See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/354445695

Size–frequency distribution of shallow landslides in the Black Forest,

Germany

Article in Earth Surface Processes and Landforms · September 2021

DOI: 10.1002/esp.5237

CITATIONS READS

0 122

4 authors, including:

Jakob Wilk Stefan Hergarten

University of Freiburg University of Freiburg
36 PUBLICATIONS 101 CITATIONS 147 PUBLICATIONS 2,146 CITATIONS

SEE PROFILE SEE PROFILE

Frank Preusser
University of Freiburg
303 PUBLICATIONS 7,925 CITATIONS

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Drilling Overdeepened Valleys in the Alps (DOVE) View project

PhD thesis with title "Determining the reservoir effect of marine and terrestrial mollusks in SE-Arabia with respect to time and dependence on species" View project

All content following this page was uploaded by Frank Preusser on 10 October 2021.

The user has requested enhancement of the downloaded file.

Received: 4 May 2020 Revised: 12 August 2021 Accepted: 13 August 2021

DOI: 10.1002/esp.5237

RESEARCH ARTICLE

Size–frequency distribution of shallow landslides in the Black

Forest, Germany

Matthias Büschelberger | Jakob Wilk | Stefan Hergarten | Frank Preusser

Institut für Geo- und

Umweltnaturwissenschaften, Albert-Ludwigs- Summary
Universität Freiburg, Albertstrasse 4 5 23b,
Landslides are a major hazard in mountainous regions, represent a threat to human
79100 Freiburg i. Br., Germany
life, and cause substantial economic costs. While some landslide hazard assessments,
Correspondence including hazards maps, are available for Germany, their spatial coverage is not uni-
Jakob Wilk, Institut für Geo- und
Umweltnaturwissenschaften, Albert-Ludwigs- form. In the Black Forest, several landslides have recently attracted public attention,
Universität Freiburg, Albertstrasse 4 5 23b, but the landslide hazard in this region has received limited consideration in literature
79100 Freiburg i. Br., Germany.
Email: [email protected] to date. This study focuses on the spatial pattern and size distribution of soil-borne
landslides in the submountainous valley of Menzenschwand. A combination of remote
sensing, geophysical surveying, and geotechnical testing was used to map and charac-
terize the landslide inventory of two selected hillslopes. In the statistical analyses, we
observe a larger proportion of small-scale landslides with size below 100 m2 than usu-
ally reported in similar studies. This effect may be related to the low cohesion of the
soil (glacial deposits). As a major result, a cutoff in the distribution at large landslide
sizes compared with the expected power-law tail of the distribution was found. While
the maximum landslide size found in this study was about 1100 m2, the largest land-
slide should theoretically be at least 2500 m2 at 95% probability. The cutoff at large
sizes is probably due to the limited soil thickness, where about 50% of the considered
area has a depth to bedrock of 1 m or less. For the considered location, this result
suggests that an increase in frequency and intensity of rainstorms should predomi-
nantly result in an increase of landslide frequency, but without increasing the size of
the largest landslides. As a more general implication, the contribution of large land-
slides to the total hazard may be overestimated if soil thickness is not considered, in
particular if the present-day hazard is projected to future scenarios.

KEYWORDS
landslide hazard, Pareto distribution, digital surface models, GPR survey, geotechnical analysis,
remote sensing, soil-borne landslides, drone photogrammetry, near-surface geophysics

1 | I N T RO DU CT I O N The latter occur in almost all mountain ranges and are often triggered
by rainstorms or rapid snow melt, so that a change of landslide hazard
The frequency and size of landslides in mountainous areas and the can be expected if precipitation and temperature patterns change in
related risk to human life and infrastructure are potentially affected the future. Shallow landslides, being bound to these mechanisms
by climate change (e.g., Crozier, 2010; Gariano & Guzzetti, 2016), (Crosta, 1997), are used synonymous with soil-borne landslides in
except for those triggered by earthquakes (e.g., Keefer, 1994; this text.
Meunier et al., 2007). This applies to landslides involving bedrock in Since landslides represent a threat to humans life and their eco-
high mountains due to the retreat of permafrost as well as to land- nomic costs are substantial (Klose et al., 2014), increasing efforts have
slides in unconsolidated sediments (i.e., soils in a geotechnical sense). been made to assess landslide hazards on both global and regional

This is an open access article under the terms of the Creative Commons Attribution-NonCommercial-NoDerivs License, which permits use and distribution in any
medium, provided the original work is properly cited, the use is non-commercial and no modifications or adaptations are made.
© 2021 The Authors. Earth Surface Processes and Landforms published by John Wiley & Sons Ltd.

Earth Surf. Process. Landforms. 2021;1–14. wileyonlinelibrary.com/journal/esp 1

2 BÜSCHELBERGER ET AL.

n et al., 2006). The ultimate goal is to produce maps

scales (e.g., Chaco the triggering mechanism (e.g., earthquakes, rainstorms, or rapid snow
showing the spatial incidence of landslides, reveal their spatiotemporal melt). These results were obtained by Malamud et al. (2004) from a
incidence and forecast their occurrence, as well their potential conse- few inventories and confirmed by Brunetti et al. (2009) when compar-
quences. Although landslides cause damage totaling about US $300 ing a larger number of datasets.
million in Germany annually (Klose et al., 2015), knowledge about Second, the size distribution of landslides in soil shows a strong
their impact is still based on limited data. In particular, the spatial cov- deficit at small sizes compared with the theoretical Pareto distribu-
erage of such information is not uniform, with a main focus on areas tion. This deficit has been attributed to incompleteness of the inven-
such as the uplands of central Germany (e.g., Damm et al., tories by Stark and Hovius (2001), while Malamud et al. (2004)
2009, 2010), the Swabian Alb (e.g., Bell et al., 2006; Neuhäuser & considered it to be a real property. As cohesion stabilizes slopes at
Terhorst, 2007), and the Alps including their foreland (e.g., Nie et al., small scales but becomes less important for large slope failures, a defi-
2017). Damm and Klose (2015) introduced a national landslide data- cit of small landslides can in principle be explained by soil mechanics.
base for Germany, although still with limited and nonuniform spatial However, recent work (Li et al., 2014; Tanyas et al., 2018) showed
coverage. In total, at least 23 national landslide databases exist in that there is still no full consensus about the origin of the deficit of
Europe (Damm & Klose, 2015). small events in landslide inventories. The same holds for the origin of
The statistical distribution of event sizes has become a central the power-law tail at large landslide sizes. Although some new model-
part of hazard assessment. Pareto distributions, also called power-law ing approaches have been presented in recent years (Alvioli et al.,
distributions, are now frequently used in the assessment of major 2014; Frattini & Crosta, 2013; Jeandet et al., 2019; Liucci et al.,
geohazards (e.g., Stark & Hovius, 2001). Whenever a certain type of 2017), there still seems to be no unique and widely accepted explana-
hazard is characterized by a universal size distribution, i.e., by a distri- tion. For further insights into the scaling properties of landslides, the
bution that is independent of the triggering mechanism and site- overview by Tebbens (2020) is recommended.
dependent conditions, an extrapolation of the hazard to event sizes This study addresses landslide occurrence and the related hazard
lying outside the range of the inventory available at a given location in the Valley of Menzenschwand in the Black Forest, a mountain
becomes possible. This is particularly helpful if the available invento- range in the southwestern part of Germany (Figure 1). While a series
ries are either restricted to a small domain or cover only a short time of smaller and moderate events in recent years have attracted atten-
span. Thus, although the statistics of large events are often insuffi- tion from the public and local authorities, to the best of the authors’
cient, their expected frequency can be estimated from the larger knowledge, no related studies have been published in international lit-
number of smaller events contained in an inventory. For Pareto- erature to date. The study area (47
840 N, 8
060 E) is situated close to
distributed hazards, however, attention must be paid to the tail of the the community of Menzenschwand (which initialized this study), at
distribution. Practically, Pareto distributions break down at very large about 900 m a.s.l. and southeast of the summit of Feldberg (1493 m
event sizes, and this breakdown or the transition to another type of a.s.l.), the highest peak in the Black Forest. The bedrock is mainly of
distribution is particularly relevant in the context of risk intrusive origin (Bärhalde-granite, GB), followed to the east by
(Hergarten, 2004). metamorphic units of paragneissic composition (pgA and pg)
Earthquakes were the first (Gutenberg & Richter, 1954) and are (Wimmenauer & Schreiner, 1981). The crystalline rocks were formed
still the most widely studied example in this context. Interest in during the Variscan orogeny of the Paleozoic and were uplifted again
power-law distributions rapidly increased after the theoretical con- in the context of the formation of the Upper Rhine Graben during the
cept of self-organized criticality was introduced (Bak et al., 1987; Cenozoic (Pflug, 1982). During Pleistocene glaciations, the valley of
Bak, 1996; Hergarten, 2002; Jensen, 1998). The power-law distribu- Menzenschwand was excavated by the Alb glacier system (Geyer &
tion of wildfires was even predicted by a simple model (Drossel & Gwinner, 2011). Evidence of these glaciations are U-shaped valleys as
Schwabl, 1992) several years before it was recognized in real data well as cirques (Zienert & Fezer, 1967). The last glaciation also formed
(Malamud et al., 1998). Nowadays, this has become an essential part terminal moraines (Rahm, 1970) such as the Kluse moraines (We in
of the discussion about the impact of climate change on wildfire haz- Figure 1) north of the Menzenschwand waterfalls (Hantke &
ard, and simple models have turned out to be able to explain some Rahm, 1976). Associated with the melting of the ice is the deposition
effects, for example, the statistical difference between natural and of a glacial debris sheet (Wm), which widely covers the resident
human-induced fires (Krenn & Hergarten, 2009). hillslopes and is mainly interpreted as till. Wimmenauer and
The first study where a power-law distribution of landslide sizes Schreiner (1981) estimated the till to be only a few meters thick upon
was found is more than 50 years old (Fuyii, 1969). However, it took the polished bedrock. According to the same source, the filling of the
almost 20 years for this topic to become popular through comprehen- valley is composed of fluvial and glaciofluvial sediments with a maxi-
sive studies mapping several thousand landslides (Hovius et al., 1997) mum thickness of about 45 m. Hence, our findings are restricted to
and modeling approaches (Densmore et al., 1998; Hergarten & regions with impermeable rock covered by unconsolidated deposits.
Neugebauer, 1998) attempting to bring landslides into the context of While the mean monthly precipitation at Menzenschwand is
self-organized criticality. The power-law size distribution of landslides about 150 mm, maximum values of more than 300 mm occur during
turned out to be more complicated than those of earthquakes and winter (Figure 2). Outstanding peaks of more than 500 mm occurred
wildfires. First, there seem to be two distinct groups of landslides. in December 2011 and in January 2018. In contrast to typical winter
Both landslides in non-consolidated layers and those involving bed- maxima, the January 2018 event was accompanied by moderate for-
rock follow power-law distributions over some range, but with mation of fresh snow. A considerable amount of precipitation was
strongly different exponents. In turn, the differences within each class contributed by a rainstorm that initiated some of the landslides inves-
seem to be small, and the distribution was found to be independent of tigated in this study.
BÜSCHELBERGER ET AL. 3

