Assessing the Link between Wildfires, Vulnerability, and Climate Change: Insights from the Regions of Greece

Abstract

Forests provide a wide range of ecosystem services which are important for achieving sustainable development. Anthropogenic climate change has led to the increased frequency and severity of forest fires, which imply losses of valuable ecosystem services. This paper provides a methodological framework based on Multiple Criteria Decision Aid methods for determining relative regional vulnerabilities associated with forest fires. Different notions of ex-post vulnerability are defined based on the forest area burned and the frequency of forest fires, and their relation to the regions’ area and certain socioeconomic characteristics. The climatic drivers of forest fire occurrence are explored by linking forest fires with summer and spring temperatures and precipitation, using econometric count data analysis. The methodology is applied to Greece and its administrative regions for the period 2000–2022. Ex-post vulnerability of regions to forest fires based on physical and socioeconomic characteristics is calculated, and expected changes in the frequency of fires of specific size classes conditional on the evolution of mean seasonal regional temperature and precipitation according to IPCC scenarios are predicted. Relative vulnerability estimates and the impact of specific climatic drivers on forest fires will be useful in designing policies for preserving forests as natural capital and promoting sustainability.

1. Introduction

Human-induced climate change has, according to the IPCC Sixth Assessment Report [1], increased the frequency, intensity and/or duration of extreme weather events such as droughts, wildfires, terrestrial and marine heatwaves, cyclones and floods and has caused widespread and severe loss and damage to human and natural systems. Moreover, the same report states that since the writing of the IPCC Fifth Assessment Report [2], the impacts of climate change and extreme weather events have adversely affected or caused loss and damage to human health, incomes and livelihoods, and security and inequality, indicating that climate change has been evolving over the past decade.

Abatzoglou et al. [3] provide evidence indicating that increased fire weather becomes more widespread as a function of global temperature change, which suggests that an increase in fire weather conditions is related to anthropogenic climate change. In a review of numerous scientific papers regarding the link between climate change and fire risk, Jones et al. [4] (p. 117) find that “Human-induced warming has already led to a global increase in the frequency and severity of fire weather, increasing the risks of wildfire. This signal has emerged from natural variability in many regions, including the western US and Canada, southern Europe, Scandinavia and Amazonia”. For Southern Europe and the Mediterranean, Jones et al. [4] indicate that the impact of anthropogenic climate change on fire weather extremes and the length of fire seasons started in the 1990s (see also [3]). In a recent report regarding wildfire activity across Europe, Pronto et al. [5] identify a significant increase in wildfire activity and associate this increase in wildfire risks with prolonged heatwaves, droughts, and strong or unusual wind patterns. For a general view of the impacts of fire on forests, see, e.g., Robinne [6].

In Greece, forests and woodlands cover 49.5% (6.532 million hectares) of the total land [7]. Kassomenos [8] points out that intense forest fires have been occurring with increasing frequency in Greece since the 1970s. He associates wildfire events and wildfire persistence with interactions between low- and high-pressure systems and anticyclonic conditions associated with high temperatures, low humidity, and moderate winds. Papadopoulos et al. [9] also conclude that high temperatures, which strongly desiccate the land and dry anticyclonic activity, encourage wildfires in Greece. Koutsias et al. [10] studied the long-run relationship between forest fires and weather conditions using data for the period 1894–2010. Their results indicate a statistically significant positive trend between the number of fires and air temperature—which was related to the occurrence of summer heat waves—after the mid-1970s. Sarris et al. [11] predict escalation of wildfires with global warming, based on a case study of Mt. Taygetos in southern Greece.

Forests provide a wide range of ecosystem services, which include provisioning services (e.g., food, timber, water), regulatory services (e.g., climate regulation, flood control, carbon dioxide sequestration), cultural services (e.g., recreation, tourism, spiritual and moral values) and the supporting services necessary for the production of all other ecosystem services (e.g., soil creation, photosynthesis, conservation of species) [12]. Thus, the contribution of forest ecosystem services to human well-being is significant. Furthermore, these services are important for the livelihoods of poor rural households [13]. Therefore, conserving forest ecosystems could play an important role in poverty reduction.

De Groot et al. [14] provide extensive information on the valuation of forest ecosystem services, with aggregate values for tropical forests and temperate forests, respectively, being 119,076 and 5383 Int$/hectare-year in 2020 price levels, while the value of ecosystem services for woodland and shrubland is 769 Int$/hectare-year in 2020 price levels. A recent study of the resilience of Greek forests—which are Mediterranean forests characterized, in general, by low timber productivity due to the climate and poor pasture production [15]—estimated the value of ecosystem services to be greater than 726 EUR/hectare-year in 2020 price levels with a probability of 90%, with the mean and median of the value distribution being 1262 and 1145 EUR/hectare-year in 2020 price levels, respectively, not including loss of human life and private property [16].

The discussion above indicates that climate change and the associated changes in temperature and precipitation have an impact on forest fires, and that the occurrence of forest fires implies a loss in ecosystem services which are important for human well-being at both the local and global levels. In the sustainability context, preservation of ecosystem services provided by forests requires the preservation of forests, which are regarded as natural capital. Forest fires degrade this type of natural capital and could harm the prospects for sustainability, with sustainable development defined, following Arrow et al. [17], as the non-declining productive base of the economy (defined as the sum of produced, human, and natural capital, valued at appropriate accounting prices). Thus, understanding the link between forest fires and climate change could be important in designing adaptation policies which would protect forests as natural capital and would support sustainable development. Since forest fires in each country are distributed across different administrative regions according to geographical characteristics, the associated losses in ecosystem services have a spatial distribution. Given that the loss of ecosystem services affects well-being, there is a corresponding spatial distribution of welfare losses as well.

Characterizing the spatial distribution of these losses by assessing relative regional vulnerabilities is expected to address research gaps in the approach of calculating regional vulnerabilities based on physical and socioeconomic characteristics. Such estimates could help in the design of policies to remedy these vulnerabilities in the future.

To address these gaps, the present paper provides different notions and insights into the vulnerability of a region to forest fires relative to the rest of the regions of the country, using Multiple Criteria Decision Aid (MCDA) methods. The MCDA method is used in order to determine, via various ranking procedures, the degree of relative vulnerability of regions to forest fires. Alternative vulnerability types, expressed as indices, are introduced in this paper. These vulnerability types are obtained by linking the area burned and the frequency of fires in a given region over a year to specific regional characteristics such as area, gross domestic product (GDP), population level and density, and regional GDP per capita relative to the country’s GDP per capita. This relative vulnerability concept could be useful for prioritizing adaptation policies when trade-offs emerge because of funding constraints. For example, whether regions which are more vulnerable in terms of regional income distribution should have priority over regions which are more vulnerable in terms of population density.

Although the link between fire weather and climate change has been established, as discussed in the beginning of this section, an open research issue is the quantitative prediction of the impact of climate change on the future frequency of forest fires, especially in specific locations (see, e.g., [18]). This paper seeks to address this issue by providing a framework for predicting the expected frequency of forest fires of certain size when climate change evolves according to IPCC scenarios.

