Introduction

Volcano instability is defined as the condition for a volcanic edifice to be destabilized to a degree sufficient to increase the likelihood of the structural failure of all or part of the edifice (Voight and Elsworth 1997). Landslides in volcanic terrains comprise a wide spectrum of phenomena, from shallow debris remobilization (small rock falls, dry or wet debris flows) to large-scale slope deformation, eventually evolving into large lateral collapse of the volcanic edifice with the emplacement of rockslide-debris avalanches of tens to hundreds of km3 (Siebert 1984; Sosio et al. 2012). These events may potentially cause severe damage to human society, directly or through secondary events like tsunamis and lahars (Siebert 1984; Major et al. 2000; Bonaccorso et al. 2003). Smaller landslides at active volcanoes are often associated with the precursory activity of larger collapses (Voight et al. 2002). Even when they are not the main hazard issue, mass-wasting may cause long-lasting hazard factor in volcanic setting, due to the persistency of non-equilibrium slope conditions (Major et al. 2000).

Slope failure may occur in response to active deformation (magma intrusion, gas overpressure, substrate spreading) or may result over a long period due to over-steepening, over-loading, or peripheral erosion (Voight and Elsworth 1997). Slope stability is diminished by magmastatic and magma overpressures that accompany intrusion in unstable flanks (Tibaldi et al. 2009). Excess pore pressures in potential failure zones can be generated as a result of intrusion-related mechanical or thermal straining of the rock-fluid medium, pressurized retrograde boiling in high level magma chambers, or hydrothermal fluid circulation (Voight and Elsworth 1997), reducing the resistance to sliding. In addition, tectonic activity can generate earthquakes that may aid slope instability through inertial forces and shaking-induced pore pressure generation (Hazlett et al. 1991).

In this study, landslide susceptibility (LS) was estimated considering the combination of morphological features (e.g., slope angle, slope direction, etc) and geomechanical rock properties characterizing the Island of Stromboli (Fig. 1). Moreover, volcanic susceptibility (VS, i.e., the probability of new vent opening) was estimated. The combination of LS and VS was used as proxy of intrusion-related landslide susceptibility. The effects of other triggers (rainfall, sea waves, seismic shaking) were not considered here. The landslides predisposing factors were estimated by random forest (RF) technique (Catani et al. 2013; Ciampalini et al. 2015), while the probabilistic volcanic vent opening distribution, as proxy for magma injection, was evaluated by QVAST tool (Bartolini et al. 2013). The reasons to apply this methodology to Stromboli volcano are based on the following facts: (i) it has experienced moderate to huge landslides (Tibaldi 2001), (ii) it is geomorphologically prone to instability events (Nolesini et al. 2013), (iii) it is affected by active intense volcanic activity that can significantly affect the stability of slopes (Di Traglia et al. 2014a) and (iv) its flank instability could affect areas inhabited and intensely frequented by people for touristic activities (Nave et al. 2010). This work is based on a large amount of existing information, mainly derived from scientific papers, and from elaboration of topographic data, both collected in a dedicated geo-database. This paper is focussed on testing the impact of the application of the proposed method on the final result, rather than on the comparison of existing methodologies.

Fig. 1
figure 1

a Geographic setting of the Island of Stromboli. b SPOT satellite image (2009) of the Island of Stromboli, highlighting the two settlements (Stromboli and Ginostra villages) and the main active volcanological features (the crater terrace and the Sciara del Fuoco)

Full size image

Geological setting

Stromboli Volcano, located in the Tyrrhenian Sea off the southern coast of Italy (Fig. 1), is the north-eastern most tip of the Aeolian volcanic arc, comprising islands and seamounts of volcanic origin. The Aeolian Islands are characterized by a typical semi-arid Mediterranean condition, with a mean annual precipitation of 602 mm, mostly falling in autumn and winter (Kokelaar and Romagnoli 1995).

