This research has been supported by the “Topo-Europe project” and the “European Science Foundation (ESF)” called SedyMONT project (Timescales of Sediment Dynamics, Climate and Topographic Change in Mountain Landscapes). The authors acknowledge the valuable input from anonymous referees and GIE’s editorial team.

1. Introduction

1Each year, mass movements cause considerable financial damage to alpine societies (Hilker et al., 2009) and are a major driver of landscape changes and evolution by transferring sediment from sources to sinks (Guzzetti et al., 2005). They pose a serious threat to the populations by repeatedly destructing settlements, disrupting transportation corridors, or even leading to the loss of life. The occurrence of mass movements has recently become a topic of major interest for both research and administration, especially in relation with the assessment of landslide hazards and risks (Magliulo et al., 2008). The increasing interest in landslides certainly reflects the increasing awareness of the socio-economic significance of landslides (Aleotti and Chowdhury, 1999) but also indicates quite clearly that human pressure on the environment has become more important for land development and urbanization (Petrascheck and Kienholz, 2003). Global statistics show that damage from landslides has risen for the last 30 years in mountain areas (Alexander, 2008). This trend is linked both to an increase in the occurrence of hazardous events, attributed to climate change (Gariano and Guzzetti, 2016) and to larger populations living in constantly growing Alpine settlements (Petrascheck and Kienholz, 2003).

2An appropriate assessment of existing and potential future landslide hazards requires, among others, detailed determination of the spatial and temporal occurrences of landslides at the site level (Claessens et al., 2006; Thiery et al., 2007; Corominas and Moya, 2008). The spatial extent and failure type of individual landslides can be further clarified via field investigations and surveys of tension cracks, secondary scarps, and scars around the landslide (Cruden and Varnes, 1996). Landslide activity can be inferred from multi-temporal remote sensing measurements or repeat field surveys (Cruden and Varnes, 1996; Montgomery et al., 2000).

3However, such data are not normally available with satisfying spatial resolution over long enough timescales and as a continuous record, landslide chronologies thus play a key role because they provide very essential information on past activity and thereby contribute substantially to hazard assessment (Šilhán and Stoffel, 2015). In previous studies, historical records were reconstructed for single landslides or landslide-prone regions and estimates were usually derived from existing archives such as narrations, historical documents, terrestrial or aerial photographs, remote sensing series, incidental statements or, more rarely, from public historical databases (Brunsden et al., 1976; Coe et al., 2000; Crovelli, 2000; Martin et al., 2002). Yet, the temporal window of such records only rarely spans more than a few decades and almost never covers centuries (Lopez-Saez et al., 2013a). In addition, and even more importantly, archival data on landslides have not normally been recorded for geomorphic purposes. As a result, they lack spatial completeness, resolution, and precision and invariably emphasize events that caused damage to human structures (Mayer et al., 2010). At the same time, they tend to typically overrepresent the largest and will, at the same time, miss the smaller events (Carrara et al., 2003). Finally, considerable problems exist in interpretation because of the changing standards and criteria of reporting in archival records over time (Ibsen, 1996). As a consequence, past research thus focused more on landslide susceptibility, see Guzzetti (2000), and references therein for a review, rather than on the documentation of landslide hazards.

