Delineation of landslide hazard areas using frequency ratio, logistic regression and artificial neural network model at Penang Island, Malaysia. Environ Earth Sci

Abstract

This paper summarizes findings of landslide hazard analysis on Penang Island, Malaysia, using frequency ratio, logistic regression,
and artificial neural network models with the aid of GIS tools and remote sensing data. Landslide locations were identified
and an inventory map was constructed by trained geomorphologists using photo-interpretation from archived aerial photographs
supported by field surveys. A SPOT 5 satellite pan sharpened image acquired in January 2005 was used for land-cover classification
supported by a topographic map. The above digitally processed images were subsequently combined in a GIS with ancillary data,
for example topographical (slope, aspect, curvature, drainage), geological (litho types and lineaments), soil types, and normalized
difference vegetation index (NDVI) data, and used to construct a spatial database using GIS and image processing. Three landslide
hazard maps were constructed on the basis of landslide inventories and thematic layers, using frequency ratio, logistic regression,
and artificial neural network models. Further, each thematic layer’s weight was determined by the back-propagation training
method and landslide hazard indices were calculated using the trained back-propagation weights. The results of the analysis
were verified and compared using the landslide location data and the accuracy observed was 86.41, 89.59, and 83.55% for frequency
ratio, logistic regression, and artificial neural network models, respectively. On the basis of the higher percentages of
landslide bodies predicted in very highly hazardous and highly hazardous zones, the results obtained by use of the logistic
regression model were slightly more accurate than those from the other models used for landslide hazard analysis. The results
from the neural network model suggest the effect of topographic slope is the highest and most important factor with weightage
value (1.0), which is more than twice that of the other factors, followed by the NDVI (0.52), and then precipitation (0.42).
Further, the results revealed that distance from lineament has the lowest weightage, with a value of 0. This shows that in
the study area, fault lines and structural features do not contribute much to landslide triggering.

KeywordsLandslide-Hazard-Frequency ratio-Logistic regression-Artificial neural network-GIS-Malaysia

Full text

ORIGINAL ARTICLE
Delineation of landslide hazard areas on Penang Island,
Malaysia, by using frequency ratio, logistic regression,
and artificial neural network models
Biswajeet Pradhan ÆSaro Lee
Received: 11 November 2008 / Accepted: 7 July 2009 / Published online: 25 July 2009
ÓSpringer-Verlag 2009
Abstract This paper summarizes findings of landslide
hazard analysis on Penang Island, Malaysia, using frequency
ratio, logistic regression, and artificial neural network
models with the aid of GIS tools and remote sensing data.
Landslide locations were identified and an inventory map
was constructed by trained geomorphologists using photo-
interpretation from archived aerial photographs supported
by field surveys. A SPOT 5 satellite pan sharpened image
acquired in January 2005 was used for land-cover classifi-
cation supported by a topographic map. The above digitally
processed images were subsequently combined in a GIS with
ancillary data, for example topographical (slope, aspect,
curvature, drainage), geological (litho types and lineaments),
soil types, and normalized difference vegetation index
(NDVI) data, and used to construct a spatial database using
GIS and image processing. Three landslide hazard maps
were constructed on the basis of landslide inventories and
thematic layers, using frequency ratio, logistic regression,
and artificial neural network models. Further, each thematic
layer’s weight was determined by the back-propagation
training method and landslide hazard indices were calculated
using the trained back-propagation weights. The results of
the analysis were verified and compared using the landslide
location data and the accuracy observed was 86.41, 89.59,
and 83.55% for frequency ratio, logistic regression, and
artificial neural network models, respectively. On the basis
of the higher percentages of landslide bodies predicted in
very highly hazardous and highly hazardous zones, the
results obtained by use of the logistic regression model were
slightly more accurate than those from the other models used
for landslide hazard analysis. The results from the neural
network model suggest the effect of topographic slope is the
highest and most important factor with weightage value
(1.0), which is more than twice that of the other factors,
followed by the NDVI (0.52), and then precipitation (0.42).
Further, the results revealed that distance from lineament has
the lowest weightage, with a value of 0. This shows that in the
study area, fault lines and structural features do not con-
tribute much to landslide triggering.
Keywords Landslide Hazard Frequency ratio 
Logistic regression Artificial neural network GIS 
Malaysia
Introduction
Landslides are a recurrent problem throughout most of
Malaysia, where they are serious threats to settlements and
the structure that supports transportation. Occasionally they
also result in loss of life. Damage to people and property was
particularly high in the most recent years of 2000, 2001, 2002,
2004, 2006, 2007, 2008, and 2009. These landslides have
caused substantial damage to highways, waterways, proper-
ties, and livestock (Sassa 2008). Although most of these
landslides occurred on cut slopes or embankments alongside
roads and highways in mountainous areas, there are records of
landslides in other relatively low-slope areas. A few
B. Pradhan (&)
Faculty of Forestry, Geo and Hydro-Science,
Institute of Cartography, Dresden University of Technology,
01062 Dresden, Germany
e-mail: Biswajeet.Pradhan@mailbox.tu-dresden.de;
biswajeet@mailcity.com
S. Lee
Geoscience Information Center,
Korea Institute of Geoscience and Mineral Resources (KIGAM),
30, Kajung-Dong, Yusung-Gu, Daejon, Korea
e-mail: leesaro@kigam.re.kr
123
Environ Earth Sci (2010) 60:1037–1054
DOI 10.1007/s12665-009-0245-8

