Array processing of teleseismic body waves with the USArray

摘要

We introduce a novel method of array processing for measuring arrival times and relative amplitudes of teleseismic body waves recorded on large aperture seismic arrays. The algorithm uses a robust stacking algorithm with three features: (1) an initial 'reference' signal is required for initial alignment by cross-correlation; (2) a robust stacking method is used that penalizes signals that are not well matched to the stack; and (3) an iterative procedure alternates between cross-correlation with the current stack and the robust stacking algorithm. This procedure always converges in a few iterations making it well suited for interactive processing. We describe concepts behind a graphical interface developed to utilize this algorithm for processing body waves. We found it was important to compute several data quality metrics and allow the analyst to sort on these metrics. This is combined with a 'pick cutoff function that simplifies data editing. Application of the algorithm to data from the USArray show four features of this method. (1) The program can produce superior results to that produced by a skilled analyst in approximately 1/5 of the time required for conventional interactive picking. (2) We show an illustrative example comparing residuals from S and SS for an event from northern Chile. The SS data show a remarkable +10s residual pattern across the USArray that we argue is caused by propagation approximately parallel to the subduction zones in Central and South America. (3) Quality metrics were found to be useful in identifying data problems. (4) We analyzed 50 events from the Tonga-Fiji region to compare residuals produced by this new algorithm with those measured by interactive picking. Both sets of residuals are approximately normally distributed, but corrupted by about 5% outliers. The scatter of the data estimated by waveform correlation was found to be approximately 1/2 that of the hand picked data. The outlier populations of both data sets are likely produced by cycle skips, but the distribution of hand picked data show a more diffuse departure from a Gaussian error model.