4To compute accurate probability maps of landslide reactivation at the local scale, available for disaster prevention and the generation of risk maps, an approach is thus required that offers both an adequate temporal and spatial resolution. In forested shallow landslides, the analysis of growth disturbances contained in tree-ring records (Alestalo, 1971; Astrade et al., 2012; Stoffel and Corona, 2014) can greatly help the documentation of past events and may allow reconstruction of accurate chronologies of landslide reactivations over considerable periods in the past (Lopez-Saez et al., 2013a; Corona et al., 2014). As tree-ring series provide a continuous record over the lifespan of the tree and, collectively, over the lifespan of the sampled population (Procter et al., 2011), they offer a unique spatio-temporal resolution of past process activity. According to Carrara and O'Neill (2003), the first investigators using tree rings to date landslides were McGee (1893) in Tennessee and Fuller (1912) in Mississippi. However, modern dendrogeomorphology dates back to the early 1970s (Alestalo, 1971) and the information contained in tree-ring records has been used extensively in the United States (Jensen, 1983; Carrara, 2007) ever since. In Europe, tree rings have been used to document landslide reactivations in the French (Astrade et al., 1998; Lopez Saez et al., 2011, 2012, 2013a) and Italian Alps (Fantucci, 1999; Stefanini, 2004), the Spanish Pyrenees (Corominas and Moya, 1999), the Flemish Ardennes (Van Den Eeckhaut et al., 2009) or more recently in Czech Republic (Silhan et al., 2016; Silhan, 2017). Whereas these studies focused on the overall activity or possible triggers of landslides, they did neither define the temporal frequency of reactivation for specific areas nor address the probability of future events to occur in certain compartments on the landslide body. However, the location of past and potential future landslide reactivation along with a detailed assessment (i.e. annual resolution) of actual landslide triggers appears key for a better understanding of the process and for the management of sites at risk. The purpose of this study therefore is to provide a high-resolution, spatio-temporal chronology of reactivations on a forested, shallow landslide body located in the Schimbrig area (Entlebuch, Canton of Lucerne, Switzerland). The specific goals of this study are (i) to derive periods of landslide reactivation with annual resolution using the dendrogeomorphic record of 197 spruce trees (Picea abies (L.) Karst.). In addition, (ii) single-point data on past landslides was then compiled to derive a high-resolution landslide return period map for 24 geomorphic compartments identified on the landslide body and, in a final step, (iii) to quantify and map the probability of landslide reactivation for the coming 5, 20, 50, and 100-yrs, using a Poisson distribution.

2. Regional setting

5The study area is the Schimbrig landslide (8°5'58"E; 46°56'39"N; 1100-1440 m) located in the northern foothills of the central Swiss Alps (Clapuyt et al., 2015), in the UNESCO world heritage site of Entlebuch (Canton of Lucerne). The landslide is part of the Entlen catchment whose outlet is located near the town of Entlebuch. It consists in a complex 5‑10 m thick earth slide (Schwab et al., 2008), almost entirely lying in the Flysch domain, which promotes hillslope instability due to the low mechanical strength of the geological material (fig. 1E). With a maximum length of 1,280 m and a maximum width of 530 m, the landslide body occupies a surface of approximately 300,000 m² with an average slope of 16°. The landslide body is made up of cm- to dm-large clasts that are embedded in a matrix of silt and mud. The head of the landslide is translational, whereas the central part and the toe show sinusoidal downslope profiles. The flysch landslide mass currently moving is located between the Schimbrig ridge, formed by Early Cretaceous and Early Tertiary suite of marls, limestones, siliceous limestones, and quartzites of the Helvetic thrust nappes, and two longitudinal NE‑SW striking compartments made up of Molasse conglomerates (Mollet, 1921). According to Schwab et al. (2008) and Savi et al. (2013), a detailed field survey of the local morphology reveals that the landscape of the Schimbrig area is currently modulated by complex slab slides with shallow surface ruptures, multiple cracks and creeping surfaces.

6Climatic conditions at Schimbrig are characterized by a subarctic climate without dry season « Dfc » according to the Köppen-Geiger classification (Rubel and Kottek, 2010). Present-day climatic conditions are humid with average precipitation rates of approximately 1,500 mm.yr^‑¹ (MeteoSwiss rain gauge at Entlebuch). Mean annual temperature is 4.6°C and, on average, the maximum temperature is below the freezing point 60 days per year (Schwab et al., 2008).

7The upper area of the landslide is characterized by steep slopes (unsuitable for agriculture) and partially covered by forest. The tree stand is mainly composed of spruce (Picea abies (L.) Karst.) intermixed with younger deciduous trees. Spruce forms nearly homogeneous stands outside the surfaces affected by the scarps and recent earth slides. The tilted and deformed trees also clearly indicate that the Schimbrig landslide has been affected by multiple reactivations in the past (fig. 1B-C).

