Exploring the Signal Content of Continuous Seismic Recordings at Station NKC (West-Bohemia/Vogtland) With a Self-Supervised Vision Transformer

Marcel van Laaten , Roman Esefelder

Continuous seismic recordings contain a wide variety of signal types (earthquakes, quarry blasts, induced events, and diverse background and anthropogenic noise) but sorting them into meaningful categories usually relies on manual inspection. Here we take an exploratory approach and ask a deliberately simple question: what kinds of signals are actually present in the continuous record of station NKC, and can we let the data organise themselves without imposing labels beforehand? Station NKC lies in the West-Bohemia/Vogtland earthquake swarm region and records a mixture of earthquakes and non-earthquake transients. We select waveform windows with an STA/LTA trigger, represent each window as a continuous-wavelet-transform spectrogram of all three components, and use a self-supervised vision transformer (DINOv3) to group the windows by similarity. Because the model learns its representation directly from the unlabelled spectrograms, the resulting structure reflects the data rather than a predefined catalogue. We then examine the groups that emerge and compare them, where possible, against earthquake catalogues and against the waveforms themselves. The learned representation separates earthquakes clearly from the background, and unsupervised clustering recovers groups that are strongly enriched in earthquakes and that concentrate in the known swarm episodes. Equally informative is what the clustering reveals about the rest of the record. A large fraction of triggered windows have no catalogue counterpart, and inspection shows that many of these are themselves earthquake-like signals. We present the signal groups identified and discuss their likely physical interpretation.