GIS-Based Modeling for Vegetated Land Fire Prediction in Qaradagh Area, Kurdistan Region, Iraq

Abstract

This study aims to estimate the susceptibility of fire occurrence in the Qaradagh area of the Iraqi Kurdistan Region, by examining 16 predictive factors. We selected these predictive factors, dependent on analyzing and performing a comprehensive review of about 57 papers related to fire susceptibility. These papers investigate areas with similar environmental conditions to the arid environments as our study area. The 16 factors affecting the fire occurrence are Normalized Difference Vegetation Index (NDVI), slope gradient, slope aspect, elevation, Topographic Wetness Index (TWI), Topographic Position Index (TPI), distance to roads, distance to rivers, distance to villages, distance to farmland, geology, wind speed, relative humidity, annual temperature, annual precipitation, and Land Use and Land Cover (LULC). To extract fires that occurred between 2015 and 2020, 121 scenes of satellite images (most of them are scenes of Sentinel-2) were used, with the aid of a field survey. In total, 80% of the data (185,394 pixels) were used for the training dataset in the model, and 20% of the data (46,348 pixels) were used for the validation dataset. Conversely, 20% of these data were used for the training dataset in the model, and 80% of the data were used for the validation dataset to check the model’s overfitting. We used the logistic regression model to analyze the multi-data sites obtained from the 16 predictive factors, to predict the forest and vegetated lands that suffer from fire. The receiver operating characteristic (ROC) curve and the area under the curve (AUC) were used to evaluate the accuracy of the proposed models. The AUC value is more than 84.85% in all groups, which shows very high accuracy for both the model and the factors selected for preparing fire zoning maps in the studied area. According to the factor weight results, classes of LULC and wind speed gained the highest weight among all groups. This paper emphasizes that the used approach is useful for monitoring shrubland, grassland, and cropland fires in other similar areas, which are located in the Mediterranean climate zone. Besides, the model can be applied in other regions, taking the local influencing factors into consideration, which contribute to forest fire mitigation and prevention planning. Hence, the mentioned results can be applied to primary warning, fire suppression resource planning, and allocation work. The mentioned results can be used as prior warnings of the outbreak of fires, taking the necessary measures and methods to prevent and extinguish fires.

1. Introduction

Vegetated lands are important geomorphological phenomena. They play a significant role in surface processes, such as erosion and landslides [1,2,3]. Vegetation is a vital natural resource, which aids in the preservation of environmental balance [4]. Wildfires are the greatest serious hazard to the world’s grasslands and forests, causing severe ecological, economic, and social consequences [5]. Wildfires may occur as a result of long dry seasons; however, the majority of these events are caused by anthropogenic actions, such as accumulated fuel load in unmanaged vegetated environments as well as vegetation removal and clearing for increased pasture and agricultural lands [5,6]. They have not, yet, been described adequately because of the complexity of the processes determining fire behavior, the great amount of data required, and the difficulty of the extraction and collection of the data [7]. Every year, wildfires cause a great deal of damage to natural vegetation, property, the biotic elements of a given ecosystem, and the environment itself, thereby threatening people’s lives and significant human resources [8]. Fire leads to environmental degradation, by reducing vegetation coverage and, consequently, food resources for fauna, while modifying soil properties leading to degradation, mass movement, and diversity exhaustion [9].

It is vital to identify the factors affecting the spread and occurrence of fires in vegetated areas, to reduce the fire risk in a given environment [10]. Generally, the causes of fires in the vegetation include natural (environmental) and artificial (anthropogenic) factors. Environmental factors affecting fire include biological, physiographic, and climatic factors [11]. Vegetation is one of the most common biological factors affecting fire occurrence, which exerts its effects depending on the species, density, humus, and moisture [12,13,14,15,16]. The factor of topography is composed of the sub-factors of slope, aspect, altitude, plan curvature, and Topographic Position Index (TPI). TPI is another vital factor affecting fire occurrences [17,18,19,20,21].

Among climatic factors, temperature increases, relative humidity decreases, precipitation amounts, and wind speed and direction affect fire occurrences [19,22,23]. Aside from natural factors, anthropogenic factors as fire ignition sources have caused many fires worldwide [24,25], such as roads, farmlands, and residential areas located within or around forests [14,26,27]. In addition, many fires are produced as a means of clearing lands for agriculture or afforestation purposes [28,29,30]. Therefore, we carried out an inventory of previous research that has a similar topographic environment to the study area, to select the effective factors in fire occurrence. These factors can be divided into two categories: environmental and anthropogenic factors. Environmental factors include slope, slope aspect, elevation, topographic Wetness Index (TWI), TPI, annual temperature, precipitation, wind speed, Normalized Difference Vegetation Index (NDVI), distance from streams, relative humidity, and geology. On the other hand, anthropogenic factors are the distance from roads, the distance from settlements, the distance from farmlands, and LULC.

