You are on page 1of 28

Geomorphology 72 (2005) 272 – 299

www.elsevier.com/locate/geomorph

Probabilistic landslide hazard assessment at the basin scale


Fausto Guzzetti *, Paola Reichenbach, Mauro Cardinali,
Mirco Galli, Francesca Ardizzone
IRPI CNR, via Madonna Alta 126, 06128 Perugia, Italy
Received 8 November 2004; received in revised form 24 May 2005; accepted 16 June 2005
Available online 15 August 2005

Abstract

We propose a probabilistic model to determine landslide hazard at the basin scale. The model predicts where landslides will
occur, how frequently they will occur, and how large they will be. We test the model in the Staffora River basin, in the northern
Apennines, Italy. For the study area, we prepare a multi-temporal inventory map through the interpretation of multiple sets of
aerial photographs taken between 1955 and 1999. We partition the basin into 2243 geo-morpho-hydrological units, and obtain
the probability of spatial occurrence of landslides by discriminant analysis of thematic variables, including morphological,
lithological, structural and land use. For each mapping unit, we obtain the landslide recurrence by dividing the total number of
landslide events inventoried in the unit by the time span of the investigated period. Assuming that landslide recurrence will
remain the same in the future, and adopting a Poisson probability model, we determine the exceedance probability of having
one or more landslides in each mapping unit, for different periods. We obtain the probability of landslide size by analysing the
frequency–area statistics of landslides, obtained from the multi-temporal inventory map. Assuming independence, we obtain a
quantitative estimate of landslide hazard for each mapping unit as the joint probability of landslide size, of landslide temporal
occurrence and of landslide spatial occurrence.
D 2005 Elsevier B.V. All rights reserved.

Keywords: Landslides; Mathematical model; Hazard; Susceptibility; Frequency; Magnitude; GIS; Maps

1. Introduction meteorological changes, such as intense or prolonged


rainfall or snowmelt, and rapid tectonic forcing, such as
Landslides are important natural hazards and an earthquakes or volcanic eruptions. Human disturbances
active process that contributes to erosion and landscape include land use changes, deforestation, excavation,
evolution. Different natural phenomena and human dis- changes in the slope profile, irrigation, etc. On Earth,
turbances trigger landslides. Natural triggers include the volume of mass movements spans 15 orders of
magnitude, and landslide velocity extends over 14
* Corresponding author. Tel.: +39 075 501 4413; fax: +39 075 orders of magnitude, from millimetres per year to hun-
501 4420. dreds of kilometres per hour. Similarly, mass movements
E-mail address: Fausto.Guzzetti@irpi.cnr.it (F. Guzzetti). can occur singularly or in groups of up to several thou-
0169-555X/$ - see front matter D 2005 Elsevier B.V. All rights reserved.
doi:10.1016/j.geomorph.2005.06.002
F. Guzzetti et al. / Geomorphology 72 (2005) 272–299 273

sands. Multiple landslides, for example, occur almost Guzzetti et al., 2002a,b). The latter models attempt to
simultaneously when slopes are shaken by an earth- predict bwhenQ a landslide will occur by establishing
quake, or over a period of hours or days when failures the exceedance probability of landslide occurrence
are triggered by intense rainfall or snow melting. Land- during an established period. Most commonly, the
slides can involve flowing, sliding, toppling or falling exceedance probability is obtained from catalogues
movements, and many landslides exhibit a combination of historical landslide events, which are lists showing
of these types of movements (Cruden and Varnes, the time (or period) of occurrence of single or multiple
1996). The extraordinary breadth of the spectrum of slope failures. No single measure of landslide magni-
landslides makes it difficult to define a single metho- tude exists (Guzzetti, 2002). For certain landslide
dology to ascertain landslide hazard (Guzzetti, 2002). types, including slides and complex failures, landslide
Many methods have been proposed to evaluate area is a reasonable proxy for landslide magnitude.
quantitatively landslide hazard geographically at the Recently, information has become available on the
basin scale (Carrara, 1983; Carrara et al., 1991, 1995; frequency–area statistics of landslides (Hovius et al.,
van Westen, 1994; Soeters and van Westen, 1996; 1997; Stark and Hovius, 2001; Guzzetti et al., 2002a,b;
Chung and Frabbri, 1999; Guzzetti et al., 1999 and Guthrie and Evans, 2004a,b; Malamud et al., 2004).
references herein). Such models are best classified as This information can be used to determine the
susceptibility models (Brabb, 1984), because they pro- expected probability of landslide area and magnitude.
vide an estimate of bwhereQ landslides are expected. A In this paper, we propose a probabilistic model to
few attempts have been made to establish the temporal determine landslide hazards at the basin scale. The
frequency of slope failures (Keaton et al., 1988; Lips model exploits information obtained from a multi-
and Wieczorek, 1990; Coe et al., 2000; Crovelli, 2000; temporal inventory map to predict where landslides

RIVANAZZANO
N
0 1 2 4 km

VARZI

Lombardy

Piedmont MILAN

TURIN Emilia-Romagna
BOLOGNA

M. CHIAPPA
Fig. 1. Location of the study area, the Staffora River basin.
274 F. Guzzetti et al. / Geomorphology 72 (2005) 272–299

will occur, how frequently they will occur, and how ciated to the formation of the Apennines mountain
large they will be. We test this model in the Staffora chain. A compressive phase of Cretaceous to Eocene
River basin, in the Northern Apennines, and we dis- age produced large, north–east verging thrusts with
cuss the results obtained. associated anticlines, synclines and transcurrent faults,
and was followed by an extensional tectonic phase of
Oligocene to Holocene age, which produced chiefly
2. The study area normal faults. The lithological and the structural set-
tings control the morphology of the area, which fea-
The study area extends for 275 km2 in the southern tures steep and asymmetric slopes, dissected by a
Lombardy region, in northern Italy (Fig. 1). Elevation dense, locally actively eroding stream network. Land-
in the area ranges from about 150 m at Rivanazzano to slides are abundant in the area, and range in type and
1699 m at M. Chiappa. The Staffora River, a tributary size from large rotational and translational slides to
of the Po River, drains the area. In the 42-year period deep and shallow flows. Some of the landslides are
from 1951 to 1991, annual rainfall in the area ranged presumably very old in age. Very old landslides are
from 410 to 1357 mm, with an average value of 802 mostly relict or dormant, and are partially concealed
mm. Precipitation is most abundant in the autumn and by forest and the intensive farming activity.
in the spring (Fig. 2).
Marine, transitional and continental sedimentary
rocks, Cretaceous to Holocene in age (Servizio Geo- 3. Mathematical model
logico Nazionale, 1971) outcrop in the Staffora River
basin. Marine sediments include: (i) sequences of Varnes and the IAEG Commission on Landslides
layered limestone, marly-limestone, marl and clay, and other Mass-Movements (1984) proposed that the
with ophiolites, (ii) disorganized, and highly fractured definition adopted by the United Nations Disaster
marl and clay, overlaid by massive sandstones, and Relief Organization (UNDRO) for all natural hazards
(iii) shallow marine sediments pertaining to the Ges- be applied to the hazard posed by mass movements.
soso-Solfifera Fm. Transitional deposits feature con- Varnes and his co-workers defined the landslide
glomerates, with lenses of marl and sand, which are hazard as bthe probability of occurrence within a
Oligocene in age. Fluvial and terraced deposits, Holo- specified period of time and within a given area of a
cene in age, represent the continental deposits and potentially damaging phenomenonQ. Guzzetti et al.
outcrop along the main valley bottoms. (1999) amended the definition to include the magni-
The area has a complex structural setting resulting tude of the event. In contrast to other natural hazards,
from the superposition of two tectonic phases asso- no unique measure of landslide magnitude is available
(Hungr, 1997). For earthquakes, magnitude is a mea-
sure of the energy released during an event. For land-
Percentage of mean annual rainfall (%)

14
100 slides, a measure of the energy released during failure
Mean monthly rainfall (mm)

12 is difficult to obtain. Hungr (1997) proposed destruc-


80 tiveness to be a measure of landslide magnitude.
10
Cardinali et al. (2002b) and Reichenbach et al.
8 (2005) defined landslide destructiveness as a function
60
of the landslide volume and of the expected landslide
6
velocity, the latter obtained from the landslide type.
40
4 For large areas, landslide volume and velocity are
difficult to evaluate systematically, making the
20 2 approach impracticable. Alternatively, where slope
Jan

Feb

Mar
Apr
May
Jun

Jul
Aug
Sep

Oct
Nov
Dec

failures are chiefly slow-moving slides and slide


Fig. 2. Monthly rainfall values (left y-axis) and percentages (right y-
earth-flows, destructiveness can be related to the
axis) for the period between 1951 and 1990 for the Varzi rain gauge, area of the landslide, information that is readily avail-
416 m a.s.l. able from accurate landslide inventory maps.
F. Guzzetti et al. / Geomorphology 72 (2005) 272–299 275

The definition of landslide hazard incorporates the Guzzetti et al., 2002a,b; Guthrie and Evans, 2004a,b;
concepts of location, time and size. To complete a Malamud et al., 2004).
hazard assessment one has to predict (quantitatively) Malamud et al. (2004) analysed three recent and
bwhereQ a landslide will occur, bwhenQ or how fre- well documented landslide inventories from California
quently it will occur, and bhow largeQ the landslide (Harp and Jibson, 1996), central Italy (Cardinali et al.,
will be. In mathematical terms, this can be written as: 2000; Guzzetti et al., 2002a,b) and Guatemala (Buck-
nam et al., 2001), and found the probability density
HL ¼ P½AL zaL in a time interval t; given
function (PDF) of landslide area, A L, to be in good
fmorphology; lithology; structure; land use; . . . g agreement with a truncated inverse gamma distribu-
ð1Þ tion. These authors proposed that the PDF of landslide
area can be estimated as (Malamud et al., 2004):
where, A L is the area of a landslide greater or equal
 qþ1  
than a minimum size, a L, measured, for example, in 1 a a
square meters. For any given area, Eq. (1) expresses pðAL ; q; a; sÞ ¼ exp
aCðqÞ AL s AL s
landslide hazard as the conditional probability of land-
ð4Þ
slide size, PA L, of landslide occurrence in an estab-
lished period t, P N, and of landslide spatial where: C(q) is the gamma function of q, and q N 0,
occurrence, S, given the local environmental setting. a N 0, and s V A L b l are parameters of the distribution.
Assuming independence among the three probabil- In Eq. (4), q controls the power–law decay for medium
ities, the landslide hazard, i.e. the joint probability is: and large landslide areas, a primarily controls the
HL ¼ PAL  PN  S ð2Þ location of the maximum of the probability distribu-
tion, and s primarily controls the exponential decay for
From a geomorphological point of view, the small landslide areas. Fitting Eq. (4) to the three avail-
assumption that the three probabilities (i.e., the three able inventories, Malamud et al. (2004) found
components of landslide hazard) are independent is q = 1.40, a = 1.28d 103 m2, and s = 1.32d 102 m2
strong and may not hold always and everywhere. In (determination coefficient r 2 = 0.965).
many areas we expect slope failures to be more fre- Using Eq. (4), PA Lis given by:
quent (time component) where landslides are more Z l Z l  qþ1
abundant and landslide area is large (spatial compo- 1 a
PAL ¼ pðAL ; q; a; sÞdAL ¼
nent). However, given the lack of understanding of the aL aL aCðqÞ AL s
 
