Exploratory data analysis for continuous seismograms

René Steinmann

Seismic data contain a wealth of crucial information about active geological structures such as faults or volcanoes. The growing amount of seismic data collected nowadays cannot scale with manual investigation, suggesting automatic algorithms for scanning continuous data streams. We develop a strategy based on artificial intelligence to scan continuous seismic data and infer patterns automatically. The strategy involves three major steps: first, a scattering network retrieves the scattering coefficients of the continuous seismogram, a well-suited data representation for pattern recognition tasks. Second, principal or independent component analysis are applied to the scattering coefficients and obtain a low-dimensional and meaningful feature space. At last, we perform hierarchical agglomerative clustering in the feature space, revealing the hierarchical structure of the seismic signal classes present in the data. Both the features and the output of the clustering are then used to explore the content of the seismic data. We show applications of this strategy to seismic data recorded in the vicinity of faults, volcanoes and cities, revealing patterns and signals unique to each environment.

René Steinmann is currently a PhD student at the Université Grenoble Alpes, working at the intersection between seismology and artificial intelligence.