F I G U R E 1 Overview of the study area. (A) Geological map of Menzenschwand with the boundaries of hillslopes H1 and H3 as well as profile
lines of the ground-penetrating radar survey. Geological units digitized after Wimmenauer and Schreiner (1981). (B) Location in Europe.
(C) Oblique view with texture of the geological units from (A). Topography not exaggerated

F I G U R E 2 Monthly precipitation data in the

Menzenschwand area (DWD, 2019)

The particular situation with steep slopes where the bedrock is So far, only moderately sized landslides have been observed at
covered by a presumably thin layer of unconsolidated deposits raises Menzenschwand, but with high overall landslide frequency. However,
several questions. The first is (i) whether or not the landslide sizes fol- for practical hazard assessment, the size of the largest landslides to be
low the typical distribution for soil-covered slopes discussed above. expected is particularly relevant. This leads to the next question of
4 BÜSCHELBERGER ET AL.

(ii) whether landslides that cause serious damage to infrastructure or cannot be transferred to the uncertainty of the digital surface model.
even threaten human life could also take place. This question is partic- Nevertheless, combined with the fact that the raw data were obtained
ularly relevant in the context of changing climatic conditions, e.g., in from a multi-perspective scan via DroneDeploy, it can be assumed
case the contribution of rainfall to winter precipitation increases. This that the relative axis distortion of the projected model is close to real-
study combines digital elevation models (DEMs) derived from photo- ity. This surely will not affect the absolute georeferencing of the data,
grammetry in addition to light detection and ranging (LiDAR) data as but since only the relative sizes of mapped objects from the model
well as near-surface geophysics and laboratory soil tests to provide a (volume, scar surface area) are of interest, this will not affect the
comprehensive overview of the spatial pattern of landslides and their quantitative values of the derived size distributions, even if the true
statistical distribution in the Menzenschwand area. With regard to the precision of the DTM remains undefined.
specific geological setting, i.e., impermeable bedrock covered by thin Using the Zevenberg–Thorne approach for slope- and hillshade
layers of unconsolidated deposits, such as glacial sediments or rego- calculations in Q-GIS, the resulting raster maps enabled the identifica-
sols, this study could have implications for similar regions, beyond the tion of 246 recent and historical landslide scars and for the mapping
scope of regional hazard assessment. of the respective detachment areas. Allochthonous landslide deposits
were not considered. Since this optical method yields the surface
including vegetation, the survey was restricted to nonforested areas.
2 | FIELDWORK AND DATA ACQUISITION The 246 detected landslide scars were verified in the field and clas-
sified into rotational and translational type according to the scheme of
2.1 | Landslide mapping Varnes (1978). Rotational slides were recognized by their short trans-
port distance of a few decimeters and by the property that the sliding
Recent and historical landslides were mapped on two slopes of the masses remained in the failure cavity. Failure escarpments are often
valley in September and October 2018. The southeastern slope has characterized by a slightly converging shear plane with a crescent-like
an extension of about 0.17 km2 and is labeled H1 in Figure 1, while shape in plan view and terrace-like curve in cross-sectional view. The
the northwestern slope is 0.24 km2 in size and is labeled H3. Surface sliding masses often tend to become swampy, thus exhibiting notice-
morphologies and current land use are similar within H1 and H3 ably green vegetation cover in comparison with surrounding grasslands.
Two separate flight plans (one for each hillslope) for a DJI Mavic In turn, translational slides were identified by their tendency for higher
Pro drone were generated using the Android app DroneDeploy transport distances and a hollow form with relatively uniform failure
v2.81.0. By default, sampling rates and paths were computed using depth. In plan view, their detachment areas mostly show irregularly,
the software’s internal methods. The frontal and side overlap were elongated shapes. Some of the structures form multiarmed shapes.
kept constantly at 75% and 65% at a cruising altitude of about 110 m Figure 3 shows an example of a fresh translational slide and an older
above the launch location, which was arranged at medium slope ele- slide where the scar is already covered by vegetation.
vations of about 950 m a.s.l. This resulted in the construction of In total, 136 out of 246 slides could be assigned to one of these
216 camera views on H1 and 320 camera views on H3. By the use of two types (Table 1). The location and the spatial extent of the land-
a high-performance desktop computer, a local build of Agisoft slide scars are shown in Figure 4. Note that the landslides were not
Metashape Professional v1.5.2 was chosen for mesh and surface classified by age. Only B61 and B105 were evidently created during
model generation. The depth reconstruction for each aligned camera the heavy rainfall of January 2018. Escarpments of rotational land-
led to the formation of point clouds representing H1 with about slides that are not covered by vegetation were found at only a few
230,000 data points and H3 with about 2.4 million data points. The locations, suggesting the rather recent activity of landslides B2, B27,
following triangulation step produced a three-dimensional mesh with B42, B73, and B94. Hence, the predominant portion of the landslide
approximately 700,000 faces on H1, whereas H3 could be digitized inventory is considered to be historical.
with roughly 1.4 million faces. From this model, GIS-compatible
Geotiffs were derived with pixel resolution of ≲ 10 cm for both
slopes. The orthophotos with spatial resolution of ≲ 5 cm were deliv-
ered by texturing the cloud from a weighted-average mosaic
(see Agisoft LLC, 2018, for further technical details).
As usually applied in drone photogrammetry, ground control
points (GCPs) are key control features to ensure the accuracy of the
resulting point clouds and elevation models with respect to real-world
objects and their absolute coordinates. Since this study was carried
out due to the interest from a small community (Village of
Menzenschwand, population 500 people), the lack of financial
resources meant that we had to discard the option of investing in bor-
rowing high-precision continuous global positioning system (CGPS)
units. The drone itself provided absolute georeferencing in terms of
three-dimensional (3D) coordinates, whose precision is far more
uncertain than those obtained using more dedicated devices. The F I G U R E 3 Characteristic shapes of translational landslides B69
root-mean-square (RMS) reprojection error of the point cloud has a and B61 at hillslope H1. Landslide B61 occurred during the rainstorm
reasonably low target value of 1.4 pixels. Of course, this parameter in January 2018. EPSG code for map projection: 32632
BÜSCHELBERGER ET AL. 5

