We present an automatic P‐ and S‐wave onset‐picking algorithm, using kurtosis‐derived characteristic functions (CF) and eigenvalue decompositions on three‐component seismic data. We modified the kurtosis CF (Saragiotis et al., 2002) to improve pick precision by computing the CF over several frequency bandwidths, window sizes, and smoothing parameters. Once phases are picked, our algorithm determines the onset type (P or S) using polarization parameters, removes bad picks using a clustering procedure and the signal‐to‐noise ratio (SNR) and assigns a pick quality index based on the SNR.

We tested our algorithm on data from two different networks: (1) a 30‐station, 100×100  km array of mostly onland wideband seismometers in a subduction context and (2) a four‐station, 7×4  km array of ocean‐bottom seismometers over a midocean ridge volcano. We compared picks from the automatic algorithm with manual and short‐term average/long‐term average (STA/LTA)‐based automatic picks on subsets of each dataset. For the larger array, the automatic algorithm resulted in more locations than manual picking (133 versus 93 locations out of 163 total events detected), picking as many P onsets and twice as many S onsets as with manual picking or the STA/LTA algorithm. The difference between manual and automatic pick times for P‐wave onsets was 0.01±0.08  s overall, compared with −0.18±0.19  s using the STA/LTA picker. For S‐wave onsets, the difference was −0.09±0.23  s, which is comparable to the STA/LTA picker, but our picker provided nearly twice as many picks. The pick accuracy was constant over the range of event magnitudes (0.7–3.7 Ml). For the smaller array, the time difference between our algorithm and manual picks is 0.04±0.17  s for P waves and 0.07±0.08  s for S waves. Misfit between the automatic and manual picks is significantly lower using our procedure than using the STA/LTA algorithm.

You do not currently have access to this article.