Towards a standard typology of endogenous landslide seismic sources

The objective of this work is to propose a standard classification of seismic signals generated by gravitational processes and detected at close distances (<1 km). We review the studies where seismic instruments have been installed on unstable slopes and discuss the choice of the seismic instruments and the network geometries. Seismic observations acquired at 13 unstable slopes are analyzed in order to construct the proposed typology. The selected slopes are affected by various landslide types (slide, fall, topple and flow) triggered in various material (from unconsolidated soils to consolidated rocks). We investigate high-frequency bands (>1 Hz) where most of the seismic energy is recorded at the 1 km sensor to source distances. Several signal properties (duration, spectral content and spectrogram shape) are used to describe the sources. We observe that similar gravitational processes generate similar signals at different slopes. Three main classes can be differentiated mainly from the length of the signals, the number of peaks and the duration of the autocorrelation. The classes are the “slopequake” class, which corresponds to sources potentially occurring within the landslide body; the “rockfall” class, which corresponds to signals generated by rock block impacts; and the “granular flow” class, which corresponds to signals generated by wet or dry debris/rock flows. Subclasses are further proposed to differentiate specific signal properties (frequency content, resonance, precursory signal). The signal properties of each class and subclass are described and several signals of the same class recorded at different slopes are presented. Their potential origins are discussed. The typology aims to serve as a standard for further comparisons of the endogenous microseismicity recorded on landslides.

1 Introduction

Seismology can be used to record (remotely and in a noninvasive way) ground deformation processes and to measure stress–strain conditions through the hydromechanical interactions occurring in the media. Seismology is widely used to understand the physical processes taking place on tectonic faults or volcanoes, to investigate fluid reservoir circulation, and more recently to analyze the dynamics of Earth surface processes such as glaciers (Podolskiy and Walter, 2016), snow avalanches (Lacroix et al., 2011; Leprettre et al., 1996; Pérez-Guillén et al., 2016; Sabot et al., 1998; Surin et al., 2000) and landslides (Deparis et al., 2008; Ekström and Stark, 2013; Gomberg et al., 1995; Rouse et al., 1991). In this manuscript, the term landslide describes a wide variety of processes resulting from the downslope movement of slope-forming materials by falling, toppling, sliding or flowing mechanisms (Hungr et al., 2014). Thus, landslides cover a large range of deformation processes, that can be differentiated in terms of sizes and volumes (smaller than 1 m³ up to more than 10⁷ m³), in terms of displacement rates (mm yr⁻¹ to m s⁻¹), and in terms of mobilized material (hard or soft rocks, debris, poorly consolidated soils, and artificial fills).

With the increasing number of seismic sensors deployed worldwide and the development of automatic seismological processing chains, the construction of landslide catalogs using seismology is now possible, especially at the regional scale (e.g., Switzerland; Dammeier et al., 2016; Hammer et al., 2013 or France; Deparis et al., 2008). However, the forecast of a particular landslide rupture or acceleration is still challenging at the slope scale, which is the focus of this work. In the 1960s, Cadman and Goodman (1967) observed an increase in acoustic emissions (AEs) generated by slopes tilted towards failure at both laboratory and field scales. AEs are high-frequency (10–1000 kHz) body waves generated by the release of strain energy through grain rearrangement (Michlmayr et al., 2012). Further studies confirmed these results for several slopes (Dixon et al., 2015, 2018; Rouse et al., 1991; Smith et al., 2014), where correlations between AE, surface displacement and heavy rainfall were documented. AEs record deep deformation processes before signs of displacement are identifiable at the surface. However, AEs are rapidly attenuated with the distance to the sources. The location of the sensors and the type of waveguide are also critical to capture the slope behavior. Recent developments of fiber optic distributed acoustic systems (FO-DAS) offer the opportunity to overcome attenuation limitations and deploy measures over long distances (Michlmayr et al., 2017). More recently, several studies focused on the analysis of the microseismicity (MS) observed on unstable slopes. MS studies analyze the seismic waves generated by the release of strain energy in the ground at a larger scale than the grain-to-grain interactions in the frequency range of 1 to 500 Hz. The method offers the opportunity to remotely record the spatial distribution of the deformation through time (BRGM, 1995; McCann and Forster, 1990) and is less sensitive to attenuation than AE methods. Gomberg et al. (1995) installed seismometers on the Slumgullion slow-moving landslide (Colorado, USA) in order to understand the mechanical processes taking place during landslide deformation. Further studies used the same method for several slope configurations (hard or soft rocks, soils, very slow to rapid movements) but also investigated the possible links between the displacement rate and the seismic energy release (Helmstetter and Garambois, 2010; Spillmann et al., 2007; Tonnellier et al., 2013; Walter et al., 2012, 2013b). Helmstetter and Garambois (2010) correlated the seismic response of the Séchilienne rockslide with the surface displacement rate and the rainfall amount. The analysis of the seismic waves generated by landslides allows for monitoring spatiotemporal changes in the stress–strain field in the material from the scale of microscopic internal damage (Dixon et al., 2003; Michlmayr et al., 2012; Smith et al., 2017) to the initiation (e.g., pre-failure) of large ruptures (Amitrano et al., 2005; Poli, 2017; Schöpa et al., 2018; Yamada et al., 2016b). Both the failure and surface processes (e.g., rockfall, debris flow) generate seismic waves. Physical properties (mass, bulk momentum, velocity, trajectory) of the landslide can be inferred from the analysis of the seismic signals (Brodsky et al., 2003; Ekström and Stark, 2013; Hibert et al., 2014a; Kanamori et al., 1984; Lacroix and Helmstetter, 2011; Levy et al., 2015; Tang et al., 2015). On clayey landslides, drops in shear-wave velocity have been observed before acceleration episodes. This shear-wave variation through time has been documented using noise correlation techniques for laboratory experiments (Mainsant et al., 2012b), and for a few cases in the field at Pont Bourquin landslide (Switzerland; Mainsant et al., 2012a), at Harmaliére landslide (France; Bièvre et al., 2017) and at Just-Tegoborze landslide (Poland; Harba and Pilecki, 2017). Precursory seismic signals are also expected and documented before large failures. Precursory increase in microseismic activity (in terms of event rates and/or average amplitudes) has been observed first before the fall of a coastal cliff (Mesnil-Val, France; Amitrano et al., 2005) and was interpreted as the propagation of a fracture. More recently, repeating events have been detected before the Rausu landslide (Japan; Yamada et al., 2016b) and the Nuugaatsiaq landslide (Greenland; Poli, 2017). These events are likely associated with the repeated failure of asperities surrounded by aseismic slip, driven by the acceleration of the slope displacement during the nucleation phase of the landslide rupture. Schöpa et al. (2018) recorded harmonic tremors that started 30 min before the failure of the Askja caldera landslide (Iceland) with temporal fluctuations of resonance frequency around 2.5 Hz. This complex tremor signal was interpreted as repeating stick-slip events with very short recurrence times (less than 1 s) producing a continuous signal. However, the characterization of the size of the asperity and the velocity of the ruptures associated with these precursory signals are difficult to invert mostly because of the lack of a dense seismic network at close proximity to the slope instability (Schöpa et al., 2018). Therefore, the monitoring of endogenous MS may represent a promising approach, especially with the advent of robust, cheaper and portable seismic sensors and digitizers. It is now possible to install dense sensor networks close to the unstable slopes and record low amplitude signals in broad frequency bands. A wide variety of unstable slopes are currently monitored (i.e., through permanent or campaign installations) with seismic networks of different sizes and instruments (Table 1).

