You are on page 1of 15

Science of the Total Environment 762 (2021) 143121

Contents lists available at ScienceDirect

Science of the Total Environment

journal homepage: www.elsevier.com/locate/scitotenv

Reassessing wildfire susceptibility and hazard for mainland Portugal


Sandra Oliveira, Ana Gonçalves, José Luís Zêzere ⁎
Centre for Geographical Studies, Institute of Geography and Spatial Planning, Universidade de Lisboa, Portugal

H I G H L I G H T S G R A P H I C A L A B S T R A C T

• We present new wildfire susceptibility


and hazard maps for mainland Portugal.
• Predisposing factors are landcover,
slope angle and elevation.
• We applied the Likelihood Ratio method
and used fire history from 1975 to 2018.
• Landcover favourability scores are ob-
tained for different timeframes.
• Structural maps can support spatial
planning and risk reduction strategies.

a r t i c l e i n f o a b s t r a c t

Article history: Recent large wildfires have caused serious environmental and social impacts in different areas of the world, such
Received 29 June 2020 as Portugal, Greece and Australia. Sociodemographic changes in rural areas and climate change issues create un-
Received in revised form 9 October 2020 precedented and challenging circumstances that call for an adjustment of fire management strategies and plan-
Accepted 11 October 2020
ning tools. In this research, we revise the procedure to assess wildfire susceptibility and hazard for mainland
Available online 20 October 2020
Portugal in a structural perspective. Based on a dataset of burned areas covering 44 years and a set of predisposing
Editor: Paulo Pereira factors related to topography and landcover, we tested several susceptibility models using the Likelihood Ratio
method, to calculate favourability scores. Landcover classes were analysed using national detailed sources and
Keywords: considering different timeframes. Hazard levels were obtained by combining the susceptibility model with the
Wildfires best accuracy and the probability to burn estimated from fire history (1975–2018). The resulting hazard values
Susceptibility model were classified in 5 levels based on the breaks of the success curve, instead of quantiles as in previous maps, to
Hazard map ensure territorial continuity at different spatial scales. The results show where the combination of terrain features
Structural approach is more favorable to fire propagation. The higher favourability scores were found for shrubland-type vegetation,
Spatial planning
whereas agricultural areas, cork and holm forests show lower scores. Eucalyptus and maritime pine forests show
similar intermediate scores and have increased since 2007. The two highest hazard classes classify correctly 90%
of the burned area over 44 years, demonstrating the high accuracy of the model. If integrated in spatial planning
instruments, in conjunction with municipal plans, the two most hazardous classes may hinder the expansion of
built-up areas. About 2% of the municipalities have more than 90% of their territory classified as hazardous,
whereas 32% of the municipalities have, instead, less than 10% of hazardous area. These structural maps are a
useful baseline for a long-term approach and can be complemented by estimations regarding the behavior and
severity of wildfires, which should be further explored.
© 2020 Elsevier B.V. All rights reserved.

⁎ Corresponding author.
E-mail address: zezere@campus.ul.pt (J.L. Zêzere).

https://doi.org/10.1016/j.scitotenv.2020.143121
0048-9697/© 2020 Elsevier B.V. All rights reserved.
S. Oliveira, A. Gonçalves and J.L. Zêzere Science of the Total Environment 762 (2021) 143121

1. Introduction (DL 327/80, 26th August). In 1981, a regulation established the zoning
of the mainland in four levels of wildfire risk, according to seven vari-
In the past two decades, wildfires have become increasingly destruc- ables: forest species distribution and their vulnerability to fire; combus-
tive around the world (Bowman et al., 2017; Tedim et al., 2020). tibility and flammability levels of understory species (shrubland);
Australia, Greece, Portugal and the USA, amongst others, have been af- terrain morphology; aspect; mean of maximum temperature and
fected by extreme wildfires that caused massive damages and human mean relative humidity between May and September; a population
losses (Moreira et al., 2020; Nolan et al., 2020; Ribeiro et al., 2020). In index (DR 55/81, 18th December). Two decades later, Pereira and
Portugal, wildfires are a recurring hazard, where they burned an annual Santos (2003) developed a different classification system, which was
average of 147,000 ha of land between 2000 and 2018 (San-Miguel- published by the National Forest Services (currently, Institute for Con-
Ayanz et al., 2019). The recent events of 2017, fueled by extreme servation of Nature and Forests, ICNF). The authors applied a classifica-
weather conditions occurred outside the typical main wildfire season, tion and regression tree algorithm to a set of variables, namely:
have brough devastating social and economic consequences (Gómez- landcover, elevation, slope, number of days with temperatures above
González et al., 2018; Sánchez-Benítez et al., 2018; Turco et al., 2019), 25 °C and number of days with rainfall between May and September,
and have flagged the need to adapt strategies and procedures to new the resident population by civil parish and the burned areas between
and challenging conditions. Driven by climatic changes and framed by 1990 and 1999, obtained through satellite imagery. They presented
the sociodemographic evolution of rural areas, which resulted in land the results in 5 classes of wildfire risk, with different probabilities over
abandonment and impacts farming activities, land use and fuel manage- a period of 30 years and with a spatial resolution of 1 km. Although it
ment (Moreira et al., 2011; Nunes et al., 2016; San-Miguel-Ayanz et al., was called risk, these results reflect, in fact, wildfire hazard levels,
2012; Tedim et al., 2013; Turco et al., 2014), Portugal, as other countries since they represented the spatial (susceptibility) and temporal (proba-
in Southern Europe, is now faced with unparalleled circumstances, bility) dimensions of wildfire occurrence and do not include potential
where large and uncontrollable severe wildfires become more frequent, impacts or consequences (vulnerability and value). Currently, it is
surpassing the capabilities of existing suppression resources and in- widely accepted that the concept of wildfire risk encompasses these dif-
creasing the exposure of communities (Fernandes et al., 2016; Moreira ferent components, with equivalent approaches being applied also for
et al., 2020). In these settings, wildfire management and mitigation ap- other environmental hazards; risk is, therefore, defined as the probabil-
proaches need to be adjusted and to incorporate the most up-to-date in- ity of a wildfire occurring at a specified location under given circum-
formation and modelling results that can efficiently support land use stances and its expected outcomes, measured by the impacts on the
planning and risk reduction practices. affected elements (Bachmann and Allgöwer, 2001; Hardy, 2005; IPCC,
Wildfire susceptibility and hazard mapping is a valuable tool to sup- 2012; Julião et al., 2009; Poljansek et al., 2017; UNDRO, 1979).
port wildfire management decisions, either as individual components or In the aftermath of the extreme wildfire years of 2003 and 2005
integrated in a risk assessment procedure. This can be done at multiple (Amraoui et al., 2015; San-Miguel-Ayanz et al., 2013; Trigo et al.,
temporal and spatial scales, depending on its purposes and applications 2006), new regulations and measures were implemented (DL 156/
(Jappiot et al., 2009; Miller and Ager, 2013). When used to identify fire- 2004, 30th June), and the zoning of the mainland was based on the clas-
prone areas based on factors that remain constant for at least one sea- ses defined by Pereira and Santos (2003). Two years later, this law was
son, or even for years, the assessment is carried out in a structural or revoked, in light of required conceptual adjustments, the experience of
long-term perspective, being suitable for the definition of prevention the prior extreme wildfire seasons and the development of new guide-
strategies, for land use planning and for the allocation of wildfire re- lines regarding land management and risk reduction, although the zon-
sources prior to the start of a fire season (Chuvieco et al., 2010; ing procedure remained the same (DL 124/2006, 28th June). In parallel,
Oliveira et al., 2012; Parente and Pereira, 2016; Sebastián-López et al., between 2006 and 2009, the Portuguese Geographic Institute, now
2008). This approach differs from the hazard assessment that is carried called Directorate-General of the Territory, has developed another
out in a dynamic perspective, based on conditions that change daily, methodology, which originated the Forest Fire Risk Cartography
such as weather and fuel moisture, which are used instead to define (Cartografia de Risco de Incêndio Florestal, CRIF), applied at regional
local fuel treatments, for surveillance and in suppression activities and national levels (IGP, 2008). It included initially seven variables, spe-
(Ager et al., 2010; Fernandes, 2009; Scott et al., 2012). Both approaches cifically landcover, slope, aspect, density of roads, distance to roads,
coexist, in an autonomous but associated way, and dynamic methods population density and visibility of the surveillance stations, but the lat-
have been applied in areas where the structural approach indicates ter was eventually dropped (Verde, 2015). The analysis was based on a
high hazard levels. Indeed, these are the areas where a finer-scale and hierarchical analytical process and the weighting of the variables re-
short-term analysis can give useful details to fire managers, to comple- sulted from expert judgement. In view of the conceptual approach
ment the averaged long-term perspective given by the structural maps. here applied, CRIF is a susceptibility map. The zoning of the country in
This has been done, for example, with simulations of burn probability different classes of susceptibility (or hazard) define the type of wildfire
and fire propagation considering the effects of a fuel break network prevention and mitigation measures to be implemented. More rigorous
and different forest management options at the local scale (Oliveira actions are done in the most hazardous areas, such as constraints in the
et al., 2016, 2020). circulation of people and vehicles or the mandatory creation of safety
Considering a structural perspective, susceptibility represents the perimeters around villages through fuel management.
spatial dimension of wildfire hazard; based on terrain features, such as In the last decade, the research on wildfire risk has refined the dis-
topography and landcover, it evaluates the predisposition of each spa- tinction between the different risk components, providing a more sys-
tial unit (grid cell) to burn. When susceptibility is combined with the tematized procedure that facilitates their assessment as separated,
probability of wildfire occurrence, measured with the wildfire recur- though associated, modules. Verde and Zêzere (2010) used a set of var-
rence in each grid cell and representing a temporal dimension, hazard iables representing terrain predisposing factors, specifically landcover,
is obtained (Parente and Pereira, 2016; Verde and Zêzere, 2010). using the dataset of Corine Land Cover, CLC 2000 (Büttner et al.,
These tools allow to increase our understanding of the propensity of dif- 2004), elevation, slope, mean annual rainfall and mean number of
ferent territories to be affected by a wildfire and to estimate the likeli- nights above 20 °C temperature, together with wildfire history from
hood of future occurrences under equivalent terrain conditions. 1975 to 2004, to assess wildfire susceptibility and hazard in mainland
The assessment of wildfire susceptibility and hazard in Portugal has Portugal, at 80 m resolution. The burned areas from 1975 to 1994
been part of legal and planning instruments since 1980, typically incor- were used to calculate the favourability scores of the different variables
porated in risk assessment procedures with zoning purposes, in the per class, and the remaining dataset (burned areas from 1995 to 2004)
framework of the system of prevention and detection of forest fires was used for validation. The model which included only landcover,

