Fire Susceptibility Mapping in the Northeast Forests and Rangelands of Iran using New and Ensemble Data Mining Models

Fires have increased in the northeastern Iran as its semiarid climate landscape is being desiccated by human activities. To combat re outbreaks in any region, one must map re susceptibility with accurate and ecient models. This research mapped re susceptibility in the forests and rangelands of northeastern Iran’s Golestan Province using new data mining models. Fire effective factors data describing elevation, slope angle, annual mean rainfall, annual mean temperature, wind effect, topographic wetness index (TWI), plan curvature, distance from river, distance to road, and distance to village were obtained from several sources. The relative importance of each variable was determined with a random forest algorithm. Fire susceptibility maps were produced in R 3.3.3 software using GAM, MARS, SVM algorithms and a new ens emble of the three models: GAM-MARS-SVM. Validation of the four re susceptibility maps was performed with the area under the curve. Results show that distance from village, annual mean rainfall and elevation were of greatest importance in predicting re susceptibility. The new GAM-MARS-SVM ensemble model achieved the highest re susceptibility mapping precision. The re susceptibility map produced using the GAM-MARS-SVM ensemble model best detected the high re risk areas in Golestan Province.


Introduction
Forest res, whether natural or human-induced, have many negative impacts on environmental, social, and economic conditions in most environments (Dimitrakopoulos and Mitsopoulos 2006). One of the solutions to prevent res is protective management in critical re-prone and human-occupied areas (Eskandari et al. 2015a). Therefore, to map re susceptibility and to identify high-risk areas accurately is of utmost importance.
Iran's climates range from arid to semi-arid. Global warming, regional climate change, and human activities in Iran's natural ecosystems have caused wild res throughout a large portion of the forests and rangelands in the northern and northeastern parts of country in the recent years (Mazandaran Natural Resources Administration 2017; Golestan Natural Resources Administration 2018). Golestan Province is one of the regions most affected by wild res in the recent years (Golestan Natural Resources Administration 2018). This region is already prone to re due its normal, historical climate. Based on some reports and studies, res in Golestan Province usually occur in areas experiencing decreasing rainfall, dehydration, and leaf litter accumulation ( In Iran speci cally, re danger mapping has been conducted at a national scale (Eskandari and Chuvieco 2015) and at regional scales (Eskandari et  Province in recent years, no comprehensive study for re susceptibility assessment has been performed at the Province scale. Highly accurate re susceptibility maps can be very useful for guiding use and management in the zones of highest risk in the Province. The aims of this study are: (1) to map past res in the Golestan Province; (2) to evaluate the importance of the effective factors in the prediction of re susceptibility; (3) to map re susceptibility using new and ensemble data mining models; and (4) to validate the re susceptibility maps produced using AUC to identify the best modeling method for the study area.

Data
An accurate re susceptibility map is vital for re prevention, mitigation, and response in re-prone areas (Tehrany et al. 2018; Eskandari et al. 2020). Selecting the factors that are most important predictors of re susceptibility is crucial to the modeling of an accurate and reliable re susceptibility map. In this study, a DEM was used to determine elevations, slope angles, topographic wetness indices (TWI), plan curvatures, distances to roads, distances to villages, and distances to rivers of each of the locations of previous res. These effective topographic and anthropogenic factors for re susceptibility were The DEM of Golestan Province was generated from an ASTER-GDEM (30m-resolution) available from the USGS (https://earthexplorer.usgs.gov) (Fig. 2). Slope angle was calculated from the DEM. TWI is a secondary DEM feature obtained from the 30m-resolution DEM (Beven and Kirkby 1979): where, α is the cumulative upslope area of drainage through a point, and tan β is the slope angle at that point. TWI was expected to be an important re-promotion factor. The wind-effect map was constructed from three variables: DEM, wind direction (degree), and wind speed (m/s) in SAGA GIS The locations of roads, rivers, and villages in Golestan Province were extracted from 1:25,000-scale topographical maps. Distances to roads, distances to villages, and distances to rivers were then determined in ArcGIS 10.6.1. Annual mean rainfall and annual mean temperature maps were acquired from the Golestan Meteorological Administration. Maps of each of these re susceptibility effective factors are shown in Fig. 3.

