analysis of dynamics of vulcanian activity of ubinas ... · seismic data and ground deformation...

Journal of Volcanology and Geothermal Research 270 (2014) 35–52

Contents lists available at ScienceDirect

Journal of Volcanology and Geothermal Research

j ourna l homepage: www.e lsev ie r .com/ locate / jvo lgeores

Analysis of dynamics of vulcanian activity of Ubinas volcano,using multicomponent seismic antennas

L.A. Inza a,b,d,⁎, J.P. Métaxian b, J.I. Mars a, C.J. Bean c, G.S. O'Brien c, O. Macedo d, D. Zandomeneghi b

a GIPSA-LAB/DIS/UMR 5216 Institut Polytechnique de Grenoble, Franceb Institut des Sciences de la Terre IRD R219 CNRS, Université de Savoie, Campus Scientifique, 73376 Le Bourget du Lac cedex, Francec School of Geological Sciences, University College Dublin, Belfield, Dublin 4, Irelandd Instituto Geofísico del Perú, Lima, Peru

⁎ Corresponding author at: GIPSA-LAB/DIS/UMR 52Grenoble, France.

E-mail address: [email protected] (L.A. Inza).

0377-0273/$ – see front matter © 2013 Elsevier B.V. All rihttp://dx.doi.org/10.1016/j.jvolgeores.2013.11.008

a b s t r a c t
a r t i c l e i n f o
Article history:Received 30 May 2013Accepted 19 November 2013Available online 27 November 2013

Keywords:Volcano seismic monitoringSeismic arrayVulcanian activityUbinas

A series of 16 vulcanian explosions occurred at Ubinas volcano betweenMay 24 and June 14, 2009. The intervalsbetween explosions were from 2.1 h to more than 6 days (mean interval, 33 h). Considering only the first nineexplosions, the average time interval was 7.8 h. Most of the explosions occurred after a short time interval(b8 h) and had low energy, which suggests that the refilling time was not sufficient for large accumulation ofgas. A tremor episode followed 75% of the explosions,which coincidedwith pulses of ash emission. The durationsof the tremors following the explosionswere longer for the two highest energy explosions. To better understandthe physical processes associated with these eruptive events, we localized the sources of explosions using twoseismic antennas that were composed of three-component 10 and 12 sensors. We used the high-resolutionMUSIC-3C algorithm to estimate the slowness vector for the first waves that composed the explosion signalsrecorded by the two antennas assuming propagation in a homogeneous medium. The initial part of theexplosions was dominated by two frequencies, at 1.1 Hz and 1.5 Hz, for which we identified two separatedsources located at 4810 m and 3890 m +/− 390 altitude, respectively. The position of these two sources wasthe same for the full 16 explosions. This implies the reproduction of similar mechanisms in the conduit. Basedon the eruptive mechanisms proposed for other volcanoes of the same type, we interpret the position of thesetwo sources as the limits of the conduit portion that was involved in the fragmentation process. Seismic dataand ground deformation recorded simultaneously less than 2 km from the crater showed a decompressionmovement 2 s prior to each explosion. This movement can be interpreted as gas leakage at the level of the capbefore its destruction. The pressure drop generated in the conduit could be the cause of the fragmentationprocess that propagated deeper. Based on these observations, we interpret the position of the highest sourceas the part of the conduit under the cap, and the deeper source as the limit of the fragmentation zone.

© 2013 Elsevier B.V. All rights reserved.

1. Introduction

The physical mechanisms behind vulcanian events remain achallenge in our community, with andesitic volcanoes characterizedby violent and unpredictable eruptions. It is known that the vulcanianphenomenon is controlled by the physical and chemical propertiesof the magma. Vulcanian episodes can be generated by magma withintermediate properties between basaltic and rhyolitic, which comesfrom the deepest parts of the Earth, and which is less dense than thesurrounding rock (Sparks, 2000).

Fragmentation of viscous magma by brittle failure is thought to beresponsible for explosions in silicic volcanoes (Alidibirov and Dingwell,2000). Melnik and Sparks (2002) have modeled unsteady conduit flow

16 Institut Polytechnique de

ghts reserved.

in explosive eruptions after unloading through dome collapse. Thesemodels have been applied to the episodes of explosive activity thatoccurred shortly after dome collapse at Soufrière Hills volcano. Theunloading triggers gas exsolution and magma rising in the conduit,increasing its internal pressure. An explosion is triggered when theinternal pressure equals the tensile strength of themagma. Overpressureis responsible for magma fragmentation generating gas-particledispersion, which propagates to the exit of the conduit to forma volcanic column in the atmosphere (Melnik and Sparks, 2002).Other studies also suggest that fragmentation occurs when a criticaloverpressure or critical elongation strain rate of magma is reached(Alidibirov and Dingwell, 1996; Papale, 1999; Alidibirov and Dingwell,2000).

Geophysical measurements have been performed to study this phe-nomenon for andesitic volcanoes. Iguchi et al. (2008) focused on seismicobservations and ground deformation. Thus, a common sequence ofphenomena associated with vulcanian eruptions, in particular occurredwith inflation and deflation of the edifice, has been compared at three

http://crossmark.crossref.org/dialog/?doi=10.1016/j.jvolgeores.2013.11.008&domain=f

http://dx.doi.org/10.1016/j.jvolgeores.2013.11.008

mailto:[email protected]

http://dx.doi.org/10.1016/j.jvolgeores.2013.11.008

http://www.sciencedirect.com/science/journal/03770273

UB14800

5400

5000

4600

4800

4600

5200

5000

5200

4800

5400

5400

5200

5000

4800

5200

UB4

WUBI

NUBI

UB2

UB3

Fig. 1.Mapof theUbinas crater. Thewhite dots show locations of the three-component seismometers of theWUBI andNUBI arrays, and thewhite triangles show the Instituto Geofísico delPerú seismic network.

36 L.A. Inza et al. / Journal of Volcanology and Geothermal Research 270 (2014) 35–52

andesitic volcanoes (Sakurajima, Japan; Suwanosejima, Japan; Semeru,Indonesia).

Another approach was analyzed by Yokoo et al. (2009), who usedboth an infrasound network and video image time-series observationsof the vulcanian eruption that occurred at Sakurajima Volcano (Japan)in January 2002. These observations suggested that the volumetricincrease of the gas pocket caused a swelling of the surface of the craterbottom, and its subsequent failure. When the expansion velocityexceeded a threshold level, themain impulsive compression phase radi-ated with a high velocity through the sudden release of the pressurizedgases. This volume change indicated that the vertical displacement ofthe swelling ground was of the order of 1 m, assuming that the radiusof the lava plug was approximately 10 m.

Druitt et al. (2002) surveyed the eruption of the Soufriere Hillsvolcano in Monserrat in 1998, and reporting on 88 vulcanian episodesthat revealed the vulcanianmechanism. This mechanism can be brokendown into the following scenario. The explosion starts when the pres-sure in the conduit goes over a threshold trigger of the cap of degassedcrystal-richmagma. A fragmentationwave goes down the conduit into aregion of pressurized magma with an approximate velocity of 50 m/s,which results in an upward speed of around 140 m/s.

On the other hand, small-aperture seismic arrays have been usefulto locate seismic sources in volcanic unrest, as seismic waveformshave a lack of clear body-wave phase arrivals. Emergent onset in long-period and tremor seismic waves makes it extremely difficult to solvea source localization from classical hypocenter methods based onphase picking and calculated travel-times. Therefore, different methodsof source localization have been applied to array data recorded fromvolcanic unrest (Saccorotti and Del Pezzo, 2000; Almendros et al.,2001a,b; Métaxian et al., 2002; La Rocca et al., 2004; Di Lieto et al.,2007; Inza et al., 2011; O'Brien et al., 2011). Also, high resolutiontechniques for multicomponent array data have been developed that

can be applied to the locating of seismic sources (Miron et al., 2005;Paulus and Mars, 2006).

A field experiment was carried out fromMay to July 2009 at UbinasVolcano (Peru) that was carried out by a research team from the FrenchInstitut de Recherche pour le Développment, University College Dublin,Ireland (Volume project) and the Instituto Geofísico del Perú, with twosmall-aperture seismic arrays that were composed of three-componentseismometers deployed on two flanks of the Ubinas Volcano.

Inza et al. (2011) presented a source localizationmethodMUSIC-3C,based on the use of 3C seismic arrays. MUSIC-3C provided realisticestimates of the depth of volcanic sources, performed on syntheticdata. Also, an explosion earthquake and a LP event recorded duringthe field experiment were located by using MUSIC-3C. The locations ofthese events were at an altitude of 4200 m for the explosion and2240 m for the LP event respectively. The explosion earthquake andthe LP event were interpreted as resulting, respectively, from a frag-mentation process and shear-fracturing of magma at the conduit walls.

In the following three sections, we focus our study towards theapplication of seismic arrays to locate and interpret the vulcanian eventsrecorded at Ubinas volcano. In Section 2 it described the data recordedduring the field experiment. The source localization analysis inSection 3 will unfold the source localization method MUSIC-3Cby using real and synthetic data set. Finally, in Section 4, results willbe discussed taking into account the methodology and observationaspects.