Spillmann et al. (2007)Helmstetter and Garambois (2010); Lacroix and Helmstetter (2011); RESIF/OMIV (2015)Palis et al. (2017); RESIF/OMIV (2015)Roth et al. (2008)Lenti et al. (2013)Brückl et al. (2013)Helmstetter et al. (2017b); Manconi and Coviello (2018)Lotti et al. (2015)Doi et al. (2015)Yamada et al. (2016a)Harp et al. (2008)Chen et al. (2005)RESIF/OMIV (2015)Bièvre et al. (2017)Voisin et al. (2013)RESIF/OMIV (2015)RESIF/OMIV (2015); Tonnellier et al. (2013); Vouillamoz et al. (2018); Walter et al. (2012)Larose et al. (2015); Mainsant et al. (2012a)Tonnellier et al. (2013)Vouillamoz et al. (2018)Gomberg et al. (1995, 2011)Larose et al. (2017)Walter et al. (2011)Hartzell et al. (2017)Amitrano et al. (2005); Senfaute et al. (2009)Norman et al. (2013)Amitrano et al. (2010); Occhiena et al. (2012)Colombero et al. (2018)Bottelin et al. (2013b); Lévy et al. (2010)Bottelin et al. (2013a, b)Bottelin et al. (2013b)Bottelin et al. (2013b)Bottelin et al. (2013b)Le Roy et al. (2017, 2018)Dietze et al. (2017a, b)Zimmer and Sitar (2015)Bottelin et al. (2014)Hibert et al. (2017a)Vilajosana et al. (2008)Hibert et al. (2011, 2014a, 2017c); Levy et al. (2015)Bessason et al. (2007)Abancó et al. (2012, 2014); Arattano et al. (2014); Hürlimann et al. (2014)Navratil et al. (2012)Coviello et al. (2015); Navratil et al. (2012)Arattano et al. (2016)Berti et al. (2000); Galgaro et al. (2005)Arattano and Moia (1999)Arattano et al. (2016)Suwa et al. (2009)Kogelnig et al. (2014); Schimmel and Hübl (2016)Burtin et al. (2014); Walter et al. (2017)Schimmel and Hübl (2016)Schimmel and Hübl (2016)Schimmel and Hübl (2016)Yin et al. (2011)Huang et al. (2007)Huang et al. (2007)Burtin et al. (2013)Itakura et al. (2000)Marcial et al. (1996)Vázquez et al. (2016); Zobin et al. (2009)Lavigne et al. (2000)Worni et al. (2012)Kumagai et al. (2009)Lube et al. (2012)Kean et al. (2015)

Table 1 Table of the instrumented sites. The bolded names correspond to the sites investigated in the present paper to establish the typology.

G: geophone (f = [0.1–10] kHz); SP: short-Period (f = [0.1–100] Hz); BB: broadband (f = [10⁻²–100] Hz); A: accelerometer; P: permanent monitoring; RC: repetitive campaigns; SC: single campaign.
OPVF/IPGP: volcanological observatory of the Piton de la Fournaise/Institut de Physique du Globe de Paris. USGS: United States Geological Survey.

Understanding the possible mechanisms generating these seismic signals needs to be achieved. The discrimination of the endogenous landslide seismic signals is difficult and needs to be established. The objective of this paper is thus to propose a typology of the landslide microseismic signals recorded in the field. The proposed typology is based on the analysis of observations from 13 monitored sites. The typology includes all of the seismic sources recorded at near distances (<1 km) and in the frequency range of MS studies (1–500 Hz), and generated by landslides (1) developed in hard or soft rocks and soils, and (2) characterized by fragile (i.e., rupture) and ductile (i.e., viscous) deformation mechanisms.

In our work, we first discuss all the physical processes that occur on landslides and may generate seismic signals. We further present the available seismic sensors, the most commonly used network geometry and the instrumented sites. Then we establish a classification scheme of the landslide seismic signals from relevant signal features based on the analysis of the datasets of 13 sites. We further discuss the perspectives and remaining challenges of monitoring landslide deformation with MS approaches. The seismic signals associated with very large rock/debris avalanches and slides observed at regional distances are out of the scope of this work.

2 Description of landslide endogenous seismic sources

This section describes the possible hydromechanical processes observed on landslides that are susceptible to generate seismic sources. We present the conditions controlling their occurrences (type of material, topography), their sizes and their mechanical properties.

2.1 Fracture-related sources

The term fracture denominates any discontinuous surface observed in consolidated media and originating from the formation of the rocks (i.e., joint) or the action of tectonic (i.e., schistosity), gravitational or hydraulic loads. In the case of slow-moving landslides, the propagation of the material also creates fractures on the edge and at the base of the moving material. Fractures occur in all types of materials at different scales ranging from grain rupture to metric faults. The term fissure is sometimes used to describe fractures affecting the surface of the ground and for fractures affecting poorly consolidated material. We here include all these surface discontinuities under the general term “fracture”. Fractures are generated in three basic modes (I: opening, II: sliding and III: tearing) depending on the movement of the medium on the sides of the fracture plane. They result from either brittle failure of the media or from desiccation effects forming polygonal failures during soil drying. On landslides, most of the fractures occur in a tensile mode because of the low tensile toughness of the landslide material and the shallow depth (Stumpf et al., 2013). The formation of fractures can also be generated in depth by progressive degradation of the rock through ground shaking and/or through weathering and long-term damage due to gravitational load. At the base and on the edges of the landslide, the movement is assumed to develop fractures in shear mode, creating sliding surfaces. Shearing on the fracture plane and tensile fracture opening/closing generate seismic signals. Shearing takes place at different scales ranging from earthquakes on tectonic plates to grain friction and generates a variety of seismic signals (Zigone et al., 2011). The unstable regime leads to stick-slip behavior where the stress is regularly and suddenly released generating impulsive seismic events. Tremor-like signals or isolated impulsive or emergent events are also generated during plate motions. A variety of signals are observed during glacier motion. Deep icequakes are usually associated with basal motion (Helmstetter et al., 2015a, b; Podolskiy and Walter, 2016; Pratt et al., 2014; Roeoesli et al., 2016a; Winberry et al., 2011). Tremor-like signals are also recorded during glacier motion (Lipovsky and Dunham, 2016). They are characterized by long duration signals of low amplitudes with no clear phase onsets. They are associated with repetitive stick-slip events on the fracture plane. Tensile fracture opening/closing generate similar signals on glaciers at the surface and at depth (Helmstetter et al., 2015b; Podolskiy and Walter, 2016; Walter et al., 2013a). Knowledge of the focal mechanism and location of the source allows differentiation between the tensile and shear mechanism.

2.2 Topple and fall related sources

On vertical to subvertical slopes, mass movement occurs as the topple of rock columns or as the free fall (and possibly bouncing and rolling) of rocky blocks (Hungr et al., 2014). In the case of toppling, the movement starts with a slow rotation of the rock blocks under the effects of water infiltration or ground shaking and ends with the free fall of larger blocks. Rockfalls, during the propagation phase, impact the ground at some location along their trajectory. These impacts generate seismic waves that can be recorded remotely by seismometers. The range of rockfall volumes can be very large, varying from less than one cubic meter to thousands of cubic meters.

2.3 Mass-flow-related sources

Mass flows gather different run-out processes of debris or of a mixture of water and debris. They cover a large range of volumes from large rock avalanches of several millions of cubic meters to small (hundreds of cubic meters) debris falls and flows (Hungr et al., 2001). They can occur in wet or dry conditions. The contacts of the rock/debris fragments with the bedrock and in the mass flow generate seismic radiations (Abancó et al., 2012; Burtin et al., 2009, 2013; Hibert et al., 2011, 2017b; Kean et al., 2015; Levy et al., 2015; Schneider et al., 2010; Suriñach et al., 2001; Vázquez et al., 2016). The seismic signal is hence a combination of grain contacts within the granular flow and of grain-to-ground surface contacts and hence generate a complex seismic signal.

2.4 Fluid-related sources

Hydrological forcing (e.g., precipitation, snowmelt) is one of the most common landslide triggers. The presence of fracture networks, water pipes and the heterogeneity of the rock/soil media result in the development of preferential water flow paths (Hencher, 2010; Richards and Reddy, 2007). These preferential flows induced local saturated area where the increase in pore water pressure may destabilize shallow or deep shear surfaces. In soils, the dissolution of material into finer granular debris creates weak zones prone to collapse either by suffusion (i.e., non cohesive material wash out under mechanical action) or by dispersion (i.e., chemical dissolution of fractured clay soils; Richards, Jones, 1981). In rocks, pipes may develop by erosion. In these saturated fracture networks, hydraulic fracturing can occur creating earthquakes and harmonic tremors related to flow migration in the fractures (Benson et al., 2008; Chouet, 1988; Derode et al., 2015; Helmstetter et al., 2015b; Tary et al., 2014a, b).

