Guria Rajkumar, Mishra Manoranjan, Mohanta Samiksha, Paul Suman
Department of Geography, Fakir Mohan University, Vyasa Vihar, Nuapadhi, Balasore, 756089, Odisha, India.
Environ Sci Pollut Res Int. 2025 Jan 30. doi: 10.1007/s11356-025-35976-6.
Forests play a vital role in environmental balance, supporting biodiversity and contributing to atmospheric purification. However, forest fires threaten this balance, making the identification of forest fire probability (FFP) areas crucial for effective mitigation. This study assesses forest fire trends and susceptibility in the Similipal Biosphere Reserve (SBR) from 2012 to 2023 using four machine learning models-extreme gradient boosting tree (XGBTree), AdaBag, random forest (RF), and gradient boosting machine (GBM). A forest fire inventory was created using the delta normalized burn ratio (dNBR) index, and 19 conditioning factors were incorporated after rigorous collinearity testing. FFP maps were generated and evaluated using ROC-AUC, MAE, MSE, and RMSE metrics. The frequency ratio (FR) model was also applied to assess the importance of variables. The results show that approximately 40.85% of the study area is high to very high susceptible to forest fires, with the RF model achieving the highest accuracy (AUC = 0.965). An average analysis across all models revealed that high susceptibility areas accounted for 23.08% of the study area, the largest among all classes. Moderate susceptibility zones covered 16.19%, while very high susceptibility areas comprised 18.23%. Interestingly, very low and low susceptibility zones together represented 42.50%, indicating a large portion of the area is at relatively low fire risk. Temporal analysis identified 2021 as the peak year for fire incidents, with 94.72% of the fires occurring during March and April. The buffer zone experienced the highest number of incidents, with a significant anthropogenic influence. Using the FR model, variable importance analysis showed that land use and land cover (LULC), NDVI, and NDMI were the most influential factors in fire susceptibility. This study contributes to forest fire management by integrating the dNBR index with machine learning models and FR analysis to generate precise FFP maps. These findings provide valuable insights for policymakers and conservationists, enabling targeted interventions in high-risk zones and enhancing fire management strategies to reduce the impact of forest fires.
森林在环境平衡中发挥着至关重要的作用，支持生物多样性并有助于大气净化。然而，森林火灾威胁着这种平衡，因此确定森林火灾概率（FFP）区域对于有效缓解至关重要。本研究使用四种机器学习模型——极端梯度提升树（XGBTree）、AdaBag、随机森林（RF）和梯度提升机（GBM），评估了2012年至2023年西姆利帕尔生物圈保护区（SBR）的森林火灾趋势和易感性。使用增量归一化燃烧比（dNBR）指数创建了森林火灾清单，并在经过严格的共线性测试后纳入了19个调节因子。使用ROC-AUC、MAE、MSE和RMSE指标生成并评估了FFP地图。还应用频率比（FR）模型来评估变量的重要性。结果表明，研究区域约40.85%对森林火灾的易感性为高到非常高，其中RF模型的准确率最高（AUC = 0.965）。对所有模型的平均分析表明，高易感性区域占研究区域的23.08%，在所有类别中占比最大。中度易感性区域占16.19%，而非常高易感性区域占18.23%。有趣的是，极低和低易感性区域合计占42.50%，表明该区域大部分地区的火灾风险相对较低。时间分析确定2021年为火灾事件的高峰期，94.72%的火灾发生在3月和4月。缓冲区的事件数量最多，受到显著的人为影响。使用FR模型进行的变量重要性分析表明，土地利用和土地覆盖（LULC）、归一化植被指数（NDVI）和归一化水分指数（NDMI）是火灾易感性的最有影响因素。本研究通过将dNBR指数与机器学习模型和FR分析相结合，生成精确的FFP地图，为森林火灾管理做出了贡献。这些发现为政策制定者和保护主义者提供了有价值的见解，能够在高风险区域进行有针对性的干预，并加强火灾管理策略，以减少森林火灾的影响。
