The past decade has witnessed a number of unprecedented extreme wildfires across parts of the world (e.g. Australia in 2019–2020, Canada in 2023, or California in 2020 and 2025), causing widespread impacts on societies, ecological environments, and human life. In 2022, southwestern Europe also faced an extreme fire season due to a persistent anticyclonic anomaly causing widespread soil moisture deficit , and record burned area in some regions including parts of France. At the national level, more than 55 000 ha of forests and other natural vegetation were burned – an area 6–7 times larger than the average over the preceding decade. In southwestern (SW) France specifically, the burned area was even more than 14 times larger than the regional average (Fig. b). This extensive burned area resulted in substantial biomass losses in Atlantic pine forests and was largely driven by a small number of large wildfires. In particular, three events alone accounted for more than 45 % of the total annual burned area in France in 2022 and over 80 % of that in SW France. On 12 July 2022, two wildfires started simultaneously within the Gironde department: La Teste-de-Buch wildfire burned approximately 5709 ha over 12 d, while the second one in Landiras burned 12 552 ha over 14 d, due to frequent wind shifts causing spread in multiple directions (Office National des Forêts, personal communication, October 2025). After this wildfire (hereafter Landiras-1) was brought under control, it reignited 15 d later, on 9 August 2022 (hereafter Landiras-2), and spread over six days, driven by northerly winds. When combined, the Landiras-1+2 wildfire burned over 19 676 ha, which makes it the largest wildfire in France since the Landes forest fire of August 1949 .

Those wildfires in SW France provided a glimpse of future projections across the region, featuring a spatial expansion of the potential fire niche with climate change towards western and northern latitudes , a fire niche historically limited to the southeastern Mediterranean region. The 2022 fire season was indeed concomitant with a broader context of global and regional climate warming. Copernicus data indicated that July 2022 was among the three warmest Julys recorded globally, exceeding the 1991–2020 average by about 0.38 °C. In France, Météo-France recorded an average annual temperature of 14.5 °C, approximately 2.9 °C higher than the 1959–2000 baseline. This warmer atmosphere and elevated atmospheric aridity have contributed to a reduced fuel moisture content, thereby increasing landscape flammability.

Attribution analysis is essential to better understand how global warming is currently altering the likelihood of extreme events and associated impacts . Quantifying the likelihood of such impacts may enhance awareness and encourage adaptation efforts. Attribution studies employ both observational and simulated climate datasets to quantify the extent to which human emissions alter the probability of a given extreme weather event. So far, most of those studies have typically focused on meteorological events such as heatwaves , droughts , or extreme rainfall . However, fire weather conditions (combining multiple meteorological variables) have received less attention, although a number of efforts have been made in the western US , Canada , Australia , and France . Recently, conducted an attribution study of the 2022 fire season in SW France using multiple standardized climate indices. Their findings suggest that climate change doubled the likelihood of the climate conditions observed during the month of July. However, the analysis was performed over relatively broad spatial (the entire SW region of France) and temporal (the whole month of July) scales, while attribution scores have been shown to be sensitive to the selection of spatial and temporal scales . Finally, the attribution scores were limited to the year 2022, with no projections on how those conditions might change in the future. Such projections may help plan adaptation strategies.

Building on prior attribution studies, we use here a complementary multi-scale framework to (i) provide a broader context of the spatiotemporal variability of fire weather conditions across France over the whole observational period from 1959 to 2023, (ii) quantify fire weather anomalies conducive to the 2022 wildfires, (iii) estimate return periods (RPs) of fire weather conditions associated with the three largest wildfires, considering a combination of different and complementary spatial and temporal scales, and (iv) estimate the extent to which those fire weather conditions have become, in 2022, more or less likely under anthropogenic-plus-natural forcings than under the natural-forcing-only climate, and how this contrast is likely to further increase in the future.

