Automatic detection of volcano-seismic events by modeling state and event duration in hidden Markov models

Highlights

• Volcano event durations are modeled in the HMM framework.

• State duration are modeling with max and min bounds.

• Event duration are modeled with max and min bounds and with a gamma distribution.

• New event penalization is incorporated to reduce false positives.

Abstract

In this paper we propose an automatic volcano event detection system based on Hidden Markov Model (HMM) with state and event duration models. Since different volcanic events have different durations, therefore the state and whole event durations learnt from the training data are enforced on the corresponding state and event duration models within the HMM. Seismic signals from the Llaima volcano are used to train the system. Two types of events are employed in this study, Long Period (LP) and Volcano-Tectonic (VT). Experiments show that the standard HMMs can detect the volcano events with high accuracy but generates false positives. The results presented in this paper show that the incorporation of duration modeling can lead to reductions in false positive rate in event detection as high as 31% with a true positive accuracy equal to 94%. Further evaluation of the false positives indicate that the false alarms generated by the system were mostly potential events based on the signal-to-noise ratio criteria recommended by a volcano expert.

Introduction

Active volcanoes are continuously monitored to observe the underlying volcano-seismic activities. Each volcano-seismic activity or event is associated to a source process and gives an insight into the current state of the volcano that could potentially be used as precursors of an eruption. Automatically detecting (or classifying) these events has gained importance because of the growing need to monitor a high number of active volcanoes, in addition to providing an efficient, consistent and reliable machine-based framework.

Hidden Markov Models (HMMs) can be a powerful and efficient tool that has therefore also been used within multiple applications (Rabiner, 1989). Among others, multiple HMM based volcano-seismic event and earthquake detection and classification techniques have been proposed in recent years (Ohrnberger, 2001, Alasonati et al., 2006, Gutiérrez et al., 2006, Gutiérrez et al., 2009, Benítez et al., 2007, Beyreuther et al., 2008, Beyreuther et al., 2012, Beyreuther and Wassermann, 2008, Beyreuther and Wassermann, 2011, Ibáñez et al., 2009, Bicego et al., 2013). From the volcanology point of view, HMM is a very convenient method because it can offer a mathematical framework to model signals that are composed of non-stationary events. Gutiérrez et al. (2006), and later Ibáñez et al. (2009), presented an HMM based classification system for volcano events recorded at Stromboli and Etna volcanoes. Data was used from field surveys of years 1997 and 1999. The authors conclude that the recognition system is very efficient but acknowledged that for an improved performance reliable labeling by an expert is vital. Benítez et al. (2007) proposed a classification system developed based on the seismic events recorded during the 1994–1995 and 1995–1996 surveys at the Deception Island Volcano, Antarctica. The events included volcano-tectonic earthquakes, long period events, and volcanic tremor. The authors claimed that the system provides a classification accuracy of 90% and is suitable for real-time operation. The system was also tested on another data set, mainly with LP events, observed during the 2001–2002 field survey, and an accuracy of around 95% was achieved. In Beyreuther et al. (2008), HMMs were employed to detect and classify volcano-tectonic or tectonic earthquakes in continuous seismic data. HMMs were built for events and noise with the HTK software. In the implementation two classifiers were running in parallel, one containing all models of events and noise, and another one with only noise models. Valuable information for the seismological analyst to evaluate the detections was provided by using confidence measure. Beyreuther and Wassermann (2008) proposed the use of Discrete Hidden Markov Modeling (DHMM) for the detection of small to medium size earthquakes. The seismic signals were recorded with three stations of the Bavarian Earthquake Service. The performance of their algorithm was compared with a recursive LTA/STA detector, within a continuous one-month period. The detection rate was 81% and 90% for the proposed scheme and the LTA/STA respectively, in 69 earthquakes. A drawback when using DHMMs, compared to continuous HMMs, is that time series of real-valued feature vectors have to be converted to discrete valued time series via a vector quantization step as DHMMs are only capable to evaluate discrete symbol sequences. Also, continuous HMMs are more flexible than DHMM. Beyreuther and Wassermann (2011) proposed to use Hidden semi-Markov Models (HSMMs) applied to earthquake detection and classification. HSMMs extend the double stochastic HMMs by integrating a more realistic duration of the target waveforms. When using ordinary HMMs the probability of the duration for a single part in the HMM (called state) is an exponentially decaying function in time which is an unrealistic representation for the duration of earthquake classes or speech units (Oura et al., 2008). HSMMs use the more realistic Gaussians as state duration probability distributions. State transition probabilities and distributions are estimated jointly and automatically by an Expectation Maximization (EM) algorithm. Weighted Finite State Transducers (WFSTs) were built for the classification as the standard Viterbi algorithm cannot be employed because it relies strongly on the intrinsic HMM design. Detection and classification process using this approach was extremely slow (1/2 h CPU time for 1 h data). To avoid this problem, Beyreuther et al. (2012) employed an HMM based system detection with state clustering where states containing similar Gaussians are tied together according to a given metric. By doing so, a more efficient HMM detection was achieved when compared with HSMM. This methodology was used for detection of anthropogenically induced earthquakes and earthquake classification at Mt. Merapi volcano, Indonesia. This approach also incorporates minimum state duration similarly to the HSMM technique. However, the maximum state duration is not restricted. The HMM topology is limited to self-transitions and transitions to the next state. State transition probabilities and distributions are estimated by an EM algorithm as in the previous paper. Studying volcano San Cristóbal, Nicaragua, Gutiérrez et al. (2009) proposed an HMM based automatic volcano event detection and classification system. Data of over 600 h from field survey February to March 2006 was used and the events therein were manually labeled by an expert. They report a classification accuracy of around 80%. Bicego et al. (2013) proposed a new HMM based classification technique by enhancing the HMMs with a generative embedding scheme. The generative embedding uses the models to map signals into a vector space called the generative embedding space. In such a space, any discriminative vector-based classifier (e.g. kernel-based SVMs) can be applied. Experiments were performed on pre-triggered signals recorded at Galeras Volcano in Colombia, and indicated that the proposed approach generally performs better than the standard HMM scheme.

This paper studies the Llaima volcano, one of the biggest by volume in Chile and one of the most active volcanoes in South America. The seismo-volcanic events considered in this study are: Long Period (LP), consisting in transient, volumetric signals; and, Volcano-Tectonic (VT), corresponding to ordinary earthquakes in the brittle rock within a volcanic edifice or in the crust beneath it. As will be shown later, duration distribution depends on the volcano events. As a consequence, duration models could provide useful complementary information to increase the accuracy in the volcano event detection problem. In this paper we therefore propose an HMM based volcano event detection system enhanced with duration modeling. This approach allows introducing minimum and maximum duration of HMM states. Furthermore, minimum and maximum duration constraints of the whole volcano event are also incorporated. Moreover, to reduce false events that may be detected at a frequency higher than true volcano event rate, a penalization is applied to every volcano activity detected. Finally, to reduce the computational load, the Viterbi algorithm, which is a very efficient scheme for HMM decoding (i.e. to find the optimal sequence of model and states for a given recorded signal), was adapted to incorporate the state and event duration constraints, and the penalization to new detected events. It is worth highlighting that the approach and analysis presented in this paper have not been applied to the problem of volcano event detection in the specialized literature.