Power-Law Correlations of Landslide Areas in Central Italy

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 15

Earth and Planetary Science Letters 195 (2002) 169^183

www.elsevier.com/locate/epsl

Power-law correlations of landslide areas in central Italy


Fausto Guzzetti a; *, Bruce D. Malamud b , Donald L. Turcotte c ,
Paola Reichenbach a
a
CNR, IRPI Perugia, via della Madonna Alta 126, Perugia 06128, Italy
b
Department of Geography, Kings College London, Strand, London WC2R 2LS, UK
c
Department of Earth and Atmospheric Sciences, Cornell University, Ithaca, NY 14853-1504, USA
Received 12 June 2001; received in revised form 5 November 2001; accepted 20 November 2001

Abstract

We have studied the frequency^area statistics of landslides in central Italy. We consider two data sets. Data set A
contains 16 809 landslide areas in the Umbria^Marche area of central Italy; they represent a reconnaissance inventory
of very old, old, and recent (modern) landslides. The noncumulative frequency^area distribution of these landslides
correlates well with a power-law relation, exponent 32.5, over the range 0.03 km2 6 AL 6 4 km2 . Data set B contains
4233 landslides that were triggered by a sudden change in temperature on 1 January 1997, resulting in extensive melting
of snow cover. An inventory of these snow-melt-triggered landslides was obtained from aerial photographs taken
3 months after the event. These landslides also correlate well with a power-law relation with exponent 32.5, over the
range 0.001 km2 6 AL 6 0.1 km2 . We show that the correlation of data set B is essentially identical to the correlation of
11 000 landslides triggered by the 17 January 1994 Northridge, California earthquake. We attribute a rollover for small
landslides in data set A to incompleteness of the record due to erosion and other processes, and to limitations in the
reconnaissance mapping technique used to complete the inventory. On the other hand, we conclude that rollovers for
small landslides in data set B and the California earthquake data are real and are associated with the surface
morphology. We conclude that the power-law distribution is valid over a wide range of landslide areas and discuss
possible reasons. We also discuss the contribution of the snow-melt- and earthquake-triggered landslide events to the
total landslide inventory. ß 2002 Elsevier Science B.V. All rights reserved.

Keywords: landslides; power law; distribution; fractals; Umbria Italy; Marches Italy

1. Introduction losses due to landslides are greater than losses


from other natural hazards such as £oods, earth-
In many countries, the casualties and economic quakes, volcanic eruptions, and severe storms [1].
Considerable progress has been made in under-
standing the physical processes associated with
the landslide hazard (i.e., hydrology, surface mor-
* Corresponding author. Fax: +39-75-501-4420. phology, triggering mechanisms). However, less
E-mail addresses: f.guzzetti@irpi.pg.cnr.it (F. Guzzetti),
bruce@malamud.com (B.D. Malamud),
work has been done on the probabilistic hazard
turcotte@geology.cornell.edu (D.L. Turcotte), assessments of landslides [2]. Important questions
p.reichenbach@irpi.pg.cnr.it (P. Reichenbach). concerning probabilistic assessments include:

0012-821X / 02 / $ ^ see front matter ß 2002 Elsevier Science B.V. All rights reserved.
PII: S 0 0 1 2 - 8 2 1 X ( 0 1 ) 0 0 5 8 9 - 1

EPSL 6085 11-2-02


170 F. Guzzetti et al. / Earth and Planetary Science Letters 195 (2002) 169^183

(1) gWhat fraction of triggered landslide events 2. Computer models and self-organized criticality
are associated with speci¢c triggers (i.e., intense
or prolonged rainfall, earthquake, snow-melt) ? Many empirical observations have shown that
(2) What is the frequency of occurrence of a land- both earthquakes and landslides obey power-law
slide of a speci¢ed size in a speci¢ed region and (fractal) frequency^size statistics. However, is
time period? (3) Can the rate of occurrence of there an underlying physical basis for this behav-
small landslides in a region be extrapolated to ior? A number of authors [17,18] have argued
predict the rate of occurrence of large landslides ? that the power-law behavior of earthquakes is a
In this context, it is of interest to compare land- consequence of self-organized critical behavior. A
slides with earthquakes. The physical processes rigorous de¢nition of self-organized criticality is
associated with both phenomena are extremely elusive, but a working de¢nition is that the `input'
complex. Despite these complexities, the fre- to a complex system is nearly constant, whereas
quency-magnitude statistics of earthquakes satisfy the `output' is a series of events or `avalanches'
the Gutenberg^Richter relation [3]: that follow a power-law (fractal) frequency^size
distribution.
log N CE ˆ 3bm ‡ a …1† Slider block models are considered to be a sim-
ple analog for regional seismicity [19]. In these
where NCE is the cumulative number of earth- models, slider blocks in an array are connected
quakes that occur in a speci¢ed region and time to each other and to a driver plate by springs.
interval with magnitudes greater than or equal to The blocks interact frictionally with a surface
m, a and b are constants. This relation is valid and are dragged over the surface by the constant
over a wide range of earthquake sizes both re- motion of the driver plate. The blocks exhibit
gionally and on a worldwide basis [4]. stick^slip behavior and they move forward in a
It can be shown [5] that Eq. 1 is entirely equiv- sequence of slip events. Under many circumstan-
alent to the power-law (fractal) relation: ces, the noncumulative frequency^area statistics
K of the slip events satisfy the power-law (fractal)
N CE ˆ CA3
E …2† relation:
L
with C a constant, AE the earthquake rupture N E ˆ C 0 A3
E …3†
area, and K = b from Eq. 1. The frequency^area
statistics of earthquakes satisfy power-law (frac- where NE is the (noncumulative) number of slip
tal) statistics. events with area AE , the number of blocks that
There is accumulating evidence that the fre- participate in the event, and CP and L are con-
quency^area statistics of landslides are also stants. A number of authors have shown that
power-law [6^14]. Inventories of landslides usually for a variety of slider block models L is in the
give the total area of the landslide, including the range 1.0^1.3 [20]. However, some nonconserva-
depletion area and run out. It can be argued that tive models [21] have been proposed in which L is
it would be preferable to use either landslide fail- a function of the parameters in the model, with
ure area or landslide volume, but these quantities values obtained for L as large as 2.5.
are di¤cult to determine. Where landslide volume For a noncumulative power-law distribution
has been determined, the frequency^volume statis- (Eq. 3) with L = 1.0, the corresponding cumulative
tics also obey a power-law function [15,16]. In this distribution obtained by summing or integrating
paper, we will show that two landslide inventories will be logarithmic. For a noncumulative power-
from central Italy provide further strong support law distribution (Eq. 3) with exponent L s 1.0, the
for the hypothesis that landslide frequency^area corresponding cumulative distribution obtained
statistics are power-law (fractal) for landslides by integration or summing will be a power-law
that exceed a minimum length scale. (Eq. 2) with exponent K = L31.0. We use the ex-

