Forest Fire Susceptibility and Risk Mapping Using Social/Infrastructural Vulnerability and Environmental Variables

Forests fires in northern Iran have always been common, but the number of forest fires has been growing over the last decade. It is believed, but not proven, that this growth can be attributed to the increasing temperatures and droughts. In general, the vulnerability to forest fire depends on infrastructural and social factors whereby the latter determine where and to what extent people and their properties are affected. In this paper, a forest fire susceptibility index and a social/infrastructural vulnerability index were developed using a machine learning (ML) method and a geographic information system multi-criteria decision making (GIS-MCDM), respectively. First, a forest fire inventory database was created from an extensive field survey and the moderate resolution imaging spectroradiometer (MODIS) thermal anomalies product for 2012 to 2017. A forest fire susceptibility map was generated using 16 environmental variables and a k-fold cross-validation (CV) approach. The infrastructural vulnerability index was derived with emphasis on different types of construction and land use, such as residential, industrial, and recreation areas. This dataset also incorporated social vulnerability indicators, e.g., population, age, gender, and family information. Then, GIS-MCDM was used to assess risk areas considering the forest fire susceptibility and the social/infrastructural vulnerability maps. As a result, most high fire susceptibility areas exhibit minor social/infrastructural vulnerability. The resulting forest fire risk map reveals that 729.61 ha, which is almost 1.14% of the study areas, is categorized in the high forest fire risk class. The methodology is transferable to other regions by localisation of the input data and the social indicators and contributes to forest fire mitigation and prevention planning.


Introduction
Forest fires are a natural hazard that significantly affects Iran's forest areas due to their widespread environmental, economic, and social impacts. Forest fires are also considered as the leading force damaging forest resources, and they happen periodically in different severities [1]. Additionally, forest fires have increased due to an increase in the global temperature, the population, and human activities in forest areas [2]. In our study area, upward trends in both the number of forest fire events and burned areas are confirmed from 2012 until 2017. Forest fires cause some permanent alterations to forest areas, such as a reduction of plant communities and biodiversity, which can accelerate the deforestation processes [3]. However, there are also some advantages of forest fires, e.g., the elimination of harmful microorganisms, fungi, insects, and herbal disease, and soil enrichment by the nutrients and Applying the artificial neural network (ANN) method for the spatial prediction of forest fire susceptibility mapping. Applying the GIS-MCDA method for the social/infrastructural vulnerability mapping. Validating the performances of the ANN method using the receiver operating characteristics (ROC) curve and the root mean square error (RMSE).
The methodologies (see Figure 1) and the experimental results are organized in the following sections. Further descriptions and discussions regarding the resulting maps can be found in the discussions and conclusion section. Validating the performances of the ANN method using the receiver operating characteristics (ROC) curve and the root mean square error (RMSE).
The methodologies (see Figure 1) and the experimental results are organized in the following sections. Further descriptions and discussions regarding the resulting maps can be found in the discussions and conclusion section.

Study Area
The case study was the forestry area of Amol County in the Mazandaran province of northern Iran ( Figure 2). Our study area has an expanse of 646 km 2 and is essential for its natural forests and as one of the common recreational centres of the country. The study area attracts tourists from all over the country. The altitude ranges from 100 m in valleys to approximately 2500 m above mean sea level in mountainous areas in the southern part. Other topographical and environmental characteristics are described in Table 1. The study area is potentially vulnerable to forest fires, which are considered as a common problem in the region. There are more than 20 villages in the study area, mostly in valleys and some located in remote forest areas. The most common activity among the population is animal husbandry. However, there are also orchards, including walnuts and apple gardens, and even some agricultural activities.

Study Area
The case study was the forestry area of Amol County in the Mazandaran province of northern Iran ( Figure 2). Our study area has an expanse of 646 km 2 and is essential for its natural forests and as one of the common recreational centres of the country. The study area attracts tourists from all over the country. The altitude ranges from 100 m in valleys to approximately 2500 m above mean sea level in mountainous areas in the southern part. Other topographical and environmental characteristics are described in Table 1. The study area is potentially vulnerable to forest fires, which are considered as a common problem in the region. There are more than 20 villages in the study area, mostly in valleys and some located in remote forest areas. The most common activity among the population is animal husbandry. However, there are also orchards, including walnuts and apple gardens, and even some agricultural activities.