In this context, the contribution of the present paper is twofold. First, it provides a methodological framework based on Multiple Criteria Decision Aid methods for determining relative regional vulnerabilities associated with the forest area burned and the frequency of forest fires that occurred during the period 2000–2022 in Greece as a whole and its administrative regions. Area burned and frequency were related to the regions’ area and certain socioeconomic characteristics to define different notions of ex-post vulnerability based on observed forest fires. Second, by linking the forest fire events during 2000–2022 with summer and spring temperatures and precipitation, and by using econometric count data analysis, the paper explores the climatic drivers of forest fires and forecasts the expected number of fires of specific size conditional on the evolution of temperature in the regions of Greece according to the IPCC representative concentration pathways (RCP) scenarios RCP4.5 [19] and RCP8.5 [20]. This can be regarded as providing insight into a forward-looking measurement of vulnerability, and the forecasts can be used to study the expected vulnerability associated with the evolution of climate change and to design adaptation policies.

Vulnerability analysis to forest fires based on historic data and predictions of the expected number of forest fires using climatic drivers and the evolution of climate change act in a complementary way in this paper. The determinants and the structure of historic vulnerabilities can be used to project future vulnerabilities based on the predicted expected frequency and size of forest fires. This would be helpful for designing adaptation policies to climate change.

The rest of this paper is organized as follows. Section 2 presents the MCDA method used and the formulation of the socioeconomic sustainability indices, the approach for modeling the climatic drivers of forest fires and forecasting them according to the IPCC scenarios, and the data used in this paper. Section 3 presents the results regarding the vulnerability indices, the estimated count data models, and the forest fire forecasts according to IPCC scenarios using a Poisson regression framework. Section 4 discusses the results and policy implications, and Section 5 concludes.

2. Materials and Methods

2.1. Regional Vulnerability Assessment and Multicriteria Methods

The decision process for assessing the vulnerability of a region to a specific hazard depends on many factors. The selection of the appropriate assessment method can therefore be considered as a Multi-Criteria Decision-Making (MCDA) problem. MCDA methods (see, e.g., [21,22]) have been extensively used in vulnerability assessment [23,24,25]. In this paper, we use the Technique of Order Preference Similarity to the Ideal Solution (TOPSIS) method, which assigns vulnerability attributes or criteria—characteristics of regional impacts of fires in this case—to regions. TOPSIS, which is one of the classic MCDA methods, can be used to make an analytical decision based on collected data reflecting a set of attributes. TOPSIS is based on finding an ideal and an anti-ideal solution and comparing the distance of each of the alternatives to them. A decision matrix is then created, and weights are associated with different attributes. Suppose, for example, we have a problem with m alternative regions, $A_i$, and n vulnerability attributes, C_j. Then, the decision matrix shown in

Table 1. The decision matrix.

Vulnerability Attributes	${\mathit{C}}_1$	${\mathit{C}}_2$	${\mathit{C}}_3$	…	${\mathit{C}}_{\mathit{n}}$
	Weights
Regions	$w_1$	$w_2$	$w_3$	…	$w_n$
$A_1$	$x_{11}$	$x_{12}$	$x_{13}$	…	$x_{1n}$
$A_2$	$x_{21}$	$x_{22}$	$x_{23}$	…	$x_{2n}$
…	…	…	…	…	…
$A_m$	$x_{m1}$	$x_{m2}$	$x_{m3}$	…	$x_{mn}$

is constructed. The weights $\left\{w_j\right\}, j=1,\dots ,n$ reflect the relative importance associated with a specific attribute, while the values of the entries $\left\{x_{ij}\right\}, i=1,\dots ,m, j=1,\dots n$ provide a numerical value of attribute $j$ to region $i$. In order to derive objective weights, we use information entropy so our method is a modified TOPSIS referred to as the entropy weight TOPSIS (EW-TOPSIS) [23,26,27].

For the case of Greece, which has thirteen regions, the following steps are taken:

• Step 1: We calculate the normalized decision matrix P with elements

$$p_{ij}=\frac{x_{ij}}{{\sum }_{i=1}^m{x}_{ij}}, j\in \left[1,\dots ,n\right].$$

The process of normalization allows us to remove the units of measurement, which makes the various features comparable to each other.

• Step 2: We calculate the weight of each criterion using Equations (1)–(3). Equation (1) represents entropy which, in information theory, is a measure of the amount of uncertainty presented by a discrete probability distribution:

$$e_j=\frac{-1}{ln\left(m\right)}{\sum }_{i=1}^m{p}_{ij}\mathrm{l}\mathrm{n}\left(p_{ij}\right).$$

(1)

Equation (2) is interpreted as the degree of diversity of the information contained in each criterion and can be calculated as

$$d_j=1-e_j, j\in \left[1,\dots ,n\right].$$

(2)

Finally, Equation (3) is the objective weights for each criterion that are given by

$$w_j=\frac{d_j}{{\sum }_{j=1}^n{d}_j}, j\in \left[1,\dots ,n\right].$$

(3)

Step 2 provides an objective way of calculating the weights so there are no subjective biases that could distort the importance of the criteria. If an element $p_{ij}$ of the normalization matrix is zero, then the corresponding term of the sum in the definition of $e_j$ will be zero since $\underset{x\to 0}{\lim }x\mathrm{l}\mathrm{n}\left(x\right)=0$.

• Step 3: We calculate the ideal ($A^{+}$) and anti-ideal ($A^{-}$) solution as

$$A^{+}=\left(p_1^{+},p_2^{+},\dots ,p_m^{+}\right)$$

$$A^{-}=\left(p_1^{-},p_2^{-},\dots ,p_m^{-}\right)$$

where

$$p_j^{+}=\left\{p_{ij},j\in J_1; p_{ij},j\in J_2\right\}$$

$$p_j^{-}=\left\{p_{ij},j\in J_1; p_{ij},j\in J_2\right\}.$$

J₁ is the set of benefit criteria, and J₂ is the set of cost criteria. This basically tells us that the ideal $A^{+}$ solution consists of the maximum of each provision criterion and the minimum of each cost criterion. In contrast, the anti-ideal solution $A^{-}$ consists of the minimum of each provision criterion and the maximum of each cost criterion.

• Step 4: We compute the weighted Euclidean distances between each alternative $A_i$ and $A^{+}$, and between each alternative $A_i$ and $A^{-}$ as

  $$d_i^{+}=\sqrt{\sum _{j=1}^n{w}_j{\left(d_{ij}^{+}\right)}^2}$$

  $$d_i^{-}=\sqrt{\sum _{j=1}^n{w}_j{\left(d_{ij}^{-}\right)}^2}$$

  where

  $$d_{ij}^{+}=p_j^{+}-p_{ij},i\in \left[1,\dots ,m\right]$$

  $$d_{ij}^{-}=p_j^{-}-p_{ij},i\in \left[1,\dots ,m\right].$$

• Step 5: We calculate the relative proximity of each alternative to the ideal solution. The relative proximity of alternative $A_i$ with respect to $A^{+}$ is defined as

$${\xi }_i=\frac{d_i^{-}}{d_i^{+}-d_i^{-}}, i\in \left[1,\dots ,m\right].$$

This relative proximity tells us how close each alternative $A_i$ is to the ideal $A^{+}$. By ranking the alternatives $A_i$ from largest to smallest based on relative closeness, we can see which alternative is closest to the ideal solution and which is farthest away.

