Wildre Severity Zoning Through Google Earth Engine and Fire Risk Assessment: Application of Data Mining and Fuzzy Multi-Criteria Evaluation in Zagros Forests, Iran

The arid and semi-arid regions of Zagros forests in the Middle East are constantly exposed to wildre due to ecological conditions, and support systems are inecient in controlling wildres due to managerial and social weaknesses. Remote sensing and assessment tools are suitable for rapid prevention and action to identify the severity and location of a wildre. This study investigated the natural resource management of Zagros Forestry in terms of protecting wildre and combating forest wildres using the NASA re spatial data and the wildre severity in the Google Earth Engine (GEE) platform. The land-use of the study area is produced by applying the Random Forest (RF) classication method and data from the Sentinel 2 satellite imagery for 2019. To separate the types of cultivation and vegetation of the region, the method of extracting the average vegetation index of the seasons is extracted from GEE. To evaluate re risk, eleven human and ecological factors and two assessment models are applied to classify the probability re risk therein. Furthermore, the outcome of AUC conrmed the Logistic Regression (LR) model; the accuracy of the LR (AUC=0.875049) model is satisfactory and is suitable for re risk mapping in Zagros Forestry. Six high-risk areas of the wildre were identied by MOLA, which overlap with protected areas. Out of a total of 20469.17 Ha of wildre, 10426.41 Ha belong to these protected areas. 3826 Ha of this area were in the forests of Amygdalus spp, Quercus brant ii, pistacia Atlantica, and Quercus Infectoria, and 6600.41 Ha of it were in rangelands. Accordingly, an executive order was developed for the decision support system that reduces the risk of wildre and helps extinguish the wildre.


Introduction
Forests are the main natural resources and are an indicator of the prevailing ecological situation in the area. Forest ecosystems are constantly changing. These changes can be due to human development or part of the evolution of nature itself (Dimopoulou and Giannikos, 2001). The considerable damages due to wild res are directed to the environment, human health, and property (GS, 2003). Wild re as the most important disturbing factor in ecosystems leads to the most dramatic changes in the structure and function of forest (González-Pérez et al., 2004). The degree of degradation of the ecosystem and its function due to wild res as opposed to landscapes as differences in the severity of wild res from local to regional scales, and this wild re-induced ecological change, is a major focus of many studies worldwide (Naderpour et al., 2019;Parks et al., 2014). These studies often depend on network metrics that pre-and post-re images use to estimate the rate of change caused by re, and the most common matrix is the normal delta burn ratio, which is used to calculate dNBR, RdNBR, and RBR, and for large processing, it is better to use GEE (Key and Benson, 2006). There exist many different methods and models involved in evaluating forest re risk (FFR) in different areas at different scales and different e ciencies. In some studies, the Dong model is applied in predicting high-risk re areas in the forests (XU et al., 2005;E et al., 2004;Eskandari et al., 2013), and in some, the Analytic Hierarchy Process (AHP) or fuzzy sets are applied in modeling FFR (Chuvieco and Congalton, 1989;Vadrevu et al., 2010;Sowmya and Somashekar, 2010;A et al., 2012).
The forest wild re process is typical, nonlinear, and complex, and it is in uenced by many ecological and human factors. This fact, in turn, makes the task of seeking high accuracy prediction modes di cult (Pettinari and Chuvieco, 2017). In the research that has been done in this regard, Ngoc-Thach, Nguyen et al. (2018) ran a study where the advanced machine learning models like Support Vector Machine classi er (SVMC), Random Forest (RF), and Multilayer Perceptron Neural Network (MLP-Net) were applied. They, rst, established a GIS database of 564 forest re locations and then considered ten variables for the study area. Next, they applied the Pearson correlation method in assessing the correlation between variables and forest re and then applied the MLP-Net model (Pourghasemi et al., 2020). Using three machine learning algorithms, satellite imagery, and ten in uential factors, they have modeled, predicted, and evaluated the accuracy of the models in South Zagros (Nami et al., 2018a). By analyzing spatial patterns and ve tree-based classi cations, decision making includes alternating decision tree (ADT), classi cation and regression tree (CART) (Gayen and Pourghasemi, 2019), functional tree (FT), logistic model tree (LMT) (Kim et al., 2018), and Naïve Bayes tree (NBT) for wild re pattern, and the ADT classi er performed best (Jaafari and Pourghasemi, 2019) The Frequency Ratio (FR) and AHP models are applied for FFR mapping in a comparative study run on Melghat Tiger Reserve forest, central India by Kayet et al. (2020); the results obtained from applying FR and AHP indicate that though the trends were similar, FR model has signi cantly higher accuracy compared with the AHP.
According to studies on wild re assessment and modeling, researchers have only been looking for a way to assess wild re and compare their methods, the results of which only examine the accuracy of the models. In the present study, there is an accurate method of extracting the burnt area where the wild re severity was estimated, and by applying the normalized difference vegetation index of Landsat satellite imagery, before and after, the wild re is discussed in Google Earth Engine platform (Parks et al., 2018). Using MCE fuzzy (Eskandari and Miesel, 2017;Kahraman et al., 2014) and LR (Pourtaghi et al., 2016;Were et al., 2015;Satir et al., 2016), wild re and risk zoning is done (Guo et al., 2016). Multi-Objective Land Allocation (MOLA) (Canova, 2006) was used to identify high-risk areas, investigate the extent of plant species degradation, and provide management strategies to combat and prevent wild re. Assuming that satellite images and spatial data can be used to extract the severity and risk of wild re, using the nal risk map, it is possible to extract high-risk areas of wild re according to the history of the wild re and the importance of the area. Which machine learning method offers better decision making?
How can potential protection zones be extracted using decision support methods?

