Understanding human movements in the face of natural disasters is critical for disaster evacuation planning, management, and relief. Despite the clear need for such work, these studies are rare in the literature due to the lack of available data measuring spatiotemporal mobility patterns during actual disasters. This study explores the spatiotemporal patterns of evacuation travels by leveraging users’ location information from millions of tweets posted in the hours prior and concurrent to Hurricane Matthew. Our analysis yields several practical insights, including the following: (1) We identified trajectories of Twitter users moving out of evacuation zones once the evacuation was ordered and then returning home after the hurricane passed. (2) Evacuation zone residents produced an unusually large number of tweets outside evacuation zones during the evacuation order period. (3) It took several days for the evacuees in both South Carolina and Georgia to leave their residential areas after the mandatory evacuation was ordered, but Georgia residents typically took more time to return home. (4) Evacuees are more likely to choose larger cities farther away as their destinations for safety instead of nearby small cities. (5) Human movements during the evacuation follow a log-normal distribution.

Tracking spatial and temporal trends of events (e.g. disease outbreaks and natural disasters) is important for situation awareness and timely response. Social media, with increasing popularity, provide an effective way to collect event-related data from massive populations and thus a significant opportunity to dynamically monitor events as they emerge and evolve. While existing research has demonstrated the value of social media as sensors in event detection, estimating potential time spans and influenced areas of an event from social media remains challenging. Challenges include the unstable volumes of available data, the spatial heterogeneity of event activities and social media data, and the data sparsity. This paper describes a systematic approach to detecting potential spatiotemporal patterns of events by resolving these challenges through several interrelated strategies: using kernel density estimation for smoothed social media intensity surfaces; utilizing event-unrelated social media posts to help map relative event prevalence; and normalizing event indicators based on historical fluctuation. This approach generates event indicator maps and significance maps explaining spatiotemporal variations of event prevalence to identify space-time regions with potentially abnormal event activities. The approach has been applied to detect influenza activity patterns in the conterminous US using Twitter data. A set of experiments demonstrated that our approach produces high-resolution influenza activity maps that could be explained by available ground truth data.