Exploring Earth’s Vibrations Using Machine Learning: From Giraffe Footfalls to Volcanic Eruptions
Our planet undergoes continuous unrest across a broad spectrum of scales, ranging from the gentle footfalls of giraffes to the immense forces unleashed during volcanic eruptions and megathrust earthquakes. Seismic sensors bear witness to this diverse unrest, capturing valuable insights into active geological features and surface dynamics. However, the complexity and volume of seismic data pose challenges to efficiently extracting meaningful information hidden within vast datasets. Automated algorithms analyzing continuous data streams offer a promising solution, capable of uncovering hidden signal patterns and providing new perspectives on the task at hand.
In this seminar, I will introduce machine learning strategies leveraging wavelet transforms (a scattering network) to derive meaningful and continuous patterns from seismograms, identifying groups of seismic signals in a data-driven manner. Through the application of uniform manifold approximation and projection (UMAP), seismic data are then transformed into a two-dimensional representation – termed a seismogram atlas – facilitating visualization and interpretation of signal content within extensive seismic datasets (see attached image). I will present several case studies revealing insights into 1) the signal-modifying effects of surface freezing, 2) a continuously evolving magmatic plumbing system (see attached image), and 3) the distinctive footfall signal characteristics exhibited by various mammal species in the African savanna. The study cases show that continuous seismograms are indeed a goldmine of information, spawning new areas of inquiry within seismology.