2
S. Oliveira, A. Gonçalves and J.L. Zêzere Science of the Total Environment 762 (2021) 143121

slope and wildfire probability showed the best predictive behavior. The 0 m a.s.l. along the coastline, to 1993 m a.s.l. in the Estrela mountain
resulting values were then classified in 5 levels, based on quantiles. This range in the central part of the country. Mountains, plateaus and
methodology, generally called CSP (Cover, Slope, Probability) was deep incised valleys with steep slopes dominate in the north and
adopted by the National Forest Services, to generate annual hazard central parts of the country, whereas large plains and low-lying
maps at the national level. This same scheme was adopted as a reference hills dominate in the south (Pereira et al., 2020). The climate is Med-
to the municipal plans of forest defence against fires, since 2012 (AFN, iterranean with warm-summer in the north (Köppen class Csb) and
2012). The modelling procedure was further tested with a dynamic ap- hot-summer in the south (Köppen class Csa). As a rule, the natural
proach, showing the consistency of the variables used to assess suscep- vegetation is evergreen, resilient to drought and pyrophytic
tibility (Verde, 2015). In another recent study to assess structural (Pereira et al., 2018).
wildfire risk in Portugal, Parente and Pereira (2016) have updated and Generally, the natural conditions of the country are prone to wildfire
improved the susceptibility and hazard maps proposed by Verde and and these have been aggravated in the last decades by demographic and
Zêzere (2010). They used a digital terrain model with better resolution socio-economic factors. Most rural areas suffered extensive depopula-
(at 25 m), added 10 more years of burned areas (until 2013), and used tion since the middle of the 20th century, as well as the progressive
landcover data for 3 different periods (Corine Land Cover 1990, 2000, abandonment of cultivated lands, which were converted to forest plan-
2006). The authors showed that the predictive values of the five suscep- tations or reverted into shrublands (Pereira et al., 2018).
tibility classes improve with the new data, while the spatial patterns of
the classes distribution remain the same throughout the country. Few 2.2. Data collection and pre-processing
years later, Leuenberger et al. (2018) compared the approach first
used by Verde and Zêzere (2010), with two stochastic methods based The dependent variable derived from fire history, specifically data
on machine-learning, namely extreme learning machine (ELM) and on burned area (BA). The information was obtained from the na-
random forest (RF), for a region in Central Portugal. This research tional burned areas mapping database for the mainland Portugal, be-
followed the recent tendency to apply machine-learning methodologies tween 1975 and 2019, provided by the National Forest Services
in wildfire susceptibility assessment, as used in several studies for dif- (ICNF, 2020). The fire perimeters of individual events are recorded
ferent regions (Arpaci et al., 2014; Cao et al., 2017; Gholamnia et al., in vector format; to ensure a minimum consistency of the records,
2020; Hong et al., 2018; Pourtaghi et al., 2016; Tehrany et al., 2019; since the minimum detectable area for remote sensing data varied
Tonini et al., 2020). along the considered period (Tonini et al., 2017), we retained only
In the case of Portugal, the authors found that the results were sim- the fire perimeters above 30 ha until 1983, and those above 5 ha re-
ilar amongst the three methods and the ranking of importance of the corded afterwards. The fire perimeters were aggregated by year,
variables was comparable, showing that both the deterministic and converted in raster format and assembled in two different layers:
the stochastic approaches are suitable to assess wildfire susceptibility i) the number of times each grid cell (at 25 m resolution) was
in the country, in a structural perspective (Leuenberger et al., 2018). burnt between 1975 and 2018, for the total period of 44 years; ii)
Following the disastrous events occurred in 2017 (Gómez-González the number of times each grid cell was burnt between 1995 and
et al., 2018; Turco et al., 2019), the National Forest Services spurred the 2018, as this timeframe matches the available landcover data. The
revision of the existing wildfire susceptibility and hazard maps for area burnt in 2019 was used for the independent validation as it
mainland Portugal, to be integrated in the risk assessment procedures was not used for modelling.
that feed the spatial planning instruments, and to inform the medium The independent variables were based on topography and
to long-term strategies of risk reduction. The main purpose of this re- landcover, as prior studies have shown their influence in wildfire in-
search is to produce new wildfire susceptibility and hazard maps for cidence in Portugal (Carmo et al., 2011; Nunes et al., 2016; Oliveira
mainland Portugal, built from prior experience while incorporating up- and Zêzere, 2020; Verde and Zêzere, 2010). The variables tested,
dated data and improved methods. Although the prior approaches were which represent the terrain predisposing factors, were based on
accurate enough considering the data available at the time, this update slope, elevation, aspect and landcover. The topographic data were
was required because a more recent landcover dataset at national level, obtained from the Digital Elevation Model (DEM) available from
and new wildfire data (burned areas since 2015, including the severe the European Environment Agency (EEA), at 25 m resolution and in
year of 2017) were available and needed to be incorporated. The hazard raster format; this DEM is a hybrid product based on SRTM and
map should inform fire managers about the propensity of different ter- ASTER GDEM data fused by a weighted averaging approach (EEA,
ritorial units to burn, based on the spatial distribution of the most rele- 2020). Based on prior studies of wildfire susceptibility in Portugal
vant predisposing factors, combined with yearly probability to burn, (Verde and Zêzere, 2010; Verde, 2015), slope was reclassified in suc-
and it is expected to feed spatial planning and wildfire management cessive steps of 5 degrees (0–5°; 5–10°; 10–15°; 15–20°; > 20°),
regulations The mapping procedure should be consistently applied at whereas elevation was reclassified in 11 categories, each with
different spatial scales (from local to national) and able to be routinely 100 m until the 800 m threshold, and with larger intervals after-
updated. Here, we present the procedure applied to achieve these ob- wards (0–100 m; 100-200 m; 200-300 m; 300-400 m; 400-500 m;
jectives and discuss the results of the different models tested, based 500-600 m; 600-700 m; 700-800 m; 800-1000 m; 1000-1500 m;
on three major assumptions: i) the spatial distribution of wildfires in 1500–2000 m). Aspect was classified in 9 categories, based on the 8
Portugal's mainland, as evidenced by burned area, is not random; ii) different geographical directions plus a flat option (flat areas,
the predisposition to burn can be quantified through statistical relation- north, northeast, east, southeast, south, southwest, west and north-
ships between past burned areas and a relevant set of spatial data; iii) west) (Table 1).
the conditions under which wildfires occur can be characterized by The landcover data were obtained, in vector format, from the
the spatial data selected, which represent predisposing factors in the Directorate-General of the Territory (DGT, 2020), and represent sev-
modelling procedure. eral years: 1995, 2007, 2010 and 2015. We considered only the
landcover types representing forested or rural areas, and excluded
2. Materials and methods all artificial and water surfaces, after a required harmonization of
the classes according to the latest dataset of 2015, because the cate-
2.1. Study area gories varied slightly in the different maps. We obtained 35 distinct
landcover classes, including 10 representing forests, 4 of shrubs,
Mainland Portugal is located in the southwest part of the Iberian grasses and sparse vegetation, 7 of agroforests and 11 of agricultural
Peninsula and extends over 89,015 km2. The elevation ranges from areas, including pastures (Table 2).

3
S. Oliveira, A. Gonçalves and J.L. Zêzere Science of the Total Environment 762 (2021) 143121

Table 1 Table 2
Likelihood Ratio scores (LRi) of slope, elevation and aspect classes, considering two Likelihood Ratio scores (LRi) for landcover classes integrated in wildfire susceptibility and
timeframes. hazard assessment.

Slope angle (degree) LRi (1975–2018) LRi (1995–2018) Categories Landcover classes LRi LRi LRi LRi Weighted
0–5 0.442 0.446 1995 2007 2010 2015 LRi mean
5–10 1.216 1.229
Agricultural Temporary irrigated 0.103 0.201 0.149 0.250 0.149
10–15 1.939 1.940
areas crops
15–20 2.405 2.382
Temporary dryland crops 0.211 0.399 0.200 0.166 0.225
>20 2.666 2.582
Temporary crops and/or 0.138 0.191 0.761 0.368
Elevation (meter) LRi (1975–2018) LRi (1995–2018) pastures associated with
0–100 0.228 0.247 olive groves
100–200 0.440 0.467 Temporary crops and/or 0.769 0.217 0.277 0.375
200–300 0.699 0.776 pastures associated with
300–400 1.138 1.206 orchards
400–500 1.656 1.683 Temporary crops and/or 0.182 0.113 0.227 0.168
500–600 1.924 1.839 pastures associated with
600–700 2.017 1.901 vineyards
700–800 2.012 1.906 Complex cropland 0.262 0.211 0.213 0.747 0.326
800–1000 2.642 2.430 systems
1000–1500 2.665 3.279 Olive groves 0.304 0.179 0.213 0.255 0.261
1500–2000 1.390 1.714 Orchards 0.149 0.174 0.138 0.155 0.151
Vineyards 0.109 0.097 0.154 0.160 0.126
Aspect (classes) LRi (1975–2018) LRi (1995–2018) Permanent pastures 0.438 0.306 0.143 0.094 0.303
Flat 0.058 0.040 Agriculture with natural 0.644 1.129 0.661 0.766 0.728
N 1.021 1.038 and semi-natural areas
NE 0.975 0.977 Agroforests Agroforests of holm oak 0.087 0.119 0.064 0.037 0.078
E 1.017 1.006 Agroforests of cork oak 0.192 0.051 0.097 0.087 0.137
SE 1.033 1.029 Agroforests of cork and 0.122 0.063 0.078 0.065 0.096
S 0.955 0.958 holm oak
SW 0.941 0.936 Agroforests of other oak 0.573 2.030 0.310 0.385 0.669
W 1.024 1.018 trees
NW 1.106 1.115 Agroforests of stone pine 0.091 0.001 0.123 0.010 0.073
Agroforests of other 0.569 0.788 0.599 0.397 0.574
species
Agroforests of mixed 0.308 0.039 0.068 0.050 0.181
2.3. Data analysis forest species
Forests Forests of holm oak 0.469 0.666 0.226 0.139 0.388
Forests of cork oak 0.750 0.179 0.472 0.207 0.530
The methodological flowchart to assess wildfire susceptibility and Forests of chestnut trees 2.973 0.953 0.537 0.305 1.768
hazard for the entire mainland Portugal is showed in Fig. 1 and is ex- Forests of other oaks 1.454 2.816 2.002 1.204 1.697
plained in detail in the next subsections. Forests of eucalyptus 1.638 0.705 1.313 2.099 1.531
Forests of other 0.772 2.453 1.285 1.436 1.200
broadleaved
2.4. Wildfire Probability Forests of maritime pine 1.546 1.036 1.418 2.205 1.566
Forests of stone pine 0.255 0.217 0.355 0.193 0.261
The annual probability to burn was calculated as the number of years Forests of other conifers 1.087 1.700 0.924 0.837 1.088
each grid cell (at 25 m resolution) was burnt between 1975 and 2018, Forests of invasive 0.559 1.724 1.107 3.959 1.386
species
calculated with the following equation:
Shrub & Shrubland 2.744 3.474 3.443 2.002 2.857
Grass Other woodland 2.636 1.413 1.429 1.724
Natural herbaceous 1.319 0.916 0.598 0.911
f
wp ¼ ð1Þ vegetation
N Sparsely vegetated areas 3.457 6.019 3.786 2.191 3.635
Other Fire breaks 0.742 0.243 0.867 1.051 0.757
Forest nurseries 0.000 0.000 0.020 0.007
Where: Greenhouses and 0.017 0.051 0.041 0.040
wp – wildfire annual probability. nurseries

