Automatic discrimination of earthquakes and false events in seismological recording for volcanic monitoring

This paper reports on the classification of earthquakes and false events (thunders, quarry blasts and man-made undersea explosions) recorded by four seismic stations in the Vesuvius area in Naples, Italy. For each station we set up a specialized neural classifier, able to discriminate the two classes of events recordered by that station. Feature extraction is done using both the linear predictor coding technique and the waveform features of the signals. The use of properly normalized waveform features as input for the MLP network allows the network to better generalize compared to our previous strategy applied to a similar problem [2]. To train the MLP network we compare the performance of the quasi-Newton algorithm and the scaled conjugate gradient method. On one hand, we improve the strategy used in [2] and on the other hand we show that it is not specific to the discrimination task [2] but has a larger range of applicability