TABLE 1 Number of slope failures on both considered slopes The resulting distribution curves (Supplementary Figure S4) were
Landslide type H1 H3 Total dominated by gravel with a cumulative mass fraction of
48.4  10.9 wt.% and sand with 38.8  7.3 wt.%. Silt contributed a
Rotational 45 46 91
minor fraction of 10.0  3.9 wt.%, whereas clay was almost absent
Translational 17 28 45
with a content of less than 5 wt.%.
Unassigned 57 53 110
Since the grain size analyses revealed no fundamental differences
Total 119 127 246
among the samples, only a single triaxial testing set consisting of three
experiments was performed in order to estimate the effective shear
parameters after D.I.N.E.N.I.S.O. 17892-9 (2018). To obtain some kind
2.2 | Characterization of the subsurface of upper limit for the cohesion, the sample was taken from location
B2 with the highest content of clay and silt (about 20% in total)
For the assessment and calculation of the local stability of slopes, according to the grain size analyses. Fitting of the shear parameters to
insights into the architecture of the subsurface are essential. Based on the Mohr circles of the three experiments (Supplementary Figure S3)
the facies of the glacial debris cover (Wimmenauer & yielded an angle of internal friction of ϕ ¼ 38 ∘ and even a negative
Schreiner, 1981), it is expected that the maximum grain size exceeds cohesion of C ¼ 1 kPa. As a test of uncertainty, the parameters were
63 mm. Investigation of the soil strata by percussion drilling was also determined using the combinations of two experiments. Values
therefore not feasible. Instead, ground-penetrating radar (GPR) was of C ¼ 13:0 kPa, 6.6 kPa, and 8.2 kPa were obtained. So, there is an
used to mark the boundary to the bedrock. A detailed review of this uncertainty of several kPa even for this single triaxial testing set. As
method is provided by Neal (2004). the grain size sample from B2 had the highest content of clay and silt,
In total, 15 GPR surveys with an overall track length of 1.3 km the cohesion should be even lower at the other locations. As a rough
were carried out in September 2019 and January 2020. A bistatic, estimate, the cohesion should not be higher than 10 kPa over the
common-offset, sled-mounted, unshielded 100-MHz antenna entire considered domain.
equipped with a pulseEKKO Pro transmitter and a pulseEKKO Ultra
receiver unit (Sensors & Software) was used. Preceding in situ mea-
surements with a 200-MHz antenna suffered from rapid signal attenu- 3 | DATA ANALYSIS AND RESULTS
ation and strong dispersive diffraction. Due to the expected lower
propagation velocities of 0.04–0.06 m ns1 from this survey and liter- 3.1 | Spatial distribution of landslides
ature data on glacial deposits (Lukas & Sass, 2011), 100-MHz anten-
nas were chosen for this study. The antennas were oriented The areas of H1 and H3 exposed to landsliding are located at the foot
perpendicular broadside with separation of 0.5 m. Profiles were taken of the respective hillslope. Even though the planar surfaces of these
preferentially parallel to the contour lines with a few cross profiles to mapping areas feature slope angles within a quite narrow range
enable subsequent corroboration of depth migration of individual pro- between about 20
and 25
, the mapped landslides are not distributed
files. Traces along the GPR profiles were recorded at a spacing of homogeneously. This is revealed by the nesting and staggering of sev-
0.2 m, triggered by an odometer wheel attached to the rear of eral rotational and translational landslides, such as B1 plus B2 and B49,
the sled. B61, plus B69 at H1 as well as B90 to B94 and B105 and B108 at H3
The data were recorded with a 240 ns time window, trace stac- (Figure 4). The clustering of these mass movements is most likely trig-
king of 4096 measurements, and in situ time-zero corrections before gered by the accumulation of water both above and underneath the
the acquisition of each profile. Postprocessing steps were applied surface and is most likely related to formation of gullies (Figure 5).
using the pulseEKKO software as follows: (i) threshold-based time- As an example, B61 at H1 has a concise, crescent-shaped recess
zero adjustment, (ii) dewow and band-pass filter, (iii) muting of air and at the southern upper edge (Figure 3), suggesting that the fracture for-
ground waves, and (iv) fitting of the velocity function to the shape mation initially spread from this location since the preexisting cavity
and size of visible point source reflections and diffraction. The loca- of B49 did not provide any resistance to the detachment. A similar
tion of diffraction hyperbolas was carefully noted and the profiles f–k mechanism took place at B105 at H3. Over several heavy rainfall
migrated. For interpretation, Spreading & Exponential Compensation events, this process supports the ongoing erosion of gullies at both
(SEC) gain with attenuation of 0.9 db m1 was applied, and the topog- hillslopes. Due to the intensified water inflow into the sediment layer,
raphy was corrected by elevation static correction based on LiDAR a single hollow can cause the activation of newer detachments in the
and drone photogrammetry data acquired along the profiles. surroundings and even initiate the formation of a new coherent
channel system over long times.

2.3 | Geotechnical characterization

3.2 | Landslide size distribution
Geotechnical characterization was not a focus of this study. Thus, only
a few analyses were performed, to verify the characteristics of the Figure 6 shows the probability density of the landslide size distribu-
sediment described by Wimmenauer and Schreiner (1981). Grain size tion estimated from the 246 mapped landslides, where the area of
analyses by sieving and settling were conducted according to DIN EN ablation is considered as size. The data points were obtained by
ISO 17892-9 (2016). For this investigation, 12 samples were taken in aligning the data in logarithmic bins where the bin size increases by a
pffiffiffi
total at depths from 0.75 to 0.1 m below the subsurface. factor of 2 from one bin to the next.
6 BÜSCHELBERGER ET AL.

FIGURE 4 Orthophotos derived from quadrocopter overflights with mapped ablation areas of historical and recent landslides

FIGURE 5 Characteristic shapes of gullies at hillslope H3 in the surroundings of historical landslide events

Due to the limited range of areas, a power-law behavior maximum at A ≈ 50 m2. The behavior at very small sizes A < 10 m2 is
(corresponding to a straight line on a double-logarithmic plot) at large probably owing to the small numbers (four events in the first bin and
sizes is only recognizable over less than one decade in area. In turn, one event in the second bin), although one of the three large datasets
the rollover at small sizes (Malamud et al., 2004; Stark & analyzed by Malamud et al. (2004) even showed a systematic increase
Hovius, 2001) is clearly visible. The probability density achieves a local at very small sizes.
BÜSCHELBERGER ET AL. 7