Various studies have been done on the factors affecting fire intensity and occurrence [2,5,31,32,33,34,35,36]. The geographical distribution of fires in central Spain has been studied, in terms of environmental factors. This study showed a significant relationship between the area of the burned regions and the topography of those regions. In the same vein, Lozano et al. [37] showed a high correlation between fire occurrence and environmental factors on different geographical scales. Zumbrunnen et al. [38] showed that climate, roads, and livestock play a significant role in fire occurrence in Switzerland. In addition, a non-linear relationship was established between fire occurrence and anthropogenic factors, with the fire occurrence rate being higher in warmer climates. In another study led by Justus [39], the vital factor affecting forest fire distribution is topography, followed by anthropogenic factors.

Humans cannot control fire completely, but they can reduce its intensity and the damage to natural resources, to a great extent. To this end, each area can be zoned, given the points available in terms of fire likelihood. In other words, areas with high fire likelihood must be identified; then, necessary measures should be adopted to prevent or fight fires [40,41]. Several studies have been conducted, so far, to model and assess fire risks in the world, using different methods [42,43,44,45]. These studies have used the analytical hierarchical process (AHP) [2,5,11,13,43,44,45], fuzzy methods [21,46,47,48,49], Analytical Network Process (ANP) [50], logistic regression [51,52,53,54], artificial neural networks [20,55], support vector machine c, and random forests [15,56] to model fire risks. These methods require accurate preparation and implementation of the produced models, for the factors affecting fire occurrence. Given the complexity of the fire processes in various regions and the effects of various factors on its occurrence, each of the mentioned methods has potential and limitations. Moreover, in many cases and places, researchers may not have the necessary efficiency for the accurate prediction and geographical zoning of fire risks [57]. Therefore, examining the capabilities of these methods and employing them in different regions can help with choosing the best and most effective method. Moreover, the use of the capabilities of the Geographic Information System (GIS) and remote sensing has facilitated the zoning and modeling of fire occurrence [2,12,14,47,58].

In recent years, frequent fires have occurred in the vegetated lands of the Iraqi Kurdistan Region. In fact, fires are among the major causes that destroy these vegetated areas, and few studies have been reported so far. Thus, given the limited natural resources, including shrublands, in this region, the importance of trying to protect them against fires is maximized. However, despite the occurrence of frequent fires, no research has been conducted on the factors affecting the occurrence, intensity, and distribution of fires in this region. Accordingly, in this study, the shrublands of the Qaradagh area were studied, as a sample of the Kurdistan Region forests. Our study investigates the effects of natural and anthropogenic factors on fire risks, by integrating satellite images, logistic regression, and GIS techniques. Therefore, fire risk potentials were mapped and modeled. Preparing the map of fire potentials using a suitable method plays an important role in understanding and predicting fire occurrence in vegetated areas of this region, thereby helping environmental agencies, farmers, and land managers to prevent and deal with this problem through better supervision of their natural resources.

2. Materials and Methods

2.1. Study Area

The Qaradagh area encompasses the southern part of Sulaymaniyah governate, in the Kurdistan Region of Iraq (Figure 1). It is an important area in terms of ecology, biodiversity, and recreation, extending along with the two-mountain series from the northeast to the southeast. The Qaradagh, Sagrma, and Gulan mountain series (anticlines) forms the southern and western parts of the region, while Baranan Mountain (anticline) forms the northern and eastern parts. The region’s altitude ranges between 337 m and 1853 m above sea level. This region has a Mediterranean climate [59,60]. According to data from the Tropical Precipitation Measuring Mission (TRMM), for the period of 1998 to 2017 [61], the average annual precipitation of the study area varies between 591 and 703 mm; whereas, the average annual temperature varies between 17 and 19 °C, for the period of 1979 to 2014 [62].

Fires constitute one of the most important challenges that threaten the ecosystem and biodiversity in the Qaradagh area. Through the available satellite images of fires (between May 2015 and October 2020), we surveyed the area and recorded (660) vegetation fires (Appendix B). As a result, 12.35% of the whole study area was exposed to fire during the mentioned period. Sagrma Mountain, at the western part of the study area, was exposed to fire five times, whereas Baranan Mountain, at the eastern part of the study area, was subjected to fire four times.

2.2. Material and Factors Affecting Fire Susceptibility

Many natural and socioeconomic factors affect fire susceptibility in vegetated and pastures of a region, with the impact of each of these factors being specific and different in each region [49]. However, due to the large number and high variety of the factors affecting fire occurrences and data availability, and, also, to facilitate the better modeling and understanding of the calculation process, part of these factors—the main factors affecting fire occurrences in shrublands and pastures—are selected. In the present study, the factors were selected after the comprehensive review of 57 papers (Appendix A) selected from the Scopus (Elsevier) database and related to fire susceptibility in forests and pastures in different countries, published between 2008 and 2020. Accordingly, 16 predictive factors, confirmed in most sources, were employed to produce a fire susceptibility map.

