Volcano-seismic signals can help for volcanic hazard estimation and eruption forecasting. However, the underlying mechanism for their low frequency components is still a matter of debate. Here, we show signatures of dynamic strain records from Distributed Acoustic Sensing in the low frequencies of volcanic signals at Vulcano Island, Italy. Signs of unrest have been observed since September 2021, with CO2 degassing and occurrence of long period and very long period events. We interrogated a fiber-optic telecommunication cable on-shore and off-shore linking Vulcano Island to Sicily. We explore various approaches to automatically detect seismo-volcanic events both adapting conventional algorithms and using machine learning techniques. During one month of acquisition, we found 1488 events with a great variety of waveforms composed of two main frequency bands (from 0.1 to 0.2 Hz and from 3 to 5 Hz) with various relative amplitudes. On the basis of spectral signature and family classification, we propose a model in which gas accumulates in the hydrothermal system and is released through a series of resonating fractures until the surface. Our findings demonstrate that fiber optic telecom cables in association with cutting-edge machine learning algorithms contribute to a better understanding and monitoring of volcanic hydrothermal systems.

<p>Since September 2021, signs of unrest at Vulcano Island have been noticed after four years of quiescence, along with CO<sub>2</sub> degassing and the occurrence of long-period and very long-period events. With the intention of improving the monitoring activities, a submarine fiber optic telecommunications cable linking Vulcano Island to Sicily was interrogated from 15 January to 14 February 2022. Of particular interest has been the recording of 1488 events with wide range of waveforms made up of two main frequency bands (from 3 to 5 Hz and from 0.1 to 0.2 Hz).</p> <p>With the aim of the automatic detection of seismic-volcanic events, different approaches were explored, particularly investigating whether the application of machine learning could provide the same performance as conventional techniques. Unlike many traditional algorithms, deep learning manages to guarantee a generalized approach by automatically and hierarchically extracting the relevant features from the raw data. Due to their spatio-temporal density, the data acquired by the DAS can be assimilated to a sequence of images; this property has been exploited by re-designing deep learning techniques for image processing, specifically employing Convolutional Neural Networks.</p> <p>The results demonstrate that deep learning not only achives good performance but that it even outperforms classical algorithms. Despite providing a generalized approach, Convolutional Neural Networks have been shown to be more effective than traditional tecniques in expoiting the high spatial and temporal sampling of the acquired data. </p>