RAPID LANDSLIDE SUSCEPTIBILITY MAPPING

Andrew J. W. McDonald
British Geological Survey
Kingsley Dunham Centre
Keyworth, Nottingham, UK
NG12 5GG
+44 (0)115 9363225, +44 (0)115 9363474 ( f a )
a.mcdonald@bgs.ac.uk

Eugene A. O'Connor
British Geological Survey
Kingsley Dunham Centre
Keyworth, Nottingham, UK
NG12 5GG
+44 (0)115 9363416, +44 (0)115 9363474 ( f a )
e.o'connor@bgs .ac.uk

David Greenbaum
British Geological Survey
Kingsley Dunham Centre
Keyworth, Nottingham, UK
NG12 5GG
+44 (0)115 9363227, +44 (0)115 9363474 ( f a )
d.greenbaum@bgs.ac.uk

Key Words Landslide susceptibility; remote sensing; GIS; Papua New Guinea; Fiji; Jamaica; Slovakia

Abstract

Landslides are a common natural hazard and have a major impact on life and property. Conventional
geotechnical methods of producing regional landslide hazard maps can be slow and costly. Remote sensing
and spatial analysis using a GIS provide potentially usehl tools for the rapid preparation of such maps.
Pilot studies were carried out in Papua New Guinea and Fiji using satellite images, aerial photographs and
existing geoscientific information. Models of landslide susceptibility were created by correlating the various
information themes with the known distribution of landslides, interpreted from the imagery, and by
combining the information. A refined method, using a probability based weights-of-evidence analysis, was
applied to hrther test areas in Jamaica and Slovakia. The resulting maps display good correspondencewith
the known distribution showing that the technique can map landslide susceptibility in a range of vulnerable
environments.

Introduction

Landslides are a common natural hazard which, together with earthquakes, volcanic eruptions and floods,
have a major impact on life and property. Although most landslides are small and cause few fatalities, the
cumulative losses world-wide account for around 25 per cent of annual deaths from natural hazards
(Hansen, 1984). A knowledge of the causes and incidence of landsliding can help planners: (1) make
contingency plans to prepare for, and mitigate against, the effects of landslide events on infrastructure,
housing and people; and (2) avoid or minimise the risks associated with new developments.

The general term landslide is used here to describe the mass movement of soil and rock downslope under
the influence of gravity. Landslides result from natural slope instabilities and a combination of other
factors. They possibly represent the single most important mechanism of erosion, especially in tropical
environments. The term hazard refers to the probability of Occurrence of an event in an area within a
specified period of time (Varnes, 1984), although it is commonly used in a relative rather than an absolute
sense. The distribution of landslide hazards is usually presented on a map as simple zones (e.g. high,
intermediate, low) indicating the relative likelihood, or susceptibility, of landsliding.

Conventional geotechnical investigations, involving materials testing and extensive fieldwork, are slow,
specialised and costly. Whereas they are appropriate for detailed studies in urban areas, they cannot in
general be justified, nor are they economically feasible, at the regional scale, particularly in many
developing countries which do not have the resources to carry out such work. As a result, hazard maps are
only available for a few areas although many parts of the world experience landsliding. Given the
magnitude of the task world-wide, it is important that alternative methods are developed to provide
provisional hazard information cheaply and quickly.

Aims and Ob-iectivesof the Study

The initial aims of the study were: (1) to establish, through pilot studies in test areas, methods by which
remote sensing, combined with spatial analysis in a Geographic Information System (GIS), could be used
to produce a preliminary hazard zonation map; and (2) to provide advice and training in these techniques to
geologists in the countries where the pilot studies were conducted (Greenbaum, 1995).

The wider objective was to develop the basis of a generic methodology for rapid landslide susceptibility
mapping that could be implemented in developing countries world-wide. The overriding purpose was to
identify approaches that could lead to a rapid, cost-effective and practical solution. The use of empirical,
less rigorous techniques inevitably results in less precise information but there are many situations where
such preliminary information can provide a valid interim solution.

Study Areas

Test sites for the first phase of the project were chosen in two countries where landsliding is a significant
problem and represents a threat to life and property: Papua New Guinea and Fiji (Greenbaum and others,
1995a, 1995b). In both countries, hazard maps are an important and urgent requirement. National agencies
are involved in landslide hazard mapping but the size of the problem, in terms of area to be covered, is
large compared with resources. If significant progress is to be made, this must involve the use of rapid,
more cost-effective techniques. The test sites chosen provide a range of different situations requiring
somewhat different techniques.