Figure  illustrates the first two dominant modes of May–September FWI variability over the 1959–2023 period, with the EOF loadings (left panels) and their corresponding PCs (right panels). The loadings characterize the spatial structure of a given mode, identifying regions where FWI anomalies vary either in phase (same sign) or in opposition. The PCs reflect the temporal evolution of each mode, highlighting years during which the associated spatial structure is either amplified or dampened. In other words, the initial FWI time series in a specific grid cell featuring a high positive loading will strongly look like the associated PC. A negative loading will in turn indicate that the FWI time series varies in opposition to the PC. Together, the first two modes explain about 75 % of the total variance (61.69 % and 12.75 % for EOF-1 and EOF-2, respectively). EOF-1 exhibits positive loadings throughout France, albeit with smaller coefficients along the Mediterranean (Fig. a). This mode presents a strong interannual variability with an upward underlying trend, featuring an increasing frequency of higher FWI years in recent decades with global warming (Fig. b). We found that PC-1 was strongly correlated with both temperature and rainfall anomalies over a large portion of western Europe (see Fig. S4). Note that the highest amplitude is seen in 2022, followed by 1976, a notoriously warm and dry year in France. By contrast, after removing the influence of PC-1, EOF-2 shows a slightly unbalanced north–south dipole, with a near-zero band straddling central France. In other words, when positive FWI anomalies occur preferentially in the south, negative anomalies are seen in the north and vice versa. This mode of variability correlates with a larger continental-scale dipole in rainfall anomalies, as well as with temperature anomalies south of 45°N (see Fig. S5). Like the first mode, the second mode presents a long-term trend (Fig. d), reflecting an increasing occurrence of years with higher FWI in southern France and lower FWI in northern France. Note that the following modes were not analyzed due to their little variance explained and their lack of consistency across space.

We then restricted our attention to local FWI conditions associated with actual wildfires across SW France. Figure a shows that, on average, FWI increases until the wildfire day and decreases in the following days, with higher FWI values for larger wildfires. Figure c shows positive anomalies three months before wildfires (note that 100% indicates that FWI was twice as high as expected from average local conditions), reaching 71 %, 106 %, and 155 % for small, medium, and large fires, respectively. These persistent pre-wildfire positive anomalies may reflect not only prolonged antecedent hot and dry conditions, but also, to some extent, an earlier seasonal onset of the fire weather season. A similar signal was observed in 2022 (Fig. b and d), but FWI anomalies were that time higher during the previous months and were 119 %, 137 %, and 180 % higher than mean conditions on the starting days of small, medium and large wildfires, respectively (Fig. d).

We then examined the return periods (RPs) of the three largest wildfires of 2022. Figure  (left panels) indicates that the annual maxima of the moving-average (MA) FWI are consistently highest when computed at the finest spatiotemporal resolution (i.e. the fire-duration window at the 64 km2 SAFRAN grid cell fire level). These annual maxima decrease when the temporal window is lengthened to 30 d or when FWI conditions are spatially averaged over SW France region (∼4.9×104 km2). This pattern holds for all three wildfires (Fig. , left panels). In every case, the absolute maximum occurs in 2022, underscoring the exceptional FWI conditions during that year. Overall, the rarity of those conditions also increases with the resolution (Fig. , right panels), with the best-estimate RPs increasing from ≈6 to ≈34 years, from ≈22 to ≈38 years, and from ≈6 to ≈101 years when moving from the coarsest to the finest spatiotemporal scale for Landiras-1 (Fig. b), Landiras-2 (Fig. d), and La Teste-de-Buch (Fig. f) wildfires, respectively, illustrating how sensitive the RPs are to the chosen scales.

Finally, we examined how ACC altered the probability of those FWI conditions. Figure  illustrates, for one model (NorESM2-LM, r1i1p1f1) and one spatiotemporal set-up (local and fire duration), how pNAT varies relative to the reference probability pALL (=pOBS by construction). In this spatiotemporal set-up, pNAT slightly exceeds pALL until the 1970s although pALL remains within the bootstrapping confidence interval. After the 1970s, pNAT becomes systematically lower than pALL and tends toward zero after the 2040s, reflecting the growing inability of the NAT climate to produce FWI conditions that remain at the same extreme quantile in the warming ALL climate. We then computed the risk ratio (RR) for the Landiras-1 wildfire (Fig. ) using all models. The RR exhibits a consistent increase from the late 20th century for each model and across all temporal and spatial scales. The multi-model ensembles exceed the reference line RR=1 (indicating that pNAT=pALL) around 1970–1980, with earlier emergences for finer resolutions (Fig. d). In 2022, three of the four models yield best-estimated RR>1 across spatiotemporal scales: IPSL-CM6A-LR ≈1.5–3, CanESM5 ≈6–20, and NorESM2-LM ≈2–130; by contrast, MIROC6 presents an RR<1 (≈0.4–0.6). The multi-model median RR lies between ≈2 and 10 in 2022 across the four scales. In the latter half of the century, this ratio increases by 1 to 2 orders of magnitude. Note the large confidence intervals within and across models and scales, reflecting the strong sensitivity to internal variability, parameter uncertainty and model uncertainty. Also, RR scores may sometimes exceed 10 000 as from the mid-to-late 21st century due to very low pNAT approaching zero as shown in Fig. .