In general, the factors effective in fire occurrence in this study were divided into two categories: environmental and anthropogenic factors. Environmental factors include slope, slope aspect, elevation, Topographic Wetness Index (TWI), TPI, annual temperature, annual precipitation, wind speed, NDVI, distance from streams, relative humidity, and geology. On the other hand, anthropogenic factors are the distance from roads, the distance from settlements, the distance from farmlands, and LULC. From among the aforementioned factors, the three discrete factors, i.e., slope aspect, LULC, and geology, are classified into 9, 5, and 10 classes, respectively. Thus, each class in these three factors is considered a variable, and the total number of layers was 37 factors.

We selected the factors used in the previous papers at the rate of ≥10% (16 factors for this study). In other words, we eliminated factors decreasing the accuracy of vegetation fire susceptibility mapping and retained those increasing it.

The required and available information were employed, in light of the factors impacting vegetation fires that were discovered and mentioned (Figure 2). In total, 121 pieces of Sentinel-2 images were downloaded from the EO-Browser website, to map burned areas at their burning time, during the period 2015–2020. The Sentinel-2 imagery was utilized to determine the NDVI, using Sentinel Application Platform (SNAP) software (

Table 1. Sentinel-2 spectral bands used to determine the NDVI.

Bands	Central Wavelength (µm)	Resolution (m)
Band 4—Red	0.665	10
Band 8—NIR	0.842	10

). NDVI is defined based on Equation (1) [63]:

$$\mathrm{NDVI}=\frac{\mathrm{NIR} - \mathrm{R}}{\mathrm{NIR} + \mathrm{R}}$$

(1)

where NIR and R values are the infrared and red portion of the electromagnetic spectrum, respectively.

In addition, the Digital Elevation Model (DEM) of the Shuttle Radar Topography Mission (STRM), with 1 arc-second resolution, was employed to derive slope, slope aspect, elevation, TWI, and TPI maps. The TWI and TPI are defined, based on Equation (2) [64] and Equation (3) [65], respectively:

$$\mathrm{TWI}=\ln \left(\frac{\alpha }{\tan \beta }\right)$$

(2)

$$\mathrm{TPI}={\mathrm{E}}_{\mathrm{c}}-\left(\frac{1}{\mathrm{nM}}{{\displaystyle \sum }}_{\mathrm{i}\in \mathrm{m}}{\mathrm{E}}_{\mathrm{i}}\right)$$

(3)

where α is the area of the drained zone, and tan β is the slope angle in degrees. Where E_c is the elevation of the central pixel, E_i is the elevation of the grid within the local window, and n is the total number of surrounding points employed in the evaluation.

Moreover, GIS world imagery was used to create the road, settlements, and farmlands maps, to extract the maps of distance from roads, distance from settlements, and distance from farmlands. More details of the datasets used and their specifications are fully described and presented in

Table 2. Factors and sub-factors effective in fires as well as their source.