Data Used in the Analysis
To generate the forest fire susceptibility map of our study area, 16 relevant factors including environmental variables were considered as forest fire occurrence data.
The relevant topographic factors are slope aspect, slope, altitude, plan curvature, landform, and the topographic wetness index (TWI). The meteorological factors are annual temperature, potential solar radiation, and wind effect. The considered vegetation factor was the normalized difference vegetation index (NDVI). The NDVI factor was created for the high vegetated summer period, applying Landsat-8 Operational Land Imagery (OLI) 30 m resolution, retrieved from the USGS archive (http://earthexplorer.usgs.gov). The human-made factors are the distance to the nearest village, land use, and the distance to roads and recreational areas. The hydrological factors are annual rainfall and the distance to streams (see Table 1).
The topographic factors were derived from the Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER) (NASA, California Institute of Technology, USA) Global Digital Elevation Model (GDEM) with an approximate resolution of 30 m. Related topographic factors are considered as important forest fire factors that impact the fire's separate pattern in forests [47]. These factors have been widely used for forest fire susceptibility mapping [48]. The slope factor is essential because increasing in slope angle results in fire spreading more quickly [27]. Since usually northfacing slopes are colder and moister than those of south, the risk of forest fire on south-facing slopes is higher [49]. The presence of higher moister at higher altitude shows the importance of this factor [50]. The factor of wind effect was generated by three different factors, including the degree of wind direction, wind speed (m/s), and altitude layer [51]. The solar radiation varies from nearly 0.3 KWH m, ⁄ to more than 1.8 KWH m ⁄ and this range is classified in five classes based on natural

Data Used in the Analysis
To generate the forest fire susceptibility map of our study area, 16 relevant factors including environmental variables were considered as forest fire occurrence data.
The relevant topographic factors are slope aspect, slope, altitude, plan curvature, landform, and the topographic wetness index (TWI). The meteorological factors are annual temperature, potential solar radiation, and wind effect. The considered vegetation factor was the normalized difference vegetation index (NDVI). The NDVI factor was created for the high vegetated summer period, applying Landsat-8 Operational Land Imagery (OLI) 30 m resolution, retrieved from the USGS archive (http: //earthexplorer.usgs.gov). The human-made factors are the distance to the nearest village, land use, and the distance to roads and recreational areas. The hydrological factors are annual rainfall and the distance to streams (see Table 1).
The topographic factors were derived from the Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER) (NASA, California Institute of Technology, USA) Global Digital Elevation Model (GDEM) with an approximate resolution of 30 m. Related topographic factors are considered as important forest fire factors that impact the fire's separate pattern in forests [47]. These factors have been widely used for forest fire susceptibility mapping [48]. The slope factor is essential because increasing in slope angle results in fire spreading more quickly [27]. Since usually north-facing slopes are colder and moister than those of south, the risk of forest fire on south-facing slopes is higher [49]. The presence of higher moister at higher altitude shows the importance of this factor [50]. The factor of wind effect was generated by three different factors, including the degree of wind direction, wind speed (m/s), and altitude layer [51]. The solar radiation varies from nearly 0.3 KWH/m 2 to more than 1.8 KWH/m 2 and this range is classified in five classes based on natural breaks. In terms of human-made factors, various distance measures were considered. The distances and intervals for the classified layers were defined based on their importance regarding forest fires, the radius of human activities, the literature, and expert experiences [24] (see Figure 3).  Acronyms stand for: the state wildlife organization of Amol County (SWOAC) [52], the state meteorological organization of Amol County (SMOAC) [53]. (ASTER DEM) The advanced spaceborne thermal emission and reflection radiometer (ASTER) onboard the national aeronautics and space administration's (NASA's) Terra spacecraft provides digital elevation model (DEM) [54].

Data Generation for Training and Testing
In this subsection, we summarise the generation process of our training and test data. The polygons of the forest fires in the study area were generated based on GPS data (obtained from the state wildlife organization) along with the hotspots of Moderate Resolution Imaging Spectroradiometer (MODIS) (NASA, Goddard Space Flight Center, Maryland, USA). Using MODIS fire detection data is common, and used in several current studies, to specify the location and the time of forest fires [55][56][57]. Our investigation resulted in 34 forest fire GPS polygons that were evaluated by the MODIS sensor during the 2012-2017 timeframe (see Figure 1). Of these forest fire polygons, we randomly selected 70% (12,195 pixels) as training data and 30% (5226 pixels) for the validation section using ten-fold cross-validation (CV) (see Figure 4). We used the ten-fold CV approach to deal with the randomness adverse effects on the forest fire susceptibility modelling [58]. The forest fire polygons were randomly divided into ten folds of approximately 1742 pixels. Technically speaking, if we have the dataset D and divide it randomly into k number of the same folds of , , …, . Then, the model is trained k times and each time t ∈ (1, 2, …, k). For the time of t, the model is run with the dataset of D lacking of the subset of , and validated with [59]. Thus, Figure 3. Sixteen input forest fire factors used in the artificial neural network (ANN) method: (a) normalized difference vegetation index (NDVI), (b) distance to village, (c) topographic wetness index (TWI), (d) annual rainfall, (e) land form, (f) altitude, (g) distance to stream, (h) plan curvature, (i) slope aspect, (j) land use, (k) wind effect, (l) annual temperature, (m) slope degree, (n) distance to road, (o) potential solar radiation, (p) recreational areas.

