ASSESSMENT OF LANDSLIDE HAZARD GEOMORPHOLOGICAL AND HISTORICAL DATA IN ASSESSING LANDSLIDE HAZARD

ASSESSMENT OF LANDSLIDE HAZARD 1125

Copyright © 2003 John Wiley & Sons, Ltd. Earth Surf. Process. Landforms 28, 1125–1142 (2003)

Earth Surface Processes and Landforms

Earth Surf. Process. Landforms 28, 1125–1142 (2003)Published online in Wiley InterScience (www.interscience.wiley.com). DOI: 10.1002/esp.545

GEOMORPHOLOGICAL AND HISTORICAL DATA IN ASSESSING

LANDSLIDE HAZARD

ALBERTO CARRARA,1* GIOVANNI CROSTA2 AND PAOLO FRATTINI2

1 CNR-IEIIT, Viale Risorgimento 2, 40136, Bologna, Italy2 Department of Geological Sciences and Geotechnologies, University of Milan-Bicocca,

Piazza della Scienza 4, Milan, Italy

Received 9 April 2002; Revised 27 January 2003; Accepted 24 February 2003

ABSTRACT

Traditionally, earth scientists assess landslide occurrence on the basis of geomorphological investigations carried out throughaerial photograph interpretation and fieldwork. Conversely, local administrators primarily evaluate the impact of naturalcatastrophes, such as landsliding, on the basis of historical records and data. Owing to the substantial difference in thestructure and spatial density of these two types of information, it is difficult to compare them directly and few investigatorshave attempted this.

We compared landslide information derived from geomorphological mapping and historical data in a pilot area (theStaffora river basin, northern Italy). To do this we generated two multivariate statistical models where the dependent variablewas either the mapped landslide deposits (geomorphological model), or the historical sites affected by landslide-induceddamage (historical model). By quantitatively comparing these two model maps, we demonstrate that the geomorphologicalmodel performs better in terms of percentage of terrain units correctly predicted as stable or unstable. The historical modelunderestimates landslide hazard mainly where human structures are lacking. However, it highlights slopes where landslidemovements take place with a high frequency at the temporal scale of human life. Hence, the joint use of these two modelsmay facilitate the knowledge of the overall instability conditions of a given region and the identification of the landslidesthat are most frequently reactivated. Copyright © 2003 John Wiley & Sons, Ltd.

KEY WORDS: historical data; landslide; hazard; statistical model; Apennine; Italy

INTRODUCTION

During the past twenty years, population growth and the expansion of infrastructure into hazardous areas have

increased the impact of natural disasters in both developed and developing countries (Alexander, 1989, 1995;

Rosenfeld, 1994). Among such catastrophes, landslides play a major role causing considerable loss of lives and

properties (Schuster and Fleming, 1986; Glade, 1998; Guzzetti et al., 1999). In Italy, for example, the annual

mortality rate (i.e. the number of casualties per 100 000 people every year) over the last 50 years due to

landslides ranges from 0·18 to 0·14; the yearly mortality rate for the same period due to earthquakes or floods

is 0·16 and 0·04, respectively (Guzzetti, 2000). Hence, the assessment of landslide hazard, the implementation

of warning systems and land planning strategies are dramatic issues for many governments worldwide (IDNHR,

1987; Brabb and Harrod, 1989; UNDRO, 1991; Schuster, 1995).

Among earth scientists, landslide occurrence and the associated hazard are generally evaluated on the basis

of geomorphological investigations carried out through aerial photograph interpretation and fieldwork. Today,

a general consensus exists on the fact that this is the best technique to obtain information on the spatial

distribution of slope failures that have taken place over a long time span (1000 to 10 000 years, depending on

landslide type and geological–geomorphological setting) in a given study area (Varnes and IAEG, 1984; Carrara

et al., 1991; Brunsden, 1993; Hutchinson, 1995). The limitations of the approach refer to the intrinsic subject-

ivity of the operation whose quality and reliability are essentially dependent on the skill and experience of the

* Correspondence to: A. Carrara, CNR-IEIIT, Viale Risorgimento 2, 40136, Bologna, Italy. E-mail: [email protected]

1126 A. CARRARA, G. CROSTA AND P. FRATTINI


investigator or the type and age of the slope failures (Carrara et al., 1992; van Westen et al., 1999; Ardizzone

et al., 2002).

Among many public administrators, natural catastrophes, such as landslides, are primarily evaluated on

the basis of historical records (Regione Lombardia, personal communication). Historical information can

fairly accurately date slope-failure occurrence; it does not require any skill in the earth science discipline.

However, the approach in general allows one to detect not the landslide itself, but the damage that it produced.

In addition, the temporal window of the record seldom spans more than a few decades and rarely as long as a

century. Most importantly, historical data on landslides lack spatial completeness, resolution and precision and

invariably emphasize events that caused damage to human structures, whereas they tend to underestimate

failures, even large ones, which took place in unpopulated areas (Guzzetti et al., 1994; Ibsen and Brunsden,

1996).

Despite the lack of consensus on the reliability and usefulness of historic information, some investigators have

attempted to reconstruct historical records for single landslides or landslide-prone regions (Eisbacher and Clague,

1984; Harty, 1989; Guzzetti et al., 1994; Wieczorek and Jäger, 1996; Ibsen and Brunsden, 1996, Cruden, 1997;

Guzzetti, 2000; Glade, 2001; Calcaterra et al., 2003). According to these authors, the results are encouraging

and appear to be useful for the evaluation of landslide hazard at various scales.

To date, few investigators have tried to compare quantitatively the potential of these two sources of landslide

information or to integrate them into a coherent framework of systematic research. In this investigation we have

attempted this for the Staffora basin of northern Italy, which has a surface area of nearly 280 km2 (Figure 1).

