A Robust Framework for Clustering Variable-Length Seismic Events: A Cryogenic Case Study

Long-term seismic monitoring provides a unique insight into glacier and ice-shelf dynamics. However, the extraction of meaningful cryoseismic information from continuous multi-year records remains challenging. Icequake events frequently show unclear or overlapping signals due to harsh environmental conditions and persistent background activity. While the utilisation of variable event lengths can be instrumental in the avoidance of merging multiple events into a single window, most unsupervised learning methods require fixed input durations. This emphasises the necessity for a novel unsupervised clustering approach that can handle time-variant events of varying lengths while robustly detecting outliers. The method should be physics-based to accommodate the limited prior knowledge of icequake characteristics. It should also operate directly on large event catalogues, without reliance on handcrafted features.

To address this issue, the incrementally buffered dynamic time warping clustering is introduced. This is a new approach to clustering dynamic time warping (DTW) distances of events and it incorporates the requirements stated above. The method starts with an initial k-medoids clustering on a pairwise DTW distance matrix of a subset of events, thereby generating initial k clusters. The subsequent addition of new samples is based on a statistically robust distance threshold from the distribution in within-cluster distances of the initial step. Each event is compared only to existing medoids, assigned to the nearest cluster if the DTW distance falls below the threshold, or temporarily placed in a buffer when classified as an outlier. The promotion of buffered events to new clusters is only permitted when the criteria for similarity and minimum sample count are met, thus preventing the formation of spurious clusters from isolated noise events. Lastly, a final global reassignment step is performed. This step involves the recomputation of all event-to-medoid distances. The purpose of this is to stabilise cluster boundaries and refine the catalogue. The combination of these steps results in a scalable and transparent algorithm that is well-suited to the analysis of extensive environmental time-series data.

The present study applies this framework to a time span of several years of vertical-component seismic data from the 16-sensor Watzmann array at Neumayer Station III, Antarctica. Preliminary results indicate the presence of numerous persistent families of icequakes. These are analysed with regard to their correlation with environmental conditions, including tidal modulation and wind.