[论文解读] Constructing Evacuation Evolution Patterns and Decisions Using Mobile Device Location Data: A Case Study of Hurricane Irma

Aref Darzi, Vanessa Frías-Martínez+3 more|arXiv (Cornell University)|Feb 24, 2021

Human Mobility and Location-Based Analysis参考文献 34被引用 5

一句话总结

本研究利用佛罗里达州飓风玛丽亚期间超过110亿次的匿名手机定位数据，对疏散行为进行建模，发现强制疏散区57.92%的居民选择撤离，而无指令区仅为32.98%。该框架将个人移动历史（如出行频率和空间覆盖范围）整合进选择模型，显著提升了基于现实行为模式的疏散决策预测准确性。

ABSTRACT

Understanding individuals' behavior during hurricane evacuation is of paramount importance for local, state, and government agencies hoping to be prepared for natural disasters. Complexities involved with human decision-making procedures and lack of data for such disasters are the main reasons that make hurricane evacuation studies challenging. In this paper, we utilized a large mobile phone Location-Based Services (LBS) data to construct the evacuation pattern during the landfall of Hurricane Irma. By employing our proposed framework on more than 11 billion mobile phone location sightings, we were able to capture the evacuation decision of 807,623 smartphone users who were living within the state of Florida. We studied users' evacuation decisions, departure and reentry date distribution, and destination choice. In addition to these decisions, we empirically examined the influence of evacuation order and low-lying residential areas on individuals' evacuation decisions. Our analysis revealed that 57.92% of people living in mandatory evacuation zones evacuated their residences while this ratio was 32.98% and 33.68% for people living in areas with no evacuation order and voluntary evacuation order, respectively. Moreover, our analysis revealed the importance of the individuals' mobility behavior in modeling the evacuation decision choice. Historical mobility behavior information such as number of trips taken by each individual and the spatial area covered by individuals' location trajectory estimated significant in our choice model and improve the overall accuracy of the model significantly.

研究动机与目标

• 利用大规模移动设备定位数据，理解重大飓风期间个体的疏散行为。
• 识别疏散指令和地理风险（如低洼地区）对个体决策的影响。
• 通过整合个人移动行为（如出行频率和空间范围）来建模疏散决策。
• 利用现实世界定位数据提取的行为与情境特征，提升疏散选择预测的准确性。
• 构建一个数据驱动的框架，实现实时分析灾害期间疏散演变模式与决策。

提出的方法

• 使用飓风玛丽亚期间佛罗里达州超过110亿次匿名手机位置服务（LBS）数据。
• 应用时空聚类与轨迹重建方法，识别个体移动模式与疏散事件。
• 采用逻辑回归构建选择模型，预测疏散决策，整合个人移动特征（如出行次数与轨迹空间覆盖范围）。
• 将疏散区划分为强制、自愿和无疏散指令三类，以分析政策影响。
• 通过模型准确率指标，实证评估历史移动行为对疏散决策的预测能力。
• 利用基于定位轨迹推导的出发与返回日期真实数据，对模型进行验证。

实验结果

研究问题

• RQ1 在飓风玛丽亚期间，不同类型的疏散区（强制、自愿、无指令）中，有多少比例的个体选择撤离？
• RQ2 个体的历史移动模式（如出行频率与空间范围）如何影响其疏散决策？
• RQ3 在低洼居民区，疏散指令的存在在多大程度上提高了疏散的可能性？
• RQ4 不同人口统计与地理群体的出发与返回时间分布有何差异？
• RQ5 移动行为特征是否能显著提升疏散决策预测模型的准确性？

主要发现

• 强制疏散区57.92%的居民选择撤离，而无指令区为32.98%，自愿疏散区为33.68%。
• 历史移动性更高的个体（以出行次数与轨迹空间覆盖范围衡量）更可能选择撤离。
• 整合移动行为特征后，疏散决策预测模型的整体准确率显著提升。
• 出发与返回日期的分布呈现出明显的时序模式，疏散高峰出现在风暴登陆前。
• 官方疏散指令的存在显著影响了疏散决策，尤其在高风险低洼地区更为明显。
• 模型表现出强劲的预测性能，表明移动设备的行为数据可有效捕捉现实中的疏散动态。