Stromboli was affected by several mass-wasting events during its geological history, mainly located in the NW part of the volcano that produced a depression called “Sciara del Fuoco” (SdF). Stromboli is characterized by intermittent mild explosions occurring at a typical rate of 1–10 events per hour (Strombolian activity; Blackburn et al. 1976). The ordinary Strombolian activity is occasionally interrupted by the occurrence of higher-intensity explosions, usually referred to as “major” or “paroxysmal” events, based on the explosion intensity (Rosi et al. 2013). The explosive activity is confined in the crater terrace, comprising the NE Crater zone (NEC), Central Crater zone (CC) and SW Crater zone (SWC), located at ~750 m above sea level (a.s.l.) at the headwall of the SdF (Calvari et al. 2014). Ephemeral vents fed flank eruptions (1 every 5 years; Rosi et al. 2013), which opened at different heights but almost always confined in the SdF (Tibaldi et al. 2009), and were often associated with tsunamis (1 every 20 years; Rosi et al. 2013). The last flank eruption occurred inside the SdF and lasted 4 months from August to November 2014 (Di Traglia et al. 2015; Rizzo et al. 2015; Zakšek et al. 2015), whereas outside the SdF, the last episode dates back to the Roman age (Arrighi et al. 2004). The SdF depression is filled with volcaniclastic deposits and lavas (Apuani et al. 2005a), emitted from the strombolian activity and lava flows from the summit crater terrace and the ephemeral vents within the SdF (Rosi et al. 2013).

Slope instability phenomena at Stromboli (Table 1; Fig. 2) are common and frequently associated with the effusive phases. At the onset of the 1985, 2002–2003, 2007 and 2014 flank eruptions, gravel/debris slides evolved in dry gravel/debris flows along the SdF following the opening of eruptive fissures on the NE flank of the summit cone (De Fino et al. 1988; Calvari et al. 2005, 2010; Carlà et al. 2016). During the 2012–2013 high-intensity strombolian activity, overflows occurred and were often associated with the sliding of the NEC crater-rim (Di Traglia et al. 2014b; Calvari et al. 2016). During the 2002–2003 flank eruption, a rock rotational slide (25–30 × 106 m3) occurred (Bonaccorso et al. 2003). This phenomenon was triggered by the injection of a lateral intrusion (Neri et al. 2008) and caused two tsunamis, which affected the coastline with a maximum run-up of 6–7 m at Stromboli village (Tinti et al. 2005). The submarine flanks of Stromboli volcano are characterized by mass-wasting, affecting a large part of the entire submarine area and often influencing the stability of subaerial slopes (Casalbore et al. 2010).

Table 1 Classification of slope instability phenomena at Stromboli volcano
Full size table
Fig. 2
figure 2

Examples of mass-wasting at the island of Stromboli. a Rock falls affecting the coastal cliffs, mainly in the western part of the island as consequence of sea erosion acting on the lava units emplaced during the Neo-Stromboli period (26–13 ka). In b, the same area as observed from PLEIADES-1A panchromatic image (may 2016). c Rock fall evolving in dry gravel/debris flows within the Sciara del Fuoco depression, classified as deep-seated gravitation slope deformation

Full size image

Methods

Geo-database generation

The geo-database of the Stromboli Island was developed in three different phases: (i) ancillary data collection, (ii) database design and (iii) product generation (Ciampalini et al. 2015). All the ancillary data were included in a dedicated database (Fig. 3) comprising different thematic maps (geological, land cover and landslide inventory maps, digital elevation models (DEM), very high-resolution (VHR) optical photos). Morphometric data of the island was produced using DEMs with a 10 × 10-m resolution generated by the Istituto Nazionale di Geofisica e Vulcanologia (Tarquini et al. 2007, 2012). Landslide inventory map (LIM; Fig. 4) was constructed using VHR (0.5 × 0.5 m) DEM (Salvatici et al. 2016), and two Quickbird VHR-optical images were also used. The geology of the study area was retrieved using the geological maps of the Island of Stromboli (scale 1:10,000; Tibaldi 2010) in digital format. The geological information was classified in order to produce a litho-technical map. Stratovolcano is composed of rock layers with inequality distributed elastic properties (primarily Young’s modulus), and volcanic slope instability depends partly on how fracture resistant the edifices are (Apuani et al. 2005a,b; Schaefer et al. 2015). In this work, the following classification for the Stromboli rocks was used (Apuani et al. 2005a,b):

  1. 1.

