Options
Predictive analysis of the seismicity level at Campi Flegrei volcano using a data-driven approach
Author(s)
Language
English
Obiettivo Specifico
5V. Sorveglianza vulcanica ed emergenze
Publisher
Springer International Publishing Switzerland 2014
Status
Published
Pages Number
133-144
Refereed
Yes
Issued date
2014
Alternative Location
ISBN
978-3-319-04128-5
Abstract
This work aims to provide a short-term tool to estimate the possible
trend of the seismicity level in the area of Campi Flegrei (southern Italy) for
Civil Protection purposes. During the last relevant period of seismic activity,
between 1982 and 1984, an uplift of the ground (bradyseism) of more than 1.5
m occurred. It was accompanied by more than 16,000 earthquakes up to
magnitude 4.2 which forced the civil authorities to order the evacuation of
about 40,000 people from Pozzuoli town for several months. Scientific studies
evidenced a temporal correlation between these geophysical phenomena. This
has led us to consider a data-driven approach to obtain a forecast of the
seismicity level for this area. In particular, a technique based on a Multilayer
Perceptron (MLP) network has been used for this intent. Neural networks are
data processing mechanisms capable of relating input data with output ones
without any prior correlation model but only using empirical evidences
obtained from the analysis of available data. The proposed method has been
tested on a set of seismic and deformation data acquired between 1983 and
1985 and then including the data of the aforementioned crisis which affected
the Campi Flegrei. Once defined the seismicity levels on the basis of the
maximum magnitude recorded within a week, three MLP networks were
implemented with respectively 2, 3 and 4 output classes. The first network (2
classes) provides only an indication about the possible occurrence of
earthquakes felt by people (with magnitude higher than 1.7), while the
remaining nets (3 and 4 classes) give also a rough suggestion of their intensity.
Furthermore, for these last two networks one of the output classes allows to
obtain a forecast about the possible occurrence of strong potentially damaging
earthquakes with magnitude higher than 3.5. Each network has been trained on
a fixed interval and then tested for the forecast on the subsequent period. The
results show that the performance decreases as a function of the complexity of
the examined task that is the number of covered classes. However, the obtained
results are very promising, for which the proposed system deserves further
studies since it could be of support to the Civil Protection operations in the case
of possible future crises.
trend of the seismicity level in the area of Campi Flegrei (southern Italy) for
Civil Protection purposes. During the last relevant period of seismic activity,
between 1982 and 1984, an uplift of the ground (bradyseism) of more than 1.5
m occurred. It was accompanied by more than 16,000 earthquakes up to
magnitude 4.2 which forced the civil authorities to order the evacuation of
about 40,000 people from Pozzuoli town for several months. Scientific studies
evidenced a temporal correlation between these geophysical phenomena. This
has led us to consider a data-driven approach to obtain a forecast of the
seismicity level for this area. In particular, a technique based on a Multilayer
Perceptron (MLP) network has been used for this intent. Neural networks are
data processing mechanisms capable of relating input data with output ones
without any prior correlation model but only using empirical evidences
obtained from the analysis of available data. The proposed method has been
tested on a set of seismic and deformation data acquired between 1983 and
1985 and then including the data of the aforementioned crisis which affected
the Campi Flegrei. Once defined the seismicity levels on the basis of the
maximum magnitude recorded within a week, three MLP networks were
implemented with respectively 2, 3 and 4 output classes. The first network (2
classes) provides only an indication about the possible occurrence of
earthquakes felt by people (with magnitude higher than 1.7), while the
remaining nets (3 and 4 classes) give also a rough suggestion of their intensity.
Furthermore, for these last two networks one of the output classes allows to
obtain a forecast about the possible occurrence of strong potentially damaging
earthquakes with magnitude higher than 3.5. Each network has been trained on
a fixed interval and then tested for the forecast on the subsequent period. The
results show that the performance decreases as a function of the complexity of
the examined task that is the number of covered classes. However, the obtained
results are very promising, for which the proposed system deserves further
studies since it could be of support to the Civil Protection operations in the case
of possible future crises.
Type
book chapter
File(s)
No Thumbnail Available
Name
WIRN 2013_last_published.pdf
Size
404.97 KB
Format
Adobe PDF
Checksum (MD5)
d27a955767cfeeded1f2ed626ab2835d