Professional Documents
Culture Documents
www.elsevier.com/locate/geomorph
Abstract
We present a landslide susceptibility model for the Collazzone area, central Italy, and we propose a framework for evaluating
the model reliability and prediction skill. The landslide susceptibility model was obtained through discriminant analysis of 46
thematic environmental variables and using the presence of shallow landslides obtained from a multi-temporal inventory map as the
dependent variable for statistical analysis. By comparing the number of correctly and incorrectly classified mapping units, it is
established that the model classifies 77.0% of 894 mapping units correctly. Model fitting performance is investigated by comparing
the proportion of the study area in each probability class with the corresponding proportion of landslide area. We then prepare an
ensemble of 350 landslide susceptibility models using the same landslide and thematic information but different numbers of
mapping units. This ensemble is exploited to investigate the model reliability, including the role of the thematic variables used to
construct the model, and the model sensitivity to changes in the input data. By studying the variation of the model's susceptibility
estimate, the error associated with the susceptibility assessment for each mapping unit is determined. This result is shown on a map
that complements the landslide susceptibility map. Prediction skill of the susceptibility model is then estimated by comparing the
forecast with two recent event inventory maps. The susceptibility model is found capable of predicting the newly triggered
landslides. A general framework for testing a susceptibility model is proposed, including a scheme for ranking the quality of the
susceptibility assessment.
2006 Elsevier B.V. All rights reserved.
Keywords: Landslide susceptibility; Statistically based model; Discriminant analysis; Quality; Uncertainty; Validation; Landslide prediction; Map
1. Introduction
Susceptibility is the propensity of an area to generate
landslides. In mathematical form, landslide susceptibility is the probability of spatial occurrence of known
slope failures, given a set of geoenvironmental conditions (Guzzetti et al., 2005). Assuming landslides will
occur in the future because of the same conditions that
produced them in the past (Guzzetti et al., 1999),
Corresponding author. Tel.: +39 075 5014 413; fax: +39 075 5014
420.
E-mail address: Fausto.Guzzetti@irpi.cnr.it (F. Guzzetti).
0169-555X/$ - see front matter 2006 Elsevier B.V. All rights reserved.
doi:10.1016/j.geomorph.2006.04.007
167
168
Table 1
Descriptive statistics of landslide data sets for the Collazzone study area
Inventory
Multi-temporal landslide
inventory prepared through the
interpretation of five sets of
aerial photographs (Table 2).
Landslides are older than 1941
to December 2004.
Subset of the multi-temporal
inventory showing shallow
landslides and used to prepare
the susceptibility model shown
in Fig. 3. Landslides are older
than 1941 to 1996.
Snowmelt induced landslides
occurred in January 1997
(Fig. 10A).
Rainfall-induced landslides
occurred in autumn 2004
(Fig. 10B).
Type
Number
Area
Total (km2)
Percent (%)
Minimum (m2)
Maximum (m2)
All landslides
Deep-seated landslides
Shallow landslides
2760
363
2397
12.51
7.70
6.53
15.8
9.76
8.28
51
3815
51
173,518
173,518
64,691
Shallow landslides
1759
5.77
7.31
103
43,204
413
7
406
153
1
152
0.78
0.14
0.64
0.38
0.05
0.33
0.98
0.17
0.81
0.48
0.06
0.42
78
10,199
78
51
47,884
51
44,335
44,335
9882
47,884
47,884
12,098
All landslides
Deep-seated landslides
Shallow landslides
All landslides
Deep-seated landslides
Shallow landslides
169
hydrological, lithological, structural and land-use information. We obtained 26 variables describing hydrology
and morphology from the same DTM used to perform
the subdivision of the study area into slope units.