Data Generation for Training and Testing
In this subsection, we summarise the generation process of our training and test data. The polygons of the forest fires in the study area were generated based on GPS data (obtained from the state wildlife organization) along with the hotspots of Moderate Resolution Imaging Spectroradiometer (MODIS) (NASA, Goddard Space Flight Center, Maryland, USA). Using MODIS fire detection data is common, and used in several current studies, to specify the location and the time of forest fires [55][56][57]. Our investigation resulted in 34 forest fire GPS polygons that were evaluated by the MODIS sensor during the 2012-2017 timeframe (see Figure 1). Of these forest fire polygons, we randomly selected 70% (12,195 pixels) as training data and 30% (5226 pixels) for the validation section using ten-fold cross-validation (CV) (see Figure 4). We used the ten-fold CV approach to deal with the randomness adverse effects on the forest fire susceptibility modelling [58]. The forest fire polygons were randomly divided into ten folds of approximately 1742 pixels. Technically speaking, if we have the dataset D and divide it randomly into k number of the same folds of D 1 , D 2 , . . . , D k . Then, the model is trained k times and each time t ∈ (1, 2, . . . , k). For the time of t, the model is run with the dataset of D lacking of the subset of D t , and validated with D t [59]. Thus, three of the folds were chosen each time for validation and other seven folds were used for training the model. This process was repeated for every fold in our inventory dataset of the forest fire. Although most researchers used ten folds, the number of folds in the spatial applications is often selected by the user without any empirical evidence. The authors of Reference [60] used a three-fold CV approach. Whereas, in Reference [61] they selected a five-fold for their training and validation aims. We selected ten folds for the cross validation because of the large volume of our inventory dataset. Therefore, the outputs of the forest fire susceptibility modelling in this study were based on the ten-fold CV that resulted in the highest accuracies in both the training and the validation sections (see Sections 3.1 and 4.1).
Fire 2019, 2, x FOR PEER REVIEW 11 of 26 three of the folds were chosen each time for validation and other seven folds were used for training the model. This process was repeated for every fold in our inventory dataset of the forest fire. Although most researchers used ten folds, the number of folds in the spatial applications is often selected by the user without any empirical evidence. The authors of Reference [60] used a three-fold CV approach. Whereas, in Reference [61] they selected a five-fold for their training and validation aims. We selected ten folds for the cross validation because of the large volume of our inventory dataset. Therefore, the outputs of the forest fire susceptibility modelling in this study were based on the ten-fold CV that resulted in the highest accuracies in both the training and the validation sections (see Sections 3.1 and 4.1).

Forest Fire Susceptibility Mapping Using Artificial Neural Network (ANN)
The ANN method is considered to be one of the most commonly used ML methods to assess susceptible areas of forest fires [28]. Generally, ANNs mimic human brain performance, and they can solve complex nonlinear problems with multiple numbers of components and variables [62]. ANNs also can find patterns and discover the trends of complex issues [63]. In the present study, to create an index for assessing forest fires, a model was used with the ANN method using a multilayer perceptron (MLP) architecture and trained with the backpropagation algorithm (BPA), which is considered as the most common algorithm for the ANN method. The MLP architecture consists of several hidden layers between the input and output layers. The number of hidden layers of a neural network depends on the complexity of the problem [64]. For our case, the feed-forward network was