landslide phenomena, independence is an acceptable a
first-approximation that makes the problem mathema-  exp dAL ð5Þ
AL s
tically tractable and easier to work with.
In another study of frequency–area statistics of
3.1. Probability of landslide size landslides, Stark and Hovius (2001) analyzed three
landslide datasets obtained from New Zealand and
The probability that a landslide will have an area Taiwan and found the PDF of landslide area to be in
greater or equal than a L is: good agreement with a double Pareto probability dis-
tribution. Using this distribution, PA Lis given by (Stark
PAL ¼ P½AL zaL  ð3Þ and Hovius, 2001):
Z l
and can be estimated from the analysis of the fre- PAL ¼ pðAL ; a; b; l; m; cÞdAL
quency–area distribution of known landslides, aL
obtained from landslide inventory maps. Analysis of Z l " #
b=a
b ½1 þ ðm=l Þ a 
accurate landslide inventories reveals that the abun- ¼ a 1þðb=aÞ
dance of landslides increases with landslide area up to aL l ð1 dÞ ½1 þ ðAL =l Þ 

a maximum value, where landslides are most frequent,  ðAL =l Þ ðaþ1Þ dAL ð6Þ
then it decays rapidly along a power law (Pelletier et
al., 1997; Hovius et al., 1997; Stark and Hovius, 2001; where: a N 0, b N 0, 0 V c V l V m V l, and with
276 F. Guzzetti et al. / Geomorphology 72 (2005) 272–299

h ib=a
1þðm=l Þ a
d ¼ yð cÞ ¼ 1þðAL =l Þ a
. Note that a in Eq. (6) is the events, or from a multi-temporal landslide inventory
same as q in Eqs. (4) and (5) and controls the power- map.
law decay of landslide probability for large landslide From Eq. (8), the probability of experiencing one
areas. Also, b in Eq. (6) controls the power-law or more landslides during time t (i.e. the exceedance
decays for small landslide areas. probability) is:
Use of Eqs. (4)–(6) requires the catalogue of land-
slide areas from which the distributions are derived to P½ N ðt Þz1 ¼ 1 P½ N ðt Þ ¼ 0 ¼ 1 expð kt Þ
be statistically substantially complete, i.e. that most of ¼ 1 expð t=lÞ ð9Þ
the landslides that have occurred in the region were
mapped accurately (Malamud et al., 2004). Discussing Eq. (9), Crovelli (2000) notes that for a
given period of time t, if l Y l, then P[N(t) z1] Y 0,
3.2. Temporal probability of landslides i.e. if the estimated mean recurrence interval between
successive events is very large, chances are that no
As a first approximation, landslides can be consid- slope failures will be experienced in the considered
ered as independent random point-events in time period. Also, if the estimated mean recurrence l is
(Crovelli, 2000). In this framework, the exceedance fixed, and the time interval is very long (t Y l), then
probability of occurrence of landslide events during P[N(t) z 1] Y 1, and one is certain to observe a land-
time t is: slide event.
The Poisson model allows for determining the
PN ¼ P½ N ðt Þz1 ð7Þ probability of future landslides for different times t
(i.e. for different numbers of years) based on the
where N(t) is the number of landslides that occur statistics of past landslide events, under the following
during time t in the investigated area. assumptions (Crovelli, 2000): (i) the number of land-
Two probability models are commonly used to slide events that occur in disjoint time intervals are
investigate the occurrence of naturally occurring ran- independent; (ii) the probability of an event occurring
dom point-events in time: the Poisson model and the in a very short time is proportional to the length of
binomial model (Crovelli, 2000; Önöz and Bayazit, the time interval; (iii) the probability of more than
2001). The Poisson model is a continuous-time model one event in a short time interval is negligible; (iv)
consisting of random-point events that occur indepen- the probability distribution of the number of events is
dently in ordinary time, which is considered naturally the same for all time intervals of fixed length; and (v)
continuous. The Poisson model has been used to the mean recurrence of events will remain the same
investigate the temporal occurrence of, for example, in the future as it was observed in the past. The
volcanic eruptions (Klein, 1982; Connor and Hill, consequences of these assumptions, which may not
1995; Nathenson, 2001), floods (e.g. Yevjevich, always hold for landslide events, should be consid-
1972; Önöz and Bayazit, 2001), and landslides (e.g. ered when interpreting (and using) the results of the
Crovelli, 2000; Coe et al., 2000). Adopting a Poisson probability model.
model for the temporal occurrence of landslides, the As an alternative to the Poisson model, a binomial
probability of experiencing n landslides during time t model can be adopted. The binomial probability
is given by (Crovelli, 2000): model is a discrete-time model consisting of the
ðkt Þn occurrence of random-point events in time. In this
P½ N ðt Þ ¼ n ¼ expð kt Þ n ¼ 0; 1; 2; . . . model, time is divided into discrete increments of
n!
equal length. Within each time increment a single
ð8Þ
point-event may or may not occur. The binomial
where k is the estimated average rate of occurrence of model was adopted by Costa and Baker (1981) to
landslides, which corresponds to 1 / l, with l the investigate the occurrence of floods, and by Keaton
estimated mean recurrence interval between succes- et al. (1988), Lips and Wieczorek (1990), and Coe et
sive failure events. The variables k and l can be al. (2000) to study the temporal occurrence of land-
obtained from a historical catalogue of landslide slides and debris flows.
F. Guzzetti et al. / Geomorphology 72 (2005) 272–299 277

Following Crovelli (2000), and adopting the bino- Susceptibility can be estimated using a variety of
mial probability model, the exceedance probability of statistical techniques, which include among others
experiencing one or more landslides during time t is: discriminant analysis (Reger, 1979; Carrara, 1983;
Carrara et al., 1991, 1992, 1995, 2003; Guzzetti et
P½ N ðt Þz1 ¼ 1 P½ N ðt Þ ¼ 0 ¼ 1 ð1 pÞt al., 1999; Nagarajan et al., 2000; Baeza and Coro-
¼ 1 ð1 1=lÞt ð10Þ minas, 2001; Ardizzone et al., 2002; Cardinali et al.,
2002a; Santacana et al., 2003), logistic regression
where, p is the estimated probability of a landslide
analysis (Carrara et al., 1992; Atckinson and Mas-
event in time t, and l = 1 / p is the estimated mean
sari, 1998; Rowbotham and Dudycha, 1998; Dai and
recurrence interval between successive slope failures.
Lee, 2002, 2003; Olhmacher and Davis, 2003; Lee,
As for the Poisson model, l can be obtained from a
2004; Süzen and Doyuran, 2004; Ayalew and Yama-
historical catalogue of landslides or from a multi-
gishi, 2005), and conditional analysis based on a
temporal landslide inventory map. The binomial prob-
variety of favourability functions like weight of evi-
ability model holds under the same or similar assump-
dence (Bonham-Carter, 1991; Lee et al., 2002a,b;
tions listed for the Poisson model.
Wu et al., 2004), weighting factors (Çevik and
Crovelli (2000) compared the Poisson and the
Topal, 2003), weighted linear combination of
binomial probability models, and showed that the
instability factors (Ayalew et al., 2004), likelihood
two models differ for short mean recurrence intervals
ratio (Chung and Fabbri, 2003, 2005; Fabbri et al.,
(i.e. when l is small) and for short periods (i.e. when t
2003; Lee, 2004), certainty factors (Binaghi et al.,
is small), with the binomial model over-estimating the
1998), information value (Lin and Tung, 2004), and
exceedance probability of future landslide events. For
modified Bayesian estimation (Chung and Frabbri,
long periods and large mean recurrence intervals, i.e.,
1999).
for rare events, the bimodal model coincides with the
Depending on the type of statistical technique, the
Poisson model.
meaning of the probability changes slightly. When
using discriminant analysis or logistic regression
3.3. Spatial probability of landslides
analysis, the probability assigned to any given area
(i.e. to each terrain or mapping unit) is the prob-
The spatial probability of landslide occurrence,
ability that the area pertains to one of two groups,
also known as susceptibility (Brabb, 1984), is the
namely: (i) the group of mapping units having land-
probability that any given region will be affected by
slides, P 1, or (ii) the group of mapping units free of
landslides, given a set of environmental conditions.
landslides, P 0, given the set of environmental con-
Defining:
ditions used in the analysis. At the beginning of a
L : a given region will be affected by landslides study only past landslides in a region are known.
Hence, classification of landslide-bearing and land-
ð11Þ
slide-free mapping units is made based on the known
susceptibility, S, becomes: distribution of past slope failures. A straightforward
deduction is to assume S = P[r a P 1] = 1 P[r a P 0].
S ¼ P½ L is true; given
In other words, if a region r pertains to the group of
fmorphology; lithology; structure; landuse; etc:g mapping units having known (i.e. past) landslides
ð12Þ (e.g. P 1) because of the local environmental condi-
tions, it is likely that the same region will experience
or, slope failures again in the future (even if we do not
S ¼ P½ Ljv1 ðrÞ; v2 ðrÞ; . . . ; vm ðrÞ ð13Þ know when). Equally, if a region pertains to the
group of mapping units free of (known) landslides
which is the joint conditional probability that a region (e.g. P 0) it is unlikely that the same region will
r will be affected by future landslides given the m experience mass movements.
environmental variables v 1, v 2, . . ., v m in the same Chung and Frabbri (1999) proposed to estimate
region. the probability of future landslides in any given
278 F. Guzzetti et al. / Geomorphology 72 (2005) 272–299

region, S, from the probability of past landslides in ena (Chung and Frabbri, 1999). This is an improve-
the same region, given a set of environmental vari- ment over other modelling approaches, particularly
ables. Letting: where information on past landslides is limited or
imprecise.
F : a given region has been affected by landslides; Quantitative susceptibility models can predict the
ð14Þ spatial occurrence of future landslides under the gen-
eral assumption that in any given area slope failures
the joint conditional probability of past landslides in will occur in the future under the same circumstances
a region r, given the m environmental variables v 1, and because of the same conditions that caused them
v 2, . . ., v m in the same region is: in the past. This is a geomorphological rephrase of
bthe past is the key to the futureQ, which is a direct
D ¼ P½ Fjv1 ðrÞ; v2 ðrÞ; . . . ; vm ðrÞ ð15Þ consequence of the well-known principle of unifor-
mitarianism. However, the principle may not hold for
From Eqs. (13) and (15) it follows that: landslides. New, first-time failures occur under con-
ditions of peak resistance (friction and cohesion),
P½ Ljv1 ðrÞ; v2 ðrÞ; . . . ; vm ðrÞ ¼ P½ Fjv1 ðrÞ; whereas reactivations occur under intermediate or
v2 ðrÞ; . . . ; vm ðrÞ; ð16Þ residual conditions. It is well known that terrain
gradient is an important factor for the occurrence of
or S = D. To estimate the spatial probability of past landslides. An obvious effect of a slope failure is to
landslides (and infer from it the spatial probability change the morphology of the terrain where the fail-
of future failures), the study area is partitioned into ure occurs. In addition, when a landslide moves it
unique condition units (UCU), i.e. areas character- may change the hydrological conditions of the slope.
ized by an exclusive (unique) combination of envir- It is also well known that landslides can change their
onmental conditions. This can be easily obtained in type of movement and velocity with time. Lastly,
a GIS by the geographical union of all thematic landslide occurrence and abundance are a function
layers available for the study area. The spatial of environmental conditions that vary with time at
probability of past landslide occurrence is then sim- different rates. Some of the environmental variables
ply estimated from the density of landslides in each are affected by human actions (e.g. land use, defor-
UCU. estation, irrigation, etc.), which are also highly
Chung and Frabbri (1999) showed that D is a changeable. Because of these complications, each
good descriptor of the past (known) landslides, landslide occurs in a distinct environmental context,
conditioned to the available environmental informa- which may have been different in the past and that
tion, but it may not be a good estimator of the might be different in the future. Despite these limita-
future spatial occurrence of landslides, S. They tions, in this work we assume that the principle of
proposed more efficient indexes to estimate the uniformity hold bstatisticallyQ, i.e. that in the investi-
future occurrence of landslides from the observed gated region future landslides will occur on average
past distribution of slope failures, including: (i) a under the same circumstances and because of the
Bayesian estimation under the condition of indepen- same conditions that triggered them in the past. We
dence, (ii) a regression model based on bivariate further assume that our knowledge of the distribution
conditional probabilities, (iii) a modified Bayesian of past failures is reasonably accurate and complete.
estimator under the condition of independence, and We accept these simplifications to make the problem
(iv) a modified regression model based on bivariate tractable.
conditional probabilities. The two modified models
incorporate expert knowledge, i.e. information
which is not included in the original landslide or 4. Landslide hazard assessment
thematic data, and that is used to modify the
observed frequency of occurrence of landslides to To ascertain landslide hazard we partitioned the
fit the expert’s understanding of landslide phenom- Staffora basin into mapping units (Guzzetti et al.,
F. Guzzetti et al. / Geomorphology 72 (2005) 272–299 279

