1. NH 3.8, Vienna, 02-05-2014, Ivan Marchesini
Non-Susceptible Landslide Areas in
Italy and in the Mediterranean Region
Massimiliano Alvioli1
, Francesca Ardizzone1
,
Fausto Guzzetti1
, Ivan Marchesini1
, and Mauro
Rossi1,2
1) Consiglio Nazionale delle Ricerche, Istituto di Ricerca per la Protezione Idrogeologica, via Madonna Alta 126,
I-06128 Perugia, Italy
2) Universita` degli Studi di Perugia, Dipartimento di Scienze della Terra, Piazza Universita`, 1, I-06123, Perugia,
Italy
NH 3.8, Vienna, 02-05-2014
2. NH 3.8, Vienna, 02-05-2014, Ivan Marchesini
Why this work?
● A large number of methods and techniques were
proposed and tested to ascertain landslide
susceptibility
● A few attempts were made to define landslide
susceptibility at the continental and even at the global
scale (e.g. Van Den Eeckhaut et al., 2012; Gunther et
al., 2013)
● Little effort was made to define where landslides are
not expected, i.e. where landslide susceptibility is null,
or negligible (Godt et al. 2012)
3. NH 3.8, Vienna, 02-05-2014, Ivan Marchesini
What we did
● In this work, we discuss some methods for the definition
of non-susceptible landslide areas, at the synoptic scale.
● We apply the best method in Italy and to the
landmasses surrounding the Mediterranean Sea
10° W 40° E
10° W 40° E
50° N
30° N
50° N
30° N
4. NH 3.8, Vienna, 02-05-2014, Ivan Marchesini
Data
● Digital terrain elevation:
– 40 SRTM data tiles
covering the
Mediterranean area
● Landslide information:
– 13 inventories of
polygons including
geomorphological, event,
and multi-temporal
inventory maps in Italy
5. NH 3.8, Vienna, 02-05-2014, Ivan Marchesini
SRTM
● We exploited two morphometric parameters computed from the SRTM
DEM:
– relative relief R (in meters)
– terrain slope S (in degrees)
● We computed
– R using a circular moving window with a diameter of 15 cells
– S in a 3 × 3 - cell square moving window.
6. NH 3.8, Vienna, 02-05-2014, Ivan Marchesini
Landslide inventories
● The maps cover 8.9% of the
Italian territory
● 93,538 landslides
● Mapped area: 2726 km2
● Landslide area is 10.1% of
the mapped areas
– Rotational and translational
slides,
– Earth flows
– Complex and compound
movements
7. NH 3.8, Vienna, 02-05-2014, Ivan Marchesini
Methods
● We defined the areas that are expected to be
non susceptible to landslides in Italy, using two
different methods:
1.The first method is derived from the work of Godt et
al. (2012) (method I)
2.The second method was developed specifically for
this work (method II)
8. NH 3.8, Vienna, 02-05-2014, Ivan Marchesini
Method I
● We computed the frequency distribution of the relative relief R and of the
terrain slope S for all the grid cells in each single landslide in each inventory.
0 10 20 30 40 50 60
Local Terrain Slope, S [°]
EmpiricalCumulativeProbability
0.20.40.60.81.00.0
0 20 40 60
Local Terrain Slope, S [°]
0.20.40.60.81.00.0
● For each inventory, we prepared the
Empirical Cumulative Distribution
Functions (ECDFs) for the 50th
percentile
of the two terrain variables, R and S, in all
the mapped landslides.
● Next we arbitrary chose the
5% cumulative frequency of
both slope and relief of the
ECDFs of the different
inventories
0.050.05
9. NH 3.8, Vienna, 02-05-2014, Ivan Marchesini
Method I
● Plot of the 5th
percentile pairs
(R50, S50).
● Data fitting with a
Linear Regression
model (LR):
Non
susceptibleSusceptible
S50 = 3.448 + 0.040 R50
A-M: landslide inventories
10. NH 3.8, Vienna, 02-05-2014, Ivan Marchesini
Method I
● We used the Linear Regression model (LR) to prepare the binary
zonation of the Italian territory.
● The orange color shows areas where landslide susceptibility is
expected to be null or negligible.
11. NH 3.8, Vienna, 02-05-2014, Ivan Marchesini
Method II
● 354,406 (R,S) pairs of
slope and relief values
● Corresponding to all
the cells inside the
landslide polygons
● We searched for a
lower threshold to the
cloud of points.
12. NH 3.8, Vienna, 02-05-2014, Ivan Marchesini
Method II
● Quantile regression
● We tested
– A Quantile Linear Regression
model (QLR)
– A Quantile Non-Linear regression
(exponential) model (QNL)
● We instructed the quantile
regression to model the 5th
percentiles i.e., to search for a
regression line that would
leave, below the line, 5% of the
empirical data points.
95%
of data
points
5%
of data
points
?
