Robust clustering of variable-duration seismic waveforms using an incremental dynamic time warping framework

Antonia Kiel , Vera Schlindwein, Conny Hammer

Different geophysical processes generate characteristic seismic waveforms. Identifying similar waveforms therefore helps to distinguish between different source mechanisms and to monitor seismic activity in settings ranging from volcanoes and glaciers to urban environments. In many applications, however, the seismic event types are unknown and no labelled training data is available. Then, unsupervised clustering is necessary. Most existing approaches require all input vectors to have the same length, which in seismology translates to a fixed event window. However, seismic signals can differ considerably in duration: This may require events to be cropped or padded to a common length, which may lead to a loss of important waveform characteristics. Furthermore, the approaches commonly rely on extensive feature engineering or machine-learned representations, which often require hyperparameter tuning and can reduce the physical interpretability of the resulting clusters. In addition, many clustering algorithms require the number of clusters to be specified a priori, limiting their flexibility when analysing previously unknown datasets. This study introduces Incrementally Buffered Dynamic Time Warping Clustering, a new unsupervised clustering algorithm that operates directly on seismic time-series data using a small number of physically interpretable parameters. The method is based on Dynamic Time Warping and is particularly suited to noisy and continuously active seismic environments, where overlapping signals and imperfect event catalogues are common. The buffered and incremental structure of the algorithm explicitly identifies outliers and allows new clusters to form during incremental updates, enabling continuous monitoring without reclustering the entire dataset. The algorithm is evaluated on a labelled volcano-seismic dataset, where it achieves a clustering accuracy of 69.39 per cent, comparable to previously published supervised approaches. A second application to Antarctic icequake data demonstrates its ability to identify meaningful waveform groups in a setting with limited prior knowledge of event types. These results show that the newly introduced clustering algorithm provides an interpretable and scalable framework for clustering large seismic datasets while retaining direct access to waveform similarity in the time domain.