Three landslide inventory maps have been completed for this area. A systematic historical investigation on

landslide and flood events was recently completed by the Italian National Research Council in Turin (CNR-IRPI-

Turin, 1998), which was revised and updated for this study.

In this article, the results obtained from the comparison of the statistical models developed for the Staffora

basin in order to predict either landslide deposits or sites historically affected by landslide-induced damage are

presented and briefly discussed.

GEOLOGICAL AND GEOMORPHOLOGICAL SETTINGS

The Staffora basin is characterized by a complex geological–structural setting resulting from the Alpine

overthrusting of different allochthonous units of the northwestern Apennines (Schumacher and Laubscher, 1996).

This deformation led to the formation of large SW–NE trending synclines subsequently displaced by normal

faulting of Late Tertiary–Quaternary age. In the central part of the valley the Villavermia–Varzi line, an east–

west trending sinistral transfer zone (Schumacher and Laubscher, 1996), occurs whose current activity is witnessed

by several historical seismic events.

The basin is underlain by clay-rich terrigenous sedimentary rocks of Cretaceous–Miocene age. Calcareous

flysch, made up of alternating limestones, marls and minor shale of Eocene–Paleocene age, predominates in the

southern part of the basin (Fontana et al., 1994). Clays, marly clays, shales, minor sandstones and limestones,

Cretaceous to Miocene in age, crop out throughout the area as patches of relatively small size. Most of these

rocks are highly fractured and deeply sheared. Marls, calcareous marls, marly clays and clays with alternating

sandstones of Eocene–Miocene age prevail in the northern part of the basin (Gelati et al., 1974). Well-cemented,

stiff sandstones of Miocene age crop out in the central part of the area. Pliocene sands and conglomerates form

small outcrops in the northernmost portion of the basin. Recent alluvial deposits made up of gravel, sand and

clay occupy the main channel of the Staffora river (Figure 1).

The geomorphological setting of the area is strongly controlled by the local geological and structural con-

ditions. The relative relief is greater and the slopes are generally steeper and more stable in areas underlain by

competent rocks such as limestone (mainly in the southern part of the basin) and sandstone (in the central part

of the area). On the other hand, where marls and clay predominate, the relative relief is lower, and the slopes

are generally shallower. In these areas, landslides are abundant, and the present-day landscape is essentially

controlled by mass-movement (Braga et al., 1985).

Most of these movements are earth flows and translational slides (Figures 2A and 3); large rotational

slides are essentially concentrated in the southern part of the basin (Figure 3). Complex landslides (cf. Varnes,



Figure 1. Staffora basin. Location and geological setting of the study area. Lithostratigraphic units are grouped into 15 units, with the namesgiven in parentheses

1978) mostly occur as rotational/translational slides in the upper part of the slope, with a flow component in

the toe area. A few small rockfalls associated with sandstone outcrops are present in the central part of the

basin.

Most of the active landslides are relatively small flows, whereas dormant and inactive landslides are mainly

slides or complex failures (Figure 2B and 4). A few large landslide bodies are concentrated in the southern part

of the basin; they are mainly inactive or relict rotational and translational slides, but they may be the sites of

small reactivations, especially at their toes.

The area is characterized by an average annual precipitation of 788 mm distributed with strong seasonality,

having maxima in spring and autumn (Rossetti and Ottone, 1979).

Land use in the northern hilly part of the basin mainly consists of cultivated areas with extensive vineyards

on gentler slopes. These areas are characterized by very weak soil protection and by agricultural practices such



Figure 2. Landslides in the Staffora basin: (A) landslide type; (B) landslide state of activity

as milling and hoeing that enhance water infiltration in a way that is not conducive to slope stability. The upper

part of the basin is dominated by forest and pasture.

THEMATIC MAPS AND LANDSLIDE INVENTORY

The data needed for this investigation were derived from aerial photograph interpretation, fieldwork and existing

topographic, geological and land-use maps (Crosta et al., 1999). The large variety of lithological types cropping

out in the study area was grouped into 15 classes according to their compositional characteristics and estimated

or measured mechanical properties (Figure 1). As discussed below, in order to define better the terrain unit to

be used in the statistical analysis, a further grouping of the 15 classes was carried out, obtaining eight major rock

complexes. In addition, structural data (bedding planes, joints, faults, etc.) were obtained in the field and from

existing geological maps.

Land-use data were derived from existing maps, aerial photo interpretation and field verification. Land-use

types were thus grouped into eight classes. A digital terrain model, with a ground resolution of 20 × 20 m, was

produced by interpolating digital contour lines, with a 10 m interval, derived from 1 : 10 000 scale topographic

sheets (Crosta et al., 1999; Antonini et al., 2001). A landslide inventory map for the study area was also completed,

using 1 : 20 000 colour aerial photographs acquired in 1982, and systematic field checks. As a result, 1567

landslides were mapped whose size ranged from nearly 200 m2 to over 1 200 000 m2 with a mean value equal

to 33 900 m2. The total area of landslide deposits is over 53 000 000 m2, that is, nearly 20 per cent of basin area.

A simple form was designed that was then completed in the field or laboratory for each landslide. The form

included: landslide type, degree of activity, relative age, estimated depth and an index reflecting the degree of

certainty in mapping and classifying the slope failure. In addition a code was assigned to identify each of the

sections pertaining to the same landslide, such as: crown area, accumulation area, reactivation area, etc. (Carrara

et al., 1991). These were subsequently digitized as separate polygons. All local morphometric, lithological and

geomorphological features provided by the different thematic maps could be readily assigned to each landslide

after digitization and storage in an ArcGIS database.

