A Comparative Study of Forest Fire Mapping Using GIS-Based Data Mining Approaches in Western Iran

Abstract

Mapping fire risk accurately is essential for the planning and protection of forests. This study aims to map fire risk (probability of ignition) in Marivan County of Kurdistan province, Iran, using the data mining approaches of the evidential belief function (EBF) and weight of evidence (WOE) models, with an emphasis placed on climatic variables. Firstly, 284 fire incidents in the region were randomly divided into two groups, including the training group (70%, 199 points) and the validation group (30%, 85 points). Given the previous studies and conditions of the region, the variables of slope percentage, slope direction, altitude, distance from rivers, distance from roads, distance from settlements, land use, slope curvature, rainfall, and maximum annual temperature were considered for zoning fire risk. Then, forest fire risk maps were prepared using the EBF and WOE models. The performance of each model was examined using the Relative Operating Characteristic (ROC) curve. The results showed that WOE and EBF are effective tools for mapping forest fire risks in the study area. However, the WOE model shows a slightly higher Area Under the Curve value (0.896) compared to that of the EBF model (0.886), indicating a slightly better performance. The results of this study can provide valuable information for preventing forest fires in the study area.

1. Introduction

Forest fire is inevitable and plays a vital role in vegetation sequencing and landscape transformation [1]. This type of fire is widely considered a serious risk exerting negative effects on the environment and society in many countries of the world [2]. In recent years, severe forest fires have occurred in some countries, such as the United States, Sweden, China, Indonesia, and Portugal, because of the effects of climatic events caused by climate change, including long periods of drought accompanied by dry and hot air [3]. It is forecast that forest fires will increase in the future due to continuing climate change [4]. In addition, human activities, e.g., the combustion of garbage, stubble burning, and discarded cigarettes play key roles in the ignition, combustion, and spread of forest fires [5,6].

Since forest fires follow a nonlinear and complex process under the effects of various factors, it is difficult to develop highly accurate models for predicting forest fire risks [7]. Various factors have been utilized to identify the potential of forest fire risks, among which one can refer to slope [8], distance from roads [9], aspect [10], altitude [11], rainfall [12], temperature [13], distance from rivers [14], land use [15], vegetation [16], distance from settlements, wind speed [17], curvature [18], and the topographic wetness index [19]. Given that forest fires are caused by many factors, it is necessary to consider various influencing factors as much as possible for high-precision fire risk mapping in specific regions.

Forest fire risk mapping is regarded as an effective and important tool for fire forecasting, which is essential to manage and protect forest areas [8,20]. Risk potential maps can provide planners and managers with useful information to better monitor areas with higher forest fire risk potential. In addition to considering influencing factors adequately, the adopted model is one of the most important factors in mapping fire risk. In general, models employed in zoning fire risks can be divided into (1) statistical and data-based models, (2) machine-based learning models, (3) multi-criteria decision-making models, and (4) integrated models [21]. In addition, many models have been developed to predict forest fire risks, among which one can refer to simple statistical methods, including Poisson regression [22], binary logistic regression [13], and linear regression [20]. In this regard, more complex models, for predicting forest fire risks, include Pareto distribution [23], favorability functions [24], and the numerical simulation-based approach [3], as well as complex mathematical models, such as ELMFIRE [25]. Moreover, some machine learning algorithms have been employed in recent years to predict forest fire risks, such as the evidential belief function (EBF) [13], support vector machine [14], random forest [26], and neural fuzzy algorithm [10].

