The Petrinja earthquake series located and visualised using machine learning

Earthquake detection and phase picking are crucial steps in the analysis of earthquakes. With the increasing number of seismic instruments available, large amounts of seismic data are generated, requiring the use of automatic algorithms to process earthquake series and to include events that would not be discovered with manual approaches.

The Petrinja earthquake series started with local magnitude M_L5.0 earthquake on December 28, 2020, followed by M_L6.4 earthquake one day later. In the two years of this earthquake series, human analysts picked a total of 16,000 earthquakes smaller than M2.0, 1528 with magnitudes M2.0-2.9, 156 with magnitudes M3.0-3.9, 17 with magnitudes M4.0-4.9, 2 with magnitudes M5.0-5.9 and one earthquake with magnitude greater than M6.0. While the seismic network at the onset of this sequence counted only a few instruments in the epicentral area, the rapid aftershock deployment of 5 stations in the near vicinity of the fault zone, and the further gradual yet still remarkable growth of the seismic network to more than 50 instruments, produced an extraordinary amount of data, which are perfectly suited for employing machine learning (ML) methods for seismic phase picking and earthquake detection.

In this study we present application of various ML methods to the Petrinja earthquake series. We also compare how the results change when we train a model using a subset of data from this earthquake series. Our results show that these machine learning methods are promising approaches for accurately detecting and picking phases in such earthquake series, and also delineate tectonic features responsible for generating them.