landslides have been reported near high-rise apartments and
in residential areas, causing loss of life (Cheng et al. 2008).
Landslides in Malaysia are natural degradational processes
and are largely triggered by heavy rain because of either a
single heavy thunderstorm/rain or successive days of mod-
erate rain, especially during the rainy season, causing flash
floods leading to failure of the rock surface along fracture,
joint, and cleavage planes (Sin and Chan 2004; Lloyd et al.
2001). Rainfall and subsequent movement of ground water
greatly affect slope stability. Rainfall causes a change in the
moisture content of the soil. Changes in moisture content
increase the interstitial pore water pressure, seepage pressure,
and soil weight, and reduce cohesion. The presence of heavy
boulders in the soil mass triggers the slide mechanism. In
recent years, the incidence of landslides has increased, and
this could be attributed to rapid development and land
clearing in mountainous areas. The geology of the country is
quite stable but continuous development and urbanization
leads to deforestation and weathering, with erosion of the
covered soil masses causing serious threat to slopes.
Damage and losses are regularly incurred in Malaysia
because, historically, there has been too little consideration
of potential problems in land use planning and slope man-
agement. In recent years greater awareness of landslide
problems has led to significant changes in the control of
development on unstable slope areas. Recently, the
Malaysian government and highway authorities have been
stressing the need for local planning authorities to take
landslides into account at all stages of the landslide hazard
mapping process. So far, few attempts have been made to
predict these landslides or to prevent the damage they cause.
Scientifically, there are three steps in landslide-suscepti-
bility analysis: susceptibility, possibility (often termed
‘‘hazard’’), and risk (Varnes 1984). According to Varnes
(1984)inEqs.1and 2, susceptibility is the function of
landslide and its causative factors whereas hazard analysis is
described as the function of susceptibility and associated
triggering factors. The triggering factors are normally referred
to external forces such as earthquake or rainfall. In this paper,
three different methods, frequency ratio, logistic regression,
and artificialneural networks, were developed and applied for
determination of weight and rating in landslide hazard anal-
ysis. Historical rainfall distribution data were combined with
the susceptibility map to prepare the hazard map.
Susceptibility ¼fLandslide, Landslide related factorsðÞ
ð1Þ
Possibility/Hazard ¼fSusceptibility, trigerring factorsðÞ
ð2Þ
Through this kind of prediction model, landslide
damage could be greatly reduced to some extent. In
addition, landslide-related factors were also assessed and
weightage for each of the thematic layers was calculated.
However, to be able to spatialize these statistically based
prediction models (or any other uni-dimensional model
based on measurements made on small plots), a few
assumptions are normally made, including the assumption
that the landscape under study is continuum of small units
on which the models are valid. This can be true for
homogenous land cover or soil types but it is less evident
for fragmented landscapes composed of a mosaic of units
behaving differently in terms of landslides.
Many studies have been carried out on landslide hazard
evaluation using GIS; for example, Guzzetti et al. (1999)
summarized many landslide hazard evaluation studies.
Recently, there have been studies applied using probabi-
listic models (Akgun et al. 2008; Dahal et al. 2008;
Clerici et al. 2002; Lee and Talib 2005; Lee and Pradhan
2006; Pradhan et al. 2006) for landslide hazard evalua-
tion. Logistic regression models have also been applied to
landslide hazard mapping (Lee and Dan 2005; Akgul and
Bulut 2007; Tunusluoglu et al. 2008; Lamelas et al. 2008;
Wang and Sassa 2005; Atkinson and Massari 1998;Su
¨zen
and Doyuran 2004; Dai and Lee 2002; Lee 2005; Pradhan
et al. 2008) in different parts of the world. The geo-
technical model and the safety factor model have also
been applied by many researchers (Gokceoglu et al. 2000;
Romeo 2000; Refice and Capolongo 2002; Carro et al.
2003; Zhou et al. 2003; Youssef et al. 2009). Landslide
hazard evaluation using GIS, including data mining using
fuzzy logic, and artificial neural network models have
been applied in other areas (Ercanoglu and Gokceoglu
2002; Lee and Pradhan 2007; Pradhan et al. 2009;
Lee 2007; Pradhan and Lee 2008a,2008b; Choi et al.
2009).
Previously, not much work has been done on landslide
hazard and risk analysis on Penang Island. Toh (1999) and
Ooi (1999) presented some interesting case studies of
Penang Island dealing with preventive and remedial work
for rockfalls involving granite boulders, for example the
use of wire nets and fences, cables, and rock anchors, and
the concreting together of loose boulders. From their study,
where numerous boulders appear as in a boulder field,
some of these preventive methods may not be feasible or
practicable. Recently, Lee and Pradhan (2007) have con-
ducted landslide hazard and risk analysis using a statisti-
cally based frequency ratio model on Penang Island. In this
paper, hazard analysis was further extended using logistic
regression and artificial neural network models to assess
the most suitable results for the study area. The purpose of
this study was to provide a detailed hazard analysis of the
Penang Island area using three different types of models for
assessment of the best statistical model for this particular
study.
1038 Environ Earth Sci (2010) 60:1037–1054
123

