灾害事件下人类活动群体异常探测

Developing capable anomaly detection methods for discovering spatiotemporal patterns of urban responses to disaster events from geolocation big data has become one of the research frontiers in disaster science. It contributes to understanding patterns of human behavior under the disturbance of disaster and provides innovative ways to disaster monitoring and loss reduction. Existing studies focus much on identification of the space-time points or clusters that have demonstrated significant abnormality. However, the lack of detecting collective anomalies, especially the latent anomalies which indicate potent impact of disaster in the near future, limits the ability to recognize the disaster occurrence before the outbreak of its adverse effects. To address the problem, this project aims to create a collective anomaly detection method that enables the diagnosis of the latent anomalies. The method extends the ability to detect latent anomalies by extracting spatiotemporal features and constructing a classification model for the latent anomalies. To examine the effectiveness and value of the method in disaster monitoring and management, this project plans to conduct an empirical study of typhoon and rainstorm disasters and investigate the collective anomalies in human responses to the disasters using smartphone location datasets.

利用时空异常探测方法从位置大数据中挖掘灾害事件下人类活动的响应特征是当前灾害研究的前沿方向，有助于揭示灾害事件下人群异常响应的模式与机理，并最终为实时掌握灾情，实施救援提供大数据方法支持。但现有方法主要关注已显现异常的时空点和聚簇的提取，缺乏对潜在异常的甄别及包含潜在异常的群体探测分析，限制了现有方法在异常爆发之前就对灾害事件的发生作出预判的能力。所以，本项目拟在现有异常探测研究的基础上，建立一种顾及潜在异常诊断的群体异常探测方法，通过构建潜在异常的判别指标与诊断模型，拓展潜在异常的甄别能力，实现群体异常探测的方法创新。本研究将以台风、暴雨内涝等典型城市灾害为例开展实证研究，通过手机定位大数据挖掘，揭示灾害引起的城市群体异常，检验方法提高灾害监测与应急管理能力的应用价值。

为了从蕴含人类活动信息的时空大数据中挖掘灾害事件对城市节律造成的影响，本项目构建了面向群体异常的时空探测与分析技术体系，从时序变化和空间特征建立综合异常评估指标，更好地揭示灾害导致的群发性空间异常。项目收集了全国范围的定位请求数据和含坐标微博数据，以降雨和台风为例开展了方法应用研究，通过群体异常探测方法有效识别出灾害过程中的时空异常，发现异常响应特征与灾害的发生与发展过程相一致，通过回归建模可以根据异常程度快速估算城市灾损情况。并且通过微博文本解析，可以解析灾害过程中社会关注焦点的变化，也能揭示不同城市的居民对降雨事件的感知存在敏感性差异。本项目的研究结果有利于监测和量化灾害对城市造成的社会影响，为刻画城市灾害韧性提供新方法。