Classifying Regional Seismic Signals Using TDDN-alike Neural Networks

Abstract

We present a shared-weights neural architecture for classifying spectrograms from regional seismic events. This approach does not use any high-level information available (like epicenter, magnitude, time origin etc.) on the event other than the estimated arrival times for the P-wave and eventually the S-wave. This severe restriction is to make sure that the system continues to function in the absence of high-level information. In spite of the very noisy character of the spectrograms, the classification rates obtained are comparable to those obtained by methods which do depend on the availability of high-level information (≈ 90 – 98%), though the rejection rate on non-trained events is rather high (≈ 20%).