IEEE Signal Processing - March 2018 - 28

By using a newly trained model every month, we manage to follow and learn the evolutions of the classes. Global performances
are indeed much higher in this configuration than in the previous one. However, the evolution of the observations is still visible
with sudden accuracy drops in the plot: for example, in September 2007, the accuracy drops from 95% to 38% and increases
again to 98% in October. With the sudden change in the signals'

100

60
40
20
0

Conclusions

8,000
6,000

7/2011

2/2011

9/2010

4/2010

11/2009

5/2009

12/2008

7/2008

2/2008

9/2007

4/2007

2,000
0

11/2006

4,000

6/2006

Number of Observations

Accuracy (%)

shapes, observations from September are predicted with a
model trained on data that are no longer representative of the
class. As soon as the model has been updated (in October), however, accuracy increases again. This phenomenon is observed
over the six years and is notably visible between stable phases
in the volcanic activity. Second, sporadic classes such as HYB
present a difficulty since very few observations are available
(only six for the first model, including
this class): accuracy levels are then notably low (see Figure 5). With an increasing number of observations used in
training, however (blue plot displaying
cumulative number of observations),
the accuracy level steadily increases
to very good performances (100% in
November 2011).

Time

FigUre 4. The mean accuracy evolution from 2006 to 2011. A new model (SVM) is trained every
month on a maximum of 800 observations per class.

80
60
40
20
0

Cumulative
Number of Observations

500
400
300
200

7/2011

2/2011

9/2010

4/2010

11/2009

5/2009

12/2008

7/2008

2/2008

9/2007

4/2007

100
11/2006

120
100
80
60
40
20
0

6/2006

Number of Observations

Accuracy (%)

100

Time

FigUre 5. The accuracy evolution of HYB classification from 2006 to 2011. A new model (SVM) is
trained every month on the past data (the maximum limit of 800 observations is never reached, as the
class is particularly sporadic).
28

IEEE Signal Processing Magazine

March 2018

We would like to emphasize the importance of monitoring the volcanic hazard
and of evaluating the associated risks.
With the increasing number of volcanoseismic stations that record continuous
signals, machine-learning methods have
become essential. Tools for the automatic
analysis of volcano-seismic signals have
started to appear but are often limited,
and operational systems are still very
scarce. In this article, we reviewed the
existing methods and proposed an efficient procedure for the automatic classification of volcano-seismic events. The
model is tested using cross validation on
70,856 observations and reaches 92.2%
of accuracy. Furthermore, we investigated the generalization capacities by building models in experimental conditions
close to real-time analysis and displayed
the need for more than a couple hundred
observations to represent the variety
within a considered class. Finally, such
tools can also be used to monitor the evolution of intraclass observations: we illustrated the process by analyzing six years
of volcano-seismic recordings.
Prospects for such tools are, therefore,
more than promising, and machinelearning architectures can be used to
automatize the analysis of volcanoseismic data. Challenges for the future
include the number of observations
needed to classify a new class: in many
applied cases, hundreds of signals are
not available. The lowering of this constraint is a straightforward prospect to
diffuse the use of such techniques in

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