Unsupervised Machine Learning for Analyzing Continuous Seismic Recordings: Insights from Piton de la Fournaise Volcano

Seismograms recorded near active volcanoes contain numerous volcanic earthquakes and tremors that capture signatures of diverse volcanic processes and offer insights into the state of the volcano’s plumbing system and its underlying physical mechanisms. However, the strong variability of the seismo-volcanic signals makes their interpretation in terms of associated physical processes difficult. To address this, we employ unsupervised machine learning techniques, specifically, the scattering transform and Uniform Manifold Approximation and Projection (UMAP), to extract statistically significant features from continuous seismograms and to identify meaningful patterns related to volcanic activity. This approach eliminates the need for discrete event catalogs, enabling a comprehensive analysis of seismic manifestations of the volcanic activity.

Our study focuses on Piton de la Fournaise (PdF), a highly active basaltic volcano on La Réunion island, France, which erupted 25 times between 2014 and 2024. As one of the world’s best-monitored volcanoes, PdF represents an ideal natural laboratory for testing and refining our methodology. We analyze three-component seismograms from multiple stations and validate our findings using complementary datasets, including eruption, earthquake, and tremor catalogs.

The two-dimensional UMAP representation of the analyzed seismic data reveals distinct patterns that correlate with volcanic activity. The resulting seismogram atlas shows isolated clusters of points forming continuous features, which correspond to co-eruptive tremors. During non-eruptive periods, the analyzed time windows accumulate in a dense point cloud. Within this cloud, a predominantly random distribution of points is evident. However, some points form nearly linear, continuous pathways within the cloud, correlating with periods of magmatic intrusions. Adjacent to the dense point cloud, pre-eruptive seismic swarms are grouped in a specific region of the UMAP space, suggesting a common underlying mechanism.