    Lava unit (L). Alternation of lava layers (more than 65%) and autoclastic breccia layers;

  2. 2.

    Lava–Breccia unit (LB). Alternation of lava layers (ranging from 35 to 65%) and breccia layers;

  3. 3.

    Breccia unit (B). Alternation of lava layers (less than 35%) and autoclastic breccia layers;

  4. 4.

    Pyroclastic deposits (P). Prevailing pyroclastic poorly welded breccias alternating with tuff and lapillistones.

Fig. 3
figure 3

a Workflow adopted in this study. All available data are collected in a dedicated geo-database. Selected thematic maps were extrapolated from the geo-database and used for landslides and volcanic susceptibility estimation. The final intrusion-related landslide susceptibility map, derived from the previously generated landslides and volcanic susceptibility maps, was re-integrated in the geo-database. b Workflow adopted in this study for the generation of the Stromboli geo-database. All digital data were collected in the Projected coordinate system: WGS 1984 UTM zone 33 Projection: Transverse Mercatore

Full size image
Fig. 4
figure 4

a Structural data of the Island of Stromboli. Sub-areal eruptive fissures and outcrop** sheet intrusions younger 13 ka. The distribution of the last eruptive event occurred outside the SdF depression (San Bartolo lava flow) is also reported. b Subaerial (red dots) and submarine (orange dots) vents (craters of isolated cinder cones, craters belonging to coalescent cinder cones and single craters). c Binary raster containing the distribution of landslides over the Island of Stromboli, derived from the Landslide Inventory Map (vector data)

Full size image

The land cover was deduced by the updated (2000) Corine Land Cover map (scale 1:100,000). Each polygon was classified following the three levels of the Corine Land Cover classification (http://www.eea.europa.eu/publications/COR0-landcover).

Moreover, the geo-database includes the following (Fig. 3):

  • the location of past recognizable eruptive vents and fissures, derived from eruption accounts, reports, and scientific papers (Rittmann 1931; Abruzzese 1935, 1937, 1940; Gasparini et al. 1968; Capaldi et al. 1978; Corazzato et al. 2008; Neri and Lanzafame 2009; Calvari et al. 2011; Marsella et al. 2012; Di Traglia et al. 2015);

  • the structural elements related to this volcanism such as faults identified by previous geological studies (Tibaldi 2001; Tibaldi et al. 2009), monitoring data (Bonaccorso 1998; Falsaperla et al. 1999) and geophysical and geochemical surveys (Finizola et al. 2002, 2003; Linde et al. 2014).

