RECOVAR: An unsupervised deep learning approach to seismic event detection by training on continuous waveform data

State-of-the-art machine learning models for seismic event detection, such as EQTransformer and PhaseNet, use supervised learning, which requires labeled event catalogs and curated waveforms. This dependence creates two fundamental limitations: the cost of preparing  high-quality datasets, and an annotation bias which limits the model training to event types well represented in existing catalogs. Unsupervised deep learning has the potential to overcome these limitations, but despite its prevalence in other domains, this approach remains rather under-explored for the problem of seismic event detection.

We present RECOVAR, an unsupervised deep learning method that trains directly on continuous waveform data, requiring no labeling or catalog preparation. The architecture consists of an ensemble of convolutional autoencoders, each trained independently. Detection exploits how these latent representations differ for signal versus noise: coherent seismic arrivals produce convergent representations with high cross-covariance, while stochastic noise produces uncorrelated representations. 

Since continuous recordings are dominated by noise, a naive approach to training on continuous waveforms ends up creating a model that focuses excessively on representing noise, and results in quite good but suboptimal event detection. We introduce a dynamic training pipeline that preferentially resamples low-scoring segments using the model's own cross-covariance scores, which results in strong detection performance.

RECOVAR achieves event detection ROC AUC scores of 0.97-0.99 on the STEAD and INSTANCE benchmarks, comparable to PhaseNet and EQTransformer. We demonstrate a regional application to the 2019 Istanbul Silivri earthquake sequence, training directly on continuous waveforms without any catalog preparation. We show the utility of RECOVAR as a post processing tool that filters picks by supervised methods, retaining 99% of true picks by PhaseNet while filtering half of the false positives, and with less conservative settings, removing 83% of false positives while retaining 84% of true detections.

RECOVAR provides an unsupervised deep learning alternative for seismic detection. Training directly on continuous data without labels avoids the annotation bias that is inherent to supervised methods, which potentially opens the door to detecting rare event types absent from established catalogs. As demonstrated by its post-filter performance, RECOVAR also integrates naturally within existing detection pipelines.