Using the finest spatiotemporal set-up (local over fire duration), we found that ACC contributed approximately 78 %, 73 %, and 79 % to the FWI conditions associated with Landiras-1, Landiras-2 and La Teste-de-Buch wildfires respectively, and will approach 100 % by mid-21st century (Fig. ) with very large uncertainty until 2070s. Note that the signals for Landiras-1 and La Teste-de-Buch are very similar, as the two events occurred approximately over the same period.

The spatiotemporal variability of the observed warm-season (May–September) FWI during 1959–2023 has been synthesized into two leading modes. The first mode (∼61.69 % of variance explained) shows strong interannual variability due to an alternation between warmer/drier and cooler/wetter years throughout France. This mode also features an increased frequency of years with positive FWI anomalies over time. This evolution is consistent with the long-term trend in temperature and drought in France, collectively contributing to increased fire weather conditions as observed more broadly in the Mediterranean region , Europe , and globally . The second mode (accounting for ∼12.75 % of variance) reveals a north–south dipole. This north–south contrast in climate anomalies probably relates to the summer North Atlantic Oscillation, which is, during negative phases, generally associated with blocking events producing cooler/wetter conditions in northern Europe and warmer/drier conditions in the south . An underlying long-term trend with more frequent warmer/drier years in the south in recent years was also evident, in agreement with precipitation decreases in southern Europe and slightly wetter conditions in the north as documented in observations as well as in future simulations . Interestingly, both PC-1 and PC-2 scores indicate that the year 2022 represented a combination of these two leading modes, with unprecedented FWI anomalies over a large portion of France and, orthogonal to that mode, a latitudinal dipole with higher (lower) FWI in the southern (northern) half of the country.

The exceptionally high FWI values observed in 2022 in SW France, whether sampled locally or regionally, were conducive to a series of wildfires, with larger wildfires associated with larger FWI anomalies. We found that FWI levels reached their highest amplitude on the day of ignition (or the week surrounding it), boosted by either synoptic-scale heat waves or local wind bursts, as shown in previous studies over southern Europe and France . We estimated that the conditions observed locally during the three largest wildfires were expected, on average, to occur once every 34, 38, and 101 years for the Landiras-1, Landiras-2, and La Teste-de-Buch wildfires, respectively. The last estimate, relying on extrapolation beyond the observational record, naturally involves substantial uncertainty.

Across spatial scales, Landiras-1 and La Teste-de-Buch wildfires exhibited higher RPs at the local scale than at the regional scale, reflecting the well-known effect of spatial averaging on dampening extremes . We note that the region is flat, suggesting that those spatial differences cannot be attributed to terrain-related factors. However, Landiras-2 displayed the opposite pattern, with regional RPs exceeding local RPs. This inversion may be due to the fact that Landiras-2 was a rekindling of the Landiras-1 wildfire and was driven mainly by smoldering peat rather than by fire weather conditions. Also, we noticed that some precipitation occurred locally on 13–14 August, depressing the local FWI during the 9–14 August interval. Furthermore, even though the regional-scale FWI had a longer return period, the FWI associated with Landiras-2 was still higher in absolute terms at the local scale. Across temporal scales, all three wildfires exhibited higher RPs during the fire-duration window than during the 30 d window, indicating that short-duration, but more acute FWI driven by daily or synoptic meteorological variations also contributed strongly to these wildfires. Note that the wide confidence intervals around the RPs illustrate the uncertainty inherent in GEV fits with limited sample sizes.

Our results indicated that anthropogenic warming has, since the early twenty-first century, markedly increased the probability of occurrence of FWI conditions associated with those wildfires. This is consistent with a growing body of fire-attribution research at the global scale and at regional scales across parts of the US , Canada , Australia , the Arctic , Portugal , and France . This increase is evident across temporal and spatial scales in the multi-model median, although its magnitude and timing differ substantially among models. Likewise, the timing at which the RR exceeds 1 and remains above that threshold is broadly consistent with the emergence of anthropogenic signals in simulated fire weather indices since the late twentieth century in southern Europe . Our study suggests that the multi-model median probability of such conditions increased by approximately 2–10 times in 2022 and is projected to increase by roughly 1 to 2 orders of magnitude by the end of the twenty-first century under a medium-level radiative forcing scenario. However, the pooled ensemble uncertainty range indicates substantial uncertainty across and within models (see Fig. S6). This validates, across different metrics, fire weather indices, and spatiotemporal scales, the results obtained by using a soil-moisture index. Interestingly, despite the large spread in RPs across scales, the attribution scores were not found to change substantially across regional and local scales, suggesting that the rarity of FWI conditions does not necessarily relate to the magnitude of the anthropogenic forcing . Although inter-model differences were evident in terms of amplitude and timing (i.e. the date at which the RR emerges above 1), due to model sensitivity to greenhouse gas emissions, all models point to a substantial increase in 2022-like conditions in future decades, consistent with previous projections of FWI or FWI-derived fire activity in France. Also, MIROC6 was found to cross the RR=1 line later than other models, supporting the findings of who detected a lower climate change signal in MIROC6. This probably relates to the lower climate sensitivity of MIROC6 due to the radiative forcing of aerosols and cloud feedback . Finally, our study shows that approximately 70 % of the risk that fire weather conditions reach the levels observed during those wildfires can be attributed to ACC (more than the 49 % found in over broader spatial and temporal scales, and the nearly half contribution reported by at a multi-decadal scale for Mediterranean France) and that this estimate will reach 100 % by the 2050s. These conclusions apply to other large wildfires (Landiras-2 and La Teste-de-Buch) as well.