1.1. Ubinas volcano

Ubinas volcano (16° 22′ S, 70° 54′ W; altitude, 5672 m) began toerupt on March 25, 2006, after nearly 40 years of quiescence. Situatedin the Central Volcanic Zone (CVZ, southern Peru), Ubinas Volcano isan active andesitic stratovolcano that is truncated in the upper part by

Kx 1/km

Ky

1/km

−15 −10 −5 0 5 10 15

−15

−10

−5

0

5

10

15

Fig. 2. Array response function normalized to one, of the WUBI array.

37L.A. Inza et al. / Journal of Volcanology and Geothermal Research 270 (2014) 35–52

a caldera of 600 m in diameter (De Silva and Francis, 1991). The calderafloor is a flat area that lies at an altitude of approximately 5100 m. Theactive crater is situated in the southern section; the bottom is 300 munder the caldera floor (Fig. 1). Ubinas is considered to have been themost active Peruvian volcano during the last 500 years, which hasthreatened 3500 people who live on the edge of the Ubinas Valley(Rivera et al., 2010). Arequipa Airport is situated 60 km east of theUbinas volcano, and it has had to be closed several times since 2006,due to ash emissions. Under the 6th EU Framework Programme projectknown as VOLUME, the Instituto Geofísico del Perú with the coopera-tion of the Institut de Recherche pour le Développment (France) startedseismic monitoring of Ubinas volcano to understand the activity associ-atedwith this eruptive sequence. A network of four digital 1-Hz stationswith a radio telemetry systemhas been operating there since 2006,withthe data transmitted to Arequipa Instituto Geofísico del Perú observatory.At the time of this study, the eruption was characterized by almostpermanent ash emissions. Two main types of degasing were observed:1. Exhalations that rose a few hundred meters above the crater rim;and 2. Plumes that were produced by explosions that could reach10 km above sea level (Rivera et al., 2010), and which were criticalto aircraft safety. This activity is thought to be related to a magmaticplug that is positioned at the bottom of the southern part of the calderawall (Macedo et al., 2009).

2. Data

2.1. Data collection

From May to July 2009, two small-aperture seismic arrays weredeployed at Ubinas Volcano: on the north part (the NUBI antenna)and the west part (the WUBI antenna) (Fig. 1). The NUBI antenna wascomposed of 10 instruments: eight Guralp-6TD and two Guralp-3ESPseismometers, while the WUBI antenna consisted of six Guralp-6TDseismometers and six Titan-Neomax Agecodagis instruments. Eachstation included a seismic three-component sensor and a GPS receiver,and was set up to continuously record at 100 samples/s. The twocross-shaped antennas were installed on slightly sloping surfaces,according to the topography. The slopes between the farthest sensorson each antenna were 7% and 30% for NUBI and WUBI, respectively.The distances between the seismometers at each antenna were setto approximately 50 m. The centers of the NUBI antenna (altitude4632 vm) and WUBI antenna (altitude 4732 m) were 3750 m and2567 m away, respectively, from the northern border of the crater(Fig. 1).

Themonitoring networkwasmanaged by the Instituto Geofísico delPerú and provided uswith additional data. This network was composedof four seismic stations (UB1, UB2, UB3, UB4) that were distributed onthe flanks of the volcano (Fig. 1). UB2 was equipped with a broadbandvertical sensor, and the three other stations had short-period sensors.In addition, UB2 and UB4 were equipped with a bi-axial tiltmeter with0.1 micro-radian resolution. The seismic and tiltmeter instruments ofUB2, which was located approximately 1 km north to the Ubinas crater(Fig. 1), were installed at a depth of 50 cm,with a concrete base on solidvolcanic rock. Seismic and tilt data were recorded by the same digitizer,a Reftek 130,with a sampling rate of 50 Hz. Data were recorded contin-uously fromMay 24 to July 14, 2009, and they were converted into SACformat. The tilt instrument was a bubble-type two-axis sensor. The ori-entation of one of the components, the Y-axis, was to the north, almostradial to the active crater. The X-axis component was oriented to theeast, orthogonal to the Y-axis. A positive tilt on the north componentindicated inflation of the crater area with respect to the tiltmeter. Apositive tilt on the east component indicated a tilt down towards thewest (Ferro et al., 2011). Before going further, it is necessary to describethe array response function.

2.2. Array response function

The purpose of the array response function was to evaluate thesensitivity and resolution of seismic antenna recording wavefrontswith different wavelengths. Classically, an antenna operates as a wave-number filter, or wavelength filter. Assuming that a plane wave propa-gating through the antenna with a back-azimuth θ and an incidenceangle ϕ, the wavenumber vector (k) for this wave, in Cartesian coordi-nates, has the form of Eq. (1) (Rost and Thomas, 2002):

k ¼ 2πλ

cos θð Þsin ϕð Þ; sin θð Þsin ϕð Þ; cos ϕð Þð Þ ¼ 2π fu0 ð1Þ

Where λ is thewavelength, f is the frequency, andu0 is the slownessvector. For an antenna composed of N seismometers, the wavefrontw1

recorded by the nth seismometer with the relative position of rn, can beexpressed as Eq. (2):

wn tð Þ ¼ w1 t−rn � u0ð Þ ð2Þ

Consequently, the optimal beam of the N sensors in the antenna forany slowness u relative to slowness u0 can be represented by Eq. (3):

b tð Þ ¼ 1N

XNj¼1

wj t−rj � u−u0ð Þ� �

ð3Þ

The energy E of this beam can be computed by summing the squareamplitude of b(t). According to the Parseval theorem and Fourier shiftproperties, E can be defined as a function of waveforms k, wherek = f·u, and can be written in the frequency domain as:

E k−k0ð Þ ¼Z þ∞

−∞w fð ÞÞj j2 1

N

XNj¼1

exp i � 2π k−k0½ � � rnð Þ��

��2

df ð4Þ

Consequently, the array response function A(k) is expressed as:

A k−k0ð Þj j2 ¼ 1N

XNj¼1

exp i � 2π k−k0½ � � rnð Þ��

��2

ð5Þ

In the case of surveying the array response function for the horizontalplane, the wavenumber should equal k = (kx,ky). According to thespatial Nyquist wavenumber, a minimum of two grid points per wave-length are theoretically enough for the spatial sampling of thewavefront.

-4

0

4

8

12

logR

SE

M

24 31 7May June

1

2

3

4

5

6

7 8 910 11

-4

0

4

8

12

logR

SE

M

17 24 1June

12 1314 15

16

-4

0

4

8

12

logR

SE

M

21 28 5June July

-4

0

4

8

12

logR

SE

M

15 2July

Fig. 3. Log of the real-time seismic energy measurement (RSEM) computation, with 5-min sliding time windows, during the 2009 experiment. The black inverted triangles indicate theexplosions at Ubinas volcano, according to Table 1. The horizontal dashed lines match the tremor event duration, and the black squares relate to earthquakes out of the volcano.


For the WUBI antenna, with an X-north aperture of 321 m and a Y-eastaperture of 315 m, the array response function for this antenna is shownin Fig. 2. This gives the smallest wavenumber of around 0.5 km−1, andthe maximum wavenumber without aliasing of around 18 km−1. Withthese characteristics, it is possible to discriminate thewavelengths asso-ciatedwith Ubinas Volcano. For example, for an explosive quake record-ed by the WUBI antenna, the wavenumbers were identified as 5 km−1

(f = 1.1 Hz; wave velocity, 1.4 km/s), or 6 km−1 (f = 1.5 Hz; wave

Table 1Source localization of the eight synthetic sources in the Ubinas model.

Source 1 2 3 4 5 6 7 8Depth 4972 4912 4532 3972 3912 3472 2472 1972

velocity, 1.5 km/s), which were in good agreement with the antennaresolution.

2.3. Seismic and tilt observations

Two preprocessing controls were carried out on the raw data thatwere recorded in this field study, one to deconvolve the instrumentresponse, and another to monitor the accuracy of the time-stamp onthe seismograms, as each sensor was synchronized from its own GPSclock. Time synchronization was crucial for correlating the events inthe array dataset analysis. For this purpose, we examined low frequencywaves (0.2–0.4 Hz) generated by oceanwaves near the coast and prop-agated to long distances (Longuet-Higging, 1950; Berger et al., 2004).Here, the microseism was recorded 150 km away from the coast.To check the quality of the time synchronization, we computed the

Table 2Characteristics of the 16 vulcanian events of May and June, 2009. The date and time are inuniversal time (local time was 5 h more), E is the event energy in MJ, the ‘Interval’ is thetime interval since the previous explosion, and TR is the tremor duration after the explo-sion.