8Contrariwise, the lower part is characterized by gentle slopes, allowing cattle and sheep to graze. During the past centuries this area has been mainly covered by meadows (Savi et al., 2013). From a historical perspective, landslide activity at Schimbrig has been documented since the beginning of the 1920s (Mollet, 1921). The earth slide has experienced a period of intense activity between spring 1994 and May 1995 when slip rates reached maximum values of several meters per day (Savi et al., 2013)

3. Material and methods

3.1. Collection and preparation of samples

9In a preliminary step, 24 geomorphic units were identified on the landslide body based on the topography of the slope, sedimentary fabric of deposits as well as the presence and density of cracks, backscars of secondary degradational landslides or undulating ground associated with secondary landslide movement. All these features were mapped at a scale of 1:1,000 (fig. 1D). At each unit, P. abies trees damaged by past landslide activity were sampled based on an outer visual inspection of the stem. Four cores per tree were extracted: two in the supposed direction of landslide movement (i.e. upslope and downslope cores), and two perpendicular to the slope. To gather the greatest amount of data on past events, trees were sampled within the tilted segment of the stems (Stoffel and Bollschweiler, 2008). To avoid misinterpretation, trees growing in sectors influenced by processes other than landslide or anthropogenic activity (sylviculture) were not considered for analysis. A total of 197 disturbed P. abies trees were sampled resulting in a total of 788 increment cores.

10For each tree, additional data were collected, such as (i) tree height; (ii) diameter at breast height; (iii) visible defects in tree morphology, and particularly the number of knees; (iv) position of the extracted sample on the stem; (v) photographs of the entire tree; and (vi) data on neighboring trees (following Stoffel et al., 2005). Tree coordinates were obtained with an accuracy <1 m with a Trimble GeoExplorer GPS. In addition, 20 undisturbed P. abies trees located above the landslide scarps and showing no sign of landslide activity or other geomorphic processes were sampled to establish a reference chronology. Two cores per tree were extracted, parallel to the slope direction and systematically at breast height. The reference chronology represents common growth variations in the area (Cook, 1990) and enables precise cross-dating and aging of the cores sampled on the landslide body. The samples obtained in the field were analyzed and data processed following standard dendrochronological procedures (Stoffel and Corona, 2014). Single steps of surface analysis included sample mounting on a slotted mount, drying, and surface preparation by finely sanding the upper core surface up to grit size 600. In the laboratory, tree rings were counted and rings measured to the nearest 0.01 mm using a digital LINTAB positioning table connected to a Leica stereomicroscope and TSAP-WIN Scientific software (Rinntech, 2009). The reference chronology was developed based on the growth curves of the undisturbed trees using the ARSTAN software (Cook, 1985). The two measurements of each reference tree were averaged, indexed and detrended using a double detrending procedure (Holmes, 1994) with a negative exponential curve (or linear regression) and a cubic smoothing spline function (Cook, 1990). The quality of the cross-dating was evaluated using COFECHA (Holmes, 1983) as well as the graphical functions of TSAPWin (Rinntech, 2009). Growth curves of the samples of disturbed trees were then compared with the reference chronology to detect missing, wedging or false rings and to identify reactions to mechanical stress. As no significant correlation was found between the reference chronology and 26 cores from 13 affected trees, we rejected these samples for further analysis.

3.2. Age structure of the stand

11The age structure of the stand was approximated by counting the number of tree rings of selected trees (n = 184). However, since trees were not sampled at their stem base and the piths or innermost rings of several trees were rotten, the age structure might be biased; the map (fig. 1D) thus only reflects age at sampling height, but neither inception nor germination dates. Nonetheless, it provides valuable insights into major disturbance events at the study site with reasonable precision.

3.3. Sign of disturbance in the tree-ring record

12Landslide movement induces several kinds of growth disturbances (GD) to trees, most commonly in the form of an abrupt reduction in annual ring widths and/or the formation of compression wood on the tilted side of the stem. A reduction in annual ring widths over several years is interpreted as damage to the root system, loss of a major limb, or a partial burying of the trunk resulting from landslide activity (Carrara and O'Neill, 2003). In this study, growth-ring series had to exhibit (i) a marked growth reduction (GS) in annual ring width for at least five consecutive years such that the (ii) width of the first narrow ring was 50% or less of the width of the annual ring of the previous year. The onset of compression wood (CW) is interpreted as a response to stem tilting induced by landslide pressure. Tilted trees try to recover straight geotropic growth (Mattheck, 1993) through the development of asymmetric growth rings, i.e. the formation of wider annual rings with smaller, reddish-yellow colored cells with thicker cell walls (Timell, 1986) on the tilted side and narrow (or even discontinuous) annual rings on the opposite side (Carrara and O'Neill, 2003; Panshin and De Zeeuw, 1970). The production of chaotic callus tissue (Stoffel et al., 2010) and tangential rows of traumatic resin ducts (referred hereafter as to TRD; Stoffel, 2008) are the consequences of the injury after the corrasion of tree stems by debris (Stoffel and Hitz, 2008; Schneuwly et al., 2009a, b). If impacts locally destroy the cambium, incremental cell formation will become disrupted and new cell formation will cease in the injured segment of the tree (Stoffel and Corona, 2014). Finally, geomorphic processes do not only disturb trees in their growth, but large and devastating hydrogeomorphic events can also eliminate trees along channels or landslide bodies through uprooting and stem breakage while leaving their neighbors intact (Stoffel and Corona, 2014). The elimination of neighboring trees can result in a new environment with less competition, more light, nutrients and/or water. Survivor trees will benefit from the improved conditions and respond with a growth increase (GR) and wider tree rings. For this investigation, we used the appearance of TRD, the initiation of CW, the abrupt GS and GR to determine the occurrence of landslides.