Hydrological variables included slope unit drainage
channel length, gradient, order and magnitude, and slope
unit area and upstream contributing area. Morphological
variables included slope unit mean elevation, standard
deviation of elevation, mean length, mean terrain
gradient and standard deviation of terrain gradient,
slope unit aspect (in six classes), slope unit terrain
roughness, and mean terrain gradient for the upper,
intermediate and lower portions of the slope unit. From
the latter three statistics, derivative variables describing
the shape of the slope unit profile (concave, convex,
irregular, etc.) were obtained. Since most of the
morphological variables describe average terrain conditions in a slope unit, local testing of the variables in the
field was problematic and was not performed. We
compiled lithological and structural data, including the
attitude of bedding, through detailed lithological and
structural mapping at 1:10,000 scale. The lithological
map did not show the distribution and thickness of the
soils or the colluvial deposits. We obtained information
on land use from a map compiled in 1977 by the Umbria
Regional Government, largely revised and updated by
interpreting the most recent aerial photographs (Table 2).
To determine landslide susceptibility we adopted
discriminant analysis, a multivariate technique introduced by Fisher (1936) to classify samples into
alternative groups on the basis of a set of measurements
(Michie et al., 1994; Brown, 1998; SPSS, 2004). More
precisely, the goal of discriminant analysis is to classify
cases into one of several mutually exclusive groups
based on their values for a set of predictor variables. The
grouping variable must be categorical and the predictor
variables can be interval or dichotomous. For landslide
susceptibility assessment most commonly two groups
are established, namely, (i) mapping units free of
landslides (G0, stable slopes); and (ii) mapping units
having landslides (G1, unstable slopes). The assumption
Table 2
Aerial photographs used to prepare the multi-temporal landslide
inventory map (1 to 4) (Fig. 2) and the inventory of snowmelt induced
landslides (5) (Fig. 10A) for the Collazzone area
ID
Year
Period
Type
Nominal scale
1
2
3
4
5
1941
1954
1977
1985
1997
Summer
SpringSummer
June
July
April
Panchromatic
Panchromatic
Colour
Panchromatic
Panchromatic
1:18,000
1:33,000
1:13,000
1:15,000
1:20,000
170
Variable
SDFC
SLO_ANG
ELV_STD
SLO_LEN
ANGLE3
SS
FRA
CONV
0.398
0.370
0.287
0.282
0.276
0.241
0.135
TRAVERTI
TR2
ORDER
ALLUVIO
GHIAIA
ANG_STD
MARNE
CC
CARBO
0.105
0.133
0.140
0.144
0.179
0.219
0.285
0.303
0.833
171
PC PE
lVxV1
1 PC
Table 4
Comparison between slope units classified as stable or unstable by the statistical model (Fig. 3) and slope units free of and containing shallow
landslides in the multi-temporal inventory map shown in Fig. 2
Predicted groups (model)
Group 0
Group 1
239 (66.4%)
121 (33.6%)
360 (100%)
85 (15.9%)
449 (84.1%)
534 (100%)
172
Table 5
Comparison between the proportion of slope units classified as stable or unstable by the susceptibility model (Fig. 3) and the proportion of slope units
free of and containing shallow landslides in the multi-temporal inventory map (Fig. 2)
Predicted groups (model)
Group 0
Group 1
Marginal
totals
0.267 (0.146)
0.135 (0.257)
0.403
0.095 (0.216)
0.502 (0.381)
0.597
0.362
0.638
1.000
Table 6
Statistical indexes measuring the performance of the susceptibility
model shown in Fig. 3
Index
Value
Range
0.664
0.159
0.841
0.262
0.738
0.788
0.770
0.527
0.513
0.107
0.505
0.537
0.825
[0,1]
[0,1]
[0,1]
[0,1]
[0,1]
[0,1]
[0,1]
[0,1]
[ ,1]
[ 1,1]
[ 1,1]
[0,1]
[ 1,1]
173
174
Table 7
Independent thematic variables selected, or not selected, by 50 discriminant functions as the best predictors of landslide occurrence
Variables
SDFC
Susceptibility
models
#
ELV_STD
CARBO
FRA
GHIAIA
MARNE
ANG_STD
SLO_LEN
SLO_ANG
SS
TR2
CC
ORDER
ALLUVIO
CONV
AREN
TRAVERTI
ANGLE3
BOSCO
SLO_ARE
LINK_LEN
R
ANGLE1
ELV_M
CONC
LNK_ANG
CONTI
SABBIA
MAGN
URB
REG
TRA
TR1
LEN_STD
COC_COV
IRR
ARGILLA
SA
VIG
AREAT_K
ANGLE2
COV_COC
RET
FRUTT
PASCOLO
TR3
FRA_OLD
Predictor
0.370
0.833
0.241
0.179
0.285
0.219
0.287
0.398
0.276
0.133
0.303
0.134
0.144
0.135
50
100
I
50
100
S
49
98
I
47
94
S
47
94
S
45
90
S
45
90
I
41
82
I
40
80
I
38
76
S
33
66
S
30
60
S
30
60
S
27
54
I
25
50
S
0.105
23
46
S
0.282
21
42
I
21
42
S
13
26
I
10
20
I
10
20
I
5
10
I
4
8
I
4
8
I
3
6
S
3
6
I
3
6
I
2
4
I/S
2
4
S
2
4
S
2
4
I
2
4
I
1
2
S
1
2
S
1
2
S
1
2
I
1
2
I
1
2
S
Variables were never selected as predictors of landslide
occurrence
175
The adopted approach to ascertain landslide susceptibility provides a unique value for the probability of
176
Fig. 8. Map showing estimated model error (2) for the landslide
susceptibility model shown in Fig. 3. Model error computed using Eq.
