EVALUATING ROBUSTNESS OF A HMM-BASED CLASSIFICATION SYSTEM OF VOLCANO-SEISMIC EVENTS AT COLIMA AND POPOCATEPETL VOLCANOES Guillermo Cort´ es (1) , Ra´ ul Ar´ ambula (2) , Ligdamis A. Guti´ errez (1) ,Carmen Ben´ ıtez (1) , Jes´ us Ib´ a˜ nez (3) , Philippe Lesage (2) , Isaac ´ Alvarez (1) and Luz Garc´ ıa (1) (1) Dpto. de Teor´ ıa de la Se ˜ nal, Telem´ atica y Comunicaciones, Universidad de Granada. Granada, Spain. (2) Instituto de Geof´ ısica, Universidad Nacional Aut´ onoma de M´ exico. D. F. M´ exico, M´ exico. (3) Instituto Andaluz de Geof´ ısica. Granada, Spain. ABSTRACT This work presents a continuous volcano-seismic classifica- tion system based in the Hidden Markov Models as solution to recently strong needs for automatic event detection and recognition methods in early warning and monitoring scena- rios. Furthermore, our system includes a reliable method to assign confidence measures to the recognized signals in order to evaluate the robustness of the results. Data from the two most active volcanoes have been used to probe the system reliability on a complex joint corpus achieving a recognition accuracy higher than 78% in blind recognition tests. Index Terms— volcano monitoring, classification, Hi- dden Markov Models, volcano-seismic events, reliability 1. INTRODUCTION The event classification at active volcanoes is usually a hard and time-consuming task carried out by expert technicians in a non-stop monitoring process. Nowadays, automatic classifi- cation and detection methods are being developed in order to help with this task and for building early warning robust sys- tems. In this area, the Hidden Markov Models (HMM) [1, 2] are arising as the most promising solution as it was in Auto- matic Speech Recognition (ASR) [3, 4, 5, 6] versus other a- pproaches [7]. The HMM approach is also valid when SNR is too low for the classic detection algorithms to properly work [6]. In the present paper we advance one step forward follo- wing the ideas already stated in ASR. Previous works [3, 4, 6] have probed the parallelism between speech and volcano- seismic events in terms of signal complexity and real-time requirements. Multi-speaker ASR databases lead to indepen- dent classification models able to transcribe the speech of an unknown speaker. Our aim is the design of a general recogni- tion system integrating several volcano-seismic classes from different types of volcanoes to be effortless portable to any This work has been partially supported by the VOLUME UE (FP6-2004- Global-3-018471) and HISS (CGL2008-01660) projects Fig. 1. Seismograms and spectrograms (frequency ver- sus time) of the database classes: (T1) tremor with repet- itive pulses, (T2) spasmodic tremor, (T3) harmonic tremor, (REG) regional earthquakes, (COL) collapses, (LAH) lahars, (EXP) explosions, (LP) long period events and (VT) volcano- tectonic earthquakes. volcano monitoring station. At first approach, events from two Mexican volcanoes of the same type have been mixed for building a complex database used to evaluate the robust- ness of a continuous HMM-based recognition system. For increasing system reliability, a method to assign class confi- dence scores to each recognized event has been implemented facilitating result interpretation by experts in a critical alert situation or simply in a signal cataloging or monitoring work. We also made a comparative study between the system of the mixed corpus and the ones built with standalone databases II - 1012 978-1-4244-3395-7/09/$25.00 ©2009 IEEE IGARSS 2009