Item Date time E (MJ) Interval TR

1 2009/05/24 14:43:33 3 0 02 2009/05/24 22:05:23 60 7.4 h 03 2009/05/25 00:21:56 56 2.3 h 0.5 h4 2009/05/25 02:26:47 3 2.1 h 05 2009/05/25 04:35:05 13 2.1 h 1.6 h6 2009/05/25 08:14:40 48 3.6 h 0.5 h7 2009/05/25 12:06:59 22 3.9 h 1.5 h8 2009/05/26 22:54:33 14 1 day 10.8 h 4 h9 2009/05/27 04:52:53 9 6 h 1.1 h10 2009/06/02 13:27:42 48 6 days 8.6 h 16 h11 2009/06/05 00:39:37 314 2 days 11.2 h 2 days 3 h12 2009/06/10 07:31:35 145 5 days 7.8 h 9 h13 2009/06/12 23:16:03 5 2 days 15.3 h 014 2009/06/13 13:26:02 218 14 h 20 min15 2009/06/14 05:15:24 85 16 h 0.5 h16 2009/06/14 12:15:42 383 7 h 1 day 13 h


correlation coefficients c(τ) for each pair of sensors. c(τ) is expressedby:

c τð Þ ¼X

tf tð Þh t þ τð ÞffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiX

tf 2 tð Þ

q ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiXth2 tð Þ

q ð6Þ

where f(t) and h(t) are 1-min timewindows. Good synchronizationwascharacterized by c(τ) close to 1.0 for a time lag close to zero. Using thismethod, we ensured the good quality of the time synchronization foreach array.

Sixteen vulcanian explosions, hundreds of long-period events, andseveral hours of tremors were identified during the period of our obser-vation, and these were classified in the catalog of the Instituto Geofísicodel Perú Volcano Observatory. The vulcanian explosions were small tomoderate-sized and lasting from seconds to minutes. Most of them

0

100

200

300

400

Sei

smic

ene

rgy

[MJ]

1 2 5 20 50 100 200

Time interval [hours]

2*3

4*5

6

789

10

11

12

13*

14

15

161-day

10

Fig. 4. Seismic energy and tremor duration as functions of the interval between theexplosions. The black squares are the explosion energies and the explosion intervals.The horizontal segments are the duration of the post-explosion tremor. The numbersare the explosions, according to Table 2.

(11 of 16) were followed by tremor episodes that lasted from 20 minto 2 days. A practicalway to estimate the seismic intensity is to computethe energy of the time-series data, as proposed, for example, by De laCruz-Reyna and Reyes-Davila (2001) with the real-time seismic energymeasurement. The RMSaverage squared amplitude of the seismic signalwas performed for a fixed time window (T) on an iterative process,as in Eq. (7), and it was then depicted as an energy function over time.

RSEM iTð Þ ¼ log1T

Xt¼iTþT2

t¼iT−T2

y2 tð Þ0@

1A

12

24

35 ð7Þ

where y(t) is the amplitude vector of the seismic signal, and i representsthe sliding window position. Starting from the initial vector seismicdata, the real-time seismic energy measurement function was per-formed with T = 5 min for the total period of observation, as shownin Fig. 3. In Fig. 3, the inverted triangle markers correspond to the 16vulcanian explosions, identified by numbers, the dashed horizontallines specify the duration of the tremor events, and the square markerscorrespond to tectonic earthquakes (outside the volcano edifice). The16 vulcanian events are listed in Table 2, which summarizes the dateof occurrence of each event, the seismic energy, the intervals betweenthe events, and the durations of the post-explosion tremor. The energyE was computed according to Eq. (8) (Johnson and Aster, 2005), by in-tegrating the square amplitudes of the velocities between the fullwave of the explosion confined to a window of 1 min (tremor partnot included), where ρ is the volcano density (2600 kg/m3, r is thesource-station distance (2567 m), vP = 3000 m/s is the P-wavevelocity, this equation can be further simplified by assuming: the atten-uation correction A = 1, and the site-effect constant S = 1 (homoge-neous media and no scattering phenomena).

E ¼ 2πr2ρvPS2

A

Z T

0y2 tð Þd tð Þ ð8Þ

The event energy of the explosions varied between 3 MJ and 383 MJ(Table 2). During this period of explosive activity that lasted 22 days,the interval between two successive explosions varied from 2.1 h toN6 days. The tremors that followed most of the explosions lastedfrom 0.4 days to N2 days. There were no tremors following four of theexplosions. The means of the seismic energy, time interval and tremorduration were 89 MJ, 33.5 h and 7.7 h, respectively.

We did not see any clear relationship between the interval durationand the energy. For example, on May 25, five explosions occurred withvery short and relatively similar intervals, of between 2.1 h and 3.9 h,while the energy fluctuated considerably, between 3 MJ and 56 MJ. Inthe same manner, explosion #6 occurred 3.6 h after explosion #5 withan energy of 48 MJ, while explosion #10 occurred 6days 8.6 h after ex-plosion #9with an equivalent energy, of 48 MJ. Fig. 4 shows the seismicenergy and the tremor duration as functions of the intervals betweenthe explosions. Except for explosion #16, the explosions that occurredwith relatively small time intervals compared to the mean (b8 h)were of low energy (b60 MJ). On the other hand, the explosions withgreater time intervals were of either low or high energy. Therefore, itdoes not appear that the time interval is related to the intensity of theexplosions. For the tremor duration following the explosions, it appearsthat inmost cases the following apply: 1. The tremor lasted a short timeafter low-energy explosions and for short time intervals; 2. the tremorlasted longer for high-energy explosions (#11 & #16) and for longtime intervals.

We compared the tilt and seismic data of explosion event #7(Table 2). The seismic waveforms were filtered with band-pass filtersbetween 0.03 Hz and 2 Hz, and converted to displacement (Fig. 5). InFig. 5a, NUBIZ and WUBIZ correspond to the vertical components ofone seismometer located in the middle of each antenna, and UBI2Z isthe vertical component at UB2. The tilt data were converted to micro-

10 µm

a) NUBIZ

WUBIZ

UBI2Z

390 400 410 420 430 440 450

sec

5 µrad

b) UB2Y

UB2X

416 418 420 422

sec416 418 420 422

sec

c) UB2ZUB2Z-HF

UB2Y

UB2X

-1.0

-0.5

0.0

0.5

1.0

-1.0 -0.5 0.0 0.5 1.0

µrad

-1.0

-0.5

0.0

0.5

1.0

-1.0 -0.5 0.0 0.5 1.0

µrad

d) UB2Y-North

UB2X-East

Fig. 5. Explosion #7waveforms recorded at theNUBI,WUBI andUB2 stations. a. Vertical components of the displacement filtered between 0.03 Hz and 2 Hz. b. Tilt signals recorded at UB2in μrad. c. Zoom of UB2 signals inside the selected rectangle in (a) and (b), with the high frequency (2.2–6 Hz) seismic signal waveform of UB2Z in gray. d. UB2 tilt vector evolution.The black curve corresponds to the tilt for the rectangle.


radians (Fig. 5b). An expansion of the UB2 first-arrival waveforms in thetime window of 415 s to 422 s is shown in Fig. 5c. It is worth recallingthat UB2Ywas oriented to the north, almost radial to the volcano crater,and UB2Xwas oriented to east. The gray curve of UB2-HF in Fig. 5c is thevertical seismic signal (UB2Z) filtered into the band of 2.2 Hz to 6 Hz.

The vector sum of the tilt channels (UB2Y and UB2X) yields the di-rection and magnitude of rotation with respect to the vertical gravityvector (Applied-Geomechanic Tiltmeter manual). The black curve inFig. 5d is the UB2 tilt vector for the time window of 418 s to 420.5 s(Fig. 5c, red dashed rectangle). This time window corresponds tothe signal prior to the seismic broadband first arrival (UB2). The graycurve in Fig. 5d corresponds to the UB2 tilt vector given by the signalsUB2X and UB2Y for the whole time window (Fig. 5c). The radial trendcan be described as a slight upward tilt [418–419], followed by aminor downward tilt [419–420.5], and a clear upward tilt to the maxi-mum [≈421] that coincides with the upward displacement recordedby the broadband seismometer at UB2. The tilt pattern observed onthe radial component of UB2 can be described as a slight inflationfollowed by a deflation, such that a contraction of the crater area startedat 419 s. This was followed by a strong inflation (expansion), whichcan be seen as the onset of the explosion between 420.5 s and 421 s.The beginning of the vertical seismic movement that coincided with

the maximum of the radial tilt was also observed at WUBI. This wasless clear at NUBI, which was at a greater distance from the crater. Inaddition, an emergent high-frequency signal (Fig. 5c, UB2Z-HF) startedalmost simultaneously with the UB2Y downwardmovement, at around419 s.

3. Processing

A primary objective of this study was the location of the sourcesof the 16 vulcanian events that were identified, as in Table 2, to try tobetter understand the explosive dynamics at Ubinas. With this aim,we applied the MUSIC-3C method, as proposed by Inza et al. (2011).Given the lack of a velocity model for the Ubinas volcano, the propaga-tion medium was assumed to be homogeneous.

This next section is divided into several parts. We begin by identify-ing the frequency bands where the energy of the explosions wasconcentrated. We then describe the different steps of the slowness vec-tor estimation, with its application to one of the explosions. Syntheticdata are used to validate our results. Finally, we present the resultsobtained for the whole set of explosions, an example of long-periodevents, and we discuss how our data have improved our understandingof the explosive dynamics.

a)

c) d)

b)

Fig. 6. Filtered (0.5–6 Hz) waveforms of explosion #7. a. Time-series of seven sensors of three components recorded at theWUBI antenna. b. Smoothed spectrum of the 3C components.c. Thinnest curve: average of the smoothed spectrum over all of the traces. Thicker curve: average coherence. d. Average of the smoothed spectra over each component, with the thin linefor the North components, the dashed line for the East components, and the thick line for the vertical components (Z).