3 Landslide seismic investigation

3.1 Sensors used in landslide monitoring

Body and surface mechanical waves may be generated by the sources described in Sect. 2. Body waves (primary – P, secondary – S) radiate inside the media. P waves shake the ground in the same direction they propagate while S waves shake the ground perpendicularly to their propagation direction. Surface waves only travel along the surface of the ground and their velocity, frequency content and intensity change with the depth of propagation. Acoustic waves can be generated by the conversion of body waves at the surface. These waves travel in the air at a velocity of about 340 m s⁻¹, slightly varying with temperature and air pressure. Acoustic waves are often generated by anthropogenic or atmospheric sources (gun shots, explosions, storms, etc.), but can also be generated by rockfalls, debris flows or shallow fracture events. All these mechanical waves are subject to attenuation with the travel distance; the high-frequency waves are attenuated faster than the low-frequency waves. The relatively low energy released by the landslide-related sources makes the choice of seismic instruments to deploy very important. Four types of instruments are used to record ground motion for different frequency ranges and sensitivities. For landslide monitoring, short-period (SP) seismometers and geophones, broadband (BB) seismometers, accelerometers and AE sensors are commonly installed in the field.

• Broadband seismometers are force-balanced sensors with a very low corner frequency (<0.01 Hz) that can record the ground motion with a flat response in a large frequency range [0.01–25] Hz. They require a careful mass calibration during their installation and are sensitive to temperature and pressure variations. They are mostly used to record very weak ground motion and ambient noise;

• SP seismometers are passive or force-balanced instruments with a high corner frequency (>1 Hz). They measure the velocity of the ground with high sensitivity and a flat response in the [1–100] Hz frequency band. They are recommended for volcanic and glacier monitoring among other applications. They are less sensitive to air temperature and pressure variations and do not require mass calibration. They are hence particularly suitable for landslide monitoring. Geophones are similar to SP seismometers but usually cover higher frequencies [1–600] Hz with lower sensitivity. They are mainly used for active seismic campaigns but may also be installed for the same purposes as SP seismometers;

• Accelerometers are strong motion sensors able to record high amplitudes and high-frequency seismic waves. They can resolve accelerations in the frequency bands from 0.1 to 10 kHz. The response of the sensor is proportional to ground acceleration for all frequencies (there is no corner frequency). But the noise level is important for low frequencies and the sensitivity is not as good as for velocimeters. They are used to record strong ground motion in particular when installed close to epicenters (<100 km) of large earthquakes where seismometers usually saturate. For landslides, they are usually used as inclinometers;

• AE sensors can record ground vibrations at very high frequencies (10–10 MHz) and low amplitudes. There are two types of AE sensors: the first type is very sensitive to a narrow frequency band only, while the second type is sensitive to a broader frequency band (Michlmayr et al., 2012). In the field, a waveguide is often installed together with AE sensors in order to counteract the attenuation of the signal. They are used in combination with accelerometers for structural monitoring and for laboratory experiments (e.g., loading, shear, flume tests) and can be used on landslides to monitor very low magnitude sources at the grain-to-grain interactions (Dixon et al., 2003; Michlmayr et al., 2012; Smith et al., 2017);

• In addition, microphones or infrasound sensors can be useful to detect, locate and classify landslides seismic signals (Helmstetter and Janex, 2017; Kogelnig et al., 2014; Schimmel and Hübl, 2016). The detection of acoustic waves and body waves at one point, because they propagate at different velocities, can be used to estimate the distance from the source. The relative amplitude of seismic and acoustic waves can also provide information on the depth of the source, because shallow sources generate more acoustic waves than deeper ones.

It must be noted that AE sensors only record AEs generated at very high frequencies (>10 kHz) and consequently are very sensitive to attenuation. Indeed, attenuation factor Q is estimated to range between 10⁻² and 10¹ dB cm⁻¹ (Michlmayr et al., 2012). Even with a waveguide, they must be collocated with the cracks or the sliding surfaces observed on the slope (Dixon et al., 2015). BB and SP seismometers and geophones record seismic signals in the common band of 10⁰–10² Hz and hence offer a solution to monitor more distant sources. The detection of a seismic source by MS sensors depends on the seismic energy released by the source, the sensor to the source distance and the attenuation of the media. Installation of MS sensors at the proximity of the geomorphological features of interest (e.g., scarp, faults, sliding surfaces, superficial crack networks, etc.) optimize the detection of the seismic signals generated by those processes but distant sources (>1 m) can also be recorded by MS sensors. The latter do not need to be colocated with the geomorphological features of interest. After correcting the sensor response, the signals generated by these sensors can be analyzed and compared in their common frequency range. Installation of BB seismometers can complete SP networks and enable us to investigate the low-frequency signals generated by the slope, while geophones are more adapted to explore very-high-frequency content (>100 Hz). Dense networks of the latter instruments are recommended to investigate the seismicity induced by landslide deformation while the installation of one unique BB seismometer is enough to investigate the low-frequency radiations of the landslide.

3.2 Network geometry

Several network configurations have been tested in different studies. It must be noted that the network geometry in the case of landslides is constrained by the site configuration. Indeed, the maintenance of seismic sensors may be very challenging when installed on the moving parts of the landslide; therefore, an installation on the most stable parts of the landslide or at its vicinity is often preferred for permanent monitoring (Helmstetter and Garambois, 2010; Spillmann et al., 2007; Walter et al., 2017). During field campaigns, maintenance of sensors installed on the unstable slopes is possible and often realized (Gomberg et al., 2011; Tonnellier et al., 2013; Walter et al., 2012). Therefore, the main challenges for seismic sensor installation at this scale are (1) to locate the sensor at close distance to the sources, (2) to maximize the number of stations and to locate the sensor close to each other to record the same event at different seismic station, and (3) minimize the azimuthal gap between the sensors. The number of deployed sensors plays an important role in the magnitude of completeness (M_c) of the seismic network, while the geometry of the network (i.e., inter-sensor distances, azimuthal gap) mostly controls the accuracy of source locations.

Seismic sensors can be deployed in a network of single sensors or a network of sensor arrays. The difference between seismic networks and seismic arrays is related to the distance at which the signals recorded by two sensors can be correlated. In the case of seismic arrays, the distance between the sensors is reduced to maximize the correlation of the signals recorded by each sensor. Otherwise the installation is called a seismic network (Podolskiy and Walter, 2016). Although the inter-sensor distance is often small (<1 km) in the case of landslide monitoring, decorrelation of the signals is often observed even at small distances due to the complexity of the underground structure especially at high frequencies. The use of the seismic array approach in landslide monitoring often refers to specific geometries of collocated sensors (inter-sensors distances <50 m) organized with a central sensor (often a three-component seismometer) and several satellite sensors (often vertical sensors). This kind of installation presents many advantages such as enhancing the signal-to-noise ratio and allowing the computation of the back azimuth of the source with beam-forming methods.

For the majority of the instrumented landslides, seismic networks are organized with single sensors located on or at close distance to the unstable slopes. The inter-sensors distance and the azimuthal gap are often controlled by the location of easily accessible or stable portions of the slopes. However, a specific geometry can be adopted such as (almost) a linear geometry. This is particularly the case for monitoring the propagation of debris flows in stream channels. Dense networks (number of sensors >50) can also be deployed. In this case, the sensors are installed using a grid geometry with regular inter-sensor distances. This kind of installation is probably the most optimal but is currently mostly realized during short acquisition campaigns due to the difficulty to maintain a large number of sensors over long periods (battery, data storage, possible movement of the sensor), especially when installed directly on the unstable zones of landslides. Finally, the installation of sensors at depth (>1 m) is challenging for landslides and it has currently only been realized on hard-rock slopes (e.g., Randa; Spillmann et al., 2007, or Séchilienne; RESIF/OMIV, 2015). This kind of installation is, however, very valuable to constrain the depth of the sources.

3.3 MS processing chains

One of the current challenges for landslide MS analysis is the development of dedicated processing chains able to analyze the unconventional seismic signals observed on landslides. The three steps of MS processing are, successively, the detection, the classification and the location of the endogenous seismic events. The development of robust and versatile processing chains for analyzing landslide MS is challenging because of (1) the low magnitude of the events and the attenuation of the media that results in emergent and low signal-to-noise ratio records, (2) the seismic source radiation patterns that may be single centroid source, double couple source or volumetric source, and, (3) the heterogeneity and variation in time (i.e., topography, water table levels, fissures) of the underground structure preventing the construction of precise velocity models and hence, accurate source locations.

