The goal of this project is to detect and classify volcanic events in time-series data from seismographs. To that end, we utilize a state-of-the-art machine-learning methodology that hybridizes the power of deep learning with the expertise of domain experts. We train and test these models on a data set from the southern Taupo Volcanic Zone, New Zealand: specifically, the 2012 eruptions at the Te Maari craters of Mt. Tongariro on the North Island of New Zealand. These events happened to be along the popular hiking trail Tongariro Crossing. There were no injuries in the Te Maari eruptions, but in recent years other popular volcano tourist sites, e.g., Whakaari/White Island in New Zealand and Mount Ontake in Japan, have suffered similar steam-driven (phreatic) volcanic eruptions as Tongariro, but with tragic outcomes including multiple fatalities (Yamaoka et al., 2016; Park et al., 2020). Such steam-driven eruptions are particularly hard to anticipate at present owing to an absence of large-scale motion of magma that produces signals like surface deformation that portend a future eruption. Investigating the small earthquakes associated with volcanoes could elucidate changes in the interaction of magma and fluids that lead to eruptive activity. Our preliminary work suggests that machine-learning approaches can detect small earthquakes that were previously missed by human analysts. If this conjecture is borne out by our project, the results could scaffold new insights into the dynamics of earthquake swarms that are associated with eruptions -- insights that could lead to effective prediction methods for these dangerous events.
|
 |