3.1. Determination of the useful frequency band

In this part, we focus on the analysis of explosion #7. We considerthe most relevant waveforms within the band of 0.5 Hz to 6 Hz, asshown in Fig. 6. From the top to the bottom of Fig. 6 there are thetime-series waveforms of each 3C component clustered as north (NS),east (EW) and vertical (Z), with the power spectral density for eachcomponent, computed by the smoothed periodogram method. Thecoherences were calculated between the waveforms using the dataof each clustered components shown in Fig. 6a. The thicker curve inFig. 6c shows the average coherence, where high coherence occurs atfrequencies between 0.5 Hz and 2.2 Hz, whereas less significant coher-ence is seen between 2.2 Hz and 6 Hz. A simple method to find the av-erage energy composing the wave fields that impinge on the antenna isby using the spectral average. The arithmetic average of the spectrum ofeach component depicts the global spectrum of the antenna (Fig. 6b),

which illustrates the frequency bins contributing to the seismic body.Moreover, the average of these three spectra shown in Fig. 6c revealstwo frequency bands, 0.5 Hz to 2.2 Hz and 2.2 Hz to 6 Hz; these willbe analyzed separately.

Our interest is focused on the early arrivals of the signals that are themost direct waves from the source. A time-frequency distribution isperformed to identify the frequency ranges that correspond to the firstarrivals. We used time-frequency representation to analyze the nonsta-tionary behavior of the signal that composes the explosion quake. TheHilbert–Huang transform (Huang et al., 1998; Flandrin et al., 2004)provides a clear time-frequency representation of the body waves. Thevertical component waveform of explosion #7 was analyzed using theHilbert–Huang transform. The time-frequency representation (Fig. 7a)shows the dominant frequencies between 0.5 Hz and 2.2 Hz, in thetime window between 3.8 s and 5.8 s, which are depicted in the grayregion in Fig. 7b, and which are the first-arrival seismic waves.

0

1

2

3

4

5

6

7

hz

a)

0.00

0.25

0.50

0.75

12 16

µms-

1 b)

-16

-8

0

8

16

0 4 8

sec

Fig. 7. a. The Hilbert Huang transform normalized to one, as applied to the vertical component of the seismic signal of explosion #7, as recorded by sensor 04 of theWUBI antenna. b. Thetime representation of the signal in (a), where the gray region represents the first-arrival signal.


3.2. Estimation of the slowness vector and source location

To estimate the slowness vector and to describe its variability withtime, we applied a sliding window technique over the entire wave-forms. The sliding window was set to a length of 2 s, with 85% overlap.

a)

b)

Fig. 8. a. Spectral average calculated for each component of the signals recorded by all ofthe WUBI sensors for the time-window defined in Fig. 7b by the gray zone (first-arrivalsignal). The thick line corresponds to the vertical component, the dashed line to theEast component, and the thin line to the North component. b. Averaged spectrum of thethree components, where the squares indicated the bins selected for analysis.

This allowed certain inconsistencies of the sliding window process to bemanaged correctly, such as some waveforms cannot be completelyinside the sliding window for some iterations, due to the lag time.Following the time-frequency analysis, the two frequency bands hadto be analyzed separately, as 0.5 Hz to 2.2 Hz, and 2.2 Hz to 6 Hz(Fig. 7). We first processed the data for the low frequency band(0.5–2.2 Hz). To explain the processing for each sliding window,one time window was taken from the sequence. We are mainlyinterested in the beginning of the seismic traces of the explosion,as our objective was to study the initiation of the physical processof the explosions. Then, we computed the average spectrum forone time window positioned at the beginning of the explosionquake. We first have to identify the bins where the seismic energyis concentrated (Fig. 8a and b). The bins were selected by takingthe spectral peaks above the threshold of 70% of the maximum.There were two frequencie bins at f1 = 1.1 Hz, and f2 = 1.5 Hz(Fig. 8b).

The MUSIC-3C analysis was performed by processing eachfrequency bin (f1, f2) separately. Following this method, a cross-spectral matrix was assembled using the three components. Thenthe eigendecomposition of the data cross-spectral matrix was splitinto two subspaces (signal, noise). The MUSIC spectrum is the in-verse of the projection of the steering vector onto the noise space.Ideally, the steering vector of a frequency bin is orthogonal to thenoise subspace. At this stage, the MUSIC-3C results were expressedin terms of the back-azimuth and the apparent velocity spectrum(see Inza et al. (2011) for details). The back-azimuth is the angleof the wavefront that arrives at the antenna, measured clockwisebetween North and the direction towards the epicenter in the hori-zontal plane. Fig. 9a and b shows the MUSIC-3C spectrum as it wasapplied to each frequency bin (1.1 Hz, 1.5 Hz, respectively). Thesefigures show the vertical sections of the azimuth and the apparentvelocity passing through the maximum of the peak of the spectrum.We deduced from these curves the values of the back-azimuth,the apparent velocity, and the associated errors. We obtained theback-azimuths of 123° +/− 4° and 119° +/− 5° for f1 and f2,respectively. The apparent velocities were 1356 +/− 188 m/s and1507 +/− 250 m/s, respectively. The errors were estimated by

a)

0

50

100

150

200

250

300

350

degr

ees

0.0 0.5 1.0

MUSIC3C

b)

0

1

c)

0

50

100

150

200

250

300

350

d)

0

50

100

150

200

250

300

350

degr

ees

e)

0

1

2000 4000

f)

0

50

100

150

200

250

300

350

m/s

2000 4000

m/s

Fig. 9. Music-3C spectrum normalized to one, obtained for a time window positionedat the beginning of the signal of explosion #7, represented as a function of the back-azimuth and the apparent velocity, for the frequencies of 1.1 Hz (a) and 1.5 Hz (d). b., c.Horizontal and vertical sections of the spectrum following thewhite dashed lines crossingthemaximumof the peak spectrum for the frequency of 1.1 Hz. e., f. As for (b) and (c), forthe frequency of 1.5 Hz.

a)

degr

ees

0.0 0.5 1.0

MUSIC3C

b)

2000 4000

c)

d)

0

30

60

90

120

150

180

degr

ees

e)

0

1

0

30

60

90

120

150

180

0

1

f)

0

30

60

90

120

150

180

0

30

60

90

120

150

180

m/s

2000 4000

m/s

Fig. 10.Music-3C spectrum normalized to one, obtained for a time window positioned atthe beginning of the signal of explosion #7, represented as a function of the incidenceangle and the crust velocity beneath the antenna, for the frequencies of 1.1 Hz (a) and1.5 Hz (d). b., c. Horizontal and vertical sections of the spectrum following the whitedashed lines crossing the maximum of the peak spectrum for the frequency of 1.1 Hz. e.,f. As for (b) and (c), for the frequency of 1.5 Hz.


taking the peak width of 90% of the peak maximum of the MUSIC-3Cspectrum.

Fixing the back-azimuth and the apparent velocity found previously,the MUSIC-3C algorithm allowed an estimation of the incident angle tobe obtained. TheMUSIC-3C spectrum is represented in Fig. 10a and d asa function of the incidence angle and the crustal velocity below theantenna, for each frequency bin, f1 and f2. The incidence is the anglemeasured between the direction of the wavefront moving towardsthe antenna and the vertical (in the vertical plane). The incidentangles and the velocities were estimated by taking the maximum ofthe MUSIC-3C spectra (Fig. 10b, c, e and f). This gave 75° +/− 10°(f1 = 1.1 Hz) and 89.5° +/− 15° (f2 = 1.5 Hz) for the incident

angle, and 1367 +/− 140 m/s (f = 1.1 Hz) and 1391 +/− 150 m/s(f = 1.5 Hz) for the velocity. These values of the velocity are represen-tative of the most superficial layer.

