Abstract
The high-frequency Pg/Lg discriminant is studied between frequencies of 0.5 and 10 Hz using 294 NTS explosions and 114 western U.S. earthquakes recorded at four broadband seismic stations operated by Lawrence Livermore National Laboratory. The stations are located at distances of about 200 to 400 km from the Nevada Test Site (NTS). Event magnitudes ranged from about 2.5 to 6.5, and propagation paths for the earthquakes range from approximately 175 to 1300 km. The discriminant is shown to be very effective and displays improved separation between earthquakes and explosions as frequency is increased. Because of propagation effects, it is important to apply distance corrections directly to the amplitude ratios or to the magnitude-corrected amplitudes prior to computing the ratios. Multivariate discrimination analysis using both maximum-likelihood Gaussian classifiers and a backpropagation neural network show that approximately 95% of the events can be correctly identified. Both classification procedures were designed to handle missing data filled in using a nearest-neighbor algorithm. Except for a few notable exceptions, most of the earthquake misclassifications occur for mb < 4, which is expected for events having reduced signal-to-noise ratios. All of the explosion misclassifications occur for mb < 4, suggesting a source or near-source effect rather than an effect of poor signal-to-noise ratio. The explosions that were misclassified were typically of magnitude large enough to be classified correctly by mb/Ms or Love-wave energy. The main drawback of the Pg/Lg discriminant is that, because of signal-to-noise considerations and propagation effects, the number of measurements are reduced considerably at higher frequencies. It is expected that the problem will be amplified as magnitudes are reduced and event-receiver distances are increased.