New approach targets noisy, continuous volcanic data
A new research study has demonstrated a way to extract more information from a single seismic station operating in the Campi Flegrei volcanic region near Naples, Italy. Volcanic areas generate complex ground vibrations, but separating meaningful signals from background noise and huge volumes of continuous recordings remains a major challenge for monitoring systems.
The researchers focused on building an automated method to detect and classify seismic behaviour without relying solely on pre-labelled event catalogues.
How the method works
The study combined several signal-processing and pattern-recognition tools. A Self-Organizing Map (SOM), an unsupervised learning technique, was used to group similar seismic patterns into clusters.
To represent the seismic waveforms, the analysis drew on Linear Predictive Coding (LPC), a technique that captures how a signal evolves over time. It also used STA/LTA ratios, a standard measure in seismology that compares short-term energy to longer-term background levels to help identify departures from normal noise conditions.
In addition, the researchers applied Multiscale Entropy (MSE) to measure the complexity of the signals across different time scales, aiming to distinguish structured tremor from random fluctuations.
Uncatalogued events and tremor-related anomalies
Using this combined workflow on single-station data, the study reports that the SOM-based clustering could identify seismic events that were not included in existing catalogues. It also detected anomalies linked to fumarolic tremor, a type of continuous vibration associated with gas and steam release from vents.
The clustering results were not treated as isolated outputs. The study examined how the distribution of clusters changed over time, using the variation as an indicator of evolving seismic behaviour in the area.
Links observed with CO2 emissions and rainfall
The researchers also analysed how changes in clustering corresponded with other observations. They report temporal relationships between cluster variation, measured carbon dioxide (CO2) emissions, and rainfall.
These patterns point to a possible role of environmental conditions in modulating tremor-related activity, based on the timing relationships observed in the datasets used for the study.
Early-2025 test suggests real-time potential
To evaluate whether the method could be used operationally, the trained SOM model was applied to an independent dataset from early 2025. In this test, the study reports that the approach detected intensification of tremor.
The analysis also indicated a capability to anticipate a major local earthquake recorded as Md 4.4, based on the behaviour identified ahead of the event in the processed signals.
The study concludes that entropy-based unsupervised learning, applied to continuous single-station data, can support rapid seismic characterisation and continuous monitoring in active volcanic settings such as Campi Flegrei.