Landslide susceptibility

A random forest implementation based on Matlab (Matworks, version 7.11, treebagger object (RFtb) and methods) was used as a basic model for Landslides Susceptibility Maps (LSM; Catani et al. 2013). RF classification is a machine-learning algorithm for non-parametric multivariate classification (Breiman 2001). The algorithm exploits random binary trees, which use a subset of the observations through bootstrap** techniques (Catani et al. 2013). In particular, from the original data set, a random selection of the training data is sampled and used to build the model; the data not included are referred to as “out-of-bag” (OOB; Breiman 2001). Moreover, a random selection of predictor variables is used to split each node of the trees (LSM; Catani et al. 2013). Each tree is developed to minimize classification errors, but the random selection influences the results, thus making a single-tree classification very unstable (Catani et al. 2013). For this reason, the RF-type methods make use of an ensemble of trees (the so-called “forest”) thereby ensuring model stability (Catani et al. 2013). The choice of the input parameters, i.e. how many (and which) attributes should constitute the optimal configuration of the susceptibility assessment (conditioning variable, LCV), is a fundamental step in the susceptibility assessment process (Catani et al. 2013). The use of the RF treebagger algorithm provides quantitative estimates of variable importance, and therefore an initial and expert-driven selection of the input parameters is not necessary (Catani et al. 2013). For this reason, we prepared and performed tests where a fixed structure RFtb model is repeatedly run with decreasing total number of variables (LCV#). This means running the RFtb estimator with the complete LCV set and then removing one parameter at a time (the least important in terms of increase in prediction error when the variable is permuted across the out-of-bag observations, according to the Matlab function OOBPermutedVarDeltaError – http://www.mathworks.it/it/help/stats/treebagger.oobpermutedvardeltaerror.html), so as to reduce the parameter space. Every predictor set was applied to the test points and evaluated, in terms of misclassification probability (MP), to identify the optimal configuration for each working resolution. This assessment is very important, especially when there is a limited possibility of increasing the training set along with the LCV#.

The DEM morphometric parameters, quantitatively used to characterize landforms and genetically related to geo-morphologic processes and landslide susceptibility (Catani et al. 2013), were derived from calculation in Geographical Information System (GIS) environment. Several parameters were considered (using the focal operators and raster calculator of ESRI’s ArcGIS 10.3), in order to avoid the subjectivity in the choice of explanatory variables: elevation, the terrain roughness (measured as the standard deviation of elevation, calculated on a 3 by 3 pixel kernel window), curvature (general, planar, profile), slope, slope direction (aspect operator in ESRI’s ArcGIS 10.3) and flow accumulation. Moreover, the land use and litho-technical features of the outcrop** rocks were also considered. The complete description of the methodology applied to LSM generation is reported in Catani et al. (2013).

Volcanic susceptibility

The volcanic susceptibility is a spatial forecasting that allows calculating the probability of opening of new emission vents (Felpeto et al. 2007; Martí and Felpeto 2010). It is based on available volcano-structural data (fractures, faults, location of vents, eruptive fissures, etc.), which help to identify previous behaviour of the volcanic area, and on the changes occurring in regional and/or local stress fields originating from tectonic or lithological contrasts. One available tool to determine the volcanic susceptibility is QVAST tool (Bartolini et al. 2013), which determines the probability density function (PDF) based on kernel density estimation. This probabilistic method depends on the combination of different factors such as the size of the volcanic field, the degree of clustering and the density of the volcano-structural data (Cappello et al. 2012; Bartolini et al. 2013; Becerril et al. 2013). For that reason, it requires the selection of an appropriate method for evaluating the bandwidth for the kernel function that allows calculating the volcanic susceptibility of the area, i.e. the probability of new vent opening, using direct and indirect structural data. The advantage to use QVAST lies in the possibility to choose between several approaches to determine the bandwidth for the kernel function, being this parameter the most important factor in representing the degree of randomness in the distribution of past events and determines the shape of the Gaussian kernel and, consequently, the resulting PDF.

In the case of Stromboli, the input parameters used for the spatial analysis were as follows (Fig. 4): (i) the location of past recognizable eruptive vents and (ii) the structural elements (fault, fissures, sheet intrusions) related to this volcanism, both collected in the geo-database.

The PDFs were obtained for each type of structural data with the method of the least square cross validation method (LSCV; Bartolini et al. 2013; Becerril et al. 2013) to obtain the bandwidth parameter, as it was the best method to represent the distribution of analysed data, and were combined in a non-homogeneous Poisson process to obtain the final susceptibility map. The weights assigned were 0.6 for the vents, 0.3 for the faults and 0.1 for the fissures, which provided initial indicative probability distributions associated with each PDF. In particular, we have considered assigning these weights that the high probability of new vent opening is localized in the SdF depression. On the other hand, less probability has been assigned for the Holocene activity in the rift zone, along the NE-SW structural trend (Tibaldi et al. 2009), where only the San Bartolo lava flow occurred during the Roman age (Arrighi et al. 2004). The final susceptibility map represents those areas with greater or lesser probability of hosting a new vent.

