Increased wildfire hazard along South-Central Chile under the RCP8.5 scenario as revealed by high-resolution modeling

Among Mediterranean regions, the South-Central Zone of Chile (SCZCh) portrays a landscape where wildfires constantly and historically occur, many times damaging ecosystems, lives and livelihoods. Since 2010, this zone has entered a period of unprecedented drought that has contributed to wildfire rising trends. Wildfire occurrence and intensity in this zone can be associated with three main factors: climate and land cover as conditioning factors, and human activity as a triggering factor. This paper evaluates wildfire hazard for the SCZCh, based on environmental susceptibility to wildfire occurrence, using numerical regional climate and wildfire modeling for the 2002–2005 historical period and for the mid and late 21st-century under the RCP8.5 climate change scenario. Results indicate high skill in matching spatial patterns of fire spot occurrence and density in the historical period, as well as the ability to simulate seasonal behavior in wildfire environmental susceptibility, consistent with national historical statistics. The fire hazard in SCZCh will slightly increase in all seasons for both 2041–2050 and 2091–2100 periods, especially southward, with a long-term spatial homogenization of medium levels of hazard in Central Valley and Coastal Range, between 0 and 1000 m a.s.l. These results combined with the current homogeneous extensive exotic species plantations dominated by inflammable tree species in SCZCh might facilitate the occurrence of large wildfires under the projected 21st-century climate regime.


Introduction
A wildfire is an uncontrolled expansion of fire that, regardless of its origin, spreads to rural areas or urban-rural interfaces through ligneous, shrub or herbaceous vegetation, dead or alive, with hazard or damage to ecosystems and/or population [1][2][3][4]. Wildfire ignition and propagation are controlled by the load, distribution, continuity and accumulated potential energy in the fuel; as well as topography and wind speed [5][6][7][8][9]. The probability of ignition increases with high temperatures and low humidity in the air, intensifying its effect the greater the flammability of the vegetation [10].
Ubiquitous in all world's biomes, wildfires have played a critical role in ecosystem dynamics throughout Earth's history [7,11]. For instance, as recurring phenomena in Mediterranean climates especially during summer due to the combination of dry and warm conditions [1,3,[12][13][14], wildfires have contributed to the evolution of the flora's post-fire regeneration strategies, i.e. resprouting from roots or the root crown, or resprouting from lignotuber [15]. However, in the last decades, wildfires have become a concerning issue due to their increasing intensity and recurrence [16][17][18][19]. Some of the negative impacts of wildfires on the ecosystems are changes in the structure of forests, loss of fertile soil, loss of ecosystem services, desertification, changes in microclimate and atmospheric pollution, among others [1-3, 6, 12, 14, 19-21].
Alterations in wildfires' frequency and intensity may correlate with human activities at the global and regional scale. At the global scale, anthropogenic climate change may affect humidity, precipitation and temperature regimes. Several studies show that these climatic transformations, coupled with certain land use practices that impact fuel distribution contribute to increasing occurrence and intensity of wildfires, especially in Mediterranean climates [14,[17][18][19][22][23][24][25]. Some examples of land use practices are landscape fragmentation, densification of exotic pyrophyte plantation, fire suppression and accumulation of fuel (the so-called fire paradox [22]), land abandonment and extensive grasslands, among others [8,12,18,19].
Among Mediterranean regions, the South-Central Zone of Chile (SCZCh) (32 • -39.5 • S) constitutes a case study where most of the abovementioned alterations are occurring. A decade of 20%-30% precipitation deficit unprecedented in the last millennium [7,26], concurs with almost half a century of land use practices, leading this zone to concentrate 88% of Chile's pyrophyte exotic forest plantations, 24% of native forest and 91% of farmland [27]. Furthermore, seemingly unabated warming according to elevation suggest that conditions for wildfires may be expanding [28][29][30]. In fact, over the past years, the occurrence of large wildfires in the country (>200 ha) has increased [14,31]. South of 30 • S, climate conditions are strongly determinant in the occurrence of wildfires [12]. Climate projections for the 21st-century indicate a rise in annual temperature and precipitation decrease throughout Chile [32][33][34][35]. These warmer and drier conditions may lead to a rise in the frequency of wildfires [32,36] and heat waves [35]. This suggests that conditions for the second half of the century will be similar to those experienced in recent summers when large wildfires affected extensive areas of the SCZCh [14,24], especially if substantial changes in the land use and disaster management are not accounted for [24].
In this work, we evaluate 21st-century wildfire hazard at a spatial resolution of 9 km grid cell under the RCP8.5 climate change scenario [37], using the Potential Fire Index (PFI) model [6] fed by a recently validated regional climatic projection [38]. By analyzing environmental susceptibility to wildfires in the SCZCh (32 • -38 • S) (figure 1(a)), we aim to support risk management. Due to the historical impact of wildfires in Chile, many groups have implemented models to estimate the hazard [i.e. 1,3,5,36,39,40]. However, to our knowledge, only one projection of a long-term hazard exists [36] and thus we assert our work helps in filling a gap.
In order to provide context on the importance of wildfire projection for this region, in the next section we briefly describe key characteristics of wildfire occurrence in the SCZCh. We then continue detailing the methods and results, to end up with a section of discussion and conclusions.