This analysis was then performed iteratively to each sliding win-dow over the explosion signal. Fig. 11 shows the results obtained bythe processing of several successive sliding time windows, startingfrom before the beginning of the explosion, and including the first10 s of the signal. The left part of Fig. 11 shows NUBI, and the rightpart shows WUBI. Fig. 11a and f shows the waveforms for thevertical components of one element of each antenna. Fig. 11b andg shows the logarithm of the short-term average (STA) expressedby Eq. (9), as applied to the waveforms represented in Fig. 11a and

a) NUBI

b)

c)

100

200

degr

ees

d)

50

100

degr

ees

1000

2000

3000

4000

m/s

416 420 424 428

seconds

e)

f) WUBI

g)

h)

100

200

i)

50

100

1000

2000

3000

4000

416 420 424 428

seconds

j)

Fig. 11. Time series representation for the low frequency band of explosion #7. Left, for NUBI; right, for WUBI. a., f. Waveform of the vertical component for NUBI andWUBI, respectively.b., g. STA logarithmic signature for NUBI and WUBI, respectively. c., h.; d., g.; e., j. Back-azimuth, incident angles, and velocity, for the NUBI and WUBI antennas, respectively.


f. Here, we have used the recursive algorithm of STA presented byWithers et al. (1998):

STAi ¼yi

f ststaþ 1− 1

f ststa

� �STAi−1 ð9Þ

where yi is the amplitude of the time-series signal, fs is the samplefrequency, tsta is the STA length of the timewindow, and i is the positionof the time window. The STA algorithm is widely used for real-timeseismic detection, most of the time using the formulation of the STA/LTA ratio (Trnkoczy, 1998; Withers et al., 1998), where the LTA is thelong-term average. Here, we have used STA analysis because it is moreappropriate for short time windows. To obtain the first arrival zone, itis enough to evaluate only the STA, as it surveys the signal variationfor a short time window. When the signal is noise, the STA representsthe energy noise for a time window. The negative slopes at the begin-ning of the curves in Fig. 11b and g are due to the initial values of STAbeing bigger than the noise energy. The inflection point indicates thefirst arrival waves. This is around 418 s for NUBI and 417 s for WUBI.The STA curve indicates well-defined arrivals at both antennas, andshows that the wavefield arrives first at WUBI, which is consequentlycloser to the source. The curves in Fig. 11c, h and d, i show the back-azimuth and incident angles, respectively. The time series of back-azimuth depicted in Fig. 11c and h, remained extremely stable for thefirst arrivals, with values around 180° for NUBI and 116° for WUBI.

After 424 s for NUBI and 422 s for WUBI, the back-azimuth was moreunstable, probably due to the mixing of several waves. The evolutionof incidence angles shown in Fig. 11d and i show twobranches. Focusingon the first arrival times, one set of incidence angles was delimitedbetween 75° for NUBI and 79° for WUBI, with the second source in therange of 90° for NUBI and 92° for WUBI. These two sets of incidentangles corresponded to the distinct frequency bins, of 1.1 Hz and1.5 Hz, respectively. If these two frequencies correspond to distinctsources, this result indicates that they have different positions. Anotherpiece of information given by the slowness analysis is that the positionsof the sources were separated in depth but not in the horizontal plane,as the back-azimuth remains constant whatever the frequency. Finally,Fig. 11e and j indicates the crust velocities in the most superficial layerfor the NUBI andWUBI antennas, respectively. For both of the antennas,the velocities were about 1450 +/− 125 m/s.

A similar analysis was applied to the same explosion #7 for higherfrequencies. Fig. 12 summarizes the data from theMUSIC-3C processingapplied to the NUBI and WUBI array data, with filtering at the highfrequency band between 2.2 Hz and 6 Hz. The first arrivals were identi-fied between 421 s and 422 s (Fig. 12a, f and b, g) with dominant fre-quencies around two bins of 3.6 Hz and 3.8 Hz. The back-azimuthestimations are shown in Fig. 12c and h. These data are relatively similarto those obtained in the lower frequency band, at about 180° to 185° forNUBI, and 115° and 120° for WUBI. The incident angles are shown inFig. 12d and i. The incident angles were extremely scattered in this

a) NUBI

b)

c)

100

200

degr

ees

d)

50

100

degr

ees

1000

2000

3000

4000

m/s

416 420 424 428

seconds

e)

f) WUBI

g)

h)

100

200

i)

50

100

1000

2000

3000

4000

416 420 424 428

seconds

j)

Fig. 12. Time series representation for the high frequency band of explosion #7. Left, for NUBI; Right, forWUBI. a., f. Waveform of the vertical component for NUBI andWUBI, respectively.b., g. STA logarithmic signature for NUBI and WUBI, respectively. c., h.; d., g. ; e., j. Back-azimuth, incident angles, and velocity, for the NUBI and WUBI antennas, respectively.


frequency band, so that no representative value could be deduced.Given these data, we focused our analysis only on the frequency bandfrom 0.5 Hz to 2.2 Hz.

Returning to the results obtained previously in the frequency bandof 0.5 Hz to 2.2 Hz, we estimated the source positions by crossing thesource directions obtained with both of the antennas. To do this, wefirst had to fix the values of the back-azimuth and the incident anglefor both of the antennas, with their respective errors. These angleswere taken for the selected time windows corresponding to the firstarrivals, which were defined by the two inflection points of the STAcurves (Fig. 11b and g). Hence, the average back-azimuth angles andthe corresponding errors were estimated by taking the average valuesfor the selected time windows (gray section in Fig. 11c and h).This gave 181.5° +/− 4° for NUBI, and 117.8° +/− 3° for WUBI.Two distinct values of incident angles were found for both antennas(green and orange sections in Fig. 11d and i). These correspondedto 74.2° +/− 7° and 93.1° +/− 8° for NUBI, and 74.95° +/− 8° and90.3° +/− 7.6° for WUBI.

To delimit the source location, we used a probability approach, aspresented by Inza et al. (2011). We defined a probability density func-tion (PDF) of the back-azimuth and the incident angle by representingthese two parameters as Gaussian variables with the mean includedin (0°–360°) for the back-azimuth, and in (0°–120°) for the incidentangle, and with standard deviations corresponding to the errors esti-mated in the processing. The PDF of the back-azimuth and the incident

angle is presented as a rose diagram in Fig. 13. Moreover, a conditionalprobability function was created from the product of the individual PDFof the back-azimuth and the incident angle for both of the antennas. Theresulting PDF, which includes information given by the two angles ofboth of the antennas, was represented inside the digital grid of theUbinas map as a function of the Universal Transverse Mercator coor-dinates, as presented by Inza et al. (2011). The maximum likelihood ofthe PDF yielded an estimate of the source location. The error is definedby the mean quadratic radius R of the PDF of the source position, and

is defined as R ¼ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiσ2

1 þ σ22 þ σ2

3

q, where σ1, σ2, σ3 are the principal

variances. The two sources lie below the crater, at 4.1 km and 4.85 kmin altitude, +/−250 m, as shown in Fig. 13.

These data clearly showed two source areas for explosion #7. At thispoint, the sliding-window average spectrum of the antenna data(around the first arrival wave) were shown with either one or twofrequency bins that exceeded the trigger level of 70%, at about 1 Hzand 1.5 Hz. The multiple signal classification (MUSIC) algorithm(Bienvenu and Kopp, 1983; Schmidt, 1986) identified (N-1) sources(constrained by the array response function) simultaneously (in itsspectrum), and uncorrelated sources arriving at the antenna fromdifferent locations, where N is the number of sensors. Here, MUSIC-3Cbased on MUSIC identified two sources arriving at the antenna.

Looking at Fig. 11, we observe that two different values of theincidence angle appear simultaneously. These angles correspond to

4800

5000

5000

5200

5400

5400

8192

8194

8196

latit

ude

km

294 296 298 300

3

4

5

altit

ude

km

294 296 298 300

longitude km

-5 -4 -3

-km

Fig. 13.MapofUbinas showing the slowness vector directions given byMUSIC-3C, and themaximum likelihood solution (red contour) that indicates the localization of the two sources forexplosion #7.


two different waves (frequencies) in our analysis, where the process iscontrolled by a slidingwindow of 2-s width for thewaveforms between0.5 Hz and 2.2 Hz. This would thus mean that the two waves weregenerated almost simultaneously, or at least over a time lapse of b2 s.

Sixteen explosions (Table 2) were performed following the sameabove procedure, and each one showed two sources that were locatedalmost in the same regions where explosion #7 was located. Table 3shows the details of the locations for each of the 16 explosions, theseare then transformed into a rose diagram in which an average of themis depicted in Fig. 18 (gray color) as well as the location includingtheir error bars.

Table 3Source localization given by the MUSIC-3C analysis on the 16 explosion events.

Item Date Xlon Ylat Alt1 Alt2 Error

1 2009/05/24 14:43:22 297.53 8192.12 3.96 4.79 0.422 2009/05/24 22:05:13 297.43 8192.34 4.06 4.85 0.243 2009/05/25 00:21:50 297.41 8192.41 3.97 4.81 0.274 2009/05/25 02:26:39 297.30 8192.41 4.01 4.81 0.205 2009/05/25 04:34:58 297.32 8192.41 3.92 4.78 0.316 2009/05/25 08:14:32 297.37 8192.34 3.95 4.85 0.287 2009/05/25 12:06:53 297.37 8192.20 4.10 4.85 0.258 2009/05/26 22:54:26 297.36 8192.28 3.98 4.82 0.279 2009/05/27 04:52:49 297.44 8192.17 3.99 4.85 0.4210 2009/06/02 13:27:34 297.43 8192.24 3.92 4.77 0.3511 2009/06/05 00:39:31 297.43 8192.34 4.11 4.89 0.2612 2009/06/10 07:31:25 297.37 8192.38 3.97 4.79 0.2713 2009/06/12 23:15:52 297.47 8192.39 3.97 4.77 0.2214 2009/06/13 13:25:55 297.39 8192.35 3.88 4.81 0.2815 2009/06/14 05:15:15 297.43 8192.28 3.91 4.80 0.3216 2009/06/14 12:15:35 297.32 8192.23 4.00 4.78 0.29

To ensure that our processing analysis could identify two wavesthat were emitted at the same time, and with two locations that wereseparated by 800 m, we used synthetic data.

