Contrasting large fire activity in the French Mediterranean

In the French Mediterranean, large fires have significant socioeconomic and environmental impacts. We used a long-term georeferenced fire time series (1958–2017) to analyze both spatial and temporal distributions of large fires (LFs; ≥ 100 ha). The region was impacted in some locations up to six times by recurrent LFs and 21 % of the total area burned by LFs occurred on a surface that previously burned in the past, with potential impact on forest resilience. We found contrasting patterns between the east and the west of the study area, the former experiencing fewer LFs but of a larger extent compared to the latter, with an average time of occurrence between LFs exceeding 4000 ha<7 years mostly in the eastern coastal area and >50 years in the west. This longitudinal gradient in LF return level contrasts with what we would expect from mean fire weather conditions strongly decreasing eastwards during the fire season but is consistent with larger fuel cover in the east, highlighting the strong role of fuel continuity in fire spread. Additionally, our analysis confirms the sharp decrease in both LF frequency and burned area in the early 1990s, due to the efficiency of fire suppression and prevention reinforced at that time, thereby weakening the functional climate–fire relationship across the region.


32
It is now unanimously agreed that large fires have most significant socio-economic and environmental 33 impacts, threatening or damaging infrastructures, ecosystems, and even costing human life, especially 34 in the expanding wildland-urban interfaces (

Study Area 105
The study area (total surface area of 11 157 km 2 ) is one of the most fire-prone region of SE France in 106 terms of fire frequency (i.e. number of fires) and burned area (Ganteaume and Jappiot, 2013;107 Ganteaume and Guerra, 2018). The western part is characterized by an extensive WUI where the 108 ignitions are the most frequent (47% of the total ignitions occurred in the WUI) (Ganteaume and 109 Long-Fournel, 2015). Most large fires occur in summer but their cause is often unknown and when it 110 is known, these large fires are mainly due to arson (Ganteaume and Guerra, 2018). 111 The two parts of the study area ( Fig. 1

Climate and Land Cover Data 144
We computed the daily Fire Weather Index (FWI) from the Canadian Forest Fire Weather Index 145 system using daily surface meteorological variables at a 8-km spatial resolution from the quality-146 controlled SAFRAN dataset providing maximum temperature, minimum relative humidity, 147 precipitation and wind speed over France from 1959-2017 (Vidal et al., 2009(Vidal et al., , 2010(Vidal et al., , 2012. The FWI 148 computation usually requires noon observations. However, given that SAFRAN is a daily 149 meteorological database, we calculated FWI using maximum temperature and minimum relative We estimated annual maximum burned area (AMBA) return levels in the eastern and western 174 part of the study area using the so-called block (here 1-year) maxima approach. We extracted the 175 AMBA in both areas and selected the type of distribution that best fitted both series using the Akaike 176 Information Criteria (AIC). In both areas, the gamma distribution was found to best describe the 177 AMBA series. Using this distribution, the inverse cumulative distribution was calculated allowing the 178 determination of the theoretical quantiles from which we derived the return levels (AMBA) associated 179 to different return periods ranging from 5 to 100 years. Asymmetric confidence intervals were 180 calculated using a resampling approach. This approach consists in creating new sub-samples from the 181 original sample (75% of the original sample are extracted at random) using a bootstrapping process 182 with replacement and then estimating a return level for each of the resampled data (N=1000). The 183 resulting empirical distribution can then be used to derive the 95% confidence intervals from the 184 resulting collection of estimates. Tab. 1). LF were responsible for most of the total burned area in the East (97%) as well as in the West 193 (87%), which supports the relevance of the fire-size threshold selected (100 ha). 194 Regarding the LF age distribution (Fig. 2), the most frequent LF belonged to the 31-40 year-195 class resulting in the most LF-prone decade. In the East, recent LF were mainly located on the coast 196 while the age distribution was more homogeneous in the western part. Notice that most LF growths 197 were in the main wind direction blowing from Northwest. A total surface area of 312,447 ha was 198 burned during the period studied of which 21% occurred on a surface that already burned in the past 199 (Fig. 3), due to multiple overlaps in burned areas by recurrent fires (i.e. LF occurrence on the same 200 surface). LF reburns occurred up to 6 times in the East but represented only a small part of the 201 recurrence (0.3%; Tab. 2). One to two reburns were the most frequent patterns in the western part of

Longitudinal contrast in large fire extent 212
The mean LF extent varied along a longitudinal gradient, increasing from the West to the East 213 ( Fig. 4 top). This signal contrasts with the mean summer FWI gradient decreasing towards the East but 214 is consistent with the sharp increase in biomass towards the East (Fig. 4 bottom). This suggests that LF 215 spread is not limited by climate conditions across the region but strongly fuel-limited in the West, due 216 to landscape fragmentation and the high proportion of WUI. Indeed, the landscape has undergone 217 substantial transformation with time in the western part contributing to reduce fuel cover and thereby The 95% confidence intervals were estimated using a bootstrapping approach. Bottom) Same as top 224 panel but for mean June-September FWI (in red) and the percent of biomass (in green). 225 226

Long-term trends in large fires 227
A significant decline in annual LF frequency alongside area burned by LF was found across the region 228 according to a Man-Kendall test (Fig. 5). This overall decline is consistent with a significant change 229 point in both LF metrics in 1991 as shown in previous findings (Fox et al., 2015;Ruffault and 230 Mouillot, 2015). This signal was especially evident in the eastern part (Fig. 5c) while neither a change 231 point nor a significant trend (p>0.05) were detected in the western part for both LF metrics (Fig. 5b). 232 We then examined how interannual correlations between mean June-September FWI and LF activity 233 have changed over time across both regions (Fig. 5d). Higher correlations prevailed in the western part part of the eastern area where the tourist pressure is lower. In contrast, LF were homogeneously 277 distributed in the West, regardless of their age and most reburns corresponded to WUI areas. 278 We found that the return level was higher in the eastern part of the study area although LF 279 were more frequent in the West. These contrasted regional return levels may provide critical and 280 useful information for risk assessment and local decision-making. Indeed, LF >4000 ha may occur 281 within seven years in the East against 55 years in the West. In other words, LF are less probable in the 282 east where fire ignitions are more limited but when an ignition does occur, the fire is likely to spread 283 over larger areas. This longitudinal gradient is likely due to the variation in landscape fragmentation. 284 Indeed, the western area presents a mosaic of wildlands interspersed with agricultural areas and WUI, 285 LF being thereby concentrated in natural spaces less extended than in the eastern part where large 286 forested massifs mostly located on the coast allowed fire spread. By contrast, LF were more frequent  On the whole, 21% of the total area burned by LF occurred on a surface that already burned in the 328 past, the region being impacted in some locations up to 6 times by recurrent LF (coastal areas of the 329 eastern part of the study area). LF were less frequent in the eastern part but larger than LF occurring in 330 the West mostly in WUI. This longitudinal gradient in LF extent, featuring a shorter time of 331 occurrence between LF in the East with respect to the West, contrasts with what we would expect 332 from mean fire weather conditions strongly decreasing eastwards but is consistent with larger fuel 333 cover in the East. Indeed, fuel continuity in the East allows fire to grow large and to reach on average 334 4,000 ha every 7 years, a spatial extent in burned area observed only every 50 years in the West. 335 An abrupt decline in LF was evident across the eastern part in the early 1990s, mostly due to a 336 change in fire management policy thereby contributing to the weakening of the climate-fire 337 relationship. However, despites large means allocated to fire suppression, large fire outbreak is still