Advancing Local Distance Discrimination of Explosions and Earthquakes With Joint P/S and M_L‐M_C Classification

Abstract

Classification of local‐distance, low‐magnitude seismic events is challenging because signals can be numerous and difficult to characterize with approaches developed for larger magnitude events observed at greater distances. Yet, accurate classification is important to studies of earthquake processes and detection of potential underground nuclear tests. Here, we combine two classification metrics: the three‐component ratio of high‐frequency P/S amplitudes and the difference between local and coda duration magnitudes (ML‐MC). The metrics use different parts of the high‐frequency wavefield and exhibit complementary sensitivity for classification of M ∼ 0.5–4 natural earthquakes and borehole explosions, which are the best analog for underground nuclear explosions. Using means from bootstrap resampling across four diverse geologic settings, joint classification achieves >94.4% true positives and <8.4% false positives when using ≥8 seismographs within 200 km of the source. This high performance is obtained without local site corrections, indicating that the method may be transportable for local event classification.

Combining Deep Learning With Physics Based Features in Explosion‐Earthquake Discrimination

“…The Salton Seismic Imaging Project (SSIP) was an active source seismic survey conducted in 2011 to image crustal faults and constrain rifting processes beneath the Salton Sea (Fuis et al, 2017;Han et al, 2016), during which 41 shots (M L 0.6-2.1) were conducted. The United States Geological Survey reported 76 events in this region (M L 1-3.6) during the same time of the survey, including six borehole shots mis-labeled as earthquakes, as detailed in Wang et al (2021). In total, the SSIP data set has 2,307 explosion and 4,047 earthquake three-component records.…”

“…P/S ratios are calculated from 10 to 18 Hz filtered three-component waveforms using the corresponding phases shown in Equation 1 in Wang et al (2021). Signal-to-Noise Ratio (SNRs) for the three component waveforms are also computed using the predicted P windows and noise windows (i.e., 10 s before the P arrival) for quality control purposes, with a cutoff threshold of SNR >2 to declare a valid P/S ratio of a given event-station pair.…”

“…Signal-to-Noise Ratio (SNRs) for the three component waveforms are also computed using the predicted P windows and noise windows (i.e., 10 s before the P arrival) for quality control purposes, with a cutoff threshold of SNR >2 to declare a valid P/S ratio of a given event-station pair. More details about the parameters and regional velocity models used are described in (Wang et al, 2021).…”

“…The websites for the datasets used in this study are: SPE: https://doi.org/10. 7914 Wang et al (2021). All the analysis is done in Python and the deep learning framework used here is Tensor-Flow (Abadi et al, 2016), and seismological related analysis used Obspy (Beyreuther et al, 2010;Krischer et al, 2015).…”

Seismic Discrimination Between Nuclear Explosions and Natural Earthquakes using Multi-Machine Learning Techniques

“…In this research, Statistical approaches that discriminate by examining P-wave and S-wave spectra relative to a standard earthquake source model (Allmann et al, 2008), Complexity and spectral ratio parameters, and body wave (mb) and surface wave magnitudes (Ms) (Holt et al, 2019;Koper et al, 2016;Wang et al, 2021) were applied for discrimination between earthquakes and nuclear explosions based on six distinctive features extracted from event waveforms as a first step and then applying seven ML models to these features of five global regions around the world (China, USA, North Korea, Pakistan and India) utilizing IRIS data to enhance the accuracy of manual features extraction.…”