Skip to content

Enhancing Phase Detection of Volcanic Earthquakes Through a Multi-Station Machine Learning Model

^^ Robinson, A.G^^ .1, Hammond, J.O.S.1, Steele, A.2 Lapins, S.3

  • Affiliations: 1Birkbeck, University of London, Malet St, London WC1E 7HX, arobin24@student.bbk.ac.uk 2University College London, Gower Street, London WC1E 6BT 3University of Bristol, Beacon House, Queens Rd, Bristol BS8 1QU 

  • Presentation type: Poster

  • Presentation time: Monday 16:30 - 18:30, Room Poster Hall

  • Poster Board Number: 128

  • Programme No: 3.1.20

  • Theme 3 > Session 1

Abstract

Seismic phase picking is a crucial step in many volcano monitoring workflows. Over the past decade, several deep learning algorithms have been developed for this task and have shown impressive results. However, despite many volcanoes hosting a multi-station network, existing deep learning models for phase picking still process data from each station independently. As a result, these models are 'blind' to the relationships between seismograms from different stations that capture the same underlying event. Unlike these models, human analysts frequently use contextual information from multiple stations to pick phase arrivals at stations where the signal to noise ratio is low, and to rule out spurious signals. Here, I present ongoing work to develop a multi-station deep learning phase picker that takes raw seismograms from multiple stations as input and incorporates context from neighbouring stations to update the probability of phase arrivals at individual stations. The model architecture has been designed using graph layers which allow for a variable number of station recordings as input. This is a practical necessity for scenarios in which a station drops offline or is damaged, and means the model could be used at different volcanoes, irrespective of their network configuration. Using a 14-month dataset from Nabro volcano, Eritrea, acquired following the onset of its VEI4 eruption in 2011, I will present findings on how to best encode and share signal context across stations to optimise model performance, and compare the performance of multi-station models against standard single-station model equivalents.