f – number of times the grid cell was burnt. Number of years represented by the 12 3 5 4
N – number of years (44 years in this study). dataset
For the areas that were never burnt during this period, a base value
of 0.01 was given, which corresponds to an annual probability of 1%.
This step was necessary to prevent the exclusion of the areas that
The LR represents the ratio between the probability of a wildfire to
were never burnt (and have zero probability), when multiplied with
occur and the probability that it will not occur (Lee, 2004). It is first cal-
positive values of the other component. This value is below the proba-
culated for the whole area, considering the conditions given by all grid
bility value obtained for the areas that were burnt only once in the con-
cells that compose the study area, in this case mainland Portugal and
sidered timeframe (0.023, or 2.3%).
calculated with the following formula:
2.5. Wildfire susceptibility
Si=S
Lri ¼ ð2Þ
We applied the deterministic method of Likelihood Ratio (LR, also Ni=N
called Frequency Ratio) to assess wildfire susceptibility. This method
is based on Bayesian statistics and has been widely used in the spatial Where
analysis of environmental hazards, including wildfires (Bergonse and Lri – Likelihood Ratio of the variable i (class x of predisposing factor y)
Bidarra, 2012; Ehret et al., 2010; Nóbrega et al., 2018; Pradhan et al., Si – number of burned grid cells (considering the number of times
2007; Yilmaz, 2009). each grid cell has burned in 44 years) and presence of variable i

4
S. Oliveira, A. Gonçalves and J.L. Zêzere Science of the Total Environment 762 (2021) 143121

Fig. 1. Methodological flowchart to assess wildfire susceptibility and hazard for mainland Portugal. Solid arrows – modelling steps; dashed arrows – validation and classification steps.

Ni – number of grid cells where variable i is present the case of landcover, it was assumed that, conceptually, the relation
S – total number of burned grid cells (considering the number of with the burned areas can only be reasonably tested for the year that
times each grid cell has burned in 44 years) the landcover data represents and the following ones, not in previous
N – total number of grid cells of the study area (burned and non- years. As such, landcover data of 1995 was analysed with regard to
burned) burned areas between 1995 and 2006, the landcover data of 2007
Lri values between 0 and 1 indicate a negative relation between the with the burned areas between 2007 and 2009, the landcover of 2010
presence of a variable i (class x of predisposing factor y) and the burned with burned areas between 2010 and 2014 and the landcover data of
areas, whereas Lri values above 1 indicate a positive relation, which in- 2015 was analysed considering the burned areas recorded from 2015
creases with higher values. to 2018. Therefore, the final scores obtained for landcover result from
On a second step, the Lri scores obtained for each variable (classes of the mean of these multiple values, weighted by the number of years at-
the predisposing factors) are attributed to each grid cell, by summing up tached to each landcover dataset. Due to the lack of available data prior
the scores corresponding to the variables present in each cell. The Lrj is to 1995, the burned areas between 1975 and 1994 were disregarded in
calculated per grid cell as the sum of the Lri variable classes found in the analysis of landcover. To test if the exclusion of the BA before 1995
each grid cell, calculated with: would affect the final scores of wildfire susceptibility, we calculated
the LR scores for the topographic data using the two datasets created
n
Lrj ¼ ∑i¼0 Xij Lri ð3Þ with the burned areas, one including the BA for all the 44 years, and
the other matching the timeframe of the available landcover data.
Where:
Lrj – Total Likelihood Ratio of the grid cell. 2.7. Susceptibility models
n – number of variables (classes of predisposing factors).
Xij - 1 if that class of the variable is present; 0 if that same class is Based on the LR method, we tested 4 susceptibility models, using dif-
absent. ferent sets of independent variables. Starting with a baseline model that
The results of Lri do not represent a direct probability, but instead includes only landcover and slope, we tested the introduction of the
the positioning of each variable i (the class x of each predisposing factor other variables (elevation and aspect), first individually and then com-
y) in relation to the mean density of the burned areas (the dependent bined, to evaluate their relative influence in the final scores. The final
variable), according to Bayesian theory. A Lri score above 1 indicates susceptibility value derives from the sum of the scores obtained for
that the grid cell is susceptible to the occurrence of a wildfire. The higher every variable (classes of predisposing factors) in each cell using
the score, the more susceptible that grid cell is to burn. Eq. (3). The models' performance was evaluated according to a valida-
tion process based on success-rate curves and area under the curve
2.6. Favourability scores of the independent variables (AUC), as explained in detail below (validation).

We applied the LR method to calculate, first, the score of each vari- 2.8. Wildfire hazard
able i (class x of each predisposing factor y, i.e., elevation, slope, aspect
and landcover), and then to obtain the aggregated score for the grid Hazard levels result from the combination of probability with suscepti-
cells (25 m resolution) that cover the entire mainland of Portugal. In bility, by multiplying the value of annual probability with the susceptibility

5
S. Oliveira, A. Gonçalves and J.L. Zêzere Science of the Total Environment 762 (2021) 143121

score of each cell. The susceptibility model which showed the highest AUC
with the lower number of independent variables, was selected to be inte- AUCmax ¼ 1−ððBurned Area=Total AreaÞ=2Þ ¼ 0; 5 ≤AUCmax≤1 ð6Þ
grated in hazard assessment, based on the following equation:
Where:
Hj ¼ Lr j  wp ð4Þ Burned Area – total coverage of burned areas
Total Area – size of the study area
Where: Considering that the total coverage of burned area in mainland
Hj – Hazard value for the grid cell (25 m resolution); Portugal, between 1975 and 2018, corresponds to 31% of the territory,
Lrj – Total Likelihood Ratio for the grid cell; the maximum AUC value expected is 0.845. According to (Oliveira,
wp – wildfire annual probability for the grid cell. 2012), all AUC values above 0.66 indicate an acceptable result, and
The hazard map was subsequently classified in 5 levels (very low; those values above 0.7 reflect already a very good model.
low; moderate; high; very high), as required by the Portuguese forest
authority. These classes were defined according to the breaks of the 3. Results
success-rate curve and the predictive capacity of the hazard model.
The spatial distribution of the different wildfire hazard classes was 3.1. Wildfire probability
analysed considering the area they occupy in each municipality. To
map the municipalities with the highest and lowest proportion of haz- The annual probability to burn, based on fire history, ranges between
ardous area, we aggregated the cells corresponding to the two top clas- 1 and 39% (Fig. 2). Considering the time interval used (44 years), this
ses (high and very high hazard) based on the municipal boundaries. We means that the areas with a 39% probability can burn 17 times in an
then calculated the area occupied by these two classes together, in rela- equivalent period. The areas more affected by wildfires are located in
tion to the total area of the municipality, and retrieved those where the the North and Central regions of the mainland, particularly in inland
hazardous area occupied more than 90% of the territory, or instead, less areas. In the southernmost region of Algarve, the mountainous area of
than 10% of the municipal area. the western part also shows a high wildfire probability. The flattest re-
gion of Alentejo shows the lowest probability to burn.
2.9. Validation
3.2. Susceptibility - favourability scores of the variables
The quality of wildfire susceptibility and hazard maps depends on
their ability to anticipate the future location of burned areas. The valida- The LRi (favourability) scores for the topographic variables show
tion of the models was based on success-rate curves and on the calcula- analogous values and comparable positions regarding the two
tion of the Area Under the Curve (AUC), a common procedure used to timeframes tested (1975–2018 and 1995–2018, Table 1). This indicates
test the robustness of predictive models (Bi and Bennett, 2003; Chung that the use of a shorter period does not affect the favourability scores.
and Fabbri, 2005; Fabbri and Chung, 2008; Jaafari et al., 2018; Vilar Regarding slope, favourability scores increase consecutively in the
et al., 2010)). The success-rate curve is obtained by confronting the pre- upper classes and the class below 5° is the only with a score below 1, in-
dictive model with the same burned areas (dependent variable) used to dicating a negative relationship with burned area. Elevation below
create it, which allows to measure the goodness-of-fit of the model. The 300 m is also negatively related with burned area, whereas favourability
success-rate curve was computed for both the susceptibility and hazard scores increase consecutively in the following classes until 1500 m,
models. In the case of the susceptibility model, it combines the grid cells when favourability decreases again but is kept above 1. Regarding as-
classified as susceptible, organized in descending order of susceptibility pect, this variable does not present a clear spatial relationship with
(from the highest Lrj to the lowest Lrj), with the fraction of the burned burned area, as seen in the similar favourability scores close to 1 of all
area that is consecutively validated. The procedure is similar to compute classes, except for the flat areas (Table 1).
the success-rate curve for the hazard model, and grid cells are organized Landcover favourability scores are highly variable. All the classes
in descending order of hazard (from the highest Hj to the lowest Hj) and representing agricultural areas have mean LRi scores below 1, including
crossed with the burned area. Moreover, the hazard model was crossed pastures and agroforests (Table 2), thus indicating a negative relation-
with the area burned in 2019 to verify the (independent) predictive ca- ship with burned areas. The highest mean score was obtained for
pacity of the model, as the corresponding burnt area was not used in the sparsely vegetated areas (LRi = 3.635), followed by shrubland (LRi =
modelling process. 2.857). Most forests show scores above 1, with the highest value ob-
The calculation of the AUC allows to quantify the global predictive tained for chestnut forests (LRi = 1.768). Exceptions are found for
ability of a model (Bi and Bennett, 2003). It ranges between 0 and 1, stone pine, holm oak and cork oak forests, with LRi scores of 0.261,
with higher values, close to 1, indicating a better predictive capabil- 0.388 and 0.530, respectively, indicating a negative relation of these for-
ity. A value of 0.5 means a neutral discriminating power and random est types with burned areas. Eucalyptus and maritime pine forests show
results, whereas values below 0.5 should be disregarded. a similar mean LRi value, of 1.531 and 1.566, respectively. For both for-
The formula to calculate the AUC is given by: ests, the value has increased since 2007, reaching LRi values above 2 in
2015. On the contrary, LRi has progressively decreased for chestnut for-
Xn   ests since 1995, when it reached nearly a value of 3, whereas it
ðai þ biÞ
AUC ¼ ðLsi−LiÞ  ð5Þ remained below 1 since 2007. A reverse tendency is found for forests
2
i¼1 of invasive species, whose LRi value has consecutively increased, from
0.559 in 1995 to 3.959 in 2015 (Table 2).
Where: The spatial distribution of the predisposing factors is highly irregular
Lsi – Li – fraction of the total study area corresponding to successive in the mainland territory (Fig. 3). The highest elevation and slope clas-
Lrj scores organized in descending order (Hj scores in the case of hazard ses are found in the mountainous areas of north and center regions,
model). whereas in the largest region in the south (Alentejo) the lower topo-
ai – fraction of the total burned area corresponding to Li graphic classes predominate. A clear distinction between these regions
bi - fraction of the total burned area corresponding to Lsi is also found with regard to landcover, as agricultural areas and
The maximum value of AUC depends on the coverage of the burned agroforests dominate the south. Forests of conifers, mainly of maritime
areas used to validate each susceptibility model, in relation to the study pine, are found mostly in the coastal areas of the central region, whereas
area, and its calculation uses the following equation: eucalyptus also expand towards inner areas.

