A Little Data Goes a Long Way: Automating Seismic Phase Arrival Picking at Nabro Volcano With Transfer Learning

Seismic monitoring plays a fundamental role in mitigating hazards near volcanoes, where thousands of earthquakes can occur each day during periods of unrest. These events produce a diverse range of seismic signals, and the detection of the initial primary (P-) and secondary/shear (S-) wave arrivals forms the basis of most seismic processing tasks undertaken to determine earthquake locations, magnitudes and source parameters. Because seismic networks gather vast amounts of data from tens or even hundreds of seismometers at once, real-time manual inspection of these time series is practically infeasible. Researchers have for this reason developed automated methods including a recently successful approach based on supervised deep learning models.
These machine learning models exploit convolutional neural networks (CNN), a variant of classical neural networks which employ convolution operations. Such models have advantages over traditional algorithms, yet also require substantial amounts of labeled training data to perform well on out-of-sample data. Moreover, as with practically any deep learning model, these models often perform poorly when faced with data that differs in source or distribution from their training data. These demands make these models unsuited for many real-world applications.
In a recent paper, LML External Fellow Maximilian Werner and colleagues explore an alternative approach based on so-called inductive transfer learning. They exploit knowledge acquired from training a model on millions of seismic waveforms recorded by the Southern California Seismic Network and apply it to seismograms from Nabro volcano in Eritrea. The researchers use the feature extraction layers of an existing, extensively trained seismic phase picking model to build a new all-convolutional model called U-GPD, and demonstrate that transfer learning reduces overfitting and model error relative to training the same model from scratch. This is particularly true for small training sets (e.g., 500 waveforms). As the work shows, the new U-GPD model achieves greater classification accuracy and smaller arrival time residuals than two existing, extensively-trained baseline models for a test set of 800 event and noise waveforms from Nabro volcano. When applied to 14 months of continuous Nabro data, the new U-GPD model detects 31,387 events with at least four P-wave arrivals and one S-wave arrival, which is more than the original base model (26,808 events) and an existing manual catalog (2,926 events), with smaller location errors.
The paper is available at https://agupubs.onlinelibrary.wiley.com/doi/full/10.1029/2021JB021910

Leave a Reply

Your email address will not be published. Required fields are marked *