5887 20241203110535.0 10.1093/gji/ggac469 DOI SCART-2022-0113 Yates, Alexander Assessing Similarity in Continuous Seismic Cross-Correlation Functions Using Hierarchical Clustering: Application to Ruapehu and Piton de La Fournaise Volcanoes 2023 Passive seismic interferometry has become a popular technique towards monitoring. The method depends on the relative stability of background seismic sources in order to make repeatable measurements of subsurface properties. Such stability is typically assessed by examining the similarity of cross-correlation functions through time. Thus, techniques that can better assess the temporal similarity of cross-correlation functions may aid in discriminating between real subsurface processes and artificial changes related variable seismic sources. In this study, we apply agglomerative hierarchical clustering to cross-correlation functions computed using seismic networks at two volcanoes. This allows us to form groups of data that share similar characteristics and also, unlike common similarity measures, does not require a defined reference period. At Piton de la Fournaise (La Réunion island), we resolve distinct clusters that relate both to changes in the seismic source (volcanic tremor onset) and changes in the medium following volcanic eruptions. At Mt Ruapehu (New Zealand), we observe a consistency to cross-correlation functions computed in the frequency band of volcanic tremor, suggesting tremor could be useful as a repeatable seismic source. Our results demonstrate the potential of hierarchical clustering as a similarity measure for cross-correlation functions, suggesting it could be a useful step towards recognizing structure in seismic interferometry data sets. This can benefit both decisions in processing and interpretations of observed subsurface changes. NO seismology volcanology volcano-seismology ambient noise machine learning classification clustering eruption Caudron, Corentin Lesage, Philippe Mordret, Aurélien Lecocq, Thomas Soubestre, Jean Geophysical Journal International 233 2023 472–489 thomas.lecocq@observatoire.be https://doi.org/10.1093/gji/ggac469 1475444 http://publi2-as.oma.be/record/5887/files/ggac469fig3.png Cluster output using 0.1–1.0 Hz cross-correlation functions (CCFs) for station-pair UV05-UV12 at Piton de la Fournaise (blue line in Fig. 1a). (a) Normalized spectral width measurement. Lower values indicate a more coherent seismic wavefield dominated by fewer seismic sources. Dashed white lines show frequency range of CCFs. Triangles above (red) indicate eruption start times (b) Correlogram showing amplitudes of CCFs at different lag times (red = positive, blue = negative). Dashed black lines show part of CCFs used in clustering. (c) Location of clusters in time, colour-coded according to dendrogram output in (e). (d) Apparent velocity changes, colour-coded according to correlation coefficient (CC) computed between 10-d stacks and the reference stack period (shaded grey). Vertical red bars denote timing of eruptive activity. (e) Dendrogram, with clusters defined at a distance threshold of 0.2 (dashed-red line). (f) Mean value of spectral width as a function of frequency within time period of each cluster. (g) Normalized frequency spectrum of CCFs within each cluster, computed from averaged CCF. Legend shows Scale Factor (SF) to convert from non-normalized spectrum. (h) Average of CCFs within each cluster at positive lag times. (i) Mean coherence, averaged over CCF frequency range, between different clusters at positive lag times, computed using moving window of 10-s length with 0.5 s step (95 per cent overlap). 17315 http://publi2-as.oma.be/record/5887/files/ggac469fig3.gif?subformat=icon icon Cluster output using 0.1–1.0 Hz cross-correlation functions (CCFs) for station-pair UV05-UV12 at Piton de la Fournaise (blue line in Fig. 1a). (a) Normalized spectral width measurement. Lower values indicate a more coherent seismic wavefield dominated by fewer seismic sources. Dashed white lines show frequency range of CCFs. Triangles above (red) indicate eruption start times (b) Correlogram showing amplitudes of CCFs at different lag times (red = positive, blue = negative). Dashed black lines show part of CCFs used in clustering. (c) Location of clusters in time, colour-coded according to dendrogram output in (e). (d) Apparent velocity changes, colour-coded according to correlation coefficient (CC) computed between 10-d stacks and the reference stack period (shaded grey). Vertical red bars denote timing of eruptive activity. (e) Dendrogram, with clusters defined at a distance threshold of 0.2 (dashed-red line). (f) Mean value of spectral width as a function of frequency within time period of each cluster. (g) Normalized frequency spectrum of CCFs within each cluster, computed from averaged CCF. Legend shows Scale Factor (SF) to convert from non-normalized spectrum. (h) Average of CCFs within each cluster at positive lag times. (i) Mean coherence, averaged over CCF frequency range, between different clusters at positive lag times, computed using moving window of 10-s length with 0.5 s step (95 per cent overlap). 17774 http://publi2-as.oma.be/record/5887/files/ggac469fig3.jpg?subformat=icon-180 icon-180 Cluster output using 0.1–1.0 Hz cross-correlation functions (CCFs) for station-pair UV05-UV12 at Piton de la Fournaise (blue line in Fig. 1a). (a) Normalized spectral width measurement. Lower values indicate a more coherent seismic wavefield dominated by fewer seismic sources. Dashed white lines show frequency range of CCFs. Triangles above (red) indicate eruption start times (b) Correlogram showing amplitudes of CCFs at different lag times (red = positive, blue = negative). Dashed black lines show part of CCFs used in clustering. (c) Location of clusters in time, colour-coded according to dendrogram output in (e). (d) Apparent velocity changes, colour-coded according to correlation coefficient (CC) computed between 10-d stacks and the reference stack period (shaded grey). Vertical red bars denote timing of eruptive activity. (e) Dendrogram, with clusters defined at a distance threshold of 0.2 (dashed-red line). (f) Mean value of spectral width as a function of frequency within time period of each cluster. (g) Normalized frequency spectrum of CCFs within each cluster, computed from averaged CCF. Legend shows Scale Factor (SF) to convert from non-normalized spectrum. (h) Average of CCFs within each cluster at positive lag times. (i) Mean coherence, averaged over CCF frequency range, between different clusters at positive lag times, computed using moving window of 10-s length with 0.5 s step (95 per cent overlap). published in REFERD