Contrasting large fire regimes in the French Mediterranean 1 2

10 In the French Mediterranean, large fires have significant socio-economic and environmental impacts. 11 We used a long-term geo-referenced fire time series (1958-2017) to analyze spatio-temporal variations 12 of large fires (LF; ≥100 ha) throughout a fire-prone area of this region. This area was impacted in 13 some locations up to 5 or 6 times by recurrent LF and 21% of the total area burned by LF occurred on 14 a surface that previously burned in the past. We found distinct patterns between the East and the West 15 of the study area, the former experiencing fewer LF but of a larger extent compared to the latter, with 16 an average time of occurrence between LF exceeding 4000 ha <7 years and >50 years, respectively. 17 This longitudinal gradient in LF extent contrasts with what was expected from mean fire weather 18 conditions strongly decreasing eastwards but is consistent with larger fuel cover in the East. The 19 temporal variation of LF, featuring a sharp decrease in both frequency and burned area in the early 2


Introduction
It is now unanimously agreed that large fires have most significant socio-economic and environmental impacts, threatening or damaging infrastructures, ecosystems, and even costing human life, especially in the expanding wildland-urban interfaces (WUI) (Blanchi et al., 2014;Syphard et al., 2012;Syphard and Keeley 2015;Radeloff et al., 2018).However, the definitions of what can be considered as a large fire are numerous (Shvidenko and Nilsson, 2000;Stocks et al., 2002;Barbero et al., 2014a, Stavros et al., 2014;Nagy et al., 2018;Tedim et al., 2018), the size of such fires being arbitrary or statistically assessed (Moritz, 1997).Yet, as taking a fire-size threshold can minimize the importance of smaller fires in highly fragmented landscapes; this notion has also been approached focusing on the upper part of the burned area distribution (i.e. the largest 10% of fires occurring in a region according to Nagy et al., 2018).Albeit the choice of the cutoff remains highly subjective and variable from one study to another.Usually, large fires represent only a small proportion of the total number of fires but they typically account for the bulk of burned area in many regions throughout the world (Stocks et al., 2002;San Miguel-Ayanz et al., 2013;Stavros et al., 2014, Barbero et al., 2014a, 2014b, 2015;Ganteaume and Guerra, 2018) and contribute in fact to the trend and interannual variability in the total burned area.and vegetation combined (Koutsias et al., 2012) alongside human activities (Syphard and Keeley, 2015;Syphard et al., 2017;Nagy et al., 2018).Some of the aforementioned factors were found to be non-stationarity in time.For instance, changes in fire suppression policy over the last few decades have induced sharp decreases in fires in some Mediterranean regions (Pezzatti et al., 2013, Moreno et al., 2014;Fréjaville and Curt, 2017;Ganteaume and Guerra, 2018), partially modifying the functional relationships linking fire to climate (Higuera et al., 2015;Fréjaville and Curt, 2017;Syphard et al., 2017), and thus, decreasing or increasing fire activity independently of the climate forcing (Hawbaker et al., 2013;Syphard et al., 2007).
As the European Mediterranean region, the Southeast of France is a highly populated area and is characterized by an extensive WUI.Fire prone areas along the Mediterranean coast have been extensively built up, reducing in some cases the availability of fuels but greatly increasing the probability of human-started fires (Ganteaume et al., 2013).The region includes plant communities well adapted to Mediterranean climate conditions that confer on this area a high fire risk.In this region, the largest fire on record reached 11 580 ha although most fires are generally smaller compared to other Mediterranean countries that have recently experienced larger fires such as Spain or Portugal.
However, because of the high proportion of WUI, these large fires are of major concern, especially in the most populated parts, where most fires are also concentrated.Moreover, an increase in fire recurrence and a shortening of the period between fires were shown to impact vegetation structure, especially with the decrease in mature tree cover (Ganteaume et al. 2009), including the loss of resilience of Pinus halepensis stands (Eugenio et al. 2006).
Here, we used longer time-series of georeferenced fires extending back to 1958 to identify both longterm trends and possible spatial patterns in large fire distribution, including fire recurrence, the time since the last fire and the mean time interval between fires.Finally, we sought to relate these spatiotemporal distributions of large fires to climate conditions and vegetation availability.