3.3. Synthetic sources

With the purpose of examining the robustness of the MUSIC-3Calgorithm for multiple sources, a full waveform synthetic dataset wasgenerated for eight sources beneath the crater, at different altitudesin an isotropic homogeneous medium.

Starting from a digital elevation map of the Ubinas topography, thethree-dimensional discrete numerical elastic lattice method (O'Brienand Bean, 2004) was performed to propagate waves in the structure,with the eight broadband isotropic sources listed in Table 1, and locatedbeneath the summit with the Universal Transverse Mercator coordi-nates of longitude 297.5 km and latitude 819.2 km. In this case, thedataset was calculated for the same array location as that used in thefield study in 2009.

These synthetic data were analyzed in detail in a previous studyby Inza et al. (2011), to test the MUSIC-3C algorithm and to compareit with MUSIC-1C and MUSIC-3C.

In the present study, we used only two synthetic sources, as sources1 and 4 of Table 1, which corresponded more or less to the depths ob-tained by the locating of explosion #7. The back-azimuth and incidenceangles are depicted in Figs. 14 and 15. The data show that the sourcepositions are recovered efficiently, as 4920 +/− 120 m for source 1and 3920 +/− 100 m for source 4. To test the efficiency of separatingtwo distinct sources, we mixed the two signals and applied MUSIC-3Cprocessing to the resulting signal. The data showing the back-azimuth

a) NUBI

b)

c)

100

200

degr

ees

20 24

d)

50

100

degr

ees

seconds

f) WUBI

g)

h)

100

200

20 24

i)

50

100

seconds

Fig. 14.Waveforms of theMUSIC-3C procedure overNUBI andWUBI for synthetic source 1. a., f. Seismic amplitudes. b., g. Logarithmic STA of the amplitudes. c., h. Back-azimuth, in degrees.d., g. Incidence, in degrees.


and incidence angles are shown in the Fig. 16. We observed two inci-dent angles. Two depthswere recovered in good agreement, at altitudesof 4940 +/− 160 m for source 1, and 3960 +/− 120 m for source 4.

3.4. Long-period events

A similar analysis was applied to a long-period event, with thedifferent dominant frequencies compared to the explosion quakesto check the nature of the sources of these events and to compare thiswith the explosions.

We took as an example a long-period event recorded on June 13,2009, around 19:58:59. The waveforms were filtered into the sameband as for the explosions, as 0.5 Hz to 2.2 Hz. Fig. 17 shows the resultsfrom theMUSIC-3C analysis with sliding time window of 4 s, 85% over-lap. Starting from the first inflection point, there were temporal varia-tions of the back-azimuth for the NUBI, while it was relatively stablefor the WUBI (Fig. 17c and h). The incident angle curves shown inFig. 17d and i show two sections that indicate a displacement of thesource with time, or eventually of two sources, one more superficialthan the other, but not at the same time, as was observed for the explo-sion quakes. There were two dominant bins for each section, at about0.55 Hz and 0.7 Hz. The first section can be distinguished nearbyat 3537–3539 s (f1 = 0.55 Hz), for both antennas. An average back-azimuth appeared at 187° +/− 8° for NUBI and 119° +/− 4° forWUBI. The average incidence angle was 49.9° +/− 5.9° for NUBI and37.8° +/− 3.9° for WUBI. For the second section, (Fig. 17d and i) be-tween 3541–3544 s, the averages of the back-azimuths and the errorswere 181° +/− 6° for NUBI and 119° +/− 4° forWUBI, with incidenceangles of 60° +/− 9° for NUBI and 49° +/− 3° forWUBI. Themaximumlikelihood of the PDF of the source positions is shown in Fig. 18.

The deeper source (Fig. 18, red contours) was located at longitude297.2 km and latitude 8192.2 km, with a 1.8 km altitude with a meanquadratic radius error of 490 m. The second source (Fig. 18, green con-tours) was located at longitude 297.5 km and latitude 8192 km, with a2.5 km altitude with a mean quadratic radius error of 339 m. We haveanalyzed the first seconds of the low frequencies of a LP events by filter-ing the signal. A deep source is not contradictorywith the short durationof the signal and the lack of coda at low frequencies. In fact, a coda existsand is composed of highest frequencies (2 Hz). The total duration of thisevent is 26 s. Several events, identical to this one have been recordedduring the experiment. The LP event we analyzed is not a unique case.This type of LP events seems to be characteristic of a non-destructivesource. LP events will be studied in detail in a future work.

4. Discussion and conclusions

4.1. Source location

The main result of this study is the identification of two distinctsources for each explosion, which were located at different depthsin the conduit. The seismic signal was composed of an initial low-frequency part [0.5–2.2 Hz] and a second high frequency part [2.2–6 Hz]. This second frequency band is predominant in the coda of theexplosions recorded during our experiment. Aki (1981) and Aki andChouet (1975) have attributed coda waves to the scattering of seismicwaves in the crust. The coda part of explosion waveforms is highlyfrequency dependent due to the scattering and anelastic attenuation.It can be produced by multiple scattering from randomly distributedheterogeneity according to Del Pezzo et al. (1997). The low frequencieswere dominant at the beginning of the explosion. This initial part of thesignal was composed of two distinct peaks, at 1.1 Hz and 1.5 Hz, which

a) NUBI

b)

c)

100

200

degr

ees

d)

50

100

degr

ees

20 24

seconds

f) WUBI

g)

h)

100

200

i)

50

100

20 24

seconds

Fig. 15.Waveforms of MUSIC-3C procedure over NUBI andWUBI for synthetic source 4. a., f. Seismic amplitudes. b., g. Logarithmic STA of amplitudes. c., h. Back-azimuth, in degrees. d., g.Incidence, in degrees.


we have analyzed separately. We used MUSIC-3C, based on the useof three components of the sensors and assuming a homogeneousmedium to estimate the slowness vectors for the antennas (Inza et al.,2011). Compared to other methods of estimation of slowness vectors(Saccorotti and Del Pezzo, 2000; Almendros et al., 2001a; Métaxianet al., 2002; La Rocca et al., 2004), Music-3C uses seismic informationin three dimensions, which allows the estimation of the angle ofincidence of the wave field. This is an apparent incident angle thatcorresponds to the propagation of rays in the underlying layer to thearray. The angle of incidence was determined, however, with a lowerresolution than the back-azimuth, due to the geometry of the network.Indeed, the inter-sensor distance was greater in the horizontal than inthe vertical plane.

The analysis of the initial part of the explosive events at the twofrequencies 1.1 Hz and 1.5 Hz gave a single back-azimuth value forthe antennas, but two incident angles. We interpreted this result asthe presence of two distinct sources. The back-azimuths and incidentangles estimated at both of the antennas were spatially crossed to de-fine the source location.We used the approximation of a homogeneousmedium, which can be justified by the position of the antennas inthe near-field. According to Lokmer et al. (2007), the distortion ofwave-forms is minimized when they are recorded at close distances from thesources. However, this approximation leads to an additional uncertaintyin the resolution of the location along the vertical axis. The analysis of allof the 16 explosions gave similar results.

We found two sources in each case. The two sources have the samepositions in the horizontal plane. They were located just beneath thecrater. Along the vertical axis, they were separated by about 800 m.The average altitudes of these sources were 3890 m and 4810 m, withan average error of 390 m. Since the altitude of the bottom of thecaldera was 5200 m, and the bottom of the crater was about 200 m to300 m below, the most superficial source was about 100 m to 200 m

below the superficial part of the conduit. These two sources were pro-duced simultaneously, or in any case, as our method of analysis used asliding time window set to a width of 2 s, these two sources cannot bedistinguished temporally. To check the validity of these data, we testedthe ability of our algorithm to separate two sources that were generatedsimultaneously using synthetic sources separated by 800 m in the verticalplane. The positions of the two synthetic sources were correctly foundby analyzing the signals separately, and also by stacking the waveforms.This reinforces the idea of the physical origin of these two sources.

In our previous study (Inza et al., 2011), we tested theMUSIC-3C ap-proach through the analysis of a single explosion, which correspondedto event #10 on the present list. We found only one source at a depthof 4200 m +/− 600 m. In this new study, we found two sources ataltitudes of 3920 m +/− 350 m and 4770 m +/− 350 m, respectively.In thefirst study,we analyzed the dominant peak of the unfiltered signal,which was set to 2.4 Hz. In the present study, we filtered the signalbetween 0.5 Hz and2.2 Hz, to keep the part of the signal that hadthe maximum coherency (Fig. 6c). So, we chose to analyze only thedominant peaks of the portion of the signal that had more coherency.Otherwise, this allowed us to obtain a gain of resolution in the localiza-tion error, as the value obtained in this study is two-times lower.

