The expanding seismic networks and the growing amount of seismic data underline the demand for automated workflows that enable fast and precise earthquake detection and localization. Recently, machine learning (ML)-based picking algorithms have demonstrated exceptional accuracy and efficiency as well as strong generalization capabilities. Consequently, these ML-based pickers can substantially enhance the localization accuracy of traditional sequential earthquake location methods.
In the present work, we build a distributed workflow to perform the automatic detection of earthquakes in volcanic areas, on multi-GPU nodes, based on Deep-Learning (DL) models for the phase picking and phase association of seismic waves. In particular, our geographical area of interest is the volcanic area of Mount Etna (Italy). The dataset includes seismic waveforms of about 38 seismic stations and seismic catalogs associated mainly to Volcano-Tectonic (VT) earthquakes and Long-Period (LP) events.
The first part of this master thesis deals with the standardization and selection of the seismic waveforms recorded between January 2019 and June 2020, accompanied by a VT earthquakes catalog and associated picks for benchmarking.
A VT earthquake detection workflow is then developed using the prepared Etna seismic waveform dataset as input. The workflow has been optimized for the Leonardo system by using its highly parallelized and GPU-accelerated architectures.
Initially, after reading the input data in a distributed way, our workflow calls the chosen DL models for the picking of P- and S-waves from continuous seismic traces, which represent the main ingredient for earthquake detections. We use two deep learning models available in the literature (PhaseNet and EQTransformer) and utilize an open-source platform that offers a variety of pretrained models on different seismic datasets. Several pretrained model configurations are tested with the Etna dataset, resulting in different detection performances. A series of statistical comparisons are performed to determine the most effective model capable of detecting the P and S picks associated with the VT events in the original catalog.
Further a phase associator linking the P and S picks to individual earthquakes in accordance with the seismic velocity model of the area is employed. Due to the complexity of the volcanic environment, where VT and LP events occur alongside other signals, special attention is given to this task.