Wildfire features in the SCZCh
Nationwide, 263 230 fires occurred between 1964 and 2021, affecting a total surface of more than three million hectares [31]. In this period, the number of fires has been rising as the linear trend shows in figure 1(b). This trend fitting the annual occurrence of wildfire explains 73% of the data variance, indicating a rising rate of 115 fires per year. That increase does not seem to correlate with the area affected, as the spike in the period 2016-2017 indicates an unusually severe season relative to the last 40 years. A simple ratio between affected area and number of fires per fire season (figure 1(c)) shows that the periods 1967-1968 and 2016-2017 had the highest figures, with 120 ha/fire and 108 ha/fire, respectively.
Within a predominant Mediterranean climate, with warm/dry summers and cold/rainy winters and a precipitation pattern that rises from North to South [42,43], during the austral summer season (DJF) the highest probability of fire occurrence takes place [31] (figure 1(f)). During this season, water stress of the vegetation imparts high flammability and facilitates fire propagation [26,44,45]. Since 2012 there has been an increase in the monthly affected area, especially between June and January (figure 1(g)), coinciding mainly with the megadrought period.
In addition to favorable hydrometeorological conditions, land cover change has affected fire exposure in the SCZCh. While native vegetation cover varies according to latitude from sclerophyll forests and shrublands in the North to temperate forests in the South [46], most of the SCZCh landscape has changed to exotic forest plantations dominated by the species Pinus radiata and Eucalyptus globulus, known for their flammability [47]. Larger and more severe wildfires have occurred in zones where those exotic plant species are dominant [2,14,17,24,[45][46][47][48][49]. For instance, in the summer of 2017, the SCZCh suffered the most severe wildfires on record ( figure 1(b)), affecting more than 560 000 hectares [31]. While lumbering and deforestation in Chile date back to the 16th-century [47], the forestry expansion that started in the decade of 1930, accelerated since 1974 as a result of economic incentives that facilitated forestation with exotic species [50][51][52]. Although climate and vegetation cover act as conditioning factors of fire activity in Chile, the main triggering factor for its occurrence is human activity, responsible for almost 90% of forest fires, either intentionally (∼33%) or accidentally (∼56%) [31].

Data and methods
Prediction of wildfires is difficult since it depends on the combination of various factors, both physical (meteorology, climate, estate of vegetation or fuel) and human (population density, exposition, etc) [3].
The prediction of wildfires tends to be inherently probabilistic, as with the choice of relevant variables depending on the spatial scale (i.e. the higher the spatial resolution human-related variables take greater relevance). Geographical information systems and mathematical models using logistic or multiple linear regression are common. The most representative variables are precipitation (also Day n-4 PP5 Day n-5 PP6_10 Between days n-6 and n-10 PP11_15 Between days n-11 and n-15 PP16_30 Between days n-16 and n-30 PP31_60 Between days n-31 and n-60 PP61_90 Between days n-61 and n-90 PP91_120 Between days n-91 and n-120 represented as consecutive days with no rain), temperature, relative humidity, vegetation cover (type, humidity), topography and wind direction (the last two in the case of propagation models) [i.e. 5,6,13,16,20,21,36,39,[53][54][55][56][57][58]. In this case, we implemented the PFI model [6] in Python 2.7 and used regional climatic projections to determine wildfire hazard for the 2002-2005, 2041-2050 and 2091-2100 periods. This model uses daily values of precipitation (represented with a factor of drought), maximum temperature in the air, minimum relative humidity, and type of vegetation [6,59]. The PFI calculation serves to estimate a daily wildfire hazard based on the environmental susceptibility to wildfire occurrence. The first step is to define accumulated precipitation (mm) for the 11 preceding periods (120 days) to the day to estimate (table 1). Then, for each period a precipitation factor (PF) is calculated with values ranging from 0 to 1 (table 2); these values are used in equation (1) to estimate the days of drought (DD). With equation (2) the basic risk (BR) is determined for each vegetation type; each vegetation type has a specific flammability constant (A) [59]. Afterward, minimum relative humidity and maximum surface temperature (both at 18 UTC) are considered in equations (3) and (4) respectively. All equations are based on 20 year empirical evidence of vegetation fires in Brazil [6]. After all these calculations, daily PFI is obtained with values ranging between 0 and 1 (table 3). Finally, the seasonal mean of PFI was calculated and its spatio-temporal variations were analyzed  The PFI model is used by the Instituto Nacional de Pesquisas Espaciais of Brazil in its Wildfire Monitoring Program since 2017 [59]. At a global scale, the predictive capacity of PFI is over 60% accuracy in detecting fire spots for Africa, Asia, South America, North America and the Caribbean, Southwest Pacific and Europe for the 2001-2016 period [8].

