【摘要】 相较于城市暴雨致灾事件（如内涝、洪水、泥石流等），现有研究对小粒度、多样化暴雨级联事件（如房屋损毁、地铁淹没等）风险的特征构成及其客观评估关注较少，难以适应城市精准化管理目标；同时，暴雨级联事件的风险评估模型构建面临样本数据风险特征不完备带来的模型效果约束。针对上述问题，考虑空间特征和风险特征的空间关联性，提出了融合风险特征和空间特征的城市暴雨级联事件风险评估模型构建方法。首先，面向不同暴雨级联事件的风险情景，从暴雨基层官员巡检、公民上报和社交媒体发帖数据中提炼级联事件风险特征；其次，以原始风险样本的空间定位为衔接，利用改进的边际Fisher方法从多源空间数据中挖掘空间特征，补充风险特征的缺失；最后，基于机器学习方法建立风险特征与风险类别的关联关系，构建多类别暴雨级联事件的风险评估模型。中国湖北省武汉市的实验结果表明：所提方法能够通过多源空间特征挖掘解决风险评估模型构建的特征不完备问题，实现多样化暴雨级联事件风险评估模型的有效构建，总体准确率、 F1得分以及AUC分别提升了23%、24%以及25%；同时，针对小粒度承灾体开展多样化级联事件风险评估，有助于更加精准的城市暴雨风险管理。更多还原

【Abstract】 Compared with urban rainstorm hazardous events(such as waterlogging, flood, debris flow, etc.),existing studies pay less attention to the feature composition and objective assessment of risks associated with small-scale and diversified rainstorm cascading events(such as house damage, subway inundation, etc.), making is difficult to meet the goals of refined city management. At the same time, constructing risk assessment models for rainstorm cascading events faces constraints due to incomplete risk features in sample data. To address these issues, this paper proposes a risk assessment model for urban rainstorm cascading events that integrates risk features and spatial features, considering the spatial correlation between them. Firstly, for the risk scenarios of different rainstorm cascading events, the risk features are extracted from data sources such as grassroots officials’ inspection, citizen reporting, and social media posts. Secondly, using the spatial localization of the original risk samples as a connection, an improved marginal Fisher method is employed to mine spatial features from multi-source spatial data to supplement the missing risk features. Finally, using a machine learning approach, the relationship between risk features and risk categories is established, leading to the construction of a risk assessment model for multi-category rainstorm cascading events. Experimental results from Wuhan, Hubei Province, China, show that the proposed method effectively addresses the problem of incomplete features in the construction of risk feature models through multi-source spatial feature mining, enabling the construction of diversified rainstorm cascading event risk assessment models. The overall accuracy, F1-score and AUC increased b y 2 3 %, 2 4 %, a n d 25%, respectively. Additionally, the complexity and diversity of spatial features highlighted the risks of subjective and arbitrary feature fusion, which can negatively affect the performance of machine learning model construction by adding irrelevant features. The proposed method mitigates this issue with an adaptive feature selection approach. Furthermore, grassroots officials’ inspection records contributed the most to the construction of urban rainstorm cascading event risk assessment models, followed by citizen-reported texts,and finally, social media data. Compared to traditional disaster event risk assessment methods, urban rainstorm cascading event risks have smaller risk granularity and involve more complex and diverse risk types and features. Traditional comprehensive evaluation models face challenges of subjectivity in manual evaluation,while traditional disaster loss curve methods encounter high experimental costs and data scarcity. The method proposed in this paper utilizes objective data to generate multidimensional risk features and establishes relationships between diverse risk levels, resulting in a machine learning-based risk prediction model that is more suitable for small-scale risk assessment scenarios.更多还原