, the no LUC simulations show only limited changes on the predicted wildfire risk, with small increases in parts of northern Spain, southern Italy, and some Balkan regions, while most of southern Europe exhibits little change or even a decrease. In the incl. LUC simulation, decreases are more evident across eastern parts of the domain, as well as in northern Spain and Italy. This pattern is consistent with Figure 3, where the addition of LUCs leads to cooler and wetter conditions than in the no LUC case.
Future wildfire risk in Southern Europe is clearly found to be sensitive to both the magnitude of climate warming and the representation of LUCs. The increase in time spent under wildfire risk conditions under SSP3–7.0 is consistent across most of the modeling domain and is generally larger when LUC is included, particularly in areas that already exhibit high present-day risk. This suggests that some LUC scenarios may challenge current and future fire management strategies and increase pressure on firefighting resources. In contrast, SSP1–2.6 shows relatively small changes, with a mix of slight increases and decreases depending on the region, and with LUC often reducing the time at risk compared to the static scenario no LUC. This shows that the effect of LUC is not uniform but interacts with the surrounding climatic conditions, amplifying local risk under strong warming; and that in some cases it can also dampen it under more moderate scenarios.
In summary, the observed differences between simulations considering LUC effects (incl. LUC) and those that do not (no LUC) seem to imply that failing to account for land use dynamics can lead to different estimates of both the magnitude and spatial distribution of future wildfire risk, with potential implications for the robustness of the result and on how such projections are interpreted and used for planning and risk assessments.
4. Discussion and conclusions
In this study, a data-driven wildfire risk model was developed for Southern Europe that integrates hydro-meteorological variables, land cover, topography, and population density. Four different ML models were tested, and the best-performing model was found to be the XGB model, achieving an AUC performance metric of 0.93 and an F1 score of 0.85. SHAP analysis identified soil moisture, slope, and temperature, along with grassland cover and population density, as the most influential features, underpinning the combined roles of environmental and human drivers in defining wildfire risk. These results further suggest that LUC could play an important role for future projections of wildfire risk and therefore should not be neglected.
To quantify the future risk of wildfires in Southern Europe under projected changes in climate, land cover, and population, the XGB model was applied to downscaled end-of-century scenarios (SSP1–2.6 and SSP3–7.0) using an ensemble of CRCM5 regional climate simulations. The results show contrasting outcomes depending on the different climate and land use scenarios. Under the low-emission SSP1–2.6 scenario, incorporating projected LUC consistent with the SSP1 narrative leads to a larger reduction in wildfire risk compared to simulations without it (i.e., the wildfire risk is approximately 3% across the domain when including LUC vs. a reduction of 1% when LUC is not considered). However, the increase is more pronounced in areas already prone to frequent fires when LUCs associated with this scenario are considered. These results indicate that projected wildfire risk is shaped not only by hydro-meteorological conditions but also by socio-environmental dynamics and by land management that can alter the spatial distribution and intensity of fire exposure.
While many previous studies have used ML or multivariate regression to map current wildfire risk at the regional scale, future climate projections often rely on fire weather indices as proxies for wildfire risk. The Fire Weather Index, for instance, does not account for human influences and likewise is not easily converted into a probability, which limits its use in a formal risk assessment framework. In this sense, the current study paves the way for assessing future wildfire risk from a probabilistic perspective similar to how one assesses future flood risk. A natural next step would therefore be to combine our fire risk model with a detailed fire spread model using the same LUC scenarios as discussed above, with atmospheric fields from regional climate simulations to obtain burnt areas at unprecedented spatial and temporal resolution under current and future conditions. It is very likely that it may be even more important to account for joint climate and LUC scenarios when carrying out detailed fire spread simulations under future scenario assumptions than for wildfire risk estimation.
The present study gives rise to several methodological considerations. First, in the analysis, we focused mainly on high-risk areas, using a classification probability threshold of 0.5. This inherently highlights regions in Europe where changes in risk are most pronounced but may under-represent emerging fire regimes in what is currently lower-risk areas. In other European regions or for other purposes, where the base fire risk is much less pronounced, it may be necessary to adjust this threshold or consider alternative metrics. In addition, although future wildfire risk projections here incorporated SSP1 and SSP3 population changes, population effects were not explicitly decoupled from other effects. Lower population densities were found to reduce model output and hence wildfire risk. Here, it is again important to note that the trained ML model estimates wildfire occurrence probability, not fire spread or burned area. Consequently, a reduction in risk associated with lower population does not preclude extensive fires due to limited suppression resources or capacities. Future studies should explore the impact of population changes and assumptions of fire suppression on wildfire risk.
As described above, this study is based on a single-model, multi-scenario ensemble of CRCM5 simulations that represent similar assumptions with respect to climate and land use assumptions. It is evident that an assessment of climate model uncertainty will necessitate additional climate models to strengthen confidence in the projections and facilitate proper statistical tested. Ideally, such models must be able to account dynamically for LUC like it was done within the ClimEx project, and such models are unfortunately very scarce. Such simulations are not planned for upcoming CMIP6 updates of the CORDEX multi-model ensemble, and with potential advances toward pure AI/ML-based downscaling of global climate projections, it may be critically important to obtain more of these kinds of experiments as a means of training downscaling models that are aware not just of climate forcing but also of LUC effects. Finally, in the study, we consider four different simple ML models as proof-of-concept, which prove to exhibit more or less similar performance. To increase confidence in climate projections, we propose that it would be necessary to test also other ML models, including more complex deep learning models, to assert the particular strengths and weaknesses of also other kinds of data-driven models with respect to modeling current and future wildfire risks. In the paper, we have not explored the model limitations in terms of representing out-of-sample extremes, which could be important, especially in high warming scenarios like SSP3–7.0 or SSP5–8.5. To properly account for these limitations, it is possible that other methods will be more appropriate.
Supplementary material
The supplementary material for this article can be found at http://doi.org/10.1017/eds.2026.10045.
Acknowledgments
We gratefully acknowledge Ralf Ludwig and Andrea Böhnisch, LMU Munich, and the ClimEx project for making the climate simulations available; and Beichen Zhang for technical assistance.
Author contribution
Conceptualization: All authors. Methodology: All authors. Formal analysis and Data visualization: O.M.; J.S.M.; M.D. Data curation: O.M.; J.S.M. Writing—original draft: All authors. Writing—review and editing: All authors. Funding acquisition and Project administration: M.D. All authors approved the final submitted manuscript.
Competing interests
The authors declare no competing interests.
Data availability statement
The scikit-learn Python package (Pedregosa et al., Reference Pedregosa, GaëlVaroquaux, Michel, Thirion, Grisel, Blondel, Prettenhofer, RonWeiss, Vanderplas, Passos, Cournapeau, Brucher, Perrot and Duchesnay2011) used for developing and training the ML models is available at https://scikit-learn.org, version 1.6.1. All the analysis codes used in this paper are made available via GitHub. Climate projections (see Section 2.1) were made available upon request to the ClimEx team at LMU Munich (Professor Ralf Ludwig). See Table A1 in the Supplementary Materials for further information about data availability.
Funding statement
This study was supported by the European Union’s Horizon Europe Research and Innovation Program under Grant Agreement numbers 101081358 (ACCREU) and 101137682 (AI4PEX).
Provenance statement
This article is part of the Climate Informatics 2026 proceedings and was accepted in Environmental Data Science on the basis of the Climate Informatics peer review process.
Open access
