A group of researchers, including one from the University of Michigan, have found that social media could help the public and emergency responders gauge how much damage was inflicted following a natural disaster, and subsequently predict the amount of relief funds given to the community following a natural disaster.

The study found a moderate positive correlation between Twitter activity in an area and the level of damage inflicted by Hurricane Sandy in 2012.

Engineering Prof. Pascal Van Hentenryck initiated the project as part of his optimization research on how better predictive models can be used to study complex infrastructure, natural phenomena and human behavior.  

Van Hentenryck said the idea for the research came as he was looking for alternatives to existing mechanisms currently used to assess and predict damage to power networks and electric generators caused by natural disasters.

“One of the issues that I was dealing with at that point was the assessment and repair of the power system,” Van Hentenryck said. “How can we use social media to refine the prediction of damage of the network? Maybe it can give a better estimation of what is down and what is not down.”

This curiosity about forming predictive methods for damage to electricity networks created the starting hypothesis for the research study.

“We wanted to find out if there is some correlation between Twitter activity and the damage inside a power network. The hypothesis was social media can be predictive, somewhat, of the damage done,” Van Hentenryck said.

Launched in 2012, the team consisted of a variety of scientists, researchers and graduate students from five universities. Prior to coming to the University, Van Hentenryck led a research group at National Information Communications Technology Australia. It was there that he engaged long-time friend Manuel Cebrian as co-author for the research project. Currently based in Melbourne, Cebrian works as research team leader at the Commonwealth Scientific and Industrial Research Organization.

Cebrian said using real-time data from social media to evaluate complex events that have far-reaching and long-term consequences is significant because these events are difficult to measure quickly.

“We wonder if we could actually do something that would mix two very different time scales — one that is very slow and take a long time to assess, the other one would be social media, which will be very fast,” Cebrian explained. “Natural disasters produce massive damage; it’s estimated that damage takes months to assess. When you estimate damage in one community, usually what you do is send experts, experts conduct assessment with technology, verify that damage was caused by the disaster and it takes sometimes about a year to do this. So we asked: can we infer some of this from the way people tweet during the disaster?”

For the purpose of this study, “damage” was quantified by the amount of disaster relief funds issued by the Federal Emergency Management Agency. The team analyzed 52 million geographically pinpointed tweets posted before, during and after the hurricane. In order to ensure that tweets were filtered for relevance, they searched for keywords such as “FEMA” and “hurricane.”  

“We very carefully gathered tweets about the hurricane in the east coast, and obtained data from both FEMA and private as well as state-level insurance claims,” Cebrian said. “We put everything together and then started to measure the specific correlations between them.” 

Given the large amount of data, Cebrian said the team was careful to ascertain that statistical results could be reproducible on the city, county and national levels.

“The third phase of our project looked at the question: are the signals robust? We needed to make sure that what we found was meaningful and reliable. It took almost a year to write up the results and make sure that they are presented in an intelligible fashion,” Cebrian said.

The research found that areas that tweeted more post-disaster filed the most insurance claims and received the most individual assistance from FEMA grants. Similar studies on 10 to 15 other disasters revealed the same correlation.

“There is a positive correlation between the level of twitter activity and the damage in a particular area,” Van Hentenryck said. “The interesting thing is that it’s not when the hurricane is passing through the region, it’s a couple of days after. At the peak of the event, almost every location will have a lot of tweets, even if they aren’t directly on the path of the hurricane.”

The researchers said their findings are significant because the uncovered correlation between the volume of tweets and local damages is actually stronger than those employed by fragility simulations currently used in federal emergency response efforts. Nonetheless, Cebrian stressed the research findings should be used as a complement to, rather than replacement of existing forecasting and predictive models.

“We wanted to be very cautious about this,” Cebrian said. “Correlation is positive, but moderate, can predict some of the damage that is going to happen, but not all of it. This could be due to the fact that maybe we don’t have enough data, or maybe Twitter doesn’t capture 100 percent of the damage.”

Van Hentenryck said he hopes federal emergency services will leverage on the predictive potential of social media, now that this peer-reviewed study has been published.

“I think emergency services would like to tap into that potential — they don’t really know how to do it, they don’t really know the reliability,” he said. “For me, this is exciting because it gives them positive indication that social media is another tool that they can use.”

Leave a comment

Your email address will not be published.