For the second phase of the study, the methodology was refined and applied to two further, more typical,
test areas, Jamaica and Slovakia, where landsliding is widespread.

Papua New Guinea (PNG): A large, mountainous and sparsely populated country, PNG has both
high rainfall and high seismicity. The region is tectonically active, the Highlands experiencing some
of the greatest rates of uplift in tectonic collision belts. As a result, landslides in this rugged terrain
involve considerable volumes of material. Although most occur in regions of low population density,
 their widespread distribution and size result in some loss of life and damage to infrastructure. Given
the large area involved, the requirement is for regional to national scale hazard zonation maps at
scales of 1:lOO 000 to 1:250 000.

Fiji: A small South Pacific nation comprising two main islands and many smaller ones. Although
geologically young, the islands are tectonically more stable than PNG and the relief more subdued.
The association of deep tropical weathering with high intensity rainfall, especially during tropical
cyclones, results in generally small, though locally numerous and damaging, landslides. The
requirement in Fiji is for medium to regional scale hazard maps covering at least the two main
islands.

Jamaica: In many respects, the landslide hazard on the Caribbean island of Jamaica is similar to that
on Fiji. Lying close to the northern boundary of the East Caribbean Plate, Jamaica has moderate
seismicity and is also affected by hurricanes; both acting as triggers for landslides. A relatively
stable limestone plateau forms most of the island. The Central Inlier mountain zone, the area of the
present study, is drained principally by the Minho and Thomas rivers. These rivers and their
tributaries cut down through a series of fault-bounded volcaniclastic rocks and sediments producing
ridge and slope physiography. The main requirement is for large scale (1 :50 000) maps within the
most affected areas.

Slovakia: Located in central Europe, Slovakia can be divided into two broad regions, the low, flat-
lying Danube Basin in the south, and the mountain belt of the West Carpathians in the north. The
upland areas are prone to landsliding, particularly in the Flysch Belts which comprise rhythrmc
sequences of sandstones, claystones and mark Much of the mass movement is in the form of soil
creep and earthflows resulting from seasonal rainfall and weathering of the regolith. The Javorniky
Mountains in West Slovakia, near the border with the Czech Republic, form the study area. They
comprise a series of north-east trending sandstone ridges, which tend to be stable, separated by
valleys in which there is evidence of landsliding on the lower slopes. Much of the area is covered by
coniferous forest although there is some pasture and cultivation in the valleys.

Methodology

For a landslide to occur a number of conditions must be satisfied. These conditions may be classified as (1)
controlling factors, and (2) triggering events. Controlling factors may be broadly divided into material
properties (rock/soil type; in situ and bulk strength; etc.) and terrain conditions (slope angle; slope aspect;
fracturing; cultivation; etc.), whereas triggering events include earthquakes, intense rainstorms, clearing of
vegetation and possibly new constructiorddevelopment.If all the controlling factors in an area are known, it
may be possible to say where landslides are most likely to occur given a particular triggering event.

The approach adopted here assesses the importance of assumed controlling factors. Underlying the method
is the assumption that it may be possible to predict where future events are most likely to occur by relating
the distribution of past landslides to information on geology, topography, soils, rainfall etc. This is achieved
by statistical correlation within the GIS. Satellite imagery and aerial photographs are used to provide the
basic inventory of past landsliding events. This methodology contrasts with conventional geotechnical
approaches, which involve slow and expensive ground surveys and laboratory testing to establish rock and
soil properties.

The simplest form of hazard zonation from imagery or photography is a distribution map of past landslides,
perhaps subdivided into older and more recent events. Here, concentrations of past landslides may be taken
to indicate generally unstable areas. At best, a landslide susceptibility map of this type is a general guide to
fbture hazards.

Adding information such as topography, geology, rock structures, lineaments, roads, infrastructure,
population distribution etc. will provide a more informative map. In such maps the spatial association
between landslides and these other variables may be visually apparent but the approach does not make use
of the data in any rigorous or quantifiable manner. Maps of this type require only a cartographic capability.
A vector-based GIS provides an appropriate means of storing data and producing customised, easily
updated maps as and when required.

A more quantitative approach involves ranking the different variables according to their calculated
statistical importance to landsliding and combining them to produce a hazard susceptibility map. This type
of analysis can be approached in a number of ways depending on the information held and the availability
of supporting field or gmtechnical evidence. It requires a raster-based GIS. Here, every point within the
area has a classified value for each attribute, thus making the analysis of spatial relationships easy.

Remote Sensing

Satellite imagery and aerial photography can provide rapid regional overviews. They provide information
on old and recent landslides, geological faultdfractures, bedding and other lithological structures, recent
erosionaVdepositiona1processes, habitation, cultivation and infrastructure which can be used within the
GIS analysis.