2.2. Modelling the Arrival of Forest Fires

Poisson regression is generally used to model the arrival of events such as forest fires. More specifically, in the context of Poisson regression analysis, the dependent variable or outcome $Y$ is the annual number of fires that burned an area in excess of a set threshold, which is a non-negative integer. In Poisson regression, the underlying probability distribution of the dependent $Y$ variable is assumed to be Poisson, which is a discrete distribution commonly used to model measurements of events occurring during a time interval. The dependent variable $Y$ can take any non-negative integer value. Poisson distribution has only one parameter $\mu$ equal to the mean value and the variance. The distribution is given by the following relationship:

$$\mathrm{P}\mathrm{r}\left(Y=y;\mu \right)=\frac{{\mu }^y{\mathrm{e}}^{-\mu }}{y!},y=0,1,2.$$

The aim of the Poisson regression (PR), in our case, is to adapt the observed data on the annual occurrence of forest fires to a regression equation which models the expected value of the dependent variable $Y$. This expected value $\mathrm{E}\left(Y\right)$ is a function of an $X$ vector of explanatory variables $X_1,X_2,...,X_k$ and regression parameters $\beta$. In this case,

$$\mathrm{E}\left(Y\right)=\mathrm{e}\mathrm{x}\mathrm{p}\left({{\beta }_0+\beta }_1{X}_1+\dots +{\beta }_k{X}_k\right),$$

(4)

where $\mathrm{E}\left(Y\right)$ is the number of fires expected in a year for given values of the explanatory variables $X_1,X_2,...X_k$. The calendar year is used, given that in Greece most fires occur during the hot and dry period, i.e., mostly from May to October. From (4), a unit change in $X_i$ will change the expected number of forest fires by $\mathrm{e}\mathrm{x}\mathrm{p}{\beta }_i$ keeping all other explanatory variables constant.

As it is well known, when the variance of the dependent variable of the dataset statistically exceeds the corresponding mean, that is, when overdispersion is present, then the negative binomial regression is regarded as the appropriate estimation method [28]. In this model the expected value $\mathrm{E}\left(Y\right)$ is defined as

$$\mathrm{E}\left(\tilde{Y}\right)=\mathrm{e}\mathrm{x}\mathrm{p}\left({{\beta }_0+\beta }_1{X}_1+\dots +{\beta }_k{X}_k\right)\delta ,$$

(5)

with $\mathrm{E}\left(\delta \right)=1$ and $\mathrm{E}\left(\tilde{Y}\right)=\mathrm{E}\left(Y\right)\mathrm{E}\left(\delta \right)=\mathrm{E}\left(Y\right).$ When underdispersion is present, the generalized Poisson (GPR) regression can be used [29]. When there is an excess number of zeros for the dependent variable, which is a possible case if we use annual data for large regional wildfires, then the zero-inflated Poisson regression (ZIPR) or zero-inflated negative binomial (ZINB) regression can be used (see [28], Chapter 9.6, and the STATA manual for details).

2.3. Regional Forest Fires in Greece

The total number of forest fires in Greece for the period 2000–2022 was 233,116 with the annual distribution shown in Figure 1. The fire events were grouped into five categories according to the size of the burned area: less than 10 hectares (ha); between 10 and 50 ha; between 50 and 100 ha; between 100 and 1000 ha; and more than 1000 ha. This structure provides a basis for examining vulnerability to different types of fires with respect to area burned and frequency of fires. The burned areas in Greece as a whole and in three selected regions are presented in Figure 2. The three regions were selected as examples of high vulnerability with respect to population density (Attika), high socioeconomic vulnerability (Eastern Macedonia and Thrace), and size of area burned by fires in a single year (Western Greece). (Graphs for the other ten regions of Greece are available from the authors upon request).

Figure 3 is the same as Figure 2 except for the number of fire events. Most of the fires are small, while the large mega-fires are just a small fraction of the total number of fires.

This mismatch is emphasized in Figure 4, which is a combination of Figure 2 and Figure 3 for the whole of Greece. The dominant color in the figure is light orange, which represents the area burned by the fires in the largest category. This contrasts with the small number of these fires, which is represented by the light orange part of the x-axis that is measured by the difference between the light orange arrow and the dark orange arrow on the x-axis.

The data used to produce Figure 4 for Greece as a whole, as well as for each Greek region, were used to define nine fire characteristics:

• The number of fires in each region that burned: between 10 and 50 ha; between 50 and 100 ha; between 100 and 1000 ha; above 1000 ha (four attributes);
• The total area burned for each category (four attributes);
• The total area burned for the whole region (including small fires that burned between 0 and 0.1 km²) (one attribute).

There is a direct correspondence between area burned and damages to society. These damages include the loss of ecosystem services, as stated in the introduction, but also include health effects from exposure to wildfire smoke which contains carbon dioxide and pollutants such as carbon monoxide, nitrogen, and particulate matter including that which is smaller than 2.5 μm in diameter (PM_2.5). Worldwide, wildfire smoke is estimated to kill 339,000 people a year, mostly in Asia and sub-Saharan Africa [30]. In a recent paper, Gould et al. [31] report a statistically significant association between wildfire PM2.5 and respiratory hospitalization and respiratory emergency department visits. Furthermore, evidence suggests negative mental health outcomes during and after wildfires [32]. Thus, vulnerability measured in terms of frequency and size of forest fires is directly associated with vulnerability in terms of economic damages, with economic damages defined in terms of economic valuation of both ecosystem services (e.g., [33]) and related health effects (e.g., [34]).

3. Results

3.1. Vulnerability of Greek Regions to Forest Fires

Vulnerability is a metric that can be defined in a multidimensional context. In physical terms, vulnerability for a region can be strictly defined as a function of the frequency and the size of fires. However, since forest fires induce damages which can be expressed in economic terms, vulnerability can be interpreted from the point of view of economic damages. In the same way, a social interpretation is possible if damages are associated with, for example, the regional distribution of GDP per capita.

Extending the concept of vulnerability means that the $x_{ij}$ attributes in Table 1 with $m=13, n=9$ for the case of Greece can be defined in alternative ways. We therefore define the following vulnerability types:

a.

(i) Unadjusted physical vulnerability $\left(V_{P1}\right)$, and (ii) adjusted regional area physical vulnerability $\left(V_{P2}\right)$. $V_{P1}$ is the number of fires or the area burned in a given region, and $V_{P2}$ is the same quantity divided by the area of the region. The former quantifies aggregate losses in ecosystem services, while the latter captures losses per regional unit area. Therefore, large regions are relatively less vulnerable in terms of this metric for a similar number of fires or area burned.

b.

Economic vulnerability, $\left(V_E\right)$, defined per regional GDP and obtained by dividing the attribute $x_{ij}$ by the GDP of the region. Regions with relatively higher GDP are less vulnerable in terms of this metric.

c.

Social vulnerability, $\left(V_S\right)$, defined per inhabitant of the region and obtained by dividing the attribute $x_{ij}$ by the population of the region.

d.