Recently, Geographic Information System (GIS)-based models have been developed for forest fire probability modeling. In these methods, it is assumed that conditions that have led to wildfires’ occurrence in the past, are likely to continue causing fires in the future as well [27]. GIS-based models such as weight of evidence (WOE) were initially introduced to identify and explore mineral deposits [28] and for landslide susceptibility mapping [29]. Carrara et al. [30] consider conditional probability analysis a valuable risk zoning tool, especially if appropriate factors and a good understanding of risk factors exist. Later, this model was applied to forest fire risk mapping and showed good performance in case studies, such as in China [31] and Iran [32]. The main advantages of the EBF and WOE methods are that they calculate the weighted value of the factors based on a statistical formula and thus avoid the subjective choice of weighting factors. In addition, input maps with missing data (incomplete coverage) can be accommodated in the model and under sampled data do not significantly impact on the results [29,33]. Results of studies show that data-driven approaches are the most frequently applied methods, while ensemble approaches are more accurate [21].

Iran is one of the countries in the Middle East and North of Africa that face high forest fire risk, with an annual average number of 130 fires and an annual average burnt area of 5400 ha [34]. Therefore, mapping fire risk plays a crucial role in forest fire management in Iran because knowing the location with the highest risk is essential to minimize threats to resources, life, and property. To identify fire risk potential in forests more accurately, it is vital to compare the performance of different methods and techniques as well as consider influencing factors comprehensively for a specific study. Against this background, this study aims to identify the potential for forest fire risks using the data mining approaches of EBF and WOE by considering various factors with an emphasis put on climatic variables in Marivan County in the Kurdistan province of western Iran. The results of this study will be valuable for the planning and protection of forests in Iran.

2. Materials and Methods

2.1. Study Area

This study with an area of 231,721.5 hectares was conducted in Marivan County, located in the west of the Kurdistan province, Iran. This city is located between the eastern longitude of 45°58′50″ to 46°46′6″ and the northern latitude of 35°20′56″ to 35°49′10″, with the elevation ranging from ~1000 m to 3000 m. Figure 1 shows the location and topography of the study area.

2.2. Preparing Layers and Maps

The 284 fires having occurred in Marivan County were divided into two parts, including the training data (70%; 199 incidents) and the validation data (30%; 85 incidents). The fire information was provided by the Department of Natural Resources and Watershed Management of Marivan City. Then altitude, slope percentage, slope direction, distance from roads, distance from rivers, distance from settlements, land use, slope curvature, average annual rainfall, and average temperature, which have been identified as factors affecting the potential of forest fire risks, were used to produce the fire risk zoning map. The methodology used in this study is graphically represented in Figure 2.

To classify the criteria and layers, the expert opinion method was utilized [35]. The altitude map was prepared in ArcGIS 10.3 software based on the digital elevation model (DEM) obtained from the ALOS PALSAR satellite as well as the Alaska Satellite Facility website (https://vertex.daac.asf.alaska.edu/ accessed on 31 October 2021), and was divided into six classes. Furthermore, the maps of the slope percentage, the slope, and the slope curvature were prepared based on the DEM, and were classified into six, nine, and three classes, respectively. The land use layer was obtained from the Forests, Range and Watershed Management Organization, and was divided into different classes based on varied uses. The map of distance from roads, distance from settlements, and distance from rivers, was prepared using the digital topographic map of 1:50,000 and through applying the functions of Distance in the ArcGIS environment. The monthly 20-year data (2000–2020) at eleven rainfall and climatology stations inside and outside the study area were employed to prepare the annual rainfall and maximum annual temperature map, through employing the Inverse Distance Weighting (IDW) method, which was divided into five classes. Figure 3 shows the maps of the parameters effective in identifying fire risk potential of forests.

2.3. Weight of Evidence (WOE)

WOE is a model used to integrate spatial information, being known as one of the models of the Bayesian theory within the framework of linear logarithms [36]. The weight of each factor is calculated based on Equations (1) and (2).

$$W_i^{+}=\frac{\frac{N_{pix1}}{N_{pix1}+N_{pix2}}}{\frac{N_{pix3}}{N_{pix3}+N_{pix4}}}$$

(1)

$$W_i^{-}=\frac{\frac{N_{pix2}}{N_{pix1}+N_{pix2}}}{\frac{N_{pix4}}{N_{pix3}+N_{pix4}}}$$

(2)

where:

$W_i^{+}$ represents the positive weight indicating a positive relationship between the predictable variable and forest fire. $W_i^{-}$ indicates that their correlation is negative. In addition, the difference between $W_i^{+}$ and $W_i^{-}$ is weight contrast, with magnitude showing the entire spatial association between the predictable variable and forest fire [37]; moreover, Equation (3) represents its computational equation.