First, for detecting automatically or manually the seismic events, the use of spectrograms is common. Spectrograms represent the evolution of the frequency content in time by computing the Fourier transform on small moving time windows (e.g., <1 s). Automatic detection is usually carried out with the STA/LTA (short-term average/long-term average) detector (Allen, 1982) applied on the summed energy of the spectrogram (Helmstetter and Garambois, 2010; Spillmann et al., 2007; Tonnellier et al., 2013).

Second, classifying the detected signals can be carried out automatically by discarding exogenous events with simple criteria (i.e., threshold on the signal duration, inter-trace correlation, apparent velocity) but the determination of the threshold to differentiate the class of signals may be difficult. Machine learning algorithms offer nowadays the possibility to automatize and improve this step. Dammeier et al. (2016) developed a hidden Markov model (HMM) that can detect automatically in the time series the occurrence of one particular type of events. The success rate of HMM is reasonable and this technique has the advantage of requiring only one single example to scan the time series. The Random Forest algorithm has proven its efficiency for volcanic and landslide signals classification with higher success rate and versatility (Hibert et al., 2017c; Provost et al., 2017a). New signals are successfully classified into multiple pre-defined classes and changes in the source properties may be detected by change on the uncertainties (Hibert et al., 2017c). It must be noticed that this approach requires a training set with sufficient elements to build the model. Good success rates (i.e., >85  %) are rapidly reached with 100 elements or more per class. Template-matching filters have also been used in many studies of landslide collapse and glaciers (Allstadt and Malone, 2014; Bièvre et al., 2017; Helmstetter et al., 2015a, b, 2017a; Poli, 2017; Yamada et al., 2016a) in order to detect and classify seismic signals. This method consists of scanning continuous data to search for signals with waveforms similar to template signals. It can detect seismic signals of very small amplitude, smaller than the noise level. Seismic signals are grouped in clusters of similar waveforms, implying similar source locations and focal mechanism.

Finally, the location of the sources is the most challenging step. Common location methods (such as NonLinLoc; Lomax et al., 2000, 2009) were used in combination with 3-D velocity models for locating impulsive microearthquakes occurring at the Randa rockslide (Spillmann et al., 2007). However, a certain number of recorded signals do not exhibit impulsive first arrivals and clear P- and S-wave onsets. For this kind of signal, location methods based on the inter-trace correlation of the surface waveform (Lacroix and Helmstetter, 2011) or on the amplitude (Burtin et al., 2016; Walter et al., 2017) are more suitable and easier to automatize. Other methods such as HypoLine (Joswig, 2008) aim at integrating different strategies (i.e., first arrival picking, inter-trace correlation and beam-forming) to accurately locate the epicenter under the control of an operator, while (Provost et al., 2018) developed a method combining amplitude source location and inter-trace correlation of the first arrivals in an automatic scheme. In most of the studies, the media attenuation field and/or the ground velocity are approximated to a 1-D model, and/or do not take into account the topography. Both the complexity of the landslide underground structure and of the recorded seismic signals lead to mislocation of the events that prevents the accurate interpretation of certain sources and leads to false alarms (Walter et al., 2017).

3.4 Instrumented sites

In the last two decades, seismic networks have been installed on several unstable slopes worldwide. Table 1 summarizes the unstable slopes or debris-flow-prone catchments instrumented with seismic sensors worldwide. The sites are classified in terms of landslide types (i.e., slide, fall or flow) according to the geomorphological typology of (Cruden and Varnes, 1996). Studies on snow avalanches (Biescas et al., 2003; Kishimura and Izumi, 1997; Lawrence and Williams, 1976; Sabot et al., 1998; Suriñach et al., 2001) are not integrated. Most of the instrumented sites are located in the European Alps (France, Italy and Switzerland). Short-period (SP) seismometers and Geophones (G) are the most common type of instruments. Their installation and maintenance is easy as they do not require mass calibration in comparison to broadband (BB) or long-period (LP) seismometers.

4 Data

Seismic observations from 13 sites are used to propose the typology. The sites are representative of various types of slope movements and lithology (Table 1) with four slides occurring in hard rocks, four slides occurring in soft rocks, three rockfall-prone cliffs occurring in hard and soft rocks, and one catchment prone to debris flows. The seismic instruments installed on these sites are recording the seismicity generated by the slope deformation and are installed either permanently or were acquired during short campaigns (Table 1). The Riou Bourdoux catchment is the only site where the seismic signals were manually triggered as rock blocks were thrown down the cliff and monitored with cameras, lidar and seismic sensors (Hibert et al., 2017a).

Table 2 Characteristic of the seismic network for the 13 sites analyzed in the present parer. The landslide dimensions are given for the most active area of the slope instabilities (as presented in the published studies). The total number of sensors in the seismic network are given, as well as the minimal and maximal inter-sensor distance and distance to the active zone. In the cases where only a few of the sensors have been investigated in the present study, we indicate the number of the sensors as well as the name of the station in parenthesis.

G: geophone (f = [0.1–10] kHz); SP: short-period (f = [0.1–100] Hz); BB: broadband (f = [10⁻²–100] Hz); SN: seismic network; D-SN: dense-seismic network;
SA: seismic array; L-SA: linear-seismic array; SS: single sensor; ^∗ investigated stations: FOR, MOL and GAR.

The dimensions of the unstable slopes range from 60 m × 30 m for the Chamousset cliff to 7 m × 300 m for the Saint-Eynard cliff (Table 2). The seismic networks are deployed with various geometry depending on the configuration of the slope, its activity and the duration of the installation. For most of the sites, at least one seismic sensor is deployed on the active zone or very close to it (Table 2). The maximal distance to the slope instabilities is 500 m for the Saint-Eynard cliff being the largest investigated site of our study.

The seismic network geometry of the majority of sites is a distributed seismic network where sensor locations are regularly installed over the active zone or in its vicinity. In the case of the Rebaixader catchment, the seismic network is installed at the border of the stream channel almost linearly. At the Slumgullion landslide, a dense network has been installed with regular spacing of the seismic sensors. Seismic arrays are installed at the other sites. The geometry of the seismic arrays are triangular in shape with the exception of the Séchilienne landslide where an hexagonal shape is used.

The instruments are mostly SP seismometers with natural frequencies of 1 to 5 Hz. Fewer geophones and BB seismometers are installed at the sites. The instrument response is corrected for all of the dataset. To be consistent with the sensitivity of all the sensors, we do not investigate the data below 1 Hz for BB seismometers or above 100 Hz for SP seismometers and geophones.

The dataset being analyzed is composed of either published seismic events or published catalogs. The comparison of these events and catalogs enable us to compare the signals and to compose the classes of the typology. In the case that no published events or catalogs are available, we manually analyzed the dataset to complete the number of examples for each proposed class (see Sect. 5 for detailed information).

5 Methodology

The seismic signals recorded at different sites are compared in order to identify common features. Seismic signals result from the convolution of both the wave propagation and of the seismic source mechanism. Consequently, the observation of common signal features in signals recorded at different sites can only be explained by similar source mechanisms. The proposed typology is hence based on the analysis of these common features. We then selected nine signal features in order to quantify the differences and similarities between the different classes. The nine parameters are chosen because they correspond to the criteria used by experts to analyze and classify a seismic signal and also because they can be used in automatic classification algorithms (Curilem et al., 2009; Fäh and Koch, 2002; Hammer et al., 2012, 2013; Hibert et al., 2014a, 2017c; Langer et al., 2006; Maggi et al., 2017; Provost et al., 2017a; Ruano et al., 2014). They can be computed for any signal type and present a robust framework for future comparison. The selected signal features are the following.

• The duration of the signal T (expressed in seconds) is computed on the stacked spectrogram of the traces (Helmstetter and Garambois, 2010).

• The dissymmetry coefficient of the signal (expressed in percent) is computed as

  $$\begin{array}{}\text{(1)}& s={\displaystyle \frac{t_m-t_{\mathrm{1}}}{t_{\mathrm{2}}-t_{\mathrm{1}}}}\times \mathrm{100},\end{array}$$

  where t₁, t₂ and t_m are the time of the signal onset, ending and maximum, respectively.