Socioeconomic vulnerability, $\left(V_{SE}\right)$, which is obtained by calculating the quantity

$$V_{CE}=\frac{x_{ij}}{y_j}{\left(\frac{y}{y_j}\right)}^{\eta },$$

where $x_{ij}$ is the attribute $i$—number of fires or area burned—for region $j$; $y_j$ is the per capita GDP for region $j$; $y$ is the per capita GDP for Greece; and $\eta$ is the elasticity of the marginal utility of income which reflects society’s aversion to income inequality. Higher values for $\eta$ indicate stronger societal preference for equal income distribution. Typical values for $\eta$ range from 1 to 3. The distributional coefficient ${\left(\frac{y}{y_j}\right)}^{\eta }$ indicates that regions with GDP per capita below the national average are relatively more vulnerable than regions with GDP per capita above the national average for the same $\frac{x_{ij}}{y_j}$ ratio. For $\eta =0$, vulnerability depends only on the relation between the attribute and the regional GDP per capita and is not affected by the relation between the regional GDP per capita and the overall GDP per capita of the country.

e.

Vulnerability with respect to regional population density, $\left(V_{PD}\right)$. Vulnerability per capita that reflects damages from forest fires per person might not provide the whole picture with respect to the effects of the wildfires. This is because, apart from provisioning services and some cultural services, important regulating and supporting services have public good characteristics, and their loss affects most or even the whole regional population. Furthermore, the smoke from wildfires also has public good characteristics and, in densely populated regions, affects relatively more people and therefore generates higher aggregate damages (which could become even higher if the diffusion of smoke from wildfires in neighboring regions is taken into account). Thus, a region with high population density could experience relatively higher impacts compared to regions with low population density. $V_{PD}$ is defined then as the attribute $x_{ij}$ multiplied by the regional population density (inhabitants per km²).

Forest areas that were burned can be directly associated with quantifiable damages in ecosystem services. Thus, vulnerability attributes associated with area burned can be directly mapped to economic damages, while vulnerability attributes associated with forest fire frequency can be associated with the frequency that these economic damages occur. Thus, the vulnerability types defined above can be further interpreted from the point of view of economic damages.

Physical vulnerability can be associated with regional aggregate damages or damages per square ha. Economic vulnerability reflects damages per unit of regional GDP. High economic vulnerability will indicate regions which are relatively more dependent on ecosystem services and thus more vulnerable to climate change since, as shown in Section 3.3, expected forest fires will increase under both IPCC scenarios. Social vulnerability reflects damages per person and distinguishes, for an approximately similar number of fires, between heavily populated and lightly populated regions. Socioeconomic vulnerability focuses on the impact of forest fires between relatively rich and poor regions in terms of GDP per capita and attitudes toward income equality. High socioeconomic vulnerability for a region suggests that a relatively poor region has received a relatively higher share of damages from forest fires. Finally, vulnerability with respect to population density seeks to capture aggregate environmental damages to the whole population of a region.

Combining the nine attributes for the thirteen regions of Greece (which are shown in Figure 5), the six vulnerability types were calculated.

The data for forest fires for the period 2000–2022 were used, and two sets of vulnerability indices were estimated, one covering the period 2000–2010 and the other the period 2011–2022. Results are shown in

Table 2. Results of the six vulnerability types for each of the thirteen regions of Greece.

	Vulnerability Types 1
	Physical VP2		Economic VE		Social VS		Socioeconomic VSE								Density VPD		Physical VP1
	2000–2010	2011–2022	2000–2010	2011–2022	2000–2010	2011–2022	η = 0		η = 1		η = 2		η = 3		2000–2010	2011–2022	2000–2010	2011–2022
Regions							2000–2010	2011–2022	2000–2010	2011–2022	2000–2010	2011–2022	2000–2010	2011–2022
Eastern Maced & Thrace	0.080	0.152	0.185	0.462	0.148	0.331	0.177	0.498	0.190	0.658	0.202	0.848	0.214	0.929	0.019	0.025	0.119	0.370
Central Maced	0.002	0.032	0.011	0.044	0.002	0.030	0.049	0.174	0.052	0.214	0.054	0.262	0.054	0.273	0.017	0.027	0.030	0.144
Western Maced	0.100	0.028	0.252	0.172	0.245	0.158	0.113	0.048	0.103	0.049	0.094	0.056	0.085	0.052	0.006	0.001	0.086	0.055
Epirus	0.145	0.038	0.356	0.204	0.289	0.148	0.200	0.096	0.215	0.138	0.230	0.190	0.245	0.217	0.017	0.003	0.128	0.073
Thessaly	0.147	0.047	0.248	0.152	0.199	0.111	0.315	0.179	0.318	0.222	0.320	0.273	0.321	0.290	0.049	0.011	0.211	0.144
Central Greece	0.273	0.381	0.439	0.889	0.526	0.917	0.419	0.919	0.337	0.892	0.272	0.864	0.219	0.769	0.068	0.053	0.427	0.945
Ionian Islands	0.454	0.361	0.273	0.378	0.344	0.332	0.087	0.117	0.068	0.123	0.053	0.135	0.041	0.124	0.040	0.019	0.094	0.116
Western Greece	0.667	0.139	0.744	0.289	0.764	0.210	0.857	0.337	0.868	0.424	0.878	0.529	0.887	0.569	0.265	0.027	0.864	0.266
Peloponnese	0.458	0.240	0.800	0.663	0.797	0.578	0.830	0.681	0.792	0.737	0.755	0.794	0.718	0.752	0.130	0.036	0.747	0.612
Attica	0.519	0.855	0.010	0.047	0.023	0.064	0.110	0.385	0.059	0.250	0.031	0.161	0.016	0.088	1.000	1.000	0.184	0.567
North Aegean	0.489	0.289	0.830	0.620	0.631	0.471	0.324	0.215	0.325	0.277	0.325	0.358	0.326	0.398	0.046	0.014	0.194	0.166
South Aegean	0.241	0.153	0.150	0.206	0.287	0.222	0.077	0.102	0.051	0.086	0.033	0.075	0.021	0.057	0.031	0.013	0.122	0.118
Crete	0.038	0.062	0.040	0.102	0.043	0.079	0.029	0.100	0.027	0.115	0.025	0.135	0.023	0.134	0.009	0.011	0.025	0.088

¹ The TOPSIS algorithm for calculating the relative vulnerabilities was written in Mathematica 12 and is available from the authors upon request.

• Legend: Most vulnerable; 2nd most; vulnerable; 3rd most vulnerable; Least vulnerable.

.

The relative vulnerabilities shown in Table 2 suggest that:

• Attica, which is the smallest region in terms of area but the most heavily populated both in terms of absolute number and density and has the highest GDP per capita, is the most vulnerable region in terms of density, $V_{PD}$, and per unit area, $V_{P2}$. This exemplifies the public good aspect of damages from wildfires.
• Peloponnese is highly vulnerable in almost every vulnerability type except $V_{P2}$. This indicates that the area-adjusted indices should be interpreted carefully. This is because although the region is vulnerable in most of the indices, it is not vulnerable in the area-adjusted, $V_{P2}$, because it is a large region.
• Central Greece shows a similar pattern to Peloponnese.
• Eastern Macedonia and Thrace shows higher socioeconomic vulnerability at high preferences for equal distribution ($\eta =2, \eta =3$). This suggests that damages in this region disproportionally affect relatively poor communities.
• For Central Greece, the vulnerability of almost all types increased during the period 2011–2022 relative to the previous period.

