Supervised microseismic event detection using Siamese Networks for labelling of noisy recordings

We propose a supervised classification method for data labelling that can reduce the time and expertise needed to create fully annotated large seismic datasets via the use of machine learning based Siamese networks and propose a fully data-driven methodology for selecting best class-exemplar anchors in order to better detect and classify different types of seismic events. We demonstrate the proposed network on detecting and characterising earthquakes, micro-quakes, rockfall and anthropogenic noise events. Major challenges in the analysis of seismic recordings lies in the manual labelling of collected data, which requires a great deal of time and effort from seismologists. The proposed solution shows good agreement with manually detected events while requiring little training data to be effective, hence reducing the time needed for labelling and updating models.