Variational autoencoders in earthquake detection and processing for submarine distributed optical fibre sensing

The technology of Distributed Optical Fibre Sensing (DOFS) has attracted a lot of attention for monitoring and studying geophysical processes, including earthquakes. This is partly because using DOFS earthquakes can be studied in greater detail, as Earth ground motions are captured on a synchronised network of hundreds of thousands of sensing locations along the fibre optic cable. Using DOFS, seismologists and other experts can monitor earthquakes with a sensing range spanning hundreds of kilometres and over harsh environments, e.g., across oceans.

These experimental studies are typically carried out for long time-periods, and therefore, are known to generate large amounts of data. At the same time, seismic signals captured by DOFS are typically of lower signal-to-noise ratio in comparison to the conventional seismic sensors. Hence, devising effective signal processing and machine learning technologies that can be used to process DOFS signals has become critical for fibre optic seismology studies. Of particular importance are regions whereby naturally occurring earthquakes may be rare, but there is nevertheless the need to deploy capable earthquake detection and classification systems using DOFS.

In this study, we demonstrate the concept of detecting earthquakes, using hundreds of spatiotemporal images, as obtained using a commercial DOFS system. Approximately two months of data were collected using the 55 km subsea fibre optic cable, which is located off the coast of Muroto in Japan. Each of the 9800 locations along the fibre optic cable was sampled at a rate of 500 Hz. Although earthquake events have been historically shown to occur frequently in this region, we demonstrate a monitoring capability that does not require a priori “labels” in constructing an earthquake detection model. The abundance of non-earthquake data generating sources in the region, e.g., fishing vessels, marine life, and ambient noise, allowed us to compute representations that are specific to those sources. Consequently, the amount of deviation from these representations can be calculated on new spatiotemporal images and be used to flag potential seismic activity. A Variational Autoencoder was used to obtain the representations of non-earthquake sources. Our earthquake detector correctly identifies more than 90 % of available earthquakes in our dataset using metrics such as the reconstruction error.