Item

Abstract

Analyzing large seismic datasets and extracting information is becoming more challenging with the continuously growing amount of seismic records. Machine learning (ML) techniques have been utilized as powerful statistical tools for efficient seismic processing. Among other applications, ML algorithms allow the clustering of seismic data in order to reveal different patterns in the data or to identify types of signals for further analysis. The dominant uses of clustering algorithms in seismology have been in the realm of transient earthquake signal analysis. Clustering long lasting signals like volcanic tremors is however another appealing problem that could benefit from ML techniques albeit being slightly more complicated due to their high variability in signal properties such as time duration and time-frequency content. Here we use deep clustering (combination of deep learning and clustering) in order to cluster lava fountaining episodes which are recorded as tremor episodes in the seismic waveform between 2 May and 14 June 2021 during the Fagradalsfjall eruption in Iceland. Using an autoencoder, our model simultaneously learns feature representations and assigns clusters to them. The tremor episodes show systematic changes during the eruptive periods of Fagradalsfjall eruption consisting of distinct patterns with changing tremor duration, repose time and corresponding amplitude. The relation between tremor duration and repose time and their regular changes can indicate special volcanic activity stages containing starting, evolving, and stabilizing sequences. Unsupervised deep learning techniques help to automatically identify patterns in the data, find similar/dissimilar pulses, and lead to a better understanding of the subsurface processes and eruptive activities. The primary investigation on tremor pulses clustering is promising while further analysis is ongoing.