3.4. Dating of events

13Determination of events was based on the number of samples showing GD in the same year and on the spatial distribution of disturbed trees on the landslide body (Lopez-Saez et al., 2013a). To avoid overestimation of GD within the tree-ring series in more recent years because of the larger sample of trees available for analysis, we used an index value (It) as defined by Butler and Malanson (1985):

where R is the number of trees showing a GD as a response to a landslide event in year t, and A is the total number of sampled trees alive in year t.

14Following disturbance by an initial event, a tree may not necessarily yield useful data on additional events for some time (i.e. a tree may already be forming a narrow band of annual rings such that a subsequent disturbance would not be detected) (Stoffel et al., 2010); this is why It was adjusted to only take account of trees with a useful record for year t (Carrara and O'Neill, 2003). A minimum of 10 trees exhibiting a response was required for a major reactivation to be dated so as to avoid an overestimation of relative response numbers resulting from a low number of trees early in the record (Corona et al., 2010, 2012, 2014). To minimize the risk that GD caused by other (geomorphic) processes could mistakenly be attributed to a landslide event and to take account of the sample size, the chronology of past events was also based on It >5%. However, the strictness of these thresholds and the large sample size may induce a misclassification of minor reactivation. To avoid misclassification, the annual patterns of disturbed trees for years with 5% >It >2% and GD in ≥5 trees were carefully examined (Lopez-Saez et al., 2012). Using GPS coordinates, trees were placed into a Geographical Information System (GIS; ESRI, 2005) as geo-objects, and GD were linked as attributes to each single tree. We computed autocorrelations based on the location and GD values of trees with the ArcGIS pattern analysis module and calculated yearly Moran indices (Moran, 1950) to evaluate whether the impacted trees were randomly distributed (Moran’s I = 0), dispersed (Moran’s I moving toward -1), or clustered (Moran’s I moving toward 1). Random or dispersed years were disregarded and years with a statistically significant (using the Z score with a p‑value <0.1); Moran’s I >0 were included as minor or spatially limited movements in the landslide chronology.

3.5. Calculation of landslide return periods and probabilities of reactivation

15Traditionally, the return period designates the mean time interval at which a material reaches a given point in an avalanche path (Corona et al., 2010), with frequency being usually expressed in years as a “return period” (i.e. 1/frequency). We adapted this approach and, by analogy, calculated individual return periods (Rp) for each of the 24 geomorphic units (u) identified on the Schimbrig landslide as follows:

16where P represents the period covered by the reconstruction at the geomorphic unit u, i.e. the time elapsed between the first event detected at u and 2010 and Events u the total number of events reconstructed at u.

17These return periods were used, in a second step, to perform a probabilistic landslide analysis. To investigate time series of natural events, different probability distributions have been adopted, including the Poisson, binomial, Weibull, and the mixed exponential distributions (Crovelli, 2000). Amongst these distribution models, the Poisson distribution was used frequently to investigate the temporal occurrence of landslides (Crovelli, 2000; Coe et al., 2000; Guzzetti et al., 2003, 2005, 2006) as it does not require, contrary to other distributions, a precise knowledge of the total possibilities, that is how often an event occurs and how often it does not occur (Lee and Jones, 2014). The theoretical probability for a landslide to occur at Schimbrig was therefore modeled, at each geomorphic unit, using a Poisson distribution and the probability p for an event e with a return period T to occur in a given number of years N (fixed to 5, 10, 20 and 100 years) was computed as follows:

18According to this distribution, the probability for a centennial event to occur during the next 100 years is, for example, p = 0.63.