6
S. Oliveira, A. Gonçalves and J.L. Zêzere Science of the Total Environment 762 (2021) 143121

Fig. 2. Burned areas recorded in Portugal between 1975 and 2018 (on the left); Annual wildfire probability in Portugal, at 25 m resolution (on the right). Source of data: ICNF, 2019.

3.3. Susceptibility models maximum expected value (0.845). Instead, the validation using the
area burned in 2019, that was not used in the model, did not provide
All the models tested showed AUC values above 0.7, indicating a very an equivalent result, with the AUC reaching only 0.742.
high degree of model fit (Table 3). The baseline model 1, that included To classify the hazard values in 5 qualitative levels, we identified the
only landcover and slope, showed an AUC of 0.712. This value increased thresholds of the classes based on the breaks of the success-rate curve
to 0.720 when elevation was incorporated (model 2). On the contrary, (Fig. 5), the trend of the success-rate curve segments, as well as the frac-
the integration of aspect (model 3) did not change the AUC value in re- tion of burned area correctly classified as susceptible (Table 4). The class
lation to the baseline model. The incorporation of all the independent area is described as the percentage of burnable area in the mainland,
variables (model 4) did not show differences in relation to model 2. since it includes only the landcover classes that can be affected by wild-
When tested with the burned areas dataset with the shorter timeframe fires and excluding water and urban areas (the excluded areas cover 8%
(1995–2018), the AUC value of model 2 increased to 0.730. of the total mainland area). The class that represents the highest hazard
Susceptibility values range from 0.67 to 9.96, and the higher values level (very high) covers 15.5% of the total burnable area in the mainland
are found in inland areas of the north and center regions (Fig. 4). The and validates 50% of the total burned area (Table 4). The second highest
areas more prone to wildfire have slope higher than 15° and elevation hazard class (high) covers 18% of the area and validates 40.6% of the
ranging from 600 to 1500 m. The most favorable conditions to wildfire total burned area. The third and fourth hazard classes (moderate, low)
occurrence are observed where these topographical features combine cover 18% and 21.2%, respectively, and validate only a small fraction of
with landcover containing shrublands and sparsely vegetated areas, as the total burned area (6.6% and 2.8%, respectively). The least hazardous
well as forests of chestnut trees, other oaks, eucalyptus, maritime pine class (very low) does not overlap any burned area.
and invasive species. The quality of the hazard model is confirmed by the two highest haz-
ard classes being able to correctly classify 90% of the burned area (each
burned pixel is counted once) over the entire period (1975–2018).
3.4. Wildfire hazard The hazard map was overlapped with the burned areas, per year
since 1975 to 2018 and the results are showed in Fig. 6. For each year,
Wildfire hazard was calculated by multiplying the probability values we calculated the proportion of burned area that fall in the very high
with the susceptibility scores obtained with model 2, that includes (orange line in Fig. 6) and high + very high classes (red line in Fig. 6)
landcover, slope and elevation as predisposing factors, for each grid given by the new wildfire hazard map. The annual burned area falling
cell (25 m resolution). The hazard scores varied between 0.007 and in the very high hazard class ranges from 44.9% to 90.1%, being the
2.481 (Table 4). The AUC of the hazard model reached 0.823 (Fig. 5), yearly average 76.1%. These figures increase when the very high and
which indicates a particularly high goodness of fit, close to the high hazard classes are aggregated, showing that the annual burned

7
S. Oliveira, A. Gonçalves and J.L. Zêzere Science of the Total Environment 762 (2021) 143121

Fig. 3. Spatial distribution of the predisposing factors integrated in susceptibility assessment; A) elevation, B) slope, C) aspect and D) landcover (only the major categories of the latest
COS2015 are shown). Sources: EEA, 2019; DGT, 2019.

8
S. Oliveira, A. Gonçalves and J.L. Zêzere Science of the Total Environment 762 (2021) 143121

Table 3 area correctly classified ranges from 83.5% to 99.4% (with an average of
Area Under Success-rate curves (AUC) using the complete set of burned areas 95.5%). The two highest classes of the wildfire hazard map validate al-
(1975–2018).
most perfectly the burned area recorded in years with average values
Models Variables in the model AUC of burned area (around 115,000 ha). However, the ability of the hazard
Model 1 Landcover, slope 0.712 model to capture the burned area decreases in atypical years, with
Model 2 Landcover, slope, elevation 0.720 values high above the average and characterized by extensive burned
Model 3 Landcover, slope, aspect 0.712 areas (e.g. years 2003 and 2017), as well as in those years with much
Model 4 Landcover, slope, elevation, aspect 0.720
lower values of burned area in relation to the average (e.g. years 1976
and 2008) (Fig. 6).

Fig. 4. Wildfire Susceptibility Map for mainland Portugal (original values from model 2).

9
S. Oliveira, A. Gonçalves and J.L. Zêzere Science of the Total Environment 762 (2021) 143121

Table 4 4. Discussion
Properties of the wildfire hazard classes.

Hazard Hazard Trend of the Class area (% Burned 4.1. Modelling approach
classes scores success rate curve burnable area in the area (% of
(min – mainland) total) We applied a statistically-based modelling procedure to assess
max)
wildfire susceptibility and hazard in mainland Portugal, using vari-
Very high 0.142–2.841 y = 3.23× - 1E-15 15.5 50.0 ables that represent terrain conditions, specifically topography and
High 0.050–0.142 y = 2.43× + 0.16 17.9 40.6
landcover, and the burned areas recorded since 1975. We started
Moderate 0.028–0.050 y = 0.34× + 0.80 18.0 6.6
Low 0.015–0.028 y = 0.13× + 0.91 21.2 2.8 with a baseline susceptibility model, composed solely of landcover
Very low 0.007–0.015 y = 8E-14× + 1.0 27.4 0.0 and slope data, as it is currently applied in the wildfire management
plans at the local and national levels (AFN, 2012); local plans include
the same classes of landcover and slope found for the national level
The spatial distribution of the hazard classes shows a marked dif- considering their relation with burned areas (e.g. slope above 20°
ference between the north/central regions of the mainland Portugal, have the highest weight), and the final susceptibility map is classi-
with a general pattern of high and very high wildfire hazard, in rela- fied in five levels based on quantiles, where each class included
tion to Alentejo in the south, where the very low class predominates 20% of the municipal territory. Then, we tested three other models
(Fig. 7). The highest hazard class is found in the top northwest of the by consecutively adding data on elevation, aspect and both. The
country, in the inland areas of the central region and in the western model that included also elevation showed a slightly better perfor-
part of Algarve. Most coastal areas show low to intermediate wildfire mance than the baseline one, with an AUC value of 0.720, while the
hazard levels, with few clusters of high hazard levels dependent of models that included aspect did not show higher predictive ability
the underlying landcover (e.g. forest of maritime pine), and of the in comparison to the others. The positive relation of elevation with
wildfire probability. When aggregated at the municipal level, we burned area distribution has been found by other authors
found that 2.2% of the municipalities have at least 90% of their terri- (Marques et al., 2011; Oliveira and Zêzere, 2020), and in this case,
tory classified in the high and very high hazard classes. These are lo- the positive effect of elevation in burned area is likely due to proxy
cated in inland sections of central Portugal and in the west of Algarve relationships with wind speed, which tends to increase at higher al-
(Fig. 7B). In contrast, there are 32.4% of municipalities where the two titudes and, therefore, can influence fire spread, namely when fire
most hazardous classes cover less than 10% of the territory, which propagation is driven by the wind (Costa et al., 2011). In addition,
are instead located in the northern coastal area and in the Alentejo some pastoral activities developed at higher altitudes are known to
region in the south (Fig. 7B). influence fire ignition in Portugal (Catry et al., 2009), particularly in
We calculated the consensus between our map and the previous the mountainous areas of north and central Portugal. Other possible
official wildfire hazard map produced in 2019 (ICNF, 2019), at grid reasons may be the lower accessibility of these areas to suppression
cell level. We found that 47% of the area remained in the same means, and the reduced possibility of surveillance and timely detec-
class, 38% of the area in the new map have lower hazard levels than tion of wildfires as a result of the low population density (Fernandes
the previous map, and 14% a higher hazard class. High hazard levels et al., 2016; Moreira et al., 2010). The decrease of favourability above
decrease in the south and increase in the inner north and central 1500 m, attested by the lower LRi score, can be explained by the re-
regions. duction of the available fuel load.