1999), adopting the procedure proposed by Carrara A team of three geomorphologists carried out inter-
et al. (1991). Starting from a digital terrain model pretation of the aerial photographs in a period of 6
with a ground resolution of 20 m  20 m and a months. Two team members looked at each pair of
simplified representation of the main drainage lines, aerial photographs using a mirror stereoscope (with a
specialized software (Carrara et al., 1991; Detti and magnification of 4) that allowed both interpreters to
Pasqui, 1995) was used to partition the territory map contemporaneously on the same stereo pair. The
into individual slope units, bounded by drainage third photo-interpreter independently reviewed, and
and divide lines. For each slope unit, the software where necessary corrected the interpretations of the
computed 21 morphometric and hydrological para- other two, using a high magnification (up to 20)
meters useful in explaining the spatial distribution stereoscope. For the interpretation, we used all geolo-
of landslides (Carrara et al., 1991, 1995). We gical and geomorphological information available to
further subdivided the slope units according to the us from published maps, previous works carried out in
main rock types cropping out in the basin. In this the same area, and discussion with other geologists
way, we partitioned the study area into 2243 map- (Rossetti, 1997; Antonini et al., 2000; Ardizzone et
ping units based on lithology, morphology and al., 2002).
hydrology (geo-morpho-hydrological terrain units), Landslide information collected through the inter-
which represent the mapping units used for the pretation of aerial photographs was visually trans-
hazard assessment. ferred to topographic maps at 1 : 10,000 scale. We
transferred geomorphological information from the
4.1. Landslide identification and mapping base maps onto stable, transparent sheets and scanned
these to obtain black and white, raster images of each
We obtained information on the geographical and map sheet. We used a scanning resolution of 300–400
temporal distribution of landslides and on landslide dpi, which corresponded to a ground resolution of less
size from a detailed multi-temporal inventory map, than 0.1 m. The raster representation of the geomor-
prepared through the interpretation of multiple sets phological line images was changed in a GIS into
of aerial photographs. For the study area, five sets of vector format using a semi-automatic procedure,
aerial photographs were available to us (Table 1). We which allowed us to assign attributes to each line
used each set of aerial photographs separately to segment. Polygons were then constructed and labelled
obtain different landslide inventory maps (Fig. 3). with the appropriate codes, depending on their land-
We then merged in a GIS the five individual land- slide properties.
slide maps to obtain the multi-temporal inventory In the separate inventory maps (Fig. 3A–E), we
map (Fig. 3F). classified landslides according to the type of move-
ment, and the estimated age, activity, depth, and
velocity. We defined landslide type according to
Cruden and Varnes (1996), and the WP/WLI
Table 1
Staffora River basin (1990). For deep-seated slope failures, we mapped
Code Flight Date Type Nominal scale
separately the landslide crown (depletion area) from
the deposit. Landslide age, activity, depth, and
A GAI-IGMI 18 July Black and 1 : 33,000
1955 white
velocity were determined based on the type of
B ALI FOTO Winter Black and 1 : 15,000 movement, the morphological characteristics and
1975 white appearance of the landslide on the aerial photo-
C TEM 1 Summer Colour 1 : 22,000 graphs, the local lithological and structural setting,
1980 and the date of the aerial photographs. We categor-
D Lombardy Summer Black and 1 : 25,000
1994 white
ized landslide age as recent, old or very old,
E IT 2000 22 June Colour 1 : 40,000 despite ambiguity in the definition of the age of a
1999 mass movement based on its appearance (Wiec-
Aerial photographs used to compile the multi-temporal landslide zorek, 1984). Landslides were classified active
inventory map. (WP/WLI, 1993) where they appeared fresh on
280 F. Guzzetti et al. / Geomorphology 72 (2005) 272–299

Fig. 3. Landslide inventory maps for the Staffora River basin. Capital letters indicate the year of the aerial photographs used to compile the
inventory. See Table 2 for reference. Map F shows the entire multi-temporal inventory, which includes relict and old slides (shown in grey)
identified in the 1955 aerial photographs not shown in the other maps.
F. Guzzetti et al. / Geomorphology 72 (2005) 272–299 281

the aerial photographs (of a given date). Mass move- slides. Table 3 lists the landslide types identified in
ments were classified as deep-seated or shallow, the individual maps and shown in the multi-temporal
depending on the type of movement and the estimated inventory map.
landslide volume. The latter was based on the type of
failure, and the morphology and geometry of the 4.2. Probability of landslide size
detachment area and the deposition zone. Landslide
velocity (WP/WLI, 1995; Cruden and Varnes, 1996) To ascertain the probability of landslide area, we
was considered a proxy of landslide type, and classi- selected the multi-temporal inventory map covering
fied accordingly. We acknowledge that the adopted the 45-year period from 1955 to 1999 (2390 land-
classification scheme suffers from simplifications and slides, A2–E2 in Table 2). We obtained the area of
required geomorphological deduction, but it fits the each landslide from the GIS. Care was taken to
available information on landslide types and process in calculate the exact size of each landslide, avoiding
the northern Apennines (Rossetti, 1997). topological and graphical problems related to the
Table 2 shows the number, total extent and area presence of smaller landslides inside larger mass
statistics of the landslides identified in the five sets movements. For convenience, we merged the
of aerial photographs. We identified the largest num- crown area and the deposit, and we used the total
ber of failures and the largest landslide area in the landslide area in the analysis. We used only recent
1955 photographs, which show landslide of much and active landslides, and we excluded the old and
older age. In the other flights, we identified only new relict mass movements.
and recent landslides. The entire landslide inventory Fig. 4A shows the probability density function
shows 3922 landslides, including 89 very old, relict (PDF) of landslide areas in the Staffora basin. Two
mass movements. The multi-temporal map covering estimates of the PDF are shown. We obtained the first
an undefined period from pre-1955 to 1999 (A1–E2 estimate using the truncated inverse-gamma function
in Table 2) shows 3833 landslides, and does not of Malamud et al. (2004) (Eq. (4)). For this estimate,
include the relict landslides. The multi-temporal 97.5% confidence intervals are also shown. We
inventory map covering the 45-year period from obtained the second estimate using the double-Pareto
1955 to 1999 (A2–E2 in Table 2) shows 2390 land- function of Stark and Hovius (2001) (Eq. (6)). In the

Table 2
Staffora River basin
Inventory Estimated landslide age Landslide Landslide area
Number # Density #/km2 Total km2 Percentage* % Min m2 Mean m2 Max m2
A0 Very old (relict) 89 0.32 34.72 49.30 57,300 390,100 2,384,900
A1 Older than 1955 1443 5.27 38.24 54.30 900 27,900 826,700
A2 1955 active 306 1.12 2.46 3.49 700 8000 163,400
B1 1955–1975 318 1.16 2.38 3.39 200 7500 51,000
B2 1975 active 685 2.50 4.41 6.26 100 6500 114,700
C1 1975–1980 89 0.32 1.32 1.87 400 14,800 119,100
C2 1980 active 305 1.11 2.40 3.41 500 7900 119,100
D1 1980–1994 455 1.66 2.06 2.92 500 4500 177,800
D2 1994 active 175 0.63 1.36 1.94 500 7800 77,900
E1 1994–1999 19 0.07 0.65 0.93 3600 34,300 119,100
E2 1999 active 38 0.14 0.85 1.21 1900 22,400 119,100
A0–A1 Very old to older than 1955 1532 5.57 63.22 90 900 41,300 2,384,900
A0–E2 Very old to 1999 active 3922 14.26 70.42 100 100 17,900 2,384,900
A1–E2 Older than 1955 to 1999 active 3833 13.93 46.43 66 100 12,100 177,800
A2–E2 1955 active to 1999 active 2390 8.69 12.08 17 100 3600 177,800
Landslide size and abundance. For dates of aerial photographs see Table 1.
* The percentage of landslide area is computed with respect to the total area covered by landslides (A0–E2).
282 F. Guzzetti et al. / Geomorphology 72 (2005) 272–299

Table 3
Staffora River basin
Inventory Estimated landslide age Landslide types
I II III IV V VI VII ALL
A0 Very old (relict) 0 0 0 0 0 41 48 89
0% 0% 0% 0% 0% 8.3% 6.0% 2.2%
A1 Older than 1955 121 7 15 253 55 357 635 1443
10.5% 3.6% 62.5% 25.7% 20.0% 72.8% 78.9% 36.8%
A2 1955 active 149 26 2 86 26 7 10 306
13.0% 13.3% 8.3% 8.7% 9.6% 1.4% 1.2% 7.8%
B1 1955–1975 60 4 0 93 52 30 79 318
5.2% 2.1% 0.0% 9.5% 18.9% 6.2% 9.8% 8.1%
B2 1975 active 291 32 3 254 60 29 16 685
25.3% 16.4% 12.5% 25.8% 21.8% 5.9% 2.0% 17.5%
C1 1975–1980 23 15 0 26 20 4 1 89
2.0% 7.7% 0.0% 2.6% 7.3% 0.8% 0.1% 2.3%
C2 1980 active 71 33 0 132 48 12 9 305
6.2% 16.9% 0.0% 13.4% 17.3% 2.4% 1.1% 7.8%
D1 1980–1994 293 29 0 121 5 5 2 455
25.5% 14.9% 0.0% 12.4% 1.8% 1.1% 0.2% 11.6%
D2 1994 active 132 23 4 5 2 5 4 175
11.5% 11.8% 16.7% 0.5% 0.7% 1.1% 0.6% 4.5%
E1 1994–1999 2 10 0 5 1 0 1 19
0.2% 5.1% 0.0% 0.5% 0.4% 0.0% 0.1% 0.5%
E2 1999 active 7 16 0 9 6 0 0 38
0.6% 8.2% 0.0% 0.9% 2.2% 0.0% 0.0% 1.0%
A0–A1 Very old to 121 7 15 253 55 398 683 1532
older than 1955 10.5% 3.6% 62.5% 25.7% 20% 81.2% 84.8% 39.0%
A0–E2 Very old to 1149 195 24 984 275 490 805 3922
1999 active 100% 100% 100% 100% 100% 100% 100% 100%
A1–E2 Older than 1955 1149 195 24 984 275 449 757 3833
to 1999 active 100% 100% 100% 100% 100% 91.6% 94.0 97.7%
A2–E2 1955 active to 1028 188 9 731 220 92 122 2390
1999 active 89.5% 96.4% 37.5% 74.3% 80% 18.8% 15.2% 61.0%
Number (normal text) and percentage (italic) of landslide types.
(I) single flows; (II) multiple small flows; (III) deep-seated flows; (IV) shallow slides; (V) shallow slide-earth flows; (VI) deep-seated slide-earth
flows; (VII) deep-seated slides; (ALL) all landslide types.