	Factors	The Method of Affecting Fires	Factors Used by Experts (%)	Source
1.	Slope gradient	Fire spread is higher on steep slopes, so that with a 10-degree increase in the slope angle, the propagation speed doubles [66].	54.24	DEM-30 m https://earthexplorer.usgs.gov/ (accessed on 13 August 2021)
2.	Slope aspect	Since southward slopes are warmer and drier than other areas, fire risks are higher on these slopes [67].	52.54	DEM-30 m https://earthexplorer.usgs.gov/ (accessed on 13 August 2021)
3.	Elevation	Under normal conditions, at lower and middle altitudes of an area, fires are much more probable to occur than at higher altitudes because of relatively higher temperatures, lower humidity, and ease of human access [12]. Besides, by an increase in the altitude, damage caused by fires is reduced due to the slower growth of trees compared to lands of lower altitudes and the lower accumulation of resin under the bark [68].	50.85	DEM-30 m https://earthexplorer.usgs.gov/ (accessed on 13 August 2021)
4.	Precipitation	Precipitation is an important factor contributing to the high humidity of fuels, so it is considered a negative indicator of fire spread. In fact, the higher the moisture content in a species’ tissue, the more heat and the longer time it will need to evaporate the moisture content, so the higher fire resistance will be. Therefore, the moisture content of the burnable matter is one of the major factors effective in fire occurrence [16]. Besides, rising precipitation in spring increases vegetation growth. This vegetation dries in summer and makes fire occurrence possible [53]. Thus, it is required to pay attention to the amount of precipitation, the precipitation season, and the system creating it. This is because precipitation accompanied by strong lightning can cause fires.	40.68	TRMM-NASA
5.	Distance from roads	When the distance from roads increases, fire risks decrease because of less traffic and human activity [47].	40.68	World Imagery-GIS base map
6.	Distance from settlements	There is a higher fire occurrence risk in the vicinity of settlements involving human activity [58].	40.68	World Imagery-GIS base map
7.	LULC	The effects of LULC are exerted by human activity. Thus, it uses in which there is more human activity with suitable conditions for fire, it is more probable to occur [69].	38.98	LULC https://earthexplorer.usgs.gov/ (accessed on 13 August 2021)
8.	Annual temperature	Temperature rises increase evaporation and transpiration, thereby drying combustible materials, which is considered one of the factors effective in fire occurrence [70].	38.98	Climate Data https://globalweather.tamu.edu/ (accessed on 13 August 2021)
9.	Wind speed	The higher the wind speed is, the higher the fire intensity will be. This is because when the wind moves the air, more oxygen is transferred to the burning environment [22]. In this respect, if the wind blows from the land, it will have a greater effect on fire intensity.	32.2	Climate Data https://globalweather.tamu.edu/ (accessed on 13 August 2021)
10.	NDVI	Since vegetation is the fuel itself, and NDVI shows the condition of vegetation, this factor has a strong effect on fire ignition and the spread of fire [23].	28.81	Sentinel-2 A https://earthexplorer.usgs.gov/ (accessed on 13 August 2021)
11.	Distance from streams	Human activity (particularly, camping along rivers and springs in summer, monitoring cropland, and, sometimes, visiting by tourists to this area, due to the river and the good weather), could affect vegetation fires. Thus, the increase in people’s presence along rivers is one of the serious threats [71].	23.73	World Imagery-GIS base map
12.	Geology	The parent rock determines the soil type. Soil is considered one of the major parameters in describing vegetation fires. Besides, it, indirectly, affects the entire environment of a given region [72]. The soil type demonstrates the effects of the texture and composition of soil substances on fire occurrence [15].	16.95	Iraqi Geological Survey (Scale 1: 250,000) GEOSURV-IRAQ
13.	TWI	The TWI shows the size of saturated areas of runoff generation and the effects of topography on the location. Thus, upon an increase in the value of this index, fire risks decrease [15].	16.95	DEM-30 m https://earthexplorer.usgs.gov/ (accessed on 13 August 2021)
14.	Relative humidity	In dry seasons, with an increase in the temperature as well as a decrease in relative humidity and dryness of vegetation, the probability of fire occurrence and spread increases [73].	13.56	Climate Data https://globalweather.tamu.edu/ (accessed on 13 August 2021)
15.	Distance from farmlands	Since many fires are created for clearing agricultural lands off crop residues or developing them, fire risks are greater near agricultural lands [28]. Traditionally, farmers used the above-mentioned method, which is not seen as a harmful method. On the contrary, it is viewed as a method that increases soil fertility. Nevertheless, the fire sometimes gets out of control, unintentionally.	11.86	World Imagery-GIS base map
16.	TPI	It displays the difference in height between a focal cell and all cells in the neighborhood [74]. In addition, TPI shows that flat areas are not as favorable for fire occurrence as ridges and gentle slopes [5]. Besides, the more positive or negative the curvature is, the more likely the vegetation-fire occurrence will be [4].	10.17	DEM-30 mhttps://earthexplorer.usgs.gov/ (accessed on 13 August 2021)

. Figure 3 and Figure 4 show the 16 predictive factors, which were used in this study.

Due to the low number of climatological stations in the study area, we used the TRMM monthly global precipitation data, from 1998 to 2017 [61], and air temperatures acquired from Climate Forecast System Reanalysis (CFSR), from 1979 to July 2014 [62]. The TRMM data within Iraq and the Kurdistan Region have been used and verified by several researchers, such as [75,76], where their accuracy shows a strong direct relationship with the climatological stations’ data (coefficient of determination (R²) > 0.8). Although the CFSR was not verified before, it was compared to the data obtained from Darbandikhan station, located southeast of the study area (Figure 4A). For this purpose, we used the nearest pixel that was located to the climatological stations. In Figure 4C,D, the R² indicates a strong direct relationship (>0.8) between the CFSR and the climatological stations’ data. The spatial resolution of the CFSR is ~38 km [77]. The TRMM data have a resolution of 0.1 degrees (~9 km). The TRMM and the CFSR data have been converted to vectors and interpolated with a 30 m spatial resolution.

The geological map was obtained from the Iraqi Geological Survey [78]. Whereas, the LULC and geological maps were derived from USGS EarthExplorer website and GEOSURV-Iraq [79].

2.3. Burn Scar Inventory

A map of past fires is required to investigate the relationship between fire locations and the factors influencing fire occurrence. Thus, the manual delineation of the burn scar inventory was created for 121 satellite images, for the period from 2015 to 2020, in addition to the field studies. These data have been delivered from Sentinel Hub services, which, efficiently, provide long-term analysis. Next, the locations of all burn scars, which occurred during this time interval, were observed on these images, using RGB:432 color composite. Finally, the burn scars were drawn as polygons, and were, thereafter, captured and used for modeling and validation purposes.

When all required data and information were collected from different sources, and after necessary corrections and processing were applied, all factors were converted to the raster format. Moreover, given the fact that the input layers have varied units, which are not suitable as direct inputs for the logistic regression, the input parameters were normalized within the range of 0 to 1 [80]. In the end, making use of the models and coefficients achieved from the logistic regression method, the maps were overlaid with fire risk maps of the study area produced.

