Analyzing Fire Susceptibility and its Driving Factors Using Maximum Entropy Model (Case Study: Forest and Rangeland of East Azerbaijan)

Fire plays a major role in shaping the structure of many terrestrial ecosystems and influences their functionality. With East Azerbaijan Province experiencing one of the most severe droughts over the past few years, the average size and extent of burned area had increased intensively in this region, which is significantly posing threats to the wildland, people, property, and destroying environmental infrastructure. In order to accomplish an in time effective and preventive management and also prior to fire occurrences in forest and rangeland, it was essential to identifying and classifying the susceptible areas and factors influencing on fire occurrences. Generally, the lack of knowledge about regional conditions, which has major influence on the expansion of fire occurrence and causing serious damage to environmental area, infrastructure and property, animal life and plants (Jaiswal et al, 2001 1). Therefore, generating of fire susceptibility maps and analyzing its driving factors may have a great impact on reducing and preventing the deforestation and desertification in environmental area (dong et al, 2005: 169), also it will support and assist the activities of managers and planners in forest management for planning and implantation of an early warning system. In this manner, understanding the spatial distribution of fire ignitions and learning about the key drivers behind this distribution is mostly based on the analysis of historical fire locations. Detailed information about the changes of spatial distribution of fire dangers will provides valuable guidance in the direction of forecasting the trends of environmental dangers for the government and protection agencies. Moreover, it would be very useful as a complementary approaches to fire or wildfire management. Various approaches have been applied to modeling wildfire susceptibility, including different techniques, some authors concluded that machine learning algorithms, like maximum entropy (MaxEnt), performed with high accuracy for modeling wildfire susceptibility. In this regard by using an approach based on a machine learning algorithms called the maximum entropy (MaxEnt) with correct and well-organized data in the context of geographic information system (GIS), it is possible to achieve better and more understanding of the changes in the spatial distribution in fire dangers and affecting factors. In the current study, by using of fire event data from a 9-year period, wide variety of environmental variables and MaxEnt approach, an attempt was made for investigation in fire danger and its influencing factors in forest and rangeland of East Azerbaijan province over a 9-year period. The present study area is East Azerbaijan province in Northwestern Iran, which lie between latitudes 36° to 39° N and longitudes 45° to 48° E and extends over 47,000 km2, with a wide variety of vegetation types, topographical and climatic conditions. The East Azerbaijan province environment is highly diverse in terms of climate, geology and topographic characteristics which lead to different levels of susceptibility to fire. The climate in the study area based on the De Martonne aridity index is semi-arid, average annual precipitation and temperature are 315.2 mm and 10.2 °C, respectively, and elevation ranges from 160 to 4811 m above sea level.
Three phase comprising in the methodology of this research; the first phase applies the several methods such as image processing techniques, GIS analysis, and spatial interpolation methods for producing the dependent and independent datasets. The second phase, data of fire occurrences between 2006 and 2015 associated with a range of environmental and anthropogenic independent variables in three period (2006-2008, 2009-2011 & 2012-2015), jackknife test was used to find out the main driving factors in the spatio-temporal distribution of fire and wildfire. The third and last phase includes analysis of changes in fire danger and its driving factors. Data was collected from various sources including satellite imageries, and metrological data. Fire occurrences and locations were derived from MODIS (Moderate-Resolution Imaging Spectero Radiometer) fire product, historical record of environmental protection agency, and field checking. The Collection 5, Level 3, 8-day MODIS Terra and Aqua active fire product (MOD14A1 & MYD14A1) was used during this study. In the data preparation phase, ASTER Global-DEM with 28.5 meter spatial resolution was used to generate maps of elevation, slope and aspect by using ArcGIS software. Climate maps (max and mean temperature, and precipitation) were created through the interpolation of data gathered by several meteorological stations in the region. A detailed land use/land cover map derived from LANDSAT 8 satellite images, using multi-layer perceptron (MLP) neural network method as a wildly used image classification approach. Then, with kernel density and Euclidean distance functions, anthropogenic variables were created. Eventually, map of Normalized Difference Vegetation Index (NDVI) was derived from MODIS satellite imagery. Maximum Entropy (MaxEnt) is a presence-only machine learning algorithm that iteratively contrasts environmental predictor values at occurrence locations with those of a large background sample taken across the study area. The principle of MaxEnt is to estimate the probability distribution of maximum entropy, which is below a collection of constraints (environmental and anthropogenic conditions), the most spread-out or closest to uniform. It is a sophisticated approach to modeling the probability distribution from the n-dimensional environmental space using occurrence locations data and iteratively evaluates the contrasts between the values of those occurrences and those of a background consisting of the mean occurrence over the entire study area, as sampled from a large number of points. By applying this algorithm, the most uniform distribution will be recognized and selected from several possible distributions, moreover it can specify a per-pixel susceptibility to wildfire occurrence which might be used as an essential tool for environmental hazard management in forest and rangeland. It should be noted that the MaxEnt software was used to analysis the fire susceptibility. The main outputs of this model are; prediction of fire danger in study area, Receiver Operating Characteristic curve as an indicator for accuracy assessment and jackknife test to assess the significance or relatively importance of each independent variable in model result in each period. The process of implementation models for three 3-year period have been done by using the fire occurrences data as the dependent variable and 12 anthropogenic and environmental factor as the independent variables. Assessment of the changes at different areas in terms of fire susceptibility represented upward trend in fire danger, increasing the areas in high-risk zones and extensive changes in the factors influencing the fire occurrence. According to the results most extent of changes about decreasing the areas of zones with very low risk and increasing the area of zones with very high risk and medium risk. Based on the result of accuracy assessment by using Area Under Curve (AUC), which was calculated from Receiver Operating Characteristic (ROC), our analysis indicated the maximum entropy as machine learning algorithm were efficient predictor (AUC =0.91, 0.828, 0.839) for wildfire susceptibility assessment in East Azerbaijan province in three different time period. The machine learning algorithm MaxEnt was successfully applied to analyzing forest fire susceptibility assessment in East Azerbaijan province to demonstration areas with higher fire susceptibility and analyzing its driving factors over the times. Due to the high accuracy of the suggested model (MaxEnt) which has been utilized in the study for 3- period (2006-2008, 2009-2011 and 2012-2014) and also the assessment of the previous studies, this model could considered as a proper approach to evaluating and modeling fire danger and its affecting factors. The results of this study will help to determinate and evaluate fire management objectives. Considering limited financial and human resources to suppress fires in the East Azerbaijan province the concentration on very susceptibility areas could be a way to encounter these challenges. The results can be used for preventive actions and additionally for making more scientific and appropriate decisions. Investigate the changes of fire dangers in forest and rangeland of East Azerbaijan and its upward trend, reveals the requirement for more attention to this issue and also for planning the proper strategies and activities. On the other side increasing in the impact of human factors on fire occurrences in recent years, demand the necessity for cultural proceeding, raise of public awareness as well as taking advantage of the general public potential in the field of preventive activities.