Earthquakes often occur in spatio-temporal clusters, reflecting complex physical processes within the Earth’s crust. While background seismicity results from long-term tectonic loading and typically behaves as a temporally independent Poisson process, large earthquakes can generate aftershock sequences that develop into multi-stage cascades. Seismic catalogues contain both independent (background) and dependent (clustered) events, and distinguishing between these populations is crucial for scientific interpretation and hazard forecasting. Background seismicity underpins long-term seismic hazard models, whereas clustered events provide information on stress transfer, fault interactions, and aftershock evolution. Many seismic hazard frameworks require, or benefit from, catalogues in which independent events are separated and clusters are accurately identified (Gentili et al., 2019, 2025). A major development in seismic cluster analysis in recent years has been the introduction of machine learning techniques. Among these, density-based algorithms such as DBSCAN (Ester et al., 1996) and OPTICS (Ankerst et al., 1999) have been widely used to identify arbitrarily shaped seismicity clusters based on spatial density criteria (Cesca, 2020; Piegari et al., 2022). However, classical DBSCAN does not differentiate between spatial and temporal coordinates, often requiring time to be treated independently from space, which limits its suitability for aftershock identification, where simultaneous space–time clustering is essential (Nicolis et al., 2024). An improvement to this limitation is ST-DBSCAN (Birant and Kut, 2007), which uses separate spatial and temporal search radii, allowing more flexible treatment of space–time relationships. In this preliminary study, we compare two clustering approaches applied to the New Zealand earthquake catalogue: a window-based method, commonly used in operational seismology, and ST-DBSCAN. We selected clusters containing more than 100 events and focused our analysis on pairs of clusters identified by both methods whose centroids lie within 10 km of each other. Fifteen clusters were found to overlap, providing a consistent basis for comparison. Cluster quality metrics indicate broad agreement between the two approaches, with an Adjusted Rand Index of approximately 0.79. Preliminary results show that ST-DBSCAN tends to resolve fine-scale structure more effectively, whereas the window-based method forms tighter and more distinct large-scale groupings. To assess correspondence with an independent reference, we carried out an initial analysis of the 2010–2013 Canterbury–Christchurch sequence. Using the event list from Herman et al. (2014), which includes 150 earthquakes of magnitude Mw ≥ 3.5, we compared sequence membership and structural consistency.
Spatiotemporal Clustering of the New Zealand Catalog: A Comparative Study of ST-DBSCAN and Window-Based Methods – Preliminary results
S. Gentili;L. Caravella;
2026-01-01
Abstract
Earthquakes often occur in spatio-temporal clusters, reflecting complex physical processes within the Earth’s crust. While background seismicity results from long-term tectonic loading and typically behaves as a temporally independent Poisson process, large earthquakes can generate aftershock sequences that develop into multi-stage cascades. Seismic catalogues contain both independent (background) and dependent (clustered) events, and distinguishing between these populations is crucial for scientific interpretation and hazard forecasting. Background seismicity underpins long-term seismic hazard models, whereas clustered events provide information on stress transfer, fault interactions, and aftershock evolution. Many seismic hazard frameworks require, or benefit from, catalogues in which independent events are separated and clusters are accurately identified (Gentili et al., 2019, 2025). A major development in seismic cluster analysis in recent years has been the introduction of machine learning techniques. Among these, density-based algorithms such as DBSCAN (Ester et al., 1996) and OPTICS (Ankerst et al., 1999) have been widely used to identify arbitrarily shaped seismicity clusters based on spatial density criteria (Cesca, 2020; Piegari et al., 2022). However, classical DBSCAN does not differentiate between spatial and temporal coordinates, often requiring time to be treated independently from space, which limits its suitability for aftershock identification, where simultaneous space–time clustering is essential (Nicolis et al., 2024). An improvement to this limitation is ST-DBSCAN (Birant and Kut, 2007), which uses separate spatial and temporal search radii, allowing more flexible treatment of space–time relationships. In this preliminary study, we compare two clustering approaches applied to the New Zealand earthquake catalogue: a window-based method, commonly used in operational seismology, and ST-DBSCAN. We selected clusters containing more than 100 events and focused our analysis on pairs of clusters identified by both methods whose centroids lie within 10 km of each other. Fifteen clusters were found to overlap, providing a consistent basis for comparison. Cluster quality metrics indicate broad agreement between the two approaches, with an Adjusted Rand Index of approximately 0.79. Preliminary results show that ST-DBSCAN tends to resolve fine-scale structure more effectively, whereas the window-based method forms tighter and more distinct large-scale groupings. To assess correspondence with an independent reference, we carried out an initial analysis of the 2010–2013 Canterbury–Christchurch sequence. Using the event list from Herman et al. (2014), which includes 150 earthquakes of magnitude Mw ≥ 3.5, we compared sequence membership and structural consistency.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
