Performance assessment of Deep Learning picking models at Mount Etna volcano

We benchmark several pre-trained deep learning models for automatic phase picking and discrimination of volcano-tectonic earthquakes from long-period events in the complex volcanic setting of Mount Etna, Italy. We used SeisBench, an open-source framework to evaluate PhaseNet and EQTransformer models trained on different datasets from both tectonic and volcanic environments. These configurations are integrated into an autonomous workflow for phase picking, event association, and event classification.

The tests use a dataset of seismic waveforms recorded between January 2019 and June 2020 by  INGV – Osservatorio Etneo network. Performance is assessed  throughout the workflow, using two human-compiled catalogs of volcano-tectonic earthquakes and long-period events as reference benchmarks. Event classification combines signal-to-noise analysis, network geometry, and the frequency content associated with each event.

Among the tested configurations, models trained on volcanic datasets achieved the highest accuracy in both phase picking and events association. Furthermore, the spatial and temporal distribution of classified events closely matches the patterns observed in the reference catalogs.