Feature Extraction Using Circular Statistics Applied to Volcano Monitoring C´ esar San-Martin 1,4 , Carlos Melgarejo 1,2 , Claudio Gallegos 1,2 , Gustavo Soto 3 , Millaray Curilem 4 , and Gustavo Fuentealba 5 1 Information Processing Laboratory, Department of Electrical Engineering, Universidad de La Frontera. Casilla 54-D Temuco, Chile csmarti@ufro.cl 2 Observatorio Volcanol´ogico de los Andes del Sur Dinamarca 691, Temuco, Chile cmelgarejo@sernageomin.cl, cgallegos@sernageomin.cl 3 Center for Mathematical Model, Universidad de Chile Casilla 412-3, Santiago, Chile gsoto@dim.uchile.cl 4 Department of Electrical Engineering, Universidad de La Frontera. Casilla 54-D Temuco, Chile millaray@ufro.cl 5 Department of Physics, Universidad de La Frontera. Casilla 54-D Temuco, Chile gustavo@ufro.cl Abstract. In this work, the applicability of the circular statistics to feature extraction on seismic signals is presented. The seismic signals are captured from Llaima Volcano, located in Southern Andes Volcanic Zone at 38 ◦ 40’S 71 ◦ 40’W. Typically, the seismic signals can be divided in long- period, tremor, and volcano-tectonic earthquakes. The seismic signals are time-segmented using a rectangular window of 1 minute of duration. In each segment, the instantaneous phase is calculated using the Hilbert Transform, and then, one feature is obtained. Thus, the principal hy- pothesis of this work is that the instantaneous phase can be assumed as a circular random variable in [0, 2π) interval. A second feature is ob- tained using the wavelet transform due to the fact that seismic signals present high energy located in low frequency. Then, in the range 1.55 and 3.11 Hz the wavelet coefficients were obtained and their mean en- ergy is calculated as the second feature. Real seismic data represented using this two features are classified using a linear discriminant with a 92.5% of correct recognition rate. Keywords: seismic classifications, feature extraction, circular statistic, wavelet transform. 1 Introduction An adequate study of the activity of an active volcano requires the use of indirect methods for evaluating information related to the dynamics of magma [1]. The scientific literature has shown that volcanic activity can generate a wide range of seismic signals [2]. The study of the waveforms of these signals differentiate the I. Bloch and R.M. Cesar, Jr. (Eds.): CIARP 2010, LNCS 6419, pp. 458–466, 2010. c  Springer-Verlag Berlin Heidelberg 2010