Wildfire is a major driver of terrestrial ecosystem degradation, and its impacts have intensified in the forests and rangelands of Mazandaran Province due to climate change and human activities. This study assessed wildfire susceptibility and evaluated predictive model performance using four data-mining methods: Artificial Neural Network (ANN), Generalized Linear Model (GLM), Generalized Additive Model (GAM), and Maximum Entropy (MaxEnt), applied individually and in an ensemble framework. Climatic, topographic, vegetation, and human-related variables were used as model inputs, and performance was evaluated using the Area Under the Curve (AUC), True Skill Statistic (TSS), sensitivity, and specificity. Results showed that the ensemble model achieved the highest accuracy and stability, while MaxEnt performed best among single models. Variable-importance analysis indicated that Normalized Difference Vegetation Index (NDVI) and precipitation were the dominant factors controlling the spatial pattern of wildfire susceptibility, followed by elevation and slope. The highest susceptibility occurred in areas with moderate elevation, slopes greater than 35%, moderate to dense vegetation cover, and annual precipitation below 500 mm. The ensemble map showed a widespread and relatively uniform distribution of high-risk zones, confirming the multifactorial and complex nature of wildfire occurrence and highlighting the value of multi-model approaches for preventive and sustainable natural-resource management.
Introduction
Wildfires represent one of the most significant ecological hazards affecting Iran’s natural ecosystems, resulting from the combined influences of climatic conditions, vegetation characteristics, topographic complexity, and human activities. Global evidence indicates that climate change through reduced precipitation, rising temperatures, and increased soil and vegetation dryness has intensified both the frequency and severity of wildfire events. Over the past two centuries, forests worldwide have undergone extensive degradation driven by fire, and numerous studies have highlighted the pivotal roles of natural and anthropogenic factors in shaping landscape-level fire susceptibility. In Iran, wildfires have emerged a critical environmental concern, destroying thousands of hectares of forest annually and placing the country among the most affected in the region. The ecological consequences of wildfire include biodiversity loss, alterations in soil physical and chemical properties, accelerated erosion, and increased carbon emissions. Given these impacts, the application of remote sensing, geographic information system (GIS), and data-driven modelling approaches has become essential for predicting fire occurrence and generating reliable susceptibility maps. Previous studies have demonstrated the effectiveness of neural networks, statistical methods, and machine-learning algorithms in wildfire forecasting. Building upon this body of research, the present study aims to map wildfire susceptibility in Mazandaran Province, identify the most influential environmental drivers, evaluate the performance of individual and ensemble models, and determine the most accurate predictive framework. The study is based on the hypothesis that climatic factors exert a dominant influence on wildfire occurrence and that ensemble modelling provides superior predictive performance compared to single-model approaches.
Material and Methods
To develop the wildfire susceptibility map, spatially explicit records of historical fire occurrences were collected through field surveys. After removing spatial autocorrelation, the fire locations were randomly divided into training and testing subsets. Absence data were generated using a pseudo-absence approach guided by preliminary MaxEnt outputs to ensure representative background conditions. A comprehensive set of environmental predictors was compiled, including topographic, meteorological, anthropogenic, and vegetation-related variables. Elevation, slope, aspect, topographic wetness index, and distance to rivers were derived from digital elevation models and remote-sensing data. Climatic variables—temperature, precipitation, relative humidity, and wind speed were interpolated across the study area using geostatistical methods based on observations from 32 meteorological stations. Anthropogenic factors included distance to roads and settlements, while vegetation conditions and moisture stress were characterized using the normalized difference vegetation index (NDVI) and Temperature Vegetation Dryness Index (TVDI). Wildfire susceptibility across forest and rangeland ecosystems of Mazandaran Province was modeled using four data-driven approaches: Artificial Neural Networks (ANN), Generalized Linear Models (GLM), Generalized Additive Models (GAM), and Maximum Entropy (MaxEnt). Model performance was evaluated using k-fold cross-validation, the area under the receiver operating characteristic curve (AUC), sensitivity, specificity, and the true skill statistic (TSS). To reduce uncertainty and improve predictive robustness, an ensemble model was developed using AUC-weighted averaging of the individual models, and wildfire susceptibility was classified into five categories. Finally, the relative contribution of each predictor was quantified using game-theoretic Shapley values to identify the key environmental drivers of wildfire occurrence.
Results and Discussion
The results indicate that the models exhibited varying levels of performance in predicting wildfire susceptibility. Both the MaxEnt and ensemble models achieved excellent accuracy, with AUC values of 0.92 and 0.94, respectively. Among the evaluation metrics, the ensemble model produced the highest TSS value (0.74), reflecting a strong ability to distinguish between fire presence and absence. In terms of sensitivity, the GAM and ensemble models performed best, whereas the GLM and ensemble models exhibited the highest specificity. Variable-importance analysis revealed that vegetation, climatic, and topographic factors exerted the greatest influence on wildfire susceptibility. NDVI showed high importance in the GAM and MaxEnt models, while precipitation was identified as the dominant variable by the ANN and ensemble models. Elevation and other topographic attributes also played significant roles in fire occurrence and spread, whereas anthropogenic variables had comparatively weaker effects. Spatial patterns of wildfire risk varied among models but generally converged in identifying high-risk zones. The ensemble susceptibility map showed a widespread distribution of wildfire risk across the region, highlighting the multifactorial nature of fire occurrence. Response curves indicated that increasing NDVI is generally associated with higher wildfire probability, although this relationship weakens at very high NDVI values due to reduced fine fuel availability. Fire susceptibility was highest at moderate elevations and increased with slope, particularly on slopes exceeding 10%. An inverse relationship was observed between precipitation and wildfire susceptibility, with drier areas exhibiting the highest risk.
Conclusion
The findings demonstrate that ensemble modeling provides a robust and reliable framework for wildfire susceptibility mapping, outperforming individual algorithms in terms of predictive accuracy and spatial consistency. Wildfire risk in Mazandaran Province is controlled by a complex interaction among vegetation characteristics, climatic conditions, and topographic features, underscoring the need for integrated management strategies. The susceptibility maps produced in this study offer a valuable scientific basis for land-use planning, proactive fire management, and prioritization of high-risk areas, particularly given that preventive measures are generally more cost-effective and sustainable than post-fire suppression and restoration. Overall, this study emphasizes the importance of multi-model and systematic approaches for mitigating wildfire risk in ecologically sensitive landscapes, and the ensemble-based maps developed here can serve as effective decision-support tools for regions with similar environmental conditions.