The forest fire hazard mapping using the accurate models in the fire-prone areas has particular importance to predict the future fire occurrence and allocate the resources for preventing the fire ignition. This research aimed to compare the accuracy of some individual models including boosted regression tree (BRT), classification and regression trees (CART), functional discriminant analysis (FDA), generalized linear model (GLM), mixture discriminant analysis (MDA), random forest (RF) and two new hybrid models including FDA-GLM-MDA and RF-CART-BRT for predicting the fire hazard in a fire-prone area in the northeast Iran, Golestan Province. For this purpose, a comprehensive dataset from ten effective parameters including digital elevation model (DEM), slope angle (SA), plan curvature (PC), topographic wetness index (TWI), annual rainfall mean (ARM), annual temperature mean (ATM), wind effect (WE), distance to urban areas (DTU), distance to streams (DTS) and distance to roads (DTR) was created in GIS. Furthermore, 3705 historical fire locations in the Golestan Province from 2002 to 2017 were obtained from MODIS fire product dataset. Then, the variable importance was assessed using the XGBoost machine learning (ML) technique. Finally, the individual and hybrid models were evaluated using the ROC-AUC method. The results showed that the DTU was the most important factor in modeling and mapping the fire hazard in the Golestan Province. Also, the results demonstrated that the individual random forest (RF) (AUC = 0.855) and hybrid RF-CART-BRT algorithms (AUC = 0.854) were the most accurate predictive models for mapping the fire hazard in the Golestan Province, respectively. Considering the high significance of DTU in fire occurrence in this study, the area of fire hazard classes in fourteen different counties of the Golestan Province was calculated using the most accurate model (RF model). The final results indicated that Minudasht County had the most area (35.55%) of fire hazard in the very high fire hazard class. The results of this study are very useful for local forest managers to control the future fires using the best model in the natural areas of the Golestan Province, especially in Minudasht County. The protective management of the natural areas of the Golestan Province would be performed based on the fire hazard maps produced by RF and RF-CART-BRT algorithms. We recommend applying these models for fire danger mapping in fire-prone areas around the world which have semi-arid conditions. More comparative assessment of individual and ensemble models for fire danger mapping in semi-arid areas around the world could provide a baseline for monitoring fire danger in similar conditions.

在易火地区使用准确的模型绘制森林火灾危险图对于预测未来的火灾发生和分配资源以防止火灾着火特别重要。这项研究旨在比较某些单个模型的准确性，包括增强回归树（BRT），分类和回归树（CART），功能判别分析（FDA），广义线性模型（GLM），混合判别分析（MDA），随机森林（RF）和两种新的混合模型，包括FDA-GLM-MDA和RF-CART-BRT，用于预测伊朗东北部Golestan省易火地区的火灾隐患。为此，我们提供了十个有效参数的综合数据集，包括数字高程模型（DEM），坡度角（SA），平面曲率（PC），地形湿度指数（TWI），年降雨量平均值（ARM），在GIS中创建了年平均温度（ATM），风影响（WE），到市区的距离（DTU），到溪流的距离（DTS）和到道路的距离（DTR）。此外，从MODIS消防产品数据集中获得了2002年至2017年期间Golestan省的3705个历史火灾地点。然后，使用XGBoost机器学习（ML）技术评估了变量的重要性。最后，使用ROC-AUC方法评估单个模型和混合模型。结果表明，DTU是对Golestan省火灾隐患进行建模和绘图的最重要因素。此外，结果表明，个体随机森林（RF）（AUC = 0.855）和混合RF-CART-BRT算法（AUC = 0.854）分别是用于绘制Golestan省火灾隐患的最准确的预测模型。考虑到本研究中DTU在火灾发生中的重要意义，使用最准确的模型（RF模型）计算了Golestan省14个不同县的火灾危险类别区域。最终结果表明，在极高的火灾隐患等级中，Minudasht县的火灾隐患最多（35.55％）。这项研究的结果对于当地森林管理者使用最佳模型在Golestan省自然地区（尤其是Minudasht县）控制未来大火非常有用。将基于RF和RF-CART-BRT算法生成的火灾隐患图，对Golestan省自然区域进行保护管理。我们建议将这些模型用于在世界范围内有半干旱条件的易火地区绘制火灾危险图。