The inverse Gamma distribution has apparently been used more

frequently than the double Pareto distribution in recent studies (e.g,
Tanyas et al., 2018; Van Den Eeckhaut et al., 2007). A few studies
have compared these two distributions (Hurst et al., 2013; Li et al.,
2016; Tebbens, n.d.) and found differences in the exponent ρ of up to
about 0.4, where the values obtained from the inverse Gamma distri-
bution tend to be higher than those obtained from the double Pareto
distribution. However, the small dataset considered in this study pre-
vents a contribution to answering this question of which distribution is
more appropriate. As the range of the power-law tail is very small here,
both are better than fitting a Pareto distribution to a small part of the
dataset only. In the following, we use the inverse Gamma distribution.
Equation (1) can be fit to the data by applying the maximum-
likelihood method. This means that the joint probability density of the
F I G U R E 6 Landslide size distribution. The points show the
data is interpreted as the likelihood of the parameter combination
probability density estimated by aligning the mapped landslide sizes in
(ρ, a, s),
logarithmic bins. The curves correspond to the inverse gamma
distribution (Equation (1)) with different parameter values Y
Lðρ, a, sÞ ¼ pðAi Þ, ð4Þ
i

Early studies (e.g., Hovius et al., 1997) fit a Pareto distribution

(a straight line on a double-logarithmic plot) to the part of the inven- where the product expands over all areas Ai in the dataset. The combi-
tory above a given minimum landslide size. The rollover at small sizes nation (ρ, a, s) that maximizes L is considered to be the most likely
leads to systematic underestimation of the exponent (the negative parameter set. Practically, the maximization is performed by
slope on a double-logarithmic plot) and thus an overestimation of the minimizing
frequency of large landslides. Although this problem may not be cru-
X
cial if the focus is on comparing different datasets (e.g., from different lnLðρ, a, sÞ ¼  lnpðAi Þ: ð5Þ
i
lithologies Hurst et al., 2013), it has become popular to use statistical
distributions that take such rollover into account, so that the fit can
include the entire dataset. Based on the idea that the rollover arises Application to the 246 mapped landslides yields an exponent ρ ¼
from incompleteness of the inventory, Stark and Hovius (2001) pro- 2:0 and the other values given in Table 2. The problem is, however,
posed a double Pareto distribution with an increasing power-law part practically underdetermined due to the limited range of landslide
at small sizes and a decreasing power-law tail. This distribution has sizes. As Malamud et al. (2004) suggested a universal value of ρ ¼ 1:4,
also been used in other studies (e.g., Guthrie & Evans, 2004a, 2004b). we also tested the fit when assuming this fixed value, so that only
Malamud et al. (2004) were the first to suggest that the deficit in the a and s are adjustable parameters. The resulting values are also pres-
frequency of small landslides compared with a Pareto distribution is ented in Table 2, while the resulting probability densities are plotted
not a matter of data incompleteness but rather a real property of land- in Figure 6. Both curves can hardly be distinguished over the range
slides in soil. They proposed an inverse gamma distribution as a heu- covered by the data. The respective values in lnL differ by only 0.013,
ristic description of the entire distribution. The probability density of so that the Bayes factor (the ratio of the two L values) is 1.013. This
the inverse gamma distribution in terms of area A reads means that the combination with ρ ¼ 2:0 is formally more likely by a
factor of 1.013 than the version with ρ ¼ 1:4, but this difference is so
1  a ρþ1  a 
small that the two must be considered equivalent. Therefore, there is
pðAÞ ¼ exp , ð1Þ
aΓðρÞ A  s As
no evidence that the exponent deviates from the universal value ρ ¼
1:4 suggested by Malamud et al. (2004).
where ρ, a, and s are parameters, and Γ is Euler’s gamma function. The distribution as a whole, however, deviates from the predicted
Similarly to the double Pareto distribution, the inverse Gamma distri- distribution at large landslide sizes. The largest landslide in our dataset
bution approaches a power law for A ! ∞, has an area of Amax ¼ 1096 m2. The probability density estimated from
the binned data (points in Figure 6) follows the inverse gamma distri-
pðAÞ  Aðρþ1Þ : ð2Þ bution reasonably well up to about this size and then suddenly drops
to zero. In contrast, the cumulative probability PðAmax Þ (Equation (3))
The cumulative distribution, i.e., the probability that a given land- is 3.7% for ρ ¼ 1:4, so the inverse gamma distribution predicts about
slide has an area ≥A, also turns into a power law for A ! ∞, nine landslides in the inventory with sizes A ≥ Amax . Thus, the distribu-
tion of the data is rather a truncated inverse gamma distribution with
∞
PðAÞ ¼ ρ
pðuÞdu  A , ð3Þ a cutoff at a given maximum landslide size. As the maximum-likelihood
ð A estimate assumes a full inverse gamma distribution without a cutoff,
the lack of large landslides also explains the overestimation of the
but with an exponent of ρ instead of ρ + 1. exponent.
8 BÜSCHELBERGER ET AL.

The lack of large landslides on the considered slopes can be quan-

tified with the help of the extreme value distribution where Pe ðAÞ is
the probability that the largest landslide in an inventory of n landslides
has a size of at least A. This probability can be computed by consider-
ing the inverse case that all n landslides are smaller than A, i.e.,

1  Pe ðAÞ ¼ ð1  PðAÞÞn : ð6Þ

 n
Using the relation 1  nx ! expðxÞ for n ! ∞, the extreme value
distribution can be written in the form

Pe ðAÞ ¼ 1  expðnPðAÞÞ ð7Þ

for large n. This distribution is shown in Figure 7, while some of its

characteristic values are presented in Table 2. F I G U R E 7 Extreme value distribution of landslides sizes for an
The probability that the largest event will be larger than the larg- inventory of n ¼ 246 landslides
est observed landslide is Pe ðAmax Þ ¼ 99:99% for ρ ¼ 1:4. Even for the
estimate ρ ¼ 2:0 that is already biased by the cutoff, it is 98.60%, so
an inverse gamma distribution without an upper cutoff can be effect of alterations in topography. The required width of the buffer
excluded for the size of the largest detected landslide in any case at was estimated visually for each of the 22 cavities.
more than 95% confidence level. For ρ ¼ 1:4, the largest landslide Figure 8 compares the computed volumes with empirical relations
should be even larger than A95 ¼ 2510 m2 at 95% probability, which is between area and volume, namely the relation
about 2.3 times as large as the largest detected landslide. The
expected size of the largest landslide in an inventory of 246 landslides V ¼ 0:224A1:262 ð8Þ
should be Amax ¼ 17,660 m for an inverse gamma distribution with-
2

out an upper cutoff at ρ ¼ 1:4, which is even 16 times larger than (in meters) obtained by Larsen et al. (2010) from an analysis of 1617
observed. landslide scars and the relation

V ¼ 0:074A1:45 ð9Þ
3.3 | Landslide volume and depth
obtained by Guzzetti et al. (2009) from 677 landslides. Although the
Detached volumes were estimated for 22 sufficiently preserved cavi- exponents in Equations (8) and (9) differ strongly, both relations finally
ties. The cookie-cutter tool originally developed for glacial bedforms differ by less than 20% in the range between A ¼ 150 m2 and
(Smith et al., 2009) was used for this purpose. This tool first removes A ¼ 1000 m2. Klar et al. (2011) obtained a range from 1.32 to 1.38 for
the disturbed area from the DEM and then fills the void with a thin the exponent from theoretical considerations, which lies about in the
plate spline function. As this kind of interpolation is sensitive to eleva- middle of the two values mentioned above. In the following, we use
tions and to local slopes close to the boundary of the void, it relies on Equation (8) because it was obtained from a larger dataset.
the preservation of the topography around the cavity. In total, 22 out If the mean landslide depth d is defined as the ratio of volume to
of the 246 mapped landslides appeared to be visually preserved suffi- area, this relation becomes
ciently well. A buffer around the cavity was introduced to reduce the
V
d¼ ¼ 0:224A0:262 : ð10Þ
A