Study area
The study zone is the Zagros Mountain chain in West Iran (at 46 ° 28' N to 46 ° 22' N and 45 ° 52' E to 47 °5 8' E), one of the sub-basins of the western rivers, which is covering 1342387.14 Ha (Fig.1). The climate there is semi-arid and Mediterranean with a temperature average of 15.6 °C. The annual precipitation mean is 503 mm ( Table 1). The vegetation is of semi-arid type in the sparse distribution of trees and short weed and grass (Sadeghifar et al., 2020). From 2012 to 2019, 622 wild res are recorded in 1840 points in the wild re information resource management system in the study zone (NASA FIRMS, 2019[1]).
In this study, according to the set goals, questions, and hypotheses, the conceptual framework was designed in 6 stages according to Fig. 2. In the rst step, information was collected from responsible sources and organizations. The second step is to process the information received or extract the information. In the third step, LR and fuzzy MCE, Wild re zoning models were performed and then the models were evaluated in the next step. The fth step was to assess the risk of wild re with selected model.
Step six is to select high-risk locations and develop management scenarios.

Data gathering
Identi cation of factors involved in forest wild re is essential in constructing a model to assess its re risk. In this study, these factors are extracted from the available studies and the formal reports of the local state authorities. In general, forest wild re has to do with the climatic conditions, vegetation dryness, zone topography, and human activities (Eskandari, 2017;Hong et al., 2017;Valdez et al., 2017).
The NDVI, slope, aspect, and land use constitute the important variables necessary to be addressed in evaluating the re risk (Nami et al., 2018b;Parisien et al., 2012). Moreover, the data on land cover and human accessibility are contributive to the analyzing process, because human is in uential in the spatial setup and the frequency of forest wild re. By manipulating nature in this case, humans make natural vegetation vulnerable to wild re occurrences (Parisien et al., 2016). Accordingly, eleven parameters in uence forest wild re: slope (%), NDVI, wind speed (km/h), precipitation (mm), temperature (°C), land use speci cation, distance from road (m), distance from cities (m), distance from villages (m), aspect, and Mean Sea Level (MSL) (m) (Fig.6). For NDVI and land use, these factors are measured through the GEE. For a more detailed survey of vegetation and land use, in this article, the NDVI is calculated in two ways: 1) the last growing season of the study area (Rouse et al., 1974) (Fig. 4) and 2) the variations in vegetation coverage. The average NDVI of four seasons in the region is calculated and extracted as the NDVI of the seasons (Link) (Fig. 3). Consequently, the types of vegetation that have grown in the area over a year can be identi ed. This method can be adapted to identify and classify the type of cultivation and vegetation segregation. To calculate land use, the Sentinel 2 satellite imagery, NDVI seasons, and Landsat Urban product in GEE are applied (Pesaresi et al., 2015). All the data in this study are extracted at 30 meters resolution. The land-use consists of six factors including forests, water, bare land, grassland, and urban and agricultural classes, which are obtained through the RF algorithm (Li et al., 2020) (Link) (Fig. 5).

