Computationally Efficient Search for Similar Seismic Signals in Continuous Waveform Data over a Seismic Network

Cross-correlating continuous seismic data streams with waveform templates has proven to be a sensitive, discriminating detector of repeating seismic signals; however, template matching requires a priori knowledge of the signals we wish to detect. Detection of unknown sources is possible with autocorrelation, which searches for similar waveforms within all pairs of short overlapping windows from continuous data. Unfortunately, naïve application of autocorrelation scales quadratically with time, which limits its use to short duration time series.

We previously developed an efficient, correlation-based approach to find similar seismic waveforms. We avoid comparing most non-similar signals by first developing compact, discriminative “fingerprints” of waveforms, and then assigning signals to sub-groups (buckets) using locality-sensitive hash functions (LSH). The probability that two signals enter the same bucket increases monotonically with similarity. LSH trades space for speed, requiring near-linear storage, yet yielding near-constant query time, and avoiding nearly all of the unproductive pair-wise computation of autocorrelation.

Our method previously detected uncataloged earthquakes 40 times faster than autocorrelation when applied to 24 hours of single-channel continuous data from one station in the Northern California Seismic Network. Here we extend our method to incorporate multiple channels of continuous data from the distributed network of HRSN stations at Parkfield. Our goal is to detect low frequency earthquakes (LFEs), and to compare our detections to LFE events previously identified using waveform templates. This is a challenging test since LFEs are more difficult to find than earthquakes due to their non-impulsive nature, low snr, and only modest waveform similarity.