LANDSLIDE HISTORICAL DATA

A historical archive of the damage caused by landslide and flood events occurring during the past 150 years

(from 1851 to 1998) was available for the study area (CNR-IRPI-Turin, 1998). This archive, which includes up

to 300 landslide events in nearly 200 sites, was produced using sources of information such as newspapers,

scientific journals and reports.

In order to validate and integrate these data, three other historical archives were examined: the first two were

also produced by CNR-IRPI of Turin (Govi et al., 1979; Tropeano et al., 1999) and the third was compiled by

the Geological Survey of Lombardy Region (Regione Lombardia, personal communication). As a result, 231

sites at which one or more events causing damage to human structures have taken place in the last century



were identified over the study area and stored as point features in the GIS database. A total of 389 historical

records was amassed.

Significant limitations were encountered while using this data set. First the location of the events was

frequently uncertain and sometimes clearly incorrect; moreover, the distribution of the sites is strongly controlled

by the presence of man-made structures liable to be damaged by the landslides. When historical sites are plotted

upon the map displaying roads and villages, over 75 per cent of these sites fall within a 100 m buffer drawn

around these structures (Figure 5). In order to minimize the uncertainty associated with these shortcomings,

Figure 3. Staffora basin. Spatial distribution of landslide types



Figure 4. Staffora basin. Spatial distribution of landslide state of activity

we carried out field checks and interviews with local residents. This enabled us to locate better the position of

some events and to attribute to each site and event a degree of reliability that could be used in the subsequent

analyses.

An analysis of temporal recurrence of landslide events was carried out in order to verify the reliability of

historical records and to understand the temporal distribution of the events. Climatic factors are considered

particularly relevant to the incidence of landslides in this environmental setting. Hence, rainfall data were

collected for the period 1921–2000. The annual frequency of historical records was therefore compared with



Figure 5. Staffora basin. Historical sites located nearer than 100 m to built structures (black dots) and further than 100 m from them(white dots)

rainfall indices such as annual precipitation, maximum daily rainfall and maximum 30-day rainfall. The annual

landslide frequency gave the best correlation with maximum 30-day precipitation (Figure 6).

Since the 1950s, the frequency of landslide events shows a good agreement with rainfall values. The major

events, such as those recorded in 1951, 1959, 1976/7, 1993 and 1997, were associated with widespread slope

Figure 6. Staffora basin. Frequency of historical landslide events plotted along with the rainfall amount of the most rainy 30-day periodfor each year



movements that caused damage to human structures (roads and dwellings). As a response to a period of

relatively low rainfall, a significant decrease in landslide frequency was observed during the 1980s. The correla-

tion between landsliding and heavy rainfall does not hold true for the period before 1950. Most probably this

is due to the smaller availability of information in newspapers that did not have sections of regional and local

information (cf. Guzzetti et al., 1994).

MODEL BUILDING

Owing to the substantial difference in structure (polygon versus point features) and spatial density (1567 land-

slide ‘polygons’ against 231 sites) of the landslide inventory and historical data sets, it is difficult or impossible

to compare them directly. However, the comparison can be performed much better between the maps displaying

the outcomes of the statistical models that predict either landslide deposits or the location of sites where

historical information recorded damage induced by landsliding.

Geomorphological model

By using techniques and tools developed, tested and refined during almost twenty years of investigations

(Carrara, 1983; Carrara et al., 1995, 1999; Guzzetti et al., 1999), a predictive model of landslide occurrence was

developed. To accomplish this task, the basin area was automatically partitioned into main slope-units (i.e. the

left and right sides of elementary sub-basins) using a specifically designed software module. Starting from a

DTM of high accuracy, this generates fully connected and complementary drainage and divide networks, and

a wide spectrum of morphometric parameters of channels and slopes (Carrara et al., 1995, Guzzetti et al., 1999).

Using this module, 1429 main slope-units were generated. These units were subsequently subdivided according

to the eight classes of the main rock complexes cropping out in the basin. Hence, the study basin was partitioned

into 2245 morphological–lithological terrain units whose mean size (85 580 m2) was evaluated to be suitable for

spatially analysing the landslides that have a mean size of 33 900 m2.

Under the assumption that both the mapping errors and the uncertainty decrease with the size of the terrain

unit, all the discriminant analyses were weighted by the square root of the terrain unit area. Likewise, the area

of the landslide deposits was weighted according to their estimated degree of activity and the degree of certainty

associated with their identification and mapping.

Using a stepwise procedure, 40 geological–morphological factors were selected as predictors, and the presence

or absence of landslide deposits within each terrain unit was used as the predicted or dependent variable of a

discriminant function (Table I). Since the discriminant function assumes a negative (−0·75) value at the centroid

of the groups of unstable terrain units and a positive (0·67) one at that of the stable units, predictors with

negative and positive coefficients are in agreement and disagreement with landslide occurrence, respectively.

Among the lithological variables, the factors with the largest standardized discriminant function coefficients

were the presence within the terrain unit of alluvial deposits (ALLUVIO), of massive sandstones (AR_BIS) and

calcareous marls (MR_AN_LO). Of the land use and land cover variables, the most relevant were the presence

of tilled fields (SEM), of denuded surfaces (INC) and pastures (PRA) and river-beds and rocky cliffs (ALV).

Among the morphometric variables, terrain unit mean slope angle (SLO_ANG) and its square (SLO_ANG2)

have large coefficients that indicate a curvilinear relationship between steepness and landslide frequency.

That is, landsliding first increases with slope angle up to a threshold value above which the relation does not

hold true any more (Carrara et al., 1995). Terrain unit area (SLO_AR) and local relief (ELV_STD) have also

large coefficients.

Outcomes of the analysis indicate that such a mix of environmental factors is capable of predicting, with a

reliability of the 77 per cent, which terrain units could either be affected by or be free of landslides (Figure 7,

Table I).