Forest Fire Susceptibility Mapping Using Artificial Neural Network (ANN)
The ANN method is considered to be one of the most commonly used ML methods to assess susceptible areas of forest fires [28]. Generally, ANNs mimic human brain performance, and they can solve complex nonlinear problems with multiple numbers of components and variables [62]. ANNs also can find patterns and discover the trends of complex issues [63]. In the present study, to create an index for assessing forest fires, a model was used with the ANN method using a multilayer perceptron (MLP) architecture and trained with the backpropagation algorithm (BPA), which is considered as the most common algorithm for the ANN method. The MLP architecture consists of several hidden layers between the input and output layers. The number of hidden layers of a neural network depends on the complexity of the problem [64]. For our case, the feed-forward network was made up of an input layer with 16 neurons (input factors), one hidden layer (28 neurons) and one output layer as a network structure of 16-28-1 (see Figure 5). The number of 28 hidden layers was determined through 3-fold CV on a range of (12, 18, 28 neurons). made up of an input layer with 16 neurons (input factors), one hidden layer (28 neurons) and one output layer as a network structure of 16-28-1 (see Figure 5). The number of 28 hidden layers was determined through 3-fold CV on a range of (12, 18, 28 neurons). The training dataset consists of forest fire and non-forest fire areas containing sixteen inputs, namely, slope aspect, slope degree, altitude, plan curvature, landform, TWI, annual temperature, potential solar radiation, wind effect, NDVI, distance to village, land use, distance to road, recreational areas, annual rainfall, and distance to stream. Mapping forest fire susceptibility using the described ANN method consists of a three-stage process, namely, training, allocation, and testing the values of each pixel of an input layer were presented to the neural network along with the values of the training dataset. Each of the fifteen neurons in the input layer ( , , …, ) characterized an attribute of the input data and transferred the signal to the corresponding connected neurons ( , ,…, ) in our hidden layer. The total input signal to the neuron (in the hidden layer) can be defined as Equation (1) [65].
where, is the weighted value that came from the input neuron and is the corresponding bias value. The activation of each neuron in this layer is computed separately and transferred to the next neuron, which is the output neuron. Equation (2) presents how the output neuron O catches the values from the hidden layer.
where, zj is the j th resulting value from Equation (1), wj is the weight and wo is the corresponding bias value [66]. The calculated output value from the ANN method is compared with the forest fire pixel values, which are considered as the target value t. The comparison is done for every terrain unit until the overall error between t and the derived output value o is acceptable (minimum). The forest fire pixel values, which are considered as the target value t, were not used for training data earlier in the process. The momentum parameter avoids divergence during the process, and has a useful role in increasing the speed of convergence [67]. All weightings were updated by the backward process during each cycle in order to minimize the error. The common activation functions used for the ANN The training dataset consists of forest fire and non-forest fire areas containing sixteen inputs, namely, slope aspect, slope degree, altitude, plan curvature, landform, TWI, annual temperature, potential solar radiation, wind effect, NDVI, distance to village, land use, distance to road, recreational areas, annual rainfall, and distance to stream. Mapping forest fire susceptibility using the described ANN method consists of a three-stage process, namely, training, allocation, and testing the values of each pixel of an input layer were presented to the neural network along with the values of the training dataset. Each of the fifteen neurons in the input layer (y 1 , y 2 , . . . , y 16 ) characterized an attribute of the input data and transferred the signal to the corresponding connected neurons (z 1 , z 2 , . . . , z 9 ) in our hidden layer. The total input signal to the z x neuron (in the hidden layer) can be defined as Equation (1) [65].
where, u ix is the weighted value that came from the i th input neuron and u x is the corresponding bias value. The activation of each neuron in this layer is computed separately and transferred to the next neuron, which is the output neuron. Equation (2) presents how the output neuron O catches the values from the hidden layer.
where, z j is the j th resulting value from Equation (1), w j is the weight and w o is the corresponding bias value [66]. The calculated output value from the ANN method is compared with the forest fire pixel values, which are considered as the target value t. The comparison is done for every terrain unit until the overall error between t and the derived output value o is acceptable (minimum). The forest fire pixel values, which are considered as the target value t, were not used for training data earlier in the process. The momentum parameter avoids divergence during the process, and has a useful role in increasing the speed of convergence [67]. All weightings were updated by the backward process during each cycle in order to minimize the error. The common activation functions used for the ANN method are hyperbolic tangent and the logistic functions. For our case, the logistic function (see Equation (3)) was selected, as the output of the function is an interval value varying from 0 to 1, which is more preferable for a susceptibility map.
The transfer function f was a nonlinear sigmoid function that was used to the weighted sum of the input data before the next layer. Different learning rates were applied from 0.1 to 0.9 and the learning rate of 0.9 got a higher accuracy in forest fire susceptibility mapping. Several iterations were tested, and 1000 iterations were selected as the optimal one in terms of both of accuracy and cost. Finally, the hazard map was generated from the output values of the ANN method. The pixels with values close to 1 have a higher probability of forest fires occurring and pixels with low values indicate safer areas.

