DISCUSSION OF “CLUSTERING ON DISSIMILARITY REPRESENTATIONS FOR DETECTING MISLABELLEDSEISMIC SIGNALS AT NEVADO DEL RUIZ VOLCANO” BY MAURICIO OROZCO-ALZATE, AND CÉSAR GERMÁNCASTELLANOS-DOMÍNGUEZ

Mehmet

Description

The authors are to be congratulated for a systematic investigation of the accurate and non subjective classifying approach in seismic research. The authors have conducted several clustering algorithms to the seismic
event records from Volcanological and Seismological Observatory at Manizales. Their objective was to improve the grouping of seismic data (i.e., volcano-tectonic earthquakes, long-period earthquakes and icequakes) digitized at 100.16 Hz sampling frequency. Their study seems adding new approach to their previous work of Langer et al. (2006) who applied different classification techniques to seismic data.

Sujets

• Earthquake

Informations

• Publié par: erevistas
• Publié le: 01 janvier 2008
• Nombre de lectures: 17
• Langue: EnglishEnglish

Extrait

ppp
Composite 133 lpi at 45 degrees
EARTH SCIENCES
RESEARCH JOURNAL
Earth Sci. Res. J. Vol. 12, No. 2 (December 2008): 265-268
DISCUSSION OF “CLUSTERING ON DISSIMILARITY
REPRESENTATIONS FOR DETECTING MISLABELLED
SEISMIC SIGNALS AT NEVADO DEL RUIZ VOLCANO” BY
MAURICIO OROZCO-ALZATE, AND CÉSAR GERMÁN
CASTELLANOS-DOMÍNGUEZ
1 2 3Mehmet C. Demirel , Ercan Kahya and Diego Rivera
1 Ph.D. Student, Department of Water Engineering and Management, University of Twente,
PO Box 217, 7500 AE Enschede, The Netherlands. E-mail: [e-mail]
2 Civil Engineering Department, American University of Sharjah,
PO Box 26666, Sharjah, United Arab Emirates (corresponding author).
E-mail: [e-mail]
3 Professor, Department of Water Resources, University of Concepcion, Vicente Méndez 595,
Chillán, Chile. E-mail: [e-mail]
which are (i) selection of variables, (ii) selec-The authors are to be congratulated for a systematic in-
tion of standardization technique (if necessary),vestigation of the accurate and non subjective classify-
(iii) dissimilarity metric, (iv) selection of an ap-ing approach in seismic research. The authors have
propriate method, and (v) test of stability orconducted several clustering algorithms to the seismic
validation (Demirel 2004; Everitt 1993; Greenevent records from Volcanological and Seismological
et al. 1990). These steps are difficult to distin-Observatory at Manizales. Their objective was to im-
guish in the present study by Orozco-Alzateprove the grouping of seismic data (i.e., volcano-tec-
and Castellanos-Domínguez (2007). The users`tonic earthquakes, long-period earthquakes and
experience and preferences in these steps mayicequakes) digitized at 100.16 Hz sampling frequency.
radically affect the resultant cluster structures.Their study seems adding new approach to their previ-
For instance, the selected distance metric wasous work of Langer et al. (2006) who applied different
not clearly mentioned in the text. Did the nota-classification techniques to seismic data.
tion refer to the correlation coefficient be-kl
The discussers have the following suggestions to
tween entities k and l? How many stations were
improve the author’s investigation on Ruiz volcano
selected near the Olleta crater and the glacier at
data and to be a guide for similar future studies.
Nevado del Ruiz volcanic complex? Were
1. There are five empirical steps that should be there any scale issues in the dataset which may
followed in the application of cluster analysis perturb the dissimilarity matrices? Henceforth
265
NOV 10-DISCUSSION.prn
D:\GEOCIENCIAS V-12-2-DIC 2008\GEOCIENCIAS-VOL 12-2 DIC.vp
martes, 10 de noviembre de 2009 20:50:03ppp
Composite 133 lpi at 45 degrees
MEHMET C. DEMIREL, ERCAN KAHYA AND DIEGO RIVERA
the span of the data and some statistical infor- unclear points should be justified in the manu-
mation on the data structure should be ex- script.
plained for easy follow of readers and to avoid
5. The clustering results were not given in the
the aforementioned questions. Standardization
text. The labels of clusters and statistics (i.e.,
priori to analysis phase is necessary when the
variance, mean) of each cluster should be sum-
scale differences emerged in a dataset (Demirel
marized in the result section. Only the averaged
et al. 2008; Everitt 1993; Gnanadesikan et al.
numbers of mismatches between class labels
1995; Milligan and Cooper 1988).
were presented but this was not adequate for
the readers to have appropriate insights regard-2. The authors applied several algorithms on their
ing the main objective of the study. Since thedata and reasoned the following statement: “the
article is about signal clustering, it would havelack of a single appropriate clustering algo-
been very illustrative to put 2 figures: Onerithm”. However most of the algorithms were
graph including 3 representative signals (e.g.already tested in the literature and the relevant
Langer et al., 200) and one figure representingshortcomings are given in many text books