2.4. Data Analysis by Logistic Regression

After the parameters were prepared, fire susceptibility maps were prepared, using the logistic regression method. Logistic regression is a statistical method devoted to the group of generalized linear statistical models, which predicts the probability of an event (fires) using independent variables. The key point in logistic regression is that the dependent variable is Boolean (a two-state variable that can be only number 0, meaning non-occurrence, or number 1, indicating the occurrence of an event). Concerning fire probability mapping, logistic regression is used to find the best model for describing the relationships between the presence or absence of a dependent variable (vegetation fires) and a set of independent variables. Logistic regression utilizes the maximum likelihood estimation (MLE) method to find the set of parameters fitting the model best. The model’s output yields coefficients between 0 and 1, which through fuzzy theory assigns value 1 to probabilities more than 0.5 (fire occurrence) and value 0 to probabilities less than 0.5 (non-fire occurrence). Finally, it produces a Boolean map for fire susceptibility.

We used the fitting generalized linear models (glm), with a binomial family and logit link in R. Logistic regression is used, by assuming that the probability of the dependent variable being 1 follows the logarithmic curve. Besides, its value is estimated by Function 4 [81], as follows:

$$P\left(y=\frac{1}{X}\right)=\frac{\exp \left(\mathrm{SBX}\right)}{1+\exp \left(\mathrm{SBX}\right)}$$

(4)

where, P is the possibility of being 1 for the dependent variables, X represents the independent variables, B is the estimated parameters, and y represents the dependent variables. To linearize the function above, the following transformation is applied:

$$lo{g}_e\left(\frac{P}{1+P}\right)=b_0+b_1{x}_1+b_2{x}_2+b_3{x}_3\dots +b_k{x}_k+error term$$

(5)

where, P is the possibility of being 1 for the dependent variables, x₁, x₂ …, x_k represents the independent variables, b₀ is the coefficient of the regression equation, and b₁, b₂, …, b_k are the coefficients of each independent variable.

This logarithmic change makes the predicted probability become continuous within the range 0 to 1 and makes the model’s output be presented as a geographical prediction map for risk probability [82].

2.5. Preparation of Training and Validation Dataset

In total, 103 scenes of satellite images were taken from Sentinel-2, 6 scenes from Landsat-8, and 3 scenes from Landsat-7 (Appendix B). The Landsat images were used to avoid cloudy days and cover the time before the Sentinel-2 launch in 2015. All burned locations, during the mentioned period, were merged to show all areas exposed to the fire on one map. We separated the burned area randomly into training and validation data subsets, as recommended in the literature [83]. A percentage of 80% (185,394 points) was used for the training dataset in the model and 20% (46,348 points) for the validation dataset. A similar number to the training points (185,394 points) from the burned areas were picked at random from the unburned areas. Hence, 185,394 points from the burned areas and 185,394 points from the unburned areas were entered into the model as a training dataset.

2.6. Omission of the Factors Less Effective in Fire Occurrence to Map Fire-Susceptible Areas

To map the fire-susceptible areas, all the training data collected from the 16 main factors were entered as a table into RStudio. The table includes a column that shows the case of the burn scar that exists. An existed burn was coded as 1 and a non-existed burn was coded as 0. The logistic regression model was employed to analyze the data and find the estimation constants (α and β) for probability calculation.

Accordingly, it was observed that in most of the studies, factors, such as slope, slope aspect, and altitude, were used. However, factors such as relative humidity, distance from agricultural lands, and TPI (less than 20% of the reviewed articles) were utilized in fewer studies. Apart from a map, in which all of the mentioned factors were used, we prepared other maps, from which we removed one of the less important factors from the total factors every time. Based on their number of frequencies in the studies, factors of less importance included geology, TWI, relative humidity, distance from farmlands, and TPI. Besides, another map was prepared, from which all six factors were removed. Less important factors were omitted, to stress the importance of the impact of the factors examined on preparing the fire susceptibility map. This is because the careful selection of factors is more important than the methods used for the purpose of zoning [75]. In the end, we will compare the aforementioned maps.

2.7. Validation

Several performance metrics, such as the Roc Curve, Mean Absolute Error (MAE), Relative Error (RE), and Percentage Error (PE) have been used to assess the predictive accuracy of the implemented model in the current study.

Validation or accurate evaluation is required in modeling, to determine the scientific significance or evaluation of a study [3,84]. The efficiency of the model presented in the present study was evaluated, making use of the receiver operating characteristic (ROC) curve and the area under the curve (AUC). This is the most common method used for evaluating modeling efficiency, being used in various research fields [85]. The ROC method is the graphical representation of the balance between the False Positive Rate and the True Positive Rate, for every value of probability. The area beneath the AUC-ROC curve indicates the value predicted by the model, by describing its capability of accurately estimating events or their non-occurrence. In general, an AUC value of 1–0.8 indicates very high efficiency, 0.7–0.8 indicates high efficiency, 0.6–0.7 denotes medium efficiency, and 0.5–0.6 indicates low efficiency, of the model.

