Infrasonic arrays are a powerful tool for volcanic monitoring and hazard assessment. Explosions were recorded at Stromboli using a small aperture array of 4 infrasonic stations, allowing precise vent location. The acoustic signals were delayed-and-summed, revealing the existence of two main groups of infrasonic waves. The NE crater produces short (<3 s) high amplitude (20–80 Pa) pressure waves while the SW crater produces small acoustic pressure (10–30 Pa) with long (5–15 s) coda. The two groups reflect different in explosive styles and similar spectral content, centered on 5–6 Hz. When stacked together, acoustic waveforms for each crater reveal the same pressure pulse, which indicates a common source process. We infer that the acoustic onset at both craters is generated by the burst of a large gas bubble while the acoustic coda is controlled by a sustained pressure release.
 Seismic networks, videocameras and thermal sensors are commonly used as monitoring tools for risk mitigation and are deployed on volcanoes to observe the evolution of the explosive activity and to track changes in the source mechanism. However, carefully designed infrasound sensors provide the most direct and reliable measurement of the excess pressure driving volcanic eruptions.
 On explosive volcanoes characterized by a large number of vents, such as Stromboli, changes in explosive style or in the activity level are considered a crucial information, interpreted in terms of precursory behavior of large explosive eruption.
 Changes in tremor amplitude and precursory seismic events have been often recognised before strong volcanic eruptions [Chouet et al., 1994]. However, the analysis of the seismic wavefield is highly complicated by topography [Ohminato and Chouet, 1997; Neuberg and Pointer, 2000] and path effects [Gordeev, 1993; Kedar et al., 1996] and generally leads to a poor source location. Only the use of a large number of broad band seismic stations gives reliable locations and realistic source dynamics [Chouet et al., 1999]. From this point of view, volcano seismology represents a major tool in volcanic hazard assessment, but the application of real time array techniques to a large deploy of broadband sensors is still too complicated and expensive to be wide-world applied.
 Videocameras and thermal sensors can provide a powerful real time monitoring system [Bertagnini et al., 1999; Harris et al., 1997], because changes in the explosive activity can be visually detected. However, videocameras and thermal sensors are strongly affected by weather conditions, which reduces their efficiency in case of poor visibility caused by clouds or gas plumes.
 We present here how the use of small aperture infrasonic array could represent a major tool in volcano monitoring. The array analysis of infrasonic signals produced by volcanic explosions or degassing activity allows a precise source location of the explosive vents. At short ranges, pressure perturbations propagate in the atmosphere with almost no constraint by path effects, and thus contain direct information on the source process. The acoustic wavefield is much simpler to analyze than the seismic and can provide important constraints on source dynamics [Buckingham and Garcès, 1996; Vergniolle and Brandeis, 1996; Ripepe and Gordeev, 1999].
2. Array Configuration
 An array of 4 infrasonic sensors was deployed near the summit of Stromboli volcano (Italy) in September 1999, at ∼350 m from the active craters. Each infrasonic sensor consists of a Monacor condenser microphone MC2005 with a sensitivity of 46 mV/Pa in the range from 2 to 20 Hz. Data were recorded by a 5 channel, 16 bit digital acquisition system at a sampling interval of 18.5 ms. Pressure sensors were deployed in a triangular geometry with external sides of 120 meters. This size of the array was chosen considering the sound speed in the atmosphere (∼340 m/s, at 20°C and 900 m a.s.l.), the frequency content of the infrasound (1–20 Hz) and the sampling interval. We estimated that this array aperture is large enough to guarantee a good discrimination between infrasonic onsets at the 4 stations and it is small enough to record coherent infrasonic waves (Figure 1).
 During the experiment three vents in the crater terrace were active: one in the NE crater and two in the SW crater. Activity at the NE crater was characterised by short-lived (4–5 s) explosions, highly energetic and with a large amount of scoria. The explosions at the SW crater were long lasting (10–20 s) and rich in cold volcanic ash. This explosive behavior at the SW and NE craters has been described several times as characteristic [Ripepe et al., 1993; Chouet et al., 1999] and seems to be quasi-stationary in time.
 We applied array analysis to locate and to discriminate signals produced by this different explosive activity.
