Direct Empirical Estimation of Arrival-Time Picking Error from Waveform Cross-Correlations

Error in picking arrival times of seismic phases by human analysts has concerned the seismological community since the inception of earthquake location via the Geiger (least-squares) method. Arrival-time picking error and the inaccuracy of the location model both contribute to errors in hypocentral estimates, but in such a way that they are usually not separable. Model error can be attenuated greatly by joint tomography inversions for 3D velocity and revised hypocenters. This article attempts to define picking error in a more rigorous way than the subjective judgments of human analysts concerning their accuracy and the estimates made from tomography. A precise way to estimate interevent arrival times at stations is possible through cross-correlation of waveforms. This method obtains an accurate interevent time against which interevent time from human picking can be compared, and a statistical treatment of picking error emerges from a large set of cross-correlations. Estimates of picking error can be compared with final data error estimates from joint tomography inversions and to analysts' own estimates of picking error. This article uses data from the Southern Great Basin Digital Seismic Network (SGBDSN) for the years 2000-2007. From waveform cross-correlations with this data, the standard deviation of picking error for good quality signals is 0.020 sec for P and 0.028 sec for S. The analysts' own estimates of picking error considerably exceed the values determined from cross-correlation of very good waveforms, by roughly a factor of 2 to 3 routinely. Joint structure/hypocenters tomography on a local scale (model extent < 100 km), as gleaned from the literature, reduces the standard deviation of travel-time errors to roughly 0.08 sec. The standard deviation of residuals for a tomography study using the SGBDSN data was 0.06 sec, larger by a factor of roughly three than the picking error estimate from cross-correlation of waveforms.