EPSL 6085 11-2-02


F. Guzzetti et al. / Earth and Planetary Science Letters 195 (2002) 169^183 171

ponent K for linearly binned noncumulative versal, indicating an important underlying


power-law distributions and the exponent L for physics.
cumulative distributions. A logarithmically It has been proposed that self-organized critical
binned noncumulative distribution is equivalent behavior can be explained as a self-similar inverse
to a cumulative distribution. cascade [18,26,27]. This explanation is given in
The robust power-law behavior of the slider terms of metastable regions. In terms of land-
block model has been associated with self-orga- slides, the metastable region is the region over
nized criticality and this is then associated with which a landslide will propagate once initiated.
earthquakes [17]. Although both the model and As the landscape evolves, metastable regions
the actual earthquakes yield robust power-law fre- grow, principally by the coalescence of smaller
quency^size distributions under a wide variety of metastable regions. This growth by coalescence
circumstances, the power-law exponent in Eq. 2 is the self-similar inverse cascade. We will return
for earthquakes is KV1.0, equivalent to LV2.0 in to this explanation after we have presented the
Eq. 3, compared to LV1.0 for most slider block landslide inventory data.
models.
The concept of self-organized criticality was
¢rst given by Bak et al. [22,23] in order to explain 3. Landslides in central Italy
the behavior of their `sandpile' model. In this
model, there was a square grid of boxes and at We ¢rst turn our attention to landslides in the
each time step a particle was dropped into a ran- Umbria^Marche area of central Italy [28]. The
domly selected box. When a box accumulated Umbria region lies on the west side of the Apen-
four particles, the particles were redistributed to nines mountain chain and is drained by the Tiber
the four neighboring boxes, or in the case of edge River that £ows into the Tyrrhenian Sea. The
boxes, were lost from the grid. Redistributions of Marche region lies on the eastern side of the
particles could lead to further instabilities with Apennines divide and the region drains into the
`avalanches' of particles lost from the grid. Be- Adriatic Sea. These regions have a long history of
cause of this avalanche behavior, this was called catastrophic landslides that date back to the
a `sandpile' model. The noncumulative frequency^ Etruscan and Roman periods.
area distribution of model avalanches was found A variety of rock types are found in the out-
to satisfy the noncumulative frequency^area crops in these regions [28]. Hard rocks include
power-law distribution (Eq. 3) with N the number layered and massive limestones, cherty limestones,
of avalanches and A the number of particles lost sandstones, pyroclastic deposits, travertines and
from the grid, with again LV1. conglomerates. Weak rocks include marls, shales,
Shortly after the `sandpile' model was pro- sands, silty clays, and overconsolidated clays. Soft
posed, a number of laboratory studies were rocks are marine and continental clays, silty clays,
undertaken to determine whether actual sandpiles and shales. The morphology and structural setting
exhibit self-organized criticality. A variety of fre- are the result of two tectonic phases. A compres-
quency^size statistics for `avalanches' were found sive phase, late Miocene to early Pliocene in age,
[24,25]; in some cases the results were consistent was followed by an extensional phase of Pliocene
with a power-law relation (Eq. 3), but in other to recent age. The compressive deformation pro-
cases they were not. duced large anticlines, corresponding to major di-
Noever [9] and many other authors have asso- vides, and synclines associated with thrusts and
ciated the apparent power-law frequency^area transcurrent faults. The extensional tectonic phase
statistics of landslides with the `sandpile' model. produced normal faults that formed intra-moun-
However, there is clearly a large extrapolation tain basins and valleys. Landslides include falls
from the model to actual landslides. Neverthe- and topples in hard rocks, soil slips in the collu-
less, the power-law behavior of landslides above vial cover, rotational slides in homogeneous soft
a minimum length scale appears to be quite uni- rocks, translational slides in well-bedded soft and

EPSL 6085 11-2-02


172 F. Guzzetti et al. / Earth and Planetary Science Letters 195 (2002) 169^183

hard rocks, earth £ows, and complex and com- [29,30]. The total of all landslides mapped is 1831
pound slides where alternating layers of hard km2 , or about 10% of the Umbria^Marche study
and soft rocks crop out [28]. area. Table 1 gives the distribution of the type,
We ¢rst consider a regional inventory of 16 809 number and size of the mapped landslides. Land-
landslides in the Umbria^Marche area, which is slides were mapped according to a simpli¢ed ver-
slightly larger than the political regions of Umbria sion of Varnes classi¢cation [29^31].
and Marche. The total study area is about 18 000 The frequency^area statistics for landslides are
km2 . The inventory (Figs. 1 and 2) was obtained often presented using cumulative statistics; the
by the systematic analysis of some 2000 black and cumulative number of landslides NCL with areas
white vertical aerial photographs taken at a nom- greater than AL is plotted as a function of AL .
inal scale of 1:33 000 and supplemented by de- However, the frequency^area distributions of
tailed geomorphic investigations at selected sites sandpile model avalanches, presented in Eq. 3

Fig. 1. Spatial distribution of landslides mapped in the Umbria^Marche (central Italy) area. Shown is the spatial location of
16 809 landslides (data set A) identi¢ed through the interpretation of about 2000 black and white 1:33 000 aerial photographs
£own 1954^1956. This ¢gure is a reduced version of 1:100 000 color landslide inventory maps [29,30], an example of which is
given in Fig. 2 (represented as a small square in the upper left of the above ¢gure). Boundaries for the Umbria (left) and Marche
(right) political regions are shown as thick black lines.