Fig. 5. Success-rate curves of the susceptibility and hazard model and independent validation using the area burned in 2019. In colour: Hazard classes defined from the breaks of the
success-rate curve and correctly classified burned area. VH – Very high; H – High; M – Moderate; L – Low; VL – Very low.

10
S. Oliveira, A. Gonçalves and J.L. Zêzere Science of the Total Environment 762 (2021) 143121

Fig. 6. Retrospective evaluation of the predictive capacity for the very high and high hazard classes, showing the proportion of burned area that fall in the very high (orange line) and high
+ very high classes (red line) per year.

Climatic variables were not integrated, as previous studies showed (Freire and DaCamara, 2018; Pinto et al., 2016; Salis et al., 2016), but los-
how their inclusion did not increase the performance of the susceptibil- ing predictive power when applied in a structural perspective. Extreme
ity model (Verde and Zêzere, 2010; Verde, 2015). Meteorological data is weather conditions may be the reason why the model's predictive abil-
essential in dynamic approaches, for the daily or short-term assessment ity decreased in years that fall outside standard situations, when burned
of wildfire hazard and for fire behavior simulations, being suitable for area was either very low or very high in relation to average values, as
operational responses during surveillance and suppression phases suggested by prior research (Costa et al., 2011; Oliveira and Zêzere,

Fig. 7. A) Wildfire Hazard Map for mainland Portugal, classified with the breaks of the success-rate curve. B) Municipalities where the classes of high and very high hazard cover more than
90% (in dark red) and less than 10% (in yellow) of their total area.

11
S. Oliveira, A. Gonçalves and J.L. Zêzere Science of the Total Environment 762 (2021) 143121

2020). Nevertheless, a structural wildfire susceptibility map should be hazard levels, because they burn often, may not be the ones with the
considered as a benchmark that represents the dominant spatial pat- highest associated risk and potential for damages, precisely because a
terns of burned area distribution over a long period of time, which can higher frequency of wildfires prevents the accumulation of high fuel
be complemented by additional, and more specific, analysis when loads, capable of feeding mega-fires. On the other hand, areas that
needed. have a very high susceptibility, but burn infrequently, must be moni-
The use of a simple model with few variables has several advan- tored rigorously, since the potential for large fires exists, especially
tages; on the one hand, it is easy to understand and to process, facilitat- where large and continuous extensions of forest and shrubland are as-
ing its application in different domains and by different users; on the sociated with unfavourable topographic conditions.
other hand, since it includes variables that are commonly available, its Another drawback of this work relates to the fact that the hazard
application and update are not hindered by the lack of appropriate model was validated independently using the burned area of a single
data. This parsimonious model is, therefore, a reasonable compromise year (2019) and the predictive capacity of the models measured by suc-
between a good predictive ability and simplicity, providing a stable cess rates is likely overestimated. Verde and Zêzere (2010) and Verde
tool that is easy to generalize (Catry et al., 2010a, 2010b; Leuenberger (2015) have shown that the drop of AUC of prediction-rate curves ob-
et al., 2018; Parente and Pereira, 2016; Verde and Zêzere, 2010; Vilar tained with independent validation can be larger than 0.1 in relation
et al., 2010) and can be easily replicated in other countries. to the AUC of the corresponding success-rate curves. In this work we
Compared to the previous versions of hazard mapping for main- found an equivalent drop when comparing the AUC corresponding to
land Portugal (Parente and Pereira, 2016; Verde and Zêzere, 2010), the independent validation of the 2019 burned area (AUC = 0.742)
the procedure here presented underwent several changes. Specifi- with the AUC of the success rate of the model (AUC = 0.845). The con-
cally, the use of landcover data obtained from national sources, the ditions that frame wildfire occurrence show high interannual variabil-
calculation of favourability scores using four landcover datasets ity, which is reflected in annual changes in the predictive capacity of
with specific timeframes, the integration of elevation data and the the model, as shown in Fig. 6. The hazard map can be further validated
definition of hazard classes based on the breaks of the success-rate when new burned area data becomes available in the following years
curve, instead of quantiles. The quantile classification method is con- and, considering the straightforward approach applied, it can be easily
sidered neutral and suitable for linearly distributed data, grouping upgraded if needed.
the same number of features in each class; when the CSP method
was applied at the municipal level the aggregation of the resulting 4.2. The relation of landcover with burned areas
maps at regional level showed inconsistencies amongst adjacent
municipalities, due to the inherent distortions of the classification, Overall, shrubland and sparsely vegetated areas showed the highest
since different features can be placed in the same class, or con- mean LRi scores and remained high over time. The positive relationship
versely, equivalent features can be distributed by different classes. of shrubland-type vegetation with burned areas is widely recognized, in
Therefore, it is expected that the new procedure will allow a better Portugal (Barros and Pereira, 2014; Carmo et al., 2011; Nunes et al.,
integration of the data when applied at different scales, because 2016; Oliveira and Zêzere, 2020) and in other southern European coun-
the thresholds of the classes are always the same. tries (Lampin-Maillet et al., 2011; Moreira et al., 2011; Oliveira et al.,
The assessment of hazard levels included historical burned areas in 2014). Shrubland-type vegetation can rapidly colonize abandoned
two different steps of the procedure; to obtain the favourability scores farmland and recently burned areas (Moreira et al., 2011; Pereira
of each class of the predisposing factors, and to calculate the annual et al., 2014), growing every year fostered by precipitation, which is
probability. The dual use of the same dataset does not compromise the higher in areas with irregular topography. In addition, the economic
results, as the difference between susceptibility and hazard is clear. value of shrublands is very low (Canadas et al., 2016) and there are no
The introduction of the annual probability allows to upgrade suscepti- incentives to their management, besides the mandatory wildfire hazard
bility (that reflects the spatial distribution of fire-prone factors) to haz- reduction activities.
ard levels and allows individualizing those susceptible areas that are Out of the 10 forest types, seven show a positive relation with
also most frequently affected by wildfires. Conceptually, even if a certain burned areas, stronger for invasive species, maritime pine and eucalyp-
area is more susceptible to wildfires due to its topographic and tus. This relation has varied over time, with invasive species gaining
landcover conditions, this is not sufficient to justify a high fire incidence, high favourability in the latest survey (2015), whereas the chestnut for-
since factors linked to ignition causes, land management activities and ests scores have strongly decreased. These results are likely associated
suppression strategies are also relevant drivers of burned area. More- with the abundance of the species in the different timeframes consid-
over, these maps do not show where wildfires can be more severe, al- ered, with invasive species expanding extensively in recent years
though these should coincide with the areas with higher susceptibility throughout the Portuguese territory (Morais et al., 2017). The LRi
but not necessarily with the most hazardous ones, since higher fire fre- score of maritime pine forests has followed a constant trend and was al-
quency can reduce the fuel loads and hinder the occurrence of future ways above 1, being on average slightly higher than the score of euca-
fires. Additional work is needed to obtain a deeper understanding re- lyptus forests. Although the abundance of eucalyptus species has
garding the link between susceptibility, hazard and severity, bearing greatly increased in Portugal in the last decades (Alves et al., 2007), re-
in mind that extreme fires are harder to control and may overpower cent research has shown that this expansion has not increased burned
the suppression capabilities in place, therefore increasing the risk to area nor modified the fire regime (Fernandes et al., 2019). On the
people and communities. other hand, eucalyptus can be naturally established in burnt areas
One of the caveats of this study is the exclusion of the severity of the (Águas et al., 2014) and can expand rapidly if the land is abandoned
fires and fire behavior parameters, due to lack of data in a format com- (Fernandes et al., 2019), therefore these areas colonized by eucalyptus
patible with a structural approach applied at the national level. Wildfire can contribute to increase wildfire hazard when left unmanaged.
severity measures the effects of wildfires on vegetation, visible for ex- Three of the forest types showed a LRi value below 1, namely cork
ample, in the crown tree consumption and in vegetation recovery and holm oak, and stone pine forests. This pattern may be linked to
rates, and it can be estimated with the Normalized Burn Ratio (NBR), the level of forest management implemented in these forests, rather
as done previously in Portugal (Catry et al., 2010a, 2010b; Maia et al., than to the characteristics of the species. Landcover data does not pro-
2012; Teodoro and Amaral, 2019). These have been applied to specific vide information on the level of forest management implemented,
and localised fire perimeters and are not available for most burned although it is expected that the removal of understory vegetation and
areas, which hinders their integration at the national level (Tedim related fuel treatments, could decrease wildfire hazard (Mirra et al.,
et al., 2013). In this respect, we admit that the areas with the highest 2017; Salis et al., 2018).

