1 Introduction

The Stromboli volcano (Italy) is one of the most active volcanoes in the world, with moderate and persistent explosive activity. Its recent history is characterized by two effusive eruption episodes (Calvari et al., 2005; Martini et al., 2007) that were accompanied by dangerous phenomena, including a tsunami (La Rocca et al., 2004; Pino and Boschi, 2009; Tinti et al., 2005, 2006) on December 30, 2002, and paroxysmal explosions that occurred on April 5, 2003 (D’Auria et al., 2006), and on March 15, 2007 (Martini et al., 2007). These two effusive phases were associated with instability of the slope known as Sciara del Fuoco (Baldi et al., 2005; Falsaperla et al., 2006; Tibaldi, 2001). This structure is a weakness zone of the volcanic edifice that fractures when the explosive activity increases, giving rise to the effusive activity. In particular, during the 2002–2003 effusive phase, there was a partial collapse of this side that caused a tsunami, which spread to the nearby coast with waves up to 10 m high. After this episode, the northwest flank became unstable, and as many as 50 landslide signals were recorded per day (Martini et al., 2007).

The monitoring of the stability of Sciara del Fuoco is an important topic both from the scientific point of view and for Civil Protection purposes. In the framework of the real-time monitoring of the Stromboli volcano, different systems have been implemented to detect the instability of the Sciara del Fuoco. All these techniques rely on geodetic measurements. Their target is to detect an acceleration in the deformation rate of the volcano flank, in order to forecast its possible catastrophic failure. Puglisi et al. (2005) implemented an integrated geodetic monitoring system based on EDM and GPS measurements. Using a statistical analysis of EDM data, Nunnari et al. (2008) developed an early warning system to evaluate the level of hazard on Sciara del Fuoco.

Casagli et al. (2009) propose another geodetic monitoring system based on ground-based synthetic aperture radar (SAR) interferometry (GB-InSAR). This system is able to provide real-time images of the deformation rates on the Sciara del Fuoco flank and was shown to be an effective monitoring tool during the 2007 effusive eruption. However, the cited methods may be less sensitive to very local deformations, such as those that are associated with minor landslides that affect only the superficial part of the slope of the Sciara del Fuoco. In this case, the analysis of seismic signals recorded from the proximal stations can help significantly to identify signs of instability of the slope.

In January 2003, during the aforementioned effusive eruption, the Istituto Nazionale di Geofisica e Vulcanologia (INGV) installed a broadband seismic network to monitor the volcanic activity and to enhance the surveillance of the island (De Cesare et al., 2009). Due to these systems, changes in the very long period (VLP) seismicity of the island were later recorded, from about a month before the 2007 effusive eruption (Barberi et al., 2009; Giudicepietro et al., 2009; Martini et al., 2007). In addition, the new seismic monitoring system highlighted that the first seismic signal associated with a landslide occurred at around 09:00 UTC on February 27, 2007, and was followed by other similar signals. This behavior was recognized as an anomaly, and it was communicated to the Civil Protection authorities before the beginning of the effusive eruption. At 12:39 UTC on February 27, 2007, Stromboli volcano started an effusive phase (Barberi et al., 2009; Giudicepietro et al., 2009; Ripepe et al., 2009).

For this reason we believe that, together with geodetic monitoring, seismic signal analysis could also provide a valuable contribution to the monitoring of Sciara del Fuoco stability, with special reference to the opening of eruptive vents and in general with respect to an early warning of effusive eruptions.