$N_{pix1}$ denotes the number of fire occurrence pixels in each class.

$N_{pix2}$ represents (the total number of fire occurrence pixels in each map) − (the number of fire occurrence pixels in each class).

$N_{pix3}$ shows (the number of pixels in each class) − (the number of fire occurrence pixels in each class).

$N_{pix4}$ represents (the total number of pixels in each map) − (the total number of fire occurrence pixels in each map) − (the number of pixels in each class) + (the number of fire occurrence pixels in each class) [38].

$$C_i=W_i^{+}-W_i^{-}$$

(3)

To obtain the final weight for each factor, positive and negative weights of various classes of each factor are added together. If a factor’s weight is positive, it will play a role in the fire occurrence; in contrast, if its weight is negative, it will have no role in the fire occurrence. In addition, some factors have a negligible effect on the fire occurrence, with their weight being 0 or close to 0. Upon entering the weights in the ArcGIS environment on thematic maps, the weighted thematic map was produced. In addition, by integrating all these maps, the fire occurrence potential was forecast.

2.4. Evidential Belief Function (EBF)

This model incorporates four basic functions of the degree of belief (Bel), degree of disbelief (Dis), degree of uncertainty (Unc), and degree of plausibility (Pls) within the range of 0 and 1. The data extracted from the function EBF estimates both the spatial correlation between effective factors and fire occurrence, and the spatial correlation between classes of each of the effective factors [39]. Equations for the four above mentioned functions are estimated by Equations (4)–(8) as follows

$$Bel=\frac{Be{l}_1+Be{l}_2+\dots \dots +Be{l}_n}{1-{\sum }_{i=2}^n\left(Be{l}_{i-1}Di{s}_i-Di{s}_{i-1}Be{l}_i\right)}$$

(4)

$$Dis=\frac{Di{s}_1+Di{s}_2+\dots \dots +Di{s}_n}{1-{\sum }_{i=2}^n\left(Be{l}_{i-1}Di{s}_i-Di{s}_{i-1}Be{l}_i\right)}$$

(5)

$$Unc=\frac{{\sum }_{i=2}^n(Un{c}_{i-1}Un{c}_i+Be{l}_{i-1}Un{c}_i+Be{l}_{i}Un{c}_{i-1}+Di{s}_{i-1}Un{c}_i+Di{s}_{i}Un{c}_{i-1})}{1-{\sum }_{i=2}^n\left(Be{l}_{i-1}Di{s}_i-Di{s}_{i-1}Be{l}_i\right)}$$

(6)

$$Pls=Bel+Unc$$

(7)

$$Bel+Unc+Dis=1$$

(8)

In these equations, n denotes the number of factors. One of the features of the EBF model is that Bel, Dis, and Unc functions are within the range of 0 and 1. Thus, if Unc equals 1, Bel and Dis will equal 0, and in the case of the sum of Bel and Dis equaling 1, the degree of Unc will equal 0.

First, the functions of the models were formulated in Excel software. Then the weights calculated in this step were added to the variables in the ArcGIS environment. Finally, the weighted variables were summed together in the ArcGIS environment and a forest fire susceptibility map was prepared.

Classification was performed based on natural breaks methods. Finally, the prepared map was divided into five hazard classes, namely, very low, low, moderate, high, and very high risk.