12
S. Oliveira, A. Gonçalves and J.L. Zêzere Science of the Total Environment 762 (2021) 143121

All types of agricultural areas, as well as all agroforests, showed a LRi 5. Conclusion
value below 1, a constant trend over time. The lower fire-proneness of
agricultural areas has been evidenced by prior research in different Wildfire management and mitigation activities must be framed
areas (Oliveira and Zêzere, 2020; Sil et al., 2019) and the farmland within territorial regulations. Since 1980, Portugal has incorporated
around villages remains as a potential buffer against wildfires, except hazard and risk zoning procedures in spatial planning instruments,
in the case of extreme wildfire events, when selectivity of fire towards which evolved over time. The reassessment of wildfire susceptibility
specific landcover types is drastically reduced (Barros and Pereira, and hazard in mainland Portugal here presented, brings new elements
2014). in relation to prior versions. First, it updated the inventory of burned
areas until 2018, thereby also including the records of 2017, the worst
4.3. Implications to spatial planning and fire risk mitigation year so far regarding area burnt. Second, it incorporated landcover
data from national and more detailed sources, and the new procedure
According to the national system of forest defence against fires to determine the favourability scores of the diverse landcover types en-
(DL 124/2006, 28th June), the fire protection network includes a abled the use of four datasets covering different timeframes. Thirdly, the
wildfire hazard map, designed at the municipal level, which should susceptibility model now integrates elevation data, besides slope and
comprise, similar to the national hazard map, five hazard classes landcover; the spatial distribution of elevation is more uniform, as com-
(very high, high, medium, low and very low). These classes have pared with slope, which results in more homogeneous classes of suscep-
been defined based on quantiles, which incorporate the same pro- tibility and hazard, with added benefits for territorial management as
portion of the territory in each class (20%). In addition, the wildfire the resulting terrain units are less spatially fragmented. The susceptibil-
hazard map, must be incorporated and regulated within the munici- ity and hazard maps that resulted from this research should be consid-
pal plans for land use, namely the municipal master plan. In particu- ered as baseline for a relatively long period, typically between 5 and
lar, the law establishes that, in order to mitigate the risk outside the 10 years. They can be updated when new landcover data becomes avail-
consolidated built-up areas, the construction of new buildings is not able, using the scores previously defined. The predictive ability of the
allowed in areas classified with high and very high wildfire hazard in model can, thus, be further tested, by adding in the validation process
the municipal map. With the classification system currently imple- new burned areas recorded after 2019.
mented, in each Portuguese municipality there is 40% of the territory Our findings can contribute to define silvicultural activities for wild-
forbidden to new built-up areas, due to the combination of this rule fire prevention, and are of particular use when articulated with spatial
with the quantile classification system applied at the municipal planning strategies due to their structural nature. Wildfire susceptibility
scale (20% very high class +20% high class). With the approach and hazard zoning provides much needed guidance to the regulation of
used here, it is considered that the very high and high hazard classes, built-up areas and socioeconomic activities in hazardous conditions.
defined at the national level, should be the ones selected to integrate
the spatial planning instruments at the municipal scale. These two
classes correspond in total to 33.4% of the agricultural and forestry CRediT authorship contribution statement
areas of the country and validate 90.6% of the burned area recorded
between 1975 and 2018. The spatial distribution of these two hazard Sandra Oliveira: Conceptualization, Formal analysis, Methodology,
classes, representing the most dangerous area regarding wildfires, is Writing - review & editing. Ana Gonçalves: Data curation, Formal anal-
not uniform in the Portuguese municipalities, as it is also verified in ysis, Investigation, Visualization. José Luís Zêzere: Conceptualization,
the patterns of wildfire distribution. The area covered by the high Formal analysis, Methodology, Writing - review & editing.
and very high wildfire hazard classes, where the construction of
new buildings will eventually be prohibited, exceeds 90% of the mu- Declaration of competing interest
nicipal area in 2.2% of the municipalities of the mainland. In contrast,
the most hazardous and restricted area is below 10% of the municipal The authors declare that they have no known competing financial
territory in 32.4% of the municipalities, and below 0.5% in 17 of them. interests or personal relationships that could have appeared to influ-
In comparison with the former borderline of 40% based on quantiles, ence the work reported in this paper.
the proposed hazardous area is higher in 41.4% and lower in 58.6% of
the municipalities. Based on this new approach, the area restricted to Acknowledgments
new built-up developments will be higher in the cluster of munici-
palities where the high + very high classes exceed 40% of the total This work was financed by national funds through FCT—Portuguese
area. However, as a general pattern, the need of new built-up area Foundation for Science and Technology, I.P., under the framework of the
in these municipalities is likely to be rather low, because, according project “People&Fire: Reducing Risk, Living with Risk (PCIF/AGT/0136/
to PORDATA (2020), they are losing population (−23.6% on average 2017) and by the Research Unit UIDB/00295/2020 and UIDP/00295/
between 1981 and 2018) and the aging index (number of elderly per 2020. The authors are grateful to ICNF and POSEUR for allowing the
100 young people) is very high (233.7 in average). In contrast, the use of data obtained in the framework of the project “Análise da
municipalities less constrained by the spatial extension of wildfire Suscetibilidade de Incêndio Rural em Portugal Continental” (POSEUR-
hazard are growing in population (2.4% in average between 1981 02-1810-FC-00504-ASIRPC). Ana Gonçalves was supported by Pahl
and 2018) and the aging index is substantially lower (171.2 in aver- Consulting.
age), thus the need for new construction can be justified by the pop-
ulation dynamics. In any case, the areas restricted to new built-up References
developments due to high wildfire hazard can be reserved for other
uses and alternative economic activities compatible with fire safety AFN, 2012. Plano Municipal de Defesa da Floresta Contra Incêndios (PMDFCI). Guia
guidelines. Also, the built-up clusters currently located in highly haz- Técnico.
Ager, A.A., Vaillant, N.M., Finney, M.A., 2010. A comparison of landscape fuel treat-
ardous areas can be regulated according to other protection rules ment strategies to mitigate wildland fire risk in the urban interface and preserve
and mitigation measures, such as fuel load reduction around build- old forest structure. For. Ecol. Manag. 259, 1556–1570. https://doi.org/10.1016/j.
ings, the creation of fuel discontinuities in the perimeter of the set- foreco.2010.01.032.
tlements or the implementation of a fuel-break network at the Águas, A., Ferreira, A., Maia, P., Fernandes, P.M., Roxo, L., Keizer, J., Silva, J.S., Rego,
F.C., Moreira, F., 2014. Natural establishment of Eucalyptus globulus Labill. In
landscape level, adjusted to the spatial distribution of hazard levels burnt stands in Portugal. For. Ecol. Manag. 323, 47–56. https://doi.org/10.1016/
within the territory of each municipality. j.foreco.2014.03.012.

