# A near real-time framework for monitoring very-long-period signals at volcanoes

## TL;DR

This paper introduces a near real-time system for detecting and analyzing very-long-period seismic signals at volcanoes, improving monitoring and early warning capabilities.

## Contribution

A novel automatic and near real-time detection method for VLP seismicity using polarization and spectral analysis with machine learning optimization.

## Key findings

- The algorithm accurately captures temporal evolution of VLP activity with 23% false alerts and 27% missed alerts.
- Automatic detections strongly correlate with manually derived daily VLP rates over 16 years of data.
- The method successfully integrates into volcano monitoring strategies for long-term activity tracking.

## Abstract

Real-time seismological applications are essential for monitoring active volcanoes, offering valuable tools for the early detection of volcanic unrest and eruption. Very Long Period (VLP) seismicity, commonly observed at open-vent volcanoes with mild and persistent explosive activity, is a key indicator of volcanic activity intensity as changes in the rate of occurrence and VLP event magnitude can be a signal of impending unrest. In this study, we introduce a new method for the automatic and near real-time detection and characterization of VLP seismicity. Our approach was tested on Stromboli Volcano (Italy), where VLP seismic activity has been well-documented for over two decades. The detection algorithm is based on three-component amplitude analysis, derived from waveform polarization and spectral characteristics of continuous seismic records. It extracts key parameters such as detection time, event duration, azimuth, and incidence (polarization) angles. VLP events are distinguished from other signals through a single-station statistical analysis of polarization parameters, providing a reliable near–real-time catalog of VLP detections. Optimal detection thresholds for each station were determined using a machine-learning hyperparameter optimization approach. Here, we focus on the year 2007, which was characterized by highly variable VLP activity, including a major effusive eruption at Stromboli. The algorithm’s performance was validated using an independent, manually inspected dataset from 2007, yielding a false alert rate of 23% and a missed alert rate of 27% for the best-performing station. The results show that the method accurately reproduces the temporal evolution of the different activity phases throughout the year, with clear implications for enhancing and integrating VLP detection into existing volcano monitoring strategies. We applied the method to 16 years of seismic data (2009–2024), successfully reconstructing the temporal evolution of the VLP event rate in close agreement with manual inspections. The automatic detections show a strong correlation with manually derived daily rates, demonstrating that our automatic VLP detection time series reliably captures long-term fluctuations in volcanic activity over the entire period of investigation.

The online version contains supplementary material available at 10.1038/s41598-025-25636-7.