Abstract

In volcano monitoring, effective recognition of seismic events is essential for understanding volcanic activity and raising timely warning alerts. Traditional approaches rely on manual analysis, which is inherently subjective and resource-intensive, as continuous monitoring necessitates the uninterrupted presence of analysts. Furthermore, current automatic approaches typically address either detection or classification in isolation and predominantly rely on single-station data, limiting their robustness, adaptability, and real operational implementation. We introduce an end-to-end framework that uses a transformation of 1-D multistation waveforms into compact 2-D image patches. This representation allows deep semantic-segmentation models to generate per-pixel softmax activation maps that simultaneously detect and label events within a fixed time window, while also supporting a sliding-window setup for continuous data streams. We evaluated five architectures (UNet, UNet++, DeepLabV3 +, SwinUNet, and PhaseNet) using 24 493 labeled windows from four Chilean volcanoes spanning five event classes, as well as a 10-h continuous trace. All 2-D models demonstrated strong noise robustness and adaptability to new datasets. UNet achieved the best performance, with a mean $F1$ -score of 0.91 and IoU of 0.88 on labeled windows, and a mean $F1$ -score of 0.68 when applied to the continuous trace. Depending on the architecture and refresh rate, the system can process a 10-h trace in as little as 4 s and up to 5 min on an NVIDIA RTX 3060 GPU. We believe the proposed approach provides a practical and effective solution for continuous, real-time volcanic surveillance, as it minimizes preprocessing, leverages spatial correlations across stations, and successfully adapts to unseen volcanoes.