study area the two functions provide very similar 4.3. Frequency of occurrence of landslides
results, and differ slightly in the slope of the tail of
the distribution (for inverse gamma, q + 1 = 2.77, std. The proposed model for landslide hazard requires
dev. = 0.08, for double Pareto, a + 1 = 2.50, std. an estimate of the temporal probability of slope
dev. = 0.05). Fig. 4B shows the probability of landslide failures. The availability of a multi-temporal land-
size, i.e. the probability that a landslide will have an slide inventory map (Fig. 3) allowed us to estimate
area smaller than a given size (left axis), or the prob- the frequency of landslide occurrence in each map-
ability that a landslide will have an area that exceeds a ping unit. To obtain an estimate of the frequency of
given size (right axis). Fig. 4B also shows the prob- landslide occurrence, we first counted the number of
ability that a landslide in the Staffora River basin landslides in each mapping unit. Considering only
exceeds an area of 2000 m2 and an area of 10,000 the new or reactivated landslides (A2–E2 in Table 2),
m2 (1 ha), which are found 0.75 and 0.15, respec- we prepared a map of the total number of landslide
tively. We will use these values to ascertain the land- events (occurrences) in the 45-year period between
slide hazard. 1955 and 1999, the dates of the oldest and the most
F. Guzzetti et al. / Geomorphology 72 (2005) 272–299 283

102 103 104 105 106


10-3 10-3
A
-4
10 10-4
Probability density, p (m-2)

10-5 10-5

10-6 10-6

10-7 10-7

Inverse Gamma
10-8 10-8
Double Pareto

10-9 10-9
1100 1,000 10,000 100,000 1,000,000

Landslide Area, AL (m2)

102 103 104 105 106


1.0 0.0

0.9 Inverse Gamma


B 0.1

Double Pareto 10,000 m2


0.8 0.2

0.7 0.3
Probability, P[AL < aL ]

Probability, P[AL ≥ aL]


0.6 0.4

0.5 0.5

0.4 0.6

0.3 0.7
2
0.2 2000 m 0.8

0.1 0.9

0.0 1.0
100 1,000 10,000 100,000 1,000,000
Landslide Area, AL (m2)

Fig. 4. Probability density (A) and probability (B) of landslide area in the Staffora River basin. Solid black line is a truncated inverse gamma
function (Malamud et al., 2004). Dotted line is a double Pareto function (Stark and Hovius, 2001). Error bars show 97.5% confidence
intervals.

recent aerial photographs. For each mapping unit, unit (from 1955 to 1999), assuming the rate of
based on the past rate of landslide occurrence we slope failures will remain the same for the future,
obtained landslide recurrence, i.e. the expected time and adopting a Poisson probability model (Eq. (9)),
between successive failures. Knowing the mean we computed the exceedance probability of having
recurrence interval of landslides in each mapping one or more landslides in each mapping unit. Fig. 5
284 F. Guzzetti et al. / Geomorphology 72 (2005) 272–299

Fig. 5. Exceedance probability of landslide occurrence obtained computing the mean recurrence interval of past landslide events from the multi-
temporal inventory (Fig. 3), assuming it will remain the same for the future, and adopting a Poisson probability model (Eq. (9)). Shades of grey
show exceedance probability for different periods: A) 5 years, B) 10 years, C) 25 years, D) 50 years. Square bracket indicates class limit is
included; round bracket indicates class limit is not included.

(A–D) shows the exceedance probability for different time. Based on the historical record of landslides
periods, from 5 to 50 years. Similar maps can be obtained from the multi-temporal inventory map,
prepared for any period. As expected, the probability after 50 years most of the slopes in the basin have a
of having one or more slope failures increases with medium to high probability of experiencing mass
F. Guzzetti et al. / Geomorphology 72 (2005) 272–299 285

movements. We will use the obtained estimates to


ascertain landslide hazard.
For the Staffora River basin, an historical catalogue
of damaging landslides is available (Regione Lombar-
A dia, 2002). The catalogue was compiled through the
systematic screening of newspapers and local
archives, and covers the period from 1850 to 1997,
listing 389 slope failures at 243 sites (Fig. 6A). A
comparison between the expected time between fail-
ures obtained from the multi-temporal inventory map,
and the information listed in the historical catalogue is
not straightforward. Most of the information listed in
the historical catalogue refers to landslides that caused
damage in urban areas or along the road network.
With this respect, the historical catalogue is less sys-
tematic than the mapping obtained through the inter-
pretation of aerial photographs, which provides a
more comprehensive view of the occurrence of land-
slide events in the basin. The historical catalogue
N covers a longer period of time (148 years) than the
0 1 2 4 km multi-temporal inventory (45 years). However, com-
pleteness of the historical catalogue varies with time.
Lack of historical landslide events before 1950 (Fig.
6B) is due to incompleteness in the catalogue, and to
Cumulative number of landslides

35 400
B the lesser number of vulnerable elements (e.g. roads,
30 350
houses) in the study area in the early period of the
Number of landslides

300 catalogue.
25
250 We attempted a quantitative comparison between
20
200 the two sources of information on past landslides. For
15 each mapping unit, we selected the events listed in the
150
10
historical catalogue that occurred during or after 1950,
100
and we computed the mean recurrence of failures. We
5 50 then compared the result with the mean recurrence of
0 0 landslides obtained from the multi-temporal inventory
1850

1870

1910
1920
1930
1940
1950
1960
1860

1880
1890
1900

1970
1980
1990

map (Fig. 6C). On average, mean recurrence is larger


for the multi-temporal inventory, indicating a more
0.7 systematic mapping. Inspection of Fig. 6C reveals that
C
Recurrence from Catalogue

0.6 a few mapping units exhibit greater recurrence for the


historical catalogue than for the multi-temporal inven-
0.5

0.4

0.3 Fig. 6. Historical landslide events in the Staffora River basin in the
period between 1850 and 1997. A) Map showing the location of 243
0.2 sites affected by 389 damaging slope failures in the considered
0.1 period. B) Annual number (histogram) and cumulative number
(line) of landslide events in the considered period. C) Comparison
0.0 between landslide recurrence obtained from the multi-temporal
0.0 0 .1 0.2 0 .3 0.4 0 .5 0.6 0.7 inventory (x-axis) and the historical catalogue of landslide events
Recurrence from Inventory ( y-axis).
286 F. Guzzetti et al. / Geomorphology 72 (2005) 272–299

tory map. Most of the slope failures listed in the (Fig. 7A) using only landslides identified in the 1955
historical catalogue and located in these mapping photographs (Fig. 3A). These landslides included
units (60%) occurred from 1956 to 1974, a period recent (in 1955) and old landslides that were visible
for which aerial photographs were not available. Spe- in the aerial photographs used to compile the inven-
cific triggers (e.g., in 1959, 1960 and 1985) produced tory, but not the very old and relict landslides present
many historical damaging landslides. For these trig- in the study area (A1–A2 in Table 2). We then added to
gers, aerial photographs were also not available. In the inventory map the new landslides identified in the
addition, damaging slope failures in these mapping 1975 aerial photographs (Fig. 3B) and we obtained a
units were small or very small, making their recogni- new estimate of the probability of spatial landslide
tion on the aerial photographs very difficult. occurrence (Fig. 7B). We repeated the same procedure
adding the slope failures that we identified and
4.4. Landslide susceptibility mapped using the 1980, 1994, and 1999 aerial photo-
graphs (Fig. 3C–E). Results are shown in Fig. 7C, D
The model for landslide hazard involves a quanti- and E, respectively. At each step, we obtained a
tative estimate of the probability of spatial landslide different susceptibility map, i.e. a different estimate
occurrence, i.e. of susceptibility. We obtained land- of the probability of spatial landslide occurrence.
slide susceptibility through discriminant analysis of Table 4 lists the variables entered into the five dis-
46 thematic variables, including morphology (24 vari- criminant models. In Table 4, the standard discrimi-
ables derived from a 20 m  20 m DTM), lithology nant function coefficients (SDFC) show the relative
(14 variables), structure (3 variables) and land use (5 importance of each variable in the discriminant func-
variables). Using GIS technology, we computed the tion as a predictor of slope instability. Variables with
percentage of the individual thematic variables in each large coefficients (in absolute value) are strongly
mapping unit. The obtained values were the indepen- associated with the presence/absence of landslides.
dent variables in the statistical analysis. We then The sign of the coefficient tells if the variable is
computed the percentage of landslide area in each positively or negatively correlated to the stability of
mapping unit. Very old landslides (A0 in Table 2) the mapping unit.
were excluded from the landslide inventory, and con- Twenty-six of the 46 thematic variables (58%)
sidered as a thematic variable, i.e., as an additional entered in all five susceptibility models, confirming
independent variable describing the strength of the their importance in explaining the geographical distri-
lithological types. Following the procedure adopted bution of past landslides. Variables entered in all five
by Carrara et al. (1991), we selected a threshold of 3% models include (Table 4) morphological (ORDER,
of landslide area to establish if a mapping unit was: (i) LINK_LEN, SLO_AREA, R, SLO_ANG, SLO_
free of landslides (V 3%), or (ii) contained slope fail- ANG2, ANG_STD, LNK_ANG, CONV, COC_COV,
ures (N3%). We selected the threshold to account for RET, CC, TR1), lithological (ALLUVIO, AR_
possible mapping, drafting and digitizing errors in the BIS, AR_R_M_P, DETRITO, MR_AN_LO, MR_
compilation of the landslide inventory map (Carrara et B_R_C, MR_BOSM, MR_P_R_B), and land use
al., 1991). (BD, BMD, INC, PRA, SEM, REG) parameters, and
Fig. 7 shows the results of five statistical models the presence of very old, relict landslide deposits
prepared using the same set of environmental vari- (FRA_OLD). Inspection of Table 4 reveals that 8
ables, and changing incrementally the landslide inven- thematic variables entered all five models with
tory map. We prepared the first susceptibility model large standard discriminant function coefficients

Fig. 7. Landslide susceptibility models obtained through discriminate analysis of the same set of independent thematic variables (Table 4) and
changing the landslide inventory map (dependent variable, Fig. 3 and Table 2). A) Using landslides identified in the period A1–A2 (Table 2); B)
using landslides in the period A1–B2; C) using landslides in the period A1–C2; D) using landslides in the period A1–D2; E) using landslides in
the period A1–E2. Shades of gray indicate spatial probability, in 5 classes. Square bracket indicates class limit is included; round bracket
indicates class limit is not included.
F. Guzzetti et al. / Geomorphology 72 (2005) 272–299 287
288 F. Guzzetti et al. / Geomorphology 72 (2005) 272–299