Data presentation and discussion

The estimation of the landslides, volcanic and intrusion-related landslide susceptibility was developed in three different phases (Fig. 3): (i) generation of a dedicated geo-database, (ii) estimation of landslide susceptibility using the RF technique (Catani et al. 2013) and (iii) estimation of the probabilistic volcanic vent opening distribution, as proxy for magma injection, based on available volcano-structural data (fractures, faults, location of vents, eruptive fissures, etc.; Fig. 4a, b), using the QVAST tool (Bartolini et al. 2013).

In the Landslides Inventory Map (Fig. 4c), 179 areas affected by mass-wasting phenomena were identified. Landslide dry gravel/debris slides evolving in dry gravel/debris flows are mainly located within the SdF and “Rina Grande-Le Schicciole” depressions, while they are scattered throughout the island and are for the most part related to the presence of Lava-Lava Breccia units. On the opposite, the sea-cliffs constituted by Lava unit are affected by rock-falls, related to strong coastal regression, mainly in the north-western side of the island.

The Landslides Susceptibility Map (LSM) of the Island of Stromboli (Fig. 5a) shows that the areas affected by high to very high landslide susceptibility include the Sciara del Fuoco, the upper part of the volcanic edifice and the coastal cliffs.

Fig. 5
figure 5

a Landslide Susceptibility Map. b Volcanic Susceptibility Map, showing the spatial probability distribution of future volcanic eruptions. c Intrusion-related landslide susceptibility map of the Island of Stromboli. The background topographic data are the same of Fig. 4

Full size image

The performance of the model used was evaluated by building a receiver operating characteristic (ROC) curve (Fig. 6a). The performance of the model in terms of area under curve (AUC) was analysed using a constant minimum number of samples (mS#) proportion (10% of study area). The AUC values of 0.98 highlight very good results. This good finding can be due both to the scale of the analysis (10 m) and to the fact that the existence of the Sciara del Fuoco and Rina Grande–Le Schicciole zones greatly influences the localization of areas with high landslide susceptibility. The landslide conditioning variables (LCVs) ranking and significance (Fig. 6b) permit to better identify the landslide destabilizing factors. Considering the morphometric attributes, only the slope, the slope direction (aspect) and the terrain roughness (measured using the standard deviation of the elevation in a kernel 3 × 3) seem to have influence on landslide susceptibility, while the thematic attributes (land cover and the litho-technical) clearly control the landslide attitude. In particular, the slope, the terrain roughness and land cover are classified at higher rank, meaning that the parameters describing the potential for erosion and mass-wasting always influence the landslide susceptibility. The slope facing direction is a moderate-ranked LCV, while the landslide-prone areas are well identified based on the litho-technical features of the outcrop** rocks. Most of the areas scarcely affected by mass-wasting phenomena are located at low- to moderate elevation (below 400 m a.s.l.), mainly in correspondence with lava units.

Fig. 6
figure 6

a ROC plot and AUC value for the best classifications obtained at Stromboli. b Rank–LCV# plot illustrating the variation of parameter relative importance (expressed as rank using the colour ramp on the right) with parameter space (no. of parameters used LCV#) at the island of Stromboli. Grey colours correspond to combinations of LCV# in which the parameter importance was estimated as poor or where the parameter was discarded

Full size image

Probability of vent opening

In Fig. 5b, the final susceptibility map obtained using QVAST tool is shown. The result highlights the presence of a probability of new vent opening in the NE zone of Stromboli, where the presence of previous activity in the Holocene ensures the possibility of future eruptive vents. It is also shown that the area with the highest probability of vent opening is the area closest to the eruptive vent of the SdF depression, due to the current stress field in the island (Tibaldi et al. 2009).

