A little data goes a long way: automating phase arrival picking at Nabro volcano, Eritrea, using transfer learning and a limited seismic catalog

Abstract:

Nabro volcano erupted with little warning in June, 2011. Eight broadband seismometers were subsequently deployed for 14 months, recording hundreds of volcano-seismic events per day. A core challenge with this dataset has been to accurately and efficiently identify and locate events: manual phase-arrival picking takes too long and most automated approaches require clear arrivals or existing event templates, or are computationally expensive.

To address this problem, we train and test deep learning models for automated P- and S-wave phase arrival detection using subsets of manually identified events for training (50 – 700 events). Test metrics show the most successful model to be one based on transfer learning (TL), utilising the lower-level feature extraction layers from an existing, extensively trained model as a starting point for model development. When combined with regularization techniques, the TL model requires only 500 examples of P-/S-wave arrivals and noise (~ 100 events), after which further training data provides only minor improvement. By comparison, training the same model from scratch requires 1750+ phase/noise examples (~ 350 events) before beginning to achieve similar validation loss.

When compared with its original base model (Ross et al., 2018, BSSA) and that of Zhu & Beroza (2019, GJI), the TL model achieves greater classification accuracy and smaller arrival time residuals on a test set of event and noise waveforms. By replacing the fully-connected layers from the original base model with convolutional layers, we greatly reduce the number of model parameters (more suitable for training with small datasets) and return an output with the same dimensions as the input waveform.The latter yields greater pick time precision and allows less overlap when used as a sliding window picking method, greatly reducing computational time (> 50x faster).

When run over the full 14-month deployment, the TL model picks P- and S-arrivals for > 38,000 locatable events (phase arrival picks on 4+ stations) in less than two hours of computational time on a single GPU. Even on a CPU, the model runs faster than real-time (24 hours of 3-component data in 16 secs), meaning that this approach can assist with seismic monitoring in resource poor areas.