The identified regional vulnerability differences stem mainly from climatic and socioeconomic regional characteristics.

Table 3. Regional characteristics (averages 2000–2022).

Characteristic	Region
	Attica	Central Greece	Central Macedonia	Crete	E. Macedonia & Thrace	Epirus	Ionian Islands	North Aegean	Peloponnese	South Aegean	Thessaly	W. Greece	W. Macedonia	Greece
Summer temp (°C)	25.6	23.5	24.20	24.1	22.7	21.8	23.7	24.3	23.6	25.0	23.9	23.6	21.3	23.3
Summer prec. (mm)	35.8	92.5	139.5	15.8	143.5	110.4	32.3	9.8	59.1	3.1	112.2	82.9	120.0	102.4
Spring prec. (mm)	111.6	200.0	206.2	102.0	228.1	308.7	173.8	154.9	168.4	92.6	209.9	251.5	210.6	209.6
High forests & forest lands cover (%) 1	50.9	74.75	63.57	50.06	65.76	54.55	55.25	64.78	55.88	61.48	82.07	58.64	63.57	61.48
N100 2	2.13	3.87	0.83	0.91	1.52	0.83	2.08	0.86	4.21	0.69	1.48	1.35	0.87	20.65
N50	3.13	5.91	1.65	1.61	2.22	1.61	3.96	1.13	6.70	1.22	2.82	2.48	1.30	34.13
N10	8.61	17.04	6.0	5.39	9.22	7.13	13.8	3.0	18.0	3.91	10.78	10.04	6.30	114.04
Area burned N100 (ha)	2105.41	3859.67	416.59	195.28	1233.99	490.08	1148.48	1228.67	4213.56	821.41	887.29	4053.59	355.65	20,276.89
Area burned N50 (ha)	2166.76	3992.73	464.73	236.54	1279.13	539.22	1268.03	1246.76	4369.71	851.41	965.71	4124.59	383.91	21,120.34
Area burned N10 (ha)	2270.35	4213.05	549.03	302.01	1416.82	637.21	1451.15	1280.71	4584.71	904.05	1104.68	4266.94	469.92	22,619.53
GDP per capita (€ current prices)	23.225	16.225	13.69	14.958	12.373	12.466	16.289	12.988	14.28	18.786	13.168	12.995	15.832	17.368
Area (ha)	38,080	155,490	191,410	83,360	141,570	140,370	23,070	38,360	154,990	52,860	92,030	113,500	94,510	1,319,510

¹ This is the proportional cover of the total region’s area. It includes high forests, evergreen broadleaves, grazelands, and infertile bare lands. See [35]. ² N100, N50, and N10 correspond to forest fires that burned more than 100, 50, and 10 ha per year respectively.

summarizes regional characteristics expressed as averages for the period 2000–2022. Our results indicate that relatively highly/MORE vulnerable regions tend to combine summer temperatures above the country average with spring and summer precipitation below the country average. These climatic drivers contribute to the emergence of forest fires, given the relatively high coverage of regions by high forests and forest lands. The importance of these climatic drivers is confirmed by the regression analysis presented in Section 3.3. Regions which are large in terms of area also tend to be highly vulnerable, because in large regions, large fires and multiple fires at different locations could emerge. Attica is an exception because its area is not that large but, in this case, the strong climatic drivers are combined with substantial forest coverage and high population density that increases the region’s vulnerability. In terms of socioeconomics, relatively low regional GDP per capita associated with large fires with relatively high frequency increase regional vulnerability.

To examine the sensitivity of our vulnerability results to changes in attributes, we performed sensitivity analysis by changing the attitudes by ±5% and ±10%. There were no significant changes in the relative vulnerabilities, which suggests that our algorithm provides robust results.

In order to explore potential links between regional vulnerabilities based on observed data and regional characteristics, we considered the vulnerabilities estimated in Table 2 for the periods 2000–2010 and 2011–2022 as functions of average regional population density (avg_dens) and/or average regional GDP (avg_gdp). Using panel data with the two periods of time, we performed regression analysis. The results are shown in

Table 4. Relative vulnerability versus population density and GDP.

Dependent Variable	VP1 (BE)	VP1 (PA)	VSE, η = 1 (PA)	VSE, η = 1 (RE) 1
Constant	0.1819487	0.1820646	0.7478916	0.7530574
$P>\left|z\right|$	0.03	0.0	0.010	0.0014
avg_dens	0.0004962	0.0004953
$P>\left|z\right|$	0.013	0.01
avg_gdp			−0.00003	−0.0000303
$P>\left|z\right|$			0.108	0.124
Wald χ2(1)	8.68	10.22	2.58	2.36
P > χ2	0.013	0.0014	0.1083	0.1241

BE: between estimator for panel data. PA: population-averaged estimator for panel data. ¹ The Breusch and Pagan Lagrangian multiplier test for random effects rejects the null hypothesis that the variance across regions is zero. The random effects estimator is appropriate (for estimation details; see [36]).

.

The results indicate that the relative physical vulnerability increases with average population density (avg_dens) which emphasizes the public good aspect of forest fires since it suggests that the damages from forest fires are diffused to a large population in a given area. The socioeconomic relative vulnerability is shown to decrease with average regional GDP (avg_gdp), but the result is not strongly statistically significant.

Given the health effects from forest fires, smoke might affect regional vulnerabilities. It would be interesting to include excess mortality measures or excess hospital entries during the period of forest fires in the regression analysis. This is an interesting area for further research, which is currently beyond the scope of the paper due to data limitations.

3.2. Forest Fires and Fire Weather Index

The analysis in the previous section used historical data to explore the relative vulnerability of Greek regions in terms of frequency and size of forest fires. In this section, the link between forest fires and climate change is explored.

Table 5. Pearson’s correlation between number of days per year with FWI over a threshold and number of wildfires per year in Greece.

Fires/Year That Burned…	More Than 100 ha		More Than 50 ha		More Than 10 ha
Days per Year with…	FWI > 30	FWI > 45	FWI > 30	FWI > 45	FWI > 30	FWI > 45
E. Macedonia & Thrace	0.322	0.270	0.365	0.259	0.370	0.222
Central Macedonia	0.594	0.506	0.653	0.710	0.681	0.686
Western Macedonia	0.292	0.606	0.344	0.622	0.374	0.509
Epirus	0.457	0.600	0.530	0.474	0.624	0.505
Thessaly	0.255	0.408	0.329	0.450	0.478	0.469
North Aegean	0.204	0.082	0.221	0.160	0.073	0.058
South Aegean	0.407	0.460	0.372	0.518	0.186	0.212
Central Greece	0.383	0.570	0.476	0.524	0.501	0.459
Western Greece	0.434	0.652	0.505	0.572	0.607	0.494
Peloponnese	0.594	0.563	0.690	0.688	0.736	0.741
Ionian Islands	0.208	0.068	0.414	0.061	0.375	0.308
Crete	0.024	0.173	0.061	0.232	0.240	0.364
Attica	0.329	0.413	0.510	0.620	0.342	0.491
Greece	0.561	0.772	0.652	0.709	0.686	0.552

The warmer colors denote higher correlations, starting at 0.3 (pale yellow).

presents the simple correlation between the number of fires in each category and the Canadian fire weather index (FWI).

