A Little Data Goes a Long Way: Automating Seismic Phase Arrival Picking at Nabro Volcano With Transfer Learning. Issue 7 (12th July 2021)

• Record Type: Journal Article
• Authors: Lapins, Sacha; Goitom, Berhe; Kendall, J‐Michael; Werner, Maximilian J.; Cashman, Katharine V.; Hammond, James O. S.
• Is Part Of: Journal of geophysical research. Volume 126:Issue 7(2021)
• Page Start: n/a
• Page End: n/a
• Publication Date: 2021-07-12
• Subjects: volcano seismology -- machine learning -- transfer learning -- phase arrival detection -- earthquake detection -- volcano monitoring; Geomagnetism -- Periodicals; Geochemistry -- Periodicals; Geophysics -- Periodicals; Earth sciences -- Periodicals; 551.1
• DOI: 10.1029/2021JB021910
• Languages: English
• ISSNs: 2169-9313

Abstract

Abstract: Supervised deep learning models have become a popular choice for seismic phase arrival detection. However, they do not always perform well on out‐of‐distribution data and require large training sets to aid generalization and prevent overfitting. This can present issues when using these models in new monitoring settings. In this work, we develop a deep learning model for automating phase arrival detection at Nabro volcano using a limited amount of training data (2, 498 event waveforms recorded over 35 days) through a process known as transfer learning. We use the feature extraction layers of an existing, extensively trained seismic phase picking model to form the base of a new all‐convolutional model, which we call U‐GPD. We demonstrate that transfer learning reduces overfitting and model error relative to training the same model from scratch, particularly for small training sets (e.g., 500 waveforms). The new U‐GPD model achieves greater classification accuracy and smaller arrival time residuals than off‐the‐shelf applications of 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 our existing manual catalog (2, 926 events), with smaller location errors. The new model is also more efficient when applied as a sliding window, processing 14 months of data from seven stations in less than 4 h on a single graphics processing unit. Plain Language Summary: Seismic monitoring increasingly relies on automated signal processing as the rate of data acquisition grows. Supervised deep learning models have proven to be effective for detecting and characterizing seismic events, but training such highly parameterized models generally requires large amounts of manually labeled data. Once trained, however, these models extract general seismic waveform features that can be used to train new models with more limited training data. In this work, we use the generalized knowledge of seismic data from a model trained on millions of earthquakes in California to train a new model for detecting volcanic earthquakes at Nabro volcano, Eritrea, a recently active and, prior to its 2011 eruption, poorly monitored volcano. Using a small training set of waveforms, the new model more accurately detects phase arrivals and noise than off‐the‐shelf applications of two baseline models. The new model is efficient, processing 14 months of data in less than 4 h. It is also effective, detecting more volcanic events and showing improved levels of S‐wave arrival picking. The result is smaller event location errors than even our manual picks. This level of efficiency and consistency highlights the role that machine learning can play in volcano‐seismic monitoring. Key Points: Transfer learning using existing model trained on California earthquake data produces effective new model for monitoring at Nabro volcano Nabro transfer learning model shows improved S‐wave picking resulting in smaller location errors than even manual phase picks Changing task from classification to segmentation results in more efficient model processing 14 months of data from seven stations in 4 h