Choice of Indicators
There are several social indicators, such as education level, age, gender, and access to government facilities and health services, that can indicate the level of a community's ability to be deal with and recover from natural hazards like forest fires [18,38]. In this research, 22 social vulnerability indicators were selected based on the living standards and local conditions in our study area (see Table 3). The dataset was obtained from the Iranian Bureau of Statistics 2017 census. Our study area consists of 20 villages with an estimated population of 2887. The least populated village only has 11 inhabitants, and the most populated one has a population exceeding 1650. Since all villages belong to the same county, they have approximately equal access to government support and facilities. However, there are some indicators that vary between villages, e.g., the percent of the vulnerable population under 14 years of age or 65 and older, and the percent of the unemployed people.
In this study, we also consider several susceptible areas and valued infrastructural assets and resources within the study area that are prone to forest fires (see Table 4). There are also some people who live temporarily (seasonal) or permanently in these areas. Roads and forest parks, for instance, are frequented by visitors from all over the country and, mainly, from Tehran during the summer time. On the other hand, since the occupation of most people in our study area is the livestock sector, there are several barns in the depths of the forest areas, most of which are illegal. These barns are used to keep the livestock during the fall and winter. As they are usually established in remote areas, it is more likely that they are surrounded by or even engulfed in a fire during a forest fire event. Thus, eleven infrastructural vulnerability indicators were also selected based on the situation of our study area. To consider the spatial information of both the social and infrastructure indicators, the datasets need to be transferred to a GIS environment and aggregated to layers. However, the indicators are represented by different measurement scales (i.e., percentage and number) [68]. Thus, before being aggregated, they were transformed into a continuous scale of common units with a linear standardisation method.

Aggregation of Different Indicators Using Geographic Information System Multi-Criteria Decision Making (GIS-MCDM)
A GIS-MCDA [69] approach was used to account for the relative significance of each indicator regarding forest fires in our study area. The weightings of social vulnerability and infrastructure vulnerability were calculated separately. If all of the indicators are believed to have the same importance, an equal weighting value should be used to obtain the total vulnerability [16]. Although different indicators (e.g., age, employment and education) might have approximately equal importance when we consider the general context of vulnerability, their significance depends on the type of natural hazard. MCDAs are practical approaches for a wide range of complex decision-making systems.
The set of X = (x 1 , x 2 , . . . , x n ) was considered with n number of selected social indicators. Experts compared each pair of indicators (x i and x j ) in X. A value (a ij ) can also result from the ratio of their weightings [70]. If the social indicator x i is preferred to x j, then a ij > 1 and, conversely, if x j is preferred to x i then a ij < 1. A verbal description of the importance, introduced by Saaty to compare various indicators in a pairwise manner, is given in Table 2. The values vary from 1, for equal importance, to 9, for extreme importance. For i and j = 1,2, . . . ,n, and any a ij > 0 we have a ji = 1 a ij (reciprocal property). The resulting indicator weightings (w 1 , w 2 , w 3 , . . . ,w n ) have two conditions, namely: 0 ≤ w k ≤ 1 and n k=1 w k = 1 Therefore, the pairwise comparisons are structured in a matrix, and the matrix's elements are between 1/9 and 9. The AHP method can integrate both quantitative and qualitative dimensions of a decision-making problem [71]. The method is also able to differentiate between our indicators and to support the comparison of the performance of them with few calculations [72].  If λ max is the maximum eigenvalue, the consistency ratio (CR) of the resulting weightings is determined by CR = λ max − n/RI(n − 1) [74]. Where RI is the average random index, which for n = 2, 3, 4, 5, 6, 7 and 8, RI = 0.00, 0.52, 0.89, 1.11, 1.25, 1.35 and 1.40, respectively. A CR < 0.10 indicates an acceptable consistency throughout the whole AHP process [75,76]. Using this process, we calculated the pairwise comparison matrices for the decisions of our chosen field experts, who represented the national university, the state wildlife organization, and the meteorological organization (their names and affiliations are mentioned in the acknowledgments). Both the comparison matrices of social vulnerability and infrastructure vulnerability passed the consistency test with overall CR scores of 0.055 and 0.024, respectively. Weightings are shown in the last column of Tables 3 and 4.

