A Framework for Real-Time Volcano-Seismic Event Recognition Based on Multi-Station Seismograms and Semantic Segmentation Models

TL;DR

This work tackles the challenge of real-time volcano-seismic event recognition by leveraging semantic segmentation on multi-station seismograms. It introduces Patch Stacking to transform concurrent 1-D traces from multiple stations into 2-D images, enabling the use of established 2-D segmentation models to detect and classify five event types with minimal preprocessing. Across four Chilean volcanoes, UNet achieves the best window-level performance with a mean F1 of about $0.91$ and IoU of about $0.88$, while maintaining robustness to noise and unseen data; the framework also demonstrates real-time applicability through a sliding-window pipeline and continuous-data evaluation. Overall, this approach integrates multi-station information, reduces manual preprocessing, and provides a scalable, real-time tool for volcano monitoring that can support seismologists in decision-making, though it should complement rather than replace expert analysis.

Abstract

In volcano monitoring, effective recognition of seismic events is essential for understanding volcanic activity and raising timely warning alerts. Traditional methods rely on manual analysis, which can be subjective and labor-intensive. Furthermore, current automatic approaches often tackle detection and classification separately, mostly rely on single station information and generally require tailored preprocessing and representations to perform predictions. These limitations often hinder their application to real-time monitoring and utilization across different volcano conditions. This study introduces a novel approach that utilizes Semantic Segmentation models to automate seismic event recognition by applying a straight forward transformation of multi-channel 1D signals into 2D representations, enabling their use as images. Our framework employs a data-driven, end-to-end design that integrates multi-station seismic data with minimal preprocessing, performing both detection and classification simultaneously for five seismic event classes. We evaluated four state-of-the-art segmentation models (UNet, UNet++, DeepLabV3+ and SwinUNet) on approximately 25.000 seismic events recorded at four different Chilean volcanoes: Nevados del Chillán Volcanic Complex, Laguna del Maule, Villarrica and Puyehue-Cordón Caulle. Among these models, the UNet architecture was identified as the most effective model, achieving mean F1 and Intersection over Union (IoU) scores of up to 0.91 and 0.88, respectively, and demonstrating superior noise robustness and model flexibility to unseen volcano datasets.

Figures (7)

• Figure 1: Geographical distribution of seismic stations used in this study. Left column: detailed station networks around each volcano, ordered from north to south (LDM, NChVC, VCA, and CAU). Right column: overview map of Chile showing the relative positions of the four volcanoes. Red triangles mark volcano locations, and blue circles indicate seismic stations.
• Figure 2: Real‑time event‑recognition pipeline for a continuous, multi‑station stream. In the figure $S=8$, $W=8192$ and $N=256$ represent the parameters used for most of the experiments in this work. (a) The incoming trace is sliced into overlapping windows of $W$ samples and $S$ channels. (b) Each window is reshaped into a square grayscale image by Patch Stacking. A 2‑D segmentation model (c) assigns per‑pixel class probabilities to each stacked window (d). (e) Activation Unstacking and station summation produces window-level activations of shape $C\times W$ ($C=6$ classes). (f) Activations are accumulated in a time‑aligned buffer. (g) The buffer is binarized to yield a mutually exclusive, time‑based segmentation. (h) Detected events are written to a dataframe with assigned class and start/end times. (i) A post‑processor merges adjacent segments of the same class separated by less than a user-defined threshold (in our study, 2.5 seconds; e.g. TR 1 and TR 2) and filters out events with unrealistic durations (e.g TR 3 is too short). The final dataframe records Unix timestamps (or sample indices) for the onset and end of each detected event and the predictions can be visualized over the original data (j). For clarity, start and end times are shown in seconds.
• Figure 3: Comparison of seismic signals at varying signal-to-noise ratios (SNR) for a single channel of the same seismic window. The top panel shows the clean signal (containing two events) as reference, while subsequent panels illustrate the signal at decreasing SNR levels (10 dB, 5 dB, 0 dB, -5 dB, -10 dB). Y-axis values are omitted for clarity.
• Figure 4: Class-specific $\mathrm{RSAM}_{\text{net}}$ values in the NChVC training dataset, with the dashed line indicating the smallest first-quartile value across the five classes.
• Figure 5: Classification and detection performance through F1-score and IoU metric of the five models across SNR values ranging from 10 dB to -10 dB.