Fast Discrimination of Local Earthquakes Using a Neural Approach

Abstract

In this article, we describe a neural network method for the fast discrimination between local earthquakes and regional and teleseismic earthquakes using seismic records from a single station. Neural networks are data-driven nonlinear classifiers that learn from experience and can model real-world complex relationships. For the discrimination task, we implement a twolayer feed-forward multilayer perceptron (MLP). MLP is a supervised technique that accomplishes the learning process using a preclassified dataset for the training phase. The dataset includes 70 teleseisms, 79 regional earthquakes, and 103 local earthquakes. The seismic events are recorded at a single station, equipped with a short-period sensor.We parameterize the seismograms in the frequency domain, using the linear predictive coding (LPC). This technique is mostly used in audio signal processing for efficiently encoding frequency features of digital signals in a compressed form. The obtained spectral features, or LPC coefficients, are the input to the neural model. We carry out several tests by shortening from 4 to 1 s the time-window duration used for the LPC analysis. The proposed algorithm achieves a correct classification of 98.5% and 97.7% in discriminating local versus regional and local versus teleseismic earthquakes, respectively, on a 1-s time window. These results indicate that our discrimination algorithm can be profitably exploited in automatic analyses of seismic data that require fast responses, such as seismological monitoring systems and earthquake early warning systems.