We note that the methodology developed here has some limitations. First, our analysis was based solely on meteorological forcing and therefore lacks information on fuels (e.g. forest cover, fuel breaks, and horizontal/vertical continuity). The use of a statistical model that accounts for some of these features (e.g. Firelihood; ) would produce more realistic estimates. Previous studies have also shown that some wildfires are driven primarily by wind, whereas others are driven by the dryness of climate conditions . Further efforts are thus needed to resolve the respective contributions of fuel moisture and wind forcing, based, for instance, on sub-indices of the FWI (e.g. Fine Fuel Moisture Code (FFMC), Duff Moisture Code (DMC), and Drought Code (DC)) and complementary atmospheric drivers such as the vapor pressure deficit (VPD) and wind speed, as recently implemented in a probabilistic framework . Second, our analysis does not explicitly consider ignition sources and their spatiotemporal variability. In France, approximately 95 % of ignitions are related to human activities , and the realized fire activity therefore reflects the interplay between human pressure and fire weather conditions. As a simple approximation, the “fire start probability” could be expressed as a monotonically increasing function of the FWI (i.e. an ignition probability conditional on fire weather) using a probabilistic framework such as the Firelihood model combining FWI with land use and land cover information. Also, our event-attribution analysis was based on three large wildfires in southwestern France. Extending the analysis to the French Mediterranean, which recently experienced the second largest wildfire in France since 1949 (early August 2025; approximately 17 000 ha burned at Ribaute), would be of interest. Although the spread of this wildfire was facilitated by hot and dry conditions alongside strong winds, wildfires in that region are generally associated with higher FWI levels that are expected, on average, once every year or so. Regarding climate simulations, we note that only a single ensemble member (r1i1p1f1) was used for each model. Using additional members would obviously reduce internal variability. Further work is also needed to improve the multi-model averaging by introducing weights based on each model's skill over the historical period . Finally, we used the SSP2-4.5 scenario in CMIP6, corresponding to a medium level of radiative forcing. Using higher-forcing scenarios (e.g. SSP5-8.5) would yield much higher RR in the future.

The three wildfires examined here burned over multiple days, each undergoing several complete nocturnal cycles. Overnight burning is relatively new in France, where the vast majority of wildfires were historically extinguished within a single day. Further research is thus needed to elucidate (i) the role of nighttime conditions in overnight burning and the extent to which the so-called nighttime barrier is likely to weaken under ACC , and (ii) the relative contributions of nighttime aridity vs. seasonal drought , as well as the potential role of other meteorological variables such as wind speed . Such analyses may complement the information provided by traditional daytime-based indicators (e.g. FWI) used in climate–fire or attribution studies.

This study aimed at quantifying how ACC has already, and will further, alter the probability of fire weather conditions associated with the three largest wildfires of 2022 across France (Landiras-1, Landiras-2, La Teste-de-Buch). First, we found that warm-season (May–September) FWI gradually intensified over 1959–2023, especially in southern France, making the landscape more and more flammable. Second, we found that return periods of FWI associated with extreme wildfires observed in 2022 were scale-dependent and may range for instance from about 6 (at monthly and regional scales) to more than 101 years (at fire-duration and local scales) for La Teste-de-Buch wildfire. Finally, attribution metrics indicated that FWI conditions of comparable rarity to those observed during the wildfires were substantially less likely under the natural-forcing-only climate, and that anthropogenic climate change made those conditions approximately 2–10 times more likely in 2022. Under a moderate-emissions pathway, those FWI conditions are projected to become roughly 1 to 2 orders of magnitude more probable by the end of the century, although substantial uncertainty remains across and within models.