T A B L E 2 Parameters of the inverse gamma distribution

The volumes estimated for the two fresh (January 2018) land-
(Equation (1)) and properties of the respective extreme value
distribution. The right-hand column refers to the result of the slides (B61 and B105) are in very good agreement with Equation (8),
maximum-likelihood estimate for all three parameters. The middle while the majority of the other volumes are up to about three times
column refers to a maximum-likelihood estimate of a and b only, with lower than predicted. As the raw data used by Larsen et al. (2010)
a fixed exponent ρ ¼ 1:4 as suggested by Malamud et al. (2004) show a scatter of about one decade (minimum to maximum volume at
ρ 1.4 2.0 constant area), the volumes obtained here may still be reasonable.
a (m2) 129 224 The volumes of old landslides, however, may also be systematically
2
s (m ) 16.8 29.0 underestimated. First, the cavity may have been partly filled with sedi-

lnLðρ,a,sÞ 6.159 6.146 ment over time. Moreover, thin plate splines are sensitive to the
slopes at the boundary. If the boundary of an old landslide scar has
PðAmax Þ (%) 3.7 1.7
been smoothed over time, the reconstructed original surface will be
Pe ðAmax Þ (%) 99.99 98.60
too low, so that the volume will also be underestimated. As a system-
A95 (m2) 2510 1330
atical underestimation cannot be excluded and the results for the two
‾Amax (m2) 17660 4298
fresh landslides are very close to the predicted volumes, the evidence
BÜSCHELBERGER ET AL. 9

that the landslides considered here are shallower than predicted by 3.4 | Assessment of the subsurface structure
Equations (8) and (10) is rather weak, so we assume that these rela-
tions capture the scaling properties of the landslides considered here The interpretation of the GPR profiles was approached by dis-
well. For the largest landslide in the inventory (Amax ¼ 1096 m ), the
2
tinguishing different radar facies (Jol & Bristow, 2003). Based on
predicted volume is about 1500 m3, corresponding to a mean depth of reflection patterns, three different facies were determined (Figures 9
about 1.4 m. and 10). The lowermost radar facies RF3 is interpreted as the bedrock,
characterized by a few subparallel linear reflectors, an absence of
point reflectors, and strong signal attenuation with respect to similar
two-way travel times in all the profiles. The upper section shows mul-
tiple point reflections and segments with strong echoing events
(Figure 9, profile 13 at 52–60 m). The partly very coarse-grained gla-
cial sediments contain a multitude of point reflectors that scatter the
emitted signal irregularly and cause echoes at neighboring objects,
which locally leads to a noisy signal. Also, some echoing is developed
at topographic troughs, which we interpret as filled gullies.
Since some of the profiles (e.g., Figure 10) show an additional
subhorizontal reflector, indicating stratification within the glacial sedi-
ments (Figure 10, profile 3 from 20 m on), the upper section is further
subdivided into two radar facies RF1 and RF2, with RF2 always being
separated from the lower section RF3 by a distinct, continuous reflec-
tor. We assume that the textural change from the glacial deposits to
F I G U R E 8 Estimated landslide volumes. Data points were
obtained using the cookie-cutter tool (Smith et al., 2009). The orange the underlying bedrock causes this strong contrast in relative dielec-
line shows the relation suggested by Larsen et al. (2010). The dashed tric permittivity that is responsible for the reflection (Sucre et al.,
lines correspond to given mean depths 2011; van Dam & Schlager, 2000). The onset of the lower reflector at

F I G U R E 9 GPR profiles 7 and 13 at site H3 with their respective orientation. RF2 and RF3 are the radar facies interpreted in profile 3 and
6. Note the apparent formation and refilling of gullies, as well as abundant point reflectors in the upper section
10 BÜSCHELBERGER ET AL.

F I G U R E 1 0 GPR profiles 3 and 6 at the southwestern hillslope H1. Radar facies 1 and 2 (RF1 and RF2) represent glacial sediments. Radar
facies 3 (RF3), marked by strong signal attenuation and a lack of diffraction hyperbolas, is interpreted as the bedrock. Depth to bedrock values
were determined from the solid marked boundary of RF2 to RF3

the top of RF3 is interpreted below as depth to bedrock, i.e., the tran-
sition from glacial deposits (soil) to bedrock.
Adjustment to the reflection parabolas yields two-way travel time
velocities between v ¼ 0:06 m ns1 and v ¼ 0:08 m ns1 in all profiles.
Velocities of v ¼ 0:06 m ns1 were consistently determined in RF1
and RF2, but predominantly higher apparent velocities were present
in RF3. In comparison, Tillard and Dubois (1995) determined two-way
velocities of about 0.12 m ns1 at an emission frequency of 200 MHz
in granitic rock, indicating higher characteristic EM wave velocities for
granitic rocks and effectively low penetration ability in the present
study. Based on these results, we assumed a velocity of
v ¼ 0:06 m ns1 to transform two-way travel times into depths.
For the identified contacts, the depth to bedrock was determined
in intervals of 2 m horizontally and 0.1 m vertically. Because the bed-
rock is not fully traceable through all profiles, we differentiate F I G U R E 1 1 Statistical distribution of depths to bedrock obtained
between “inferred contact” and “contact,” as indicated in Figures 9 from rasterizing GPR profiles. Area is given as a fraction, i.e., the
number of pixels per total number of pixels. Bars indicate 0.2 m bins
and 10, and discard inferred contacts from later analysis.
for the entire dataset. Lines describe the cumulative distribution,
A simple approach was chosen for analyzing the depths to bed-
i.e., the fraction of area with depth greater or equal to the value on
rock obtained from the GPR survey along the profiles. The profiles the x-axis
were rasterized using 5  5 m pixels, and a mean depth was com-
puted for each pixel that contained data. These were 95 pixels in total. et al., 2012) might help here. However, the slopes considered here are
The resolution of 5 m is a tradeoff between obtaining a sufficient topographically quite homogeneous, so the application of such a
number of pixels and reducing the influence of small-scale variations model would require an independent validation that would go beyond
in depth that are not relevant to slope stability. The obtained statisti- the scope of this study. Therefore, the results of the GPR study tenta-
cal distribution is shown in Figure 11. The largest depth found on this tively suggest that about 50% of the area has a depth to bedrock of
scale is 2.2 m. 1 m or less, about 30% a depth of 1.4 m or more, and less than 10% a
The data tentatively suggest that the depths at H1 are slightly depth of at least 2 m.
higher than at H3, but lower at the profiles outside the regions H1
and H3 (cf. Figure 1). As the outside profiles are in the upper parts of
the slopes, the latter results could reflect an increase of thickness in 4 | DI SCU SSION
downslope direction due to long-term downward movement of mate-
rial. However, note that the profiles cover only a small part of the The landslide size statistics of the two slopes considered in this study
total area and are not necessarily representative. Regression-based are overall consistent with the distribution obtained by Malamud
models for estimating soil thickness (e.g., Olyphant et al., 2016; Segoni et al. (2004) from inventories covering larger scales. Some differences,
BÜSCHELBERGER ET AL. 11