the topological structure of the clusters, e.g.
(Bacher 2002; Everitt 1993). Single linkage
dendrogram. Both figures allow analyzing in
produces chain type cluster which is not be de-
an intuitive way dissimilarities among signals.
sirable for many applications, and complete
linkage may create small and compact clusters 6. At page 135: the authors mentioned that “even
(Demirel 2004; Everitt 1993). On the other though the number of cluster is fixed, single
hand the Wards method emerged to make more linkage and average linkage find second and
successive clusters with small inner variance. third clusters of a few objects only”. The single
Hence it is herein suggested to use the Wards linkage and average linkage methods are in the
group of unsupervised clustering techniquesmethod with the squared Euclidean metric to
which has no priori knowledge on number ofget more distinct clusters in future investiga-
clusters as partitioning methods; therefore, ations.
justification should have been indicated for that
3. In the context of text indications in notation
matter (Demirel and Kahya 2007; Kahya et al.
wise, at page 133: the notation “DC” was not
2007).
explained in the text. At page 133: D(T,T) des-
7. As was noted by Morlet et al. (1982), seismicignates to distance/dissimilarity measure; how-
signal does vary in amplitude, shape, frequencyever, the notation “d” was used for the same
and phase, versus propagation time. Therefore,purpose in table 1. It is important to maintain
for clustering it is necessary to analyze signal’sconsistent use of notations for the dissimilarity
frequency content, as well as to localize in timemeasure throughout the text.
changes in both, frequency and amplitude. For
4. The mismatches in labeling were counted for this task, Wavelet transform is a joint time-fre-
the performance comparison and number of quency signal representation that can give the
runs was given as 10. The author also men- frequency content of the signal at a particular
tioned that “Hierarchical methods report the instant of time by filtering (Sheikholeslami et
same number of mismatches over the runs”. It al., 1998). It is well suited for signal whose fre-
should be noted that cluster structure in the hi- quencies change with time, but also for signal
erarchical methods do not differ in any run as containing noise and transients (Rouyer et al.,
the steps in dissimilarity calculations and clus- 2008). Also, its multi-resolution property can
ter delineation has concrete algorithm; thus, it help detecting the clusters at different levels of
is herein encouraged that issues similar to these accuracy (Sheikholeslami et al., 1998). We
266
NOV 10-DISCUSSION.prn
D:\GEOCIENCIAS V-12-2-DIC 2008\GEOCIENCIAS-VOL 12-2 DIC.vp
martes, 10 de noviembre de 2009 20:50:04ppp
Composite 133 lpi at 45 degrees
DISCUSSION OF “CLUSTERING ON DISSIMILARITY REPRESENTATIONS FOR DETECTING MISLABELLED SEISMIC
SIGNALS AT NEVADO DEL RUIZ VOLCANO” BY MAURICIO OROZCO-ALZATE, AND CÉSAR GERMÁN
CASTELLANOS-DOMÍNGUEZ
propose for further research to apply this tech- in k-means clustering”. Journal of Classifica-
nique to Ruiz volcano data. A good reference tion, 7(2), 271-285.
are Kumar and Foufoula-Georgiou (1997) and
Kahya, E., Demirel, M. C., and Piechota, T. C. “Spa-
Torrence and Compo (1998) for methods. In-
tial grouping of annual streamflow patterns in
deed, Arciniega-Ceballos et al. (2008) applied
Turkey ” 27th AGU Hydrology Days, Fort Col-
bandpass filters before clustering in seismic
lins, Colorado, 169-176.
data and Rouyer et al., (2008) applied a wave-
let-based clustering technique. Kumar, P. and Foufoula-Georgiou, E. (1997). Wave-
let analysis for geophysical applications. Re-
views of Geophysics, 35:385–412.
References
Langer, H., Falsaperla, S., Powell, T., and Thomp-
Arciniega-Ceballos A, Chouet B, Dawson Ph. and G
son, G. (2006). “Automatic classification and
Asch (2008) Broadband seismic measurements
a-posteriori analysis of seismic event identifica-
of degassing activity asociated with lava effu-
tion at Soufri?re Hills volcano, Montserrat.”
sion at Popocatépetl Volcano, Mexico. Journal
Journal of Volcanology and Geothermal Re-
of Volcanology and Geothermal Research, 170:
search, 153(1-2), 1-10.
12-23.
Milligan, G. W., and Cooper, M. C. (1988). “A studyBacher, J. (2002). “Cluster Analysis.” Lecture Notes,
of standardization of variables in cluster analy-Nuremberg.
sis.” Journal of Classification, 5(2), 181-204.
Demirel, M. C. (2004). “Cluster Analysis of
Morlet, J., Arens, G., Fourgeau, E. and Glard, D.Streamflow Data over Turkey,” Istanbul Techni-
(1982) Wave propagation and sampling the-cal University, Istanbul.
ory-Part I: Complex signal and scattering in
Demirel, M. C., and Kahya, E. “Hydrological determi-
multilayered media. Geophysics 47, 203-221.
nation of hierarchical clustering scheme by using
small experimental matrix.” 27th AGU Hydrology Orozco-Alzate, M., and Castellanos-Domínguez, C.
Days, Fort Collins, Colorado, 161-168. G