A new machine learning-based data assimilation technique to detect volcanic unrest from tremor

Linking geophysical and geochemical observables with subsurface processes is crucial to detect volcanic unrest and better anticipate eruptions. One of the most important observables to monitor pre-eruptive volcanic activity is tremor, a more or less persistent, highly periodic, ground vibration recorded near active vents. Tremor is commonly being monitored in near real-time by volcano observatories to anticipate unrest, as it may emerge, or change properties, when subsurface pressure varies. For example, it has been observed that the dominant frequency of tremor may glide towards higher or lower values before eruptions; overtones may appear or disappear; and seismic amplitude may increase or decrease. However, similar variations can be also observed during quiescence and when activity decreases. This leads to the following questions: How does tremor actually reflect the overpressure of the subsurface? Can we infer when and where the pressure beneath active vents increases or decreases by monitoring volcanic tremor? In this work, we present a new data assimilation technique that combines new physics-based models of volcanic tremor with a machine learning-based inversion algorithm to track pressure changes beneath volcanic craters in near-real time. In particular, our inversion algorithm is based on a supervised random forest classifier trained with synthetic data, whereas our physics-based model extends from Girona et al. (2019) and is based on a stop-and-go mechanism, i.e., tremor is assumed to emerge when: (i) gas is supplied randomly to shallow levels of the volcanic plumbing system; (ii) accumulates temporarily beneath permeable caps (e.g., beneath a dome or in a leaky fracture); and (iii) transfers via permeable flow to the surface. Using this machine learning-based data assimilation technique, we find that the recent 2013 unrest phase of Kawah-Ijen volcano (Indonesia) was driven by a pressure increase in the subsurface of a factor 2-to-5. This technique is currently also being applied to unveil the pressure history of the shallow vents of Pavlof and Veniaminof volcanoes (Alaska). 

Girona, T., C. Caudron, C. Huber (2019). Origin of shallow volcanic tremor: the dynamics of gas pockets trapped beneath thin permeable media. J. Geophys. Res., doi: 10.1029/2019JB017482.