however, occur at both small and large landslide sizes. We obtained Equation (12) predicts an increase by a factor of 1.08 and 1.4 at these
parameter values a ¼ 129 m2 and s ¼ 16:8 m2 describing the rollover ^
two values of C. So, potential landslides with A ¼ 50Až,m 2
are more
of the distribution at small sizes. The respective values a ¼ 1280 m2 stable by only a factor of 1.06 compared with those with A ¼ 500 m2
and s ¼ 132 m2 obtained by Malamud et al. (2004) are about one at C ¼ 1 kPa, but by a factor of 1.25 at C ¼ 5 kPa. Considering that the
order of magnitude higher. Hurst et al. (2013) obtained a ¼ 10,900 m2 range of 1–1.3 is often considered to be the conditionally unstable
and b ¼ 1910 m2 from an inventory of 8453 landslides in the UK, regime, a change by a factor of 1.25 could indeed cause a strong dif-
although for the deposit areas and with a lower value of the exponent ference in landslide frequency. Thus, a cohesion on the order of mag-
ρ. The ratio of a and s that describes the shape of the curve is similar nitude of 5 kPa could be responsible for the rollover in the
in the three inventories. Therefore, the difference mainly lies in the distribution with a maximum probability density at sizes of about
absolute scale at which landslides become less likely with decreasing 50 m2 found in this study. However, it was not conceivable at the time
size. The landslide size with the highest frequency density is about when the field and laboratory work was performed that a more pre-
^
50 m2 in our inventory, about 400Až,m 2
in the inventory of Malamud cise estimate of the cohesion would become relevant. Anyway, it
et al. (2004), and about 3000 m2 in the inventory analyzed by Hurst would have been a challenge to determine whether the cohesion was
et al. (2013). Among the inventories reviewed by Van Den Eeckhaut greater than 5 kPa for a considerable part of the domain, even with a
et al. (2007), those analyzed by Malamud et al. (2004) had the lowest larger number of samples.
rollover landslide size. In turn, the data of Van Den Eeckhaut While our results support the hypothesis that the cohesion is
et al. (2007) obtained from an inventory in the Flemish Ardennes did responsible for the rollover, the alternative hypothesis of
not show a rollover at sizes above 100 m2, but a decreasing power- undersampling cannot be discarded. As discussed above, the landslide
law section with a smaller exponent for A≤10, 000 m2. Thus, the roll- volumes estimated from the topography were consistent with the pre-
2
over in these data must be at A < 100 m , although it cannot be seen diction from Equation (8) only for the two recent landslides. The vol-
in the data restricted to sizes above 100 m2. umes of older landslides were presumably underestimated due to
So, the rollover found in our inventory occurs at sizes about one changes in small-scale topography, which could also affect the detect-
order of magnitude lower than found by Malamud et al. (2004) and ability in the field and in DEMs. Compared with other studies where a
even more than one order of magnitude lower than those found in rollover at larger sizes was found, our study refers to smaller scales
other studies. This difference may be related to the mechanical prop- with a better detectability of small landslides and to soils with a low
erties of the material involved, as the slopes considered here are cov- cohesion. As both differences probably have similar effects on the
ered by material with low cohesion. As discussed above, cohesion location of the rollover, an influence of undersampling still cannot be
avoids shallow failure and thus inhibits small landslides. The potential refuted. Nevertheless, our results support the theoretical consider-
influence of this effect on the rollover in the landslide size distribution ation of Li et al. (2016) about the influence of cohesion and the origi-
was discussed in detail by Li et al. (2014) and was even suggested to nal conjecture of Malamud et al. (2004) that the rollover is real.
be the primary reason for the rollover at small sizes. So, the low cohe- At this point, one may ask whether the cohesion in combination
sion found here provides an explanation for the shift of the rollover with variations in soil thickness may also be responsible for the con-
towards small sizes compared with other inventories, at least centration of the landslides in the lower parts of the slopes, although
qualitatively. the entire slopes are rather homogeneous in terms of slope angle. The
A rough quantitative estimate can be obtained by assuming a results of Sect. 1 suggest that flow and accumulation of water in the
slope-parallel failure plane at a depth d (normal to the surface) below subsurface may be responsible for the higher susceptibility of the
a straight slope with a slope angle α. The normal stress is then lower parts. This hypothesis is supported by the analysis of the topog-
σ ¼ wdcosα, where w is the specific weight of the soil. The critical raphy, in particular by the occurrence of gullies. However, the analysis
shear stress is thus of the subsurface structure (Sect. 1) indeed revealed a greater depth
to bedrock at the two areas H1 and H3 compared with the areas
τc ¼ σ tan ϕ þ C ¼ wd cos α tan ϕ þ C, ð11Þ above these domains (Figure 11). The mean depths of the profiles at
both H1 and H3 are about 1.2 m, while the mean depth of those out-
where ϕ and C are the angle of internal friction and the cohesion, side H1 and H3 is only 0.9 m. However, the difference in the factor of
respectively. So, cohesion results in a relative increase in the critical safety between these two depth would be only about 3% for
shear stress and thereby in the safety factor by a multiple of C ¼ 1 kPa and about 9% for C ¼ 5 kPa according to Equation (12). So,
the concentration of landslides in the lower slope segments is proba-
C bly not an immediate effect of an increase in soil depth in downslope
ϵ¼1þ : ð12Þ
wd cos α tan ϕ
direction. Although vegetation may also have an influence on the dis-
tribution, the accumulation of water in the subsurface remains the
The rollover occurs at landslide sizes of about 50 m2, and the fre- most likely reason for the strong increase of landslide susceptibility in
quency density is reduced by about one order of magnitude compared downslope direction.
with the extrapolated power law (Figure 6). Equation (10) predicts a The most important difference in the landslide statistics, however,
mean landslide depth of 0.62 m at this size. Assuming concerns the tail of the distribution, i.e., the number of large land-
w ¼ 15 kN m3 , ϕ ¼ 38 ∘ (Supplementary Figure S3), and α ¼ 22 ∘ , slides. We found a maximum size of about 1100 m2, while the
Equation (12) predicts an increase in the safety factor by a multiple of assumed inverse gamma distribution predicts a maximum size of at
1.15 at C ¼ 1 kPa and by a factor of 1.74 at C ¼ 5 kPa. In turn, no least 2500 m2 even at 95% probability (Figure 7). So, there is a clear
strong deviation from the power law is observed at A ¼ 500 m . 2
lack of large landslides despite the good agreement with the predicted
12 BÜSCHELBERGER ET AL.

distribution at smaller landslides. As discussed above, the mean depth mean soil thickness. However, these data are based on a single land-
of landslides of the maximum size found here should be about 1.4 m. slide on slopes where the soil depth has been investigated for only a
The results of the GPR measurements (Figure 11) suggest that the small fraction of the area. An analysis of the respective largest land-
rather low depths to the bedrock may be the reason for the limited slide in several inventories with known soil depths would be neces-
landslide sizes. Even on the considered 25 m2 pixel scale, less than sary to obtain a robust relationship. Such a relationship could finally
one-third of the area has a depth of at least 1.4 m, and depths of 2 m improve hazard assessment in combination with future developments
or more are very rare. The largest expected landslide with an area of such as drone-based GPR systems, satellite-based sensors, or
‾Amax ¼ 17,660 m 2
should have a mean depth of about 2.9 m improved mathematical models for predicting soil thickness.
according to Equation (10), which is even 30% larger than the maxi-
mum depth to bedrock of 2.2 m found among the considered 25 m2 CONCLUS IONS
pixels. So, it is likely that the lack of landslides much larger than This study focused on the spatial pattern and size distribution of soil-
2
1000 m is due to an insufficient depth of potentially mobile material. borne landslides in the submountainous area of the Black Forest,
A cutoff in the distribution at large landslide sizes was presumably Germany. Geophysical surveying (GPR) revealed that the two steep
discussed first by Hergarten (2012) in the context of landslides in slopes investigated bear only a thin soil layer of about 1 m, consisting
rock. Topography was suggested as the limiting factor there. Hurst of unconsolidated glacial sediments with low cohesion. While the ana-
et al. (2013) observed a deficit in the number of large landslides in lyses of size statistics derived from DEM and orthophotos are overall
soils at areas on the order of magnitude of 1 km2 and also attributed consistent with the distribution recognized in larger inventories, our
this to topography. In our study, however, the maximum landslide size study reveals differences from previous studies at both small and large
is several orders of magnitude smaller than the size of the slopes. So, landslide sizes.
the limited soil thickness introduces a cutoff at much lower scales First, we observed a rather high proportion of small landslides.
than considered by Hurst et al. (2013). The rollover in the distribution occurs at an area of about 50 m2,
From a practical point of view, the implications of this result for which is about one order of magnitude lower than the rollover found
hazard assessment are very interesting. Although we were not able to in other inventories. This difference is likely explained by the low
assign an age to the majority of the mapped landslides, it is obvious cohesion, since cohesion inhibits small landslides in general. For the
from the pattern of the landslides alone (Figure 4) that the landslide considered inventory, a low cohesion of about 5 kPa would be able to
susceptibility at the considered location is rather high. The slopes are explain the rollover. This result supports the hypothesis that cohesion
quite steep, but covered by a potentially mobile layer that is only plays a major part in the deficit of small landslides (Li et al., 2014;
weakly supported by cohesion. The observed landslides are, however, Malamud et al., 2004) observed in several studies, although an effect
rather small. While landslides of the observed sizes will typically not of undersampling still cannot be excluded.
have much effect, the question is whether larger landslides causing As a second result, we observed a distinct cutoff in the distribu-
serious damage to infrastructure or even loss of life could occur here. tion at large landslide sizes. In this study, it was found at a size of
This question becomes particularly relevant it either the frequency or about 1100 m2, which is considerably lower than the cutoff found by
intensity of heavy rainstorms increases, or if a larger part of the winter Hurst et al. (2013) in larger inventories. This cutoff is probably due to
precipitation changes from snow to rainfall. the limited soil thickness. In contrast to the increased frequency of
In general, the resulting increase in landslide frequency is accom- small landslides, the cutoff at large landslide sizes has a direct impact
panied by an increasing size of the largest landslide to be expected in on hazard assessment. In particular, this result suggests that an
a given time span. As an example, the extreme value distribution of increase in the frequency and intensity of rainstorms or a shift of win-
the inverse gamma distribution predicts an expected maximum size of ter precipitation towards rainfall should predominantly result in an
‾Amax ¼ 17,660 m2 (Table 2) for a sample size of n ¼ 246 landslides as increase of landslide frequency, but without increasing the size of the
found in our study. If we increase the sample size to 2n ¼ 492 land- largest landslides to be expected at the considered location. This
slides, the expected maximum landslide size increases to result may by transferable to locations with steep slopes and low soil
‾Amax ¼ 29,000 m2. This is, however, only true for the original distri- thickness. In this context, the contribution of large landslides to total
bution without any cutoff. Given that the observed cutoff at landslide hazard may be overestimated if soil thickness is not taken
A ≈ 1100 m is indeed imposed by the geological constraint of limited
2
into account, in particular if the present-day hazard is projected to
soil thickness, it cannot be expected that this cutoff will be strongly future scenarios. However, more research on landslide inventories
affected by climate change. In this case, an increasing frequency or and on estimating soil thickness is required on the way towards esti-
intensity of rainstorms may have an effect on landslide frequency, but mating the maximum landslide size to be expected in a given geologi-
the maximum landslide size to be expected should not increase much. cal setting.
This result should also hold for locations with a similar setting,
i.e., with steep slopes covered by a thin soil cover. AC KNOW LEDG EME NT
Determining a maximum possible landslide size or at least the size The data that support the findings of this study are available from the
above which landslides become rapidly less likely would be a major corresponding author upon reasonable request.
step in hazard assessment. As mentioned above, the largest landslide None of the authors declares any real or perceived financial con-
observed at the considered location has an expected mean depth of flicts of interest.
1.4 m, and about 30% of the area has a soil thickness of at least The authors thank Michael Vandrey, Bertram Schrade, and
1.4 m. It might be tempting to derive a relationship between the maxi- Wolfgang Gleim from HPC AG Freiburg (Germany) for the idea for
mum landslide size and any quantile of the soil thickness or even the this project, their kind support, and their valuable know-how in
BÜSCHELBERGER ET AL. 13