EPSL 6085 11-2-02


F. Guzzetti et al. / Earth and Planetary Science Letters 195 (2002) 169^183 173

Fig. 2. Portion of the landslide inventory map for the Marche Region, central Italy. This map was prepared through 1:33 000
aerial photograph interpretation. The above portrays a small portion (21U19 km) of the original 1:100 000 landslide inventory
map [30]. This ¢gure's location is given as a small square in the upper left of Fig. 1.

with power-law exponent LV1, were noncumula- method to convert cumulative distributions to
tive, i.e., the data are given in unit steps of num- noncumulative distributions. We start with the
bers of particles lost from the grid. Since we wish cumulative distribution NCL (AL ). We then de¢ne
to compare the measured landslide distributions a noncumulative distribution in terms of the neg-
with the model results, we will consider only non- ative of the derivative of the cumulative distribu-
cumulative distributions. Binning the data is one tion 3 dNCL /dAL . A close approximation of the
way to represent noncumulative distributions derivative, dNCL /dAL , is the slope of the best-¢t
(with care taken to correctly normalize the bins line to a speci¢ed number of adjacent cumulative
and specify whether bin sizes are in linear or log- data points (typically 5).
arithmic coordinates). We use here the derivative The noncumulative number^area distribution

EPSL 6085 11-2-02


174 F. Guzzetti et al. / Earth and Planetary Science Letters 195 (2002) 169^183

Table 1
Distribution of type, number and area for two landslide inventories in central Italy
Landslide classi¢cation according to Varnes [31] Number of landslides Areas of mapped landslides (km2 )
Minimum Maximum Average Total
Data set A: Regional inventory of old and recent landslides in the Umbria^Marche area.
Slides 5 745 7.00U1034 2.80U100 8.93U1032 5.13U102
Slides/£ows 1 035 7.43U1033 4.32U100 1.68U1031 1.74U102
Flows 2 652 3.07U1033 2.57U100 5.99U1032 1.59U102
Flows: debris £ows 309 3.48U1033 1.48U100 8.20U1032 2.53U101
Complex 447 1.46U1032 5.29U100 2.30U1031 1.03U102
Multiple 439 1.42U1032 5.12U100 5.96U1031 2.62U102
Sur¢cial (landslide area) 3 759 4.50U1033 3.73U100 1.09U1031 4.09U102
Indistinct (? slides) 2 423 4.66U1033 1.27U100 7.67U1032 1.86U102
All of data set A 16 809 7.00U1034 5.29U100 1.09U1031 1.83U103
Data set B: Snow-melt-triggered landslides in Umbria.
Falls 10 2.28U1034 3.34U1032 4.34U1033 4.34U1032
Slides: shallow 2 225 5.80U1035 1.20U1031 1.62U1033 3.60U100
Slides: deep-seated 1568 3.95U1035 1.56U1031 5.39U1033 8.45U100
Flows 350 8.86U1035 7.31U1033 8.67U1034 3.03U1031
Flows: multiple 27 5.20U1034 3.42U1032 4.81U1033 1.30U1031
Indistinct (? slides) 53 5.01U1034 2.07U1032 3.87U1033 2.05U1031
All of data set B 4 233 3.95U1035 1.56U1031 3.01U1033 1.27U101

of the 16 809 landslides in the Umbria and Mar-


che regions is given in Fig. 3 (data set A). These
medium and large landslides correlate well with
the power-law (fractal) relation :
dN CL L
3 ˆ C 0 A3
L …4†
dAL

taking L = 2.5 and CP = 300 (with AL in km2 ). This


data set deviates from the power-law scaling for
1=2
AL 6 1031 km2 (AL W300 m). The largest num-
ber of landslides have an area AL = 2U1032 km2
1=2
(AL W140 m).
We next consider an inventory of 4233 land-
slides in the Umbria region that were triggered
by a sudden change in temperature on 1 January
1997. Four examples of landslides triggered by
Fig. 3. Noncumulative frequency^area distributions of central this rapid snow-melt event are given in Fig. 4.
Italian landslides. The noncumulative frequency of landslides
This inventory [32] was obtained from the inter-
3dNCL /dAL with area AL is given as a function of landslide
area AL for two data sets. Data set A represents an inven- pretation of about 400 aerial photographs taken
tory of 16 809 old and recent landslides mapped in the Um- at a 1:20 000 nominal scale. The photographs
bria^Marche area (see Fig. 1 for the spatial distribution of were taken 3 months after the snow-melt event
these landslides, Fig. 2 for an example of their detailed map- and supplemented by ¢eld mapping carried out
ping). Data set B represents 4233 landslides triggered by a
at a 1:10 000 scale. The total study area is about
January 1997 rapid snow melting in Umbria (see Fig. 4 for
photographs of these landslides, Fig. 5 for their spatial distri- 2000 km2 . The spatial distribution of these land-
bution, and Fig. 6 for an example of their detailed map- slides is given in Figs. 5 and 6.
ping). The noncumulative number^area distribution

