Earthquake sensors may be helped by AI that strips out city noise

The sounds of cities can make it hard to discern the underground signals that indicate an earthquake is happening, but deep learning algorithms could filter out this noise

By Chris Stokel-Walker

A deep learning algorithm can remove city noise from earthquake monitoring tools, potentially making it easier to pinpoint when and where a tremor occurs.

“Earthquake monitoring in urban settings is important because it helps us understand the fault systems that underlie vulnerable cities,” says Gregory Baroza at Stanford University in California. “By seeing where the faults go, we can better anticipate earthquake events.”

However, the sounds of cities – from cars, aircraft, helicopters and general hustle and bustle – adds noise that makes it difficult to discern the underground signals that indicate an earthquake is happening.

To try to improve our ability to identify and locate earthquakes, Baroza and his colleagues trained a deep neural network to distinguish between earthquake signals and other noise sources.

Around 80,000 samples of urban noise and 33,751 samples of earthquake signals were combined in different forms to train, validate and test the neural network. The noise samples came from audio recorded in Long Beach, California, while the earthquake signals were taken from the rural area around San Jacinto, also in California. “We made many millions of combinations of the two to train the neural network,” says Baroza.

Running audio through the neural network improved the signal to noise ratio – the level of the signal you want to hear compared to the level of background noise – by an average of 15 decibels, three times the average of prior denoising techniques.

The research is very useful for the field, says Maarten de Hoop at Rice University in Houston, Texas. “It’s very well done, and I think beautiful work,” he says.

But he does highlight one drawback: the neural network was trained on data labelled by humans, a method called supervised learning, and the readings were all from one area. The fact that the model was supervised specifically to remove noise from sounds in California means it is less likely to be successful when presented with noise from elsewhere.

“The holy grail in this field is unsupervised learning,” says de Hoop. “If I go to one of the major cities in Japan, the chances this would work directly are pretty small, because it is supervised.”

Baroza is also unsure about how well the model would work in places other than California. “Depending on the environment, noise signatures are probably going to be different than the ones it’s trained on,” he says.

Journal reference: Science Advances, DOI: 10.1126/sciadv.abl3564