geotechnical skills. We also thank Johann Meier, Hans-Jörg Meier, Frattini, P. & Crosta, G.B. (2013) The role of material properties and land-
Adrian Probst, and Benno Kaiser from the Locality of St. Blasien as scape morphology on landslide size distributions. Earth and Planetary
Science Letters, 361, 310–319.
well as Menzenschwand for their interest in and support of this study.
Fuyii, Y. (1969) Frequency distribution of the magnitude of
The realization of the triaxial experiments was supported by expert landslides caused by heavy rainfall. Seismological Society of Japan, 22,
knowledge and professional consultation of Matthias Pamler from 244–247.
Geomation GmbH (Germany). Gariano, S.L. & Guzzetti, F. (2016) Landslides in a changing climate. Earth-
Science Reviews, 162, 227–252.
Geyer, O.F. & Gwinner, M.P. (2011) Geologie von Baden-Württemberg.
DATA AVAILABILITY STATEMENT Schweizerbart’sche Verlagsbuchhandlung: Stuttgart.
The data supporting the findings of this study are available from the Gutenberg, B. & Richter, C.F. (1954) Seismicity of the earth and associated
corresponding author upon reasonable request. phenomenon, 2. Princeton University Press: Princeton.
Guthrie, R.H. & Evans, S.G. (2004a) Analysis of landslide frequencies and
characteristics in a natural system, coastal British Columbia. Earth
ORCID Surface Processes and Landforms, 29, 1321–1339.
Jakob Wilk https://orcid.org/0000-0002-2605-9444 Guthrie, R.H. & Evans, S.G. (2004b) Magnitude and frequency of landslides
Stefan Hergarten https://orcid.org/0000-0002-4780-284X triggered by a storm event, Loughborough Inlet, British Columbia.
Natural Hazards and Earth System Sciences, 4, 475–483.
Guzzetti, F., Ardizzone, F., Cardinali, M., Rossi, M. & Valigi, D. (2009)
RE FE R ENC E S
Landslide volumes and landslide mobilization rates in Umbria, central
Agisoft LLC. (2018) Agisoft Metashape User Manual: Professional Edition, Italy. Earth and Planetary Science Letters, 279, 222–229.
Version 1.5. Available from: https://www.agisoft.com/pdf/ Hantke, R. & Rahm, G. (1976) Das frühe Spätglazial in den Quellästen der
metashape-pro_1_5_en.pdf Alb (Südlicher Schwarzwald). Vierteljahrsschrift der Naturforschenden
Alvioli, M., Guzzetti, F. & Rossi, M. (2014) Scaling properties of rainfall Gesellschaft in Zürich, 121, 293–299.
induced landslides predicted by a physically based model. Geomor- Hergarten, S. (2002) Self-organized criticality in Earth systems. Springer:
phology, 213, 38–47. Berlin, Heidelberg, New York.
Bak, P. (1996) How nature works – the science of self-organized criticality. Hergarten, S. (2004) Aspects of risk assessment in power-law distributed
Berlin, Heidelberg, NY: Copernicus, Springer. natural hazards. Natural Hazards and Earth System Sciences, 4,
Bak, P., Tang, C. & Wiesenfeld, K. (1987) Self-organized criticality. An 309–313.
explanation of 1/f noise. Physical Review Letters, 59, 381–384. Hergarten, S. (2012) Topography-based modeling of large rockfalls and
Bell, R., Kruse, J.-E., Garcia, A., Glade, T. & Hördt, A. (2006) Subsurface application to hazard assessment. Geophysical Research Letters, 39,
investigations of landslides using geophysical methods: geoelectrical L13402.
applications in the Swabian Alb (Germany). Geographica Helvetica, Hergarten, S. & Neugebauer, H.J. (1998) Self-organized criticality in a land-
61(3), 201–208. slide model. Geophysical Research Letters, 25, 801–804.
Brunetti, M.T., Guzzetti, F. & Rossi, M. (2009) Probability distribution of Hovius, N., Stark, C.P. & Allen, P.A. (1997) Sediment flux from a mountain
landslide volumes. Nonlinear Processes in Geophysics, 16, 179–188. belt derived by landslide mapping. Geology, 25, 231–234.
Chaco
n, J., Irigaray, C., Fernández, T. & El Hamdouni, R. (2006) Engineering Hurst, M.D., Mudd, S.M., Attal, M. & Hilley, G. (2013) Hillslopes record the
geology maps: landslides and geographical information systems. growth and decay of landscapes. Science, 341, 868–871.
Bulletin of Engineering Geology and the Environment, 65, 341–411. Jeandet, L., Steer, P., Lague, D. & Davy, P. (2019) Coulomb mechanics and
Crosta, G. (1997) Regionalization of rainfall thresholds: an aid to landslide relief constraints explain landslide size distribution. Geophysical
hazard evaluation. Environmental Geology, 35, 131–145. Research Letters, 46, 4258–4266.
Crozier, M.J. (2010) Deciphering the effect of climate change on landslide Jensen, H.J. (1998) Self-organized criticality – emergent complex behaviour
activity: A review. Geomorphology, 124, 260–267. in physical and biological systems. Cambridge University Press:
D.I.N.E.N.I.S.O. 17892-9. (2018) Geotechnische Erkundung und Cambridge, New York, Melbourne.
Untersuchung – Laborversuche an Bodenproben – Teil 9: Jol, H.M. & Bristow, C.S. (2003) Gpr in sediments: advice on data
Konsolidierte triaxiale Kompressionsversuche an wassergesättigten collection, basic processing and interpretation, a good practice guide.
Böden. Geological Society, London, Special Publications, 211(1), 9–27.
DIN EN ISO 17892-9. (2016) Geotechnische Erkundung und Keefer, D.K. (1994) The importance of earthquake-induced landslides to
Untersuchung – Laborversuche an Bodenproben – Teil 4: long-term slope erosion and slope failure hazards in seismically active
Bestimmung der Korngröenverteilung. Beuth Verlag, Berlin. regions. Geomorphology, 10, 265–284.
DWD (2019) Historische stündliche Stationsmessungen der Klar, A.E., Aharonov, E., Kalderon-Asael, B. & Katz, O. (2011) Analytical
Niederschlagshöhe für Deutschland Version v006. Deutscher Wetter and observational relations between landslide volume and surface
Dienst – Climate Data Center (CDC). Retrieved from: ftp://ftp-cdc. area. Journal of Geophysical Research, 116, F02001.
dwd.de. Klose, M., Highland, L., Damm, B. & Terhorst, B. (2014) Estimation of direct
Damm, B., Becht, M., Varga, K. & Heckmann, T. (2010) Relevance of tec- landslide costs in industrialized countries: Challenges, concepts, and
tonic and structural parameters in Triassic bedrock formations to case study. In Landslide science for a safer geoenvironment, Sassa, K.,
landslide susceptibility in quaternary hillslope sediments. Quaternary Canuti, P. & Yin, Y. (eds), Springer: Cham; pp. 661–667.
International, 222, 143–152. Klose, M., Maurischat, P. & Damm, B. (2015) Landslide impacts in
Damm, B. & Klose, M. (2015) The landslide database for Germany: Closing Germany: A historical and socioeconomic perspective. Landslides, 13,
the gap at national level. Geomorphology, 249, 82–93. 183–199.
Damm, B., Varga, K., Heckmann, T. & Becht, M. (2009) The impact of Krenn, R. & Hergarten, S. (2009) Cellular automaton modelling of
bedrock stratification on landslide susceptibility – an example of GIS- lightning-induced and man made forest fires. Natural Hazards and
based landslide modelling in the Bunter Sandstone areas of northern Earth System Sciences, 9, 1743–48.
Hesse and southern Lower Saxony (Germany). Die Erde, 140(1), Larsen, I.J., Montgomery, D.R. & Korup, O. (2010) Landslide erosion con-
175–193. trolled by hillslope material. Nature Geoscience, 3(4), 247–251.
Densmore, A.L., Ellis, M.A. & Anderson, R.S. (1998) Landsliding and the Li, L., Lan, H. & Wu, Y. (2014) The volume-to-surface-area ratio constrains
evolution of normal-fault-bounded mountains. Journal of Geophysical the rollover of the power law distribution for landslide size. European
Research, 103, 15203–15219. Physical Journal Plus, 129(5), 89.
Drossel, B. & Schwabl, F. (1992) Self-organized critical forest-fire model. Li, L., Lan, H. & Wu, Y. (2016) How sample size can effect landslide size
Physical Review Letters, 69, 1629–1632. distribution. Geoenvironmental Disasters, 3(1), 18.
14 BÜSCHELBERGER ET AL.

