Exploring the relative merits of signal classification methods - application to volcanic unrest

Abstract

Exploring the Relative Merits of Signal Classification Methods - Application to Volcanic Unrest ROWE, C. A., Los Alamos National Laboratory, Los Alamos, NM, char@lanl.gov; FALSAPERLA, S., Istituto Nazionale di Geofisica, Catania, Italy, falsaperla@ct.ingv.it; MACEIRA, M., Los Alamos National Laboratory, Los Alamos, NM, mmaceira@lanl.gov; LANGER, H., Istituto Nazionale di Geofisica, Catania, Italy, langer@ct.ingv.it; BEHNCKE, B., Istituto Nazionale di Geofisica, Catania, Italy, behncke@ct.ingv.it Volcanoes are among the most prolific producers of sustained and repeatable seismic signals. As such, observatories worldwide share the problem of timely evaluation of signals during periods of unrest, when the rate of incoming data can overwhelm analysts and result in incomplete cataloging as well as possible overlooking of important shifts in activity. We apply automated methods to aid in real-time signal detection, classification and evaluation for active volcanoes. Adaptive coherency-based waveform cross-correlation has been successfully applied both for retroactive classification and phase pick adjustment in large catalogs of discrete earthquakes, but also for correlation scanning (the "matched filter" method) in both volcanic and nonvolcanic settings. The subspace detector method has recently been applied successfully to isolate low frequency events within nonvolcanic tremor, as well as for discrete tectonic and volcanic earthquake signals. Here we explore the utility of the subspace detector to recognize and characterize volcanic behavior both during periods exhibiting discrete events and during episodes of tremor. We compare the adaptability and sensitivity to a parallel analysis using matched filter methods, and we discuss statistically-based decision-making approaches regarding the successful classifications of detections. For instance, application of the subspace detector to nonvolcanic tremor in Japan resulted in a projection operator whose detection threshold was best determined using a beta probability distribution, whereas scanning a continuous time series during 2004 unrest at Mount St. Helens suggested a log-normal probability distribution as a more appropriate fit to the values. We build on these earlier observations and test the robustness of the algorithm at a variety of volcanoes.