Assessing the Effectiveness of Generalized Likelihood Ratio Test Detector Schemes in Seismic Event Detection and the Avoidance of Non-Target Signals

Abstract:

Seismic event detection often relies on correlation-based techniques. Traditional correlation-based seismic detectors operate on the foundation of a binary hypothesis test where discrete-time data in question are determined to be either entirely noise or a combination of noise and signal of interest. In the context of the binary hypothesis, a false alarm is the false identification of noise as a real signal (with superimposed noise). Nuisance seismicity, or non-target signals, challenge the performance of correlation-based seismic detectors, especially when events of interest are small (magnitude < 2) [Carmichael, 2016; Carmichael and Hartse, 2016]. Non-target signals are sometimes artifact signals (e.g. data spikes), but are also often generated by real seismic sources that may differ in type and/or location from the target signals of interest. Carmichael and Hartse [2016] addressed nuisance seismicity by incorporating non-target signals into the null hypothesis. However, their method required inflated thresholds to maintain desired false alarm rates. While higher thresholds achieve acceptable false alarm rates, they also lower detection rates, especially when events of interest are small. Here, we assess the effectiveness of implementing Generalized Likelihood Ratio Test (GLRT) detectors. GLRTs compare the probability that the observed data are fit better by a more complex model (signal + noise) to the probability that the observed data are better explained by a simpler model (noise only). Successful implementation of GLRT detectors allows for the use of lower thresholds to separate signals from signals and noise (and non-target signals), as long as the suite of individual correlators spans the signal space. Thus, GLRT implementation is most suitable for scenarios involving highly repeatable seismicity such as mining [Dodge and Harris, 2019] and swarms [Maceira et al., 2010]. Dodge and Harris [2019] investigated a single mine (very compact region), and Maceira et al. [2010] searched for low frequency events occurring within an hour timeframe. We investigate the more spatially and temporally expansive dataset of the 2014 Napa aftershock sequence to assess the effectiveness of a GLRT detector scheme in detecting aftershocks while avoiding non-target signals. LLNL-ABS-812688