The Canadian FWI [37,38] is a measure of forest fire intensity and can be used for the assessment of fire occurrence risk, particularly for large fires [38,39]. The input for its calculation consists of four meteorological input parameters, namely the recent rainfall, temperature, wind speed and relative humidity, as well as information such as the date and latitude [37]. Although it was produced for use in Canada, it has been widely used by scientists across the world for decades, including in the Mediterranean region [40,41,42,43]. FWI data for the period 2000–2022 are derived from the Copernicus Climate Change Service [44]. Table 5 shows the correlation between the number of fires and the FWI for the 13 Greek regions and Greece as a whole for the period 2000–2022.

Table 5 suggests some strong correlations between the FWI and the frequency of fires of a certain size. In order to thoroughly explore the link between forest fires and climate change, a quantitative framework based on the Poisson distribution in which forest fires are treated as arrivals of events was adopted. Climate change is captured by the evolution of average summer and spring regional temperature and precipitation during the period 2000–2022, which are used as explanatory variables. Both summer and spring regional temperature and precipitation averages are calculated from the ERA-5 gridded reanalysis data [45]. The objective here is to model the arrival of forest fires in Greece from relatively small to big, so we focus on fires that burned more than 10, 50, and 100 ha.

3.3. Climate Change and Expected Forest Fires: Greece

A description of the variables for the aggregate data for Greece is shown in

Table 6. Description of variables (2000–2022).

Variable	Mean	Std. Dev.	Min	Max
N100	20.6	19.175	1.0	78.0
N50	34.1	27.501	6.0	120.0
N10	114.0	64.056	36.0	321.0
tsum	23.29	0.7067	22.24	25.014
psum	102.42	42.52	37.518	224.436
tspr	12.26	0.778	10.89	14.296
pspr	209.55	34.482	147.659	268.986

. N100, N50, and N10 correspond to forest fires that burned more than 100, 50, and 10 ha per year respectively. The explanatory variables for summer and spring were denoted respectively as $tsum, tspr$, for temperature in degrees Celsius, and $psum, pspr$ for precipitation in mm.

Table 7. Regression results for fires that burned more than 10, 50, and 100 ha in Greece ¹.

Dependent Variable	N10 (GPR)	N50 (GPR)	N100 (PR)
tsum	0.2049	0.4941	0.2227
$P>\left|z\right|$	0.071	0.023	0.0
psum	−0.0065	−0.006	−0.0086
$P>\left|z\right|$	0.0	0.015	0.0
pspr	−0.0033	−0.0023	−0.0067
$P>\left|z\right|$	0.096	0.466	0.0
Constant	1.2315	−6.99
$P>\left|z\right|$	0.659	0.187
Wald χ2(3)	35.92	17.48	4844.8
Pseudo $R^2$	0.46	0.64	0.30
MNF Est.	105.4	30.3	18.1
MNF data (Table 6)	114	34.1	20.6
2031–2060 RCP 4.5	190.387 (67.01%)	72.5404 (112.73%)	39.6955 (92.7%)
2071–2100 RCP 4.5	218.535 (91.7%)	97.783 (186.75%)	47.2403 (129.32%)
2031–2060 RCP 8.5	210.161 (84.35%)	89.9841 (163.88%)	44.9298 (118.11%)
2071–2100 RCP 8.5	394.921 (246.42%)	344.548 (910.41%)	95.4483 (363.34%)

¹ All estimations were obtained using STATA 18. GPR: generalized Poisson regression to allow for underdispersion. The constants were not significant for the N10 and N50 regression, but they are reported because the GPR estimation method used, because of underdispersion, did not allow for intercept suppression. For the N100 regression, the constant was highly insignificant and was suppressed. PR: Poisson regression. The Pseudo $R^2$ was estimated as $\mathrm{P}\mathrm{s}\mathrm{e}\mathrm{u}\mathrm{d}\mathrm{o} R^2=1-\frac{\mathrm{l}\mathrm{n} L_{fit} }{\mathrm{l}\mathrm{n} L_0}$ where $\mathrm{l}\mathrm{n} L_0$ is the log likelihood of an intercept-only model, and $\mathrm{l}\mathrm{n} L_{fit}$ is the log likelihood of the fitted model. In discrete models such as Poisson, maximum Pseudo $R^2<1$ [36]. MNF Est: mean number of fires per year estimated for the regression equation. MNF data: mean number of fires per year observed from the data; 2031–2060/2071–2100 RCP4.5/RCP8.5: predicted number of fires using the estimated regression (based on Equations (4) and (5)) for the average values of the explanatory variables for 2031–2060 and 2071–2100 under the RCP4.5 and RCP8.5 scenarios relative to the reference period (1971–2000) interpolated in the locations of the thirteen Greek regions. The proportional change in the predicted number of fires relative to the values of MNF data is shown in parentheses.

presents the results of the regression analysis for fires that burned more than 10, 50, and 100 ha. The data for observed forest fires are connected to the climate data. Subsequently, the established relationships are used to project the change in event occurrence for the three fire classes in the future. The climatic projections were based on an ensemble of climate simulations; specifically, seven combinations of global circulation models (GCMs) and regional climate models (RCMs) were used. All the simulations participated in the EURO-CORDEX experiment [46] and used the EUR-11 domain, which has a spatial resolution of 0.11 × 0.11 (approx. 12.5 km × 12.5 km). The simulations were used to produce data for the climate change in two future periods, the near future (2031–2060) and the far future (2071–2100) under the RCP4.5 and RCP8.5 greenhouse gases concentration scenarios, relative to the reference period (1971–2000) (A list of the simulation models along with a website providing detailed information is provided in Appendix A).

The RCP scenarios were the main source of input on greenhouse gas (GHG) concentrations for the IPCC Fifth Assessment Report [2]. The report states that RCP4.5 is considered to be “intermediate” while RCP8.5 is considered a “very high GHG emissions scenario”. However, Schwalm et al. [47] found that the RCP8.5 scenario was at the moment in close agreement with GHG emissions and that it was the best match with the government-stated policies for the following years.

The results suggest that mean summer temperature is positively related to the arrival of wildfires of any category, while both the summer and the spring precipitation are negatively related to the number of fires. The spring mean temperature did not seem to have any statistically significant effect on the arrival of wildfires. Also, a dummy variable used to capture the impact of the new legal framework about forests introduced in 2014 was highly insignificant.

The predictions of expected forest fires in all categories indicate an increase relative to the mean number of fires during the sample period (2000–2022) for both the RCP4.5 and the RCP8.5 scenarios. The highest percentage increase is recorded for fires that burned more than 50 ha, followed by fires that burned more than 100 ha.

To account for heterogeneities across regions, we analyzed the balanced panel of all thirteen regions with $T=23$ years (2000–2022), $n=13$ regions, and $N=n\times T=299$ observations in total, with dependent variable N100 to concentrate on relatively big forest fires.