In this study, 80% of the fires that have occurred in the study period were used as the training dataset and assigned to model production (Appendix C), while the remaining 20% of the dataset was assigned as a validation. In contrast, we used 20% of the dataset to predict the models (Appendix D), while the remaining 80% of the dataset was utilized to validate the results, to check the model’s overfitting.

We used MAE for evaluating the performance of a regression model (Equation (6)). It is a measurable indicator to validate the closeness of predicted values to actual values, as the name suggests [86]. The RE was calculated (Equation (7)), where the ratio of the absolute value to the approximate value is relative error. If the approximate error is unknown, the actual value is used [87]. Equation (8) used the reference value to calculate the PE [88].

$$\mathrm{MAE}=\frac{1}{n}{\displaystyle \sum }_{i=1}^n\left|P-O\right|$$

(6)

$$\mathrm{RE}=\frac{\left|P-O\right|}{O}$$

(7)

$$\mathrm{PE}=100*\frac{\left|P-O\right|}{O}$$

(8)

where n represents the represents the number of observations; O represents the actual value; and P represents the predicted value from the susceptibility model.

3. Results

3.1. Vegetation Fire Susceptibility Mapping

After preparing all the factors required and determining the geographical locations of the fires, which took place during the six years within the study area, logistic regression was utilized to determine the relationship between factors affecting fire occurrence and the related spatial modeling. The logistic regression relationship was established between fire occurrences, as dependent variables, and the mentioned parameters, as independent variables. Since the studied factors were selected from the literature review, the regression model was repeated seven times, once to validate the dataset and once, again, to train the dataset to examine the importance of selecting these factors (Appendix C and Appendix D). The p-values for all predictor factors used in all models are less than 0.05. Therefore, they have a statistically significant relationship with the fire occurrence in all models. In group 1, we selected those factors that were used in less than 20% of the studies in various regions, which were entered into the regression. In group 2, geology was omitted. Besides, TPI, TWI, and the distance from agricultural lands were removed in groups 3, 4, and 5, respectively. In group 6, relative humidity was omitted, and, finally, all respective factors were entered into the logistic regression in group 7 (Appendix C and Appendix D). As a result, seven zoning maps of fire susceptible areas were prepared. The importance of the order of the factors, based on the coefficients achieved in each of the seven cases, is according to the order of the following diagrams, with the maps indicating seven zoning times (Figure 5 and Figure 6).

3.2. Validation and Comparison of the Maps

Figure 7 shows the validation results for the fire potential map. This diagram shows that the success rate in all groups is quite apt, ranging from 84.85 to 97.68%. Among all groups, group 3, i.e., the group with TPI excluded from the model calculations, has the highest stability and success rates (AUC), which are 97.30% and 97.31% for both 20% and 80%, as training dataset evaluations, respectively (Figure 7). In addition, when assessing the 20% as a training dataset, group 7 had the highest accuracy (97.68%), compared to the findings of the other groups (Figure 7A), and it came in second place (96.72%) in terms of accuracy, when evaluating the 80% as a validation dataset (Figure 7B), which means less stability than that of group 3.

The results of the MAE, RE, and the PE of the seven groups show that the best model is group 7. It has a lesser error, while MAE, RE, and PE are 0.62, 2.32, and 0.64, respectively. The worst model (highest errors) come in group 4, where the MAE, RE, and PE are 2.54, 15.15, and 2.99, respectively (

Table 3. Additional statistical validation methods.

Validation Method	Group 1	Group 2	Group 3	Group 4	Group 5	Group 6	Group 7
Mean Absolute Error (MAE)%	1.19	1.01	0.68	2.54	0.86	0.76	0.62
Relative Error (RE)%	5.73	5.41	2.69	15.15	3.80	3.28	2.32
Percentage Error (PE)%	1.26	1.07	0.70	2.99	0.89	0.79	0.64

).

3.3. Key Factor

The predictive ranges of all seven models show that LULC classes, which are related to vegetated areas, are the most common factors that affect fire occurrence (Appendix C and Appendix D). This factor includes six classes, which, in order of weight importance, include: cropland, grassland, urban and built-up land, open shrub, and barren area. The wind speed factor has the second-most impact on vegetation-fire occurrence, followed by the geological formations factor, which is composed of clastic sediments (i.e., Tanjero, Aqra, and Bekhme formations).