13. NH 3.8, Vienna, 02-05-2014, Ivan Marchesini
Method II
● Quantile Linear Regression model (QLR)
resulted in:
S = 0.245 + 0.032 R
● Quantile Non-Linear regression model (QNL)
resulted in the exponential function:
S = 3.539*e(0.0028 × R)
19. NH 3.8, Vienna, 02-05-2014, Ivan Marchesini
LR, QLR, QNL
Non
susceptible
20. NH 3.8, Vienna, 02-05-2014, Ivan Marchesini
QLR zonation
● We prepared a zonation, showing
non-susceptible areas in Italy, both using the
quantile linear model QLR ...
21. NH 3.8, Vienna, 02-05-2014, Ivan Marchesini
QNL zonation
… and using the quantile non linear model QNL
22. NH 3.8, Vienna, 02-05-2014, Ivan Marchesini
Models performance
● Percentage of non-susceptible Italian territory:
– LR: 62%,
– QLR: 22%,
– QNL: 42%,
● Quantile Linear model (QLR) is very conservative respect to
the other two models
LR QLR QNL
23. NH 3.8, Vienna, 02-05-2014, Ivan Marchesini
Models performance
● Test dataset: IFFI, Italian Landslide
Inventory (Trigila et al., 2010).
●
Obtained though the IFFI WMS
service and setting ground
resolution at 5 m × 5 m
● The dataset contains:
– falls and/or topples,
– slow flows,
– rapid flows,
– complex movements,
– rotational/translational slides,
– lateral spreads,
– sinkholes,
– undefined slope movements.
From Trigila et al., 2010.
Progetto IFFI - ISPRA - Dipartimento Difesa del
Suolo-Servizio Geologico d'Italia -
www.sinanet.isprambiente.it/progettoiffi
24. NH 3.8, Vienna, 02-05-2014, Ivan Marchesini
Models performance
LR QLR QNL
FPR [%] 43.6 6.06 6.33
● We searched the proportion of
landslide cells that overlaid
non-susceptible areas: the False
Positive Rate:
– FPR = FP / (FP+TN)
● The more the FPR get close to 5%
the better is the model performance
● The QLR and QNL models performed
significantly better than the LR model
● QLR model is conservative, and so
we concluded that QNL is the bestQNL is the best
model.model.
FP TN
Nonsusceptiblearea
Landslide
25. NH 3.8, Vienna, 02-05-2014, Ivan Marchesini
QNL model performance
The QNL model performed better for
translational and rotational slides
Results are not far from the expected 5%
for slow flows, complex and undefined
movements
These landslide types represent 92% of the
IFFI landslides (in terms of covered area)
The QNL model failed to detect
non-susceptible areas for lateral spreads,
sinkholes, rapid flows and for falls and
topples.
Landslide types FPR
[%]
Rotational, translational
slides
5.3
Undefined 7.2
Slow flows 7.2
Complex movements 7.4
Falls and topples 8.3
Rapid flows 11.6
Sinkholes 13.8
Lateral spreads 20.9
26. NH 3.8, Vienna, 02-05-2014, Ivan Marchesini
QNL model performance
We further investigated the
performance of the QNL model in the
20 administrative regions in Italy.
● Size and shape deformations
depend on the IFFI landslide
density
● Orange and red colors show high
values of False Positive Rate.
● High values of FPR are frequently
associated with scarce density of the
inventory
27. NH 3.8, Vienna, 02-05-2014, Ivan Marchesini
QNL model performance
● We applied the non linear model QNL to the landmasses
surrounding the Mediterranean Sea
● Non-susceptible cells cover 3,652,683 km2, 63% of the area.
28. NH 3.8, Vienna, 02-05-2014, Ivan Marchesini
QNL model performance
We tested the synoptic-scale terrain zonation using independent landslide
information in Spain:
● Three inventories:
– Pyrenees, Murcia, and the Tramuntana range in Majorca,
– total of 521 landslides,
– total landslide area 27.24 km2.
● The resulting False Positive Rate (FPR) was: 6.11%
29. NH 3.8, Vienna, 02-05-2014, Ivan Marchesini
Conclusions
● Exploiting accurate landslide information for 13 study areas
in Italy we identified areas non-susceptible to landslides.
● We tested the Italian landslide non-susceptibility map
against independent landslide information (IFFI) and we
obtained promising results.
● We extended the application of the non-susceptibility model
to landmasses surrounding the Mediterranean Sea, and we
successfully tested the synoptic subdivision using
independent landslide information for three areas in Spain.
30. NH 3.8, Vienna, 02-05-2014, Ivan Marchesini
Conclusions
● Our work showed the importance of landslide information
for the production of maps of non-susceptible landslide
areas, and confirmed the importance of preparing
accurate landslide inventory maps.
● We expect that our synoptic-scale zonation for Italy and
for the landmasses surrounding the Mediterranean Sea
can be used for insurance and re-insurance purposes, for
large areas land planning, and in operational landslide
warning systems.
31. NH 3.8, Vienna, 02-05-2014, Ivan Marchesini
Thank you for your attention
Ivan.Marchesini@irpi.cnr.it
NHESSD Open Discussion:
I. Marchesini, F. Ardizzone, M. Alvioli, M. Rossi, and F. Guzzetti, (2014).
Non-susceptible landslide areas in Italy and in the Mediterranean region. Nat.
Hazards Earth Syst. Sci. Discuss., 2, 2813-2849, 2014