Detection of weak seismic signals in noisy environments from unfiltered, continuous passive seismic recordings

Kinali, M. and Pytharouli, S. and Lunn, R. J. and Shipton, Z. K. and Stillings, M. and Lord, R. and Thompson, S. (2018) Detection of weak seismic signals in noisy environments from unfiltered, continuous passive seismic recordings. Bulletin of the Seismological Society of America, 108 (5). pp. 2993-3004. ISSN 0037-1106

[img]
Preview
Text (Kinali-etal-BSSA-2018-Detection-of-weak-seismic-signals-in-noisy-environments-from-unfiltered-continuous-passive-seismic-recordings)
Kinali_etal_BSSA_2018_Detection_of_weak_seismic_signals_in_noisy_environments_from_unfiltered_continuous_passive_seismic_recordings.pdf
Accepted Author Manuscript

Download (1MB)| Preview

    Abstract

    Robust event detection of low signal-to-noise ratio (SNR) events, such as those characterized as induced or triggered seismicity, remains a challenge. The reason is the relatively small magnitude of the events (usually less than 2 or 3 in Richter scale) and the fact that regional permanent seismic networks can only record the strongest events of a microseismic sequence. Monitoring using temporary installed short-period arrays can fill the gap of missed seismicity but the challenge of detecting weak events in long, continuous records is still present. Further, for low SNR recordings, commonly applied detection algorithms generally require pre-filtering of the data based on a priori knowledge of the background noise. Such knowledge is often not available. We present the NpD (Non-parametric Detection) algorithm, an automated algorithm which detects potential events without the requirement for pre-filtering. Events are detected by calculating the energy contained within small individual time segments of a recording and comparing it to the energy contained within a longer surrounding time window. If the excess energy exceeds a given threshold criterion, which is determined dynamically based on the background noise for that window, then an event is detected. For each time window, to characterize background noise the algorithm uses non-parametric statistics to describe the upper bound of the spectral amplitude. Our approach does not require an assumption of normality within the recordings and hence it is applicable to all datasets. We compare our NpD algorithm with the commonly commercially applied STA/LTA algorithm and another highly efficient algorithm based on Power Spectral Density using a challenging microseismic dataset with poor SNR. For event detection, the NpD algorithm significantly outperforms the STA/LTA and PSD algorithms tested, maximizing the number of detected events whilst minimizing the number of false positives.