Input data
The climate data used here correspond to output from a 9 km, dynamical downscaling [60] for South-Central Chile (32 • -38 • S) produced using the nonhydrostatic Weather Research and Forecasting (WRF) model [61], forced by the bias-corrected version of the Community Earth System Model (CESM), a CMIP5 participating model [38]. Three periods were simulated: 1980-2005, 2040-2050 and 2090-2100. The downscaling for the hindcast period 1980-2005 adequately reproduced climatic features of the SCZCh, with slight positive biases for temperature and negative for precipitation [38], although smaller relative to previous studies [62]. More details on the implementation and validation of the downscaling are described in [38].
The periods used in this work are 2002-2005 (to coincide with Moderate Resolution Imaging Spectroradiometer (MODIS) validation data described later), 2041-2050 and 2091-2100 (according to modeling boundaries). The land cover used in PFI corresponds to data used as input in the WRF model for the dynamical downscaling with ∼1 km resolution [63] (figure 2). The projections considered the RCP8.5 scenario, that assumes a radiative forcing equal to or greater than 8.5 W m −2 where global surface temperature is likely to exceed 2 • C by the end of 21stcentury [37].

Model validation
To evaluate the efficiency of the PFI in the 2002-2005 period, each level of hazard is compared with the product MODIS Collection 6 NRT Hotspot/Active Fire Detections (MCD14ML) of 1 km resolution (figure 3), filtered according to the type of inferred focus (type = 0 or presumed vegetation fire) and confidence of the detection (confidence ⩾30; involving nominal-and high-confidence fire pixels) [64]. This product is built from a contextual algorithm that examines each pixel and assigns a class; it uses the strong emission of mid-infrared radiation from fires to detect them [64]. For potential fire pixels, valid neighboring pixels are used to estimate a background value, then a series of contextual threshold tests are performed for a relative fire detection. MODIS can detect flaming and smoldering fires about 1000 m 2 in size, although under very good or pristine observing conditions, 100 m 2 and 50 m 2 flaming fires can be detected respectively. National statistics were not used because its biases regarding fire location and fire distribution (depending on human record and access) [65]. For this regional study, MODIS fire spatialization is considered better.

Changes in temperature and precipitation
Relative to baseline (1980-2005), WRF-CESM climate projections for the mid-21st-century   Regarding temperature, climate projections for 2041-2050 indicate an increase in maximum and minimum daily mean temperatures at about 1 • C and 2 • C respectively, with the highest increase in JJA. Between 2041-2050 and 2091-2100 maximum temperature will rise between 1.5 • C and 1.9 • C, and the minimum temperature between 2.2 • C in SON and 2.6 • C in DJF, MAM and JJA. When comparing the 2002-2005 and 2091-2100 periods, maximum temperature will increase between 2.5 • C and 2.9 • C, while minimum temperature will increase in more than 3.5 • C, with the highest increases in fall and winter (+4 • C). In all periods, temperature descends from North to South in the Central Valley and Coastal Range while in Andean Range lower temperatures are North of 35 • S, according to highest altitudes ( figure 1(a)). Longitudinally, temperatures decrease, in general, from the Central Valley to the coast (west) and to the Andean Range (east).

PFI validation
In all periods, the PFI model shows two North-South strips: an Eastern Strip (ES) in which the low PFI prevails, especially North of 36 • S, and a Western Strip (WS) with higher levels of hazard (medium to high depending on the season) (figure 4). Both strips coincide with large topographic units; the ES coincides with the Andean Mountains (higher elevations, which decrease southwards from maximum altitudes of ∼4500 m a.s.l to ∼2000 m a.s.l, and lower temperatures) while the WS coincides with the Central Valley and the Coastal Range (lower elevations, higher temperatures). In DJF and MAM there is a rise of the PFI South of the ES due to decrease in Andean altitudes and increase in vegetation cover. The comparison between model results and satellite derived fire spots suggests high skill of the PFI in matching spatial patterns of fire spot occurrence and density, expressed as spots per pixel. Table 5 contains  These results indicate the model is able to simulate seasonal tendencies in wildfire environmental susceptibility, consistent with the national historical statistics [31] (figures 1 and 3).