The type of data used must relate to the scale of the problem. For the PNG study area where the landslides
are large, false colour composites from the Thematic Mapper sensor on the Landsat satellite were used to
create the landslide inventory, lineament distribution and river catchments. These data were interpreted
manually having first been digitally rectified to a base map, edge enhanced and contrast stretched.

The resolution of current satellite imagery is not appropriate for areas such as Fiji, Jamaica, and Slovakia
where the landslides are smaller than in PNG. In these areas, aerial photographs were interpreted
stereoscopically to give the distribution of landslides. Identification of these landslides relies on the
stereoscopic view to provide textural and contextual information. Combining the interpretation of the
stereo-pairs and transferring the information to the GIS can be difficult and time consuming due to the
geometric properties of the photographs. One rapid method is to identify the landslide initiation points on
aerial photographs and annotate these on suitable satellite images which have been registered to the map
projection for the study (O’Connor, 1998).

Obtaining cloud-free optical imagery is a problem especially in tropical regions. Radar imagery may
overcome this difficulty but its geometric and radiometric properties in mountainous regions may reduce its
utility. Conventional aerial photography and available topographic maps can provide an interim solution.

Landslide susceptibility analysis

Existing maps of different themes (e.g. geology, soils, topography) covering each study area were digitised
and a set of layers created within the GIS. Each layer comprises a set of classes representing the different
sub-divisions of the theme within the area. For example, the geology theme consists of classes representing
the type of rock at any point. The analysis relies on testing the spatial correlation of the landslide
distribution map with each theme.
Area1 cross-tabulation of the landslide distribution map with each theme is the basis of the analysis. For
each class this gives the area of positive match, that is the area covered by landslides within each class; the
areas of mis-match, where either landslides or the class is present; and the area of negative match where
neither are present. The overall association of the theme with landsliding can be estimated by chi-squared,
or similar, statistics although such measures can only be used for exploratory and descriptive purposes.
Chi-squared values can also be calculated for the individual classes to give a relative indication of
association with landsliding.

Dividing the results of the area cross-tabulation by the total area enables the association of classes with
landsliding to be assessed in terms of probability ( B o b - C a r t e r , 1994). Different measures of
association can be calculated for each class and compared between themes. Of interest is the measure, a,
proposed by Yule (quoted in Fleiss, 1991) which ranges in value between -1 and +1, like a correlation
coefficient, with 0 implying independence of the two patterns. Using these values as weighting factors, each
theme is reclassified as a landslide susceptibility map.

The next step is to combine themes to see if the predictability of landsliding can be improved. Logically, if
two themes (e.g. rock and soil) each individually relate to landsliding in some undefined way, then the two
themes taken together should provide a more reliable indicator. The themes, reclassified as landslide
susceptibility maps, are combined using a weighted addition.

An alternative method of combining the themes, and one which makes direct use of the probability
information, is that of weights-ofevidence. The rationale of this approach is that several classes can be
used to predict another and that the prediction can be assigned a probability. The technique has been used
effectively in mineral exploration ( B o b - C a r t e r , 1994).

A pair of weights, w' and W-, is calculated for each class from the area cross-tabulation results. Weight
w' is related to the probability of landsliding within the particular class whereas weight W is related to the
probability of there being a lidslide outside the area of the particular class. Both weights are normalised
to the regional average probability of landsliding. Each class is then treated as a binary map, indicating
presence or absence of the class, and reclassified using w' where the class is present and W-elsewhere.
These weight maps are combined to provide an estimate of the probability of a landslide occurring at any
point in the study area. The methodology is summarised in Figure 1.

Results

Papua New Guinea

Six themes were available for analysis in PNG: geology; elevation; slope angle and aspect (both calculated
from the DEM); lineaments and river catchments (interpreted from the imagery). Individual correlation
with the landslide inventory illuminates significant associations. The majority of landslides occur within
only four of the 22 rock units and on slopes with angles between 30" and 50". Landslides are common at all
heights above 750 m but show a small increase to a maximum at 2 750 m - 3 000 m. There is no
association with the slope aspect, as might be the case if the controlling factor or triggering mechanism was
directional, such as the geological strike or the prevailing direction of tropical storms; and aspect was
excluded from hrther analysis. Being line features, lineaments cannot be examined in the same manner.
Consequently, 1 000 m corridors were created around the lineaments and the cross tabulation performed.
Surprisingly, the probability of landsliding was constant within these corridors. Since it might be argued
that fractures relate more to landslide initiation than area1 extent, the analysis was modified to examine the
distribution of landslide start points to lineaments. Higher than expected values were found within 100 m of
the lineaments, reducing to regional average values at 700 m.