Figure 6 provides a scatter plot of the panel data that shows the number of fires versus temperatures and precipitation along with a fitted line. The results suggest a positive relation between number of fires and temperature for summer, and a negative relation between number of fires and precipitation. The relationship between forest fires and spring temperature is ambiguous.

For the panel data, fixed-effects estimation was used to allow for arbitrary correlation between unobserved effects, which may reflect region-specific characteristics and the observed climate variables [48]. In this sense, the fixed-effects model controls for all time-invariant differences between regions, so the estimated coefficients of the fixed-effects models cannot be biased because of omitted time-invariant characteristics [49]. The pooled data were also used with pooled feasible generalized least-squares estimation denoted by the PA estimator. In the panel data the ratio of the standard deviation over the mean for the outcome variable N100 was $\left(\frac{2.629}{1.589}\right)=1.65$. In order to consider overdispersion effects, a negative binomial regression was used. In addition, the panel data contain a large number of zeros (45.8%) which represent cases of zero fires. To allow for this, we provide estimates of the pooled data using ZINB.

Estimation results for the aggregate data and the panel data are shown in

Table 8. Estimation results ¹ for Greece, panel data (dependent variable is N100).

Variable	Panel Data
	Fixed Effects 2 13 Regions	PA Estimator 3 13 Regions	ZINB 4 13 Regions	Fixed Effects 2 9 Regions
tsum	0.3839	0.0544	0.0537	0.5052
$P>\left|z\right|$	0.0	0.0	0.0	0.0
psum	−0.0071	−0.0048	−0.0045	−0.0041
$P>\left|z\right|$	0.01	0.007	0.016	0.063
pspr	−0.0061	−0.0027	−0.0027	−0.0061
$P>\left|z\right|$	0.0	0.003	0.05	0.001
Constant	−7.2102			−9.9985
$P>\left|z\right|$	0.003			0.0
Wald χ2(3)	61.45	12.95	37.81	23.0
$\mathrm{l}\mathrm{n}\alpha$			0.338
$P>\left|z\right|$			0.049
MNF Est	1.23	1.54	1.56	1.70
MNF Data	1.59	1.59	1.59	1.84
2031–2060 RCP4.5	1.846 (16.09%)	1.73 (9.03%)	1.74 (9.61%)	2.6 (41.22%)
2071–2100 RCP4.5	2.404 (51.18%)	1.82 (14.4%)	1.83 (14.95%)	3.64 (97.95%)
2031–2060 RCP8.5	2.231 (40.29%)	1.79 (12.8%)	1.80 (13.35%)	3.31 (79.99%)
2071–2100 RCP8.5	6.951 (337.162%)	2.25 (41.31%)	2.25 (41.28%)	13.67 (643.17%)

z values in parentheses. The MNF Est and MNF Data values represent number of fires that burned more than 100 ha per region and year. The IPCC entries have the same interpretation as in Table 7. ¹ STATA was used for all the econometric estimations. For details of the methods used see [36]. ² The H₀ for the Hausman test was rejected so the fixed-effects model is used. ³ Correlation: nonstationary, Semi-robust standard errors. ⁴ Inflation model: probit. Inflation variable: tspr.

. Results are presented for the panel of all thirteen regions and for a panel of nine regions with a low proportion of zero N100 fires (less than 65%). The results suggest that the main drivers of the forest fires that burned more than 100 ha are the average summer temperature and the average precipitation in the summer and spring. The impact of the spring precipitation is interesting since it suggests that spring precipitation can be used as an early warning signal in policy design to prevent wildfires.

The intercept of the fixed-effects model can be interpreted as the average of the individual regional effects. The term is statistically significant, which indicates significant regional effects.

3.4. Climate Change and Expected Forest Fires: Greek Regions

The results of the previous section suggest that the forest fires have a statistically significant relationship with summer temperature and with spring and summer precipitation. They also suggest that individual regional effects are significant. To explore in more detail the heterogeneity among regions, individual Poisson or negative binomial regressions were estimated for the nine regions with a low proportion of zero N100 fire counts. The results are summarized in

Table 9. Regression results for the nine significant Greek regions.

Region	Attica	Central Greece	Central Macedonia	Crete	E. Macedonia & Thrace	North Aegean	Peloponnese	Western Greece	Western Macedonia
Est. Meth.	GPR	PR	PR	ZIPR	ZINB	ZIPR	PR	PR	ZINB
Dep. Var.	N100	N100	N50	N50	N50	N50	N100	N50	N50
tsum	0.4878	0.1242	0.6039	0.63524	0.04691	0.07767	0.16103	0.09753	0.152441
$P>\left|z\right|$	0.035	0.0	0.05	0.01	0.0	0.001	0.0	0.0	0.018
psum		−0.00956					−0.01409		−0.029586
$P>\left|z\right|$		0.02					0.0		0.040
pspr	−0.00993	−0.00404	−0.0073	−0.00687		−0.01198	−0.00998	−0.00574
$P>\left|z\right|$	0.029	0.095	0.077	0.111		0.01	0.0	0.012
Constant	−10.746		−12.77
$P>\left|z\right|$	0.072		0.014
Wald ${\chi }^2$	8.92	210.4		23.28		12.3	274.52	61.58	6.31
Prob > ${\chi }^2$	0.011	0.0		0.0		0.0021	0.0	0.0	0.047
Pseudo $R^2$	0.11	0.16	0.14	0.10		0.19	0.20	0.07	0.49
LR ${\chi }^2$			11.57		15.44
Prob > ${\chi }^2$			0.0031		0.001
MNF Est.	2.13	3.4	1.4	2.28	2.21	1.47	3.59	2.35	0.74
MNF Smpl.	1.91	3.8	1.6	2.04	2.09	1.13	4.21	2.48	1.3
2031–2060 RCP4.5	2.78 (45.73%)	5.39 (41.82%)	2.73 (70.45%)	1.83 (−10.54%)	3.08 (47.43%)	1.303 (15.34%)	5.98 (42.16%)	2.42 (−2.46%)	1.87 (43.54%)
2071–2100 RCP4.5	4.02 (110.48%)	6.08 (59.9%)	4.1 (156.15%)	2 (−1.95%)	3.16 (51.29%)	1.53 (35.56%)	7.61 (80.93%)	2.75 (11.08%)	1.96 (50.84%)
2031–2060 RCP8.5	3.71 (94.29%)	6.04 (59.06%)	3.48 (117.45%)	1.97 (−3.61%)	3.15 (50.58%)	1.5 (32.52%)	7.33 (74.1%)	2.72 (9.61%)	1.8 (38.23%)
2071–2100 RCP8.5	12.82 (571.1%)	9.5 (149.85%)	18.47 (1054.61%)	2.69 (31.93%)	3.55 (69.79%)	1.97 (74.13%)	15.1 (258.65%)	4.12 (66.18%)	4.11 (216.5%)

.

The results confirm that the main climatic drivers are summer temperature and summer and spring precipitation. A dummy variable used to capture the impact of the new legal framework about forests introduced in 2014 was highly insignificant.

