Exploring the performance of phase association algorithms

Seismic phase association plays an important role in earthquake detection and location workflows as it links together seismic phases detected on different seismometers into individual earthquakes. Together with improved phase picking algorithms, a phase association algorithm can generate large earthquake phase data sets and earthquake catalogs when applied to dense permanent or temporary seismic networks. Recently, many efforts have been made on improving seismic phase association performance, such as developing machine learning approaches that are trained on millions of synthetic sequences of P and S arrival times, to generate more precise and complete catalogs including more small earthquakes.

As part of project MILESTONE, which aims at the automatic creation of large microseismicity catalogs in subduction settings, the present study evaluates the performance of the deep-learning based phase association algorithm PhaseLink (Ross et al. 2019) by comparison with a traditional grid-based method and a small handpicked benchmark dataset.

We used seismic data from the IPOC (Integrated Plate boundary Observatory Chile) permanent deployment of broadband stations in Northern Chile, dedicated to the study of earthquakes and deformation at the continental margin of Chile.

For an initial calibration, we manually picked P and S phases of raw waveforms on 15 stations on two randomly chosen days. All events that were visually recognizable were picked and located, which led to a dataset of 251 events comprising 1823 P and 1468 S picks, spanning a depth range from the surface down to 240 km. We use this handpicked dataset as ‘ground truth’, and evaluate the performance of PhaseLink and the grid-based method coupled with a STA/LTA trigger against this benchmark, considering both the numbers of (correctly/falsely) associated events and the number of constituent picks per event.

In a second experiment, we compare PhaseLink and conventional phase associator using a much larger set of STA/LTA alerts from the same region, but without the additional ground truth.

The presented research represents first steps towards an integrated automated workflow for detecting, picking, associating and locating microseismicity in subduction zone settings.