13
S. Oliveira, A. Gonçalves and J.L. Zêzere Science of the Total Environment 762 (2021) 143121

Alves, A.M., Pereira, J., Silva, J.M.N., 2007. A introdução e a expansão do eucalipto em Jaafari, A., Zenner, E.K., Pham, B.T., 2018. Wildfire spatial pattern analysis in the Zagros
Portugal. In: Alves, A.M., Pereira, J.S., Silva, J.M.N. (Eds.), O Eucaliptal Em Portugal. Mountains, Iran: a comparative study of decision tree based classifiers. Ecol. Inform.
Impactes Ambientais e Investigação Científica. ISAPress, Lisboa, pp. 13–24. 43, 200–211. https://doi.org/10.1016/j.ecoinf.2017.12.006.
Amraoui, M., Pereira, M.G., DaCamara, C.C., Calado, T.J., 2015. Atmospheric conditions as- Jappiot, M., Gonzalez-Olabarria, J.R., Lampin-Maillet, C., Borgniet, L., 2009. Assessing wild-
sociated with extreme fire activity in the Western Mediterranean region. Sci. Total fire risk in time and space. Living with Wildfires: What Science Can Tell us. A Contri-
Environ. 524–525, 32–39. https://doi.org/10.1016/j.scitotenv.2015.04.032. bution to the Science-Policy Dialogue, pp. 41–47.
Arpaci, A., Malowerschnig, B., Sass, O., Vacik, H., 2014. Using multi variate data mining Julião, R., Nery, F., Ribeiro, J., Castelo Branco, M., Zêzere, J., 2009. Guia metodológico para a
techniques for estimating fire susceptibility of Tyrolean forests. Appl. Geogr. 53, produção de cartografia municipal de risco e para a criação de sistemas de
258–270. https://doi.org/10.1016/j.apgeog.2014.05.015. informação geográfica (SIG) de base municipal. ANPC, DGOTDU, IGP, Lisboa (doi:
Bachmann, A., Allgöwer, B., 2001. A consistent wildland fire risk terminology is needed. 298930/09).
Fire Manag. Today. https://doi.org/10.2307/j.ctvx078wk.11. Lampin-Maillet, C., Long-Fournel, M., Ganteaume, A., Jappiot, M., Ferrier, J.P., 2011. Land
Barros, A.M.G.G., Pereira, J.M.C.C., 2014. Wildfire selectivity for land cover type: does size cover analysis in wildland-urban interfaces according to wildfire risk: a case study
matter? PLoS One 9, e84760. https://doi.org/10.1371/journal.pone.0084760. in the South of France. For. Ecol. Manag. 261, 2200–2213. https://doi.org/10.1016/j.
Bergonse, R., Bidarra, J., 2012. Probabilidade bayesiana e regressão logística na avaliação foreco.2010.11.022.
da susceptibilidade à ocorrência de incêndios de grande magnitude. Finisterra 45, Lee, S., 2004. Application of likelihood ratio and logistic regression models to landslide
79–104. https://doi.org/10.18055/finis1353. susceptibility mapping using GIS. Environ. Manag. 34, 223–232. https://doi.org/
Bi, J., Bennett, K.P., 2003. Regression error characteristic curves. 20th International Confer- 10.1007/s00267-003-0077-3.
ence on Machine Learning (ICML), p. 8. Leuenberger, M., Parente, J., Tonini, M., Pereira, M.G., Kanevski, M., 2018. Wildfire suscep-
Bowman, D.M.J.S., Williamson, G.J., Abatzoglou, J.T., Kolden, C.A., Cochrane, M.A., Smith, tibility mapping: deterministic vs. Stochastic approaches. Environ. Model. Softw. 101,
A.M.S., 2017. Human exposure and sensitivity to globally extreme wildfire events. 194–203.
Nat. Ecol. Evol. 1, 1–6. https://doi.org/10.1038/s41559-016-0058.
Maia, P., Pausas, J.G., Vasques, A., Keizer, J.J., 2012. Fire severity as a key factor in post-fire
Büttner, G., Feranec, J., Jaffrain, G., Mari, L., Maucha, G., Soukup, T., 2004. The CORINE land
regeneration of Pinus pinaster (Ait.) in Central Portugal. Ann. For. Sci. 69, 489–498.
cover 2000 project. EARSeL eProceedings 3, 331–346.
https://doi.org/10.1007/s13595-012-0203-6.
Canadas, M.J., Novais, A., Marques, M., 2016. Wildfires, forest management and land-
Marques, S., Borges, J.G., Garcia-Gonzalo, J., Moreira, F., Carreiras, J.M.B., Oliveira, M.M.,
owners’ collective action: a comparative approach at the local level. Land Use Policy
Cantarinha, A., Botequim, B., Pereira, J.M.C., 2011. Characterization of wildfires in
56, 179–188. https://doi.org/10.1016/j.landusepol.2016.04.035.
Portugal. Eur. J. For. Res. 130, 775–784.
Cao, Y., Wang, M., Liu, K., 2017. Wildfire susceptibility assessment in Southern China: a
Miller, C., Ager, A.A., 2013. A review of recent advances in risk analysis for wildfire man-
comparison of multiple methods. Int. J. Disaster Risk Sci. 8, 164–181. https://doi.
agement. Int. J. Wildl. Fire 22, 1–14. https://doi.org/10.1071/WF11114.
org/10.1007/s13753-017-0129-6.
Carmo, M., Moreira, F., Casimiro, P., Vaz, P., 2011. Land use and topography influences on Mirra, I.M., Oliveira, T.M., Barros, A.M.G., Fernandes, P.M., 2017. Fuel dynamics fol-
wildfire occurrence in northern Portugal. Landsc. Urban Plan. 100, 169–176. https:// lowing fire hazard reduction treatments in blue gum (Eucalyptus globulus) plan-
doi.org/10.1016/j.landurbplan.2010.11.017. tations in Portugal. For. Ecol. Manag. 398, 185–195. https://doi.org/10.1016/j.
Catry, F.X., Rego, F.C., Bação, F.L., Moreira, F., 2009. Modeling and mapping wildfire igni- foreco.2017.05.016.
tion risk in Portugal. Int. J. Wildl. Fire 18, 921–931. https://doi.org/10.1071/WF07123. Morais, M., Marchante, E., Marchante, H., 2017. Big troubles are already here: risk assess-
Catry, F.X., Rego, F., Moreira, F., Fernandes, P.M., Pausas, J.G., 2010a. Post-fire tree mortality ment protocol shows high risk of many alien plants present in Portugal. J. Nat.
in mixed forests of central Portugal. For. Ecol. Manag. 260, 1184–1192. https://doi. Conserv. 35, 1–12. https://doi.org/10.1016/j.jnc.2016.11.001.
org/10.1016/j.foreco.2010.07.010. Moreira, F., Catry, F.X., Rego, F., Bacao, F., 2010. Size-dependent pattern of wildfire igni-
Catry, Filipe X., Rego, F.C., Silva, J.S., Moreira, F., Camia, A., Ricotta, C., Conedera, M., 2010b. tions in Portugal: when do ignitions turn into big fires? Landsc. Ecol. 25,
Fire starts and human activities. Towards Integrated Fire Management Outcomes of 1405–1417. https://doi.org/10.1007/s10980-010-9491-0.
the European Project Fire Paradox, pp. 9–21. Moreira, F., Viedma, O., Arianoutsou, M., Curt, T., Koutsias, N., Rigolot, E., Barbati, A.,
Chung, C.F., Fabbri, A.G., 2005. Systematic procedures of landslide hazard mapping for risk Corona, P., Vaz, P., Xanthopoulos, G., Mouillot, F., Bilgili, E., 2011. Landscape - wildfire
assessment using spatial prediction models. In: Glade, T., Anderson, M., Crozier, M. interactions in southern Europe: implications for landscape management. J. Environ.
(Eds.), Landslide hazard and risk. Wiley, pp. 139–174. Manag. 92, 2389–2402. https://doi.org/10.1016/j.jenvman.2011.06.028.
Chuvieco, E., Aguado, I., Yebra, M., Nieto, H., Salas, P.J., Martín, M.P., Vilar, L., Martínez, J., Moreira, F., Ascoli, D., Safford, H., Adams, M.A., Moreno, J.M., Pereira, J.M.C., Catry, F.X.,
Martíin, S., Ibarra, P., De la Riva, J., Baeza, J., Rodríguez, F., Molina, J.R., Herrera, M.Á., Armesto, J., Bond, W., González, M.E., Curt, T., Koutsias, N., McCaw, L., Price, O.,
Zamora, R., 2010. Development of a framework for fire risk assessment using remote Pausas, J.G., Rigolot, E., Stephens, S., Tavsanoglu, C., Vallejo, V.R., Van Wilgen, B.W.,
sensing and geographic information system technologies. Ecol. Model. 221, 46–58. Xanthopoulos, G., Fernandes, P.M., 2020. Wildfire management in Mediterranean-
https://doi.org/10.1016/j.ecolmodel.2008.11.017. type regions: paradigm change needed. Environ. Res. Lett. 15, 011001. https://doi.
Costa, L., Thonicke, K., Poulter, B., Badeck, F., 2011. Sensitivity of Portuguese forest fires to org/10.1088/1748-9326/ab541e.
climatic, human, and landscape variables: subnational differences between fire Nóbrega, L.O., Lazzarini, G.M.J., Viola, M.R., Batista, A.C., de Carvalho, E.V., Giongo, M., 2018.
drivers in extreme fire years and decadal averages. Reg. Environ. Chang. 11, Forest fire susceptibility index for assessing the history of fire occurrences in the in-
543–551. https://doi.org/10.1007/s10113-010-0169-6. digenous land of Kraholândia, Brazil. Adv. For. Sci 34, 223–232.
Ehret, D., Rohn, J., Dumperth, C., Eckstein, S., Ernstberger, S., Otte, K., Rudolph, R., Nolan, R.H., Boer, M.M., Collins, L., Resco de Dios, V., Clarke, H., Jenkins, M., Kenny, B.,
Wiedenmann, J., Xiang, W., Bi, R., 2010. Frequency ratio analysis of mass movements Bradstock, R.A., 2020. Causes and consequences of eastern Australia’s 2019–20 season
in the Xiangxi catchment, Three Gorges Reservoir area, China. J. Earth Sci. 21, of mega-fires. Glob. Chang. Biol. 26, 1039–1041. https://doi.org/10.1111/gcb.14987.
824–834. https://doi.org/10.1007/s12583-010-0134-9. Nunes, A.N., Lourenço, L., Castro Meira, A.C., 2016. Exploring spatial patterns and drivers
Fabbri, A.G., Chung, C.J., 2008. On blind tests and spatial prediction models. Nat. Resour. of forest fires in Portugal (1980–2014). Sci. Total Environ. 573, 1190–1202. https://
Res. 17 (2), 107–118. doi.org/10.1016/j.scitotenv.2016.03.121.
Fernandes, P.M., 2009. Combining forest structure data and fuel modelling to classify fire Oliveira, S.C., 2012. Incidência espacial e temporal da instabilidade geomorfológica na
hazard in Portugal. Ann. For. Sci. 66, 415. https://doi.org/10.1051/forest/2009013. bacia do Rio Grande da Pipa (Arruda dos Vinhos). Universidade de Lisboa.
Fernandes, P.M., Pacheco, A.P., Almeida, R., Claro, J., 2016. The role of fire-suppression Oliveira, S., Zêzere, J.L., 2020. Assessing the biophysical and social drivers of burned area
force in limiting the spread of extremely large forest fires in Portugal. Eur. J. For. distribution at the local scale. J. Environ. Manag. 264, 110449. https://doi.org/
Res. 135, 253–262. https://doi.org/10.1007/s10342-015-0933-8. 10.1016/j.jenvman.2020.110449.
Fernandes, P.M., Guiomar, N., Rossa, C.G., 2019. Analysing eucalypt expansion in Portugal
Oliveira, S., Oehler, F., San-Miguel-Ayanz, J., Camia, A., Pereira, J.M.C., 2012. Modeling spa-
as a fire-regime modifier. Sci. Total Environ. 666, 79–88. https://doi.org/10.1016/j.
tial patterns of fire occurrence in Mediterranean Europe using Multiple Regression
scitotenv.2019.02.237.
and Random Forest. For. Ecol. Manag. 275, 117–129. https://doi.org/10.1016/j.
Freire, J.G., DaCamara, C.C., 2018. Using cellular automata to simulate wildfire propagation
foreco.2012.03.003.
and to assist in fire prevention and fighting. Nat. Hazards Earth Syst. Sci. Discuss.,
Oliveira, S., Pereira, J.M.C., San-Miguel-Ayanz, J., Lourenço, L., 2014. Exploring the spatial
1–17 https://doi.org/10.5194/nhess-2018-227.
patterns of fire density in Southern Europe using Geographically Weighted Regres-
Gholamnia, K., Gudiyangada Nachappa, T., Ghorbanzadeh, O., Blaschke, T., 2020. Compar-
sion. Appl. Geogr. 51, 143–157. https://doi.org/10.1016/j.apgeog.2014.04.002.
isons of diverse machine learning approaches for wildfire susceptibility mapping.
Symmetry (Basel) 12, 604. https://doi.org/10.3390/sym12040604. Oliveira, T.M., Barros, A.M.G., Ager, A.A., Fernandes, P.M., 2016. Assessing the effect of a
fuel break network to reduce burnt area and wildfire risk transmission. Int. J. Wildl.
Gómez-González, S., Ojeda, F., Fernandes, P.M., 2018. Portugal and Chile: longing for sus-
Fire 25, 619. https://doi.org/10.1071/wf15146.
tainable forestry while rising from the ashes. Environ. Sci. Pol. 81, 104–107. https://
doi.org/10.1016/j.envsci.2017.11.006. Oliveira, S., Gonçalves, A., Benali, A., Sá, A., Zêzere, J.L., Pereira, J.M., 2020. Assessing risk
Hardy, C.C., 2005. Wildland fire hazard and risk: problems, definitions, and context. For. and prioritizing safety interventions in human settlements affected by large wildfires.
Ecol. Manag. 211, 73–82. https://doi.org/10.1016/j.foreco.2005.01.029. Forests 11, 859.
Hong, H., Tsangaratos, P., Ilia, I., Liu, J., Zhu, A.X., Xu, C., 2018. Applying genetic algorithms Parente, J., Pereira, M.G., 2016. Structural fire risk: the case of Portugal. Sci. Total Environ.
to set the optimal combination of forest fire related variables and model forest fire 573, 883–893. https://doi.org/10.1016/j.scitotenv.2016.08.164.
susceptibility based on data mining models. The case of Dayu County, China. Sci. Pereira, J.M.C., Santos, M.T., 2003. Cartografia das áreas queimadas e do risco de incêndio
Total Environ. 630, 1044–1056. https://doi.org/10.1016/j.scitotenv.2018.02.278. em Portugal Continental (1990–1999).
IGP, 2008. Cartografia de Risco de Incêndio Florestal. Relatório do Distrito de Santarém. Pereira, M.G., Aranha, J., Amraoui, M., 2014. Land cover fire proneness in Europe. For. Syst.
IPCC, 2012. Managing the risks of extreme events and disasters to advance climate change 23, 598–610. https://doi.org/10.5424/fs/2014233-06115.
adaptation. A Special Report of Working Groups I and II of the Intergovernmental Pereira, J.M.C., Alexandre, P., Campagnolo, M., Bar-Massada, A., Radeloff, V., Silva, P., 2018.
Panel on Climate Change [Field, C.B., V. Barros, T.F. Stocker, D. Qin, D.J. Dokken, K.L. Defining and mapping the wildland-urban interface in Portugal. In: Viegas, D.X. (Ed.)
Ebi, M.D, Cambridge. ed. Advance’s in Forest Fire Research 2018, Coimbra, pp. 743–749.