3. Tracking Source Position
 Classic beamforming technique can not be applied to locate the source of the infrasonic waves recorded by the array. Beamforming is in fact based on the plane wave front assumption and aim to identify azimuth and velocity of the plane wave that propagates along the array. In our case, the distance from the source (∼350 m) is comparable to the size of the array (∼120) and hence the acoustic wavefront can not be considered plane. In the last years, semblance has been successfully applied to localize the source of tremor [Nishimura et al., 2000] and LP events [Chouet et al., 1999] also in case of spherical wavefront. We have applied a grid search procedure for the highest probability a single point in the 3-dimensional space has to be the source. Semblance is calculated between theoretical and measured distribution and is assumed to be representative of the probability that a single node has to be part of the irradiating source volume.
 We located the acoustic source using this grid searching analysis and considering a 5-meter equispaced grid over an area of 600 × 600 m, centered in the crater terrace. We calculated for each node the theoretical time-delays at each station of the array. Theoretical delay times were used to delay-and-sum the infrasonic waveforms. In our method source is located in the node that has the best semblance with the measured delay times distribution and the highest beam amplitude (Figure 2).
 Acoustic signals produced by the explosions show two recurrent back-azimuth: one is N340° and the other is N315°, consistent with the NE and SW crater direction, respectively. These directions are in good agreement with the location of the volcanic activity observed from the summit. Acoustic waves coming from the SW crater have indeed two slightly different azimuths (N310° and N314°) indicating the presence of two different vents in the same crater (Figure 3). The NE source is located at ∼780 m of elevation, while the two sources in the SW crater are at ∼820 m (Figures 2c–2d). Source elevation has been calculated assuming no topographical constraints in the grid searching procedure, and are in quite good agreement with the average elevation (∼800 m) of the crater terrace. This latter result gives strong reliability to the location method we used and shows the high degree of resolution of acoustic waves.
4. Tracking Different Source Dynamics
 Acoustic signal propagates from the source to the sensors with almost no path distortion and can be directly assumed to be representative of the source time-function. This simplicity in the acoustic wave can be used to discriminate and to track changes in the dynamics of the source at the same vent.
 The source location allowed us to separate infrasonic signals produced by the two active craters. For each explosion, we first delayed the signals recorded at the array and then, after the correction for the geometrical spreading, we summed them. In this way we reduced possible topographic reflections around the station and the noise produced by secondary sources, such as wind. Acoustic waveforms are significantly different from crater to crater (Figure 4) and closely reflect the observed eruptive styles. The long lasting emission of ash and fine incandescent ejecta at the SW crater produces long (5–15 s) infrasonic signal with low pressure (10–30 Pa). Instead, the explosions at NE crater are short living (1–3 s), with abundant emission of lava fragments, and produce high-pressure (20–80 Pa) infrasonic waves.
 These general characteristics of the acoustic waves remain stable in time, but the single waveforms are slightly different from explosion to explosion.
 Acoustic waves generated by the NE crater (Figure 4a) are similar to the acoustic signals described by Vergniolle and Brandeis  and explained as generated by the oscillation of a large gas bubble at the magma free-surface, just before the explosion. In this model, the variability of the acoustic signal produced by the NE crater (Figure 4a) could be justified as different gas overpressure, thickness of the bubble film and volume of the gas slug.
 While, the long lasting signal of the SW crater (Figure 4b) recalls the harmonic signal used by Buckingham and Garcès  and modeled in terms of conduit resonance. Differences in the acoustic signals of the SW crater could be explained as changes in the concentration of volatile in the magma. The amount of volatile controls the sound speed in the magma, and the position of the elastic boundaries by which depend the frequency content of the acoustic signal [Garcès, 1997]. The long duration of the SW acoustic signal should be hence a consequence of the conduit resonance and reflects large reverberation of acoustic energy in the magma conduit.
 However, the short (<2 s) acoustic pulses linked to the NE explosions are more difficult to be explained in terms of conduit resonance, but are better represented by the volume oscillations of a large gas slug which fails to explain the coda of the signal produced by the SW crater. In conclusion, it seems that these two models well explain the observed acoustic signals but only for one crater at a time.