Fire severity determined through the Google Earth Engine
Naturally, every wild re incident has its address and time, which is recorded by the authorities. The GEE is applied to advance the speed of process. The date before and after every wild re incident and its location are speci ed through Landsat and or Sentinel 2 images (Mallinis et al., 2018). The normal burn ratio (NBR) is applied in designing the highlight burned areas and estimates the severity therein (Key, 2006). The NIR and SWIR wavelengths are applied in extracting the wild re scope. The fresh vegetation before the re is of high NIR and low SWIR responses, while the opposite holds true in recently burned areas.
NBR is measured for both pre-re and post-re. To obtain the difference NBR (dNBR) image (Miller and Thode, 2007), the latter is subtracted from the former (Gibson et al., 2020). According to Veraverbeke et al. (2010), dNBR is applied to assess burn severity, where the higher the dNBR volume, the more severe the damage, while vegetation regrowth is evident in areas with negative dNBR volumes. The dNBR can be classi ed according to burn severity ranges proposed by the United States Geological Survey, Table 2. High Severity +660 to +1300 +0.660 to +0.1300 The mathematical interpretation of Table 2 content is expressed through Eq. (1) introduced by Keeley (2009) and Eqs. (2) and (3)  ). This code is de ned for both Landsat 8 and sentinel 2 images and is applied as needed (Fig. 7) (Link).

Random Forest classi cation
The Random Forests (RF) classi er is a machine learning technique proposed by Breiman (BREIMAN, 2001), widely applied for classifying, regressing, and evaluating input factors with relative importance (Yu et al., 2017). The RF is an ensemble of learning approaches where a set of decision tree classi ers are developed to make prediction(s) (Belgiu and Drăguţ, 2016).. Consequently, different sub-datasets are generated by replacing the training dataset in a random manner, where each sub-dataset is applied in constructing a decision tree by the Classi cation And Regression Tree (CART) algorithm (Breiman and Friedman, 1984).

Logistic Regression (LR) Analysis
This analysis is frequently applied for prediction and explanation of the caused re by humans. The binomial logistic is performed through this regression. In this process, the input dependent variable must be binary in nature, with possible values of 0 and 1. The LR analysis is usually applied to estimate a model describing the correlation among one or more continuous independent variable(s) and the binary dependent variable. In this study, LR analysis was performed using TerrSet software 18.07. Parameters used in the model: slope (%), NDVI, wind speed (km/h), precipitation (mm), temperature (°C), land use speci cation, distance from road (m), distance from cities (m), distance from villages (m), aspect, and Mean Sea Level (MSL) (m).

Fuzzy Multi-Criteria Evaluation (MCE)
MCE is a decision support tool, based on the criterion. The basis for a decision is known as a criterion. By applying MCE, it is sought to make a combination of criteria to nd a single composite basis for a decision, with a speci c objective orientation. In this context, these developed criteria might be variables like proximity to roads, slope, exclusion of reserved lands, etc. The appropriate images may be combined with the MCE to form a single proper map from which the nal choice will be made (Bonissone and Decker, 1986).
These criteria may combine both the weighted factors and constraints. Each one of the fuzzy evaluation factors is within 1 to 255 range. The Weighted Linear Combination (WLC) is obtained by multiplying each one of the evaluation factors in AHP weight derivatives. To obtain the relative weight of each factor in the multi-criteria evaluation, the AHP weighting method is applied. The AHP weighing table was applied to this model after completing a questionnaire provided to ecologists, forestry experts, managers of environment, and natural resources. The most effective Wild re weights for NDVI were land use speci cation and wind speed. To assess the stability of AHP weights it CR, should be measured. If this rate < 0.1, the validated is acceptable. In this study, the obtained CR is 0.07, an acceptable one. The CR consistency index is calculated through Eqs. (4) and (5),(Finan and Hurley, 1997): Where, CI is the compatibility index pairwise comparison matrix, CR is the consistency rate, λ max is the maximum eigenvalue judgment matrix, RI is the random index and n is the compared components' count in the matrix (Ergu et al., 2011).

Models validation
Relative operating characteristics (ROC) are applied to validate the evaluation models, which are proper in assessing the validity of a model that predicts the location of the occurrence of a class by choosing an appropriate image for depicting the likelihood of that class occurrence.
AUC volume is applied to calculate the Area Under the Curve (Pontius Jr and Batchu, 2003), which is constitute the output of ROC. AUC value at 1 indicates that there is a perfect spatial agreement between the class map (range of re severity produced from GEE) and the appropriateness map (models output). An AUC volume of higher than 0.5 is acceptable for model validation (Gilmore Pontius and Pacheco, 2004).