Based on the results of Table 9, for the period 2031–2060 and under RCP4.5 scenario, for 6 out of the 9 regions, namely Attica, Central Greece, Central Macedonia, Eastern Macedonia and Thrace, Peloponnese, and Western Macedonia, the average expected number of fires annually are expected to increase by more than 40% relative to the reference period (1971–2000). Although these are approximate estimates, they suggest that the number of fires that will burn more than 50 or 100 ha is expected to increase along with the evolution of climate change. This will most likely increase the vulnerability of the regions to forest fires. The results are more pessimistic for the second future period (1971–2100) and for the “more aggressive”—in terms of emission and temperature anomaly—RCP8.5 scenario.

4. Discussion

This paper studied forest fires in Greece in two different contexts. The first is the study of vulnerability using historical data on forest fires. The second is the use of the same historical data and count regression techniques to link climate change in Greece with the emergence of forest fires and make predictions for the future using IPCC scenarios.

Vulnerability of forest to fires and its relation to climate change has been assessed with regard to a number of different factors. The combination of climatic data with the concepts of sensitivity of forest to fires related to flammable vegetation, exposure related to FWI, and adaptive capacity related to fire prevention policies has been used to provide a vulnerability index in qualitative terms from none to very high (e.g., [50,51]). Then the vulnerability potential of forests in relation to climate is determined on a scale ranging from limited to very high. Topographic, meteorological, vegetation, and human-made factors have been used to derive vulnerability indices by developing and training models on previous fire events (e.g., [52]). Vulnerability related to ecological factors (e.g., [53]) has been determined on a qualitative scale by combining soil erodibility, vegetation vulnerability, and water limitation.

Our approach complements and extends this literature by estimating ex-post vulnerability using a time series of data on the size and frequency of forest fires which covers a period of time long enough to capture the impact of climate change on forest fires in Greece. Since the size of forest fires is directly associated with damages due to loss of ecosystem services, our approach provides insights into the evolving impact of climate change on forest fires in Greek regions.

The different types of vulnerability developed in this paper provide information about the relative vulnerability of Greek regions per area, per inhabitant, and per euro of regional GDP. This information, in conjunction with the information about the expected size and frequency of fires—which constitute the second part of our results—could be the basis for designing forward-looking, multi-dimensional adaptation policies. In addition, our vulnerability types take into account distributional considerations and population density. This extends the potential insights of our approach to the identification of socioeconomic regional inequalities. These inequalities could point to the need for regional distributional policies associated with compensating for damages from forest fires which are diffused to the general population of the region.

For the twelve-year period from 2011 to 2022, the relatively most vulnerable regions overall were found to be Central Greece and Peloponnese, which are the easternmost and southernmost regions of continental Greece, and the ones with the most distinctly Mediterranean climate. In terms of distributional considerations, the region of Eastern Macedonia and Thrace was found to be the relatively most vulnerable region. Vulnerability differentiation according to regional per capita GDP is a factor that should be considered in the context of policy design to address regional inequality.

The population density-related vulnerability that we estimated can be interpreted as reflecting the “public bad” nature of forest fires. Fire damages to forest-related ecosystem services, especially supporting and regulating services, are non-rivalrous and non-excludable in a relatively small-sized, densely populated region like Attica, which was the most vulnerable region with respect to population density-related vulnerability. These damages are a public bad and could raise issues in the design of adaptation policies related to forest fires. When adaptation resources are limited, trade-offs may emerge when a regulatory body decides, for example, how to split these resources between regions with high density vulnerability and regions with high socioeconomic vulnerability. Our approach provides information that may assist in the efficient allocation of adaptation resources.

The vulnerability approach based on historic data and associated with damages to ecosystem services because of forest fires can be related to the more general approach of estimating damages from historical global warming [54,55]. In this context, our approach can be regarded as providing information about regional inequalities associated with the impact of historical climate change on forest fires.

The literature on the modeling of forest fire occurrence links the arrival of forest fires to natural, climatic, and anthropogenic drivers (e.g., [56,57,58,59]) using mainly count data, econometric analysis, and panel data. Our paper follows this approach which puts emphasis on climatic drivers in order to predict expected fire occurrences, using the IPCC climate change scenarios adjusted for the Greek regions.

Correlation analysis between the FWI and the number of forest fires also shows a positive correlation for some of the relatively most vulnerable regions in the easternmost and southernmost parts of Greece. This indicates a consistency between the vulnerability estimates and the FWI and is in line with results obtained by Ntinopoulos et al. [59].

Poisson and negative binomial regression analysis confirmed the scatter plot analysis and indicated that the main drivers for forest fire events for Greece as a whole were summer temperature and summer and spring precipitation. This result is in line with the finding in Ntinopoulos et al. [59] regarding the importance of precipitation as an explanatory factor for the FWI, using Poisson regression analysis. Since the FWI is related to the emergence of forest fires, our results confirmed the importance of precipitation as an explanatory factor of the forest fire events. Michetti and Pinar [58] established the importance of temperature and precipitation in explaining the number of forest fires and the area burned for Italian regions.

The use of panel data indicated heterogeneity among Greek regions regarding the occurrence of forest fire events and the impact of climate variables. Regression analysis of distinct regions confirmed the heterogeneity as well as the impact of the main climatic drivers. A similar regional heterogeneity with respect to the emergence of forest fires has been demonstrated for regions of Italy [58].

The projection of forest fires under the RCP4.5 and RCP8.5 scenarios indicated that climate change is expected to increase the number of large forest fires in Greece. There is a consensus that climate change will lead to an increase in forest fires (e.g., [3,4,5]). In this context, our approach of combining count data regression analysis with IPCC scenarios allows a quantification of this prediction for Greek regions in terms of the expected number of fires above a given size. This result provides information about the expected loss of forest ecosystem services at the regional level for each IPCC scenario, and could help in projecting future vulnerabilities, using as a starting point the regional vulnerabilities obtained from the historical data.

Predicting the expected number of fires according to IPCC scenarios is based on traditional econometric analysis. An alternative approach would be to use statistical machine learning methods where models are assessed on the basis of out-of-sample prediction using cross validation [60]. Lasso regression with the Poisson option of the STATA software, version 18 [36] could be another possible approach in which a range of data is used for training to obtain the model and another range of data is used as validation data to assess the goodness of the model. We consider this an area of further research in which daily observations would be used to generate a data set more appropriate for machine learning methods.

5. Conclusions

This paper addressed research gaps associated with the estimation of regional vulnerabilities to forest fires based on physical and socioeconomic characteristics and the estimation of changes in the expected frequency of forest fires as climate change evolves. In this respect, we used an EW-TOPSIS method to estimate regional vulnerabilities and Poisson and negative binomial regression analysis to link forest fire frequency with climatic drivers.

The most important findings of this paper include the identification of the relative vulnerability of Greek regions to forest fires that occurred during the period 2000–2022 and the use of the data from this period to predict the arrival of relatively large forest fires according to the IPCC scenarios. The importance of spring precipitation as a driver of forest fires provides an “early warning” signal that can be used in policy design. Predictions regarding future fires suggest that the regions which were relatively more vulnerable during the period 2000–2022 will remain vulnerable in the future. This result can provide some guidance for targeted regional policies in the context of climate change. Such policies could be infrastructure development as part of adaptation policies to climate change, or the creation of carbon offsets and associated voluntary carbon markets supported by carbon sequestration in protected forests in Greece. Such policies would be aimed at preserving components of natural capital related to forests and promoting sustainability.