آشکارسازی تغییرپذیری مکانی اثر عوامل مؤثر بر آتش‌سوزی جنگل‌ها و مراتع با مدل رگرسیون وزن‌دار جغرافیایی (بررسی موردی: مناطق شمالی استان ایلام)

Background and Objective: Forest and rangeland fires are among the most significant ecological hazards in the semi-arid ecosystems of the Zagros region. Their occurrence is driven by the complex interaction of climatic, physiographic, anthropogenic, and vegetation-related factors, resulting in substantial impacts on ecosystem stability, ecosystem services, and the livelihood security of local communities. In recent decades, intensified climate variability, increasing temperatures, recurrent droughts, and the expansion of human activities have heightened the susceptibility of these ecosystems to wildfire occurrence. Accurate understanding of wildfire spatial patterns requires modeling approaches capable of accounting for spatial heterogeneity in the relationships among explanatory variables. Therefore, this study aimed to investigate the spatial variability of factors influencing wildfire occurrence in forest and rangeland ecosystems and to examine the spatial differences in these relationships across the study area. Furthermore, the performance of the Geographically Weighted Regression (GWR) model was evaluated and compared with the Ordinary Least Squares (OLS) regression model to identify critical hotspots and improve management decision-making.
Methodology: A total of 441 wildfire polygons occurring during the warm seasons between 2011 and 2020 were delineated using Landsat 8 satellite imagery and MODIS hotspot data for wildfire detection and verification. Following validation against official records provided by the Ilam Province Department of Natural Resources and Watershed Management, the dataset was refined and subjected to quality-control procedures. Each wildfire polygon was considered an individual spatial analysis unit, and polygon area was used as the dependent variable representing wildfire magnitude and spatial distribution in the modeling process. A set of 21 independent variables, including climatic, topographic, proximity-related (distance to roads, villages, cities, and water resources), land-use, vegetation-cover, and socio-economic indicators, was selected. Base maps were prepared in a GIS environment, and the values of all variables were extracted for each wildfire polygon. Data preprocessing included outlier detection, scale standardization, Box–Cox transformation, and normality testing. Multicollinearity was assessed using the Variance Inflation Factor (VIF). To mitigate multicollinearity, reduce dimensionality, and eliminate internal correlations among variables, extracted components were used in subsequent modeling. Components explaining more than 80% of the total variance were retained. The primary analysis was conducted using the GWR model, which allows local coefficient estimation and the assessment of spatial variability in relationships. Moran’s I statistic was applied to model residuals to evaluate spatial autocorrelation and model validity. For comparison purposes, an OLS regression model was also implemented, and the goodness-of-fit statistics of both models were analyzed.
Results: The results demonstrated that the GWR model provided greater explanatory power than the OLS model. The local model achieved a coefficient of determination (R²) of 0.32, while a substantial reduction in the corrected Akaike Information Criterion (AICc = 1753) and the standard error of estimation indicated superior model performance. Analysis of local coefficients revealed that wind speed was among the most influential climatic factors affecting the spatial variation of wildfire occurrence. From a physiographic perspective, slope aspect was identified as a significant determinant of wildfire occurrence across many parts of the study area; however, both the magnitude and direction of its effects varied spatially. Regarding land use and vegetation cover, rainfed agricultural lands, dense forests, and moderately covered rangelands exhibited the strongest associations with wildfire occurrence, highlighting the considerable influence of human activities on fire regimes. The GWR results further indicated that the importance and effects of wildfire-driving factors were not spatially constant across the region; rather, the relationships among variables varied considerably from one location to another. This spatial non-stationarity enables the identification of critical wildfire-prone areas and supports the development of location-specific management strategies.
Conclusion: The application of Geographically Weighted Regression (GWR) provides an effective framework for investigating the spatial variability of relationships between wildfire occurrence and environmental, physiographic, anthropogenic, and socio-economic factors. Compared with the Ordinary Least Squares (OLS) model, GWR demonstrated a greater capacity for explaining the spatial heterogeneity of these relationships. The findings revealed that both the magnitude and direction of the effects of wildfire-related factors vary substantially across the study area rather than remaining spatially uniform. These results emphasize the necessity of adopting region-specific and locally adapted approaches to forest and rangeland fire management. Furthermore, the local coefficient maps and spatial patterns derived from the GWR model can serve as a scientific basis for identifying vulnerable areas, prioritizing management interventions, designing preventive strategies, and improving wildfire monitoring and management at the regional scale.