PFI projections 4.3.1. Summer (DJF)
Under the RCP8.5 summers for the period 2041-2050 show a decrease in minimum, low and high PFI while increasing medium values (table 6), with a slightly noteworthy concentration of high pixels at ∼33.5 • S ( figure 4). The spatial distribution of medium values remains similar relative to the validation period, prevailing in less than or equal to ∼1000 m altitudes ( figure 5).
In 2091-2100, while low and medium PFI decrease relative to 2041-2050, high and critical values increase, demonstrating a sustained spread of these categories across the study area, concomitant with projected temperature rise and precipitation decrease expected for later this century [32-35, section 4.1]. North of 36 • S, the ES low hazard zone narrow towards higher altitudes (>3000 m) increasing the environmental susceptibility in the Andean foothills. South of that latitude, medium PFI continues expanding to the East. High PFI concentration migrates toward the South and become more ubiquitous in the WS. For this period, the boxplot indicates that low PFI has a greater distribution of altitudes in the study area, while the rest of the categories predominate in altitudes <1000 m and the pixels to the East (higher altitudes) are represented as outliers, similar to the other periods.

Fall (MAM)
For the period 2041-2050 high PFI increases in the North while minimum PFI decreases, relative to 2002-2005. In this season the projection suggests more areas with medium values than higher ones, coinciding with a seasonal decrease of environmental susceptibility. North of ∼35 • S a wider concentration of medium values is recognized. Low and medium pixel proportions remain similar relative to the historical period. The ES remains quite similar, as the boxplot and maps show.
For the late 21st-century, model output predicts a rising in medium and high values along with a decrease in minimum and low PFI. High pixels continue in the northern part of the study area while South of ∼35 • S medium values rise. PFI distribution according to elevation is similar to the previous period, although the interquartile range of minimum PFI becomes narrower due to expansion on the higher environmental susceptibility in the WS.

Winter (JJA)
In 2041-2050 low and medium PFI slightly increase relative to 2002-2005 and its altitudes' distribution remains similar. Some pixels of minimum PFI in the ES change to low values, indicating a rise of the environmental susceptibility in those zones while concentrating in higher altitudes. As in 2002-2005, low PFI generally remain in the South and East of the study area.
The projection for the 2091-2100 period shows that minimum and low values diminish and medium PFI (present North of ∼36 • S as in the other periods) rises, both with slight changes. The boxplots show a similar distribution of PFI to the previous period.

Spring (SON)
Increase of medium PFI, clustering between ∼33 • S and ∼36 • S and a westward rise in WS is projected for 2041-2050. While low and minimum PFI diminishes. SON boxplot for this period shows a slight rise of the interquartile range of medium values to higher altitudes, similar to MAM and JJA.
In 2091-2100, minimum and medium PFI rise relative to the previous period; medium values increase in WS, especially South of ∼36 • S, thus delivering a more evenly distributed level of hazard in this entire strip, while low values keep localized in the ES. In terms of altitudinal distribution, low and medium values remain similar to the previous period. Minimum PFI however, shows a significant decrease in its interquartile ranges and mean, to lower altitudes.