5. Looking for a Common Source
 In order to find common features in the acoustic process generated by each vent, signals sharing the same back-azimuth have been first corrected for the transfer function of the microphone and then staked. We obtained two acoustic signals (Figure 5) each representative for the most stable source process at both craters. Explosions at the NE crater are represented by a single short (∼0.4 s) high pressure (∼30 Pa) wavelet (Figure 5a), while the SW crater shows a low (∼15 Pa) pressure wavelet followed by a long (∼15 s) infrasonic coda (Figure 5c). This indicates that the long infrasonic coda is generated by the source process. The two signals share the same frequency content (Figures 5b and 5d) between 2 and 10 Hz. The main impulse for both wavelets perfectly overlaps (Figure 6) and this indicates that source process was roughly constant during the experiments.
 We considered the oscillation of a gas bubble [Vergniolle and Brandeis, 1996] as source model to calculate gas volume and gas overpressure associated with the two acoustic pulses. Assuming a magma density of 2700 Kg m−3, a viscosity of 400 Pa s, and a thickness of the magma film above the bubble of 0.04–0.05 m [Vergniolle and Brandeis, 1996], we estimate a gas volume of ∼20 m3 for the NE and of ∼35m3 for the SW. While, gas overpressure is ∼4 × 105 Pa for the NE and ∼0.5 × 105 Pa for the SW. When the bubble breaks, gas overpressure will drop down to the equilibrium pressure i) almost instantaneously, if the overpressure is high, or ii) with a slow gradient, if the gas overpressure is low.
 Consequently, large gas volumes at low overpressure will generate long lasting explosion with low-pressure waves propagating in the atmosphere. The duration of the explosion is thus mainly a function of the gas overpressure, which directly controls the gas jet velocity [Wilson and Head, 1981]. High gas overpressure will produce large pressure perturbations in the atmosphere, but the duration of the explosion itself will be short, as for the explosions at the NE crater. However, low regimes of gas overpressure will generate long living explosions with small acoustic amplitude as in the case of the fountain-like explosions of the SW crater.
 The correlation of the acoustic waveforms with eruptive styles is the evidence that mass discharge rate is controlling the acoustic emission of the infrasonic coda. Mass discharge measured at Stromboli [Ripepe et al., 1993] is not steady but varies between i) a sharp mass release and ii) a long fluctuating mass discharge. In harmony with the observed explosive styles, we suggest that the source of the acoustic coda is somehow related to the time the gas-fragments mixture is ejected during the explosion. Infrasonic coda could more simply coincide with the duration of the gas flux during the explosion.
 Infrasonic array represents a major tool in monitoring the activity of explosive volcanoes. We demonstrated that a small aperture array could easily track the position of the source of the sound produced by volcanic activity.
 We showed how a grid searching technique based on semblance allows to identify different active vents in the same crater with a resolution of few tens of meters. On volcanoes such as Stromboli the possibility to control the evolution of the explosive activity at each single crater, is believed to be crucial for a correct risk assessment. This goal can not be achieved by locating the seismic activity.
 The location of the sound source lead to define, by stacking the infrasonic signals, the main characteristics of the acoustic wavefield produced by each crater. The NE crater produces short (<1 s) and strong (∼40 Pa) acoustic waves, while explosions at the SW crater generate a sharp low-pressure (∼16 Pa) onset followed by a long (∼15 s) acoustic coda. These acoustic waves reflect the observed eruptive styles.
 We found that, in spite of the longer duration (>10 s) of the SW crater, the acoustic waves of the two craters show very similar onsets indicating that the source process is the same at the two crater.
 Assuming that acoustic onset (Figure 6) is generated by the volume expansion of a gas bubble, we calculated that explosions at the NE crater are characterized by larger (8 times more) gas overpressure and smaller (2 times less) gas volume than explosions at the SW crater. Gas overpressure controls the gas jet velocity and consequently large volumes at low pressure will generate long lasting explosion and low-pressure acoustic waves.
 We suggest that the long acoustic coda recorded at the SW crater can be associated with the mass discharge rate and reflects fluctuations of a sustained pressure release.
 We are grateful to Milena Moretti, Massimo Della Schiava, Dario Delle Donne for support in the field. Anthony Finizola provided his DEM of the crater terrace. The manuscript was improved by the suggestions of an anonymous referee and the comments of Milton Garces helped to clarify our arguments and to smooth some sentences. This research has been partially supported by GNV-CNR founds.