Study Area
The study area (total surface area of 11 157 km 2 ) comprised two of the 15 French administrative districts that composed southeastern (SE) France and which are among the most impacted by fires in terms of fire frequency (i.e.number of fires) and burned area (Ganteaume and Jappiot, 2013;Ganteaume and Guerra, 2018).The western part is characterized by an extensive WUI where the ignitions are the most frequent (47% of the total ignitions occurred in the WUI; Ganteaume and Long-Fournel, 2015).Most large fires occur in summer but their cause is often unknown and when it is known, these large fires are mainly due to arson (Ganteaume and Guerra, 2018).
The two parts of the study area (Fig. 1), located on a West-East gradient of the Mediterranean, share most climate characteristics albeit the amount of annual precipitation increases eastwards (Ruffault et al., 2017).These areas also differ in the structure of landscapes; forested massifs are larger in the eastern zone while the proportion of WUI and the urbanization are higher in the western area (respectively, 15% vs 7%, Ganteaume comm.pers., and 394 vs 174 inhabitants km -2 , https://www.geoportail.gouv.fr),as well as in the main flammable fuel types, due to the nature of the bedrock (acidic soils being mainly located in the East contrary to limestone-derived soils in the West).All these differences are hypothesized to affect the spatio-temporal pattern of large fires.
Figure 1: Map of the study area.Fuel cover in green was extracted from the "BD Forêt 2014" of the National Geograhic Institute (https://www.geoportail.gouv.fr).

Fire Data
Large fires in SE France have already been studied in previous works using shorter time series based on the gridded regional fire database Prométhée that recorded fires since 1973 (Fréjaville and Curt, 2015;Ruffault and Mouillot, 2017;Ruffault et al., 2018).However, this gridded data provides neither the fire perimeter nor the temporal length needed to assess return periods in large fires.Here, we used the georeferenced fire perimeter database of the Directions Départementales des Territoires et de la 1958 to 2016 in the eastern part.We focused on large fires ≥ 100 ha (hereafter LF), representing only 28% of the total number of fires ≥1 ha (N=1277) but accounting for 94% of the total burned area.
Compared to large fires considered in other works (i.e.200 ha in Canada according to Stocks et al., 2002; 405 ha in the USA according to Dennison et al., 2014, 500 ha in Portugal according to Moreira et al., 2011, and1000 ha in Australia according to Bradstock et al., 2009), this detection threshold is lower but within the range of thresholds used in other works in SE France ranging from 30 ha (Ruffault and Mouillot, 2017) to 250 ha (Ruffault et al., 2017).

Climate and Land Cover Data
We computed the daily Fire Weather Index (FWI) from the Canadian Forest Fire Weather Index system using daily surface meteorological variables at a 8-km spatial resolution from the qualitycontrolled SAFRAN dataset providing minimum and maximum temperature, relative humidity, precipitation and wind speed over France from 1959-2017 (Vidal et al., 2009(Vidal et al., , 2010(Vidal et al., , 2012)).Although the FWI was empirically calibrated for estimating whether atmospheric conditions and fuel moisture content are prone to wildfire development in Canada (VanWagner, 1987), the FWI has already proven useful in Mediterranean regions (Dimitrakopoulos et al., 2011;Fox et al., 2018;Lahaye et al., 2017).
Grid cells of the FWI lying within the study area were first averaged across the June-September season and then averaged across all latitudes spanning the region of interest to form a longitudinal crosssection of mean summer FWI conditions.
We extracted fuel cover data from the "BD Forêt 2014" of the National Geograhic Institute (https://www.geoportail.gouv.fr)and regridded the data onto a 8-km spatial grid.The percentage of land area covered by fuel was computed across all latitudes spanning the region of interest to form a longitudinal cross-section as described above.

Spatial Analyses
We defined the LF regime in terms of the time interval since the last fire (LF ranging from "recent": less than one decade to "ancient": more than four decades).LF recurrence was calculated by counting the overlaps of LF polygons to quantify the number of times each location has been burned across the period 1958-2017.For each LF, its georeferenced location and perimeter as well as the year of occurrence were used to derive a fire return level in the western and eastern part of the study area, a recurrence on a given location and the age of the last burned area.
Comparisons of means in burned areas due to LF were performed using a non-parametric Mann-Whitney test and a Chi2 test was used to test the difference in number of LF between the two parts of the study area.

