IEEE Signal Processing - March 2018 - 27

With the increasing
number of volcanoseismic
stations that record
continuous signals,
machine-learning
methods have
become essential.

extracted from only one of the three domains of representation: 86.1%, 83.0%, and
79.4% for temporal, spectral, or cepstral
features, respectively. The choice of features
describing the signal general shape in various domains is, therefore, well-founded.

physical reason such as a volcanic phase
leading to a change in the Ubinas structure
and/or an internal activity.

Third result: Continuous analysis

To work with continuous analysis on the six
years of recordings, we propose to build an
evolving model where a new classifier is
trained each month. This kind of specification is very important because it can be
seen as a real-time analysis. To reach this analysis, we propose to train the models with a maximum of 800 past observations for each class. For very frequent classes such as LP,
only a few days are necessary to gather 800 observations: new
observations are therefore considered for each newly trained
model. For less frequent classes such as TR, few months are
needed to gather the 800 observations: the first models will be
then trained on fewer than 800 examples, but once this limit
is reached, only the 800 most recent observations will be used
for training. Finally, for very sporadic classes such as HYB, the
whole data set does not contain 800 observations; therefore,
each monthly model will be trained on the available observations. For the first months, the number of observations can be
very low (# 10).
Global accuracy for this configuration reaches 80.3% of
correct classification. The evolution of accuracy over the six
years of recordings can be seen in Figure 4. It is interesting to interpret the performance decay compared with crossvalidation results: first, and seen in the section "Result: Data
Evolution," observations within a same class do evolve with time.

Second result: Data evolution

IEEE SIgnal ProcESSIng MagazInE

March 2018

7/2011

2/2011

9/2010

4/2010

11/2009

5/2009

12/2008

7/2008

2/2008

9/2007

4/2007

11/2006

6/2006

Number of Observations

Accuracy (%)

Depending on the volcanic activity and the
volcano structure which could significantly
evolve with time, observations within the same class can have
a significant variability and change over time. To estimate this
evolution, we propose to train a model on the first 800 observations of each class (fewer, if 800 are not available) and to
test it on the six years of recordings where the mean accuracy
reaches 59.7%. More specifically, we propose to study the
evolution of the dominant class: LP, with 95,094 events. We
point out that the first 800 LP observations are recorded in
few days only (June 2006). LP mean accuracy is 61%, and
its evolution with time is presented in Figure 3. This plot is
interesting: we can see a clear change in the trend around
May-June 2007, which can be related to a change in the
observations compared to the samples used for training. The
accuracy drops from more than 95% in average to less than
10%. A drop in the number of LP can be also noticed after
August 2007.
After communicating those results to the experts in
charge of the Ubinas monitoring (from Instituto Geofísico
del Perú, Arequipa), they analyzed the data set to interpret
the drop in accuracy. This additional work reveals significant
changes in the data set labels, mostly
due to the a posteriori analysis of
signals and the current better understanding of the volcano-seismic sig100
natures by the experts with respect to
80
2007. First, the number of LP events
in the new classification system is ten
60
times fewer than in the original data
40
set (from several hundreds per week
20
to several dozen only). Second, many
observations originally labeled as LP
0
are now considered to be VT events after
7,000
May 2007. This result is of capital impor5,000
tance, since it shows that the system we
propose was able to detect dissimilari3,000
ties between signals and incongruence
in the manual labeling. Regarding the
1,000
physical interpretation, LP come from
0
superficial sources while VT are emitted by much deeper sources. The seismic
activity therefore becomes deeper from
Time
the end of May, with the increase in VT
probably meaning a recharge of the system in depth, which leads to a new volca- FigUre 3. The accuracy evolution of LP classification from 2006 to 2011. The model is trained on
nic phase of activity. Figure 3 is a proof the first 800 LP observations (all recorded in June 2006) and is then used to analyze the six years
of change starting in May 2007, with a of recordings.

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