Regional seismic event classification at the NORESS array: Seismological measurements and the use of trained neural networks

Abstract The results are presented from a two-part study of regional earthquakes and chemical explosions recorded by the NORESS seismic array. The first part of the study examines various signal parameters extracted from Pn, Sn , and Lg phases with regard to discrimination capability. These parameters include familiar spectral discriminants and other spectral measures that quantify high-frequency content, spectral complexity, and shear wave generation. Part two of the study focuses on an application of backpropagation learning to the problem of automatic event classification through the use of trained neural networks. Of the 95 events examined, 66 were selected for the classification study based on high signal-to-noise ratio and positive identification in local seismicity bulletins. Events are located in eastern Europe, southern and western Norway, Sweden, the western Soviet Union, and the Norwegian and Greenland Seas. Local magnitudes range from 1.4 to 4.7, and epicentral distances for most events are less than 1000 km. Results from the discrimination analysis indicate that the wide-band spectral ratios Pn/Sn and Pn/Lg provide good discrimination capability between earthquakes and mining explosions, although there are anomalous events in both populations and a region of overlapping event types. Mining explosions can frequently be identified by their spectral complexity as measured by the cepstra of Pn, Sn , and Lg . This complexity is assumed to occur due to a combination of ripple-firing of the charges and reverberations within the shallow source region. In part two of the study, an artificial neural network employing the backpropagation learning paradigm was trained with input vectors formed by the two spectral amplitude ratios and the mean cepstral variance. A length-2 output vector was binary coded to identify each input vector as an explosion or earthquake. Two hidden layers were used, consisting of 8 and 2 units, respectively. The network was trained first using input vectors from the entire data set. This resulted in 100 per cent correct classification when the events were processed with the trained network. This is compared to the optimum planar decision surface which resulted in 5 errors and 19 uncertain classifications. In a control experiment, the network was trained with half of the events and tested with the remaining half. This resulted in 5 errors and 2 uncertain classifications. This compares with 3 errors in training, 2 errors in testing, and a total of 19 uncertain events obtained by the optimal linear classifier. One apparent advantage of the neural network over the linear classifier is the network9s ability to determine complex patterns in the data, thus reducing the number of uncertain events.