Seismic Waveform Separation using Signal Processing Strategies from the field of Music Information Retrieval

Abstract:

Inspired by similarities of seismic and acoustic signals we take advantage of the expertise developed in the field of music information retrieval. We adopt a modified version of the repetition/similarity method being used in musical signal processing to extract harmonic parts from a musical piece by analyzing repetitive patterns in time-frequency domain. Harmonic tremor can be interpreted as a superposition of rhythmically excited signals in the volcanic edifice in close analogy to harmonic music signal generation. We show that the periodic pattern of harmonic tremor can be successfully extracted using a similarity matrix in time-frequency domain allowing to separate spectrograms of repeating and non-repeating signal patterns. The harmonic tremor can be reconstructed from the spectrogram of repeating pattern by adding phase information. Likewise the non-repeating spectrogram reveals transient events. Integrating amplitudes over frequency at each time step we derive a characteristic function suitable for extracting transient events.

For verification we apply the proposed method to synthetically generated harmonic signal contaminated with colored noise. The test confirms that we are able to reconstruct the underlying tremor signal down to a Signal to Noise Ratio (SNR) of 0.1 with respect to the contaminating colored noise. In order to investigate the ability of the proposed method for earthquake detection we have additionally created sets of semi-synthetic data by combining synthetic harmonic, seismic noise and real earthquake waveform sections. The result shows that for SNR above 0.1 we can detect more than 78 percent of the events, however below SNR=0.3 there are higher number of false picks up to 30 percent. Currently we are comparing the number of detected earthquakes using our method and one month of continuous data during the Holuhraun 2014-2015 eruption in Iceland for single station single component data. So far we achieve a detection rate of 70% in comparison to the bulletin presented in Ágústsdóttir et al. (2019) when considering matching phase picks to have deviations less than 2 seconds.