4. Discussion

4.1. Model Comparison

This study aimed to examine the factors affecting fire occurrence by using the logistic regression method to zone fire occurrence in the Qaradagh area of the Iraqi Kurdistan Region. Based on the AUC of the seven models of the logistic regression (Figure 7), group 3, i.e., the group in which the factor of TPI was not considered, had the highest success stability rate (97.30% and 97.31%) for both evaluations. This reveals that it is not necessary to include the TPI factor in the applied models, to find the susceptible areas to fires. This group ranks second, behind group 7, based on the accuracy results from the MAE, RE, and PE methods, which indicates the effectiveness of all the selected criteria in determining the sensitive areas to fires, as confirmed by the outputs of Figure 7A (97.68%). After groups 3 and 7, the accuracy of all used methods revealed that the values of groups 6 and 5 (excluding relative humidity and distance from agricultural land) were stable, indicating that the effect of the factors on vegetation fires was maintained at the same level and sequence. In terms of stability and sequence, groups 2 and 1 were similar to groups 6 and 5, although with less accuracy, as shown in Figure 7 and Table 3. As for groups 2 and 1 (group 2 with the geology factor removed, and group 1 with all factors that were used in less than 20% of past case studies removed), they were similar to groups 6 and 5 in terms of stability and sequence, but with less accuracy. The low level of accuracy is due to the importance of the geological factor, and its impact on the spatial distribution of plants that are exposed to fires. The accuracy of all the applied statistical methods shows that the least accurate model was represented by group 4, from which the TWI factor was removed. This highlights the effectiveness and strength of the impact of this factor on the accuracy level of the model, which must take into account the extent of its importance. As demonstrated, the values of the statistical accuracy methods show the very strong performance of the models, as well as the selection of apt factors, for preparing a fire zoning map in the studied area. However, it is crucial to be careful enough, when selecting factors affecting fire occurrence. The low success rate in group 4 implies that the factors being more important in fire zoning should not be removed to increase the model’s success rate. This methodology is also applicable to other areas with similar geographical and environmental conditions, and decision-makers could benefit from such an approach.

4.2. Factors Impacting Vegetation Fire

The spatial relation strength of affecting and impacting factors on the event was shown by the logistic regression model. It determined the relative importance of each factor influencing the vegetated fire in the study area. Hence, it revealed that the increase in wind speed, NDVI, and geological formation, which includes vegetation, as well as LULC class, which is related to vegetated areas, are the most common factors affecting fire occurrence (Appendix C and Appendix D). According to the factor weight results (Figure 5, Figure 6 and Figure 7), classes of LULC and, then, wind speed gained the highest weight among all groups, with their weight being significantly different from that of other factors. Among LULC classes, the highest weight has been assigned to the cropland class, which can be considered the key factor that enables us to better anticipate and, even, prevent land fires, requiring, therefore, some proper land management practices such as reducing fuel loads periodically and, also, installing firebreaks. Similarly, the results of Pourtaghi et al. [34] show that LULC is the most effective factor in fire occurrence. The presence of adequate fuel is the main cause of most fires occurring in pastures, forests, and agricultural lands. Apart from natural fires in these areas, the role of anthropogenic factors trying to eliminate agricultural residues, repel destructive pests from agricultural products, and speed up the revival of grass species in pastures by creating deliberate fires, is noteworthy. Besides, some unofficial reports relate many wildfires to ethnic disputes, which need more in-depth investigation. Given these reasons, and considering that the highest weight in all seven groups was related to the agricultural land class, one can say that humans are the major cause of fires in this particular region.

After fire occurrence, natural factors can positively or negatively affect fire spread. For example, wind speed is one of the most effective factors in its spread. Besides, wind affects both the fire triangle of fuel, heat, and oxygen as well as the fire behavior triangle of fuel, topography, and climate [89]. Wind, in the fire triangle, transfers more oxygen to the fuel, thereby causing the initial fire spark. On the other hand, the climate factor is seen in the fire behavior triangle, with wind being one of the main components of climate. Wind spreads fire further to surrounding areas and makes its extinguishment more difficult.

The data about wind direction in the study area are poor, as they are missing detailed information. Local winds dominate topographic fires more than prevailing winds, due to slope and differences in the solar heating of the Earth’s surface. Forest fires are more sensitive to small changes, so slight differences in topographic winds, slopes, or sides have a greater impact on fire behavior [90]. From the general information, the effect of wind direction on vegetation fires in each pixel is close to equal, in areas characterized by an almost equal geographical distribution of vegetation density; therefore, we ignored the wind direction factor. In the future, we will consider the wind direction factor once local measurements become available for better spatial analysis.

High temperatures, as well as low precipitation and relative humidity, dry the environment, thereby providing suitable environmental conditions for fire occurrence and spread. In regions with these attributes, in case there are combustible materials with, usually, a small amount of humidity, fire occurrence potentials are very high [91]. Due to the fact that two climatic factors, temperature and precipitation, are highly effective in causing fires in terms of time, and given the investigation that was made in the studied area, they cannot cause significant climatic changes in a habitat on such a limited scale. Research on the correlation between fires and soil types indicates that the soil moisture content and its effects on the diversity of plant species are the most effective factors in controlling fire occurrence or non-occurrence [34].