This geomorphological model is essentially founded upon the data provided by the landslide inventory map

and other environmental maps. As it attempts to predict the whole set of landslides that took place over a very

large time span (say, 5000–10 000 years), a large discrepancy may be expected to occur with respect to the data

derived from the historical archive covering less that 100 years. Therefore, a second geomorphological model

was constructed (Table II) using as predictors the same group of environmental factors but selecting, as predicted



Table I. Geomorphological model. (a) List of the 40 variables selected by stepwise discriminant analysis as the bestpredictors of the occurrence in the terrain units of landslide deposits derived from the geomorphological inventory. The mostimportant standardized discriminant function coefficients (SDFC) are shown in bold. Discriminant scores at group centroidsare: stable terrain units = 0·67; unstable terrain units = −0·75. (b) Classification of stable and unstable terrain units using the

landslide deposits as the predicted variable(a)

Variable Description SDFC

AR_VA_PA % of tectonic clayey melange in the terrain-unit 0·113ALB_ZEB % of clays, marls and limestones in the terrain-unit 0·080ALLUVIO % of recent alluvium deposits in the terrain-unit 0·768

AR_BIS % of massive sandstones in the terrain-unit 0·314

AR_R_M_P % of sandstones and marls in the terrain-unit 0·047AT_PA_CA % of marly clayey chaotic complex in the terrain-unit 0·194CA_PEN % of marls and limestones in the terrain-unit −0·160DETRITO % of detritic deposits in the terrain-unit −0·096MR_AN_LO % of calcareous marls in the terrain-unit 0·391

MR_B_R_C % of marls and sandstones in the terrain-unit 0·198MR_BOSM % of marls in the terrain-unit 0·030MR_P_R_B % of clayey marls in the terrain-unit 0·127SACONG % of sands and conglomerates in the terrain-unit 0·186INSIDE % of area fractured by faulting in the terrain-unit 0·030REG % of beds dipping toward the slope free face in the terrain-unit 0·122FRA % of beds dipping away the slope free face in the terrain-unit 0·100CAO % of chaotic bedding in the terrain-unit −0·106TR3 Terrain-unit facing SW 0·158ALV % of river-bed and rocky cliff area in the terrain-unit −−−−−0·280

BD % of densely forested area in the terrain-unit 0·074INC % of denudated area in the terrain-unit −−−−−0·269

PRA % of pasture area in the terrain-unit −−−−−0·300

RIM % of reforested area in the terrain-unit −0·031SEM % of plown area in the terrain-unit −−−−−0·557

URB % of urbanised area in the terrain-unit −0·101VIG % of vineyard area in the terrain-unit −0·092MAGN Channel magnitude 0·167LINK_LEN Channel length 0·102SLO_ARE Terrain-unit area −−−−−0·259

R I° index of terrain-unit micro-relief 0·160ELV_STD Standard deviation of terrain-unit mean elevation −−−−−0·264

SLO_ANG Mean slope angle −−−−−0·310

SLO_ANG2 Mean slope angle squared 0·480

ANG_STD Standard deviation of slope angle 0·177LINK_ANG Mean channel angle 0·070LEN_STD Standard deviation of channel angle −0·073CONV Convex slope profile −0·033COC_COV Concave-convex slope profile 0·051RET Rectilinear slope profile 0·042IRR II° index of terrain-unit micro-relief 0·076

(b)

Actual groups Predicted group membership

Group 1 (stable terrain units) Group 2 (unstable terrain units)

Group 1 (stable terrain units) 73·9% 26·1%

Group 2 (unstable terrain units) 19·5% 80·5%

Terrain units correctly classified: 77·0%



Figure 7. Staffora basin. Multivariate model of landslide occurrence based on a discriminant function in which 40 geological–morphologicalfactors were selected as predictors, and the dependent variable was the presence or absence of landslide deposits (solid black) withineach terrain unit. Light grey indicates terrain units with a probability (P) of landslide occurrence of less than 0·5. Conversely, dark grey

indicates P > 0·5

variable, the subset of landslides that were mapped and classified as active (665 failures out the whole data set

of 1567). Notice that the signs of the group centroids are inverted with respect to those of Table I (stable terrain

units = −0·39; unstable terrain units = 0·93).

Among the lithological variables, the factors with the largest standardized discriminant function coefficients

were the presence within the terrain unit of arenaceous flysch (AR_SCA), of alluvial deposits (ALLUVIO) and

clay, marls and minor limestone (ALB_ZEB). Out of the land use and land cover variables, the most relevant

were the presence of ploughed fields (SEM), denuded areas (INC), pastures (PRA), and river-beds and rocky

cliffs (ALV). Among the morphometric variables, terrain unit mean slope angle (SLO_ANG) and its square

(SLO_ANG2) have the largest coefficients. Terrain unit area (SLO_AR) and local slope angle variability

(ANG_STD) have also large coefficients.

The results of this analysis indicate that the group of environmental factors capable of predicting the subset

of active landslides is rather similar to that entered in the previous model (Table I). The main difference between

the models refers to the lower predictive power of the second (73·6 per cent, Table II) as compared to the first

(77·0 per cent, Table I). Owing to the fact that active landslides are mainly reactivations of portions of large



Table II. Geomorphological model of active landslides. (a) List of the 34 variables selected by stepwise discriminant analysisas the best predictors of the occurrence in the terrain units of active landslide deposits derived from the geomorphologicalinventory. The most important standardized discriminant function coefficients (SDFC) are shown in bold. Discriminantscores at group centroids are: stable terrain units = −0·39; unstable terrain units = 0·93 (notice that the centroid signs areinverted with respect to those of Table I). (b) Classification of stable and unstable terrain units using the landslide deposits

as the predicted variable(a)