The catchment theme does not truly represent a controlling factor but provides a representation of the
landslide density variation across the study area. It was included in the analysis to make the high
probability areas correspond more closely to the mapped distribution.

Six themes were also available for Fiji: geology; elevation; slope angle and aspect; soils and forestry land
use. The majority of landslides were related to six different rock units,four of which were conglomerates of
different types. Most of the slopes in this region are gentle with the majority of landslides occurring on
slopes less than 25'. The highest risk areas are where slopes are between 10" and 20'. There is a slight
association with slope aspect with a tendency for landsliding on north or northeast facing slopes. Some of
the 42 soil types show an association with landslides but it is difficult to generalise this information. Most
landslides occur between 50 m and 350 m, with the highest risk in the range 250 m - 300 m. There is an
association between landslides and areas used for hardwood forestry but it is unclear whether there is a
causal relationship between forestry and landslides or whether it is simply the best use for land subject to
landsliding.

Jamaica

The analysis for Jamaica was based on the distribution of landslide initiation points rather than the area
covered by debris as in the earlier test cases. Six themes were also available for Jamaica: geology; soils;
elevation; slope angle and aspect; and distance to faults. The majority of landslides were developed on four
different, predominately conglomeratic, rock units.Seven of the 2 1 soil types show an association with
landslides. Two of these, the Pennants clay loam and the C u e Gully sandy loam, have stronger
associations (approximately by a factor of two) than the rest. Most landslides occur between 370 m and
660 m but the highest associations in the range 730 m - 970 m. Most of the slopes in this region are gentle
with the majority of landslides occurring on slopes between 10" and 25". The areas of highest association
occur on slopes between 40' and 45'. There is some association with slope aspect with a tendency for
landsliding on south-east to west facing slopes. This is consistent with a directional trigger mechanism such
as intense rainfall from tropical storms and hurricanes which most frequently track across the island from
the south-east or south. The majority of initiation points were less than 1 500 m from a mapped geological
fault. The landslide susceptibility map for the upper Rio Minho watershed from the weights-of-evidence
analysis is shown in Figure 2. The current analysis has made no attempt to differentiate between different
types of mass-movement. In the Jamaican context, the distinction of two types of landslide seems important
and further analysis is required.

Slovakia

Only preliminary analysis of five themes has been performed for Slovakia. The themes are: geology;
elevation; slope angle and aspect; and distance to lineaments interpreted from aerial photographs, which
may relate to faults. There appears to be a clear association between landslide initiation points and
claystone formations with a smaller relationship with marly limestones. Nearly all of the landslides occur
below an altitude of 700 m and on slopes of less than 20'. It is possible that the coniferous woodlands on
slopes above 700 m have a stabilising effect accounting for the lack of landsliding above this height. Again,
there is a directional association with most of the landslides being on slopes facing east through to south-
west. The majority of landslides occur within 400 m of a photo-lineament. The preliminary landslide
susceptibility map for the Javorniky Mountains from the weights-of-evidence analysis is shown in Figure 3.

Conclusions and Discussion

The test studies show that a combination of remote sensing and spatial analysis can be used to map
landslide hazards in a range of vulnerable environments. Remote sensing can provide a reliable and rapid
means of mapping landslide-relatedterrain features. A GIS is an appropriate and efficient tool for storing
and displaying the data; analysing the relationships between attributes and producing susceptibility maps.

There are, however, some difficulties with the approach. The imagery used must match the scale of the
landslides and needs to show a representative distribution of recent landslides. To satisfy the latter
requirement the imagery has to be acquired at a suitable time - usually soon after a triggering event such as
the period following intense rainfall. During dry periods in tropical areas, landslide scars are difficult to
identify due to lack of movement and rapid re-vegetation.

In theory, radar imagery has two advantages over optical information. The first is that it is unaffected by
weather conditions and the second that it can be acquired at any time of the day. This means the chance of
acquiring images for an area at a suitable time is higher. Examination of radar images within this study
show that their geometric and radiometric characteristics are such that landslides cannot be easily
identified. The utility of radar imagery may lie in the production of DEMs or for direct detection of mass
movement; both approaches require sophisticated image processing procedures.