This explains the differences in the data. The 2.4 Hz peak still exists(Fig. 6c), but with less energy, because of the filtering. That said, thesingle source found in the first study was positioned between the twosource positions defined in the current study. Hence, the two resultsdo not contradict one another. It also shows that there might be severalseismogenic zones in the conduit that correspond to different frequen-cies of the signal. The location of several seismic sources in a conduitwas demonstrated by Thomas and Neuberg (2012) in the case of low-frequency activity at Montserrat. In our case, we found two sources,with perhaps the more important result being that the sources weresimultaneous, or were almost simultaneous. We discuss in the next

a) NUBI

b)

c)

100

200

degr

ees

d)

50

100

degr

ees

20 24 28

seconds

f) WUBI

g)

h)

100

200

i)

50

100

20 24 28

seconds

Fig. 16. Waveforms of MUSIC-3C procedure over NUBI and WUBI for the sum of the synthetic sources 1 and 4. a., f. Seismic amplitudes. b., g. Logarithmic STA of amplitudes. c., h. Back-azimuth, in degrees. d., g. Incidence, in degrees.


section the possible interpretations based on models of vulcanianexplosions proposed in the literature.

4.2. Explosion mechanism

Kanamori et al. (1984), and more recently Iguchi et al. (2008), pro-posed a macroscopic model of a dynamic explosion process in whichoverpressures that triggered vulcanian explosions are generated bygas accumulation below a plug positioned at the top of the conduit.Iguchi et al. (2008) examined processes of inflation and deflation thatwere associated with explosions recorded at three different volcanoes:Sakurajima and Suwanosejima in Japan, and Semeru in Indonesia. Theauthors interpreted the seismic and ground-deformation observationsas inflation before the explosions that was associated with gas accumu-lation at the top of the conduit, and then deflation that began at the startof the explosion, which was interpreted as the ejection of the magmaand draining of the conduit. They also observed a slight and shortcontraction just before the explosion (4–6 s before in the case ofSemeru). This was interpreted as the response to gas leakage atthe cap as a result of the pressure exerted by the gas pocket at the topof the conduit. We examined this model on the basis of our data.

The formation of a gas pocket in the conduit is a very credible hy-pothesis, as it has been shown by Traversa et al. (2011) that the numberof earthquakes increased the long period events before an explosion atUbinas. This study with 143 explosions recorded during the period of2006 to 2008 showed an accelerated rate of the long period for 2–3 hbefore the explosion. The acceleration of seismicity was consistentwith overpressure in the conduit. We examined the seismic data fromthree sites, one of which (UB2) also had a tiltmeter (see: Fig. 5). Theseismic signal represented as displacement (filtered between 0.03 Hzand 2 Hz) indicated the initiation of an explosion by a slight movementup (inflation), which was followed by a strong downward movement

that corresponded to the decompression. Themovement of decompres-sion was not as marked as for the case analyzed by Iguchi et al. (2008).This can be explained by a greater station-to-crater distance at Ubinas(≈2 km) compared to, for example, Semeru (b0.5 km). For the samereason, we did not see slow inflation several minutes before the explo-sion for the tilt data, or the seismic data, as observed by Semeru andSuwanosejima. However, the recordings made in 2006 in the calderawith a CMG-40T seismometer (30 s) 500 m from the crater showed sig-nificant slow vertical movement that started 40 s before the explosion.However, we clearly saw a slight movement of deflation with the tiltcomponent normal to the crater. This movement was consistent witha decompression of the superficial part of the conduit. It started 2 s be-fore the explosion, and it lasted a littlemore than 1 s. Iguchi et al. (2008)and Yokoo et al. (2009) estimated onset time of surface explosion byusing infrasounds generated by vulcanian eruptions. This moderatedeflation was quite similar to that observed by Iguchi et al. (2008) forSemeru, where it was interpreted as the failure of the cap at the top ofthe conduit. The duration of this moderate deflation was variable onthe other volcanoes, as 1 min to 2 min for Sakurajima, 0.2 s to 0.3 sfor Suwanosejima, and 2 s to 3 s for Semeru. Yokoo et al. (2009) ob-served expand signals 0.5–0.7 s prior to the main phases at Sakurajimain 2007.

The observationsmade at Ubinas are consistent with these data. Theassumption of the cap rupture just before the start of the explosion canbe strengthened by the observation of the higher frequency seismic sig-nal (2.2–6 Hz). Indeed, this signal appeared 2 s before the explosion,and a few tenths of a second before the decompression movementobserved on the tilt (Fig. 6c).

We saw two peaks at 1.1 Hz and 1.5 Hz that dominated the first fewseconds of the seismic signal of the explosion, and these signalscorresponded to sources located at depths in the conduit that were sep-arated by 800 m. How should these two sources be interpreted? Iguchi

a) NUBI

b)

c)

100

200

degr

ees

d)

50

100

degr

ees

1000

2000

3000

4000

m/s

3536 3544 3552

seconds

e)

f) WUBI

g)

h)

100

200

i)

50

100

1000

2000

3000

4000

3536 3544 3552

seconds

j)

Fig. 17.Waveforms ofMUSIC-3C procedure overNUBI andWUBI for the long-period event. a., g. Seismic amplitudes. b., f. Logarithmic STA of amplitudes. c., h. Back-azimuth, in degrees. d.,i. Incidence, in degrees. e., j. Crust velocity.


et al. (2008) also referred to two sources. The first corresponded to adecreasing pressure source that started from the gas pocket, and thesecond was associated with increasing pressure, which correspondedto outgassing of water-saturated magma deep in the conduit. Thesecond source was deeper, and it was positioned at 2 km in depthat Sakurajima, and 500 m in depth for Semeru. The two sources ofexplosions at Ubinas might be associated with these two processes;the distance between the two sources was consistent with the resultsfrom other volcanoes. For Montserrat, Druitt et al. (2002) proposedconduit depths of between 500 m and more than 2 km.

Other mechanisms have been proposed to explain how the pressur-ization exerted by a cap on a column of magma can generate vulcanianexplosions. Burgisser et al. (2011) tested four mechanisms by combin-ing data of pressure and porosity of a pre-explosive magma columninto a physical model that reconstructed a depth-reference density pro-file of the column. This studywas performed from samples produced byvulcanian explosions at Montserrat between August and October 1997(Burgisser et al., 2010). The mechanisms tested corresponded to:(1) gas accumulation, which is similar to the mechanism proposed byIguchi et al. (2008); (2) conduit wall elasticity; (3) microlite crystalliza-tion; and (4) magma flow. Depending on these mechanisms, Burgisseret al. (2011) provided vertical layering of the conduit prior to an explo-sion, composed of: (1) a dense and strongly degassed plug from ameterto a few tens of meters in thickness; (2) a shallow transition zone thatis characterized by complex mingling between vesicular and dense

magma, with a thickness varying from 200 m to 500 m, or from400 m to 700 m; and (3) at greater depth, a more homogeneous, low-porosity zone that takes the total column length from 2.5 km to≈3.5 km. If we compare our data to these models, the two sourcesthat we have defined might correspond to the boundaries of the transi-tion proposed by Burgisser et al. (2011). Assuming a cylindrical conduitwith a radius of 25 m to 30 m, as estimated by lava dome observed in2006 (Rivera et al., 2010), the volume of the transition zone was ofthe order of 5 × 105 to 5.5 × 105 m3. This volume corresponds to themaximum volume emitted, assuming that all of the conduits emptiesduring an explosion.

The locations were extremely stable over time. The errors in thedeterminations of the positions of the sources were on average 35%of the distance between the two sources. This might explain whyvariations in the height of the zone of fragmentation of this orderwere not detectable. It is also possible that the geometry of the conduitconstrained the eruptive dynamics. At depth, it is possible that theshape of the conduit did not allow an extension of the transition zone.

We analyzed only a few long-period events from the whole catalog,which includes hundreds of events during the period of the explosivephase. These events with the dominant frequencies of 0.55 Hz and0.7 Hz were located 1.5 km to 2 km deeper than the explosions. Thesources for these two frequency peaks were located at different depths,but contrary to the explosions, the two sources were not simultaneous.The deepest one (0.55 Hz) corresponded to the beginning of the signal,

4800

5000

5200

5400

5400

8192

8194

8196

latit

ude

km

294 296 298 300

2

3

4

5

km

294 296 298 300

longitude km

-5 -4 -3 -2

-km

Fig. 18.Map of Ubinas volcano showing the slowness vector directions, and themaximum likelihood solution (red, green contour), indicating the localization sources for the long-periodevent. The 16 explosion localizations are shown as gray points with their errors bars.


or preceded it slightly, while the more superficial source (0.7 Hz) wasgenerated a few seconds later. At this level, we cannot make furtherinterpretations about the origins of these signals. A larger number ofsignals must be analyzed in a future work. The temporal relationshipswith the explosions should also be analyzed in detail.

Acknowledgments

We are grateful to Alain Burgisser for their discussions on vulcanianeruptions. The appropriate commentsmade by an anonymous reviewergreatly contributed to improve the manuscript.

References

Aki, K., 1981. Source of scattering effects on the spectra of small local earthquakes. Bull.Seismol. Soc. Am. 71, 1687–1700.

Aki, K., Chouet, B., 1975. Origin of the coda waves: source, attenuation and scatteringeffects. J. Geophys. Res. 80, 3322–3342.

Alidibirov, M., Dingwell, D., 1996. Magma fragmentation by rapid decompression. Nature380, 146–148.

Alidibirov, M., Dingwell, D., 2000. Three fragmentation mechanisms for highly viscousmagma under rapid decompression. J. Volcanol. Geotherm. Res. 100, 413–421.

