Insights into a Deep Learning Workflow for Automatic Detection of Seismic Signals at Mount Etna Volcano

Abstract We present a workflow for the automatic detection of seismic events in volcanic areas, leveraging deep learning (DL) tools for the picking and association of seismic phases. Specifically, we employ PhaseNet and EQTransformer via the SeisBench open-source platform to perform automated picking of P- and S-wave arrivals from continuous seismic waveforms. To evaluate performance, we systematically test several pretrained models available within SeisBench using seismic data from Mount Etna (Italy), encompassing both volcano-tectonic (VT) and long-period (LP) events recorded from January 2019 to June 2020. Each model configuration is quantitatively assessed through statistical analyses, comparing results against existing seismic catalogs to identify the most effective approach for the Etna region. Our findings show that models trained on volcano-specific datasets significantly outperform those based on tectonic-only data. To further classify VT and LP signals, we apply a frequency-based algorithm using a modified frequency index, which effectively distinguishes VT, LP, low spectrum, mixed, and unclassified type events based on their spectral characteristics. This classification demonstrates strong agreement with the reviewed catalogs. Overall, the proposed workflow performs well in the complex setting of Mount Etna, highlighting the potential of pretrained DL models for semi-automated volcanic monitoring. At the same time, our results underscore the need for site-specific model validation and expert oversight to ensure reliable outcomes.