There remains a question about the valilty of any landslide susceptibility map produced by this method
due to errors of omission and commission. Since the technique uses only data that are readily available,
some information concerning a critical controlling factor for landslides may not be included in the analysis.
For example, the hydrology or the depth of weathering may be more important themes than the underlying
geology. Conversely, themes may be included that have only marginal association with landsliding.
Although the analysis should indicate which themes these are, defining the threshold level of significance is
subjective. There is no statistically rigorous technique to assess the validity of results other than to compare
various combinations with the landslide distribution map. Again, measuring the goodness of the model in
this manner is subjective to some degree. The estimates of probability given by the weights-of-evidence
approach reduces this subjectivity to some extent.

Entering data into a GIS can be time consuming; some workers suggest that 90 per cent of the GIS effort is
concerned with building the database. Available information is often based on different map projections
about which there is insufficient detail to perform a rigorous co-registration. Creating a DEM, which is
central to the analysis, can also require significant effort.

A landslide susceptibility map is not an end in itself the information must be made available and
understandable to development and emergency planners. The information must be targeted at specific
audiences. For example, transport departments need information about the effect on roads whereas water
authorities will be interested in how rivers and water supplies will be affected. Using a GIS enables the
appropriate details to be included easily. For example, as more information becomes available, either as
new or updated themes, the analysis can be quickly revised and new maps produced. These can be tailored
to suit particular needs by inclusion of specific information such as the road network. The way in which the
information is presented also affects whether it is used effectively. Although the maps produced can show
all the necessary information they can be difficult to interpret. Visualisation techniques, including
perspective views and animations, appear to be useful ways by which the landslide susceptibility
information can be presented to best effect.

Acknowledgements

The work reported here has been undertaken by the British Geological Survey under the Knowledge and
Research Programme of the Department for International Development and is part of the British
Government’s provision of aid to developing countries. Many people have contributed to the study and their
help and interest is gratefully acknowledged. In particular, we would like to thank our collaborators in the
Geological Survey of Papua New Guinea, the Mineral Resources Department (Fiji), the Geological Survey
of the Slovak Republic, and the Department of Geography and Geology of the University of the West
Indies (Kingston, Jamaica). The support of the Directors and Heads of Departments is gratefully
acknowledged. This paper is published by permission of the Director, BGS (NERC).

References

Bonham-Carter, G.F., 1994, Geographic Information Systemsfor Geoscientists: Modelling with GIS.
Pergamon Press (Elsevier Science Ltd.), Oxford, 398p.

Fleiss, J.L., 1991, StatisticalMethods for Rates and Proportions. Wiley & Sons, New York, 321p.

Greenbaum, D., 1995, Project Summary Report: Rapid methods of landslide hazard mapping. British
Geological Survey Technical Report WCl95130, 12p.

Greenbaum, D. and others., 1995a, Rapid methods of landslide hazard mapping: Papua New Guinea case
study. British Geological Survey Technical Report WCl95127, 112p.

Greenbaum, D. and others., 1‘995b,Rapid methods of landslide hazard mapping: Fiji case study. British
Geological Survey Technical Report WCl95128, 103p.

Hansen, A., 1984, “Landslide hazard analysis.” In Slope Instability, eds. D. Brunsde and D.B. Prior,
Wiley & Sons, New York, Chap. 13, pp. 523-602.

O’Connor, E.A., 1998, Interpretation of slope movements from aerial photography: Jamaica and Slovakia.
British Geological Survey Technical Report WO/98/12R, 9p.

Varnes, J., 1978, “Slope movement types and processes.” In Landslides: analysis & control, eds. R.L.
Schuster and R.J. Krizek, Special Report 176, Transportation Research Board, National Academy of
Sciences, Washington D.C., pp. 11-33.
 Interpret imagery Landslide Distribution Map

rw
r Cfoss tabulation
v
Assess degree of
Geology association with landslides
v
-
Digitise existing
Reclassify maps as
landslide susceptibility
v
map information Combine susceptibility maps
- simple addition
- weights-of-evidence

Landslide Susceptibility Map I Visualisation I

Figure 1. Schematic summary of the methodology

m
N
#
m
N
v1
.r(
v1
,h
B

8 8 a
8
N
8
N
N
!i
P
N

3E

8
E
2
0
!2 %
N c4
E

3!
N
8
0
0
8
0
0
80
0
8
N

N
E z
.

I I I I

83 84
d *
Y
e
n
.-
Q
0
m
Q
0
m
m m
0 P

8 0
0
8N 8
C
Y
,m :

!
-
m
8-
m
P P

0 0

i
m
d
8
m
d

!
Q)

N 0 0 0
8
0)
N
rs 0
0
0 0 P
0 0
0 0 0
OD (D m
d
Ln 2 2