(2). Square bracket, class limit included; round bracket, class limit not
included. Larger values indicate increased uncertainty in the
probabilistic estimate of landslide susceptibility.
following quadratic equation, obtained by linear regression fit (least square method):
y 0:309x2 0:308x
0VxV1 r2 0:605
177
178
Fig. 10. Recent landslide event inventory maps. (A) 406 shallow landslides triggered by rapid snowmelt in January 1997. (B) 152 shallow landslides
triggered by heavy rainfall in the period from October to December 2004. Original maps at 1:10,000 scale.
179
180
Level
181
To appraise the fitting performance and the prediction skill of a landslide susceptibility model, Chung and
Fabbri (2003) proposed comparing the proportion of
landslide area in each susceptibility class (AL) with the
proportion of the susceptibility class (AS) in the study
area. For a successful classification, the effectiveness
ratio AL/AS should be greater than one in the areas
predicted as landslide prone by the model, and less than
one in the areas predicted as stable by the model. A very
effective prediction class has a ratio close to zero or very
large, depending whether the class predicts stability or
instability. Where the effectiveness ratio is near one, the
proportion of landslides in the susceptibility class is not
different from the average landslide density in the study
area, and the performance of the susceptibility class in
determining the known (fitting performance) or the
future (prediction skill) location of landslides is weak.
Chung and Fabbri (2003) considered effective a
susceptibility class with a ratio larger than at least 3
(unstable areas) or less than at most 0.2 (stable areas),
and significantly effective a susceptibility class with a
ratio larger than at least 6 or less than at most 0.1. We
regard these criteria as very hard to match, particularly
in complex areas where landslides are large and
numerous, and where the landscape exhibits considerable geomorphological variability. We consider effective a susceptibility class with an effectiveness ratio
larger than 1.5 or smaller than 0.5, corresponding to a
50% increase or a 50% decrease from the expected
proportion of landslides in the susceptibility class.
Fig. 13 shows the efficacy of the susceptibility model
shown in Fig. 3 in describing the known distribution of
landslides (fitting performance, Fig. 13A), and the
location of future landslides (prediction skill, Fig.
13B and C). Based on the established criteria, 12 of the
20 landslides susceptibility classes are effective in
explaining the distribution of the known (past) landslides used to construct the model. In Fig. 13A, the black
and the white bars exceeding the 1.5 and the 0.5
thresholds, respectively, show these effective classes. In
the figure, the three cross-hachured bars represent
terrain units classified as unstable (spatial probability
in the range from 0.55 to 0.70) where landslides were
not abundant in the multi-temporal inventory map.
Comparison of Fig. 13B and C indicates that the
individual susceptibility classes were better predictors of
the presence (black bars) or the absence (white bars) of
the snowmelt-induced landslides than of the rainfallinduced landslides. For the latter, the number of
ineffective classes is also larger.
It should be clear that the proposed acceptance
thresholds are not absolute or fixed. The proposed limits
182
183
184