Unsupervised clustering of ambient seismic noise in an urban environment

In this study, a semi-automated unsupervised clustering workflow is developed to classify ambient seismic noise from datasets acquired in an urban region, specifically using a pair of Tromino3G+ seismographs simultaneously recording high-gain velocity along two axes (north-south and east-west). A multi-day acquisition is used as a reference ambient seismic noise characterization of the area and to assess clusterability of the dataset. Additional 20-min datasets, acquired during and after anomalous urban activities are analyzed for a case study. The anomalous urban activity is due to street parties (large gatherings of people) during the Queen's University Homecoming celebrations in Kingston, Ontario. Unsupervised clustering methods are applied to classify ambient seismic events (coherent amplitude peaks in time-frequency space caused by uncontrolled natural and anthropogenic sources) directly from statistical properties of the event features (amplitude, frequency, bandwidth, duration, and azimuth relative to source). For urban ambient seismic noise, the signal sources may be numerous and unspecified a-priori and therefore unsupervised clustering methods are sought. Three clustering algorithms (k-Means, spectral, and agglomerative) are applied to characterize the urban region during the data collection campaigns. A probability density-based threshold modeled from the baseline dataset leads to the identification of event anomalies resulting from street party activity. The results provide insight into the utility of unsupervised clustering of ambient seismic noise for urban activity monitoring and indicate future work should include assessing the dependency of the workflow on the number of clusters and the selection of clustering methods.