Based on the seismological observations collected in the past 9 years, the present study proposes a system for the automatic detection of landslides that uses a multi-layer perceptron (MLP) neural network (Bishop, 1995). The output of this system is used here to implement a decision-making method that allows changes in the landslide rate to be highlighted. Several methods have been used for seismic event discrimination based on spectral analysis (Hartse et al., 1995; Gitterman et al., 1999) and cross-correlation technique (Joswig, 1990; Rowe et al., 2004). Also the neural networks have been successfully applied in the field of seismology for the detection and discrimination of different types of seismic signals (Cercone and Martin, 1994; Dowla et al., 1990; Dowla, 1995; Del Pezzo et al., 2003; Scarpetta et al., 2005; Tiira, 1999; Wang and Teng, 1995). In particular, for the Strombolian seismicity, it has already been demonstrated that it is possible to distinguish between landslides, explosion quakes and tremor signals using a three-class MLP network (Esposito et al., 2006). Furthermore, among the supervised methods different from the MLP network, the support vector machines (SVM) (Schffolkopf and Smola, 2002) have also often used for seismic patterns classification (Giacco et al., 2009; Masotti et al., 2006; Langer et al., 2009). A performance comparison between these two techniques showed that they tend to perform very similarly (Barabino et al., 1999; Osowski et al., 2004).

Since the aim of this study was the detection of landslides with respect to the other main seismic signals recorded by the monitoring network, the dataset examined was divided into two classes of signals: the first is called “landslide”, and it contains only the landslides; the second is called “other”, and it collects the other two typologies of events that mostly contribute to the seismic wavefield at Stromboli, namely, explosion quakes and volcanic tremor.

In the following, a brief introduction to the Stromboli monitoring network and the dataset is first presented, then in Sect. 3 the methods for the data parameterization are described. In Sect. 4 we explain how the landslide detection was realized by using a two-class MLP-based system. Section 5 describes the elaboration of the neural net output for the setting-up of a decision-making algorithm, and lastly, Sect. 6 is dedicated to our conclusions.

2 The Monitoring System of Stromboli and the Seismic Data

Currently, the Stromboli seismic network consists of 13 digital broadband stations (Fig. 1), equipped with Guralp CMG-40T sensors with frequency response of 0.02–60 s. It has a geometry that provides fine coverage over the island, except on the Sciara del Fuoco side, which represents a gap in the azimuthal distribution of the sensors because of its unapproachable terrain. However, there are some stations very close to the margin of this flank, such as STR8, STRA, STRB (see Fig. 1), that allow any signals that are generated from it, and in particular the landslides, to be well recorded. Thus, these stations are more sensitive to this type of signals which attenuates away from the Sciara del Fuoco.

Fig. 1
figure 1

The monitoring seismic network of Stromboli volcano. The 13 seismic stations are indicated by black triangles. The scale is expressed as UTC coordinates. The inset shows the location of Stromboli Island with respect to the Italian peninsula

Full size image

The data are acquired by the monitoring network at a 50-Hz sampling frequency. The signals that are typically recorded include: volcanic tremor, as a continuous signal with a 1–3 Hz frequency range; explosion quakes, which have no distinct seismic phases and a 1–6 Hz frequency range; and landslides, with an emergent onset and a 1–10 Hz frequency range. Figure 2 shows a record of two and half hours at the STR8 station of the Stromboli network (the E-W horizontal component). On the seismogram, it is possible to distinguish some landslide signals (Fig. 2, red ellipses), explosion quakes (Fig. 2, red arrows) and background volcanic tremor (Fig. 2, blue rectangles). The landslide signals have a peculiar waveform, with a cigar-shaped amplitude envelope, and a spectral content higher than the typical explosion quakes and tremor signals. As we see below, these are the features that were considered during the parameterization phase, and then used by the neural net to identify landslides. In addition, other types of seismic signals are recorded, such as long-period earthquakes, rare volcano–tectonic earthquakes, VLP events (often associated with explosion quakes; Chouet et al., 2003), and teleseisms.

Fig. 2
figure 2

A record of two and half hours at the STR8 station along the E-W horizontal component: the red ellipses highlight the landslides, the red arrows indicate the explosion quakes, and the blue rectangles show portions of the background volcanic tremor

Full size image

The dataset considered contains a total of 537 events recorded by four stations of the network (STRA, STRB, STR1, STR8; see Fig. 1) along the various components. Of these, 267 signals were classified as “landslide” and 270 as “other”, the latter of which included 130 explosion quakes and 140 tremor signals.

