Humans outperform AI in identifying earthquakes and tremors from seismic recordings, finds study
This opens up avenues for roping in more citizens to work with geologists
A small army of earthquake detectives has outperformed artificial intelligence in identifying tremors or earthquakes from oceans of seismic recordings, according to a new study. This opens up avenues for roping in more citizens to work with geologists.
It is not possible to predict earthquakes yet. An instrument called a seismometer detects ground motions and generates recordings or data. Scientists will have to scour through them to identify quakes, volcanic eruptions and explosions. Though artificial intelligence has been of help, it has failed to detect tremors from the data. Tremors are weak vibrations that are believed to play a part in the origin of earthquakes.
"My aim was to receive help with detections of these special seismic events because I felt overwhelmed by the rapidly growing mountain of data I was investigating for my PhD research," lead author Vivian Tang, a graduate student at the Department of Earth and Planetary Sciences of Northwestern University, said. "With Zooniverse and the Earthquake Detective team, we provide people everywhere with a simple and engaging way to help further scientific research," she added.
Students and faculty at Northwestern University kickstarted a project named Earthquake Detective in 2019 to enable volunteers to contribute. "Classifying this data will help us paint a more complete picture of when larger earthquakes may trigger smaller ones or tremors," Suzan van der Lee, a seismologist and professor of Earth and planetary sciences, had earlier said in a statement. "Then we can learn what conditions are favorable and what factors need to align to allow them to happen, which ultimately also might inform the reverse of how small seismic events interact with big ones."
In this study, Tang and her colleagues enrolled 2,000 citizen scientists who were given an hour's training. Then, the team tested participants by making them listen to different recordings. The task was to classify them into four categories: earthquake, tremor, background noise, or none of the above. The researchers also looked at consensus, which is the number of times the volunteers agreed on a particular category. Machine learning algorithms were also evaluated.
The researchers describe seismic recordings of earthquakes as something that sounds like a slamming door. Tremors, on the other hand, are more like a train going over railway tracks. Background noise is like whistling wind, crinkling tin foil, or radio station, they added.
Among the various classifications, the participants showed more agreement while detecting earthquakes than tremors and background noise. The researchers then checked if the collective decisions were accurate: the citizens were 85% accurate in identifying quakes, higher than the machine learning algorithm's 76%. Artificial intelligence failed at identifying tremors, making volunteers particularly invaluable in the field of seismology.
"Earthquake Detective can be a resource for other researchers in this field who are interested in receiving input from an impressive group of volunteer scientists. We strongly encourage these researchers to point us to seismograms they would like to see classified so that we can include them in Earthquake Detective, and return the volunteer classifications to the researchers," says Tang. The study was published in Frontiers in Earth Science.