Discussion and conclusions
Anticipating the changes of fire regimes and future hazards becomes a useful tool for policy development on forest adaptation and fire management [67]. The combination of regional climate modeling and the PFI model provides more detailed scenarios relative to the coarse resolution of global climate models, as they increase the spatial resolution and better represent local climate features [38,62]. The dynamical downscaling output used in this work [38] provides the highest resolution climate change projection for the SCZCh with this methodology [33][34][35], along smaller validation biases relative to previous studies [62]. In spite of new statistical downscaling projections available with lower absolute biases [68], the low density of meteorological stations and its absence at high altitudes could be a limitation for high resolution wildfire hazard modeling at a regional level. Thus, our work should be considered as a first step to future more detailed studies, using new climatic projections.
Although summer is the season when the majority of wildfires occur in the study area, we analyzed all seasons since national historical statistics show an extension of the fire season through part of autumn and spring in the last decades.
Application of the PFI for the SCZCh (32 • -38 • S) revealed two North-South strips that coincide with the Chilean macrorelief and the distribution of fire spots according to MCD14ML product. PFI output for 2002-2005 shows the high dependency of wildfire seasonality on temperatures and precipitation annual cycles. In the ES low PFI dominates, while the WS concentrates the categories of major hazard. These patterns are similar to those determined for 1985-1988 period between 32 • and 43 • S based on climatic, vegetation, topographic and anthropogenic variables at 610 km 2 resolution [5]: three North-South strips correlated with the macrorelief, maximum hazard located in coastal and Central Valley areas, and considerably less hazard in the foothills and the Andean mountains. It is interesting to notice that MCD14ML fire spots shows a preferent pattern that the PFI model shows in all seasons: higher hazard in summer, medium for MAM and SON, and lower wildfires hazard in JJA. Therefore, the model shows predictive capacity as it correlates well with fire spots mapped with MCD14ML. Although in DJF, MAM and SON, all categories of hazard have fire spots, the PFI is able to map with the most hazardous zones where spots occur more regularly.
The ubiquitous presence of fire spots in spite of differences in modeled environmental susceptibility (PFI levels of hazard) could be associated with human activity, the main triggering factor for wildfires occurrence in Chile [31]. This kind of less predictable behavior is hard to incorporate in models; although some agent-based approaches are being developed for fire management, incorporating human behavior into climatic projections, and hence coupled wildfire modeling, remains a challenge [69].
To our knowledge, no studies on wildfire hazard in Chile have analyzed altitudinal distribution and thus our work provides information on unexplored pattern shifts that may support better planning. Results showed that these patterns are fairly consistent over time, i.e. differential behavior in each strip. In summer and fall months, the larger coverage of medium PFI in the ES South of ∼36 • S relative to the North can be associated with vegetation and altitude gradients. The lower Andean elevation in this zone likely determines that temperature remains relatively high as summits rarely surpass the 0 • C isotherm in summer [29,70]. On the other hand, medium and higher levels of hazard are observed in the WS, where altitudes decrease.
Our modeling results show that the strongest impact of climate change on wildfire susceptibility would be seen in the period 2041-2050 relative to 2002-2005, while towards the late 21st-century this rise will become the new regime. These outcomes are similar to previous studies focusing on Mediterranean biomes [67,[71][72][73][74][75]. For example, in Mediterranean areas of Italy (2071-2100, 25 km resolution) [72] and France (1995-2098, 8 km resolution) [73], the probability of wildfires will increase as the Canadian Fire Weather Index increases. Furthermore, a comparison of several wildfire hazard or wildfire activity projections for southern Europe shows that future fire hazard, burned area, number of fires and fire season will increase [67].
Wildfire hazard estimation from the PFI model shows that seasonal occurrence of medium levels of hazard becomes spatially homogeneous with time. Under future climatic scenarios of warming and drying, as revealed by the dynamical downscaling used in this study, the chances of megafires in the WS will probably increase in the future. Indeed, the area affected by wildfires in Chile between 2012 and 2021 increased relative to the 2002-2011 period (figure 1(g)) suggesting that more extreme weather conditions could facilitate that small wildfire become extreme wildfires [76]. Recent studies have proven a more accurate PFI model (PFIv2) for extra-tropical latitudes [9,77], thus an evaluation of this model for SCZCh is recommended.
As the PFI model does not incorporate vegetation evolution, it is difficult to assess its actual effect on susceptibility under the RCP8.5 scenario. Currently, extensive, homogeneous and continuous patches of forestry pyrophyte plantations are abundant in the SCZCh. If the area covered by exotic plantations remains or increases, this region could be even more prone to fires in the future. The inclusion of more realistic vegetation cover in the modeling (in type, dominance and continuity) has to be a priority for future analysis considering the close relationship between the activity of wildfires and vegetation cover in Chile, that some native forests in the SCZCh are less prone to wildfires [78] and the above-mentioned situation of the forestry activity.
As future wildfires hazard in Mediterranean regions has become a concerning issue due to climate change impacts, this work brings novel results since the SCZCh constitutes the major Mediterranean climate zone in South America. Regional analysis in these biomes can bring new perspectives both at global and local levels for discussion, planning, policy and decision-making processes, as well as in the development of wildfire hazard modeling.

Data availability statement
The data that support the findings of this study are available upon reasonable request from the authors.
de Valparaíso and the Department of Agriculture Engineering from Universidad Federal de Viçosa. This work was supported by FONDECYT 1171065 and ANID/FONDECYT 11190530, and is part of IC's Master Thesis on Regional Sciences, Universidad de Concepción. We would like to thank PhD student Carlos Carrasco Godoy for his help.