3 Data Parameterization

In the present study, we have set up a detector with a binary output, “landslide” or “other”, that is suitable for real-time analysis on sliding windows of incoming signals. As previously mentioned, the analysts can distinguish landslides from the other signals on seismograms due to their well-known frequency content and cigar-like waveform. The parameterization provides compact and robust signal encoding, and it appropriately describes the data. For each event, a 24-s-long recording is taken. As in Esposito et al. (2006), the spectral features are obtained by applying the linear predictive coding (LPC) technique (Makhoul, 1975), that gives the spectral envelope of a signal in compressed form. The LPC predicts a signal sample through a linear combination of its p previous samples:

$$ s{*(n)} \approx c_{1} s(n - 1) + c_{2} s(n - 2) + \cdots + c_{p} s(n - p), $$
(1)

where s (n) is the signal sample at time n, s*(n) is its prediction, and p is the model order, the value of which is problem dependent. The prediction coefficients c i , for i = 1,…, p, are estimated through an optimization procedure that reduces the error between the real signal and its LPC estimate. In our case, we extracted p = 6 LPC coefficients from each of the eight 5.12-s-long Hanning windows into which we divided the signals, with an overlap of 2.56 s. In this way, the LPC-spectrogram of each signal file is described by a vector of 48 spectral features.

The time-domain information of the signal is also preserved, through computing a function defined as the normalized difference between the maximum and minimum signal amplitudes within a 1-s long window. Thus for a 24-s long signal, a vector of 24 temporal elements is obtained.

As a result, the pre-processing returns a reduction in the initial file of 1,200 samples into a 72-element vector that is composed of 48 spectral coefficients, obtained from the LPC analysis, plus 24 time features, obtained from the waveform parameterization.

Figure 3 shows the parameterization of the signals belonging to the class “landslide” (panel A) and those in the class “other” (panel B), which contains explosion quakes and tremor signals. A seismogram is illustrated for each signal, along with the LPC-spectrogram and the vector of the 24 temporal features.

Fig. 3
figure 3

An example of the data parameterization. a Signals of the class “landslide”. b Signals belonging to the class “other”, i.e., explosion quakes and volcanic tremor. For each signal the seismogram (upper panel), the LPC-spectrogram (bottom left), and the temporal features (bottom right) are shown

Full size image

4 Landslide Detection

The landslide identification is performed by a MLP network (Bishop, 1995). The net structure includes: an input layer that receives the 72-dimensional vectors x i to be analyzed; a hidden layer and an output layer with a single neuron (Fig. 4).

Fig. 4
figure 4

The architecture of a MLP network

Full size image

The activation function of the hidden nodes is the hyperbolic-tangent expressed as:

$$ \tanh (a/2) \equiv \frac{{1 - e^{ - a} }}{{1 + e^{ - a} }}, $$
(2)

while the output unit uses the logistic sigmoidal function computed as:

$$ \sigma (a) \equiv \frac{1}{{1 + e^{ - a} }}. $$
(3)

The net output y is calculated as:

$$ y = \sigma \left( {\sum\limits_{j} {w_{jk} \tanh \left( {\sum\limits_{i} {W_{ij} x_{i} } } \right)} } \right), $$
(4)

where x i is the ith input unit, and W ij and w jk are the weights on the connections from the input to the hidden layer and from the hidden layer to the output layer, respectively. The output value ranges within [0–1], indicating the probability that the analyzed signal segment is a landslide.

The MLP carries out supervised learning using the quasi-Newton algorithm (Bishop, 1995), and performs a weight optimization through the minimization of the cross-entropy error function (Bishop, 1995), which for such a binary problem (“landslide/other” discrimination) is defined as:

$$ E = - \sum\limits_{n} {\{ t^{n} \ln y^{n} + (1 - t^{n} )\ln (1 - y^{n} )\} } , $$
(5)

where n is the number of training samples, t is the target vector, and y is the net output.