Liucci, L., Melelli, L., Suteanu, C. & Ponziani, F. (2017) The role of topogra- Tanyas, H., Allstadt, K.E. & van Westen, C.J. (2018) An updated method
phy in the scaling distribution of landslide areas: A cellular automata for estimating landslide-event magnitude. Earth Surface Processes and
modeling approach. Geomorphology, 290, 236–249. Landforms, 43, 1836–1847.
Lukas, S. & Sass, O. (2011) The formation of Alpine lateral moraines Tebbens, S.F. (2020) Landslide scaling: A review. Earth And Space
inferred from sedimentology and radar reflection patterns: a case Science, 7(1), e2019EA000662. https://doi.org/10.1029/2019EA
study from Gornergletscher, Switzerland. Geological Society, London, 000662
Special Publications, 354(1), 77–92. Tillard, S. & Dubois, J.-C. (1995) Analysis of GPR data: Wave propagation
Malamud, B.D., Morein, G. & Turcotte, D.L. (1998) Forest fires: an example velocity determination. Journal of Applied Geophysics, 33(1–3),
of self-organized critical behavior. Science, 281, 1840–1842. 77–91.
Malamud, B.D., Turcotte, D.L., Guzzetti, F. & Reichenbach, P. (2004) Land- van Dam, R.L. & Schlager, W. (2000) Identifying causes of ground-
slide inventories and their statistical properties. Earth Surface Pro- penetrating radar reflections using time-domain reflectometry and
cesses and Landforms, 29, 687–711. sedimentological analyses. Sedimentology, 47(2), 435–449.
Meunier, P., Hovius, N. & Haines, A.J. (2007) Regional patterns of Van Den Eeckhaut, M., Poesen, J., Govers, G., Verstraeten, G. &
earthquake-triggered to ground motion landslides and their Demoulin, A. (2007) Characteristics of the size distribution of recent
relationto ground motion. Geophysical Research Letters, 34, L20408. and historical landslides in a populated hilly region. Earth and Plane-
Neal, A. (2004) Ground-penetrating radar and its use in sedimentology: tary Science Letters, 256, 588–603.
principles, problems and progress. Earth-Science Reviews, 66(3–4), Varnes, D.J. (1978) Slope movement types and processes. In: Schuster, R.
261–330. L. & Krizek, R.J. (Eds.) Landslides, analysis and control. Washington,
Neuhäuser, B. & Terhorst, B. (2007) Landslide susceptibility assessment D.C.: Transportation research board, National Academy of Sciences,
using “weights-of-evidence” applied to a study area at the Jurassic pp. 11–33.
escarpment (SW-Germany). Geomorphology, 68(1), 12–24. Wimmenauer, W. & Schreiner, A. (1981) Geologische Karte von
Nie, W., Krautblatter, M., Leith, K., Thuro, K. & Festl, J. (2017) A modified Baden-Württemberg 1:25 000, Erläuterungen zu Blatt 8114
tank model including snowmelt and infiltration time lags for Feldberg (Schwarzwald). Analoge Ausgabe. Stuttgart: Geologisches
deep-seated landslides in alpine environments (Aggenalm, Germany). Landesamt Baden-Württemberg & Landesvermessungsamt Baden-
Natural Hazards and Earth System Sciences, 17, 1595–1610. Württemberg.
Olyphant, J., Pelletier, J.D. & Johnson, R. (2016) Topographic correlations Zienert, A. & Fezer, F. (1967) Vogesen-und Schwarzwald-Kare.
with soil and regolith thickness from shallow-seismic refraction con- Eiszeitalter & Gegenwart – Quaternary Science Journal, 18(1), 51–75.
straints across upland hillslopes in the Valles Caldera, New Mexico.
Earth Surface Processes and Landforms, 41, 1684–1696.
SUPPORTING INFORMATION
Pflug, R. (1982) Bau und Entwicklung des Oberrheingrabens.
Wissenschaftliche Buchgesellschaft: Darmstadt. Additional supporting information may be found in the online version
Rahm, G. (1970) Die Vergletscherungen des Schwarzwaldes im Vergleich of the article at the publisher’s website.
zu denjenigen der Vogesen. Alemannisches Jahrbuch, 1966/67,
257–272.
Segoni, S., Rossi, G. & Catani, F. (2012) Improving basin scale shallow land- How to cite this article: Büschelberger, M., Wilk, J., Hergarten,
slide modelling using reliable soil thickness maps. Nat. Hazards, 61, S. & Preusser, F. (2021) Size–frequency distribution of shallow
85–101.
landslides in the Black Forest, Germany. Earth Surface
Smith, M.J., Rose, J. & Gousie, M.B. (2009) The Cookie Cutter: a method
for obtaining a quantitative 3D description of glacial bedforms. Processes and Landforms, 1–14. Available from: https://doi.
Geomorphology, 108(3–4), 209–218. org/10.1002/esp.5237
Stark, C.P. & Hovius, N. (2001) The characterization of landslide size distri-
butions. Geophysical Research Letters, 28, 1091–1094.
Sucre, E., Tuttle, J.W. & Fox, T.R. (2011) The use of ground-penetrating
radar to accurately estimate soil depth in rocky forest soils. Forest
Science, 57, 59–66.

View publication stats