EPSL 6085 11-2-02


F. Guzzetti et al. / Earth and Planetary Science Letters 195 (2002) 169^183 175

Fig. 4. Four examples of landslides triggered by the January 1997 rapid snow-melting in Umbria. (A) To the right of the build-
ings can be seen a deep-seated slide (slump earth £ow) of about 50U180 m. (B) In the foreground, cutting across the paved
road, can be seen part of a deep-seated slide (rotational) of about 30U100 m. (C) On the slope can be seen a shallow slide (soil
slip) of about 40U40 m. (D) In the middle can be seen deep-seated slides (complex), with the largest (on the left) extending
about 100U250 m.

of these landslides is also given in Fig. 3 (data set The overlap of the two data sets illustrated in
B). Note that the two vertical scales have been Fig. 3 shows that the power-law scaling is valid
adjusted so that the two data sets overlap. These over the range of landslide areas, 1033
landslides also correlate well with the power-law km2 6 AL 6 4 km2 , i.e., for lengths scales greater
1=2
relation (Eq. 4), again taking L = 2.5 and CP = 0.3 than AL W30 m. The inventory of the snow-
(with AL in km2 ). This data set deviates from the melt-induced landslides (data set B) is certainly
1=2
power-law scaling for AL 6 1033 km2 (AL W30 more complete than the historical landslides
m). The largest number of landslides have an (data set A). We conclude that the rollover (at
1=2
area AL = 6U1034 km2 (AL W25 m). AL W2U1032 km2 ) of the landslides in the region-
In the study area, small landslides exhibit as- al inventory in data set A is due to the inability to
pect ratios (the ratio of down-slope length to the measure the areas of the smaller landslides on the
across-slope width) near unity and large landslides aerial photographs and/or due to erosion and oth-
can have aspect ratios greater than 2.5. Thus for er wasting processes.
data set B, assuming an aspect ratio of 1.0 for In data set A, the reconnaissance technique
small landslides, the deviation from the power- used to map landslides from aerial photographs
law scaling at AL V1033 km2 can be associated limited the size of features that could be recog-
1=2
with a linear dimension AL W30 m. nized. Due to the scale (1:33 000) and the quality

EPSL 6085 11-2-02


176 F. Guzzetti et al. / Earth and Planetary Science Letters 195 (2002) 169^183

Fig. 5. Spatial distribution of Umbria landslides triggered by rapid snow-melt in January 1997 [32]. Shown is the spatial location
of 4233 landslides (data set B) identi¢ed through the interpretation of 1:20 000 aerial photographs £own 3 months after the
event. These aerial photographs, supplemented by ¢eld surveys, resulted in mapping being carried out at a scale of 1:10 000, an
example of which is given in Fig. 6 (represented as the small rectangle in the lower middle portion of the ¢gure above).

of the photographs, the smallest landslide consis- freshness of the snow-melt-triggered landslides,
tently recognized had an area of about 1 hectare and the quality and scale (1:20 000) of the
(AL = 1032 km2 or about 3 mmU3 mm on the aerial photographs, the smallest landslide area
photograph). The evidence for the existence of consistently mapped is about 2.5U1034 km2
1=2
the smaller landslides in the regional inventory (AL W16 m), i.e., lower than the dimension
1=2
has been lost due to erosion and human action AL W30 m at which the data set deviates from
[28]. Data set A is thus incomplete for landslides the power-law relation. This conclusion is sup-
below AL W1032 km2 . ported by ¢eldwork carried out to map landslides
On the other hand, the rollover of the land- before the aerial photographs were taken, and by
slides in data set B is not an artifact, and is the successive ¢eld surveys performed to check the
result of the truncation of the power-law scaling quality of the aerial photographs interpretation
1=2
at a length scale of about AL W30 m. Due to the [32]. Ground surveys con¢rmed that the landslide

EPSL 6085 11-2-02


F. Guzzetti et al. / Earth and Planetary Science Letters 195 (2002) 169^183 177

Fig. 6. Portion of the landslide inventory map for the Umbria Region, central Italy. The above portrays a small portion
(4.8U4.4 km) of the inventory map created from Umbria snow-melt-triggered landslides [31]. This ¢gure's location is given as a
small rectangle in the lower middle of Fig. 5.

inventory, as interpreted from the aerial photo- with the identi¢cation of the largest failures as
graphs, is virtually complete. unique (single) landslides and that some of
Table 1 shows that the landslide areas in the the smallest landslides may have been missed by
two (combined) data sets presented in Fig. 3 our mapping, landslide areas in the combined
range from 4.0U1035 km2 (the smallest landslide data sets span a very large range. Fig. 3 shows
mapped in data set B, the snow-melt-triggered that the power-law scaling is valid from 1033
inventory) to 5.3 km2 (the largest landslide of km2 6 AL 6 4 km2 . The upper limit of this
data set A, the regional inventory). Regardless range may be constrained by the geological set-
of the fact that some uncertainty is associated ting, which controls the extent of the largest