4. Results

4.1. Age structure of the stand and growth disturbances

19After cross-dating, data on the pith age from 184 P. abies trees growing on the Schimbrig landslide suggest an average age of the sample of 109 ±43.5 yr. The oldest tree selected for analysis shows 290 rings at sampling height (AD 1720), whereas only 16 growth rings (AD 1994) were counted in the youngest tree. As can be seen from Figure 1D, the distribution of tree ages is heterogeneous and the forest stand is constituted by trees aged between 70 and 130 years with patches of old trees (>150 years) scattered within the stand.

20A total of 318 GD related to a past landslide reactivation was identified in the 184 disturbed trees. The most common reaction to landslide events was in the form of abrupt growth reductions (GS) with 47% of all GD (323 GD). The onset of compression wood (169 GD, e.i. 25%) and TRD (150, e.i. 22%) represents another common response of P. abies to landsliding. In contrast, growth release (GR: 42 GD, e.i. 6%) was by far less abundant. The earliest GD observed in the tree-ring series dates back to 1852; however, this year was not considered a landslide event as only one tree showed GD (fig. 3). In 1882, the number of GD surpassed five which was defined the threshold for GD to be considered as a landslide event.

4.2. Landslide reactivations and decadal landslide frequency

21In total, 27 years did exceed the 2% threshold for It with ≥5 trees exhibiting a GD between 1859 and 2010 (fig. 2). Major reactivations with GD>10 and It>5% were reconstructed in 17 different years, namely in 1897, 1917, 1928, 1930, 1940, 1945, 1947, 1955, 1956, 1973, 1983, 1993, 1994, 2001, 2002, 2003, and 2006 (fig. 2-3A). For the years 1859, 1860, 1882, 1890, 1903, 1906, 1922, 1936, 1962 and 1975, the number of GD was >5 and 5%> It >2%; these years could not be considered reactivation events with equal confidence and were therefore tested further with yearly Moran’s I statistics. Results point to a spatial clustering with sufficient aggregation in 1859 (0.55), 1860 (0.55), 1882 (0.37), 1890 (0.29), 1903 (0.36), 1906 (0.4), 1922 (0.42), 1936 (0.24), and 1962 (0.1); as a result, these years were considered events with minor landslide reactivation (fig. 3A). In 1975, Moran’s I statistics point to a dispersed distribution of affected trees (0.06) with no significant pattern; this year was not therefore kept for further analysis. Considering the 26 reactivations within the sampled area, the mean return period for the Schimbrig landslide is 0.17 event.yr^-¹ for the period 1859-2010. At the decadal scale (fig. 3B), 1.6 reactivations are recorded on average over the period covered by the tree-ring record. Above-average activity occurred in 1890-1899, 1900-1909, 1920-1929, 1930-1939, 1950-1959 and 1990-1999 (with 2 events for each of the periods), 1940-1949 (3 events) and 2000-2009 (4 events), whereas very limited landslide activity is observed for 1850-1869, 1880-1889, 1910-1919, and 1960-1989 with only one reactivation each. No event was recorded in 1870-1879. At the multi-decadal scale, landslide frequency does not show any clear trend. Whereas 12 landslides occurred between 1850 and 1939 (1.33 events per 10 yr), 14 events have been recorded since 1940 (2 events per 10 yr).

4.3. Return period and landslide probability maps

22When analyzed spatially, the return period of reactivations is lower in the northeastern part of the landslide body and reaches a clear minimum (13-20 years) in the geomorphic units 10, 22, 23 and 24. Conversely, they exceed 30 years in the geomorphic units 14-17 located on the southwestern margins of the earth slide to reach maximum values (>75 years) in units 3, 6, 15 and 17 (fig. 4A). Return periods of landslide reactivations were then transformed into landslide occurrence probability using a Poisson distribution. Highly resolved maps of return period derived from the 184 cross-dated P. abies trees were thereby used to represent the probability for a landslide reactivation to occur within 5, 20, 50, and 100 yrs (fig. 4B-E). As expected, the probability for a landslide to be reactivated increases from 0.33 for a 5‑year period to 0.77 for a 20-years and 0.99 for a 100-year period. On a spatial plan, the probabilities for an event to recur are highest in the northeastern part of the landslide. In other units, event probabilities are lower, yet they exceed 0.8 in 21 out of the 24 geomorphic units for the 100-year period.

