Research on AI-Based Earthquake Damage Assessment Model Using Multi-Source Data Fusion
ZHAO Lirong1,CUI Qian2,GAO Mingjie3
1.Shanxi Conservancy Technical Institute, Taiyuan 030032, China;2.Shanxi Conservancy Technical Institute, Yuncheng 044000, China;3.Shanxi Dibao Energy Co., Ltd.,Taiyuan 030045, China
The suddenness and destructiveness of earthquake disasters pose a severe challenge to the post-disaster emergency response system, and it is urgent to solve the predicament of insufficient timeliness of traditional assess⁃ment methods and poor interpretability of complex models. This study integrated multi-source data such as remote sensing images, seismic monitoring, GIS, and hazard-affected body attributes to construct a multi-dimensional data system covering disaster-causing factors, disaster-bearing environments, and hazard-affected body characteristics, and proposed a combined modeling approach of ensemble learning and single model to form a disaster damage as⁃sessment model centered on random forests. And through case verification, it is revealed that the disaster damage as⁃sessment model proposed in this study has good capabilities in disaster damage level classification, severe damage identification and spatial aggregation feature reproduction, which can provide technical support for rapid disaster identification and resource allocation in earthquake emergency response.
基于前期构建的多源融合特征向量,本研究采用集成学习与单一模型相结合的思路构建最终的灾损评估模型,其中以随机森林作为核心模型。该模型的构建过程如下:首先,将每个样本的融合特征向量Xi =( x 1 , x 2 , ..., xp )作为输入,其中p为经过特征选择后保留的特征维度。模型通过Bootstrap抽样生成k个训练子集,并行构建k棵决策树。在每棵树的节点分裂时,通过最小化基尼不纯度或信息增益等不纯度度量指标来选择最优特征及分裂点,其基尼不纯度计算遵循公式
,其中C为灾损等级数量,Pi为节点中属于第i类的样本比例,通过递归分割构建出多样化的弱分类器。对于分类任务,最终输出采用多数投票法则,即
=mode{T 1 ( X ), T 2 ( X ), ..., T K ( X )}。此集成策略有效提升了模型的泛化能力与稳健性。在模型训练阶段,通过网格搜索或随机搜索结合交叉验证对关键超参数等进行优化,以在保证模型复杂度的同时避免过拟合。为了量化模型的置信度,模型可输出所有决策树对某个预测类别的投票比例,作为该预测结果的不确定性度量。该完整模型不仅完成了从多源特征到灾损等级的映射,更通过袋外得分进行无偏估计。整个构建流程确保了模型在具备高精度非线性拟合能力的同时,保留了可通过特征重要性排序进行物理归因的可解释性,形成了一个从特征输入、集成决策到结果输出的完整、可靠的评估体系。
[11]王健,黄敏.基于数据挖掘的地震灾损率研究[C]//Wuhan University, Chung Hua University, University of Science and Technology of China, et al. Proceedings of International Conference on Engineering and Business Management (EBM2012). Shanghai: ISTP, 2012: 3926-3928.