Fire Occurrence Detection
For re susceptibility modeling, actual re data in the study area is required. All of the res that occurred Eskandari et al. 2020). In this study, all MODIS re products for Golestan were obtained from NASA (https://modis.gsfc.nasa.gov/data/). HDFView software (http://hdfeos.org/software/heg.php) was used to detect the re pixels (HDF-EOS to GeoTIFF Conversion Tool (HEG) 2017). The re products were imported to HDFView and the position of re pixels were detected. A map of the pixels that represented past res was constructed in GIS. The re pixels were divided randomly into two groups: 70% for training and 30% for validation of the re susceptibility modeling results (Fig. 1b).

Importance of the Effective Factors for Fire Susceptibility Mapping
Selection of the variables that serve as the most important re location predictors is important for the creation of reliable maps generated by proper models. In this research, the importance of effective variables on re susceptibility mapping was determined with the random forest (RF) algorithm

Importance of Effective Factors in Fire Susceptibility Mapping
Importance of the effective factors in re susceptibility mapping using the RF algorithm is shown in Fig.  4.

Multi-collinearity Analysis
The results of the multi-collinearity test among effective factors are shown in Table 1. As a result of this test, the aspect factor was deleted from analysis because of collinearity with other factors. There is no multi-collinearity among the other effective factors as VIF is <5 and tolerance is >0.1 (O'Brien 2007)  (Fig. 5). The respective areas of the four re susceptibility classes generated by the data mining models are shown in Table 2. Validation results of the re susceptibility maps from the data-mining models are shown in Fig. 6 and Table 3.

Discussion
Considering the problem of increasing res in Golestan Province, this research mapped re susceptibility using GAM, MARS, SVM and GAM-MARS-SVM data-mining models. Results of the assessment of the importance of effective factors on re susceptibility using the RF algorithm indicate that distance from nearest village, annual mean rainfall, and elevation were the most predictive of re susceptibility. Annual mean temperature, wind effects, distances to roads, and slope angle were of medium importance to re susceptibility. Previous studies have shown that annual mean temperature has a strong relationship with the number of res in Golestan Province (Eskandari 2015). Furthermore, the proximity to roads has also been identi ed as important factor in re susceptibility potential (Martinez et al. 2009 . TWI, distances to rivers, and plan curvature were the factors of least importance to re susceptibility. Therefore, using them for modeling is not recommended in future studies. Aspect was collinear with the other factors and was removed from further analysis. Based on these results, human factors (distance to villages and distance to roads), climatic factors (annual rainfall mean, annual mean temperature, and wind effect), and topographic factors (DEM and slope angle), together dictate re susceptibility in the forests and rangelands of Golestan Province. It has been con rmed based on other studies performed about re danger mapping in natural areas of Iran The ensemble GAM-MARS-SVM algorithm used for the rst time in this study generated very good results for re susceptibility mapping in the study area. At present, application of ensemble data mining algorithms for re susceptibility mapping is new and limited. The GAM-MARS-SVM algorithm is highly recommended for future efforts to map re susceptibility in other re-prone areas of Iran, such as in the Hyrcanian forests of Mazandaran and Giulan Provinces, as well as in other arid and semi-arid regions of the world. The resulting re susceptibility maps may provide helpful information to enable better prediction of future res in the resource-rich forests and rangelands of these Provinces.
The spatial analysis of re susceptibility obtained from ensemble data mining algorithm (GAM-MARS-SVM model) showed that the high and very high re-danger classes have been located in the center of Province. The individual data mining algorithms (MARS, SVM and GAM models) created different spatial pattern of re susceptibility. The important point is that high and very high re susceptibility locations in the results of all of the models are located in the center of Province where the village and road densities are highest. This seems to re ect the important role that humans play in causing wild res in the Province and this has been con rmed by others ( Analysis of re susceptibility based on most accurate algorithm (GAM-MARS-SVM model) demonstrated that 42.69% of Golestan Province has low re danger, 17.55% moderate danger, 17.70% high danger, and 22.06% by very high danger. Combined, 39.76% of the study area is high or very high re potential. These parts should, therefore, be studied and managed to mitigate and prevent ignition and augmentation of re potential. Proactive and protective management (especially against human activities) can reduce res in the Province.

Conclusions
Results of this study show that distance from village is the most important factor in re susceptibility in Golestan Province. Therefore, management to prevent res in the Province should be focused in rural areas. The high precision of the GAM-MARS-SVM ensemble model suggests that re susceptibility mapping will be enhanced by using it to forecast areas of high re danger. Using this ensemble model, we nd that 40% of the Province has high or very high re risk. Therefore, the Province will very likely experience many res in the future. The re susceptibility map that the ensemble model produced can be very useful for creating and enhancing management strategies for preventing res, particularly in the higher risk portions of Golestan Province.