Classiication of Seismic Signals by Integration of Neural Networks' Ensembles

We assess the applicability of a collection of Neural Networks' Ensembles for discrimination of Seismic Signals. The problem considered is of classifying local seismic events, into Natural Earthquakes and Man-Made Explosions, based on their waveforms recordings from a single seismometer. Several preprocessing procedures were applied and Ensembles of Neural Networks were trained on Bootstrap Sample Sets, using various network architectures and diierent signal representations. We discuss a conndence measure for the classiication, based on the agreement (variance) within the Ensembles, and present an algorithm for combining diierent classiiers on-line. Cross Validation tests and comparison with classical methods indicated that combining the results of a collection of Neural Network's Ensembles in a pluralistic environment, is a good way for handling high dimensional problems with non-stationary data distribution.