The Predictive Performances and Resulting Hazard Map
In this subsection, we sought to spatially represent regions on the study area associated with high susceptibility of forest fire. The ML component of the ANN method was used for susceptibility modelling and mapping. For this goal, the output values of the ANN method were derived through GIS spatial analysis and data aggregation models. As we trained the ANN method with burned area polygons which including also ignition locations, the resulting susceptibility maps represent a measure of the likelihood of ignition and the probability of spreading of forest fires based on considered conditioning factors. Figure 6 shows the results of the forest fire susceptibility map from the applied ML method. Concerning the spatial distribution, the probability of forest fires is higher in northern and central areas where local people and tourists have accessibility to low-elevation forests. The natural breaks classification method (or Jenks optimization) used in this study generates classes of similar values separated by breakpoints. This is a common and effective method for categorizing the susceptibility mapping results when we interpret values close to each class boundary (e.g., values between "Low" and "Moderate" probability) [77]. Table 5 depicts the area and the percentage of each class of the resulting forest fire map. The RMSE [78,79] was used to assess the model performance and Equation (5) depicts the calculation: where, e i and e i are the i th observed and results of the ANN method, respectively. The best derived RMSE value based on the ten-fold CV was 0.145. Generally, small values of the RMSE show better performance of the prediction model [80]. However, in order to validate the resulting forest fire susceptibility map, we conducted a validation process based on the 5226 forest fire pixels (see Section 2.3) that were not used for training in the utilized ANN method. The validation step aims to ensure the accuracy of the resulting hazard map [76]. The spatial success of the hazard map was examined through the receiver operating characteristics (ROC) curve, which is a common method for characterizing the quality of deterministic and probabilistic prediction methods. This curve shows the trade-off between the false positive rate and the true positive rate, by plotting the false positive rate on the X axis and the true positive rate on the Y axis [24]. If the area under the curve (AUC) is close to one, this indicates that the prediction model has a high quality. Our ROC curve consists of 5226 pixels reflecting known and observed (by MODIS) forest fires in the study area. Figure 7 shows the results of comparing the observed forest fires with the forest fire susceptibility map. The results of the ROC method for our hazard map indicated accuracies fluctuate around 74% and the highest one based on the ten-fold CV was an accuracy of more than 80%. This result was used to further analysis of the risk map generating. close to one, this indicates that the prediction model has a high quality. Our ROC curve consists of 5226 pixels reflecting known and observed (by MODIS) forest fires in the study area. Figure 7 shows the results of comparing the observed forest fires with the forest fire susceptibility map. The results of the ROC method for our hazard map indicated accuracies fluctuate around 74% and the highest one based on the ten-fold CV was an accuracy of more than 80%. This result was used to further analysis of the risk map generating.

Social/Infrastructural Vulnerability Mapping
To obtain social and infrastructural vulnerability maps, the weightings derived from the AHP method were used for data aggregation within a GIS environment. A buffer zone of 200 m was used for residential areas, and different multibuffers were used for infrastructural indicators based on their importance regarding forest fires. All indicators were prepared as GIS layers and then aggregated with a weighted overlay technique. Figure 8a and b present the vulnerability maps resulting from social and infrastructural indicators, respectively. The social vulnerability map was classified into five classes of vulnerability using the natural breaks classification method. The infrastructural vulnerability map, on the other hand, was classified into four categories of vulnerability using the same method of classification. The natural breaks classification method and the number of classes for categorisation were selected based on literature and expert knowledge. Table 5 shows the area of the respective classes for both the social and infrastructural vulnerability maps.

Social/Infrastructural Vulnerability Mapping
To obtain social and infrastructural vulnerability maps, the weightings derived from the AHP method were used for data aggregation within a GIS environment. A buffer zone of 200 m was used for residential areas, and different multibuffers were used for infrastructural indicators based on their importance regarding forest fires. All indicators were prepared as GIS layers and then aggregated with a weighted overlay technique. Figure 8a and b present the vulnerability maps resulting from social and infrastructural indicators, respectively. The social vulnerability map was classified into five classes of vulnerability using the natural breaks classification method. The infrastructural vulnerability map, on the other hand, was classified into four categories of vulnerability using the same method of classification. The natural breaks classification method and the number of classes for categorisation were selected based on literature and expert knowledge. Table 5 shows the area of the respective classes for both the social and infrastructural vulnerability maps.

Generation of the Risk Map
Forest fires will have adverse impacts on different types of elements-at-risk. Thus, it is crucial to consider the forest fire threat for different sectors [9]. Moreover, all related stakeholders should be involved in a comprehensive forest fire risk mapping [81]. For this reason, we considered 16 relevant

