Spectral classification methods in monitoring small local events by the Israel seismic network

We use the dense Israel Seismic Network (ISN) to discriminate between low magnitude earthquakes and explosions in the Middle East region. This issue is important for CTBT monitoring, especially when considering small nuclear tests which may be conducted under evasive conditions. We explore the performance of efficient discriminants based on spectral features of seismograms using waveforms of 50 earthquakes and 114 quarry and underwater blasts with magnitudes 1.0-2.8, recorded by ISN short-period stations at distances up to 200 km. The single-station spectral ratio of the low and high-frequency seismic energy shows an overlap between explosions and earthquakes. After averaging over a subnet of stations, the resolving power is enhanced and the two classes of events are separated. Different frequency bands were tested; the (1-3 Hz)/(6-8 Hz) ratio provided the best discriminant performance. We also estimated normalized r.m.s. spectral amplitudes in several sequential equal frequency windows within the 1-12 Hz band and applied multiparametric automatic classification procedures (Linear Discrimination Function and Artificial Neural Network) to the amplitudes averaged over a subnetwork. A leave-one-out test showed a low rate of error for the multiparametric procedures. An innovative multi-station discriminant is proposed, based on spectral modulation associated with ripple-firing in quarry blasts and with the bubbling effect in underwater explosions. It utilizes a distinct azimuth-invariant coherency of spectral shapes for different stations in the frequency range (1-12 Hz). The coherency is measured by semblance statistics commonly used in seismic prospecting for phase correlation in the time domain. After modification, the statistics applied to the network spectra provided event separation. A new feature of all the above mentioned procedures is that they are based on smoothed (0.5 Hz window), instrument-corrected FFT spectra of the whole signal; they are robust to the accuracy of onset time estimation and, thus well suited to automatic event identification.