Multi-Objective Land Allocation (MOLA)
MOLA is a procedure for solving single and multiple objective land allocation problems. As to multiobjective land allocation problems, a compromised solution, according to the data extracted from a set of appropriate maps, is determined by MOLA, for each objective. This solution would optimize land appropriateness for each objective, according to the assigned weights, therein. To solve the allocation problem, the user can specify either the area or maximum budget requirements. There exist options to force contiguity and compactness. The suitability maps are usually extracted from MCE. The single objective of land allocation procedure of MOLA is to solve a single-objective allocation problem. Based on the information from a single objective, or appropriateness map, the best solution given the speci ed constraints is determined. For both procedures, the user can specify spatial objectives like contiguity and compactness, non-spatial constraints like areal requirements for the objective, and maximum budget requirement based on the land price. In this study, the re risk map obtained from the logistic regression analysis method and the average models method as the base input is fed into the single objective land allocation method by applying the TerrSet s/w. For the nal location and protection prioritization, according to the maximum re in the area, the proper area is calculated and assigned. In this study, three spatial sites are applied to logistic regression analysis and average models. The spatial sites are prioritized based on the highest value of map pixels, according to the spatial location of protected areas and the type of coverage, under the Environment Organization's supervision.

Wild re severity
The range of all wild res was extracted from 2012 to 2019 using GEE, as shown in Table 3. After running and analyzing the models, in the data mining method, the logistic regression analysis model with AUC = 0.875049, (Appendix 1) and Fuzzy multi-criteria evaluation model with AUC = 0.584645 (Appendix 1) were obtained. The MCE model was rejected due to low AUC. To create a wild re probability map, LR model wAS classi ed using the natural-break classi cation method. Fig. 8 shows the percentage of the class area of probability wild re in each model. In the LR and MCE models, most of the areas are related to the Moderate-high severity class.
Furthermore, by comparing the class of wild re severity in burned areas with the classes of evaluation models, it is shown that the highest area and wild re occurred on the Low severity class (Fig. 9).

Fire risk maps
A risk map from the LR model was produced, which had the highest AUC. Then, using the number of wild res and the point density command in Arc GIS software, the wild re density map was created as a re severity map. By multiplying the LR model by wild re severity, a re risk map is produced. The re risk map is classi ed as High Risk, Miderate-high Risk, Moderate-low Risk, and Low Risk (Fig. 10).
The area of re risk classes is shown in Fig. 11. Using the number of wild res, the percentage occurrence of wild res was assessed according to Fig. 12, with the highest number of wild res occurring on the Lowrisk class. Due to the high area of this class compared to the density of the number of wild res, the risk is low. However, in both models, the high-risk class was less than 15%, and the risk of wild re is high due to the low area and high density of the wild re.

Selection of high-risk areas using MOLA
To identify the area's vulnerability to wild re and propose decision making and managerial procedures, given the largest area burned in recent wild res and using re risk map, six areas were selected by the MOLA model as high-risk areas (Fig. 13). The proposed areas are in the vicinity areas such as Kosalan, Bozin, and Marakhil Touran which protected zones by the environmental protection agency (Table 1). So here, there are a variety of animals and protected species that will be at risk of death during a wild re. By examining the buffer at a distance of 10 km from these areas, it was determined that out of a total of 20469.17 Ha of wild re, 10426.41 Ha belong to these areas. This indicates that 50% of wild res have occurred in this area (Fig. 14). The proposed areas were evaluated according to environmental factors and were classi ed as moderate-low value, miderate-high value, high value, and very high value in terms of protection importance.

Destruction rate of vegetation
To assess the vegetation type of burned areas using GIS data from Forests, Range, and Watershed Management Organization, Table 1, the extent of vegetation degradation in forest and rangeland species was investigated.

Forest species
The total area of forests in the study area is 166808 hectares, of which 6,264 hectares were burned in wild res between 2012 and 2019, and the details are shown in Fig. 15.
By assessing the distance of 10 km from the protected areas, it was determined that the total area of forests in this area is about 74,000 hectares, of which 3826 hectares were burned in wild res between 2012 and 2019, and the details are shown in Fig. 16.