(SDFCN |0.200|), equally distributed between nega- y-axis the percentage of landslide area in each sus-
tive (unstable) and positive (stable) values, including ceptibility class. In the Figure, the four curves are
the slope-unit hydrological order (ORDER), the similar, but provide different information. Curves I
slope-unit gradient (SLO_ANG, SLO_ANG2), the and II relate the percentage of landslide are used to
slope of the drainage channel (LNK_ANG), the topo- prepare the model (past landslides, A1–A2 for model
graphic profile of the slope-unit (CC), the presence of A, and A1–E2 for model B) to the predicted suscept-
alluvial deposits (ALLUVIO) and of massive sand- ibility. Curves III and IV relate the cumulative per-
stone (AR_BIS), and the presence of uncultivated or centage of landslide occurred bafter the model was
abandoned land (INC). preparedQ, to the model prediction. While the curves
Fig. 8 allows for a comparison of the models I and II measure model fit, curves III and IV provide
performance. By upgrading the landslide inventory, a quantitative measure of the model ability to predict
the total number of mapping units correctly classified, future landslides geographically, a form of model
a measure of the model fit (Chung and Fabbri, 2003), validation (Chung and Frabbri, 1999; Remondo et
increases from 74.2% to 78.9%. This indicates that a al., 2003). As expected, model fit is better than
more complete inventory improves the model fit. model performance, which decreases with the
However, we note that the enhancement is not very increase of the time span of the prediction. We
large (4.7%), indicating that even the susceptibility attribute the improved performance of the model
model prepared using solely the 1955 landslide inven- (e.g. from model III to IV) chiefly to the augmented
tory is sufficiently robust as to explain the known number of landslides (from 1719, 40.76 km2 to
distribution of the post-1955 slope failures. By adding 2722, 47.39 km2).
new landslides, the number of stable mapping units Table 5 shows similar results. In each row, corre-
that are correctly classified increases 1.6%, less than sponding to a different susceptibility model, the table
the number of unstable slope units wrongly attributed lists the percentage and the total landslide area (in
to the stable class, which decreases 5%. Correspond- square kilometres) mapped in a given period that falls
ingly, the number of misclassified stable mapping in 5 classes of landslide susceptibility, from very high
units decreases less than the misclassified unstable (VH) to very low (VL). The first block of numbers
units. By adding the landslides mapped in the 1999 (percentage and total area) in each row compares the
inventory (E1–E2 in Table 2) the model performance model forecast with the sub-set of landslides mapped
does not change. We attribute it to the small number as active in the same set of aerial photographs used to
of new landslides in the 1999 map (57), and to the obtain the inventory used for the production of the
performance of the model. A relevant proportion model. The remaining blocks show how well a model
(50% in number, 84% in area) of the new landslides was capable of predicting future landslides. Inspection
in the period 1994–1999 (E1–E2) occurred on S–SW of the table reveals that all the models correctly
facing slopes (Fig. 3E). This may explain why the classified as dlandslide proneT most the areas where
only difference between the variables entered in the active landslides were identified, and most of the areas
1994 model and those entered into the 1999 model is where future (with respect to the model) landslides
the absence of variable TR2 (slope unit facing S–SE) occurred. As an example, for model A (first row in
and the presence of variable TR3 (slope unit facing S– Table 5), which was prepared using landslides
SW) (Table 4). obtained by interpreting the 1955 aerial photography
Information is available to attempt a quantitative (A1–A2 in Table 2), the first block shows that 79% of
validation of the susceptibility model. To accomplish all landslides recognized as active in 1955 occurred
this, we compute the total area of new landslides (at within high (H, 30%) and very high (VH, 49%)
the date of the photographs) in each mapping unit. susceptibility zones, and only 9% in low (L, 8%)
We then compare the results with the susceptibility and very low (VL, 1%) susceptibility areas. Model
zoning obtained by the different discriminant models A performs less efficiently when it comes to predict
(Fig. 7). Fig. 9 shows, on the x-axis the probability the landslides occurred in the period from 1955 to
of landslide spatial occurrence, (i.e. susceptibility, 1975 (B1–B2 in Table 2). The model was capable of
ranked from most to least susceptible), and on the correctly predicting 75% of the new landslides, but
F. Guzzetti et al. / Geomorphology 72 (2005) 272–299 289

Table 4
Variables entered into the five discriminant models (see Fig. 7)
Variable description Variable Model SDFC
Model A1 Model B2 Model C3 Model D4 Model E5
Drainage channel magnitude MAGN – 0.126 0.117 0.124 0.124
Drainage channel order ORDER 0.218 0.288 0.260 0.232 0.232
Drainage channel length LINK_LEN 0.065 0.180 0.150 0.119 0.119
Slope unit contributing area AREAT_C 0.083 – – – –
Slope unit area SLO_AREA 0.282 0.162 0.191 0.246 0.246
Slope unit micro-relief R 0.177 0.125 0.133 0.139 0.139
Slope unit mean elevation ELV_M – – – 0.097 0.097
Slope unit mean terrain gradient SLO_ANG 0.919 0.470 1.754 1.456 1.456
Slope unit mean terrain gradient squared SLO_ANG2 1.116 1.165 1.236 1.265 1.265
Slope unit terrain gradient standard deviation ANG_STD 0.092 0.200 0.205 0.190 0.190
Drainage channel mean slope LNK_ANG 0.239 0.284 0.267 0.286 0.286
Slope unit length SLO_LEN 0.087 0.078 0.065 – –
Slope unit length standard deviation LEN_STD – 0.079 0.066 0.125 0.125
Slope unit terrain gradient (lower portion) ANGLE1 – 0.287 0.178 – –
Slope unit terrain gradient (middle portion) ANGLE2 0.312 – 0.437 0.186 0.186
Slope unit terrain gradient (upper portion) ANGLE3 0.246 0.427 – – –
Concave profile down slope CONV 0.164 0.204 0.211 0.208 0.208
Concave–convex profile COV_COC 0.048 – – 0.051 0.051
Convex–concave profile COC_COV 0.115 0.129 0.135 0.150 0.150
Rectilinear slope profile RET 0.066 0.047 0.064 0.048 0.048
Complex slope profile CC 0.334 0.272 0.245 0.200 0.200
Argillitic unit AG_VA_PA 0.133 – – – –
Zebedassi limestone ALB_ZEB 0.072 – – – –
Recent alluvial deposit ALLUVIO 0.694 0.584 0.571 0.592 0.592
Monte Vallassa sandstone AR_BIS 0.427 0.388 0.397 0.441 0.441
Ranzano sandstone AR_R_M_P 0.071 0.059 0.052 0.065 0.065
Scabiazza sandstone AR_SCA – 0.060 0.056 0.063 0.063
Chaotic complex AT_PA_CA 0.219 – – – –
Detritus DETRITO 0.081 0.077 0.094 0.099 0.099
Monte Lumello marl MR_AN_LO 0.293 0.130 0.122 0.146 0.146
Rigoroso marl MR_B_R_C 0.119 0.093 0.092 0.129 0.129
Bosmenso marl MR_BOSM 0.108 0.062 0.044 0.051 0.051
Monte Piano marl MR_P_R_B 0.118 0.109 0.107 0.122 0.122
Cassano–Spinola conglomerate SACONG 0.097 0.030 0.029 – –
Bare rock or soil ALV – 0.055 0.064 0.085 0.085
Dense forest BD 0.052 0.052 0.053 0.048 0.048
Woods BMD – – – 0.038 0.038
Uncultivated area INC 0.212 0.273 0.292 0.281 0.281
Pasture PRA 0.176 0.192 0.199 0.222 0.222
Cultivated area SEM 0.184 0.246 0.285 0.277 0.277
Bedding dipping into the slope REG 0.182 0.119 0.111 0.078 0.078
Bedding dipping out of the slope FRA 0.101 – – – –
Chaotic bedding attitude CAO 0.059 – – – –
Slope unit facing N–NE TR1 0.129 0.079 0.065 0.073 0.126
Slope unit facing S–SE TR2 – – 0.037 0.042 –
Slope unit facing S–SW TR3 – – – – 0.053
Very old (relict) landslide (A0) FRA_OLD 0.092 0.061 0.062 0.053 0.053
(1) Model A obtained using landslides A1–A2 (Fig. 7A).
(2) Model B obtained using landslides A1–B2 (Fig. 7B).
(3) Model C obtained using landslides A1–C2 (Fig. 7C).
(4) Model D obtained using landslides A1–D2 (Fig. 7D).
(5) Model E obtained using landslides A1–E2 (Fig. 7E).
Variables with large standard discriminant function coefficients (SDFC), in absolute value, are shown in bold.
290 F. Guzzetti et al. / Geomorphology 72 (2005) 272–299

100

90
83.9 83.8 85.0 85.0
80 80.0
78.9 78.9
Percentage of Mapping Units
77.0 77.7
74.2
70 69.0 69.0
67.4 67.1 68.4
60

50

40
32.6 32.9 31.6 31.0 31.0
30

20 20.0
16.1 16.2 15.0 15.0
10

0
A1–A2 A1–B2 A1–C2 A1–D2 A1–E2

Fig. 8. Degree of fit of the five landslide susceptibility models. In the graph, x-axis shows landslide inventories used in the analysis (see also
Table 2), and y-axis shows percentage of mapping units. Large black circles: overall percentage of mapping units correctly classified by the 5
susceptibility models. Grey and open symbols show mapping units correctly and incorrectly classified, respectively. Grey diamonds show
percentage of mapping units free of landslides classified as stable. Grey squares show percentage of mapping units having landslides classified
as unstable. Open diamonds show percentage of mapping units free of landslides misclassified as unstable (type 1 errors). Open squares show
percentage of mapping units having landslides misclassified as stable (type 2 errors).

failed to predict 12% of the slope failures that In conclusion, we take the outcome of the last
occurred in low (9%) and in very low (3%) suscept- discriminant model (Fig. 7E) as a quantitative mea-
ibility areas. Note that the percentage of landslides sure of the spatial probability of the landslides in the
falling in the very high susceptibility (34%) class is Staffora basin, and we use it to determine landslide
lower than the percentage in the high susceptibility hazard in the catchment.
class (41%), an indication of the reduced performance
of the model. Performance of model A decreases 4.5. Landslide hazard
further if one considers the landslides occurred after
1975. In particular, the model A was capable of pre- We now have all the information to determine
dicting bonlyQ 54% of the landslides identified in the quantitatively landslide hazard in the Staffora
period 1994–1999 (E1–E2 in Table 2), with a consid- basin. Fig. 10 summarizes the adopted workflow.
erable proportion of slope failures (31%) occurring in We use:
the unclassified (U) susceptibility class. Considering
the complexity of the Staffora basin, and the limited (i) the probability of landslide size, a proxy for
number of landslides that occurred in the period (57), landslide magnitude, obtained from the statisti-
we regard the model performance as very good. Data cal analysis of the frequency–area distribution
in Table 5 can also be used to determine the contribu- of the mapped landslides (Eqs. (5) and (6) and
tion of new landslides to the model performance. Fig. 4),
Considering the last column, one can see that model (ii) the probability of landslide occurrence for
A was capable of predicting 54% of the landslides established periods, obtained by computing
occurred in the period 1994–1999 (E1–E2 in Table 2), the mean recurrence interval between succes-
and model D correctly predicted 98% of the landslides sive failures in each mapping unit, and adopt-
in the same period, with the majority of the slope ing a Poisson probability model (Eq. (9) and
failures (84%) falling in the very high susceptibility Fig. 5), and
class and none of the failures occurring in the very (iii) the spatial probability of slope failures (i.e.
low susceptibility class. susceptibility) obtained through discriminant
F. Guzzetti et al. / Geomorphology 72 (2005) 272–299 291

100

90
Percentage of landslide area in the susceptibility classes
80

70

60

50

40

30 I
II
20
III
10
IV
0
0 10 20 30 40 50 60 70 80 90 100
most susceptible least susceptible
Percentage of study area in the susceptibility classes