Generation of the Risk Map
Forest fires will have adverse impacts on different types of elements-at-risk. Thus, it is crucial to consider the forest fire threat for different sectors [9]. Moreover, all related stakeholders should be involved in a comprehensive forest fire risk mapping [81]. For this reason, we considered 16 relevant forest fire factors, 22 social vulnerability indicators, and 11 infrastructural indicators, to ensure the inclusion of the most important factors that play a role in the forest fire risk assessment in our study area. The hazard map was generated using an ANN method because we had access to sufficient inventory data for training and testing matters. However, social and infrastructural vulnerability maps were created using a GIS-MCDA method, which is one of the more common methods for vulnerability and susceptibility mapping in a wide range of disciplines [82]. Mapping areas with a high probability of forest fire risk can be carried out by aggregating the hazard map and the social/infrastructural vulnerability maps. The GIS-MCDA approach was, again, used to specify the relative significance of indicators and to aggregate the forest fire susceptibility, infrastructural vulnerability, and social vulnerability to generate the final risk map. By using the AHP method, we calculated the weights resulting from the pairwise comparison matrices for the three resulting maps and the corresponding classes. The pairwise comparison matrices were filled by the same experts who were mentioned in Section 3.3. All resulting maps were prepared as GIS layers and then overlaid with the respective AHP weights, which were almost 54% for the forest fire susceptibility, 30% for the social vulnerability, and 16% for the infrastructural vulnerability layer (see Table 5). The final forest fire risk map was classified by natural breaks ('Jenks') and is represented in Figure 9 via four classes. This classification method yields classes that are based on natural groupings of the pixel values. Class breaks are specified in a way that the resulting groups are most similar and the differences between classes are maximized. Area and percentage of the forest fire risk map classes are represented in Table 6. The forest fire risk map reflects both the forest fire susceptibility (including the likelihood of ignition and the probability of spreading) and social/infrastructural vulnerability. The forest fire risk map shows that the slope aspect factor played a key role in the susceptibility of spreading, consequently raises the forest fire risk. For the other aspects, the susceptibility of a forest fire occurring is high due to the impact of other conditioning factors. Therefore, the applied ANN method could be trained well with all factors based on the inventory forest fire data and succeeded in predicting high forest fire susceptible areas. High temperature and low rainfall in the northern regions of the study area obviously increased the susceptibility of the forest fire ignition and spread. Although steep areas are generally more susceptible for forest fires, the human activity close to the main road (even with more moderate slopes in the northern parts) had a primary effect on the susceptibility and the risk map. The interaction of settlement areas (high social vulnerability) and infrastructural vulnerable areas with the high and moderate forest fire susceptibility resulted in high risk regions in the forest fire risk map (see Figure 9).

Discussion and Conclusion
Mapping the forest fire risk posed to human life and properties is an essential component of emergency land management, forest fire prevention, mitigation of adverse impacts, and response and recovery management [83]. As already discussed in the introduction section, forest fire susceptibility maps have often been used to priorities investments in forest fire prevention, and also for preparation planning. However, even considering social vulnerabilities leads to more effective forest fire risk management [84]. Since such risk maps can illustrate the location of the elements-atrisk, besides communities who have a lower capacity to deal with the fire and the spatial correlations between communities and fires, they are also more useful for land managers and firefighters for emergency planning in combating fires in real time. Therefore, a comprehensive forest fire mitigation and disaster management plan should concentrate on areas where there is an overlap between populations and their more vulnerable properties and a higher forest fire risk. In this regard, the forest fire susceptibility map of our study was generated using environmental variables. We considered the most available factors and prepared them as input data for the model. Training and testing data were generated using field survey GPS polygons and MODIS hotspot data. The field survey was done by the SWOAC [52]. According to the literature, a wide range of approaches have been used to predict areas with potential susceptibility to forest fires. Thus, different methods have been applied by various researchers to explore the potential role of relative factors that point to fire events. In this research, we generated a model using an ANN method with an MLP architecture that

