Hazard management requires intermittent and frequent monitoring of hazard factors (Tingsanchali & Karim, 2010); therefore, classification of wildfires, floods, and landslide hazard indicators is necessary for spatial hazard modeling (Tehrany et al., 2019). The Jack Knife test in the ME model is a method of determining the importance of variables as an alternative method in predictive models that works by deleting independent variables, which is executed by the number of variables (n), in which an independent variable is deleted (n -1), and model evaluation is performed to determine the importance and position of each variable in the model (Baldwin, 2009). The results of Jack Knife ME indicate the high importance of climatic and human factors in the occurrence of fire hazards in GNP; therefore, it can be said that these factors determine the occurrence of fire. Among these, the rainfall factor had the highest impact, and this study revealed that areas with rainfall less than 260 mm and more than 460 mm per year are most likely to have fires. Rain can affect the occurrence of fire in two ways. First, increasing the humidity in the area reduces the chance of fire. Therefore, the probability of wildfire in low rain areas is higher, which describes the increase in the risk of wildfire in areas with less than 260 mm yearly rain. Second, rainfall can affect the growth of annual and eventually perennial grass cover, which influences wildfire; in this case, the high probability of wildfire in areas with more than 460 mm annual rainfall can be attributed to the dense vegetation cover resulting from the abundance of water.
The graph of the average annual temperature shows the effect of increasing temperature and a significant increase in the probability of wildfire at temperatures above 17.5°C. Such a high temperature is observed in the northern and northeastern parts of the national park near the Qaraqum Deserts in the north of the national park. (Rahimi & Khadem, 2018) analyzed the synchronous patterns of wildfire hazards in the Golestan province forests and pointed to the adverse effects of hot desert winds from Qarahqum on wildfire events in the region. Results from the humidity variable also show that in areas with low humidity, and hence poor rangeland cover, the risk of wildfire increases (Faramarzi, Hosseini, Pourghasemi, & Farnaghi, 2021).
Vegetation factors are important variables that affect wildfires. The NDVI variable increases from west to east of the national park. So since the cover ranges from west to the center of the national park, which is the beginning of the Hyrcanian forests, there is an ecotonic cover with a variety of plant species and grass cover, the graph of this variable shows an increasing risk of wildfire in this range. In fact, the diversity and dense cover of annual and perennial forest in this area has increased the risk of wildfires. The most important land cover variable is related to shrub cover, which has been completely burned, and a low canopy forest is the least important in the occurrence of national park wildfires. This cover, despite being located at heigh elevation, is less vulnerable to human and therefore less exposed to wildfire.
Regarding the high importance of human aganets in the possibility of wildfire, it should be said that despite the transit of the road inside the national park and the traffic of passengers and passers-by in this section, the probability of wildfire is very high within 100 m of the road so that with the slightest spark, including cigarette butts thrown by drivers around the road, which provides access to the side roads inside the national park and nature walkers, tourists, and even hunters have access to a depth of about 14 km from the national park, the probability of wildfire despite this road is still high up to this distance.
The 500-meter area of villages has a high probability of wildfire, which can be attributed to the agricultural lands of these villages, which are bordered by national parks. Farmers burn agricultural waste to replant their land. In this case, the control of this wildfire is out of reach and penetrates the national park (Faramarzi et al., 2021). However, at longer distances can be attributed to the interaction of other variables and the livelihoods of these villagers are in turn tied to the forests and pastures, and in this regard, the managerial and legal problems for managing the national park. In fact, in these villages, there is a conflict between human activities and nature protection, such as turning national park areas into farmland, poachers, and grazing livestock (Ghoddousi et al., 2018). Therefore, in some cases, these people take wildfire in natural areas in retaliatory actions due to legal problems, which can continue up to 15 km in the villages.
The results of the studies of (Sevinc et al., 2020) and (Collins, Griffioen, Newell, & Mellor, 2018) also confirm the high impact of climatic and human agents on the occurrence of wildfire, which is the same as the results of the present study.
The two variables of elevation above sea and distance from the transit road are the most important variables of flood and landslide hazards according to the ME results. One of the reasons for the importance of these variables in flood hazards is that the main river of the national park, Maderso, where most of the runoff created in the national park collects and exits the park, is located at the lowest elevation of the area, which is also prone to flooding (Faramarzi et al., 2020). Also, the transit road is located exactly on the same river, so that due to the twisted state of the river, several bridges have been created on this river along this road.This situation has caused losses in the event of rising river water and flooding damage to the passengers who travel on this route. The graph of how these variables affect also shows a significant decrease in this hazard at a distance of more than 300 m from the transit road. The elevation graph also shows a greater hazard of flooding at an elevation of less than 500 m above sea level, which is actually produced at high elevation if runoff conditions are provided and finally reach each other at the lowest point of the basin, causing flooding areas.
Because of the mountainous nature of the national park, the possibility of flooding in the flat areas located on the slopes and upstream is very small, because in the first place, the volume of runoff in these areas is low and there is no possibility of flooding and the gravity of the earth; in the upstream areas, the runoff speed is high in these areas, which causes water to pass through the flat areas and reduce the risk of flooding (Khosravi, Pourghasemi, Chapi, & Bahri, 2016; Termeh, Kornejady, Pourghasemi, & Keesstra, 2018).
The graph of how elevation affects shows the occurrence probability of landslide hazard at low elevations, such that elevations 700-800 m above sea level are the most likely to occur. Meanwhile, the graph of the road distances shows that approximately 500 m is the most hazardous area in terms of landslides. The point that can be added to these two important variables is that this road has almost exceeded the elevation range affecting this hazard. In fact, high elevations are less prone to landslides because they are rocky and inaccessible, and low elevations are more likely to occur because of man-made structures and road traffic located in these places. In fact, excavations and embankments that have taken place around roads have also caused instability of the slopes and have increased this hazard.