Fig. 9. Graph comparing model fit and model performance (prediction rate). In the graph, the x-axis shows the probability of landslide spatial
occurrence (susceptibility), ranked from most (left) to least (right) susceptible, and y-axis shows the percentage of landslide area in each
susceptibility class. Curve I and II illustrate model fit. Curve I shows the ability of the model obtained using landslides identified in the period
A1–A2 (Table 2) to forecast the same set of landslides. Curve II shows the capacity of the model prepared using landslides in the period A1–E2 to
forecast recent landslides identified in the 1999 aerial photographs (E1–E2). Curve III and IV illustrate model performance. Curve III indicates
the ability of the model prepared using landslides identified in the period A1–A2 to forecast bfutureQ landslides occurred in the period B1–E2.
Curve IV indicates the ability of the model prepared using landslides identified in the period A1–B2 to forecast bfutureQ landslides occurred in
the period C1–E2.

analysis of 46 environmental variables (Eq. (13) 5. Discussion of the results


and Fig. 7).
The proposed method allowed us to determine
Assuming independence (Eq. (2)), we multiply the quantitatively landslide hazard in the Staffora River
three probabilities and we obtain landslide hazard, i.e. basin. We obtained most of the information used in
the joint probability that a mapping unit will be the analysis from a detailed multi-temporal inven-
affected by future landslides that exceed a given tory map. Production of the multi-temporal in-
size, in a given time, and because of the local envir- ventory required a total of 6 months of three
onmental setting. Fig. 11 shows examples of the land- geomorphologists. Digitization and validation of
slide hazard assessment. The figure portrays landslide the map required a total of one month of two GIS
hazard for mapping units in the central part of the experts. Statistical analysis and hazard modelling
Staffora river basin, for four periods (5, 10, 25 and 50 required three weeks of two geomorphologists. Con-
years), and for two different landslide sizes, greater or sidering the thematic information used to ascertain
equal than 2000 m2, and greater or equal than 10,000 susceptibility (i.e., DTM, geology, land use) was
m2 (1 ha). already available in digital format (Antonini et al.,
292 F. Guzzetti et al. / Geomorphology 72 (2005) 272–299

Table 5
Validation of susceptibility models
Susceptibility class Landslide set and age
A1–A2 B1–B2 C1–C2 D1–D2 E1–E2
% km2 % km2 % km2 % km2 % km2
Model A Fig. 7A VH 49 1.21 34 2.21 38 1.22 36 1.22 17 0.18
H 30 0.74 41 2.70 38 1.23 44 1.46 37 0.39
U 12 0.30 13 0.85 15 0.48 12 0.39 31 0.32
L 8 0.20 9 0.62 8 0.26 7 0.24 13 0.13
VL 1 0.01 3 0.19 1 0.05 1 0.04 2 0.02
Model B Fig. 7B VH 73 3.21 73 2.36 73 2.46 70 0.73
H 18 0.81 20 0.65 20 0.67 24 0.25
U 3 0.15 4 0.13 2 0.06 4 0.04
L 4 0.16 2 0.08 4 0.13 2 0.02
VL 2 0.07 1 0.02 1 0.03 0 0.00
Model C Fig. 7C VH 72 1.72 74 2.50 70 0.73
H 22 0.53 19 0.64 24 0.25
U 3 0.09 2 0.06 4 0.04
L 2 0.04 4 0.13 2 0.02
VL 1 0.02 1 0.02 0 0.00
Model D Fig. 7D VH 84 1.15 84 0.87
H 11 0.15 14 0.15
U 2 0.02 0 0.00
L 3 0.04 2 0.02
VL 0 0.00 0 0.00
Model E Fig. 7E VH 80 0.68
H 19 0.16
U 0 0.00
L 1 0.01
VL 0 0.00
Susceptibility classes are: VH, very high [1.0–0.8]; H, high [0.8–0.6]; U, uncertain [0.6–0.4]; L, low [0.4–0.2]; and VL, very low [0.0–0.2]. See
text for explanation.

2000; Ardizzone et al., 2002), production of the will occur in the future under the same circum-
multi-temporal inventory was the most difficult stances and because of the same factors that pro-
and time consuming-operation. duced them in the past; (ii) landslide events are
The proposed probability model allows for its independent (uncorrelated) random events in time;
temporal verification. Providing an estimate of the (iii) the mean recurrence of slope failures will
period of landslide hazard, the forecast is verifiable remain the same in the future as it was observed
at any moment in time, provided accurate information in the past; (iv) the statistics of landslide area are
on landslides is available. This can be obtained by correct and will not change in the future; (v) land-
systematically compiling information on landslide slide area is a reasonable proxy for landslide mag-
occurrence, and particularly after each main land- nitude; and (vi) the probability of landslide size, the
slide-triggering event (i.e. a heavy rainstorm, a pro- probability of landslide occurrence for established
longed rainfall period, a snowmelt event, or an periods, and the spatial probability of slope failures,
earthquake). When new information on landslides are independent. We now discuss these assumptions
becomes available, new spatial (S), temporal ( P N) for the Staffora basin.
or size ( PA L) models can be prepared, and the hazard That landslides will occur in the future under the
model revised. same conditions and because of the same factors
The proposed landslide hazard model holds that triggered them in the past is a recognized
under a set of assumptions, namely: (i) landslides postulate for all functional (statistically based) sus-
F. Guzzetti et al. / Geomorphology 72 (2005) 272–299 293

DATA MODELS RESULTS


Environmental Thematic
Variables Susceptibility
Discriminant
assessment
Analysis
(Figure 7E)
(eq. 16)
Multi-temporal landslide “Where”
inventory map
(Figure 3)

Exceedance
Poisson model probability
(eqs. 7, 8, 9) (Figure 5)
“When”

Probability of
Inverse Gamma
landslide area
Double Pareto
(Figure 4 A-B)
(eqs. 4, 5, 6)
“How large”

Landslide Hazard
Assessment
(Figure 11)

Fig. 10. Block diagram exemplifying the work flow adopted to determine landslide hazard. Rectangles indicate input data. Diamonds indicate
individual models, for landslide susceptibility, for the temporal probability of landslides, and for landslide size. Ellipses indicate intermediate
results. Hexagon indicates the final result.

ceptibility assessments (Carrara et al., 1991; Hutch- able. Inspection of Table 4 indicates that 42 of the 47
inson, 1995; Aleotti and Chowdhury, 1999; Chung thematic variables entered into the susceptibility
and Frabbri, 1999; Guzzetti et al., 1999). We dis- model are not expected to change significantly in
cussed the general limitations of this assumption the considered period. However, other variables and
when we presented the susceptibility model. We most notably land use types may change significantly
will now focus on the validity of the postulate in in the period. Qualitative estimates indicate a reduc-
the Staffora basin. tion of about 25% of the forest coverage in the period
The difficulty with the principle of uniformitarian- from 1955 to 2000, in favour of cultivated land. In
ism lays in the fact that the environmental conditions the same period, agricultural practices have changed,
(predisposing factors) that caused landslides bmust largely aided by large and powerful mechanical
remain the same in the futureQ in order to cause equipment. In the central and southern Apennines,
similar slope failures. Our hazard model has an areas recently deforested for agricultural purposes
expected validity of 50 years. The problem is to are more prone to shallow landslides (Cardinali et
investigate the possibility that the predisposing fac- al., 2000). If this will be the case for the Staffora
tors will change in the considered period. It is safe to basin, some of the environmental variables consid-
assume that geological factors (e.g. lithology, struc- ered in the susceptibility model will change, possibly
ture, seismicity) will not change (significantly) in hampering the validity of the model, and new vari-
such a short time. Local morphological modifications ables showing areas of land use change should be
are possible in the period, due chiefly to stream considered to describe the initiation of shallow slope
erosion, landslides and human actions, but extensive failures. We note that the susceptibility model does
(widespread) morphological changes are not foresee- not consider the landslide triggering factors, i.e. rain-
294 F. Guzzetti et al. / Geomorphology 72 (2005) 272–299

Landslide size Landslide size In the Apennines, evidence exists that where abun-
≥ 2,000 m2 ≥ 10,000 m2 dant clay, marl and sandstone crop out, landslides
exhibit spatial persistence, i.e. they tend to occur
where they have occurred in the past (Cardinali et
al., 2000). If this is the case for the Staffora River
5 years basin, the assumption that landslide events are uncor-
related random events in time will be violated. Ana-
lysis of the multi-temporal inventory map (Fig. 3)
indicates that in the study area 40% of all the land-
slides identified in the period from 1955 to 1999 (A2–
E2 in Table 2) occurred inside landslides mapped on
the 1955 aerial photographs (A0–A1 in Table 2).
Considering only the 2390 landslides occurred in the
10 years 45-year period from 1955 to 1999 (A2–E2 in Table 2),
12% of the slope failures occurred in the same area of
other landslides triggered in the same period. Analysis
of the historical record of damaging landslides indi-
cates that the 389 events listed in the catalogue
occurred at 332 different sites, with only 38 sites
affected two or more times. The same bsiteQ was
affected on average 1.2 times, indicating a low rate
25 years of recurrence of events at the same site. All this
concurs to establish that for the period of our hazard
assessment (50 years), in the Staffora River basin
landslides can be considered uncorrelated random
events in time.
Analysis of the record of damaging landslides
reveals that of the 248 events (63.7%, 96.0% of
50 years
which after 1950) for which the triggering mechan-
ism is known, 210 (84.7%) were the result of intense
rainfall, 16 (6.5%) to a combination of intense rain-
fall and snow melting, infiltration, irrigation or bro-
ken pipes, 14 (5.6%) to erosion at the base of the
0 1 km
slope, and eight (3.2%) to other causes. Earthquake-
induced landslides were not reported. The analysis
Fig. 11. Examples of landslide hazard maps for four periods, from 5 indicates that most of the landslides in the Staffora
to 50 years (from top to bottom), and for two landslide sizes, River basin are rainfall-induced. If the rate of occur-
A L z 2000 m2 (left) and A L z 10,000 m2 (right). Shades of gray rence of the meteorological events that trigger land-
show different joint probabilities of landslide size, of landslide slides changes, the mean rate of slope failures will
temporal occurrence, and of landslide spatial occurrence (suscept-
ibility). For improved readability, the scales of landslide hazard
also change. If the intensity (amplitude and duration)
differ for the two landslide sizes. of the rainfall will change, the rate of slope failures
might change, in a way that is not easily predictable.
For the coming decades, south of the Alps models
fall, seismic shaking or snow melting. Changes in the of global climate change forecast the same total
frequency or intensity of the driving forces will not amount of yearly rainfall concentrated in a fewer
affect (at least not in the considered period) the number of high intensity events (Bradley et al.,
susceptibility model. However, it may affect the rate 1987; Brunetti et al., 2000; Easterling et al., 2000;
of occurrence of landslide events. IPCC, 2001). This may result in more abundant
F. Guzzetti et al. / Geomorphology 72 (2005) 272–299 295