Rangeland species
The land cover of the study area contains scattered forests with dense rangelands background and 13 species. The rangelands area is 408233 hectares, of which 10136 hectares were burned in wild res between 2012 and 2019, and the details are shown in Fig. 17.
Rangelands area in 10 km of protected areas is about 107857 hectares, of which 3285 hectares were burned between 2012 and 2019, and the details are shown in Fig. 18.

Discussion
In this study, we have tried to develop re risk assessment models and wild re probability zoning, and also help improve the natural resource management process by bringing the results closer to reality. By resorting to the documents and data available in domestic and international organizations and applying remote sensing techniques together with the algorithms available in GEE and satellite images of high spatial distinction ability this assessment is accomplished. The effective features in this context are selected carefully and are weighted. Among the available methods and classi cation techniques, the most valid and accurate ones are applied in evaluating the potential risks in wild re occurrence. The strong point of this study in relation to its counterparts consists of 1) applying the GEE platform with a vast supportive data, high processing speed, reduced human error coe cient, and accuracy in results and 2) applying accurate evaluating and location detecting methods, multi-criteria evaluation, and neural network. High-risk wild re is overlapping with protected areas under the support of the Environment Organization (Kozlan, Bozin, and Marakhil Turan), and 50% of all wild res have occurred within a 10-kilometer range of these areas. These areas include forest species such as Amygdalus spp, Quercus brant ii, Pistacia Atlantica, Quercus Infectoria, and pastures (Grasses, Forbs, Astragalus, Acantholimon, Psathyrostachys, and Daphne) and animal species such as 117 species of birds, 23 mammals, and 17 species of reptiles. Furthermore, by examining the type of vegetation in the whole region, the burned areas, and the protected areas, it was determined separately which of the plant and forest species are endangered.
Due to the existence of Quercus forests in this region, restrictions on human non-presence in the last few decades have led to the accumulation of a high volume of dry weed and grass, i.e., high potential to wild re occurrence. According to the obtained results, in order to accurately assess the probability and risk of wild re, it is necessary to carefully examine the natural and human factors in the wild re and use wild re zones to introduce the wild re sample to the evaluation models and it is recommended to have serious controlling measures here like in-situ wild re extinguishing services, watchtowers, etc. In addition to online intelligent wild re sensors, communication stations and properly designed over ground connecting networks/paths are recommended. Given that in recent wild res, a number of volunteer re ghters have been killed, staff-training, modern equipment, aerial wild re extinguishing equipment like helicopters, and arti cial water reservoirs-resources constitute the major components in this context. Training the neighboring rural and urban population and wildlife tourist guides can be a preventive measure in wild re occurrence prevention. In a similar study, Halofsky et al. (2020) assessed the wild res in the NW Paci c Ocean forests with respect to climate change and found that they occur due to warming and humidity reduction in weather. Naderpour et al. (2019) recommend planting trees in colder and more humid micro-sites to protect species on the verge of extinction. Combined methods based on GIS to model forest wild re and their classi cation into statistical data-oriented models yield more accurate results. Among these, the data-oriented methods are the most common methods (Parks et al., 2018). Pourghasemi et al. (2020) introduced land use, precipitation, and slop as the criteria in wild re intensity according to Landsat to extract the dNBR, RdNBR, and RBR, which can be practical in evaluating wild re. Hajehforooshnia et al. (2011) andParks et al. (2014) used multi-objective land allocation (MOLA) to identify priorities and sensitive areas for the shelter during a study to expand the Qomishlu Wildlife Sanctuary.

Conclusion
The results of this study indicate that 12% of the study area is forest and 30% is rangelands. 1.52% of the total study area is affected by wild re, which includes 3.7% of forests and 2.5% of rangelands. According to the objectives of this study, the risk assessment model was selected according to the AUC coe cient, and high re risk areas were identi ed using the MOLA model. Due to the overlap of MOLA results with protected areas, these areas were selected as hotspot wild res, accounting for 47% of forests and 26.42% of the region's rangelands, and 50% of all wild res in the region have occurred in and around these areas, which is an answer to the assumptions and questions of the present study. Given that the protected areas are exactly on the border between Iran and Iraq, the choice of high-risk areas of wild re as a re ghting base could cover the protection of forests and rangelands internationally. One of the limitations of this study and similar studies at the time of the wild re, in order to extract the re zone, is the time interval of at least 15 days that we have to wait to receive satellite images after the wild re. With the method presented in this study, researchers can assess the extent and severity of wild res in the shortest possible time on a large scale. In addition, the online decision support system can be developed for use on a variety of scales and times.