• The number of peaks of the signal envelop N_peaks, computed as the number of local maximum above 50 % of maximal value of the signal envelop. The envelop of the signal is computed as the absolute value of the Hilbert transform of the signal. The envelop is smoothed by computing the average on a moving window of length $\mathit{\delta }t=\frac{\mathrm{100}}{f_{s}T}$.

• The duration of the signal autocorrelation is defined as

  $$\begin{array}{}\text{(2)}& T_{\text{corr}}={\displaystyle \frac{t_c}{T}},\end{array}$$

  with

  $$\begin{array}{}\text{(3)}& t_c=\underset{t}{max}\left(C\right(t)<\mathrm{0.2}\times max(C\left)\right),\end{array}$$

  where C is equal to the signal autocorrelation. T_corr is expressed in percent (%) and represents the duration of the signal correlating with itself. As an example, a signal with a rapid and abrupt change in frequency content will rapidly be uncorrelated (low T_corr), while a signal with a constant frequency content will have a long autocorrelation (high T_corr).

• The mean frequency (expressed in hertz) is computed as

  $$\begin{array}{}\text{(4)}& F_{\text{mean}}={\displaystyle \frac{{\sum }_{i=\mathrm{1}}^N\text{PSD}\left(f_i\right)f_i}{{\sum }_{i=\mathrm{1}}^N\text{PSD}\left(f_i\right)}}\end{array}$$

  with the power spectral density (PSD) defined as

  $$\begin{array}{}\text{(5)}& \text{PSD}\left(f\right)={\displaystyle \frac{\mathrm{2}\left|\text{FFT}\right(y){|}^{\mathrm{2}}}{N{f}_s}},\end{array}$$

  where f_s and N are the sampling frequency of the signal and the number of samples, respectively. The mean frequency is chosen here as it is more representative of the signal spectrum energy and less sensitive to noise than the frequency of maximum energy. (Farin et al., 2014).

• The frequency corresponding to the maximal energy of the spectrum is denoted F_max (expressed in hertz).

• The frequency bandwidth F_w is defined as

  $$\begin{array}{}\text{(6)}& F_w=\mathrm{2}\sqrt{{\displaystyle \frac{{\sum }_{i=\mathrm{1}}^N\text{PSD}\left(f_i\right)f_i^{\mathrm{2}}}{{\sum }_{i=\mathrm{1}}^N\text{PSD}\left(f_i\right)}}-F_{\text{mean}}^{\mathrm{2}}}.\end{array}$$

• The minimal frequency of the signal spectrum is computed as

  $$\begin{array}{}\text{(7)}& {\displaystyle }{f}_{\text{min}}=\underset{f}{min}\left(\text{PSD}\right(f)<\mathrm{0.2}\times max(\text{PSD}\left)\right).\end{array}$$

• The maximal frequency of the signal spectrum is computed as

  $$\begin{array}{}\text{(8)}& {\displaystyle }{f}_{\text{max}}=\underset{f}{max}\left(\text{PSD}\right(f)<\mathrm{0.2}\times max(\text{PSD}\left)\right),\end{array}$$

  which is the maximal frequency of the signal spectrum f_max (not to be confused with parameter F_max defined above).

The signal features are always computed on the trace with the maximal amplitude passed in the band [f_c–50] Hz (f_c: natural frequency). This enables is to limit the influence of the wave propagation and to compare signals with different sampling frequencies (i.e., 120 to 1000 Hz).

Based on already published events and further interpretations, we propose a standard classification of landslide endogenous seismic sources. The unpublished datasets are used to investigate the presence of these signals at other sites and to increase the number of examples for different contexts. Numerous signals were analyzed to draw the proposed classification and selected examples are further presented to describe the different classes.

6 Seismic description of the signals – typology

The typology of the signals is based on the duration and the frequency content of the seismic signals. The signals are classified into three main classes: “slopequake” (SQ), “rockfall” (RF) and “granular flow” (GF). For slopequake, subclasses are proposed and discussed based on the frequency content of the signals. Several examples of signals recorded at different sites are presented and the sources are discussed in the corresponding section.

Figure 1 Conceptual scheme of the landslide endogenous seismic sources with (a) wet granular flow, (b) dry granular flow, (c) rockfall, (d) tensile fracture opening, (e) tensile cracks opening, (f) shearing and (g) fluid migration in fracture.

Figure 2 Example of one controlled rockfall (mass = 430 kg) at the Riou Bourdoux catchment (Hibert et al., 2017a) recorded by a SP seismometer located at 50 m from the rock departure (a) and recorded by a BB seismometer near the rock arrival (b). The waveforms of the vertical traces are plotted in the upper part of the figure. The amplitude are normalized on the trace with the maximal amplitude (black), the signal recorded by the other sensors (when available) are represented in color below. The maximal amplitudes (A_max) of all the traces are plotted on the subplot in nm s⁻¹. The spectrogram is plotted on the middle part of the figure and normalized to the maximal energy. The lower part of the figure represents the PSD of the most energetic trace and the frequency corresponding to the maximum and the mean of the PSD are plotted in red and gray, respectively.

6.1 Rockfall (RF)

Figure 2 displays the seismic waves recorded for a single block fall at the Riou Bourdoux catchment (French Alps). The block was manually launched in the catchment and recorded with seismic sensors and cameras (Hibert et al., 2017a). The signal is characterized by successive impacts visible both on the waveform and on the spectrograms and lasts around 20 s. The spectral content contains mostly frequencies above 10 Hz but energy below 10 Hz is present for certain impacts (Fig. 2a). At closer distance, very high frequencies can be recorded up to 100 Hz (Fig. 2a). The autocorrelation remains large over time due to the similitude of the individual impact signals (T_corr>10 %). P and S waves are hardly distinguishable on the record and the signals recorded at the seismic sensors are dominated by surface waves (Dammeier et al., 2011; Helmstetter and Garambois, 2010; Hibert et al., 2014a; Levy et al., 2015).

Figure 3 Rockfall events recorded at (a) and (d) Super-Sauze (France; Provost et al., 2017a), (b) at the Séchilienne (France; Helmstetter et al., 2011; RESIF/OMIV, 2015), (c) Chamousset (Levy et al., 2011), (e) Aaknes and (f) Mount Saint-Eynard slopes (Le Roy et al., 2018). See Fig. 2 for a description of the figure.

Seismic signals of natural masses detaching from cliffs are presented in Fig. 3. They present similar characteristics to the artificially triggered rockfall. Depending on the height of the cliff, the signal ranges from 5 seconds up to tens of seconds. The symmetry of the signal ranges from 0 to 80 % depending on the cliff configuration. In general, the most energetic impacts are recorded at the middle or after the middle of the signal (dissymmetry coefficient >50 %). The highest measurable frequency depends on the source-to-sensor distance and can be very high (>100 Hz). The spectral energy is concentrated in frequencies above 5 Hz, with the largest PSD values (F_max) ranging from 20 to 40 Hz. Generally, the PSD energy is low below 10–15 Hz with the exception of one case (Fig. 5c) where spectral energy can be observed. The initial falling masses can themselves break into smaller units during propagation. In this case, the signal does not return to the noise level between the impacts due to developing granular flow (Fig. 3b, e, f) leading to a decrease in the duration of the autocorrelation of the signal. When several blocks are falling at the same time, impacts may overlap, and so do the peaks of the signals. In certain cases, the first rock free fall is preceded by a signal that can be associated with the rock detachment. An example of this precursory signal can be observed in Fig. 3a, f and in the data reported by Hibert et al. (2011) and Dietze et al. (2017b). The seismic signals of rockfalls contain information on the physics of the process. The seismic energy of rockfall signals is proportional to the volume (Farin et al., 2014; Hibert et al., 2014a). Scaling laws have also been established between seismic energy, momentum, block mass and velocity before impacts (Hibert et al., 2017a). The frequency content is mainly controlled by the block mass. The frequency of the spectral maximum energy decreases when the block mass increases (Burtin et al., 2016; Farin et al., 2014; Huang et al., 2007). If the rockfalls are well isolated, each impact generates impulsive waves. In the case of multiple rockfalls or short distances between the seismic sources and the sensors, the first arrivals may be emergent due to simultaneous arrivals of waves generated by impactors of different sizes impacting the ground at closely spaced time intervals (Hibert et al., 2014a; Levy et al., 2015).