shallow landslides, and in less frequent deep-seated slide area a good measure of landslide destructiveness
slope failures (Buma and Dehn, 1998; Malet et al., in the Staffora basin? Analysis of the historical cat-
2005). Modifications in land use induced by changes alogue of damaging slope failures reveals that infor-
in agricultural practices may also change the rate of mation on the size (area, length and width) of
occurrence of landslides. landslides is available for 26 events (6.7%), which
Determining the statistics of landslide areas is no range from 600 m2 to 600,000 m2 (mean = 58,000 m2,
trivial task (Malamud et al., 2004). Only a handful of std. dev. = 13,500 m2). Damage caused by these land-
inventories are available which are substantially slides was mostly to the road network and, sub-
complete and have enough cartographic detail to ordinately, to private homes and to the aqueduct.
allow for determining unambiguously the PDF of Casualties were not reported. Information on the land-
landslide areas (Stark and Hovius, 2001; Guzzetti slide type is available for 28 events (7.2%), of which
et al., 2002a,b; Guthrie and Evans, 2004a,b; Mala- 15 were slides, 6 flows and 5 falls. Slides and flows
mud et al., 2004). The (scant) available information caused the most severe damage, and falls produced
indicates that the frequency–area statistics of land- only minor interruptions along the roads. Information
slide areas does not change significantly across litho- on landslide velocity is available for five events, and
logical or physiographical boundaries. Malamud et ranges from 1.2  10 1 to 5.8  10 3 mm/s (moderate
al. (2004) showed that three different populations of landslide velocity (Cruden and Varnes, 1996)). As a
landslides produced by different triggers (i.e. seismic whole, the available historical information on dama-
shaking, intense rainfall, rapid snow melting) in ging slope failures suggests that: (i) damage in the
different physiographical regions (southern Califor- Staffora River basin is caused mostly by slow to
nia, central America, central Italy), exhibit virtually rapid moving slides and flows, i.e. the type of failures
identical PDFs. Unpublished work conducted in cen- considered in the hazard assessment, and (ii) large
tral Italy indicates that for the same physiographical landslides tend to produce larger damage, particularly
region the PDF of landslide area does not change in to roads.
time. It is, therefore, safe to assume that in the The last assumption of the proposed model is that
Staffora River basin the frequency–area statistics of the probabilities of landslide size, of temporal occur-
landslide area will not change in the period of the rence, and of spatial incidence of mass movements are
hazard assessment. It is also justified to use a single independent. The legitimacy of this assumption is
PDF for the entire basin. Since the most abundant difficult to prove quantitatively. We have shown that
landslides in the catchment are small (~2000 m2, Fig. the probability of landslide area is largely independent
4A), great care must be taken in mapping accurately from the physiographical setting. As a first-approxi-
(and completely) the small slope failures. The slope mation, it is safe to conclude that the probability of
of the heavy tail of the PDF in Fig. 4A is controlled landslide area is independent from susceptibility. The
by a limited number of landslides. There are 16 susceptibility model was constructed without consid-
landslides larger than 50,000 m2 (5 ha) and only ering the driving forces (meteorological or else) that
one landslide larger than 100,000 m2 (10 ha). Care control the rate of occurrence of slope failures in the
must be taken in mapping the largest landslides, and Staffora River basin. We conclude that the rate of
in deciding whether they represent an individual landslide events is independent from susceptibility.
slope failure or the result of two or more coalescent The catalogue of historical damaging landslides
landslides. reveals that landslides occurred in all sizes. We con-
Hungr (1997) argued that no unique measure of sider this an indication that the rate of failures is
landslide magnitude is available, and proposed to independent from landslide size.
adopt destructiveness as measure of landslide mag- Finally, the main scope of a landslide hazard
nitude. In this work, we have taken landslide area as assessment is to provide quantitative expertise on
a proxy for landslide destructiveness (and of land- future slope failures to planners, decision-makers,
slide magnitude). We obtained the area of the in- civil defence authorities, insurance companies, land
dividual slope failures from the multi-temporal developers, and individual landowners. The pro-
landslide inventory in GIS format. However, is land- posed method allowed us to prepare a large number
296 F. Guzzetti et al. / Geomorphology 72 (2005) 272–299

of different hazard maps (Fig. 11), depending on the ported by CNR GNDCI (publication number 2887)
adopted susceptibility model, the established period, and CNR IRPI grants. Early part of the work con-
and the minimum size of the expected landslide. ducted under contract from the Regione Lombardia.
How to combine such a large number of hazard
maps efficiently, producing cartographic, digital, or
thematic products useful for the large range of inter- References
ested users, remains an open problem that needs
investigation. Aleotti, P., Chowdhury, R., 1999. Landslide hazard assessment:
summary review and new perspectives. Bulletin of Engineering
Geology and the Environment 58, 21 – 44.
Antonini, G., Ardizzone, F., Cardinali, M., Carrara, A., Detti, R.,
6. Conclusions Galli, M., Guzzetti, F., Reichenbach, P., Sotera, M., Tonelli,
G., 2000. Rapporto conclusivo: tecniche e metodologie idonee
We have proposed a probabilistic model to ascer- alla produzione di carte della pericolosità e del rischio da
tain landslide hazard that fulfils the definition of frana in aree campione rappresentative del territorio della
regione Lombardia. CNR IRPI unpublished report, 139 p.
landslide hazard given by Varnes and the IAEG
(in Italian).
Commission on Landslides and other Mass-Move- Ardizzone, F., Cardinali, M., Carrara, A., Guzzetti, F., Reichenbach,
ments (1984), amended by Guzzetti et al. (1999) to P., 2002. Uncertainty and errors in landslide mapping and land-
include the magnitude of the landslide event. The slide hazard assessment. Natural Hazards and Earth System
model expresses landslide hazard as the joint prob- Science 2 (1–2), 3 – 14.
Atckinson, P.M., Massari, R., 1998. Generalised linear modelling of
ability of landslide size, considered a proxy for land-
susceptibility lo landsliding in the Central Apennines, Italy.
slide magnitude, of landslide occurrence in an Computers & Geosciences 24 (4), 373 – 385.
established period of time, and of landslide spatial Ayalew, L., Yamagishi, H., 2005. The application of GIS-based
occurrence given the local environmental setting. We logistic regression for landslide susceptibility mapping in the
tested the model in the Staffora River basin, for Kakuda–Yahiko Mountains, Central Japan. Geomorphology 65
(1–2), 15 – 31.
which we obtained a quantitative estimate of land-
Ayalew, L., Yamagishi, H., Ugawa, N., 2004. Landslide suscept-
slide hazard. We obtained most of the information ibility mapping using GIS-based weighted linear combination,
used to determine landslide hazard from a detailed the case in Tsugawa area of Agano River, Niigata Prefecture,
multi-temporal inventory map. The model proved Japan. Landslides 1 (1), 73 – 81.
applicable in the test area, and we believe is ap- Baeza, C., Corominas, J., 2001. Assessment of shallow landslide
susceptibility by means of multivariate statistical techniques.
propriate in similar areas, chiefly where a multi-
Earth Surface Processes and Landforms 26 (12), 1251 – 1263.
temporal landslide inventory captures the types, Binaghi, E., Luzi, L., Madella, P., Pergalani, F., Rampini, A., 1998.
sizes, and expected recurrence of slope failures. Slope instability zonation: a comparison between certainty fac-
Improvements to the model may include a more tor and Fuzzy Dempster–Shafer approaches. Natural Hazards
robust definition of the spatial probability of land- 17, 77 – 97.
Bonham-Carter, G.F., 1991. Integration of geosceintic data using
slide occurrence, a better temporal model, and im-
GIS. In: Goodchild, M.F., Rhind, D.W., Maguire, D.J. (Eds.),
proved number-size statistics, particularly for the Geographic Information Systems: Principle and Applications.
smallest landslides. Lastly, in this exercise we have Longdom, London, pp. 171 – 184.
prepared tens of different hazard maps. How to com- Brabb, E.E., 1984. Innovative approach to landslide hazard and risk
bine such a large number of maps efficiently, produ- mapping. Proceedings of the 4th International Symposium on
Landslides, Toronto, vol. 1, pp. 307 – 324.
cing outputs useful for planners and decision-makers
Bradley, R.S., Diaz, H.F., Eischeid, J.K., Jones, P., Kelly, P., Good-
remains an open problem. ess, C., 1987. Precipitation fluctuations over Northern Hemi-
sphere land areas since the mid-19th Century. Science 237,
171 – 175.
Acknowledgements Brunetti, M., Buffoni, L., Maugeri, M., Nanni, T., 2000. Precipita-
tion intensity trends in Northern Italy. International Journal of
Climatology 20, 1017 – 1032.
We acknowledge the help of Florisa Melone in Bucknam, R.C., Coe, J.A., Chavarria, M.M., Godt, J.W., Tarr, A.C.,
formalizing the hazard model. We are grateful to Bradley, L.-A., Rafferty, S., Hancock, D., Dart, R.L., Johnson,
Bruce Malamud for reviewing the paper. Work sup- M.L., 2001. Landslides triggered by Hurricane Mitch in Guate-
F. Guzzetti et al. / Geomorphology 72 (2005) 272–299 297