ALB_ZEB % of clays, marls and limestones in the terrain-unit 0·238

ALLUVIO % of recent alluvium deposits in the terrain-unit −−−−−0·333

AR_R_M_P % of sandstones and marls in the terrain-unit 0·141AR_SCA % of arenaceous flysch in the terrain unit 0·367

CA_AN_CA % of marly clayey chaotic complex in the terrain-unit 0·119CA_PEN % of marls and limestones in the terrain-unit 0·197DETRITO % of detritic deposits in the terrain-unit 0·049MR_BOSM % of marls in the terrain-unit 0·045INSIDE % of area fractured by faulting in the terrain-unit 0·034FRA % of beds dipping away the slope free face in the terrain-unit 0·109CAO % of chaotic bedding in the terrain-unit 0·073TR3 Terrain-unit facing SW −0·047ALV % of river-bed and rocky cliff area in the terrain-unit 0·247

BMD % of fairly densely forested area in the terrain-unit −0·057INC % of denudated area in the terrain-unit 0·439

PRA % of pasture area in the terrain-unit 0·426

RIM % of reforested area in the terrain-unit −0·067SEM % of plough area in the terrain-unit 0·561

URB % of urbanised area in the terrain-unit 0·079VIG % of vineyard area in the terrain-unit 0·068MAGN Channel magnitude −0·073LINK_LEN Channel length −0·144SLO_ARE Terrain-unit area 0·398

R I° index of terrain-unit micro-relief −0·166ELV_STD Standard deviation of terrain-unit mean elevation 0·152SLO_ANG Mean slope angle 0·519

SLO_ANG2 Mean slope angle squared −−−−−0·513

ANG_STD Standard deviation of slope angle −−−−−0·244

LNK_ANG Mean channel angle −0·123SLO_LEN Slope length −0·082LEN_STD Standard deviation of slope length 0·179CONC Concave slope profile 0·093CONV Convex slope profile 0·071COV_COC Concave-convex slope profile 0·150

(b)








old dormant landslide deposits, their spatial distribution is rather similar (Figure 4). Hence, It is not surprising

that the two models are fairly similar.

Historical model

The historical model was generated using the same set of predictors and the presence or absence of historical

sites within each terrain unit as predicted variable of the discriminant function (Table III, Figure 8). Because

of the previously discussed uncertainty related to the location of the sites, a circular buffer of 50 m radius was

generated around each site point. This value was selected because it is close to the minimum distance between

different sites.

Sites located near the boundary of two or more terrain units were assigned to each unit. In addition, each site

was weighted according to the degree of its locational reliability. Through a stepwise procedure, 25 predictors

Table III. Historical model. (a) List of the 25 variables selected by stepwise discriminant analysis as the best predictors ofthe occurrence in the terrain units of historical sites affected by landslide damage. The most important standardized discri-minant function coefficients (SDFC) are shown in bold. Discriminant scores at the group centroid are: stable terrain unit =−0·14; unstable terrain units = 0·75. (b) Classification of stable and unstable terrain units by discriminant analysis using the

historical sites affected by landslide damage as the predicted variable(a)


AG_VA_PA % of tectonic clayey mélange in the terrain-unit −0·159ALLUVIO % of recent alluvium deposits in the terrain-unit −−−−− 0·606

AR_SCA % of clayey marls and sanstones in the terrain-unit 0·112CA_PEN % of marls and limestones in the terrain-unit 0·251

MR_P_R_B % of clayey marls in the terrain-unit 0·076SACONG % of sands and conglomerates in the terrain-unit −0·097INSIDE % of area fractured by faulting in the terrain-unit 0·152TR2 Terrain-unit facing SE 0·098ALV % of river-bed and rocky cliff area in the terrain-unit 0·218

BMD % of forest area in the terrain-unit −0·068INC % of denudated area in the terrain-unit 0·206

PRA % of pasture area in the terrain-unit 0·165RIM % of reforested area in the terrain-unit −0·119SEM % of plown area in the terrain-unit 0·771

URB % of urbanised area in the terrain-unit 0·520

VIG % of vineyard area in the terrain-unit 0·380

SLO_ARE Mean slope length in the terrain-unit 0·084ELV_STD Standard deviation of terrain-unit mean elevation 0·518

SLO_ANG Mean slope angle 0·121LNK_ANG Mean channel angle −−−−−0·423

SLO_LEN Slope length 0·194LEN_STD Standard deviation of channel angle −0·141CONV Convex slope profile 0.166RET Rectilinear slope profile −0·079IRR II° index of terrain-unit micro-relief 0·076

(b)








Figure 8. Staffora basin. Multivariate model of sites historically affected by landsliding based on a discriminant function in which 25geological–morphological factors were selected as predictors, and the predicted variable was the presence or absence of sites (solid blackdots) within each terrain unit. Light grey indicates terrain units with a probability of landslide occurrence of less than 0·5. Conversely, dark

grey indicates P > 0·5

were entered into the discriminant function, of which the most important are the following. Among the lithological

variables, the factors with the highest coefficient are the presence of alluvial deposits (ALLUVIO) and the

presence of marls and limestones (CA_PEN). Many variables related to land-use were entered into the discri-

minant function with large coefficients, namely, the presence of denudated lands (INC), river-beds and rocky

cliffs (ALV), ploughed areas (SEM), urbanized areas (URB) and vineyards (VIG). Lastly, among morphometric

variables, only two factors show large coefficients, the mean channel angle (LNK_ANG) and the local relief

(ELV_STD). The model with this selection of predictors yielded 67 per cent of correctly classified terrain units.