The factors of the distance from roads, residential areas, and rivers reflect humans’ role in fire occurrence [92]. These factors can play a dual role in fire occurrence and spread. In the vicinity of these factors, due to more human traffic and activity, fire occurrence is more likely. However, the presence of these factors can be effective in preventing fire spread, after its occurrence. Roads can act as firebreaks, by preventing the fire from spreading to surrounding areas. As fire can be contained by humans, in case it occurs in the vicinity of residential areas, it will be extinguished sooner due to easy human access. In the present study, locations of residential areas were considered as a polygon, not a point, and, then, a map of the distance from these areas was prepared. In other words, one could say that population density has been somehow considered. Accordingly, the higher the population density is, the higher the fire probability will be. In the vicinity of rivers and waterways, due to the higher humidity of combustibles, the fire spread rate decreases [45].

The TPI implies that flat areas are not as susceptible to fire as sloping areas. Moreover, fire occurrence is higher in convex slopes than in concave slopes. This case can be justified, both at the beginning of the fire and at the time of its spread. Humidity is lower in sloping areas, so the fire spread speed increases with increasing the slope angle after fire occurrence. The TWI determines the effects of topography on the rate of saturated surfaces for producing runoff, which plays a significant role in the soil moisture content and slope stability. Hence, there is an inverse correlation between an increase in this index and fire occurrence and spread [93].

The high density of the Gramineae species results in increased vegetation density and the supply of fuels effective in surface fires. Moreover, fire statistics indicate that the highest frequency of fires occurs almost from early summer to early autumn. Given the increasing temperatures, decreased relative humidity, dryness of vegetation in summer, and concomitant effects of human activities, fire is intensified in the region, being consistent with the studies of Gholami et al. [94] and Rafiee et al. [95].

In the Qaradagh area, further work is necessary to estimate the local biomass and total carbon levels, which, in turn, can give an idea of the health of the available crops that could feed random fires. Therefore, we suggested using the method applied by Oliveira et al. [96] to integrate the field-observed biomass data with lidar data from various sensors, to estimate the biomass and carbon levels obtained in the dry vegetated land within our study area. Moreover, lidar metrics can be used to predict wildfires behaviors and risk assessment measures, by combining lidar data with Landsat-derived indices [96,97] to map canopy cover, canopy height, canopy base height, and canopy bulk density as well as to evaluate the aboveground carbon density through shrubland allometric relationships.

Furthermore, the addition of hyperspectral data can be beneficial in terms of distinguishing the moisture content of crops more accurately [98], which, in turn, can act as a predictor of fires rate. The implementation of such data, in addition to our model, will provide holistic yet easy-to-apply methods for agriculture managers, which will aid them in their production rates while minimizing losses through raising awareness on how to combat wildfires.

4.3. Advantages and Disadvantages

One of the major merits of this study is providing information about the statistical weight of every individual variable on the overall result. Moreover, the high accuracy achieved in the present research could have been due to the use of the majority of the factors affecting fire occurrence, which, in turn, allowed the preparation of a more accurate fire susceptibility map. The model’s limitation is that the factors impacting the vegetation fire differ in space and time; in other words, the model cannot be applied everywhere, at any time, using the same factors influencing the event. Therefore, this model can be applied in arid and semi-arid mountainous regions. For other regions that have different environments and topography, the priority should be on careful choice of the predictive factors, which is far more significant than the methods utilized. Interestingly, multitemporal measurements of Synthetic Apertura Radar (SAR) and Light Detection and Ranging (LIDAR) are suggested for further studies, to estimate carbon release and better assess soil erosion and vegetation recovery.

5. Conclusions

Fire events, caused by natural or anthropogenic factors, are a major threat to landscape resources, so they must be controlled and prevented. To minimize this threat to forests, managers should be fully knowledgeable about past fire processes. Accordingly, spatial analysis of fires must be performed in every region, and conservation strategies should be formulated for every region, after a thorough understanding of temporal–spatial processes. The prerequisite for such endeavor is the ability to make predictions, perform spatial analyses, and prepare a fire susceptibility map. The use of the logistic regression method is one of the risk-modeling methods, which helps to determine the relationship between independent and dependent variables. Additionally, combining remote sensing data with geographic information data and statistical models permits fire managers and personnel to predict ‘where and when’ vegetation fires will occur.

In the current study, a fire susceptibility map was prepared for the Qaradagh district, using the logistic regression method. To examine the importance of the factors selected, the regression model was repeated seven times. Each time, one of the factors used less frequently in previous studies was omitted. Moreover, all of the factors were entered once into the regression for modeling. The AUC value was shown to be higher than 96%, in all groups. This indicates the remarkable performance of the model and the selection of suitable factors, for preparing a fire zoning map in the study area. According to the results, to increase the model’s success rate, factors less effective in fire zoning should be removed. Based on the coefficients obtained, it was found out that the land-use factor exerted the greatest impact on fire occurrence, among all seven cases, in the current study area.

This approach can be applied to other areas having the same geographical conditions. However, this requires selecting the variables and their weights that work for the tested region, based on its environmental conditions.

More detailed land cover classification (particularly classifying other vegetation types and trees) can increase the understanding and accuracy of vegetated land fire prediction and is encouraged for future studies. Finally, we suggest considering the Qaradagh district as a natural protection area, since forest resources are rare and vegetation fires are high in the Kurdistan region.