14
S. Oliveira, A. Gonçalves and J.L. Zêzere Science of the Total Environment 762 (2021) 143121

Pereira, S., Santos, P.P., Zêzere, J.L., Tavares, A.O., Garcia, R.A.C., Oliveira, S.C., 2020. A land- Sil, Â., Fernandes, P.M., Rodrigues, A.P., Alonso, J.M., Honrado, J.P., Perera, A., Azevedo, J.C.,
slide risk index for municipal land use planning in Portugal. Sci. Total Environ., 2019. Farmland abandonment decreases the fire regulation capacity and the fire pro-
139463 https://doi.org/10.1016/j.scitotenv.2020.139463. tection ecosystem service in mountain landscapes. Ecosyst. Serv. 36, 100908. https://
Pinto, R.M.S., Benali, A., Sá, A.C.L., Fernandes, P.M., Soares, P.M.M., Cardoso, R.M., Trigo, doi.org/10.1016/j.ecoser.2019.100908.
R.M., Pereira, J.M.C., 2016. Probabilistic fire spread forecast as a management tool in Tedim, F., Remelgado, R., Borges, C., Carvalho, S., Martins, J., 2013. Exploring the occur-
an operational setting. Springerplus 5, 1205. https://doi.org/10.1186/s40064-016- rence of mega-fires in Portugal. For. Ecol. Manag. 294, 86–96. https://doi.org/
2842-9. 10.1016/j.foreco.2012.07.031.
Poljansek, K., Marín Ferrer, M., De Groeve, T., Clark, I., 2017. Science for Disaster Risk Man- Tedim, F., Leone, V., McCaffrey, S., McGee, T.K., Coughlan, M., Correia, F.J., Magalhães, C.G.,
agement 2017: Knowing Better and Losing Less. ETH Zurich. 2020. Safety enhancement in extreme wildfire events. Extreme Wildfire Events and
Pourtaghi, Z.S., Pourghasemi, H.R., Aretano, R., Semeraro, T., 2016. Investigation of Disasters. Elsevier, pp. 91–115.
general indicators influencing on forest fire and its susceptibility modeling Tehrany, M.S., Jones, S., Shabani, F., Martínez-Álvarez, F., Tien Bui, D., 2019. A novel en-
using different data mining techniques. Ecol. Indic. 64, 72–84. https://doi.org/ semble modeling approach for the spatial prediction of tropical forest fire susceptibil-
10.1016/j.ecolind.2015.12.030. ity using LogitBoost machine learning classifier and multi-source geospatial data.
Pradhan, B., Suliman, M.D.H. Bin, Awang, M.A. Bin, 2007. Forest fire susceptibility and risk Theor. Appl. Climatol. 137, 637–653. https://doi.org/10.1007/s00704-018-2628-9.
mapping using remote sensing and geographical information systems (GIS). Disaster Teodoro, A., Amaral, A., 2019. A statistical and spatial analysis of Portuguese forest fires in
Prev. Manag. An Int. J. 16, 344–352. https://doi.org/10.1108/09653560710758297. summer 2016 considering Landsat 8 and Sentinel 2A data. Environ. - MDPI 6, 36.
Ribeiro, L.M., Viegas, D.X., Almeida, M., McGee, T.K., Pereira, M.G., Parente, J., https://doi.org/10.3390/environments6030036.
Xanthopoulos, G., Leone, V., Delogu, G.M., Hardin, H., 2020. Extreme wildfires and di- Tonini, M., Pereira, M.G., Parente, J., Vega Orozco, C., 2017. Evolution of forest fires in
sasters around the world: lessons to be learned. Extreme Wildfire Events and Disas- Portugal: from spatio-temporal point events to smoothed density maps. Nat. Hazards
ters, pp. 31–51. 85, 1489–1510. https://doi.org/10.1007/s11069-016-2637-x.
Salis, M., Laconi, M., Ager, A.A., Alcasena, F.J., Arca, B., Lozano, O., Fernandes de Oliveira, A., Tonini, M., D’Andrea, M., Biondi, G., Esposti, S.D., Trucchia, A., Fiorucci, P., 2020. A machine
Spano, D., 2016. Evaluating alternative fuel treatment strategies to reduce wildfire learning-based approach for wildfire susceptibility mapping. The case study of the Li-
losses in a Mediterranean area. For. Ecol. Manag. 368, 207–221. https://doi.org/ guria region in Italy. Geosciences, 10 https://doi.org/10.3390/geosciences10030105.
10.1016/j.foreco.2016.03.009. Trigo, R.M., Pereira, J.M.C., Pereira, M.G., Mota, B., Calado, T.J., Dacamara, C.C., Santo, F.E.,
Salis, M., Del Giudice, L., Arca, B., Ager, A.A., Alcasena, F.J., Lozano, O., Bacciu, V., Spano, D., 2006. Atmospheric conditions associated with the exceptional fire season of 2003
Duce, P., 2018. Modeling the effects of different fuel treatment mosaics on wildfire in Portugal. Int. J. Climatol. 26, 1741–1757. https://doi.org/10.1002/joc.
spread and behavior in a Mediterranean agro-pastoral area. J. Environ. Manag. 212, Turco, M., Llasat, M.C., von Hardenberg, J., Provenzale, A., 2014. Climate change impacts
490–505. https://doi.org/10.1016/j.jenvman.2018.02.020. on wildfires in a Mediterranean environment. Clim. Chang. 125, 369–380. https://
Sánchez-Benítez, A., García-Herrera, R., Barriopedro, D., Sousa, P.M., Trigo, R.M., 2018. June doi.org/10.1007/s10584-014-1183-3.
2017: the earliest European summer mega-heatwave of reanalysis period. Geophys. Turco, M., Jerez, S., Augusto, S., Tarín-Carrasco, P., Ratola, N., Jiménez-Guerrero, P., Trigo,
Res. Lett. 45, 1955–1962. https://doi.org/10.1002/2018GL077253. R.M., 2019. Climate drivers of the 2017 devastating fires in Portugal. Sci. Rep. 9,
San-Miguel-Ayanz, J., Rodrigues, M., Oliveira, S.S. de, Pacheco, C.K., Moreira, F., Duguy, B., 1–8. https://doi.org/10.1038/s41598-019-50281-2.
Camia, A., 2012. Land cover change and fire regime in the European Mediterranean UNDRO, 1979. Natural Disasters and Vulnerability Analysis, Report of Expert Group Meet-
Region, in: Vallejo, V.R., Arianoutsou, M., Moreira, F. (Eds.), Post-Fire Management ing 9–12 July 1979.
and Restoration of Southern European Forests. Springer Netherlands, pp. 21–43. Verde, J.M., 2015. Wildfire Susceptibility Modelling in Mainland Portugal. Inst. Geogr. e
doi:https://doi.org/10.1007/978-94-007-2208-8. Ordenam. do Territ. Univ, Lisboa.
San-Miguel-Ayanz, J., Moreno, J.M., Camia, A., 2013. Analysis of large fires in European Verde, J.C., Zêzere, J.L., 2010. Assessment and validation of wildfire susceptibility and haz-
Mediterranean landscapes: lessons learned and perspectives. For. Ecol. Manag. 294, ard in Portugal. Nat. Hazards Earth Syst. Sci. 10, 485–497. https://doi.org/10.5194/
11–22. https://doi.org/10.1016/j.foreco.2012.10.050. nhess-10-485-2010.
San-Miguel-Ayanz, J., Durrant, T., Boca, R., Libertà, G., Branco, A., Rigo, D. de, Ferrari, D., Vilar, L., Nieto, H., Martín, M.P., 2010. Integration of lightning- and human-caused wildfire
Maianti, P., Vivancos, T.A., Oom, D., Pfeiffer, H., Nuijten, D., Leray, T., 2019. Forest occurrence models. Hum. Ecol. Risk. Assess. 16, 340–364. https://doi.org/10.1080/
Fires in Europe, Middle East and North Africa 2018. EUR 29856 EN. doi:https://doi. 10807031003670469.
org/10.2760/1128. Yilmaz, I., 2009. Landslide susceptibility mapping using frequency ratio, logistic re-
Scott, J., Helmbrecht, D., Thompson, M.P., Calkin, D.E., Marcille, K., 2012. Probabilistic as- gression, artificial neural networks and their comparison: a case study from
sessment of wildfire hazard and municipal watershed exposure. Nat. Hazards 64, Kat landslides (Tokat-Turkey). Comput. Geosci. 35, 1125–1138. https://doi.org/
707–728. https://doi.org/10.1007/s11069-012-0265-7. 10.1016/j.cageo.2008.08.007.
Sebastián-López, A., Salvador-Civil, R., Gonzalo-Jiménez, J., San-Miguel-Ayánz, J., 2008. In-
tegration of socio-economic and environmental variables for modelling long-term
fire danger in Southern Europe. Eur. J. For. Res. 127, 149–163. https://doi.org/
10.1007/s10342-007-0191-5.

15

You might also like