Creating an Inventory of Seismic Signals at Vulcano Island, Italy, using Unsupervised Learning Techniques

Abstract

The island of Vulcano gives its name to the so-called “Vulcanian eruptions”, an eruptive style with strong explosive characteristics and observed there for the first time. The last eruptive activity occurred between 1888 and 1890. Starting from mid-September 2021, an unrest, marked by relevant variations in geochemical and geophysical parameters, affected the island. Here, we analyze the seismic signals recorded from the onset of the unrest until December 2022. An increasing number of Very Long Period events was detected from September 2021 onwards, enhancing concerns linked to other measured anomalies, such as increasing CO2 emissions and fumarole temperatures. Numerous types of signals were generally recorded on the island, partly caused by various man-made sources, such as the close-by passage of ships, dropping anchors, etc. The large variety of the seismic signals made standard amplitude-based monitoring techniques, such as RSAM, questionable. We therefore focused on creating an inventory of the recorded signals exploiting unsupervised machine learning techniques, namely Self-Organizing Maps and Cluster Analysis. We were able to identify various classes of seismic events related to volcanic dynamics and to distinguish exogenous signals, such as anthropic noise. This allowed us to visualize the development of signal characteristics efficiently. This classification can help build an effective alert tool to automatically identify different types of seismic signals, useful for surveillance purposes. Furthermore, it is a preparative step for other studies, such as event location and source process modeling.