Temporal Analyses
Monotonic trends in LF frequency and in burned area due to LF were assessed using the nonparametric Mann-Kendall test (Kendall, 1975) and a change point detection test (Standard Normal Homogeneity Test (SNHT); Alexandersson and Moberg, 1997) was used to identify potential abrupt changes in the time series.
We estimated LF return levels in the eastern and western part of the study area using the socalled block (here 1-year) maxima approach.We extracted the annual maximum LF size in both areas and selected the type of distribution that best fitted both series using the Akaike Information Criteria (AIC).In both areas, the gamma distribution was found to best describe the annual maximum LF size series.Using this distribution, the inverse cumulative distribution was calculated allowing the determination of the theoretical quantiles from which we derived the return levels (LF extent) associated to different LF return periods ranging from 5 to 100 years.Asymmetric confidence intervals were calculated using a resampling approach.This approach consists in creating new subsamples from the original sample (75% of the original sample are extracted at random) using a bootstrapping process with replacement and then estimating a return level for each of the resampled data (N=1000).The resulting empirical distribution can then be used to derive the 95% confidence intervals from the resulting collection of estimates.western part, the distribution of LF according to their age was more homogeneous (Fig. 2).Notice that most LF growths were in the main wind direction blowing from Northwest.

Spatial Patterns of LF
A total surface area of 312 447 ha was burned during the period studied of which 21% occurred on a surface that already burned in the past (Fig. 3   This suggests that LF spread is not limited by climate conditions across the region but strongly fuellimited in the West, due to landscape fragmentation and the high proportion of WUI. Figure 4: Top) Longitudinal cross-section of LF extent computed over 30-km sliding windows.The 95% confidence intervals were estimated using a bootstrapping approach.Bottom) Same as top panel but for mean June-September FWI (in red) and the percent of fuel cover (in green).

Temporal patterns of LF
Figure 5 shows the annual LF frequency alongside area burned by LF in the entire region and in the two parts of the study area separately.For both parts, 1979 and 1989 were the years presenting the highest frequency of LF (respectively 11 for both years in the eastern area and 20 and 12 in the western area).These years were also the most impacted in terms of area burned by LF in the western area (respectively, 14 324 and 14 033 ha burned) as opposed to the eastern area more impacted in 1990 (24 920 ha burned).A significant change point in LF frequency as well as in burned area by LF was detected in 1991 in agreement with previous findings (Fox et al., 2015) while it occurred around 1986 (Ruffault and Mouillot, 2015) in a slightly different area.This signal was especially evident in the eastern part (Fig. 5c) while neither a change point nor a significant trend (p>0.05) were detected in the western area for both LF metrics (Fig. 5b). Figure 6 shows the annual maximum burned area in each part of the study area and the Gamma distribution models that were found as the best fit to the data.Estimates of LF return intervals show that a LF >4000 ha occurs on average every 7 years in the eastern zone and every 55 years in the western area.

Discussion
Improving our understanding of large fire regimes is of upmost importance to fire prevention and management to mitigate their impacts.Here, we presented a comprehensive analysis of spatial and temporal patterns of LF in the French Mediterranean.To our knowledge, the fire database compiled and analyzed in this framework provides for the first time a detailed description of the LF regime recorded on geo-referenced long time series.Although previous works (Nagy et al., 2018) argued that using a single absolute size threshold to define a large fire was not a consistent indicator of ecological and economic risks across a large area (smaller fires may have stronger impacts than larger ones depending on the location), we opted for a fixed threshold of 100 ha as fires reaching or exceeding this size contributed to 94% of the total burned area and are likely to threaten ecosystems and/or the society.We, however, acknowledged that other metrics, such as fire intensity or fire damage (when available) may also provide additional insights on fire impacts (Tedim et al., 2018).

