Automatic classification of volcanic earthquakes by using multi-layered neural networks

The application of neural networks as classifiers of seismic events is described with the aim of developing an automatic system for the classification of 'explosion quakes' at the Stromboli volcano. The architecture of the network that we trained to identify four different classes of shocks was a Multi-Layer Perceptron, using the Back Error Propagation algorithm. Five different approaches for representing the information embedded in the seismograms, both in the time and in the frequency domain, were considered, and the results compared. The direct use of the time series of the shocks was not satisfactory. The auto-correlation function worked well, but in some cases it was misleading. A better performance was obtained with a frequency domain representation. Finally, the use of the envelope function did not work well. Combining parameters such as the auto-correlation and envelope functions can improve one source of error, but it may introduce new ones. The performance obtained highlights the importance of the data attributes used for the training of the network. Topologies with eight neurons in a single hidden layer gave, on average, the best results among the considered neural network structures. The overall results provide a large number of events (89% with the best performance) correctly classified, indicating that this automatic technique is reliable, and encouraging further applications in the field of volcanic seismology.