Pictures of streams and other bodies of water uploaded to the photo-sharing site Flickr could help scientists predict floods before they inundate communities, according to a new study. Though the approach doesn't provide as many data as traditional flood-forecasting models, it could become more powerful as social media becomes ever-more pervasive.
Hydrologists usually predict floods by monitoring storm conditions and the amount of rainfall. They also use wireless electronic sensors placed in and near bodies of water to measure water levels and drainage in real time. But sensor data aren't available from everywhere, and scientists are always looking for ways to extend how far in advance they and local officials can forecast a flood.
Enter social media. In recent years, researchers have used sites such as Twitter to track the extent of a flood by identifying and mapping tweets describing inundations of water. Scientists have also turned to Flickr, a photo-sharing website that allows anyone to upload images marked with tags and captions that also have corresponding location coordinates. In one study, for example, researchers noted a strong correlation between the change in atmospheric pressure during Hurricane Sandy and images uploaded to Flickr marked with tags related to the storm. As air pressure decreased—the mark of a hurricane—more people uploaded images with tags such as “Sandy.”
The new study, published last month in PLOS ONE, goes one step further by using tags to forecast extreme flooding days in advance. Researchers at the University of Warwick in Coventry, U.K., studied images and videos uploaded to Flickr during a 10-year period that were tagged with general water-related words such as “river” and “water” and matched to known flood events. The researchers then turned to a neural network, a computer model that can learn to spot patterns in large sets of data. By studying the 5-day period before and after known floods, they noticed that as a flood builds, the water-related tags are used more often with directly flood-related words like “flood” and “floodplain.” Watching for the same tags to emerge in real time could indicate when a flood is on its way.
“It’s actually quite powerful local knowledge that is currently not being exploited by the traditional early warning systems,” says study author and environmental and computer science researcher Nataliya Tkachenko.
Nonflash floods can already be predicted days in advance using traditional methods, and the Warwick team is not yet sure how accurate their tag-based forecasting method is in comparison. But it is among the first methods for using social media to predict flooding instead of simply observing it. It could also prove useful in areas where photography is more accessible than sensor-based water data.
"This is an interesting application of social media analytics … with many potential societal benefits,” says Brian Goode, a data scientist at Virginia Polytechnic Institute and State University in Blacksburg. However, he’s concerned that the team’s method used an inappropriate mathematical equation to model the relationship between the generic tags like “river” to the more specific tags such as “flood.” That could limit any prediction’s accuracy, he says.
What’s more, traditional flood-forecasting data provide concrete information such as the amount of rainfall and current water levels. The neural network can’t provide as many specifics; it simply indicates when it notices a pattern that, in the past, indicated a flood was on the way. Still, Tkachenko says her team’s approach could be combined with existing forecasting models to improve advance warning times.