EPSL 6085 11-2-02


178 F. Guzzetti et al. / Earth and Planetary Science Letters 195 (2002) 169^183

slopes where landslides can occur. It is worth not- events certainly have occurred, and landslides
ing that the area of the 10 largest landslides in are also triggered by rainfall and earthquakes
data set A (AL s 3 km2 ) is of the same order of [28]. One interpretation is that the regional land-
magnitude as the area of the largest hillslopes in slide inventory represents landslides that have oc-
the Umbria^Marche area. Due to the present curred in the last 10 000 yr (i.e., since the last
physiographic setting, single landslide areas glaciation).
much larger than 4^5 km2 are not expected in An important question for understanding the
this region. geomorphological evolution of an area where
What is the contribution of the snow-melt-trig- slope failures actively shape the landscape, as
gered landslides to the total landslide inventory? well as for landslide hazard assessment, is the rel-
In the two available landslide inventories, the ative importance of triggered landslide events in
snow-melt-triggered landslides (data set B) have the long-term landslide inventory. Certainly earth-
a total landslide area of 12.7 km2 and represent quakes, rapid snow-melt, and high-intensity or
0.7% of the total landslide area of 1831 km2 for prolonged rainfall can each trigger many land-
the long-term (regional) landslides (data set A). slides. But, what are the frequency-magnitude sta-
However, the frequency^area distributions as pre- tistics for the resulting landslide events? Are most
sented in Fig. 3 tell a di¡erent story. Assume that landslides included in the largest triggered land-
both inventories are complete for the larger land- slide events or does the regular background of
slides. Evidence that this is true can be seen in landslides dominate the landslide inventory ? The
Fig. 3 where the larger landslides for both data comparison made in Fig. 3 provides a rational
sets A and B have the same power-law distribu- basis for quantifying the intensity of a triggered
tions. Comparing the CP values from Eq. 4 for landslide event. Simply taking the number of
both distributions (Fig. 3), data set B has counted landslides is inappropriate; doing this
CP = 0.3 and data set A has CP = 300; the ratio is for the comparison given in Fig. 3 would lead
1:1000. The area under the two frequency^area to a serious error, as data set B is relatively com-
distributions represents the relative total landslide plete (i.e., all or a high percentage of the triggered
area for each data set. Changing CP in Eq. 4 by landslides are counted) and A is incomplete.
1000 is the same as changing the area under Based on the actual landslide inventories, the rel-
the frequency^area curve by a factor of a 1000. ative intensities (number of landslides) of the two
Therefore, based on the frequency^area distribu- data sets would be 4233/16 809 or approximately
tions, the total area of the snow-melt-triggered 1/4. This is very di¡erent from our previous con-
landslides (data set B) represents 0.1% of the clusion based on the power-law distributions for
total area of the long-term (regional) landslides each data set, and the ratio of 1:1000 for the
(data set A). The lower value of 0.1% (vs. 0.7% two values of CP, that the relative intensities are
as discussed at the beginning of this paragraph) is 1/1000.
a re£ection of the fact that data set A is incom- Power-law correlations can begin to provide the
plete. basis for quantifying triggered landslide events. If
Based on empirical observations [32], we esti- the landslide inventory is complete (such as in
mate that the snow-melt-triggered landslide event data set B), then the intensity of the landslide
(the 4233 landslides in data set B) has a return event will be the number of landslides that occur
period of about 25 yr, i.e., there have been two as part of the event. However, if the landslide
snow-melt-triggered landslide events in the last inventory is incomplete (such as data set A),
50 yr. Using 20^25 yr as a very rough estimate then as we have shown, one can calculate the
for the return period of one snow-melt-triggered landslide intensity of the inventory from CP and
landslide event (data set B), 1000 of these trig- L in Eq. 4. Quantitative estimate of the extent and
gered landslide events (data set A) would intensity of triggered landslide events might aid in
occur over a period of 20 000^25 000 yr. This is improving our ability to assess landslide hazard,
only an estimate since larger triggered landslide and the associated risk.

EPSL 6085 11-2-02


F. Guzzetti et al. / Earth and Planetary Science Letters 195 (2002) 169^183 179

4. California landslides

For comparison, we now consider the distribu-


tion of 11 000 landslides over an area of 10 000
km2 triggered by the 17 January 1994 Northridge
(California) earthquake. An inventory of these
landslides was carried out by Harp and Jibson
[33,34]. They compared 1:60 000 scale aerial pho-
tographs taken the morning after the earthquake
with aerial photographs taken previously. The
digitized photographs were supplemented by ¢eld-
work. They estimated that the inventory is nearly
complete for landslides with a length scale greater
1=2
than AL W5 m. The noncumulative number^area
distribution of these landslides is given in Fig. 7.
Fig. 7. Noncumulative frequency^area distribution of 11 000
These landslides correlate well with the power-law landslides triggered by the 17 January 1994, Northridge, Cal-
relation (Eq. 4) taking L = 2.3 and CP = 1.0 (AL in ifornia earthquake [33,34]. The noncumulative frequency of
km2 ). This data set deviates from the power-law landslides 3dNCL /dAL with area AL is given as a function of
1=2
scaling for AL 6 1033 km2 (AL W30 m). The landslide area AL .
largest number of landslides have an area
1=2
AL = 6U1034 km2 (AL W25 m). Hovius et al. [13] have given a number^area in-
The data for these earthquake-triggered land- ventory of 4984 landslides in the montane zone
slides in California are remarkably similar to the east of the Alpine fault in New Zealand. They
snow-melt-triggered landslides in central Italy. estimated that these landslides occurred over a
The best-¢t power-law exponent is 2.3 for the 40^60-yr period. Their logarithmically binned
California data and is 2.5 for the Italian data. data correlated well with the power-law relation
The rollovers for small landslides occur at essen- (Eq. 2) taking K = 1.17. This correlation is valid
1=2
tially the same landslide areas for the two data over the range AL = 7U1033 km2 (AL W84 m) to
sets. The relative intensities of the snow-melt-trig- AL = 1 km2 . Logarithmic binning is equivalent to
gered landslides can be obtained from the corre- a cumulative distribution, therefore the equivalent
lations given in Figs. 3 and 7. A comparison made noncumulative power-law exponent (Eq. 4) is
at AL = 1032 km2 shows that the intensity of the L = 2.17.
California landslide event was about twice the in- Hovius et al. [14] have given a number^area
tensity of the Italian landslide event. Since both inventory of 1040 fresh landslides in the Ma-An
inventories appear to be relatively complete, the and Wan-Li catchments on the eastern side of the
relative intensities are proportional to the number Central Range in Taiwan. They estimated the
of landslides, i.e., 11 000/4233 = 2.6. landslides have an age of less than 10 yr. Their
logarithmically binned data correlate well with the
power-law relation (Eq. 2) taking K = 1.66. This
5. Discussion correlation is valid over the range AL = 1033
1=2
km2 (AL = 30 m) to AL = 5U1032 km2 . The
We now compare the results given above with equivalent noncumulative power-law exponent
previous studies. Fujii [35] obtained a cumulative (Eq. 4) is L = 2.66. It is interesting to note that
number^area inventory of 800 landslides caused the power-law exponent and the deviation from
by a heavy rainfall event in Japan. An excellent power-law scaling for this data set and the two
correlation with the power-law relation (Eq. 2) landslide events we gave above (Italy and Califor-
was found taking K = 0.96. The equivalent noncu- nia) are very similar.
mulative power-law exponent (Eq. 4) is L = 1.96. Noncumulative number^area distributions for