Discussion and Conclusion
Mapping the forest fire risk posed to human life and properties is an essential component of emergency land management, forest fire prevention, mitigation of adverse impacts, and response and recovery management [83]. As already discussed in the introduction section, forest fire susceptibility maps have often been used to priorities investments in forest fire prevention, and also for preparation planning. However, even considering social vulnerabilities leads to more effective forest fire risk management [84]. Since such risk maps can illustrate the location of the elements-at-risk, besides communities who have a lower capacity to deal with the fire and the spatial correlations between communities and fires, they are also more useful for land managers and firefighters for emergency planning in combating fires in real time. Therefore, a comprehensive forest fire mitigation and disaster management plan should concentrate on areas where there is an overlap between populations and their more vulnerable properties and a higher forest fire risk. In this regard, the forest fire susceptibility map of our study was generated using environmental variables. We considered the most available factors and prepared them as input data for the model. Training and testing data were generated using field survey GPS polygons and MODIS hotspot data. The field survey was done by the SWOAC [52]. According to the literature, a wide range of approaches have been used to predict areas with potential susceptibility to forest fires. Thus, different methods have been applied by various researchers to explore the potential role of relative factors that point to fire events. In this research, we generated a model using an ANN method with an MLP architecture that was trained with the BPA through ten-fold CV. The main limitation that we faced in this study was the dataset used. The fire inventory was from MODIS with a resolution of 1000 m while the resolution of the condition factors was 30 m. The dataset from the SWOAC includes the polygons with detailed borders. However, this documentation does not include small forest fires. Therefore, for the cases of the extensive wildfires, we considered the GPS polygons rather than those of from MODIS. Still, for the small events, the MODIS dataset was applied. Thus, the dataset of the MODIS is more accurate in terms of the number of the forest fires. The GPS dataset is more reliable regarding the shape of the polygons of burned areas because of the low resolution of MODIS. Consequently, the integration of both sources resulted in a more reliable inventory dataset of the location and the extension area of the wildfires. Nevertheless, the resolution of the resulting inventory dataset was not the same as that of the input data. Therefore, this difference in the resolution resulted in some uncertainties, which are difficult to quantify.
The model was developed and trained based on the previous forest fire events between 2012 and 2017, and the relevant factors contributing to the forest fire. The performance of the model was evaluated by the RMSE, and the resulting forest fire susceptibility map was validated using the ROC curve. Both of these measures showed acceptable accuracies of the results of the model. The resulting forest fire susceptibility map with the highest accuracy revealed that the southern part of our study area is more susceptible to forest fires in the future. This matter illustrates that the factors of annual rainfall and annual mean temperature play an important role by creating extreme dryness, which is a crucial factor determining forest fires [85]. The areas with high forest fire susceptibility also have a close spatial correlation with the previously recorded forest fire events.
Although environmental variables contributing to the forest fire played an essential role in the assessment of susceptible forest fire areas, they are not the only reason for both forest fire susceptibility and risk. As some of the forest fires in in our case study are reported to be a result of human activity and the local people in particular SWOAC [52], considering the areas and some specific sites of human activity such as recreational areas can help to have a better understanding of the results of the forest fire susceptibility and risk. The distance from population concentration is considered to be relevant to human activity. Therefore, the closer regions to the settlements and population concentration can indicate a higher presence of human activity and consequently more susceptibility and risk of forest fire [31]. Thus, the population density and social vulnerability indicators were used, which were gained from different sources and, mostly, from census data. We also consider the areas with infrastructural valued assets and resources that are recognized as public/private properties. All indicators were weighted and ranked using expert knowledge and pairwise comparison matrices. Mostly, local experts were asked to contribute to this study, and they are mentioned in the acknowledgements section. The reason for selecting local experts to weight the indicators is that they are more familiar with the existing situation in our study area. These experts have different field backgrounds, like geography, geomorphology, meteorology, and wildlife. The selection of social vulnerability indicators may have some limitations, and the authors cannot be sure to have considered all relevant indicators that determine social vulnerability. However, we tried to localise the information that defines vulnerability in our case study. The social vulnerability generally depends on different characteristics, e.g., the country that the study area is located in, government support, and social features. Thus, the indicators may vary among different nations. The average life expectancy of houses, for instance, is a range between 10 and 15 years in our case study area. As this measure depends on the materials and different standards that are used for buildings, it may vary in other countries and even states. However, the concept of social vulnerability is a stable concept, and a large segment of the indicators are the same in different communities (e.g., age, gender and occupation). As we used the most relevant factors regarding forest fires and a large number of social/infrastructural vulnerability indicators, the performed methodology can easily be generalized and adapted to different locations like Australia, California, and Spain -i.e., fire-prone areas. Thus, the transferability of the method only requires minor changes and localization regarding social vulnerability indicators.
The novelty of our study is using both data-based ML and knowledge-based multi-criteria decision analyses for producing the forest fire risk map. Although there are several instances of using these methodologies separately for risk map generation, we provide an integration of using both methods for this aim. The reasons for using different methods for forest fire susceptibility mapping and the production of social/infrastructural vulnerability maps were described in Section 4.3. Our results emphasize the importance of considering a social/infrastructural vulnerability assessment when creating forest fire risk maps. Both resulting susceptibility and vulnerability maps were overlayed and analysed to understand how they correspond.
We propose that our future research can consider biodiversity and its contributing role in forest fires. For the detection of different tree species, convolutional neural networks can be applied. This method is a deep learning method, but it consists of a more significant number of various layers and requires a massive amount of training data. We also want to use other ML methods such as support vector machines and random forest for forest fire susceptibility modelling and mapping purposes.