5. Discussion

5.1. Reliability of the tree-ring reconstruction

23Dendrogeomorphic analysis of 788 increment cores taken from 184 P. abies allowed reconstruction of 26 events for the Schimbrig landslide since 1859 yielding a return period of 0.17 event yr^-¹. The reconstructed time series represents a minimum frequency of reactivation events for the Schimbrig landslide in the recent past as the reconstruction was limited by tree age and sample depth. In addition, several limitations are apparent as to the potential of tree-ring analysis to detecting past periods of landslide activity. Reactivation of the landslide body must be, on one hand, powerful enough to damage a sufficiently large number of trees through stem topping, tilting or root damage. At the same time, more violent and destructive events are likely to kill trees and evidence of this category of events is not likely to be available to the investigator, as dead trees will disappear rather soon after an event. Despite these limitations, the It and GD thresholds as well as the spatial analysis of event-response maps minimized the risk of GD resulting from non-landslide events to be included in the chronology. The thresholds also allowed rejection of GD related to creep or fall processes which have been shown to affect a rather limited number of trees per event (Stoffel and Perret, 2006; Trappmann and Stoffel, 2015).

24Moreover, historical archives and ancient maps confirm the reliability of our reconstruction. The topographic map (fig. 5A), dated to 1892, does not show a continuous forest in the Schimbrig area and therefore supports our data suggesting tree germination and the establishment of a forest at the end of the nineteenth century. Similarly, tree demography, represented indirectly through sample depth, is confirmed by the topographic map of 1941 (fig. 5B), on which a rapid afforestation of the catchment can be observed. The resolution of topographic information reported on the 1892 map does not, at the same time, permit to attest to the presence of a clear scarp or crown at Schimbrig at the end of the 19^th century. According to Mollet (1921), however, farmers in the region were aware of the difficulty to use the lands in this sector, and reported several incidents of slide events, which is clearly in agreement with the reactivations reconstructed in 1897 and 1917. Our results also reveal that landslide activity at Schimbrig predates the 20^th century, and that the movements described by contemporary landowners refer to local reactivations rather than to the initial triggering of the landslide.

25For the period 1962-2007, the diachronic analysis of aerial photographs provides additional evidence for the spatio-temporal accuracy of the dendrogeomorphic reconstruction presented in this paper. The reactivation of 1973, deciphered from tree-ring records, is thus confirmed by the slight extension of bare areas observed in the central part of the landslide body between 1962 and 1980 (fig. 5C-D). Similarly, our reconstruction reveals that the prominent earth slide of spring 1994, which occurred after periods of enhanced precipitation rates and wet autumns (Liniger and Kaufmann, 1994) and removed a total of approximately 322,000 t of sediments supplied to the channel network through by soil creeping, slumps and earth flows (Schwab et al., 2008), had damaged a large number of trees in 13 out of the 24 geomorphic units identified in the landslide body. According to aerial photographs, this event also created bare areas, secondary scarps and cracks (fig. 5D-E). Interestingly, growth disturbances – mainly in the form of compression wood – were already observed in trees scattered throughout the landslide body in the growing season 1993. These disturbances suggest that ground displacements, sufficient to modify stem verticality, were in place in the year before peak instability and can thus be seen as precursor signals of the high-magnitude event of 1994. These findings also confirm a technical report realized for the site in which localized ground deformation was observed as early as in spring 1992 (Liniger and Kaufmann, 1994).

26Finally, between 1998 and 2007 (fig. 5E-F), bare areas that appear preferentially in the central part of the landslide body, corroborate the most recent events reconstructed to 2001, 2002, 2003 and 2006.

