Abstract In recent years, land uses have been changing and aridity has been increasing in the forests and rangelands of central Koohdasht which is a region in the forests of the Zagros Mountains in western Iran. Consequently, the number of fires has also increased. This study employs data-mining techniques to model fire danger using information regarding land cover, climate, topography, and other fire-danger influencing factors. A land cover map was prepared using Sentinel-2A satellite images and a maximum likelihood (ML) algorithm. Digital data describing other factors that influence fire danger (slope angle, aspect, elevation, climate, topographic wetness index, and distances from rivers and roads) were compiled from several sources and imported into a GIS. The locations of past fires in the study area were also determined from MODIS satellite images and data acquired from the region’s fire service. The quantitative and qualitative spatial relationships between effective factors and patterns of fires were investigated to model fire danger. A new machine-learning algorithm (the Boruta algorithm) was used to assess the relative importance of the fire-danger factors. Fire danger maps were created using several new data-mining algorithms including support vector machine (SVM), generalized linear model (GLM), functional data analysis (FDA), and random forest (RF). All were run in R 3.3.3 software. Finally, the fire danger maps were validated with several indices to determine the model that best predicts the fire danger in Koohdasht County. The results reveal that fire locations were determined mostly by elevation (low), aspect (south and southwest facing slopes), and aridity (semi-arid regions). Most fires occurred in non-natural landscapes: residential areas (46.74% of fires), agricultural lands (25.77%), and gardens (5.42%). In total, 77.93% of fires occurred in non-natural landscapes and within 500 m of roads. Only 22.07% of fires occurred on rangelands and forests. Three factors (distance from roads, climate, and aspect) were the strongest predictors of fire locations in the study area. Furthermore, area-under-the-curve (AUC) values indicate that the FDA (0.777) and GLM (0.772) algorithms generated the most accurate fire danger maps. These results have practical implications for fire danger management in the Zagros forests and provide baseline information for forest managers about the most important factors affecting fire danger in the similar regions. This methodology can be used by forest managers to predict the areas with greatest fire danger to prevent future fires through land use management, planning, and strategic decision-making. The results enable forest managers to find the best methods to monitor, manage, and control fire occurrence based on fire danger maps in the forests of western Iran, or in forests of other regions with similar conditions.

中文翻译：

---

伊朗自然区域土地覆盖、地形和气候与火灾发生的关系：应用新的数据挖掘技术对火灾危险进行建模和绘制

摘要 近年来，伊朗西部扎格罗斯山脉森林地区库赫达什特中部的森林和牧场的土地利用发生了变化，干旱程度不断增加。因此，火灾的数量也有所增加。本研究采用数据挖掘技术，利用有关土地覆盖、气候、地形和其他火灾危险影响因素的信息对火灾危险进行建模。土地覆盖图是使用 Sentinel-2A 卫星图像和最大似然 (ML) 算法制作的。描述影响火灾危险的其他因素（坡度、坡向、海拔、气候、地形湿度指数以及与河流和道路的距离）的数字数据从多个来源汇编并导入 GIS。研究区域过去火灾的位置也是根据 MODIS 卫星图像和从该地区消防部门获得的数据确定的。研究了有效因素和火灾模式之间的定量和定性空间关系，以模拟火灾危险。一种新的机器学习算法（Boruta 算法）用于评估火灾危险因素的相对重要性。火灾危险地图是使用几种新的数据挖掘算法创建的，包括支持向量机 (SVM)、广义线性模型 (GLM)、功能数据分析 (FDA) 和随机森林 (RF)。所有都在 R 3.3.3 软件中运行。最后，使用多个指标验证火灾危险地图，以确定最能预测 Koohdasht 县火灾危险的模型。结果表明，火灾地点主要由海拔（低）、坡向（朝南和西南的斜坡）和干旱（半干旱地区）决定。大多数火灾发生在非自然景观中：住宅区（火灾的 46.74%）、农田（25.77%）和花园（5.42%）。总共有 77.93% 的火灾发生在非自然景观和道路 500 m 以内。只有 22.07% 的火灾发生在牧场和森林。三个因素（与道路的距离、气候和坡向）是研究区域火灾位置的最强预测因子。此外，曲线下面积 (AUC) 值表明 FDA (0.777) 和 GLM (0.772) 算法生成了最准确的火灾危险地图。这些结果对扎格罗斯森林的火灾危险管理具有实际意义，并为森林管理者提供有关影响类似地区火灾危险的最重要因素的基线信息。森林管理者可以使用这种方法来预测火灾危险最大的区域，以通过土地使用管理、规划和战略决策来预防未来的火灾。结果使森林管理者能够根据伊朗西部森林或其他类似条件的森林中的火灾危险地图，找到监测、管理和控制火灾发生的最佳方法。规划和战略决策。结果使森林管理者能够根据伊朗西部森林或其他类似条件的森林中的火灾危险地图，找到监测、管理和控制火灾发生的最佳方法。规划和战略决策。结果使森林管理者能够根据伊朗西部森林或其他类似条件的森林中的火灾危险地图，找到监测、管理和控制火灾发生的最佳方法。