Model comparison

By overaying the maps of Figures 7 and 8, it is easily possible to make a quantitative comparison of the

models (Table IV, Figure 9). A large proportion of terrain units predicted as stable by the geomorphological

model is also classified as stable by the historical model (43 per cent); a smaller degree of agreement exists

between models for the unstable terrain units (29·2 per cent). Most importantly, almost 20 per cent of the

terrain units are predicted as unstable only by the geomorphological model, whereas the opposite occurs in a



much smaller number of cases (8 per cent). To summarize, the level of mismatch is nearly 30 per cent with

a relevant number of terrain units classified as unstable and stable by the geomorphological and historical

models, respectively.

When the comparison is confined to the historical model and the model derived from the subset of active

landslides set, outcomes are very similar: the percentage of agreement increases to 48·4 per cent for stable terrain

Table IV. Degree of agreement between the geomorphological (G_) and historical (H_) models(Figures 5 and 6, respectively)

Terrain-units classified by Terrain-units classified by Per cent ofthe geomorphological mode as the historical model as agreement

G_stable H_stable 43·0G_unstable H_unstable 29·2G_stable H_unstable 8.0G_unstable H_stable 19·8

Figure 9. Staffora basin. Comparison of geomorphological (G_) and historical (H_) models



units and decreases to 23·4 per cent for unstable ones. The overall disagreement between models is again close

to 30 per cent (Table V).

DISCUSSION AND CONCLUDING REMARKS

The geomorphological model shows a fairly high percentage of correctly classified terrain units (77·0 per cent)

obtained through a stepwise procedure using the 40 predictors selected from more than 50 input variables

(Table I). The predictive power of the model is better witnessed by the high percentage (80·5 per cent) of terrain

units classified as unstable on which geomorphologists detected landslide deposits. The historical model shows

a lower percentage of correctly classified terrain units (66·6 per cent) obtained using 25 predictors selected

from the same set of input variables (Table III). Also the percentage (69·6 per cent) of terrain units classified

as unstable on which historical sites occur is relatively low.

However, the different success of the two discriminant functions would be better evaluated by taking into

consideration the fact that in the two discriminant functions, classification was accomplished using a con-

servative a priori probability equal to 0·5; that is, each terrain unit was assigned the same a priori probability

of belonging to either stable or unstable groups. In the geomorphological model the two groups are almost equal

in size (1185 versus 1060 terrain units), while in the historical model they are very unequal (1900 versus 345

terrain units). Hence, the success of the first model is necessarily greater than that of the second.

In the geomorphological model the set of predictors of landslide deposits includes a mix of lithological,

morphological and land-use characteristics. Among the variables with high predictive power (say, SDFC greater

than 0·2; Table I), some have a clear and simple physical meaning. For example, the alluvial deposits (ALLUVIO),

which are a lithological class virtually landslide free, have the highest standardized coefficient in the function

(SDFC = 0·768; Table I), and are the best predictor of stable terrain units. Likewise, ploughed fields (SEM),

which are the land-use class most affected by seasonal surficial slope movements, rank second for the ability

to predict unstable slopes (SDFC = −0·557; Table I). Other variables exhibit large coefficients but do not easily

reveal their physical meaning. An example is given by the standard deviation of the terrain unit elevation

(ELV_STD) that has a SDFC equal to −0·264. Most probably, the variable which expresses the terrain unit local

relief is well correlated with both slope steepness and slope length.

In general, the predictors entered in the discriminant function agree well with current knowledge of the causes

of slope failure in the central section of the clay-rich Apennine mountain range (Carrara et al., 1991; Guzzetti

et al., 1996). In this environment, a strong curvilinear correlation (SLO_ANG, SLO_ANG2, Table I) exists

between slope angle and landslide occurrence (Carrara et al., 1995). Likewise, the relations between bedding

attitude and slope aspect (REG, FRA, CAO; Table I) are always significant (Guzzetti et al., 1999).

In the historical model, such relations are poorly expressed because relevant variables either were not entered

into the function (i.e. SLO_ANG2, REG, FRA) or have small coefficients (i.e. SLO_ANG). Despite these

discrepancies, of the 25 predictors entered in the historical model, 21 are common to the geomorphological one.

Hence, the two discriminant functions exhibit an overall similarity that it is greater than expected in light of the

large difference in the properties of the respective dependent variables (very many landslide polygons versus

very few site points).

Table V. Degree of agreement between the geomorphological model of active landslides (GA)and historical model (H)

Terrain units classified by Terrain units classified by Per cent ofthe geomorphological model the historical model as agreementof active landslides as

GA_stable H_stable 48·4GA_unstable H_unstable 23·4GA_stable H_unstable 12.8GA_unstable H_stable 15·2



The comparison of the maps (Figures 5 and 6) displaying outcomes of the two models indicates that in

the Staffora basin historical information underestimates landslide hazard. The total mismatch between the

two models is nearly 30 per cent (G_stable versus H_unstable equal 8 per cent and G_unstable versus H_stable

19·8 per cent; Table IV, Figure 9). The underestimation is greater in areas that are sparsely populated and

lacking in built structures (i.e. roads and dwellings). This is also confirmed by the fact that over 75 per cent

of the sites fall within a 100 m buffer drawn around built structures (Figure 5). All this confirms one of the

well-known intrinsic limitations of historical data, namely, that landslide events which are not associated

with damage or casualities are rarely reported. In addition, the historical data cannot report landslides

that are inactive for long periods of time. This is the case in the upper part of the basin where large dormant

or inactive, rotational or translational slides occur. Since most of these landslides did not reactivate in the

last century, their record is totally lacking in the historical database. Hence, the time window of the historical

data (generally 100 years or less) cannot be readily compared with the time window of the inventory maps

(say, over 1000 years). On the other hand, in the study area the proportion of slope units affected by landslid-

ing is very high, namely, over 47 per cent. Hence, the probability that historical sites fall, by chance, in an

unstable unit is not negligible. Despite all these factors that have influenced the outcomes of the analyses,

the capability of the historical archive correctly to predict almost 70 per cent of unstable terrain units is

remarkable.