2.5. Assessment of Models

The performance of the WOE and EBF models was examined using the Relative Operating Characteristic (ROC) curve. In the ROC curve, being a graph, the ratio of pixels for predicting the fire occurrence or nonoccurrence on the horizontal axis (true positive or I-specificity) is shown against their complementary values, i.e., the ratio of pixels predicted incorrectly (false positive or sensitivity) on the vertical axis. This curve was calculated and drawn in SPSS Statistics Version 16.0 (IBM, Chicago, IL, USA). The area under this curve is named the “Area Under Curve (AUC)”. In addition, the model with the largest AUC value shows a relatively better performance. An AUC value equaling 0.5 is considered the neutral model. Moreover, the closer this value is to 1, the more the models’ efficiency will increase [40]. In this part, in order to validate the models, data on 85 fire occurrences in the study area were considered the validation data.

3. Results

Table 1. Number of pixels, number of fire points, parameter class scores in the WOE and EBF.

Variable	Class	Number of Pixels in the Domain	Number of Fire Points	Belief Function Components				WOE
				Bil	Dis	Unc	Pls
Aspect	F	151,653	2	0.11	0.11	0.78	0.89	−0.001
	N	1,593,046	19	0.10	0.11	0.79	0.89	−0.013
	NE	1,746,340	18	0.09	0.11	0.80	0.89	−0.296
	E	1,753,597	24	0.11	0.11	0.78	0.89	0.021
	SE	1,924,512	18	0.08	0.12	0.81	0.88	−0.406
	S	2,071,541	30	0.12	0.11	0.77	0.89	0.088
	SW	2,083,404	33	0.13	0.11	0.76	0.89	0.194
	W	1,832,805	30	0.14	0.11	0.76	0.89	0.229
	NW	1,656,081	25	0.13	0.11	0.77	0.89	0.016
Slope	0–10	1,718,356	29	0.20	0.16	0.64	0.84	0.262
	10–20	1,805,695	30	0.20	0.16	0.64	0.84	0.246
	20–30	2,357,440	30	0.15	0.17	0.68	0.83	−0.064
	30–40	2,837,046	40	0.17	0.16	0.67	0.84	0.060
	40–50	2,712,025	35	0.15	0	0.85	1	−0.049
	>50	3,382,417	35	0.12	0	0.88	1	−0.327
Altitude	1196–1500	3,542,781	108	0.59	0.10	0.31	0.90	1.330
	1500–1800	4,756,054	73	0.30	0.16	0.55	0.84	0.205
	1800–2100	3,848,867	14	0.07	0.21	0.72	0.79	−1.533
	2100–2400	1,870,520	4	0.04	0.19	0.77	0.81	−0.115
	2400–2700	641,156	0	0.00	0.00	1	1	−0.044
	2700–3153	170,770	0	0.00	0.00	1	1	−0.012
Proximity to settlements (m)	0–500	3,068,853	56	0.33	0.18	0.49	0.82	0.307
	500–1000	5,699,514	88	0.28	0.18	0.54	0.82	0.099
	1000–1500	3,562,422	42	0.21	0.21	0.58	0.79	−0.038
	1500–2000	1,451,488	10	0.12	0.21	0.66	0.79	−0.051
	>2000	1,047,806	3	0.05	0.00	0.95	1.00	−0.058
Proximity to rivers (m)	0–500	6,486,671	59	0.11	0.25	0.65	0.75	−0.612
	500–1000	4,299,851	48	0.13	0.21	0.66	0.79	−0.250
	1000–1500	2,351,038	54	0.27	0.17	0.56	0.83	0.681
	1500–2000	1,045,874	30	0.34	0.18	0.48	0.82	0.090
	>2000	646,649	8	0.15	0.00	0.85	1.00	−0.004
Proximity to roads (m)	0–500	4,445,101	79	0.28	0.17	0.55	0.83	0.431
	500–1000	3,200,966	67	0.33	0.17	0.50	0.83	0.612
	1000–1500	2,233,710	23	0.16	0.21	0.63	0.79	−0.305
	1500–2000	1,531,765	15	0.15	0.21	0.64	0.79	−0.031
	>2000	3,418,541	15	0.07	0.00	0.93	1	−0.184
Land use	Wet farming	128,077	0	0.00	0.08	0.92	0.92	−0.009
	Garden	216,657	0	0.00	0.08	0.92	0.92	−0.015
	Dense forest	1,469,558	40	0.08	0.07	0.85	0.93	0.825
	Dry farming	1,706,907	33	0.06	0.07	0.87	0.93	0.422
	Good range	3,128,854	6	0.01	0.09	0.90	0.91	−2.154
	Low forest	1,033,863	58	0.17	0.06	0.77	0.94	1.701
	Wet and dry farming	525,611	15	0.09	0.07	0.84	0.93	0.795
	Mod forest	3,311,325	34	0.03	0.08	0.89	0.92	−0.335
	Mod range	2,836,793	6	0.01	0.09	0.90	0.91	−2.031
	Poor range	242,930	0	0.00	0.08	0.92	0.92	−0.017
	Settlements	37,376	6	0.49	0.07	0.43	0.93	2.508
	Water body	52,693	1	0.06	0.08	0.87	0.92	0.346
	Wetland	114,355	0	0.00	0.08	0.92	0.92	−0.008
Rainfall (mm)	535–600	954,661	0	0.00	0.22	0.78	0.78	−0.067
	600–700	3,011,210	2	0.01	0.25	0.73	0.75	−3.223
	700–800	2,955,632	19	0.13	0.23	0.64	0.77	−0.858
	800–900	3,388,449	47	0.28	0.20	0.52	0.80	0.043
	900–991	4,520,204	131	0.58	0.00	0.42	1.00	0.710
Temperature (°C)	33.2–33.5	221,914	0	0.00	0.21	0.79	0.79	−0.015
	33.5–34	1,199,104	1	0.02	0.22	0.76	0.78	−2.858
	34–34.5	2,495,006	2	0.02	0.24	0.74	0.76	−0.174
	34.5–35	5,430,397	58	0.28	0.23	0.49	0.77	−0.111
	35–35.4	5,483,735	138	0.67	0.00	0.33	1.00	0.721
Curvature	Concave	3,317,490	36	0.28	0.35	0.37	0.65	−0.266
	Flat	8,225,265	116	0.36	0.32	0.33	0.68	0.115
	Convex	3,287,393	47	0.36	0.33	0.31	0.67	0.082