EPSL 6085 11-2-02


180 F. Guzzetti et al. / Earth and Planetary Science Letters 195 (2002) 169^183

several regional landslide inventories have been bined, the error is reduced to L = 2.5 þ 0.10, and
given by Malamud and Turcotte [4]. Results for the power-law behavior is valid for 0.001
1130 landslides from the Challana Valley, Bolivia km2 6 AL 6 4 km2 . This combination suggests
correlate well with the noncumulative power-law that the power-law distribution of landslides
relation (Eq. 4) taking L = 2.6; 3243 landslides may be valid over a wider range of values than
from the Akishi Range, central Japan correlate shown by previous studies. Furthermore, the com-
well taking L = 3.0; and 709 earthquake-induced bined data set provides an estimate of the impor-
landslides from Eden Canyon, Alameda, CA, tance of individual triggering events. However,
USA, correlate well taking L = 3.3. due to the error bars for the di¡erent landslide
Hungr et al. [15] have given cumulative fre- data sets we have compared, it is di¤cult to con-
quency^volume inventories for 1937 rock falls clude that the variations in published values of L
and rock slides along the main transportation cor- are necessarily associated with geology. It is
ridors of southwestern British Columbia. The clearly desirable to obtain more landslide invento-
data correlate reasonably well with a power-law ries to determine whether the values of L are re-
relation taking the slope to be 30.5 þ 0.2. Assum- lated to type of physiographic setting. It is also
ing the volume V correlates with the area accord- important to emphasize studies of fresh landslides
ing to VVA3=2 , the equivalent cumulative fre- where inventories are complete and the smaller
quency^area power-law exponent (Eq. 2) is landslides have not been degraded by mass-wast-
K = 0.75 þ 0.30, and the equivalent noncumulative ing processes.
frequency^area power-law exponent (Eq. 4) is Another important question concerning land-
L = 1.75 þ 0.30. slides is the contribution of landslides triggered
Dai and Lee [16] have given cumulative fre- by an event (by intense or prolonged rain, earth-
quency^volume inventories for 2811 landslides in quake, or snow-melt) to the total regional land-
Hong Kong that occurred during the period slide inventory. This is a di¤cult question to ad-
1992^1997. The data correlate reasonably well dress at this time since there are very little data on
with a power-law taking the slope to be 30.8. the intensities of landslide events. We have shown
Again assuming VVA3=2 , the equivalent cumula- that the total number of landslides during an
tive frequency^area power-law exponent (Eq. 2) event is a measure of the intensity of the event
is K = 1.2, and the equivalent noncumulative if the inventory of landslides is complete down
frequency^area power-law exponent (Eq. 4) is to scales of 5^10 m. It is clearly desirable to sys-
L = 2.2. tematically carry out inventories of triggered land-
Although there is certainly variability, many slide events in order to establish the expected re-
landslide inventories appear to satisfy noncumu- currence rates of such events. An important
lative power-law frequency^area statistics, with an question is whether these events occur on a regu-
exponent L = 2.5 þ 0.5. An important question is lar basis, both regionally and globally.
whether this relatively large scatter in values of L Using the Gutenberg^Richter relation (Eq. 1),
is due to scatter in the data or is due to di¡erent the recurrence rate of earthquakes can be speci¢ed
values of L associated with di¡erent physiographic both regionally and globally. The rate of occur-
settings (i.e., changes in lithology, geology, mor- rence of small earthquakes can be extrapolated to
phology, climate history, etc.). For a single data predict the occurrence rate of larger earthquakes
set, the error bar on L can be relatively large. For [37]. If landslides occur on a regular basis, it may
instance, depending on where the tail is ¢t, Stark be possible to use the frequency of occurrence of
and Hovius [36] ¢nd variations on the order of small landslides to forecast the risk of catastroph-
L = 2.88 þ 0.22. Other authors' data sets that we ically large landslides. However, if triggered land-
have discussed have even larger variability. This slide events are dominant, then the risks of the
variation is also quite evident in our Fig. 3, where trigger themselves ^ earthquakes, snow-melt, or
for each data set an error bar of L = 2.5 þ 0.25 is intense rain ^ will dominate the forecasts.
reasonable. But, when the two data sets are com- The rollover of the data away from the power-