In the next step of the investigation, the historical model was compared with a new geomorphological model

based only on landslides classified as active in the inventory. In this case, the dependent variables of the two

discriminant functions should be temporally and spatially comparable: the records of damage due to landslides

that occurred in the last century, and landslides that took place or reactivated in recent times (say, 10–50 years).

This is confirmed by the fact that the mismatch is nearly equal for GA_stable versus H_unstable (12·8 per cent)

and GA_unstable versus H_stable (15·2 per cent) classes (Table V). In addition to the inherent errors related to

historical data sets, the reliability of the landslide inventories is frequently unknown. At present, three landslide

maps are available for the Staffora basin. The first was produced by the authors; the second by the CNR-IRPI

in Perugia; the third by the University of Pavia. The comparison of these three landslide inventories showed that

the positional mismatch between the maps amounts to nearly 60 per cent (Ardizzone et al., 2002). The outcome

confirms that landslide inventories produced by different investigators invariably differ in terms of landslide

spatial distribution, classification and estimated degree of activity (Carrara et al., 1992).

Notwithstanding, when landslide deposits are aggregated into morphologically meaningful terrain units, such

as the main slope units used in this investigation, map mismatch reduces to values in the range of 20–25 per

cent. Furthermore, the comparison of terrain units, classified as stable and unstable by the discriminant models

applied to the landslide deposits derived from the inventories produced by the teams of the University of Milano

and the CNR-IRPI in Perugia, indicates that map disagreement reduces to less than 16 per cent (Ardizzone

et al., 2002), which is close to the errors that affect many of the measurements of data from the environment.

Hence, the loss of spatial resolution introduced by aggregating all data into physically based terrain units, is well

compensated by the gain in agreement among different sources of information.

In conclusion, the mismatch between the geomorphological and historical models should be evaluated and

interpreted in the light of the many and different sources of error and uncertainty affecting both types of data.

Future investigation should attempt to increase the quality, completeness and reliability of both landslide

inventories and historical archives. By integrating these two types of data, it would be possible to improve

significantly landslide hazard assessment in the framework of a feasible environmental management.

ACKNOWLEDGEMENTS

The authors are grateful to David Alexander, University of Massachusetts at Amherst, for having critically read

the manuscript; M. Luino, CNR-IRPI Turin, for having made available the historical data regarding the landslide

events in the Staffora basin; Mauro Cardinali, Fausto Guzzetti and Paola Reichenbach, CNR-IRPI Perugia, for

useful suggestions. The research was supported by funds from the Gruppo Nazionale per la Difesa dall Catastrofi

Idrogeologiche (GNDCI) of the National Research Council (CNR) and the Geological Survey of the Regione

Lombardia.



REFERENCES

Alexander DE. 1989. Urban landslides. Progress in Physical Geography 13: 157–191.Alexander DE. 1995. A survey of the field of natural hazards and disaster studies. In Geographical Information Systems in Assessing Natural

Hazards, Carrara A, Guzzetti F (eds). Kluwer: Dordrecht; 1–19.Antonini G, Ardizzone F, Cardinali M, Carrara A, Detti R, Galli M, Guazzetti F, Reichenbach P, Sotera M, Tonelli G. 2001. Tecniche e

Metodologie idonee alla produzione di Carte della Pericolosità e del Rischio da Frana in aree campione rappresentative del territorio

della Regione Lombardia. Final Report. Regione Lombardia, CNR-IRPI-Perugia, CNR-CSITE-Bologna: Perugia [in Italian].Ardizzone F, Cardinali M, Carrara A, Guzzetti F, Reichenbach P. 2002. Impact of mapping errors on the reliability of landslide

hazard maps. Natural Hazards and Earth System Sciences 2: 3–14.Brabb EE, Harrod BL (eds). 1989. Landslides: Extent and Economic Significance. Balkema: Rotterdam.Braga G, Braschi G, Calculli S, Caucia F, Cerro A, Colleselli F, Grisolia M, Piccio A, Rossetti R, Setti M, Spalatro A, Soggetti F,

Veniale F. 1985. I fenomeni franosi nell’Oltrepo Pavese: tipologia e cause. Geologia Applicata e Idrogeologia 20(2): 621–666 [in Italian].Brunsden D. 1993. Mass movements; the research frontier and beyond: a geomorphological approach. Geomorphology 7: 85–128.Calcaterra D, Parise M, Palma B. 2003. Combining historical and geological data for the assessment of the landslide hazard: a case study

from Campania, Italy. Natural Hazards and Earth System Sciences 3: 3–16.Carrara A. 1983. Multivariate models for landslide hazard evolution. Mathematical Geology 15: 403–427.Carrara A, Cardinali M, Detti R, Guzzetti F, Pasqui V, Reichenbach P. 1991. GIS techniques and statistical models in evaluating landslide

hazard. Earth Surface Processes and Landforms 16: 427–445.Carrara A, Cardinali M, Guzzetti F. 1992. Uncertainly in assessing hazard and risk. ITC Journal 2: 172–183.Carrara A, Cardinali M, Guzzetti F, Reichenbach P. 1995. GIS technology in mapping landslide hazard. In Geographical Information

Systems in Assessing Natural Hazards, Carrara A, Guzzetti F (eds). Kluwer: Dordrecht; 135–176.Carrara A, Cardinali M, Guzzetti F, Reichenbach P. 1999. Use of GIS technology in the prediction and monitoring of landslide hazard. In

