Hurricanes are a recurrent disaster that significantly impact the United States.  This research examined how people affected by the flooding, resulting from Hurricane Harvey and Irma, used private and public social media to issue requests for help. We examined these practices in two phases. Phase I used field-based data collection methods to better understand social media posting, the use of 9-1-1, and the actions of rescuers and rescuees during these disasters. Private social media posts were captured to facilitate human and machine coding.  Phase II used the images captured during Phase I to develop strategies to classify image attributes. We also used public social media to further develop our classification schemes and compare those results to what we learned through private social media.  The project as a whole combined social science research with computer machine learning methods to develop new ways of thinking about data in disasters. 

Intellectual Merit     

The results from this project advance the frontiers of knowledge within and across fields the following ways.  First, we identified how rescuees used their mobile and social media so their calls for help were answered. We developed a model that explains how people make rescue decisions through a process that is socially constructed.  Specifically, some people post images of their disaster situation, and trusted others convince them to evacuate or seek help.  Once images are posted, they can be shared, amplified, and used as persuasion tools in efforts to save lives.  Second, we developed an archetype of behind-the-scene and physical-rescue volunteers who used mobile and social media in novel ways to rescue others. Third, we demonstrated that people had three key reasons to bypass 9-1-1 and post their calls for help on social media. (1) They could not get through to 9-1-1. (2) They heard 9-1-1 could not be reached so they did not try. (3) They felt that others needed more help, so they did not contact 9-1-1 to keep the phone lines open for those in serious need.         

Once we captured private social media images that reflected real calls for help, we used a series of machine learning studies to understand image attribute classification. These studies advance knowledge by identifying a process to create a supervised machine learning model that was more accurate than human coders at classifying who posted the images: rescuers or rescuees.  Second, we used a small human-coded data set to train our unsupervised machine learning models to successfully classify publicly posted images according to their urgency, relevancy, and damage.  What makes these findings meaningful is that past research often focused on text-based analyses, and our findings suggest that computer-vision-derived attributes offer promising results when the data is image-based rather than text.     

Broader Impacts

By answering the question ?Can a machine identify features in an image as well as a human?? in crisis-related data, we contribute to computer science and computer engineering by providing evidence that high-signal data is useful in and of itself to computer vision. Furthermore, we find that private social media data contains much less noise than public social media data, and the private data can yield much higher classification successes than public social media data.  Our qualitative findings contribute to disaster sociology, geography, communication, and engineering disciplines like civil engineering.  Understanding the human side of disasters is increasingly important considering that engineering and social science disciplines are regularly teaming to produce more comprehensive results from studies of technology. The developed models articulate how disaster decision making often is socially constructed, and thus social media helps to make situations visible and facilitates the rescue process. 

This project has broad impact on society because almost all people experience some type of natural disaster during their lifetime.  Furthermore, social media is being used prolifically by people of diverse backgrounds. The computer models we developed have potential application as an alternative or supplement to emergency response systems, highlighting groups or individuals who are potential rescuers or who are in need of rescuing.  The human behavior and communication models developed offer a more complete understanding around issues of using 9-1-1 for emergencies, as well as demonstrate the role that mobile and social media will continue playing during disaster situations around the globe. 

Finally, this project has supported opportunities for undergraduate and graduate student research in communication, computer science and computer engineering.  Our team?s outputs offer clear examples of how disaster-related innovations require interdisciplinary research.   

Last Modified: 10/31/2019
Modified by: Keri K Stephens