基于可解释机器学习模型的南宁市野火灾害易发性研究

野火易发性评价对野火灾害的前期预防以及灾害管理决策的制定至关重要。目前野火易发性的研究主要集中于提高模型的预测精度，而往往忽略对模型的内部决策机制进行解释分析。为此，构建了一种基于可解释机器学习的野火易发性模型，并详细分析了各因子对野火易发性预测结果的影响。以南宁市历史野火样本为基础，综合考虑样本的空间分布特征，选取高程、归一化植被指数(normalized difference vegetation index, NDVI)、年均降雨和平均气温等18项评价因子，利用分类和回归树(calssification and regression tree, CART)、随机森林(random forest, RF)、轻量的梯度提升机(light gradient boosting machine, LGBM)和极致梯度提升(extreme gradient boosting, XGBoost)4种机器学习模型构建野火易发性预测模型。基于性能最优的易发性模型，运用沙普利加和解释(shapley additive explanations, SHAP)方法完成特征全局性解释、依赖性分析和典型样本...

Study of Wildfire Hazard Susceptibility in Nanning Based on Interpretable Machine Learning Model

The assessment of wildfire susceptibility is crucial for the early prevention of wildfires and the development of disaster management strategies. Currently, research on wildfire susceptibility mainly focuses on improving the predictive accuracy of models while often neglecting the analysis and interpretation of the internal decision mechanisms of the models. Therefore, this study aims to construct an explainable machine learning-based wildfire susceptibility model and analyze in detail the influence of each factor on the wildfire susceptibility prediction results. Based on historical wildfire samples from Nanning, considering the spatial distribution characteristics of the samples, 18 evaluation factors including elevation, normalized difference vegetation index (NDVI), annual precipitation, and average temperature were selected. Four machine learning models, namely Classification and Regression Tree (CART), Random Forest (RF), Light Gradient Boosting Machine (LGBM), and Extreme Gradient Boosting (XGBoost), were employed to construct wildfire susceptibility prediction models. Based on the best-performing susceptibility model, the SHAP (SHapley Additive exPlanations) interpretable method was applied to achieve global feature explanations, dependency analysis, and local analysis of typical samples. The results showed that the XGBoost outperformed other models in terms of predictive performance, and the extremely high susceptibility zones were located in the northwest, east, and south of Nanning, accounting for 39.113 % of the total area. Wildfire susceptibility was mainly influenced by nine factors, including NDVI, annual precipitation, and soil type. The local interpretability results for typical historical wildfire samples can provide targeted references and guidance for wildfire disaster management in specific regions of Nanning.