Figure 4 Dry granular flow events recorded at (a) Séchilienne and (b) the Piton de la Fournaise caldera. See Fig. 2 for a description of the figure.

6.2 Granular flow (GF)

Granular flows are characterized by cigar-shape signals lasting between tens to thousands of seconds. They are subdivided in two classes.

• Dry granular flow (Fig. 4): These signals are characterized by cigar-shape waveforms of long duration (<500 s). Due to the absence of water, the source generally propagates over small distances. The duration of autocorrelation is very weak (T_corr≈0 %) and no seismic phase can be distinguished. No distinguishable impacts can be observed in the waveform nor in the spectrogram unlike rockfall signals. The signal onset is emergent, P and S waves are hardly distinguishable and the signal is dominated by surface waves (Dammeier et al., 2011; Deparis et al., 2008; Helmstetter and Garambois, 2010; Hibert et al., 2014a; Levy et al., 2015). The dissymmetry coefficient of the signal varies between 30 % and 75 % and depends on the acceleration and the volume of mass involved in the flow through time (Hibert et al., 2017b; Levy et al., 2015; Schneider et al., 2010; Suriñach et al., 2001, 2005). The frequency ranges from 1 to 35 Hz. The maximal frequency of the PSD varies between 5 and 10 Hz and can be larger (up to 20 Hz) when the seismic sensors are located close to the propagation path. The PSD values are significantly low below 3 Hz and increase rapidly between 3 and 20 Hz.

• Wet granular flow (Fig. 5): These signals range from several thousands of seconds to several hours and correspond to debris flows. They occur during rainfall episodes when fine material and boulders propagate downstream over long distances (>500 m). Like dry granular flow, the duration autocorrelation is very weak (T_corr=0 %) and no seismic phase can be distinguished. The seismic sensors are often installed at very close distance to the flow path so high frequencies up to 100 Hz may be recorded (Abancó et al., 2014; Burtin et al., 2016; Walter et al., 2017). Little energy is present in the low frequencies (<10 Hz) depending on the amount of water and the size of the rocky blocks integrated in the flow (Burtin et al., 2016). The signal is emergent and the amplitude variation depends on the mass involved in the flow passing in the vicinity of the sensor. Debris flows are very often divided into a front with the largest boulders and the highest velocity followed by a body and a tail where the sediment concentration and the velocity decrease (Pierson, 1995). The seismic signal amplitude hence increases progressively as the front is passing in the vicinity of the sensor (Abancó et al., 2012; Burtin et al., 2016; Hürlimann et al., 2014; Walter et al., 2017), and decreases progressively as the front is moving away from the sensor (dissymmetry coefficient >50 %). Large spikes and low frequencies may be observed in the seismic signal corresponding to the front of the debris flow generated by large boulder impacts. The frequency content also changes and, progressively, energy in the lower frequencies decreases (Fig. 5a).

Figure 5 Wet granular flow events recorded at Rebaixader torrent (Abancó et al., 2012; Arattano et al., 2016; Hürlimann et al., 2014). See Fig. 2 for a description of the figure.

6.3 Slopequake (SQ)

The “slopequake” class gathers all the seismic signals generated by sources located within the slope at the subsurface or at depth such as fracture-related sources or fluid migration (cf. Sect. 2). Different names have already been proposed for this kind of signals: “slidequakes” (Gomberg et al., 2011), “micro-earthquake” (Helmstetter and Garambois, 2010; Lacroix and Helmstetter, 2011), “quakes” (Tonnellier et al., 2013; Vouillamoz et al., 2018) or “landslide micro-quake (LMQ)” (Brückl, 2017). We here proposed the term “slopequake” as a general name for these events. They are characterized by short duration (<10 s) and are subdivided into two classes “simple” and “complex”.

Figure 6 Low-frequency slopequakes recorded at the (a) Slumgullion (Gomberg et al., 2011), (b) Pont Bourquin, (c) La Clapière and (d) Aiguilles-Pas de l'Ours slopes. See Fig. 2 for a description of the figure. Note that for the Slumgullion signals (a), the amplitudes are expressed in counts.

6.3.1 Simple slopequake

“Simple slopequake” signals are of short (<2 s) to very short duration (<1 s). Their main feature is the triangular-shape of the spectrogram with largest amplitudes being recorded in the first part of the signal (dissymmetry coefficient <50 %). The first arrivals contain the highest frequencies of the signal and are followed by a decrease in the frequencies. Depending on the frequency content, these signals can be subdivided into three classes.

• Low-frequency slopequake (LF-SQ) (Fig. 6): The signal lasts between 1 and 5 s. The maximal amplitude of the signal waveform occurs at the beginning or at the center of the signal (15 % < dissymmetry coefficient <50 %). The waveform presents only one peak and most of the first arrivals are emergent. Phase onsets are difficult to identify. The signals are mostly dominated by surface waves. Consequently, the duration autocorrelation of the signals is large (>10 %). The largest PSD values are observed between 5 and 25 Hz with a mean frequency ranging between 10 and 15 Hz.

• High-frequency slopequake (HF-SQ) (Fig. 7): The signal lasts between 1 and 5 s. The maximal amplitude of the signal waveform occurs close to the beginning of the signal (dissymmetry coefficient <30 %). The waveform presents only one peak and the first arrivals are mainly impulsive. Different phases may be observed (Lévy et al., 2010; Spillmann et al., 2007): P arrivals are detected at the beginning of the signal and correspond to the high-frequency waves, surface waves are then observed at the time the frequency decreases. However, in general the short sensor-to-source distance makes the differentiation between the different seismic phases difficult. The autocorrelation of these signals is hence lower than for LF-SQ (<10 %). In most of the cases, the picking of the different wave onsets is made difficult because of the sensor-to-source distances and the low-frequency sampling. The largest PSD values are observed between 3 and 45 Hz with a mean frequency ranging between 20 and 30 Hz.

• Hybrid slopequake (Hybrid-SQ) (Fig. 8): The signal lasts between 1 and 2 s. It presents the characteristics of the two precedent signals. The brief first arrivals are very impulsive and last less than 1 second. They are followed by a low-frequency coda similar to the LF-SQ. The maximal amplitude of the signal waveform occurs close to the beginning of the signal (dissymmetry coefficient <40 %). The waveform presents only one peak and the first arrivals are impulsive.

Figure 7 High-frequency slopequakes recorded at the (a) Super-Sauze (Provost et al., 2017a), (b) Séchilienne (Helmstetter et al., 2011; RESIF/OMIV, 2015), (c) Pont Bourquin, (d) La Clapière, (e) Aaknes and (f) Slumgullion (Gomberg et al., 2011) slopes. See Fig. 2 for a description of the figure. Note that for the Slumgullion signals (f), the amplitudes are expressed in counts.

These signals are suspected to be associated with boundary or basal sliding (Gomberg et al., 2011; Helmstetter and Garambois, 2010; Tonnellier et al., 2013; Walter et al., 2013b) or fracturing of the slope (Colombero et al., 2018; Helmstetter and Garambois, 2010). Currently, only few studies have proposed inversion of the source tensor (Lévy et al., 2010). To the best of our knowledge, for soft-rock landslides, no source mechanism was modeled. Therefore, it remains difficult to see if the observation of LF- and HF-SQs is due to attenuation of the high frequencies with the distance or to the source mechanism. Indeed, the rupture velocity may explain the difference in frequency content and low-frequency earthquakes are observed on tectonic faults (Brown et al., 2009; Shelly et al., 2006; Thomas et al., 2016). They are characterized by low magnitude (M_w<2) and short duration (<1 s) and constitute at least part of the seismic tremor signal. Therefore, the main assumption for the source of these events is slow rupture (Thomas et al., 2016). Another interpretation for the low-frequency quakes dominated by surface waves is crevasse opening (at the surface) as observed in glaciers (Deichmann et al., 2000; Mikesell et al., 2012). Colombero et al. (2018) analyzed AE at laboratory scales generated during thermal fracturing. During this experiment, high-frequency AEs are recorded during the heating stage up to the failure of the rock sample and are interpreted as thermal cracking events (Colombero et al., 2018). Low-frequency AEs are recorded during the cooling stage (after failure) and are associated with stick-slip events (Colombero et al., 2018).