Spatial patterns of LF
We found that LF were larger but less frequent in the East compared to the West of the study area.
Indeed, LF >4000 ha may occur within seven years in the East against 55 years in the West.In other words, LF are less probable in the east where fire ignitions are more limited but when an ignition does occur, the fire is likely to spread over larger areas.This longitudinal gradient is likely due to the variation in landscape fragmentation.Indeed, the western area presents a mosaic of wildlands interspersed with agricultural areas and WUI, LF being thereby concentrated in natural spaces less extended than in the eastern part where large forested massifs mostly located on the coast allowed fire spread.In contrast, LF were more frequent in the West where population density, the proportion of WUI, and of infrastructures (railroads and roads) are the highest, this result agreeing with previous works (Keane et al., 2008;La Puma, 2012;Alexandre et al., 2016;Nagy et al., 2018).Ruffault and Mouillot (2017) showed that fuel fragmentation (i.e.due to a high proportion of WUI or road density) was one of the most important factors limiting the occurrence of large fires in the French Mediterranean agreeing with our results showing that the mean LF extent was more limited in the West.However, our results also suggest that LF were slightly more frequent in the West despite the fuel fragmentation.Fox et al. (2015) showed that, in an area located to the East of our study area, neither WUI characteristics (despite the 60% increase between 1964 and 2009 in this area) nor summer weather were major drivers of fire frequency and burned area, the climate control becoming less important as the fire regime shifted to more frequent human-started fires (Zumbrunnen et al., 2009).
In the western part, the most recent LF were mainly clustered along the coast while the more ancient fires were located in the central and northern part.In contrast, LF were homogeneously distributed in the East, regardless their age.LF recurrence (number of LF occurring on a same surface during the period studied) was up to 5 times in the west and up to 6 times in the East.In the East, most LF occurred only once on the same location and the largest areas were burned by ancient LF, while, in the West, non-recurrent LF and especially two recurrent LF were the most frequent between 1975 and 1984.
Some recent studies across the Euro-Mediterranean countries emphasized that large fire preferentially occured under specific synoptic patterns associated with high temperature (Pereira et al., 2005;Trigo et al., 2013;Hernandez et al., 2015).In southern France, large fires were also facilitated by wind events blowing from Northwest (Ruffault andMouillot, 2015, 2017).The shapes of LF which were more elongated in the wind direction in the western part support the results of Ruffault et al.
(2018) pinpointing that the main wind-driven large fires that had occurred in 2016 were located in the western part while the main heat-driven large fires that occurred in 2003 were located in the East of the area.

Temporal patterns of LF
The decreasing trend in both LF frequency and burned area observed over the last 6 decades is in agreement with previous works (Ruffault et al., 2016;Turco et al., 2016) that highlighted a decrease in fire activity across parts of southern Europe in response to an increased effort in fire suppression, especially since the end of the 1990s in the French Mediterranean (Mouillot and Field, 2005).Indeed, the region was highly impacted by fires during the 1970-1990 period and developed a thorough fire suppression and prevention system in the beginning of the 1990s, allocating more means for fire management that allowed faster reactivity in case of fire start (the strategy became extinguishing the fires at their initial stage by massive attack to prevent their spread).The decrease in both LF frequency and burned area since 1991, especially evident in the eastern part of the region, is likely due to this change in firefighting policy and fire prevention regulations (the fire suppression could be more intense in the East because the fires were larger).This result was also highlighted in previous works across the region (Curt and Fréjaville, 2017;Fox et al., 2015;Ruffault and Mouillot, 2015) as well as in other countries (such as Switzerland; Pezzatti et al., 2013).
Climate projections suggest that atmospheric conditions conducive to large fire will increase in the future, the warming and drying trends facilitating their probability of occurrence and their severity (Stavros et al., 2014;Wang et al., 2015;Barbero et al., 2015), at least where fuel and ignitions are not limiting.This trend towards more extreme fire weather conditions is likely to overcome prevention efforts (Turco et al., 2016;Lahaye et al., 2018) in a region where expanding forests (Abadie et al. 2017) are increasing fuel loading and may offer opportunities for future fire spread.

Conclusions
This work, based on long-term geo-referenced large fire time series          and in the eastern part (in red) for different return periods ranging from 5 to 100 years.The 95% confidence intervals were estimated using a bootstrapping approach.

Figure 2 :
Figure 2: Time since the last LF (cat_age in years).

Figure 5 :
Figure 5: a) Annual number of LF (in black) and area burned by LF (in red) across the region.Significant change points at the 5% confidence level according to a Standard Normal Homogeneity Test (SNHT) in both metrics are indicated.Horizontal solid lines indicate the overall mean observed before and after the change point.b) Same as a) but for the western part.c) Same as a) but for the eastern part.

Figure 6 :
Figure 6: a) Time series of the annual maximum burned area in the western part (in gray) and in the 253 eastern part (in red).b) Return levels in annual maximum burned area in the western part (in gray) and 254 in the eastern part (in red) for different return periods ranging from 5 to 100 years.The 95% 255 confidence intervals were estimated using a bootstrapping approach.256 257The correlation between mean June-September FWI and LF activity was computed over 31-258 year sliding windows (Fig.7) and showed that much higher correlations in the western part than in the 259 eastern part.The relationship strongly weakened after 1990-1991, a weakening that is more 260 pronounced in the western part.261

Figure 7 :
Figure 7: Sliding correlations on 31-year windows between mean June-September FWI and a) annual LF frequency and b) and annual burned area due to LF.The horizontal dashed lines indicate different significance levels of the Pearson correlations.Correlations are indicated for the middle of the sliding windows.