Because of the high impact of rainfall on the occurrence of flood hazards, it can be said that the amount and manner of rainfall play an important role in the formation of these areas; therefore, heavy rains can cause runoff production in a short time and the occurrence of this risk. In fact, production runoffs do not have the opportunity to penetrate the soil and increase flooding hazards.
Faults play an effective role in increasing the instability of slopes; therefore, seismic events are more intense near faults, and these faults in crushing, surrounding rocks, and water penetration into the crushed masses (Abedini & Tulabi, 2018) Therefore, by moving away from them, the probability of slope instability decreases.
The history of landslide hazard studies also shows the importance of road distance, elevation, slope, and distance from faults in the occurrence of this hazard (Costanzo, Rotigliano, Irigaray, Jiménez-Perálvarez, & Chacón, 2012; Pourghasemi et al., 2019; Pradhan & Lee, 2010), which is consistent with the results of this study.
The largest area related to the overlap of hazards with 3% of the area is related to the two wildfire and landslide hazards, which the study of (Pourghasemi et al., 2019) confirmed the simultaneous occurrence of these two hazards.
The highest hazard area, accounting for 17.2% of the area, is related to landslide hazards. This hazard occurred because of its location on different faults and the mountainous nature of the national park and its high slope, as well as the existence of roads and waterways. Meanwhile, the fire hazard was more than 14%, and floods accounted for 7.1% of the national park area. In fact, the high slope and mountainous nature of the national park have caused production runoff to exit and reduce the flood risk area. Since more than 47% of the national park wildfire area is related to the impact of the transit road (Faramarzi, Hosseini, Pourghasemi, & Farnaghi, 2019), it can be said that approximately 14% about 6.6% of the national park wildfire area is the route of this road. In addition to the excavations and embankments that have taken place around it, have caused instability of the slopes, which has led to landslide hazards, and this has caused an increased overlap of the hazards of landslides and wildfires. With the loss of vegetation by wildfire, it can be considered an agent in creating a landslide hazard, as (Pourghasemi & Kerle, 2016) also pointed out the effect of forest fires in increasing the risk of landslides in mountainous areas.
ROC is a common method for evaluating the accuracy of models; therefore, in this method, the area under curve (AUC) analysis from the value 0.5 to 1 is used to evaluate the model. The value of this characteristic in the BRT model for the fire hazard map was 0.97, flood risk was 0.98, and landslide hazard was 0.93, indicating the excellent prediction of this model in identifying hazards. In fact, the high stability of this model against data turbulence has led to a reduction in its forecasting error, as many studies have shown that it performs well in data mining, classification, forecasting, and cluster analysis(Cutler et al., 2007). The results of evaluating the accuracy of this model using 30% of the data that were not included in the model and the use of value range analysis under the curve in the detection of relative performance indicate the excellent performance of this model in identifying potential hazards so that the results Studies by (Rahmati et al., 2019), (Thüring, Schoch, van Herwijnen, & Schweizer, 2015), (Pozdnoukhov, Purves, & Kanevski, 2008), (Pourghasemi & Kerle, 2016), (Rahmati et al., 2019), (Arabameri et al., 2019) and (Pourghasemi et al., 2019), (Shabani, Pourghasemi, & Blaschke, 2020). In a study conducted by (Bui, Tuan, Klempe, Pradhan, & Revhaug, 2016) on landslide hazard assessment using five models of machine learning techniques, the highest accuracy was obtained for the BRT model.
If the AUC value is between 0.9 to 1 excellent prediction, 0.8 to 0.9 very good, 0.7 to 0.8 good, 0.6 to 0.7 moderate and less than 0.6 (Pourghasemi & Kerle, 2016). Therefore, other machine learning techniques also exhibit excellent performance. The SVM model has the complex advantages of nonlinear and insensitive relationships to parasites; thus, the study records also show the high accuracy of this model in assessing landslide and flood risks (Pourghasemi, Jirandeh, Pradhan, Xu, & Gokceoglu, 2013; Tehrany et al., 2019). One of the most important features of the GAM is its similarity to machine learning methods and the possibility of easy interpretation of this algorithm(Goetz, Guthrie, & Brenning, 2011) and which has a great ability to analyze data and determine nonlinear relationships. Different variables (Hanspach, Kühn, Pompe, & Klotz, 2010). The results of the present study show the excellent accuracy of this model in risk assessment, which is the same as the study by (Abeare, 2009). In the RF model, data fragmentation with one variable is performed randomly, and this process continues with other predictor variables in a suitable set to grow classification trees (Collins et al., 2018). In fact, the high stability of this model against data turbulence has led to a reduction in its forecasting error, as many studies have shown that it performs well in data mining, classification, clusters, forecasting, and analysis (Cutler et al., 2007).
Finally, we compared the hazard map obtained in the present study and the distribution of important mammals in the national park, which indicates that the habitat of these mammals is endangered. Thus, the Mirzabaylo area, which is the habitat of deer, is at risk of flooding and the habitat of marals and leopards. There is a possibility that all three environmental hazards and the habitat of ewes are at risk of fire and bears are at risk of landslides, so the hazard maps from the present study can be used to manage and protect the habitat of these important species. In addition, because of the location of the bridges built on the Madersu River and Cheshmeh-e-Zav, it has increased the vulnerability of these areas, and the path of these rivers should be laminated in these sections as much as possible.
According to the results of this study, the implementation of the transit road transfer policy outside the national park due to its high impact on all hazards is felt more than before, and it also suggests controlling these hazards by creating hazard monitoring systems and preventing the construction of human structures in high-risk areas, especially the Zaw area, which is being developed for tourism programs.