Evaluation of Cross-correlation Methods on a Massive Scale for Accurate Relocation of Seismic Events

Lamont-Doherty Earth Observatory (LDEO) at Columbia University is evaluating a method of locating seismic sources (earthquakes, explosions) based on the use of waveform cross-correlation (WCC) measurements instead of using the conventional measurements of seismic wave arrival time (phase picks). WCC measurements have been demonstrated to be 10 to 100 times more accurate, where they can be used. The principal issue we are exploring is the extent to which a significant fraction of seismicity can be located using WCC measurements. We have organized the work into two projects: (1) the application of WCC methods to relocate earthquakes and explosions in China and neighboring regions, and (2) the application of WCC methods to relocate earthquakes in parts of North America. Using a very sparse network it was discovered that about 10% of earthquakes in and near China between 1985 and 2000 with M ≥ 3 were repeating events that generated essentially the same signals from sources which could not be more than 1 km from each other. The estimated location precision is a few 100 m. A specialized case study for the 1999 sequence of events in Xiuyan, China, found that WCC on Lg-waves combined with the double-difference technique significantly improved the epicentral locations. Several reflectivity synthetic experiments have been conducted to understand how depth, distance to the station, and mechanism influence the similarity of the Lg waveforms. In the synthetic Lg waveforms there is greater sensitivity to depth, than to epicentral distance. In North America, we are studying the Charlevoix region in eastern Canada and the New Madrid seismic zone in the central United States. Datasets have been assembled from scratch working in conjunction with regional network operators. For Charlevoix courtesy, of the Geological Survey of Canada, we now have 2,472 events with corresponding catalog, phase, and waveform data. For New Madrid, two datasets have been acquired. The first is from the PANDA deployment between 1989 and 1992, which consists of 884 events with bulletin and waveform information. The second is from the Center for Earthquake Research and Information (CERI) network, operated by Memphis State University, which we currently have waveform data from 1995-2003, and catalog and phase data from 2000-2003 for 680 events, with catalog and phase data from 1995-1999 expected shortly. Preliminary WCC results of the PANDA network indicate that 68% (597 out of 884 events) correlate with cross-correlation coefficients above 0.7 at four or more stations. Four stations are the minimum required to obtain a location estimate. Both Pand S-waves are correlated on all three components. The window lengths are 1 s and the lags searched over are also 1 s. It appears in a few examples that similar correlations are possible over 1 to 10 kilometer inter-event separation distances due to a site resonance from soft sediments underneath certain stations. Subsequent work will continue the correlation analysis for the other regions and then perform locations using the doubledifference algorithm. OBJECTIVES To evaluate a method of locating seismic sources (earthquakes, explosions) that is based on use of waveform crosscorrelation (WCC) measurements instead of using the conventional measurements of seismic wave arrival time. WCC measurements are ten or a hundred times more accurate, where they can be used. The principal issue we shall explore is the extent to which a significant fraction of seismicity can be located using WCC measurements. RESEARCH ACCOMPLISHED We have organized the work into two Projects. Project 1 concerns application of WCC methods to relocate earthquakes and explosions in China and neighboring regions. Project 2 concerns application of WCC methods to relocate earthquakes in parts of North America. Project 1 For Project 1, we studied 14,000 earthquakes in China for the years 1985 to 2000, and in particular 130,000 regional seismograms recorded at a sparse network of stations up to 20 degrees distance. We made the surprising discovery that in some cases the complex, highly scattered Lg-wave is remarkably similar for clusters of events. We analyzed in detail a subset of 28 events out of 90 from the 1999 Xiuyan sequence associated with a damaging earthquake that was successfully predicted in Liaoning Province, northeast China. Excellent relative locations could be obtained using only four or five stations 500 to 1000 km away. This approach was expanded to cover all of China and surrounding regions, and we found that about 10% of seismic events in and near China for the fifteenyear period were repeating events not more than about 1 km from each other. Specifically we found a set of 1301 seismic events with the property that any one of them wrote regional seismograms (in many cases lasting for hundreds of seconds, and using a band from 0.5 to 5 Hz) almost exactly like those from at least one other event. Among these events, 950 waveform doublets could be identified, which can used to evaluate the precision of phase picks and standard catalogs. This work has been written up and accepted for publication in two papers (Schaff and Richards, 2004a,b). We continued a detailed study of the Xiuyan sequence of earthquakes in Liaoning Province, China. Specifically, we used a reflectivity code to generate synthetic regional waves in the passband 0.5 to 5 Hz, at a distance of 750 km (corresponding to the distance at which we have data), and found that the synthetics are far more sensitive to changes in source depth than they are to changes in epicentral distance. This is of interest in the context of interpreting event clusters that have approximately the same epicenter centroid, yet have different waveforms. We speculate that in this case the waveforms are different and do not cross-correlate well between clusters, because the clusters are centered at different depths. Project 2 For Project 2, we are carrying out detailed studies of the seismicity of two regions in North America, namely New Madrid in the Central United States, and Charlevoix, East Canada. Data Compilation For New Madrid, three datasets have been acquired. The first is from the PANDA (Portable Array for Numerical Data Acquisition) network deployment from Oct. 1989 Aug. 1991, which consists of 884 events with bulletin and waveform information. The second is from the network operated by CERI (Center for Earthquake Research and Information, University of Memphis) and we currently have waveform data from 1995 to 2003 and catalog and phase data from 2000 to 2003 for 680 events, with catalog and phase data from 1995 through 1999 expected shortly. The third data set is the Central Mississippi Valley seismic bulletin data from St. Louis University, which consists of over 60,000 phase picks from about 4,000 earthquakes that occurred during 1974-1998. For the Charlevoix seismic zone, courtesy of the Geological Survey of Canada, we now have 2,472 events with corresponding catalog, phase, and digital waveform data. Technique We have experimented with a correlation detector that is able to recover lags greater than half the window length. This is a new feature and different from the correlation function which was applied in our earlier work. When dealing with finite duration signals, time-domain cross correlation is computed by fixing one window on the first seismogram and moving a sliding window over the second seismogram padded with zeros (Figure 1). An equivalent result is obtained if the cross-correlation is computed in the frequency domain. Although the cross correlation function is technically defined for lags plus and minus the window length, in practice only lags less than or equal to half the window length can be recovered (Schaff et al., 2004). The reason is that beyond this point the percentage of similar energy in the two windows is less than 50%. A related effect is that the cross correlation coefficient measurement degrades with increasing initial offset of the two seismograms (Schaff et al., 2004). If instead of padding with zeros, the original data is retained in the second seismogram, both of these negative effects with correlation functions are eliminated (Figure 1). We call such an application employing a correlation detector. Now the sliding window can align arbitrarily long offsets and perfectly capture the correct correlation coefficient. 0 1 2 3 s correlation detector