shows the results of the relationship between each effective factor and fire occurrence points, using the WOE method and the EBF. According to the table, the western and southwestern directions in both models had the highest impact, yet the southeastern, northeastern, and northern directions had the least impact on the fire occurrence. Moreover, in both models, slopes of 0–10% and 10–20% had the highest impact, yet slopes higher than 50% had the lowest impact on fire occurrence. In addition, the highest number of fires, in both models, occurred in the altitude class of 1196–1500 m, yet the lowest number of fires in the two models of WOE and EBF occurred in the altitude classes of 1800–2100 and 2400–3153 m, respectively. Regarding distance from settlements, the fire probability decreased with an increase in the distance from settlements. The highest fire risk in the two models of WOE and EBF was observed at the distances of 1000–1500 and 1550–2000 m from rivers, respectively. In addition, the lowest fire risk in both models was observed at a distance of 0–500 m from rivers. Additionally, given the weighting results, there was a reverse relationship between distance from roads and fire occurrence potential; in other words, the shorter the distance from roads, the greater the fire probability would be. The investigation of the land use map demonstrated that fires occurred more often in the sparse forests. Given the results of both models, the incidence rate of fires increased with an increase in rainfall, with the highest fire risk having been observed in the rainfall class of 900–991 mm. Furthermore, the lowest rates of fires in the models of WOE and EBF were observed in the classes of 600–700 mm and 535–600 mm, respectively. Considering the results of both models, fire incidents increased with an increase in the maximum annual temperature, with the highest fire risk having been observed for the temperature class of 35–35.4 °C. Moreover, the lowest fire incidents in the models of WOE and EBF were observed in the classes of 33.2–33.5 °C and 33.5–34 °C, respectively. Furthermore, investigation of the slope curvature map demonstrated that the highest fire potential occurred in both models on the flat slope, with the lowest fire potential occurring in both models on the concave slope.

