Discrimination of the DPRK underground explosions and their aftershocks using the P/S spectral amplitude ratio

We have estimated the performance of discrimination criterion based on the P/S spectral amplitude ratios obtained from six underground tests conducted by the DPRK since October 2006 and six aftershocks induced by the last two explosions. Two aftershocks were detected in routine processing at the IDC. Three aftershocks were detected by a prototype waveform cross correlation procedure with explosions as master events, and one aftershock was found with the aftershocks as master event. Two seismic arrays USRK and KSRS of the IMS and two non-IMS 3-C stations SEHB (South Korea) and MDJ (China) were used. With increasing frequency, all stations demonstrate approximately the same level of deviation between the Pg/Lg spectral amplitude ratios belonging to the DPRK explosions and their aftershocks. For a single station, simple statistical estimates show that the probability of any of six aftershocks not to be a sample from the explosion population is larger than 99.996% at the KSRS and even larger at USRK. The probability of any of the DPRK explosion to be a representative of the aftershock population is extremely small as defined by the distance of 20 and more standard deviations to the mean explosion Pg/Lg value. For network discrimination, we use the Mahalanobis distance combining the Pg/Lg estimates at three stations: USRK, KSRS and MDJ. At frequencies above 4 Hz, the (squared) Mahalanobis distance, D2, between the populations of explosions and aftershocks is larger than 100. In the frequency band between 6 and 12 Hz at USRK, the aftershocks distance from the average explosion D2>21,000. Statistically, the probability to confuse explosions and aftershocks is negligible. These discrimination results are related only to the aftershocks of the DPRK tests and cannot be directly extrapolated to the population of tectonic earthquakes in the same area.

💡 Research Summary

**
The paper evaluates the feasibility of using the P‑to‑S spectral amplitude ratio as a discrimination tool for separating underground nuclear explosions conducted by the Democratic People’s Republic of Korea (DPRK) from their induced aftershocks. Six DPRK underground tests performed between October 2006 and September 2017, together with six small events that occurred shortly after the two largest tests, constitute the data set. The authors employ four regional stations: two International Monitoring System (IMS) arrays (USRK and KSRS) and two non‑IMS three‑component stations (SEHB in South Korea and MDJ in China).

Waveforms are filtered into six octave‑band Butterworth filters (1–2 Hz, 1.5–3 Hz, 2–4 Hz, 3–6 Hz, 4–8 Hz, and 6–12 Hz). For each band the maximum amplitudes of the Pg (compressional) and Lg (surface) phases are measured, and their ratio (Pg/Lg) is taken as the P/S spectral amplitude ratio. At low frequencies (≤2 Hz) the ratio shows little separation, but from 4 Hz upward a clear distinction emerges: explosions exhibit Pg/Lg ratios around 1.8–2.2, whereas aftershocks have ratios below 0.6, a three‑fold difference in the 6–12 Hz band.

Statistical analysis at the single‑station level uses the mean and standard deviation of the Pg/Lg ratios for the two populations. Z‑scores indicate that each aftershock lies more than 4.5 standard deviations away from the explosion mean (confidence > 99.996 %) at KSRS, with even larger separations at USRK. Conversely, the probability that an explosion belongs to the aftershock population is effectively zero (exceeding 20 σ).

To assess network‑level discrimination, the authors construct a three‑dimensional vector of Pg/Lg ratios from USRK, KSRS, and MDJ and compute the Mahalanobis distance (D²) between the explosion and aftershock clusters. For frequencies above 4 Hz, D² exceeds 100; in the 6–12 Hz band at USRK, D² reaches ≈2.1 × 10⁴, corresponding to a p‑value well below 10⁻⁸. Thus, even when combining data from multiple stations, the two groups are statistically distinct with negligible confusion risk.

Detection of aftershocks is enhanced by waveform cross‑correlation. While the International Data Centre (IDC) routine processing identified only five aftershocks, cross‑correlation using explosion templates recovered three additional events, and using aftershock templates uncovered a further event. Template lengths ranging from 10 to 150 s and a short‑term/long‑term average (STA/LTA) detector with STA = 0.1 s, LTA = 20 s, and a threshold of 5 were found optimal. Aftershock templates yielded higher absolute correlation coefficients (|CC|) than explosion templates, reflecting greater waveform similarity among aftershocks and enabling detection of weaker signals.

The study acknowledges limitations: the aftershocks examined are all located near the DPRK test site, so the results cannot be directly extrapolated to tectonic earthquakes in the same region. High‑frequency data (>10 Hz) are limited at some stations due to sampling constraints (e.g., KSRS and SEHB have a Nyquist frequency of 10 Hz). Moreover, the analysis relies on the assumption of multivariate normality for Mahalanobis distance calculations.

Future work should expand the station network, incorporate higher‑frequency sensors, and test the method against a broader set of natural earthquakes and other anthropogenic sources. Integrating machine‑learning classifiers that combine P/S ratios with additional features (e.g., arrival‑time residuals, moment‑tensor estimates) could further improve real‑time discrimination capabilities for the Comprehensive Nuclear‑Test‑Ban Treaty monitoring system.