mala — inventory and discussion. U.S. Geological Survey Open the Yucca Mountain region, Nevada. Journal of Geophysical
File Report 01-443 (38 pp.). Research 100, 10107 – 10125.
Buma, J., Dehn, M., 1998. A method for predicting the impact of Costa, J.A., Baker, V.R., 1981. Surficial Geology — Building with
climate change on slope stability. Environmental Geology 35 the Earth. John Wiley and Sons, New York. 498 pp.
(2–3), 190 – 196. Crovelli, R.A., 2000. Probability models for estimation of number
Cardinali, M., Ardizzone, F., Galli, M., Guzzetti, F., Reichenbach, and costs of landslides. United States Geological Survey Open
P., 2000. Landslides triggered by rapid snow melting: the File Report 00-249.
December 1996–January 1997 event in Central Italy. In: Cruden, D.M., Varnes, D.J., 1996. Landslide types and processes.
Claps, P., Siccardi, F. (Eds.), Proceedings 1st Plinius Con- In: Turner, A.K., Schuster, R.L. (Eds.), Landslides, Investigation
ference on Mediterranean Storms. Bios Publisher, Cosenza, and Mitigation, Special Report, vol. 247. Transportation
pp. 439 – 448. Research Board, Washington, D.C, pp. 36 – 75.
Cardinali, M., Carrara, A., Guzzetti, F., Reichenbach, P., 2002a. Dai, F.C., Lee, C.F., 2002. Landslide characteristics and slope
Landslide hazard map for the Upper Tiber River basin. instability modelling using GIS, Lantau Island, Hong Kong.
CNR GNDCI Publication number 2116, map at 1 : 100,000 Geomorphology 42, 213 – 228.
scale. Dai, F.C, Lee, C.F., 2003. A spatiotemporal probabilistic modelling
Cardinali, M., Reichenbach, P., Guzzetti, F., Ardizzone, F., of storm-induced shallow landsliding using aerial photographs
Antonini, G., Galli, M., Cacciano, M., Castellani, M., Sal- and logistic regression. Earth Surface Processes and Landforms
vati, P., 2002b. A geomorphological approach to estimate 28 (5), 527 – 545.
landslide hazard and risk in urban and rural areas in Umbria, Detti, R., Pasqui, V., 1995. Vector and raster structures in generating
central Italy. Natural Hazards and Earth System Science 2 drainage-divide networks from Digital Terrain Models. In: Car-
(1–2), 57 – 72. rara, A., Guzzetti, F. (Eds.), Geographical Information Systems
Carrara, A., 1983. A multivariate model for landslide hazard eva- in Assessing Natural Hazards. Kluwer Academic Publisher,
luation. Mathematical Geology 15, 403 – 426. Dordrecht, The Netherlands, pp. 35 – 55.
Carrara, A., Cardinali, M., Detti, R., Guzzetti, F., Pasqui, V., Reich- Easterling, D.R., Meehl, G.M., Parmesan, C., Changnon, S.A., Karl,
enbach, P., 1991. GIS techniques and statistical models in T.R., Mearns, L.O., 2000. Climate extremes: observations, mod-
evaluating landslide hazard. Earth Surface Processes and Land- eling, and impacts. Science 289, 2068 – 2074.
form 16 (5), 427 – 445. Fabbri, A.G., Chung, C.J., Cendrero, C., Remondo, J., 2003. Is
Carrara, A., Cardinali, M., Guzzetti, F., 1992. Uncertainty in asses- prediction of future landslides possible with a GIS? Natural
sing landslide hazard and risk. ITC Journal 2, 172 – 183. Hazards 30 (3), 487 – 503.
Carrara, A., Cardinali, M., Guzzetti, F., Reichenbach, P., 1995. GIS Guthrie, R.H., Evans, S.G., 2004a. Magnitude and frequency of
technology in mapping landslide hazard. In: Carrara, A., Guz- landslides triggered by a storm event, Loughborough Inlet,
zetti, F. (Eds.), Geographical Information Systems in Assessing British Columbia. Natural Hazards and Earth System Science
Natural Hazards. Kluwer Academic Publisher, Dordrecht, The 4, 475 – 483.
Netherlands, pp. 135 – 175. Guthrie, R.H., Evans, S.G., 2004b. Analysis of landslide frequen-
Carrara, A., Crosta, G.B., Frattini, P., 2003. Geomorphological and cies and characteristics in a natural system, coastal British
historical data in assessing landslide hazard. Earth Surface Columbia. Earth Surface Processes and Landforms 29 (11),
Processes and Landforms 28 (10), 1125 – 1142. 1321 – 1339.
Çevik, E., Topal, T., 2003. GIS-based landslide susceptibility map- Guzzetti, F., 2002. Landslide hazard assessment and risk evaluation:
ping for a problematic segment of the natural gas pipeline, overview, limits and prospective. Proceedings 3rd MITCH
Hendek (Turkey). Environmental Geology 44 (8), 949 – 962. Workshop Floods, Droughts and Landslides — Who Plans,
Chung, C.J., Frabbri, A.G., 1999. Probabilistic prediction models Who Pays, pp. 24 – 26. November 2002, Potsdam, available at
for landslide hazard mapping. Photogrammetric Engineering http://www.mitch-ec.net/workshop3/Papers/paper_guzzetti.pdf.
and Remote Sensing 65 (12), 1389 – 1399. Guzzetti, F., Carrara, A., Cardinali, M., Reichenbach, P., 1999.
Chung, C.J., Fabbri, A.G., 2003. Validation of spatial prediction Landslide hazard evaluation: an aid to a sustainable develop-
models for landslide hazard mapping. Natural Hazards 30 (3), ment. Geomorphology 31, 181 – 216.
451 – 472. Guzzetti, F., Reichenbach, P., Cardinali, M., Ardizzone, F., Galli, M.,
Chung, C.J., Fabbri, A.G., 2005. Systematic procedures of landslide 2002a. Impact of landslides in the Umbria Region Central Italy.
hazard mapping for risk assessment using spatial prediction Natural Hazards and Earth System Science 3 (5), 469 – 486.
models. In: Glade, C.J., et al., (Eds.), Landslide Risk Assess- Guzzetti, F., Malamud, B.D., Turcotte, D.L., Reichenbach, P.,
ment. John Wiley, pp. 139 – 174. 2002b. Power-law correlations of landslide areas in Central
Coe, J.A., Michael, J.A., Crovelli, R.A., Savage, W.Z., 2000. Pre- Italy. Earth and Planetary Science Letters 195, 169 – 183.
liminary map showing landslide densities, mean recurrence Harp, E.L., Jibson, R.L., 1996. Landslides triggered by the 1994
intervals, and exceedance probabilities as determined from his- Northridge, California earthquake. Seismological Society of
toric records, Seattle, Washington. United States Geological America Bulletin 86, S319 – S332.
Survey Open File Report 00-303. Hovius, N., Stark, C.P., Allen, P.A., 1997. Sediment flux from a
Connor, C.B., Hill, B.E., 1995. Three nonhomogeneous Poisson mountain belt derived by landslide mapping. Geology 25,
models for the probabilità of basaltic volcanism: application to 231 – 234.
298 F. Guzzetti et al. / Geomorphology 72 (2005) 272–299

Hungr, O., 1997. Some methods of landslide hazard intensity map- Pelletier, J.D., Malamud, B.D., Blodgett, T., Turcotte, D.L., 1997.
ping. In: Cruden, D.M., Fell, R. (Eds.), Landslide Risk Assess- Scale-invariance of soil moisture variability and its implications
ment. Balkema Publisher, Rotterdam, pp. 215 – 226. for the frequency–size distribution of landslides. Engineering
Hutchinson, J.N., 1995. Keynote paper: landslide hazard assess- Geology 48, 255 – 268.
ment. In: Bell, J.N. (Ed.), Landslides. Balkema, Rotterdam, Reger, J.P., 1979. Discriminant analysis as a possible tool in land-
pp. 1805 – 1841. slide investigations. Earth Surface Processes and Landforms 4,
Intergovernmental Panel on Climate Change, IPCC, 2001. Third 267 – 273.
Assessment Report, Working Group I. Summary for policy- Regione Lombardia, 2002. Inventario delle frane e dei dissesti
makers. Available at: www.ipcc.ch. idrogeologici della Regione Lombardia. Unità Operativa Dis-
Keaton, J.R., Anderson, L.R., Mathewson, C.C., 1988. Assessing sesti Idrogeologici, 2 CD-ROMs (in Italian).
debris flow hazards on alluvial fans in Davis County, Utah. In: Reichenbach, P., Galli, M., Cardinali, M., Guzzetti, F., Ardizzone,
Fragaszy, R.J. (Ed.), Proceedings 24th Annual Symposium on F., 2005. Geomorphologic mapping to assess landslide risk:
Engineering Geology and Soil Engineering. Washington State concepts, methods and applications in the Umbria Region of
University, Pullman, pp. 89 – 108. central Italy. In: Glade, P., et al., (Eds.), Landslide Risk Assess-
Klein, F.W., 1982. Patters of historical eruptions at Hawaiian vol- ment. John Wiley, pp. 429 – 468.
canoes. Journal of Volcanology and Geothermal Research 12, Remondo, J., Gonzalez, A., Diaz De Teran, J.R., Cendrero, A.,
1 – 35. Fabbri, A., Chung, C.F., 2003. Validation of landslide suscept-
Lee, S., 2004. Application of likelihood ratio and logistic regression ibility maps; examples and applications from a case study in
models to landslide susceptibility mapping using GIS. Environ- Northern Spain. Natural Hazards 30, 437 – 449.
mental Management 34 (2), 223 – 232. Rossetti, R., 1997. Centri abitati instabili della Provincia di Pavia —
Lee, S., Choi, J., Min, K., 2002a. Landslide susceptibility analysis Vol. 1, CNR GNDCI Publication number 1780.
and verification using the Bayesian probability model. Environ- Rowbotham, D., Dudycha, D.N., 1998. GIS modelling of slope
mental Geology 43 (1–2), 120 – 131. stability in Phewa Tal watershed, Nepal. Geomorphology 26,
Lee, S., Chwae, U., Min, K., 2002b. Landslide susceptibility 151 – 170.
mapping by correlation between topography and geological struc- Santacana, N., Baeza, B., Corominas, J., De Paz, A., Marturiá,
ture: the Janghung area, Korea. Geomorphology 46 (3–4), J., 2003. A GIS-based multivariate statistical analysis for
149 – 162. shallow landslide susceptibility mapping in La Pobla de
Lin, M.-L., Tung, C.C., 2004. A GIS-based potential analysis of the Lillet Area (Eastern Pyrenees, Spain). Natural Hazards 30
landslides induced by the Chi-Chi earthquake. Engineering (3), 281 – 295.
Geology 71 (1–2), 63 – 77. Servizio Geologico Nazionale, 1971. Carta Geologica d’Italia,
Lips, E.W., Wieczorek, G.F., 1990. Recurrence of debris Foglio 71 Voghera, scale 1 : 100,000.
flows on an alluvial fan in central Utah. In: French, R.H. Soeters, R., van Westen, C.J., 1996. Slope instability recognition
(Ed.), Hydraulic/Hydrology of Arid Lands, Proceedings of the analysis and zonation. In: Turner, A.K., Schuster, R.L. (Eds.),
International Symposium. American Society of Civil Engineers, Landslide Investigation and Mitigation, Special Report, vol. 247.
pp. 555 – 560. National Research Council, Transportation Research Board,
Malamud, B.D., Turcotte, D.L., Guzzetti, F., Reichenbach, P., 2004. pp. 129 – 177.
Landslide inventories and their statistical properties. Earth Sur- Stark, C.P., Hovius, N., 2001. The characterization of landslide size
face Processes and Landforms 29 (6), 687 – 711. distributions. Geophysics Research Letters 28 (6), 1091 – 1094.
Malet, J.-P., van Asch, T.H.W.J., van Beek, R., Maquaire, O., 2005. Süzen, M.L., Doyuran, V., 2004. A comparison of the GIS based
Forecasting the behaviours of complex landslides with a spa- landslide susceptibility assessment methods: multivariate versus
tially distributed hydrological model. Natural Hazards and Earth bivariate. Environmental Geology 45 (5), 665 – 679.
System Sciences 5 (1), 71 – 85. van Westen, C.J., 1994. In: Price, M.F., Heywood, D.I. (Eds.),
Nagarajan, R., Roy, A., Vinod Kumar, R., Mukhetjess, A., Khire, GIS in Landslide Hazard Zonation: a Review With Exam-
M.V., 2000. Landslide hazard susceptibility mapping based on ples from the Colombian Andes. Taylor and Francis, Lon-
terrain and climatic factors for tropical monsoon regions. don, pp. 135 – 165.
Bulletin of Engineering Geology and the Environment 58 (4), Varnes, D.J., IAEG Commission on Landslides and other Mass-
275 – 287. Movements, 1984. Landslide Hazard Zonation: a Review of
Nathenson, M., 2001. Probabilities of volcanic eruptions and appli- Principles and Practice. NESCO Press, Paris. 63 pp.
cation to the recent history of Medicine Lake Volcano. In: Wieczorek, G.F., 1984. Preparing a detailed landslide-inventory
Vecchia, A.V. (Ed.), U.S. Geological Survey Open-file Report map for hazard evaluation and reduction. Bulletin Association
2001-324, pp. 71 – 74. Engineering Geologists 21 (3), 337 – 342.
Olhmacher, G.C., Davis, J.C., 2003. Using multiple logistic regres- WP/WLI — International Geotechnical societies’ UNESCO Work-
sion and GIS technology to predict landslide hazard in northeast ing Party on World Landslide Inventory, 1990. A suggested
Kansas, USA. Engineering Geology 69, 331 – 343. method for reporting a landslide. International Association Engi-
Önöz, B., Bayazit, M., 2001. Effect of the occurrence process of the neering Geology Bulletin 41, 5 – 12.
peaks over threshold on the flood estimates. Journal of Hydrol- WP/WLI — International Geotechnical societies’ UNESCO Work-
ogy 244, 86 – 96. ing Party on World Landslide Inventory, 1993. A suggested
F. Guzzetti et al. / Geomorphology 72 (2005) 272–299 299

method for describing the activity of a landslide. International Wu, S., Jin, Y., Zhang, Y., Shi, J., Dong, C., Lei, W., Shi, L., Tan,
Association Engineering Geology Bulletin 47, 53 – 57. C., Hu, D., 2004. Investigations and assessment of the landslide
WP/WLI — International Geotechnical societies’ UNESCO Work- hazards of Fengdu county in the reservoir region of the Three
ing Party on World Landslide Inventory, 1995. A suggested Gorges project on the Yangtze River. Environmental Geology
method for describing the rate of movement of a landslide. 45 (4), 560 – 566.
International Association Engineering Geology Bulletin 52, Yevjevich, V., 1972. Probability and Statistics in Hydrology. Water
75 – 78. Resources Publications, Fort Collins, Colorado. 302 pp.

You might also like