FIGURE CAPTIONS Figure 1 :
FIGURE CAPTIONS

Figure 2 :
Figure 2: Map of the time since the last LF (cat_age in years).

Figure 3 :
Figure 3: Map of the LF recurrence on the 1961-2017 and 1958-2016 periods in the western and eastern parts, respectively with zooms on the areas presenting the highest recurrence.

Figure 4 :
Figure4: Top) Longitudinal cross-section of LF extent computed over 30-km sliding windows.The 95% confidence intervals were estimated using a bootstrapping approach.Bottom) Same as top panel but for mean June-September FWI (in red) and the percent of fuel cover (in green).

Figure 5 :
Figure 5: (a) Annual number of LF (in black) and area burned by LF (in red) across the region.Significant change points at the 5% confidence level according to a Standard Normal Homogeneity Test (SNHT) in both metrics are indicated.Horizontal solid lines indicate the overall mean observed before and after the change point.(b) Same as (a) but for the western part.(c) Same as (a) but for the eastern part.

Figure 6 :
Figure 6: (a) Time series of the annual maximum burned area in the western part (in gray) and in the eastern part (in red).(b) Return levels in annual maximum burned area in the western part (in gray)

Figure 7 :
Figure 7: Sliding correlations on 31-year windows between mean June-September FWI and (a) annual LF frequency and (b) and annual burned area due to LF.The horizontal dashed lines indicate different significance levels of the Pearson correlations.Correlations are indicated for the middle of the sliding windows.
), due to multiple overlaps in burned areas by recurrent fires (i.e.LF occurrence on the same surface).LF recurrence occurred up to 5 times in the West and up to 6 times in the East but represented only a small part of the recurrence (0.2% and 0.3%, respectively; Tab. 3).In contrast, one LF and two recurrent LF were the most frequent patterns in the western part of the study area (39.4 and 39.9% of the recurrence, respectively; Tab. 3) while, in the East, most LF occurred only once (46.3%) on the same surface.The surface impacted by only one LF represented 74.5% and 71.2% of the total area burned by LF in the West and the East, respectively during the period studied and, as previously shown, the burned area involved in the highest recurrence (area burned five and six times) was the lowest in both parts of the study area (0.005% and 0.008%, respectively; Tab. 3).Nat.Hazards Earth Syst.Sci.Discuss., https://doi.org/10.5194/nhess-2018-263Manuscript under review for journal Nat.Hazards Earth Syst.Sci. Discussion started: 21 September 2018 c Author(s) 2018.CC BY 4.0 License.
analyzed spatiotemporal variations of LF throughout one of the most impacted areas of the French Mediterranean.On the whole, 21% of the total area burned by LF occurred on a surface that already burned in the past, the region being impacted in some locations up to 5 or 6 times by recurrent LF.LF were less frequent in the eastern part but larger than LF occurring in the West.This longitudinal gradient in LF extent, with an average time of occurrence between LF exceeding 4000 ha <7 years and >50 years in the East and the West respectively, contrasts with what we would expect from mean fire weather conditions strongly decreasing eastwards but is consistent with larger fuel cover in the East.Recurrent LF happened mostly in the WUI (especially in the West) and on the coast (especially in the East), while non-recurrent LF were located inland in the East.Nat.Hazards Earth Syst.Sci.Discuss., https://doi.org/10.5194/nhess-2018-263Manuscript under review for journal Nat.Hazards Earth Syst.Sci. Discussion started: 21 September 2018 c Author(s) 2018.CC BY 4.0 License.
knowledge of the large fire regime is necessary to strengthen fire prevention by providing valuable information on priority areas where recurrent LF are more likely to occur.

Table 2 :
Percentages of area burned (relative to the total burned area) by LF and occurrence according 591 to the LF age classification in the two parts of the study area 592

Table 3 :
Percentages of burned area (relative to the total burned area) affected by recurrent LF and 594 percentages of recurrence relative to the number of LF recurrence in the two parts of the study area 595 Nat.Hazards Earth Syst.Sci.Discuss., https://doi.org/10.5194/nhess-2018-263Manuscript under review for journal Nat.Hazards Earth Syst.Sci. Discussion started: 21 September 2018 c Author(s) 2018.CC BY 4.0 License.Nat.Hazards Earth Syst.Sci.Discuss., https://doi.org/10.5194/nhess-2018-263Manuscript under review for journal Nat.Hazards Earth Syst.Sci. Discussion started: 21 September 2018 c Author(s) 2018.CC BY 4.0 License.