The fire potential maps using the WOE and EBF models are illustrated in Figure 4. The areas and percentage of floor areas of potential fire occurrence are listed in

Table 2. Area and area percentage of fire potential classes based on the two models of WOE and EBF.

Classes	Weight of Evidence		Evidential Belief Function
	Area of Classes (ha)	Area of Classes (%)	Area of Classes (ha)	Area of Classes (%)
Very low	20,457.19	8.85	36,318.72	15.7
Low	34,343.57	14.86	49,741.63	21.5
Moderate	58,206.4	25.19	55,997.06	24.2
High	52,523.91	22.73	40,962.98	17.7
Very high	65,566.89	28.37	48,077.14	20.9

. As Table 2 shows, the results of the two models of WOE and EBF showed that about 76 and 62% of the study area were in the moderate to very high classes, respectively.

To validate the fire potential maps, the ROC curve was utilized. The area under the ROC curve is named AUC; if its value is less than 0.5, it will indicate the inaccuracy of the model, and if its value ranges from 0.5 to 1, it will indicate the accuracy of the model for predicting the presence or lack of any fire potentials.

Table 3. AUC values for the models predicting fire potentials.

Row	Predicting Model	Validation Data
1	WOE	0.896
2	EBF	0.886

shows AUC values for the evaluated models based on the validation data presented. In addition, Figure 5 presents the ROC curve of the evaluated models based on the validation data. Among the WOE and EBF models that were studied, the highest accuracy was related to the WOE model (0.896). Hence, in terms of identifying fire potentials, the WOE model delivered a better performance than the EBF model.

4. Discussion

Fire is one of the major causes of natural disturbances in forest ecosystems, greatly affecting forest resources, climate change, and ecological sequences [41] as well as a major factor in destroying forests, and threatening such vital ecosystems [42]. In this study, the two models of weight of evidence and evidential belief function were used to identify fire risk potentials. The selection of effective parameters causing fires is a crucial issue in modeling fire risk potential in forests. A comprehensive study was conducted by considering various parameters for mapping forest fire risk in this study, including altitude, slope percentage, slope direction, slope curvature, land use, distance from settlements, distance from roads, distance from rivers, average annual rainfall, and maximum annual temperature.

The highest fire potential in the present study was observed in the lowest altitude class, which could have been due to the higher concentration of human activities in lower altitude classes. This is consistent with the results of the previous study of Dong et al. [43]. In contrast, Zhang et al. [11] and Hong et al. [14] stated that fire risks were positively correlated to altitude, with this finding being inconsistent with the results of the present study.

Most fires occurred on slopes of less than 50% in the study region. The reason for this could be the higher human activity on less steep slopes. Slope direction significantly affects local conditions, including sun exposure, dominant wind direction, rainfall amount, and morphological structure, being in turn correlated with fire events [44]. Accordingly, slope direction was another factor effective in causing fire, with the highest fire occurrence being observed in the western and southwestern directions. In fact, the western and southern directions are drier because of them receiving more sunlight. In addition, it goes without saying that dry vegetation is more susceptible to fire. Consistent with the present study, Ashtiani et al. [45] reported that the highest incidents of fire occurred in the southern and southwestern directions. There is higher human activity in the neighborhood of residential areas, so more fires occur in such places [46]. In addition, fire events decreased with an increase in the distance from residential areas. Though humans are the major cause of fire, they play a dual role in the occurrence and spread of fire [12].

