← Back to Gallery

Latent Representations of Seismic Waves With Self-Supervised Learning

Fully-supervised deep learning models and large, labeled earthquake datasets have revolutionized earthquake detection in the past five years, but the reasons why these models either perform well or do not generalize to other datasets remain opaque. In this work, we propose a latent, or internal vector space, representation of seismic waves learned from self-supervised training on a large corpus of unlabeled continuous seismic waveform data. We take inspiration from the field of speech recognition, where latent representations have become the basis of nearly all natural language processing, such as transcription or translation. Our approach differs from previous fully-supervised earthquake detection methods in that we encode a time window of recorded ground acceleration as a feature vector in a latent space, rather than as a single label, such as “P-wave”, “S-wave”, or “noise”. Our approach allows for a more nuanced encoding of complex waveform data, such as during early aftershock sequences. Using years of continuous waveform data from the Southern California Earthquake Data Center (SCEDC) AWS Public Dataset as an unlabeled training set, we train a transformer model to learn latent representations of continuous seismic waveforms using a contrastive loss function. We compare our learned representations to seismic phases identified by conventional earthquake pickers, human analysts, and deep learning-based earthquake detection models. As demonstrated recently in speech recognition, we suggest that latent representations can be used as a pre-trained encoder in a diverse set of seismological tasks requiring continuous seismic data such as earthquake detection, earthquake early warning, and ground-motion forecasting.

Session: Opportunities and Challenges for Machine Learning Applications in Seismology [Poster]

Type: Poster

Room: Ballroom

Date: 4/19/2023

Presentation Time: 08:00 AM (local time)

Presenting Author: Tim Clements

Student Presenter: No

Additional Authors