EPSL 6085 11-2-02


F. Guzzetti et al. / Earth and Planetary Science Letters 195 (2002) 169^183 181

law correlation for small landslides appears to be strength. Since the gravitational body force on
systematic and requires an explanation. This roll- the basal layer driving the landslide is proportion-
over occurs at a characteristic length scale of ap- al to h, the cohesive force will prevent small land-
1=2
proximately AL W30 m. For their two data sets slides. In this explanation, the rollover is attrib-
from Taiwan and New Zealand, Stark and Ho- uted to the transition from a resistance controlled
vius [36] provide a statistical correlation with the by friction (large landslides) to a resistance con-
entire landslide distribution, using a double Pa- trolled by cohesion (small landslides).
reto probability distribution. They argue that We believe the evidence is convincing that me-
landslides exhibit two scaling regions and that dium and large landslides consistently satisfy
the rollover scale is purely an artifact of mapping power-law (fractal) frequency^area statistics, but
resolution for the data sets that they consider. why ? One explanation is to simply invoke the
For data set B considered in this paper, the sandpile model as an analog for landslides in
rollover cannot be attributed to mapping resolu- the same way that slider block models are associ-
tion and we attribute the rollover to a termination ated with earthquakes. However, the noncumula-
in the self-similar power-law scaling valid for tive power-law exponent for landslides is L = 2.5 þ
large landslides. Pelletier et al. [12] also made 0.5 whereas the noncumulative power-law expo-
this argument and provided two alternative ex- nent for the sandpile model avalanches is LV1.0.
planations for the rollover. The ¢rst is that the In order to explain this di¡erence, Pelletier et al.
scale at which the rollover occurs, for scales less [12] combined a slope stability analysis with self-
than about 30 m, is also the scale on which well- a¤ne topography and soil-moisture content and
de¢ned stream networks form in soil-mantled found a power-law noncumulative frequency^area
landforms. The gullying associated with stream distribution with L = 2.6. Hergarten and Neuge-
and river networks would be expected to play a bauer [39] used a numerical model combining
signi¢cant role in the geometry of landslides for slope stability and mass movement and found
climatically controlled failures, such as those of an approximation to a power-law distribution
data set B, or other landslides triggered by rain- with an exponent of LV2.1. These authors [40]
fall. For climatically controlled landslides, water later used a cellular-automata model with time-
and groundwater are important issues and both dependent weakening, similar to the sandpile
relate to the size of a slope, which in turn depends model, and found a power-law distribution with
on the pattern and density of the river network. LV2.0. While it is certainly possible to develop
For seismically induced landslides, the relation- SOC-like models that reproduce the observed
ship is less clear. These landslides, and particu- power-law dependence of actual data, there is a
larly rock-falls, occur where slopes are steeper, real question whether these models are realistic in
where seismic shaking concentrates, and where terms of the governing physics. Certainly much
the rock is weaker. A second explanation for the more work remains to be done.
rollover is associated with the failure criteria [38]: A relatively simple inverse cascade model [15]
can qualitatively explain the power-law behavior
d f ˆ d 0 ‡ … b gh3 W †f …5† of both SOC models and actual landslides. The
cascade model involves metastable regions over
where df is the failure shear stress of the land- which landslides will propagate once triggered.
slide's basal (sliding) surface, d0 is the cohesive These metastable regions grow due to tectonic
strength of this surface, b is the density of the mountain building processes. The growth is pri-
rock, g is the Earth's gravitational acceleration, marily through the coalescence of smaller meta-
h is the depth to the basal surface, W is the pore stable regions and the growth of the metastable
pressure in the rock or soil, and f the coe¤cient of regions is a scale invariant process that must give
friction on the basal surface. For small, shallow power-law statistics. When a triggering event, i.e.,
landslides, bgh is small because h is small and an earthquake, occurs, the landslides propagate
failure is mostly controlled by the cohesive over the metastable regions. Since the metastable

