Collaborative Research: Leveraging Massive Smartphone Location Data to Improve Understanding and Prediction of Behavior in Hurricanes

在这个项目中，新获得的匿名智能手机位置数据将用于极大地提高对飓风期间家庭行为的理解（例如，有多少人将撤离，何时撤离，如何撤离，从哪里撤离，以及前往哪里）。尽管先前的研究提供了关于飓风中人口行为的宝贵知识，但重要的差距仍然存在。现有的模型预测未来飓风行为的能力有限。不同类型的家庭和人（如游客或没有车辆的人）之间的行为差异并不为人所知。在飓风持续期间，个人身上发生的事件的顺序和时间也不重要。这些差距很大程度上是由于支持过去研究的传统数据类型的局限性——调查、访谈和焦点小组。该项目将通过利用一种新型数据的可用性——来自智能手机的匿名位置信息——促进疏散行为建模科学的发展，从而在理解和预测飓风疏散期间人口的行为方面取得飞跃。该项目将大大提高管理未来疏散的能力，从而促进国家福利，造福社会。在飓风期间，官员们要做出许多重要的决定，包括发布官方疏散命令、向公众发布信息、开放避难所、安排物资和人员、实施特殊交通计划、为没有车辆的人群提供支持，以及准备进行救援。所有这些都直接取决于预计有多少人撤离，何时，如何，从哪里以及到哪里。通过对飓风期间的人口行为提供更准确、更细致的预测，该项目将使官员能够以更明智、更有效的方式做出决策。为了确保研究结果能够迅速有效地转化为实践，这项研究的设计使其能够纳入应急管理人员目前使用的决策工具和流程。我们来自联邦紧急事务管理局（FEMA）以及佛罗里达州和北卡罗来纳州紧急事务管理机构的从业伙伴也将帮助我们与更大的紧急事务管理界分享调查结果。这项研究将有助于制定一种程序，以便实时获取和分析其他疏散事件的类似数据。新的智能手机位置数据的可用性为改变飓风中人口行为的研究提供了难得的机会。这些数据提供了许多好处，包括比以前典型的大几个数量级的样本；提供有凝聚力的个人行为时间表；提供不受回忆或报告偏差影响的直接观察；能够在24小时内移动；并且在许多飓风中以低成本的形式持续存在。将新数据的力量与基于传统调查和访谈数据的领域专业知识相结合，将从五个方面推动这一领域的科学发展。首先，我们将通过使用更大、独立的数据集和过去不易检验的新假设来检验传统文献中的假设，从而提高知识水平。其次，在飓风过程中可能会发生多种事件，包括飓风相关事件（例如，飓风转向，加强），官方行动（例如，发布官方命令，关闭学校）和个人事件（例如，停止工作）。在飓风持续期间，每个人都会先后经历一些或所有这些事件。我们将使用顺序模式挖掘来描述关键的可观察事件和动作，它们的可能序列，不同序列的概率，以及每个事件的持续时间分布。这种对个人事件的顺序和时间的建模，以前从未做过，将阐明飓风行为、官方行动、个人决定和时间标记相互作用和展开的方式，并有助于确定有希望的疏散支持干预点。第三，我们将开发新的统计模型，以预测个人/家庭属性、官方事件、飓风、预报、时间标记和飓风形成以来的过去行为的函数来预测人们在每个时间段撤离并前往特定地理目的地的概率。这些模型将通过识别小数据集无法观察到的对行为的影响，提供改进的样本外预测能力；通过提高预测地理目的地的能力，这对估计清关时间很重要；而且，这是第一次，利用对行为的早期观察，这可能是最终行为的主要指标。第四，我们将测试用于预测通关时间的交通模型中隐含的路线选择假设，并确定道路封闭对疏散和再入期间交通模式的影响。新的数据将使测试比以前通过孤立的交通统计和调查进行的测试更详细、更全面。最后，我们将使用一般归纳方法确定未来传统研究的新行为和问题。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

In this project, newly available anonymous smartphone location data will be used to dramatically improve understanding of how households behave during hurricanes (e.g., how many people will evacuate, when, how, from where, and to where). Although previous research has provided valuable knowledge about population behavior in hurricanes, important gaps remain. Available models have limited ability to predict behavior in future hurricanes. Differences in behavior across different types of households and people, such as tourists or people without vehicles, are not well known. Neither are the sequence and timing of events that unfold for individuals over the duration of a hurricane. These gaps are largely due to limitations in the traditional types of data that have supported past research—surveys, interviews, and focus groups. This project will promote the science of modeling evacuation behavior by capitalizing on the availability of a new type of data— anonymous location information from smartphones—to make a leap forward in understanding and predicting the behavior of the population during hurricane evacuations. The project will advance national welfare and benefit society by substantially improving the ability to manage future evacuations. During a hurricane, officials make many highly consequential decisions, including issuing official evacuation orders, messaging the public, opening shelters, staging materials and staff, implementing special traffic plans, executing support for vehicle-less populations, and preparing to undertake rescues. All of these depend directly on how many people are expected to evacuate, when, how, from where, and to where. By providing a more accurate and nuanced prediction of population behavior during hurricanes, this project will enable officials to make those decisions in a more informed and effective way. To ensure findings will be translated to practice quickly and effectively, the research has been designed so that it can be integrated into the current decision-making tools and processes used by emergency managers. Our practitioner partners from the Federal Emergency Management Agency (FEMA) and the Florida and North Carolina state emergency management agencies will also help us share findings with the larger emergency management community. This study will facilitate the development of a procedure to acquire and analyze, in real time, similar data for other evacuation events.Availability of new smartphone location data offers a rare opportunity to transform the study of population behavior in hurricanes. The data offers many benefits, including samples that are orders of magnitude larger than previously typical; offering cohesive timelines of individual behavior; providing direct observations not subject to recall or reporting bias; being available within 24 hours of movement; and being available at low cost in consistent form for many hurricanes. Combining the power of the new data with domain expertise based on traditional survey and interview data will advance the science in this area in five ways. First, we will improve knowledge by testing hypotheses from the traditional literature using a larger, independent dataset and new hypotheses not easily testable in the past. Second, multiple events may happen during the course of a hurricane, including hurricane-related events (e.g., hurricane turns, intensifies), official actions (e.g., issue official orders, close schools), and personal events (e.g., released from work). Each person experiences some or all of these events in a sequence over a hurricane’s duration. We will use sequential pattern mining to describe key observable events and actions, their possible sequences, the probabilities of different sequences, and duration distributions of each event. This modeling of the sequence and timing of events for individuals, which has not been done before, will illuminate the range of ways hurricane behavior, official actions, personal decisions, and time markers interact and unfold, and help identify promising points of intervention for evacuation support. Third, we will develop new statistical models to predict the probability a person will evacuate at each time period and go to a particular geographic destination as a function of attributes of the individual/household, official events, hurricane, forecast, time markers, and past actions since the hurricane formed. These models will offer improved out-of-sample predictive power by identifying influences on behavior that are not observable with small datasets; by improving the ability to predict geographic destination, which is important for estimating clearance times; and by, for the first time, taking advantage of observations of behavior early in the event that may be leading indicators of final behavior. Fourth, we will test the route choice assumptions implicit in traffic models used to predict clearance times, and determine the effects of road closures on traffic patterns during evacuation and reentry. The new data will allow testing that is more detailed and comprehensive than previously possible through isolated traffic counts and surveys. Finally, we will identify new behaviors and questions for future traditional research using a general inductive approach.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.