Figure 8 Hybrid slopequake recorded at the (a) Pechgraben and (b) Super-Sauze landslide. See Fig. 2 for a description of the figure.

Hybrid slopequakes are very similar to the events recorded on volcanoes and glaciers with the presence of fluids in conduits or crevasses (Chouet, 1988; Helmstetter et al., 2015b). The sources of these events are assumed to be related to hydro-fracturing. The first high-frequency events corresponding to a brittle failure is followed by water flow into the newly opened cracks (Benson et al., 2008; Chouet, 1988).

The frequency content depends on the sensor-to-source distance and on the source mechanism. Observation of LF- and HF-SQ may be the signature of ongoing processes taking place within the slope instabilities justifying the three proposed classes for simple slopequakes.

Figure 9 Examples of slopequakes with precursory events recorded at the (a) Super-Sauze, (b) Séchilienne and (c) Chamousset slopes. See Fig. 2 for a description of the figure.

6.3.2 Complex slopequake

The second class of short-duration signals has the same general properties as the simple slopequakes but exhibits particular frequency content or precursory events. These additional characteristics change the possible interpretation of the sources. Consequently, these signals are gathered in the class “complex slopequake”. Three different subclasses are proposed:

Figure 10 Examples of tremor-like slopequakes recorded at the (a, c) Super-Sauze, (b) La Clapière and (d) Aiguilles-Pas de l'Ours slopes. See Fig. 2 for a description of the figure.

• Slopequake with precursors (Fig. 9): The third class of short duration signals are similar to the slopequake signals but are preceded by a precursory signal of smaller amplitude (Fig. 9). The content of the precursory signal ranges from 5 to 100 Hz depending on the site and is slightly lower than the highest frequency generated by slopequake-like event. The precursory arrival lasts up to 1.2 s in the presented examples and no clear phases are detected. The frequency content ranges from 5 to 100 Hz but varies significantly at each site. At all sites, the amplitude of the signal is significantly higher for one of the sensor (3 to 50 times higher) when considering vertical traces. The precursory signal is buried in the noise at the sensors with lowest amplitudes and the signal is similar to a LF-SQ. Such events have never been documented to our knowledge. They are likely to be generated by a strong and local source located in the very close vicinity of one of the sensors (<10 m) due to the maximal amplitude (>10⁵ nm s⁻¹) and the rapid decrease in the amplitude recorded by the other sensors. Although the signal is similar to certain earthquakes (the precursory signals interpreted as P-wave arrivals and the strong arrivals as surface waves), no earthquake location can explain the signal recorded at the time these events are recorded. Their occurrence in the nighttime makes a human activity unlikely to be the source. The most probable source would then be the detachment of a single block and its fall in the vicinity to one of the sensors. This kind of precursory signal is observed for some rockfalls (Fig. 3a) and at a the Saint-Martin-le-Vinoux quarry (France; Helmstetter et al., 2011). At the Saint-Martin-le-Vinoux underground quarry, the duration between the detachment and the signal impact is well correlated to the room height. This interpretation is coherent with the drop of amplitude before the more energetic event at the Chamousset rock column (Fig. 9c), where a progressive decrease in the precursory signal is observed. However, at the other sites (Fig. 9a, b) such a decrease is not present. The 1-second-long precursory signal has a constant amplitude and frequency content. Another interpretation could be that these precursory signals are a succession of overlapping slip or fracture events. The interpretation of these signals cannot be established with certainty and further analysis (i.e., location, time of occurrence) and other examples are needed to discriminate the mechanism at work.

• Tremor-like slopequake (Fig. 10): The last class of short-duration signals often last between 1 and 5 s (Fig. 10). They present a symmetrical waveform (S=50 %) with emergent arrivals and a slow decrease in the amplitude to the noise level. The frequency ranges from 5 to 25 Hz. High frequencies may be briefly recorded in certain events (Fig. 10c). The maximal energy of the PSD corresponds to a frequency of 8 to 13 Hz, while the mean energy corresponds to a frequency of 13 to 17 Hz. No seismic phases are identified. The signal is not recorded by all the sensors even when the sensors are organized in small arrays with short inter-sensor distances (<50 m). Their waveforms and frequency content are similar to the one of the granular flows (Fig. 4). Small debris flows have been observed at La Clapière and Super-Sauze landslides and are likely to generate seismic waves; however, small debris flows are not observed at the Pas de l'Ours landslide when these kinds of seismic signals are recorded. Another possible source mechanism for such events may also be a very rapid succession (<1 s) of shear events along the basal or the side bounding strike-slip faults (Hawthorne and Ampuero, 2017). Further investigations are needed to analyze their occurrences over time and their location to confirm one or the other assumption.

Figure 11 (a) Summary of the proposed classification with plots of the attributes for the examples presented in the precedent figures and an example waveform for each class. The convention for the attribute plot is presented in (b); n being the number of seismic signals examples used to plot the feature diagram.

7 Discussion

The proposed typology is summarized in Fig. 11. The approach consisted of comparing the datasets of different sites in order to identify the common features of the recorded seismic signals. Three main classes can be differentiated mainly from the length of the signals, the number of peaks and the duration of the autocorrelation. Figure 12 shows more examples of the signal variability for the sites where long seismic catalogs have been recorded (e.g., Aaknes, Chamousset, Séchilienne, Super-Sauze and La Clapière). Only the signals classified as rockfall, LF-SQ and HF-SQ are presented because fewer events of the other classes are present in the investigated datasets. The signal features are in good agreement with the defined classes proposed in the present classification (Fig. 11). In general, narrow variability is observed on the feature values among the different sites and consequently, the observed features are likely associated with the source mechanism.

Figure 12 Variability in the signal features of classes “rockfall”, “HF-SQ” and “LF-SQ” for five different sites: Aaknes, Chamousset, Séchilienne, Super-Sauze and La Clapière. The axes of the star diagrams are the same as in Fig. 11.

However, some variability exists for rockfall events. Indeed, the volume of the blocks and possible breaks control the frequency content and the autocorrelation duration while the height of the scarp will play a significant role in the duration of the event. Depending on the site, rockfall signals can have similar features to each other (e.g., Séchilienne, Fig. 12) suggesting a constant source mechanism or very different features suggesting the presence of different rockfall mechanisms and/or trajectories (e.g., Super-Sauze, Fig. 12). In the case of the Super-Sauze datasets, rockfall are characterized by a lack of energy in high frequencies due, in this case, to the distance between the seismic network and the scarp. Installation of additional sensors could be the easiest way to get rid of this variability. It must also be noted, that differentiating flow and fall signals may be challenging. Indeed, some of the events are very likely a mix of these two sources. Rockfalls of various blocks may generate granular flows with metric block impacts, both overlapping in the recorded seismic signals. Presence of metric rocks is also observed in debris-flow-prone torrents; for this type of event, the block impacts within the mass flows are recorded in the seismic signals (Burtin et al., 2016).

Figure 13 Examples of pure harmonic signals recorded at the (a) Pechgraben, (b) La Clapière, (c) Aiguilles-Pas de l'Ours, (d) Séchilienne, (e) Slumgullion (Gomberg et al., 2011) and (f) Super-Sauze slopes. See Fig. 2 for a description of the figure. Note that for the Slumgullion signals (e), the amplitudes are expressed in counts.

Our analysis does not allow us, at this stage, to conclude whether the frequency content of the simple slopequakes is associated with the source mechanism because complete catalogs differentiating between HF-SQ and LF-SQ are not yet available. Colombero et al. (2018) suggested that HF-SQs are the dominant class of slopequake at the Madonna del Sasso cliff (hard rock) and were generated by thermal cracking, while LF-SQs associated with frictional sliding are less frequent. Although we did not investigate the whole dataset, no LF-SQs were provided at the Aaknes or Chamousset hard-rock cliffs (Fig. 12), while LF-SQs are recorded at the La Clapière and Séchilienne hard-rock slides: (Fig. 12). This observation seems to confirm the results of Colombero et al. (2018). However, further comparison of the occurrence of the different slopequakes at specific sites in space and time must be done to improve the comprehension of these sources and confirm this statement.

