IEEE Signal Processing - March 2018 - 25

by performing detection and classification. Such an architecture
is based on HMM; see, for instance, [31], [33], [40], and [51].
Once again, however, results fluctuate.
Methods proposed in the literature for the automatic classification of volcano-seismic data are engaging but are often limited
by either
■ the results, which are promising in many cases but need
improving
■ the testing process, which involves very few signals (a few
hundred or fewer than 100 samples in some cases) or only a
small number of classes compared to the variety of signals
produced by the volcano.
The main issue of today's state of the art, however, is the lack of
operational systems: very few studies mention an applicative context, and only a few have been deployed in volcanic observatories
[31], [39], [41].

Example of automatic classification
of volcano-seismic signals
Proposed architecture

ed in the section "Proposed Representation for VolcanoSeismic Signals." Eventually, each observation is
represented by a feature vector of dimension 102 that
describes the signal in time, frequency, and cepstral
domains. For each domain, features are ordered as presented in Table 1.
3) Classification. Finally, a machine-learning algorithm is
used to train a classification model by learning an automatic decision rule. In this work, we propose to use SVMs [52].
The main idea behind SVMs is to build a hyperplane in the
feature space, thereby separating the data into classes. The
hyperplane is chosen to maximize the margin, which is the
distance between the hyperplane and the support vectors
(i.e., the samples closest to the hyperplane). Features that
are used as input of the model are also displayed in a space
of higher dimension (using Gaussian kernel), where the
data might become linearly separable. Deciding on which
learning algorithm to use should not have a significant
impact on results if features have been correctly chosen
[18]. The model can then be used to predict the class of
unknown observations.

The proposed architecture is composed of three steps and will be
assessed on volcano-seicmic data recorded at the Ubinas volcano
Results
in Peru [12]. The data set is detailed in the "Results" section.
The data on which our models are tested have been recorded
1) Preprocessing. Given the amount of data to handle, which is
in Ubinas, Peru. In this study, six years of recordings are availusually very large (six years of continuous
able (including three years of eruption).
recordings in this example), a preprocessSix classes of interest have been identified
Alternative approaches
ing step is typically necessary to standardby volcanologists. Among the six classes,
to classify known data
ize the signals. In this work, we standardize
some are quite generic for volcano-seismic
without any a priori
the sampling frequency to 100 Hz: since
signals-LP, TR, EXP, and VT-but we also
knowledge on classes and considered classes that are more specific
the recording period is very long, some
sensors, or their settings, have changed and
to the Ubinas volcano, such as hybrids
detect new classes pose
need to be compensated for. A first step to
(HYB) and tornillo (TOR). HYB have chara challenging prospect
standardize the data is therefore necessary,
acteristics
of both VT and LP events with
that would allow the
and all signals have been resampled to the
a high-frequency onset followed by low
monitoring of a volcano
highest sampling frequency (100 Hz) availfrequencies. A HYB is defined as an event
that has previously not
able in the data set. This preprocessing
with characteristics of both shear-failure
been examined.
stage also includes the semiautomatic
and resonance [7]. HYB events are typical
detection and extraction of volcano-seismic
of andesitic magma and were observed at
events from the continuous recordings. STA/LTA methods
Mount Redoubt (Alaska, United States), where they were first
along with manual analysis were used to build a labeled data
described [2]. They were also observed at Soufrière Hills volset of observations of volcano-seismic events: each event was
cano, (Montserrat, United Kingdom) [53], and Mount St. Helens
manually associated to a class by experts. In our example,
(Washington, United States) [54]. TOR, also called screw events,
n
are related to resonating fluid-filled conduits or cavities. They
N = 109, 434 observations {x j [t] t =j 1} Nj = 1 are extracted and
can be considered as a specific type of LP event with a longlabeled {l j} Nj = 1, thereby forming a labeled data set. This considerable work was done by the staff at the Ubinas observatoduration coda composed of harmonic oscillations. They were
ry. Each observation is then normalized in energy and limited
observed at Galeras volcano (Colombia) before several erupin time. The idea here is to build a classifier based on the
tions in 1992 and 1993 [55].
observed shapes rather than on their energies, as it is usually
On Ubinas, a total of 109,434 seismic events have been
the case in the state of the art. Such models would give access
manually labeled into one of those six classes and truncated
to a more precise analysis and prediction of volcanic erupinto chunks of a maximum of 5 min. We would like to undertions and are therefore needed.
line the uniqueness of this data set in terms of duration of the
2) Features extraction. As reported in the section "Signals
period of observation and diversity of the activity monitored.
Representation," the purpose of extracting features is to repAll results are presented using an SVM with a radial basis
resent the data in a space where an automatic decision rule
function kernel as the learning algorithm, with parameters
can be established. Features used for this study were presentC SVM = 10 and c = 0.01 chosen to optimize the results on a
IEEE SIgnal ProcESSIng MagazInE

March 2018

Table of Contents for the Digital Edition of IEEE Signal Processing - March 2018