Almendros, J., Chouet, B., Dawson, P., 2001a. Spatial extent of a hydrothermal systemat Kilauea volcano Hawaii, determined from array analysis of shallow long-periodseismicity. Part I: method. J. Geophys. Res. 106, 13,565–13,580.

Almendros, J., Chouet, B., Dawson, P., 2001b. Spatial extent of a hydrothermal systemat Kilauea volcano Hawaii, determined from array analysis of shallow long-periodseismicity. Part II: results. J. Geophys. Res. 106, 13,581–13,597.

Berger, J., Davis, P., Ekstrom, G., 2004. Ambient earth noise: a survey of the globalseismographic network. J. Geophys. Res. 109, B11307.

Bienvenu, G., Kopp, L., 1983. Optimality of high resolution array processing usingeigensystem approach. IEEE Trans. Acoust. Speech Signal Process. 33, 1235–1247.

Burgisser, A., Poussineau, S., Arbaret, L., Druitt, T.H., Giachetti, T., Bourdier, J.-L., 2010. Pre-explosive conduit conditions of the 1997 vulcanian explosions at Soufriere hillsvolcano, Montserrat: I. pressure and vesicularity distribution. J. Volcanol. Geoth.Res. 194, 27–41.

Burgisser, A., Arbaret, L., Druitt, T., Giachetti, T., 2011. Pre-explosive conduit conditions ofthe 1997 vulcanian explosions at Soufriere hills volcano, Montserrat: II. Overpressureand depth distributions. J. Volcanol. Geoth. Res. 199, 193–205.

De la Cruz-Reyna, S., Reyes-Davila, G., 2001. A model to describe precursory material-failure phenomena: applications to short-term forecasting at Colina volcano, Mexico.Bull. Volcanol. 63, 297–308.

De Silva, S.L., Francis, P.W., 1991. Volcanoes of the central Andes. Geological Magazine,129. Cambridge University Press 253–254.

Del Pezzo, E., La Rocca, M., Ibanez, J., 1997. Observations of high-frequency scatteringwaves using dense arrays at Teide volcano. Bull. Seismol. Soc. Am. 87, 1637–1647.

Di Lieto, B., Saccorotti, G., Zuccarello, L., La Rocca, M., Scarpa, R., 2007. Continuous trackingof volcanic tremor at Mount Etna, Italy. Geophys. J. Int. 169, 699–705.

Druitt, T., Young, S.R., Baptie, B.J., Bonadonna, C., Calder, E.S., Clarke, A.B., Cole, P.D.,Harford, C.L., Herd, R.A., Lockett, R., Ryan, G., Voight, B., 2002. Episodes of cyclic

http://refhub.elsevier.com/S0377-0273(13)00344-2/rf0005




































vulcanian explosive activity with fountain collapse at Soufriere Hills volcano,Montserrat. Geol. Soc. London Mem. 21, 281–306.

Ferro, A., Gambino, S., S., P., Falzone, G., Laudani, G., Ducarme, B., 2011. High precision tiltobservation at Mt. Etna volcano, Italy. Acta Geophys. 59 (3), 618–632.

Flandrin, P., Rilling, G., Gonçalves, P., 2004. Empirical mode decomposition as a filterbank.IEEE Signal Process. Lett. 11 (2), 112–114.

Huang, N.E., Shen, Z., Long, S.R., Wu, M.C., Shih, H.H., Zheng, Q., Yen, N.C., Tung, C.C., Liu,H.H., 1998. The empirical mode decomposition and the Hilbert spectrum for non-linear and nonstationary time series analysis. Proc. R. Soc. Lond. A 454, 903–995.

Iguchi, M., Yakiwara, H., Tameguri, T., Hendrasto, M., Hirabayashi, J.-i., 2008. Mechanismof explosive eruption revealed by geophysical observations at the Sakurajima,Suwanosejima and Semeru volcanoes. J. Volcanol. Geotherm. Res. 179, 1–9.

Inza, L., Mars, J.I., Métaxian, J.-P., O'Brien, G.S., Macedo, O., 2011. Seismo-volcano sourcelocalization with triaxial broad-band seismic array. Geophys. J. Int. 187, 371–384.

Johnson, J., Aster, R., 2005. Relative partitioning of acoustic and seismic energy duringStrombolian eruptions. J. Volcanol. Geotherm. Res. 148, 334–354.

Kanamori, H., Given, J., Lay, T., 1984. Analysis of seismic bodywaves excited by theMountSt. Helens eruption of May 18, 1980. J. Geophys. Res. 89 (B3), 1856–1866.

La Rocca, M., Saccorotti, G., Del Pezzo, E., Ibáñez, J., 2004. Probabilistic source locationof explosion quakes at Stromboli volcano estimated with double array data.J. Volcanol. Geotherm. Res. 131, 123–142.

Lokmer, I., Bean, C.J., Saccoroti, G., Patané, D., 2007. Moment-tensor inversion of LP eventsrecorded on Etna in 2004 using constraints obtained from wave simulation test.Geophys. Res. Lett. 34, L22316. http://dx.doi.org/10.1029/2007GL031902 (6 pp.).

Longuet-Higging, M., 1950. A theory of the origin of microseisms. Phil. Trans. R. Soc. A 234(875), 1–35.

Macedo, O., Métaxian, J., Taipe, E., Ramos, D., Inza, L.A., 2009. Seismicity associated withthe 2006–2008 eruption, Ubinas volcano. In: Bean, C.J., Braiden, A.K., Lokmer, I.,Martini, F., O'Brien, G.S. (Eds.), The VOLUME Project, 1, pp. 262–270.

Melnik, O., Sparks, R.S.J., 2002. Dynamics of magma ascent and lava extrusion at SoufriereHills volcano, Montserrat (in the eruption of Soufriere Hills volcano, Montserrat from1995 to 1999). Geol. Soc. London Mem. 21, 153–171.

Métaxian, J.-P., Lesage, P., Valette, B., 2002. Locating sources of volcanic tremor andemergent events by seismic triangulation. Geophys. Res. Lett. 107 (B10), 2243 (18 pp.).

Miron, S., Le Bihan, N., Mars, J.I., 2005. Vector-sensor music for polarized seismic sourceslocalization. EURASIP J. Appl. Signal Process. 2005, 74–84.

O'Brien, G.S., Bean, C.J., 2004. A 3d discrete numerical elastic lattice method for seismicwave propagation in heterogeneous media with topography. Geophys. Res. Lett. 31,L14608.1–L14608.4.

O'Brien, G.S., Lokmer, I., De Barros, L., Bean, C.J., Saccoroti, G., Métaxian, J.-P., Patané, D.,2011. Time reverse location of seismic long period events recorded on Mt Etna.Geophys. J. Int. 184, 452–462.

Papale, P., 1999. Strain-inducedmagma fragmentation in explosive eruptions. Nature 397,425–428.

Paulus, C., Mars, J.I., 2006. New multicomponent filters for geophysical data processing.IEEE Geosci. Remote Sens. 44, 2260–2270.

Rivera, M., Thouret, J.-C., Marino, J., Berolatti, R., Fuentes, J., 2010. Characteristics andmanagement of the 2006–2008 volcanic crisis at the Ubinas volcano (Peru).J. Volcanol. Geotherm. Res. 198, 19–34.

Rost, S., Thomas, C., 2002. Array seismology: method and application. Rev. Geophys. 40(1008) (27 pp.).

Saccorotti, G., Del Pezzo, E., 2000. A probabilistic approach to the inversion of datafrom a seismic array and its application to volcanic signals. Geophys. J. Int. 143,249–261.

Schmidt, R.O., 1986. Multiple emitter location and signal parameter estimation. IEEETrans. Antennas Propag. 34 (3), 276–280.

Sparks, R., 2000. The causes and consequences of eruptions of andesite volcanoes. Phil.Trans. R. Soc. A 358, 1435–1440.

Thomas, M.E., Neuberg, J., 2012. What makes a volcano tick-A first explanation of deepmultiple seismic sources in ascending magma. GEOLOGY, 40, pp. 351–354. http://dx.doi.org/10.1130/G32868.1.

Traversa, P., Lengliné, O., Macedo, O., Métaxian, J.-P., Grasso, J., Inza, A., Taipe, E., 2011.Short term forecasting of explosions at Ubinas volcano Peru. J. Geophys. Res. 116,B11301 (15 pp.).

Trnkoczy, A., 1998. Understanding and setting STA/LTA trigger algorithm parameters forthe K2. Application Note Kinemetrics, 41 (20 pp.).

Withers, M., Aster, R., Young, C., Beiriger, J., Harris, M., Moore, S., J., T., 1998. A comparisonof select trigger algorithms for automated global seismic phase and event detection.Bull. Seismol. Soc. Am. 88, 95–106.

Yokoo, A., Tameguri, T., Iguchi, M., 2009. Swelling of a lava plug associated with avulcanian eruption at Sakurajima volcano, Japan, as revealed by infrasound record:case study of the eruption on January 2, 2007. Bull. Volcanol. 71, 619–630.





















http://dx.doi.org/10.1029/2007GL031902


































http://dx.doi.org/10.1130/G32868.1











analysis of dynamics of vulcanian activity of ubinas ... · seismic data and ground deformation...

Documents