Feasibility of Twitter Based Earthquake Characterization From Analysis of 32 Million Tweets: There’s Got to be a Pony in Here Somewhere!

Monday, 15 December 2014
Paul S Earle1, Michelle R Guy1, Greg M Smoczyk1, Scott R Horvath2, Turner S Jessica1 and Douglas B Bausch3, (1)USGS National Earthquake Information Center Golden, Golden, CO, United States, (2)USGS Headquarters, Reston, VA, United States, (3)FEMA, Lakewood, CO, United States
The U.S. Geological Survey (USGS) operates a real-time system that detects earthquakes using only data from Twitter—a service for sending and reading public text-based messages of up to 140 characters. The detector algorithm scans for significant increases in tweets containing the word “earthquake” in several languages and sends internal alerts with the detection time, representative tweet texts, and the location of the population center where most of the tweets originated. It has been running in real-time for over two years and finds, on average, two or three felt events per day, with a false detection rate of 9%.

The main benefit of the tweet-based detections is speed, with most detections occurring between 20 and 120 seconds after the earthquake origin time. This is considerably faster than seismic detections in poorly instrumented regions of the world. The detections have reasonable coverage of populated areas globally. The number of Twitter-based detections is small compared to the number of earthquakes detected seismically, and only a rough location and qualitative assessment of shaking can be determined based on Tweet data alone. However, the Twitter-based detections are generally caused by widely felt events in populated urban areas that are of more immediate interest than those with no human impact. We will present a technical overview of the system and investigate the potential for rapid characterization of earthquake damage and effects using the 32 million “earthquake” tweets that the system has so far amassed.

Initial results show potential for a correlation between characteristic responses and shaking level. For example, tweets containing the word “terremoto” were common following the MMI VII shaking produced by the April 1, 2014 M8.2 Iquique, Chile earthquake whereas a widely-tweeted deep-focus M5.2 north of Santiago, Chile on April 4, 2014 produced MMI VI shaking and almost exclusively “temblor” tweets. We are also investigating the use of other social media such as Instagram to obtain rapid images of earthquake-related damage. An Instagram search following the damaging M6.9 earthquake near the Mexico, Guatemala boarder on July 7, 2014 reveled half a dozen unconfirmed images of damage; the first posted 15 minutes after the event.