How Unique is Any Given Seismogram? - Exploring Correlation Methods to Identify Explosions

Abstract:

As with conventional wisdom about snowflakes, we would expect it unlikely that any two broadband seismograms would ever be exactly identical. However depending upon the resolution of our comparison metric, we do expect, and often find, bandpassed seismograms that correlate to very high levels (>0.99). In fact regional (e.g. Schaff and Richards, 2011) and global investigations (e.g. Dodge and Walter, 2015) find large numbers of highly correlated seismograms. Decreasing computational costs are increasing the tremendous potential for correlation in lowering detection, location and identification thresholds for explosion monitoring (e.g. Schaff et al., 2012, Gibbons and Ringdal, 2012; Zhang and Wen, 2015). We have shown in the case of Source Physics Experiment (SPE) chemical explosions, templates at local and near regional stations can detect, locate and identify very small explosions, which might be applied to monitoring active test sites (Ford and Walter, 2015).

In terms of elastic theory, seismograms are the convolution between source and Green function terms. Thus high correlation implies similar sources, closely located. How do we quantify this physically? For example it is well known that as the template event and target events are increasingly separated spatially, their correlation diminishes, as the difference in the Green function between the two events grows larger. This is related to the event separation in terms of wavelength, the heterogeneity of the Earth structure, and the time-bandwidth of the correlation parameters used, but this has not been well quantified. We are using the historic dataset of nuclear explosions in southern Nevada to explore empirically where and how well these events correlate as a function of location, depth, size, time-bandwidth and other parameters. A goal is to develop more meaningful and physical metrics that go beyond the correlation coefficient and can be applied to explosion monitoring problems, particularly event identification.