Harmonic signals have also been documented at the Pechgraben and Super-Sauze landslides (Vouillamoz et al., 2018). These signals last from 1 to 5 s and may repeat during minute-long sequences. The proposed interpretation includes hydro-fracturing or repetitive swarms of microearthquakes (Vouillamoz et al., 2018), while Gomberg et al. (2011) hypothesized that these kinds of signals were caused by trapped waves along the side-bounding strike-slip fault generated by shear events. In the investigated datasets, similar signals are recorded at the La Clapière and the Aiguilles landslides with a fundamental frequency of 8±1 Hz (Fig. 13b, c). At the Séchilienne landslide, harmonic signals are also detected (Fig. 13d), mostly during the day, with different resonant frequencies between 2 and 12 Hz. Similar signals are observed at the Slumgullion and Super-Sauze sites but without clear harmonics in the PSD (Fig. 13e, f). The presence of pipes and drains on or in the vicinity of these sites could also explain the origin of these signals justifying that these signals are not included in the slopequake class as they may not be generated by a slope deformation process. The location of the source, the distribution of the amplitude, the stability of the fundamental frequency and the daily temporal occurrence of the source supports this assumption. Systematic location of these events is needed to determine if they must be integrated or not in the general typology in the case that they are generated by fluid resonance in fractures.

For certain signals, the coda is dominated by resonance frequencies (Figs. 3d, 9c) at high frequencies (i.e., 20 and 43 Hz), and well observed in the spectrogram of the signal. The resonance is not present before the beginning of the signal and hence can not be due to anthropogenic noise (i.e., motors). In the case of Chamousset cliff, Levy et al. (2011) explained the presence of this monochromatic coda by the resonance of the rock column after the occurrence of the rock bridge breakage. At Super-Sauze, a similar resonant coda is observed at the end of certain rockfalls (Fig. 4d). Considering the distance between the main scarp and the seismic arrays (>300 m), and the absence of a large fracture on the scarp, the occurrence of this kind of resonance is very surprising in this case. This signal feature could also result from the wave propagation (i.e., trapped waves).

No long-lasting tremors are presented in this study. Schöpa et al. (2018) recorded a tremor with gliding before the occurrence of the Askja caldera landslide. Similar tremors have been found on the Whillans ice stream in Antarctica during slow slip events (Lipovsky and Dunham, 2016; Paul Winberry et al., 2013), which repeat twice a day with a slip of about 10 cm lasting for about 20 min. Therefore, such signals may also occur during the nucleation phase of landslide failure. The question remains if they are not observed because landslide acceleration is aseismic due to high pore fluid pressure (Scholz, 1998) or low normal stress at the subsurface of the slope.

Difficulties still arise in providing an exhaustive description and interpretation of all the sources from the simple analysis of the proposed signal features, particularly those generating short-duration signals, in particular for the classes discriminated by the frequency content such as LF-SQ and HF-SQ. The ambiguity between propagation effect and source mechanisms prevents further interpretation due to several limitations. Firstly, the location of the sources remain difficult to establish due to the complexity of some of the signals (Gomberg et al., 2011; Lacroix and Helmstetter, 2011; Provost et al., 2018; Tonnellier et al., 2013), the size of the instrumented sites and the complexity of the underground structure that influences the polarization of the waves (Neuberg et al., 2000) and the sensors (ie. number, location and type: 1C/3C sensor) installed close to the unstable slopes (Godano et al., 2009). The location of the epicenter of most of the events seems coherent with the instability deformation field at the surface (Helmstetter and Garambois, 2010; Levy et al., 2011; Provost et al., 2018; Walter et al., 2013b), although resolving dispersion and 3-D heterogeneities of the velocity fields currently prevents us from inferring the depth of the events and their focal mechanisms. Secondly, a complementary approach to explain the origin of the sources is the analysis of their occurrence with respect to surface or basal displacement and monitoring of the water content and pore fluid pressures. It requires both exhaustive catalogs of landslide seismicity over long time periods and continuous and distributed datasets of displacements and pore fluid pressures, which remain challenging to acquire. Finally, in addition to the characteristics of seismic signals, further information on the sources processes can be obtained from the distribution of the events in time, space and size. Events that occur regularly in time with similar amplitudes are likely associated with the repeated failure of an asperity surrounded by aseismic slip, for instance, at the base of a glacier (Helmstetter et al., 2015a) or of a landslide (Poli, 2017; Yamada et al., 2016a). Signal amplitudes and recurrence times often display progressive variations in time. In contrast, events that are clustered in time and space, with a broad distribution of energies, are more likely associated with the propagation of a fracture (Helmstetter et al., 2015b). The daily distribution of an events time can also be helpful to identify anthropogenic sources, which occur mostly during the day. In contrast, natural events are more frequently detected at night, when the noise level is lower.

Simulations and models are also required to explain the current observations. Indeed, experimental results suggest an increase in AEs correlated with an increase in the slope velocity (Smith et al., 2017) or an increase in AE due to the creation of the rupture area (Lockner et al., 1991). Acceleration of pre-existing rupture surface(s) seems to be the mechanism responsible for the seismicity recorded before large rockslide collapse. Yamada et al. (2016a) and Poli (2017) argued that the high correlation between the repetitive events could only be explained by stick-slip movement of the locked section(s), while a cracking process would imply a migration of the location of the events and a change in the events waveforms. Schöpa et al. (2018) argued that the presence of gliding frequencies could only be produced by similar sources and hence close location. On the contrary, in the case of the Mesnil-Val column, Senfaute et al. (2009) interpreted the evolution from high-frequency to low-frequency events as the progressive formation of the rupture surface followed by the final rupture process immediately before the column collapse, where both tensile cracks and shearing motion on the created rupture are generated.

8 Conclusions

Over the last few decades, numerous studies have recorded seismic signals generated by various types of landslides (i.e., slide, topple, fall and flow), for different kinematic regimes and rock/soil media. These studies demonstrated the added value of analyzing landslide-induced MS to improve our understanding of the mechanisms and to progress in the forecast of landslide evolution.

In this work we propose a review of the endogenous seismic sources generated by the deformation of unstable slopes. A dataset of 14 slopes is gathered and analyzed. Each of the sources are described by nine quantitative features of the recorded seismic signals. Those features provide distinct characteristics for each type of source. A library of relevant signals recorded at relevant sites is shared as supplementary material. We propose three main classes “slopequake”, “rockfall” and “granular flow” to describe the main type of deformation observed on the slopes. Slopequakes are related to shearing or fracturing processes. This family exhibits the most variability due to the complexity of the sources. These variations are likely to be generated by different source mechanisms. “Rockfall” and “granular flow” classes are associated with mass propagation on the slope surface. They are distinguishable by the number of peaks clearly identified in the seismic signals.

Presently, several descriptions of the seismic sources are proposed for each study case. We believe that a standard typology will allow researchers to discuss and compare seismic signals recorded at many unstable slopes. We encourage future studies to use and possibly enrich the proposed typology. This also requires publication of the datasets and/or catalogs to progress towards a common interpretation. Recently, organizations such as the United States Geological Survey (USGS) or the French Landslide Observatory (OMIV) have started this work (RESIF/OMIV, 2015).

Recent arrivals on the market of relatively cheap and autonomous seismometers (e.g., ZLand^® node systems, Raspberry Shake systems) will allow the deployment of denser seismic networks of 3C sensors. The latter will certainly improve the location accuracy and enable inversion of the focal mechanism of the sources. Moreover, the recent operational applications of ground-based SAR (synthetic aperture radar) and terrestrial lidar technologies for monitoring purposes shows their relevance to monitor distributed surface displacements. Ongoing monitoring on several landslides combining those innovative approaches will certainly help to associate slopequake events to deformation processes (Dietze et al., 2017b; Provost et al., 2017b; RESIF/OMIV, 2015).

The proposed typology will help to constrain the design of new models to confirm the assumptions on the nature and the properties of the seismic sources. This will be particularly important for (1) explaining the variability of the slopequake sources observed at the sites, (2) progressing in the physical understanding of the slopequake sources, and (3) ascertaining the spatiotemporal variations in the seismic activity observed at some unstable slopes in relation with their deformation, as well as with external forcings such as intense rainfalls and earthquakes.