About 60 % of the total dataset is used to train the net, with events recorded by four stations (STRA, STRB, STR1, STR8; see Fig. 1) of the seismic network along the various components. The remaining events, different from those used for the training, are used for the testing. The percentage of correct classification of the network on the testing files is obtained by comparing the prediction vectors with the target ones (Young, 1993) and can be expressed as a percentage:

$$ P_{\text{correct}} = 100\times\left( {1 \, - \, N_{\text{err}} /N_{\text{tot}} } \right),$$

where N err is the number of signals do not correctly classified by the net and N tot is the total number of the testing signals.

Lastly, to validate the robustness of the net, we randomly changed the weight initialization and the data permutation. In this way the final network performance of 98.67 % was the average of the percentages of correct classification obtained with each experiment (Table 1) on the testing files.

Table 1 The percentages of correct classification of the network obtained in each of the six experiments on the testing files and the averaged network performance are shown. Moreover, the number of correctly classified signals on the total number of events for each class in each test is illustrated
Full size table

5 Decision-Making Algorithm

To set up a method that can highlight changes in the landslide rate, the above described MLP is applied to analyze a 24-s-long sliding window that moves along the signal with an overlap of 12 s. For each window, the net calculates the probability that it contains a landslide signal. Because the net output ranges within [0–1], we fixed a threshold value of 0.95 to detect a landslide. This threshold is used to establish the beginning and the duration of the landslide. An example of the net response versus time for a typical landslide of about 2 min duration is illustrated in Fig. 5.

Fig. 5
figure 5

The net response for a landslide signal. The dashed line indicates the fixed detection threshold value

Full size image

Then a landslide percentage index (LPI) is calculated, as the ratio between the number of windows that are classified as landslide and the total number of windows analyzed in a given time interval:

$$ {\text{LPI}} = \, \frac{{{\text{Number\;of\;windows\;with\;output}} > 0.95}}{\text{Total\;number\;of\;windows}} \times 100.$$

In particular, we chose a 10-min-long time interval of analysis, thus obtaining 49 signal windows, each one of 24 s with an overlap of 12 s.

The decision-making algorithm was applied in particular for the analysis of the data of February 27, 2007, which contains the beginning of the effusive eruption of Stromboli volcano. Moreover, for sake of simplicity, we will show the results of the MLP-based system only for the data recorded by the STR8 station, which is one of the closest to Sciara del Fuoco. Anyway the system could be implemented on any other station and temporal interval.

Figure 6 illustrates the seismogram of February 27, 2007, in which it is possible to observe the first landslide occurred at 08:57 UTC and the eruption onset at 12:39 UTC (Barberi et al., 2009; Giudicepietro et al., 2009; Ripepe et al., 2009).

Fig. 6
figure 6

The seismogram of February 27, 2007 is shown. The first red arrow indicates the occurrence of the first landslide at 08:57 UTC, while the second one points to the eruption onset at 12:39 UTC

Full size image

In Fig. 7 the LPI over 10-min periods for the interval from January 1 to April 2, 2007, the end of the effusive phase, is shown. In particular, the LPI for the February 27, 2007 is visualized in the zoom window. Here, the solid red line indicates the eruption onset at 12:39 UTC, while the dashed red line in both graphs represents the pre-warning threshold value of the LPI. Observing the LPI for February 27, 2007, it can be noted that an increase in landslides is seen from 09:00 UTC, about 4 h before the onset of the eruption. Moreover, the neural algorithm highlights the occurrence of small landslides at around 06:00, which the analysts did not identified as such because of their relatively small amplitudes compared to those of tremor. After the beginning of the eruption, the landslide index also remained high during the lava flow phase on the Sciara del Fuoco side. This is due to the fact that the seismic signal source mechanism of the two processes, landslides and lava flow, is similar: in both cases the signal is due to rolling and sliding of blocks along the slope of Sciara del Fuoco.

