Earthquakes, tremors, and other types of seismic events are the expression of deformations that occur within the Earth's crust in seismogenic zones (subducting zones, continental faulting systems, volcanoes). Detecting these events is a crucial stage to seismic hazard assessment. While many detection algorithms exist nowadays, a large amount of seismicity may still be out of reach. Traditional detection algorithms most often look for specific characteristics in the continuous seismograms to trigger detections. The use of arrays of seismic stations significantly improves the detection accuracy; however, several seismogenic regions remain poorly instrumented, preventing from having complete earthquake catalogs.
We revisit the way to detect several types of seismic activity in a data-driven fashion. We first introduce a model-free detector based on array seismic data Shannon entropy. We show how this data representation can reveal several kinds of seismicity within tectonic and volcanic contexts. We then show how deep learning can be used to recognize patterns in the continuous seismic wavefield recorded by a single station, with examples on precursory repeaters, non-volcanic tremors, and background seismic noise properties.