Techniques and Tools for Assessing and Mapping Natural Hazards, Carrara A, Guzzetti F (eds). Natural Hazards 20: 117–135.CNR-IRPI-Turin. 1998. Determinazione delle fasce fluviali di alcuni corsi d’acqua lombardi ai fini urbanistici: Torrente Staffora. Final

report. CNR-IRPI: Torino [in Italian].Crosta GB, Frattini P, Sterlacchini S. 1999. Procedura di raccolta ed elaborazione dei dati. Final Report. Regione Lombardia – Universita’

Milano Bicocca: Milano [in Italian].Cruden DM. 1997. Estimating the risks from landslides using historical data. In Landslide Risk Assessment; Proceedings of the International

Workshop, Cruden DM, Fell R (eds). Balkema: Rotterdam; 177–184.Eisbacher GH, Clague JJ. 1984. Destructive mass movements in high mountains: hazard and management. Paper 84 (16). Geological Survey

Canada.Fontana D, Spadafora E, Stefani C, Stocchi S, Tateo F, Villa G, Zuffa GG. 1994. The Upper Cretaceous Helminthoid Flysch of the Northern

Apennines; provenance and sedimentation. Memorie della Società Geologica Italiana 48(1): 237–250.Gelati R, Bruzzi D, Catasta G, Cattaneo PC. 1974. Evoluzione stratigrafico-strutturale dell’Appennino Vogherese a nord-est della Val

Staffora. Rivista Italiana di Paleontologia e Stratigrafia 80(3): 479–514 [in Italian].Glade T. 1998. Establishing the magnitude and frequency of landslide triggering rainstorm events in New Zealand. Environmental Geology

35: 2–3.Glade T. 2001. Landslide hazard assessment and historical landslide data – an inseparable couple? In The Use of Historical Data in Natural

Hazard Assessments, Glade T, Albini P, Francés F (eds). Kluwer: Dordrecht.Govi M, Mortasa G, Soriana P, Tropeano D. 1979. Sintesi dei dissesti idrogeologici avvenuti tra il 1972 e il 1974 nell’Italia settentrionale.

Bollettino dell’Associazione Mineraria Subalpina 83(2): 329–368 [in Italian].Guzzetti F. 2000. Landslide fatalities and evaluation of landslide risk in Italy. Environmental Geology 58: 89–107.Guzzetti F, Cardinali M, Reichenbach P. 1994. The AVI Project: a bibliographic and archive inventory of landslides and floods in Italy.

Environonmental Management 18: 623–633.Guzzetti F, Cardinali M, Reichenbach P. 1996. The influence of structural setting and lithology on landslide type and pattern. Journal for

the Association of Engineering Geologists 2(4): 531–555.Guzzetti F, Carrara A, Cardinali M, Reichenbach P. 1999. Landslide hazard evaluation: a review of current techniques and their application

in a multi-scale study, Central Italy. Geomorphology 31: 181–216.Harty KM. 1989. Landslide mapping, hazards, and historical landslides in Utah. Survey Notes. Utah Geological and Mineral Survey 23(4):

2–8.Hutchinson, JN 1995. Keynote paper: landslide hazard assessment. In Landslides, Bell DH (ed.). Balkema: Rotterdam; 1805–1841.Ibsen ML, Brundsen D. 1996. The nature, use and problems of historical archives for the temporal occurrence of landslides, with specific

reference of the south coast of Britain, Ventnor, Isle of Wight. Geomorphology 15: 241–258.IDNHR. 1987. Confronting Natural Disasters. An International Decade for Natural Hazard Reduction. National Academy Press:

Washington DC.Rosenfeld CL. 1994. The geomorphological dimensions of natural disasters. Geomorphology 10: 27–36.Rossetti R, Ottone C. 1979. Esame preliminare delle condizioni pluviometriche dell’Oltrepo Pavese e dei valori critici delle precipitazioni

in relazione ai fenomeni di dissesto franoso. Geologia Applicata e Idrogeologia 14(3): 83–99 [in Italian].Schumacher ME, Laubscher HP. 1996. 3D crustal architecture of the Alps-Apennines join; a new view on seismic data. Tectonophysics

260(4): 349–363.Schuster DL. 1995. Socioeconomic significance of landslides. In Landslide Analysis and Control, Schuster RL, Krizek RJ (eds). National

Academy Sciences: Washington DC; 12–35.Schuster RL, Fleming RW. 1986. Economic losses and fatalities due to landslides. Bulletin American Association Engineering Geologists

23(1): 11–28.Tropeano D, Govi M, Mortasa G, Turitto O, Soriana P, Negrini G, Arattano M. 1999. Eventi alluvionali e frane nell’Italia settentrionale

(Periodo 1975–1981). CNR-GNDCI: Torino [in Italian].UNDRO. 1991. Mitigating Natural Disasters. Phenomena, Effects and Options. United Nations: New York.



Van Westen CJ, Seijmonsbergen AC, Mantovani F. 1999. Comparing landslide hazard maps. In Techniques and Tools for Assessing and

Mapping Natural Hazards, Carrara A, Guzzetti F (eds). Natural Hazards 20: 137–158.Varnes DJ. 1978. Slope movements, types and processes. In Landslide Analysis and Control, Schuster RL, Krizek RJ (eds). National

Academy Sciences: Washington DC; 11–33.Varnes DJ, IAEG (Commission on Landslides and other Mass-Movements). 1984. Landslide Hazard Zonation: a Review of Principles and

Practice. UNESCO Press: New York.Wieczorek GF, Jäger S. 1996. Triggering mechanisms and depositional rates of postglacial slope-movements processes in the Yosemite

Valley, California. Geomorphology 15: 17–31.

Date post:	29-Nov-2023
Category:	Documents
Upload:	unimib
View:	0 times
Download:	0 times

ASSESSMENT OF LANDSLIDE HAZARD GEOMORPHOLOGICAL AND HISTORICAL DATA IN ASSESSING LANDSLIDE HAZARD

Documents