Fig. 7
figure 7

The LPI over 10-min periods for the interval from January 1 to April 2, 2007 is shown. The zoom window refers only to the February 27, 2007. Here, the solid red line indicates the eruption onset at 12:39 UTC. In both graphs the dashed red line represents the pre-warning threshold value of the LPI

Full size image

Even if the aim of this work is to show the effectiveness of MLP networks in forecasting effusive eruptions at Stromboli, we believe they could be of more general interest. Landslides have shown to be a short-term precursor of effusive eruptions linked to brittle fracture of the volcano edifice with formation of eruptive fractures. Instead, if the lava flow is due to the overflow of lava without fracturing of the volcanic edifice, then the system can be used to track the activity of the lava front.

In this regard, Fig. 8 visualizes the LPI over 10-min periods for the interval from June 1 to August 31, 2011. The zoom window refers only to the days 1–2 August. In this range, the Stromboli monitoring network recorded two continuous high-frequency signals. The first one on August 1 at 20:52 UTC (solid red line A, Fig. 8), which lasted about 30 min, and the second one on August 2 at 04:14 UTC (solid red line B, Fig. 8), which lasted about 2 h. Both these signals were related to the movement of a small lava flow that was caused by an overflow from one of the summit craters (north vent). This type of lava flow, which typically involves small volumes, is different from those that occurred in 2003 and 2007 due to the opening of an eruptive fracture, and which was not preceded by unusual landsliding.

Fig. 8
figure 8

The LPI over 10-min periods for the interval from June 1 to August 31, 2011 is shown. The zoom window refers to the days 1–2 August, 2011. Here, the two solid red lines denote the two continuous high-frequency signals recorded respectively on August 1 at 20:52 UTC (A) and on August 2, 2011, at 04:14 UTC (B), and related to the formation of a small lava flow. In both graphs the dashed red line represents the pre-warning threshold value of the LPI

Full size image

6 Conclusions

We have set up a landslide detector based on a supervised MLP network that is trained to distinguish between two classes of signals: “landslide” or “other”. Using the detector on a sliding window, it is possible to recognize not only the landslide occurrence, but also the start time and the duration of the seismic signal associated with the landslide. Our neural detector shows an averaged performance of correct classification of 98.67 %. Moreover, we have introduced an LPI, as defined in Sect. 5, that can be used as a decision-making tool for the recognition of anomalies in the landslide rate. The method we propose is based on the choice of a threshold value for the LPI. When this threshold is exceeded, a condition of instability is identified. In particular, from analyzing the case of the February 2007 Stromboli eruption, we would suggest a pre-warning threshold LPI value of 20 %, and would consider a warning as a continued LPI >20 % for more than a 30-min time period. We used the effusive eruption of 2007 as a case study because of the huge available dataset and because it is a typical effusive phase associated with the opening of an eruptive fissure. The data collected during this crisis have shown that the increase of landslide occurrence can be an important short-term precursor of this type of phenomena. The observation made in a period of about 9 years have also enabled us to highlight that lava flows due to overflow from the summit craters are not preceded by a significant increase in the occurrence rate of landslides and typically involve smaller volumes of lava. In this case, our analysis allows to highlight the seismic signals caused by rolling blocks that fall from the lava front and then to trace the activity of the lava flow.

In the future, the proposed decision-making method regarding the recognition of the precursor “increase in landslide rate” will be integrated into the real-time monitoring system of Stromboli. The latter is composed of several modules for automatic analysis of the seismic data, some of which make use of parallel computing techniques (Auger et al., 2006; D’Auria et al., 2004). All of the data processed in real time are available on the system named “EOLO” at the internet address http://eolo.ov.ingv.it/. This system shows that before the last effusive phase at Stromboli, in addition to the increase in the landslide rate that occurred from a few hours before the beginning of the effusive eruption, there were other changes that occurred in the seismicity of the Stromboli volcano from several days before the effusive eruption. These included an increase in the occurrence rate and amplitude of the VLP, changes in the polarization parameters, and an increase in the seismic amplitude (Martini et al., 2007). Therefore, in the future we aim to develop an integrated system for the recognition of all of the seismic precursors, to obtain a more meaningful tool for decision making relating to the state of Stromboli volcano.