5.2. Temporal correlations and clustering of landslide reactivations

27The reconstruction of spatio-temporal patterns of landslide activity with dendrogeomorphic techniques is recent but has been helpful for the understanding of landslide kinematics and its spatial evolution (Corominas and Moya, 2010). In our study, the exhaustive sampling of P. abies trees enabled computation of a very detailed spatio-temporal chronology of landslide reactivation at Schimbrig. Given the completeness of the reconstruction (since AD 1859), we were able to map return periods of reactivation at each of the 24 geomorphic units identified on the landslide body. Adopting a Poisson probability model (Guzzetti et al., 2005), we were also able to determine the probability of having a reactivation in each unit for time intervals varying from 5 to 100 years. Smallest return periods associated with major probabilities of reactivation are mapped on the right bank of the Rosslochbach main stream, with probabilities of reactivation that exceed 0.4 within the next 5 years and 0.8 within the next two decades. On the contrary, the southwestern units of the landslide body appear more stable with return periods >30 years since 1859 and probabilities ranging for 0.4 and 0.6 for the next 20 years.

28Our approach purposely does not include physically based modeling, as this conventional method has been shown to predict the spatio-temporal occurrence of landslides with difficulties (Jaiswal et al., 2011). Most previous work focusing on landslide mapping has been based on susceptibility maps and therefore provides an estimate of where landslides are expected to occur (Brabb, 1984; Guzzetti et al., 2005). Much less work has been done on the establishment of the temporal probability of reactivation (Coe et al., 2000; Guzzetti et al., 2005). The approach presented in this paper enables determination of quantitative probabilities of reactivation estimated directly from the frequency of past landslide events and does not require a landslide susceptibility analysis as a complete inventory of past landslide events was reconstructed with dendrogeomorphic techniques (Corominas and Moya, 2008).

29However, the temporal occurrence of reactivations is assumed to conform to a Poisson probability model which among others assumptions are: (i) the probability of an event occurring in a very short time is proportional to the time interval; (ii) the probability of more than one event in a short time interval is negligible (iii) the number of events which occur in one-time interval or region of space are assumed to be independent incidents of the number that occurs in any other disjoint time interval or region; (iv) the a posteriori probability distribution in the future are considered to be the same as those of the past (Crovelli, 2000).

30Most hazardous events, including landslides, do not probably correspond to uncorrelated time series and do not therefore occur at random (Coe et al., 2000). Witt et al (2010), who examined historical landslide time series reporting 2,255 events between 1951-2002 for an area in the Emilia-Romagna Region (Italy), demonstrated that landslide time series show significant correlations in time, and a temporal clustering of extremes over a given threshold. Clustering is attributed to a reactivation that may make the landslide more or less susceptible to future landslides, thus creating stability or instability in the future. In addition, changing land use (Lopez-Saez et al., 2016), locally changing climatic conditions (Lopez-Saez et al., 2013b; Gariano and Guzzetti, 2016) or the implementation of landslide mitigation measures, may consequently render landslide occurrence more or less likely in the future, a fact which is further jeopardizing the possibility of using an uncorrelated process to model the temporal occurrence of landslide reactivation (Chleborad et al., 2006; Witt et al., 2010). Based on these limitations, care should be taken when estimating the recurrence times of landslide events (Guzzetti et al., 2003, 2005, 2006). Nevertheless, the Poisson model is often used when no information other than the mean rate of event occurrence is known. Under such circumstances, the Poisson model provides a good first-cut estimate for the probability of event occurrence in the future (Coe et al., 2000).

6. Conclusions

31As human activities increase in mountain areas, landslides have become a more serious social and economic issue. As a consequence, improved and more detailed landslide forecasting becomes a prerequisite, even at the local scale. In this paper, we investigated the potential of extensive tree-ring analyses for landslide forecasting and show how dendrogeomorphology can add substantially to the spatio-temporal record of landslides at a study site. Many reactivations, which remained unnoticed in archival data, could be identified and thus help extend the history of landslides back to the mid-19^th century. Comparison of tree-ring data with aerial photographs clearly demonstrates the spatio-temporal accuracy of the reconstruction. In terms of land-use planning, the identification of endangered areas is of paramount importance and dendrogeomorphic reconstruction should therefore be used systematically for hazard zoning in forested areas affected by shallow landslides. Finally, if coupled with a Poisson model, dendrogeomorphic mapping can improve our knowledge about the probability of reactivation. These probability maps should be used for disaster prevention and the generation of risk maps, as well as for the detailed design phase of engineering works and for the construction of slope stabilization works, keeping in mind that the premises for a random Poisson-type process are not necessarily met.