Volcanic susceptibility map shows a probabilistic vent opening distribution elongated towards NE. This derives from the weights assigned for faults (0.3) and fissures (0.1) in the NE part of the island. The propagation of magma-filled crack (dykes) pattern is controlled by the current regional tectonic stress field and by the pre-existing structure (Anderson 1937; Takada 1994; Rubin 1995; Le Corvec et al. 2013). Since 13 ka, no sheets intruded the SW volcano flank, whereas summit eruptions took place accompanied by a series of fissure eruptions along the opposite NE flank and sheeting around the SdF and within it (Tibaldi et al. 2009). The NE–SW rift system at Stromboli acted during historic times, as suggested by the eruption of the San Bartolo lava flow occurred during the Roman age (Arrighi et al. 2004). At Stromboli, the presence of debuttressing stresses due to the SdF forces the intrusion from the centre of the volcano to be injected towards the depression (Tibaldi et al. 2009). At the same time, the re-activation of the rift trend due to changing in regional stress could account for a new intrusion in the NE part of the island.

The integration of landslides and volcanic susceptibility maps allowed to generate the first “intrusion-related landslides susceptibility map” (Fig. 5c). The areas with higher probability of intrusion-related landslides are located in the upper part of the SdF, the same area characterized by the occurrence of mass-wasting during the 1985, 2002–2003, 2007 and 2014 flank eruptions (De Fino et al. 1988; Calvari et al. 2005, 2010; Carlà et al. 2016). The other area showing moderate susceptibility to intrusion-related landslides is the Rina Grande–Le Schicciole depression. This composite landslide scar was generated by one or more collapse that occurred between 34–26 and 2 ka (Linde et al. 2014), producing rock/debris avalanches (Romagnoli et al. 2009). The moderate susceptibility in the Rina Grande–Le Schicciole depression is mainly related to the presence of destabilizing factors, mainly slope and the presence of Breccia and Pyroclastic deposit units (Apuani et al. 2005b). Finally, the NE sector of the island shows low susceptibility to intrusion-related landslides and allows estimating unlikely the development of landslides in the case of intrusion in this area.

Conclusions

Intrusion-related landslide susceptibility at the Island of Stromboli has been evaluated coupling (1) landslide susceptibility evaluation by random forest (RF) technique and (2) the probabilistic volcanic vent opening distribution, as proxy for magma injection, using QVAST tool.

The results can be summarized as follows:

  • 179 areas affected by mass-wasting phenomena were identified, mainly located within the SdF and “Rina Grande-Le Schicciole” depressions. Sea-cliffs are affected by rock-falls, related to strong coastal regression;

  • areas affected by high to very high landslide susceptibility include the SdF, the upper part of the volcanic edifice and the coastal cliffs;

  • landslide destabilizing factors are the slope, the aspect, the terrain roughness, land cover and litho-technical features of the outcrop** rocks;

  • the highest probability of a fresh vent opening is the area that is closest to the eruptive vent of the SdF depression. Moreover, the VSM highlights a probabilistic vent opening distribution elongated towards NE, highlighting the presence of a probability of new vent opening where previous activity in the Holocene ensures the possibility of future eruptive vents;

  • areas with higher probability of intrusion-related landslides are located in the upper part of the SdF, the same area characterized by the occurrence of mass-wasting during the 1985, 2002–2003, 2007 and 2014 flank eruptions.

These results have the added value of producing a model that can be adapted, modified, updated and generalized in other geological contexts and in particular to other (mainly mafic to intermediate) volcanoes affected by intrusion-related slope instability. In addition to that, the conceptual models resulting from this work have direct implications for the monitoring and the hazard assessment at Stromboli volcano.