The layer of distance from waterways plays a dual role in the occurrence of fire as well. Human activity in the vicinity of drainage networks and rivers is higher and plays a leading role in the occurrence of fire and the increase in its incidence. On the other hand, it plays a reducing role in the occurrence of fire. This is because there is more moisture in the vicinity of rivers and lakes, with this moisture having an inhibiting and reducing role in the occurrence of fire. Based on the results of the present study, a decrease in the distance from rivers led to the incidence of fire decreasing.

The results of the models show that there is a reverse correlation between the distance from roads and fire potential; in other words, the shorter the distance from roads, the higher the fire probability will be. This is due to the higher human activity near roads. The results of this part are in line with the study by Mota et al. [47] in Matogroso, Brazil. Investigation of the land use map showed that fires occur more often in lands with sparse forest use. Accordingly, Bazyar et al. [48] stated that the highest fire incidence occurred in land uses of dense forests.

Hong et al. [14] examined fire risk potential using data mining approaches. They stated that climatic parameters were less important at a local scale, but they were more important at a regional scale. Meanwhile, the attributes of topography, vegetation, and human activity were more important at a local scale. From among the climatic factors affecting fire events, one can refer to temperature and precipitation. In the Zagros forests, the highest rainfall occurs in winter and spring and the lowest rainfall occurs in summer. In general, the lowest rainfall occurs from June to September [1]. Based on the results of both models, fire incidents increased with an increase in the rainfall and maximum annual temperature. An increase in the amount of rainfall increases vegetation density, thereby increasing the amount of fuel. On the other hand, increases in the temperature make the fuel dry out, resulting in an increased possibility of fire. These findings were in line with those of the study by Mhawej et al. [49]. Zhang et al. [11] reported that the fire risk was positively correlated to temperature in the grasslands of China, being consistent with the findings of the present study.

Investigation of the slope curvature map demonstrated that the highest fire potential occurred on flat slopes, and that the lowest fire potential was observed on concave slopes, in both models. Based on the study of Nami et al. [50], the highest fire incidence was observed on flat slopes, consistent with the results of the present study. Bazyar et al. [48] obtained similar results as well.

The results of the models employed in the present study were evaluated and validated using the system performance attribute curves. Based on the evaluation criterion utilized in this study, i.e., ROC, and in accordance with the validation data, the WOE model (0.896) delivered a better performance than the EBF model (0.886). In addition, the good performance of both models in identifying forest and rangeland fire risk potential demonstrated that the coding, processing, and modeling tasks were performed successfully in this study. Furthermore, the results of the two models showed that about 76% and 62% of the study area fell in the middle to very high classes, respectively. Nami et al. [50] investigated fire risk potential in the Hyrcanian region of Iran using the EBF model and concluded that the EBF model (0.841) delivered a good performance in terms of fire risk potential. In another study, Pourghasemi [13] examined fire risk potential in Iran using the EBF model and reported that this model delivered a good performance (0.819) in identifying fire risk potential.

5. Conclusions

In the present study, the two models of EBF and WOE were employed to map the fire risk potential of forests in Marivan County of Kurdistan province, Iran. The maps produced by these two models were divided into the five potential categories of very low, low, medium, high, and very high. The largest area of regions with very high potential was observed in the WOE model. Overall, the results of this study demonstrated that 62–76% of the study area had medium to very high fire risk potentials. Thus, some measures must be adopted to prevent and address the possible dangers of fire incidents in the forests of the study region. Such measures should incorporate an increase in the monitoring of areas with high and very high fire risk potentials, building fire barriers, and allocating more funding to the firefighting departments of forests. From among other solutions, one can refer to equipping villagers, stock raisers, and users of natural areas with firefighting equipment. Most of the fires were prompted by humans, which could be caused by releasing bottles of water or glass bottles in forests. The other factor effective in the occurrence of fires was the incorrect method of turning the fire on or off. Thus, one of the ways to prevent forest fire is to create a culture in terms of the importance and method of utilizing these national resources. In the end, one can conclude that producing an accurate and reasonable location fire risk map may assist managers and planners with identifying areas with high fire risk potential to manage crises in vulnerable areas.