EPSL 6085 11-2-02


182 F. Guzzetti et al. / Earth and Planetary Science Letters 195 (2002) 169^183

regions have a power-law frequency^area distri- cations for the frequency^size distribution of landslides,
Eng. Geol. 48 (1997) 255^268.
bution, so will the resulting landslides. However,
[13] N. Hovius, C.P. Stark, P.A. Allen, Sediment £ux from a
further work is required to provide a comprehen- mountain belt derived by landslide mapping, Geology 25
sive explanation for the landslide statistics de- (1997) 231^234.
scribed in this paper. [14] N. Hovius, C.P. Stark, C. Hao-Tsu, L. Jinn-Chuan, Sup-
ply and removal of sediment in a landslide-dominated
mountain belt: Central Range, Taiwan, J. Geol. 108
(2000) 73^89.
Acknowledgements [15] O. Hungr, S.G. Evans, J. Hazzard, Magnitude and fre-
quency of rock falls and rock slides along the main trans-
This work has been supported by NASA Grant portation corridors of southwestern British Columbia,
NAG5-9067. We thank reviewers Stefan Hergart- Can. Geotech. J. 36 (1999) 224^238.
[16] F.C. Dai, C.F. Lee, Frequency^volume relation and pre-
en and Alberto Carrara, whose comments sub-
diction of rainfall-induced landslides, Eng. Geol. 59
stantially improved this paper. This paper is (2001) 253^266.
CNR-GNDCI Pub. No. 2230.[AC] [17] P. Bak, C. Tang, Earthquakes as a self-organized critical
phenomenon, J. Geophys. Res. 94 (1989) 15635^15637.
[18] D.L. Turcotte, Self-organized criticality, Rep. Prog. Phys.
62 (1999) 1377^1429.
References [19] R. Burridge, L. Knopo¡, Model and theoretical seismic-
ity, Seism. Soc. Am. Bull. 57 (1967) 341^371.
[1] R.L. Schuster, R.W. Fleming, Economic losses and fatal- [20] J.M. Carlson, J.S. Langer, Mechanical model of an earth-
ities due to landslides, Am. Assoc. Eng. Geol. Bull. 23 quake fault, Phys. Rev. A40 (1989) 6470^6484.
(1986) 11^28. [21] Z. Olami, H.J.S. Feder, K. Christensen, Self-organized
[2] F. Guzzetti, A. Carrarra, M. Cardinali, P. Reichenbach, criticality in a continuous, nonconservative cellular au-
Landslide hazard evaluation: A review of current tech- tomaton modeling earthquakes, Phys. Rev. Lett. 68
niques and their applicability in a multi-scale study, cen- (1992) 1244^1247.
tral Italy, Geomorphology 31 (1999) 181^216. [22] P. Bak, C. Tang, K. Wiesenfeld, Self-organized criticality:
[3] B. Gutenberg, C.F. Richter, Seismicity of the Earth and An explanation of 1/f noise, Phys. Rev. Lett. 59 (1987)
Associated Phenomena, Princeton University Press, 381^384.
Princeton, NJ, 1954, 310 pp. [23] P. Bak, C. Tang, K. Wiesenfeld, Self-organized criticality,
[4] B.D. Malamud, D.L. Turcotte, Self-organized criticality Phys. Rev. A38 (1988) 364^374.
applied to natural hazards, Nat. Hazards 20 (1999) 93^ [24] S.R. Nagel, Instabilities in a sandpile, Rev. Mod. Phys. 64
116. (1992) 321^325.
[5] D.L. Turcotte, Fractals and Chaos in Geology and Geo- [25] J. Feder, The evidence for self-organized criticality in
physics, 2nd edn., Cambridge University Press, Cam- sandpile dynamics, Fractals 3 (1995) 431^443.
bridge, 1997, 398 pp. [26] D.L. Turcotte, B.D. Malamud, G. Morein, W.I. New-
[6] I.E. Whitehouse, G.A. Gri¤ths, Frequency and hazard of man, An inverse-cascade model for self-organized critical
large rock avalanches in the central Southern Alps, New behavior, Phys. A 268 (1999) 629^643.
Zealand, Geology 11 (1983) 331^334. [27] A. Gabrielov, W.I. Newman, D.L. Turcotte, An exactly
[7] H. Ohmori, M. Hirano, Magnitude, frequency and geo- soluble hierarchical clustering model: Inverse cascades
morphological signi¢cance of rocky mud £ows, landcreep self-similarity, and scaling, Phys. Rev. E60 (1999) 5293^
and the collapse of steep slopes, Z. Geomorph. Suppl. 67 5300.
(1988) 55^65. [28] F. Guzzetti, M. Cardinali, P. Reichenbach, The in£uence
[8] Y. Sasaki, M. Abe, I. Hirano, Fractals of slope failure of structural setting and lithology on landslide type and
size-number distribution, J. Jpn. Soc. Eng. Geol. 32 pattern, Environ. Eng. Geol. 2 (1996) 531^555.
(1991) 1^11. [29] F. Guzzetti, M. Cardinali, Carta Inventario dei Fenomeni
[9] D.A. Noever, Himalayan sandpiles, Phys. Rev. E47 Franosi della Regione dell'Umbria ed aree limitrofe,
(1993) 724^725. GNDCI Pub., 1989, No 204, 2 sheets, 1:100 000 scale,
[10] T. Sugai, H. Ohmori, M. Hirano, Rock control on mag- (in Italian).
nitude^frequency distributions of landslides, Trans. Jpn. [30] G. Antonini, M. Cardinali, F. Guzzetti, P. Reichenbach,
Geomorph. Union 15 (1994) 233^251. A. Sorrentino, Carta Inventario dei Fenomeni Franosi
[11] Y. Yokoi, J.R. Carr, R.J. Watters, Fractal character of della Regione Marche ed aree limitrofe,. GNDCI Pub.,
landslides, Environ. Eng. Geol. 1 (1995) 75^81. 1993, No 580, 2 sheets, 1:100 000 scale, (in Italian).
[12] J.D. Pelletier, B.D. Malamud, T. Blodgett, D.L. Turcotte, [31] D.J. Varnes, Slope movement types and processes, in:
Scale-invariance of soil moisture variability and its impli- R.L. Schuster, R.J. Krizek (Eds.), Landslide Analysis

EPSL 6085 11-2-02


F. Guzzetti et al. / Earth and Planetary Science Letters 195 (2002) 169^183 183

and Control, Transportation Research Board Special Re- [35] Y. Fujii, Frequency distribution of landslides caused by
port 176, National Academy of Sciences, Washington, heavy rainfall, J. Seismol. Soc. Japan 22 (1969) 244^247.
DC, 1978, pp. 11^33. [36] C.P. Stark, N. Hovius, The characterization of landslide
[32] M. Cardinali, F. Ardizzone, M. Galli, F. Guzzetti, P. size distributions, Geophys. Res. Lett. 28 (2001) 1091^
Reichenbach, Landslides triggered by rapid snow melt- 1094.
ing: the December 1996^January 1997 event in Central [37] V.G. Kossobokov, V.I. Keilis-Borok, D.L. Turcotte, B.D.
Italy, in: P. Claps, F. Siccardi (Eds.), Proceedings 1st Malamud, Implications of a statistical physics approach
Plinius Conference on Mediterranean Storms, Maratea, for earthquake hazard assessment and forecasting, Pure
Italy, 14^16 October 1999, Bios Pub., Cosenza, 2000, pp. Appl. Geophys. 157 (2000) 2323^2349.
439^448. [38] K. Terzaghi, Stability of steep slopes on hard unweath-
[33] E.L. Harp, R.L. Jibson, Inventory of landslides trig- ered rock, Geotechnique 12 (1962) 251^270.
gered by the 1994 Northridge, California earthquake [39] S. Hergarten, H.J. Neugebauer, Self-organized criticality
US Geological Survey Open File Report, 1995, pp. 95^ in a landslide model, Geophys. Res. Lett. 25 (1998) 801^
213. 804.
[34] E.L. Harp, R.L. Jibson, Landslides triggered by the 1994 [40] S. Hergarten, H.J. Neugebauer, Self-organized criticality
Northridge, California earthquake, Seism. Soc. Am. Bull. in two-variable models, Phys. Rev. E 61 (2000) 2382^
86 (1996) S319^S332. 2385.

EPSL 6085 11-2-02

You might also like