Study area
In this research, Penang Island was selected for imple-
mentation of landslide susceptibility model analysis
because of frequent occurrence of landslides. The study
area (Fig. 1), is undergoing rapid development with land
clearing for housing estates, hotels, and apartments causing
erosion and landslides. Penang is one of the 13 states of
Malaysia and is located on the North West coast of the
Malaysia peninsula. It is bounded to the north and east by
the state of Kedah, to the south by the state of Perak, and to
the west by the Straits of Malacca and Sumatra (Indonesia).
Penang consists of the island of Penang, and a coastal strip
on the mainland, known as Province Wellesley. The island
covers an area of 285 km
2
, and is separated from the
mainland by a channel. The study area is located approx-
imately from latitudes 58150Nto58300N and longitudes
1008100E to 1008200E. The land use in the study area is
mainly peat swamp forest, plantation forest, inland forest,
scrub, grassland, and ex-mining area. The slope of the area
ranges from 0°to as much as 87°. The relief of the study
area varies between 0 and 420 meters above sea level. The
terrain consists of costal plains, hills, and mountains.
According to the Malaysian Meteorological Department,
the temperature of the northern part of Penang ranges
between 298C and 328C and mean relative humidity varies
Fig. 1 a Tectonic setting of
Peninsular Malaysia (Source:
Minerals and Geoscience
Department Malaysia). bSPOT
5 satellite image with landslide
location map of Penang Island
Environ Earth Sci (2010) 60:1037–1054 1039
123

between 65 and 96%. The highest temperature is during
April to June and the relative humidity is lowest in June,
July, and September. The amount of rainfall in the study
area varies between 58.6 and 240 mm per month.
Geologically, the study area is located on the western
flank of the Main Range, which is composed of granite.
The granite in Penang Island is classified on the basis of the
proportion of alkali feldspar to total feldspars (Ahmad et al.
2006). Ahmad et al. (2006) reported that some of the
granite and associated micro-granite are further divided
into two main groups: the North Penang Pluton, approxi-
mately north of latitude 5°230, and the South Penang Plu-
ton. In the northern part of the island, the alkali feldspars
that generally do not exhibit distinct cross-hatched twining
are orthoclase to intermediate microcline in composition.
In the southern region, they generally exhibit well-devel-
oped crosshatched twining and are believed to be micro-
cline (Ong 1993). The regional geology map of the study
area is shown in Fig. 2. The dominant granite geology,
granitic soils and boulders, and hillside development in the
study area give rise to distinct landslide and rockfall
problems. The historical incidence of rockfalls involving
the dislodgement of large granitic boulders highlight the
problem in the Penang area, in particular where rockfalls
threaten housing schemes and condominiums (Tan 1990).
The problems of rockfalls and landslides involving granitic
rocks, boulders, and granitic soils on Penang island is not
new, as documented previously by Tan (1990). Hilly roads
cutting through granitic terrains invariably cut into the deep
weathering profiles of granitic rocks and soils, exposing, in
particular, the core stones and bedrock surface, and are
prone to rockfalls and landslides. Similarly, excavation and
blasting of cut slopes in housing schemes and condomin-
ium developments also expose such deep weathering pro-
files of granite, leading to instability problems. In addition,
colluvial deposits and colluvial boulders which accumulate
in small valleys up slopes or uphill of roads and housing
schemes also pose serious rockfall hazards.
The soil of Penang Island mainly comprises eight types,
as shown in Table 1. The kuala-kedah permatuang asso-
ciation mainly consists of clay material. The steep land soil
series covers mostly the central part of the Island and is
composed of weathered granitic rocks. The rengam series
is composed of fine to medium coarse sandy clay and is
developed on igneous and high-grade metamorphic rocks
whereas selangor kangkong association is mainly com-
posed of medium-grained clay and is generated from acid
sulfate soils. Local alluvium and colluvium series consist
of fine to medium grained loamy material. More than 70%
of the landslides have been reported in the steep land soil
types.
Rainfall on Penang Island, averages between 2,670 and
6,240 mm per annum (Fig. 3). There are two pronounced
wet seasons from September to December and from Feb-
ruary to May each year. The rainfall on Penang Island
peaks between March and May and from November to
December. Recorded single-day rainfall highs range from
87 to 200 mm. It is during such times that many streams
and rivers in Penang Island overflow, flooding the sur-
rounding areas, and landslides such as debris flow occur
along the river valleys. These landslides mainly consist of
flows, rockfall, and shallow soil slips that take place during
3–4 h of high-intensity rainfall. Landslides occur when the
maximum daily rainfall exceeds 100 mm, with a maximum
Fig. 2 Regional geological
map of Penang Island (after
Ahmad et al. 2006)
1040 Environ Earth Sci (2010) 60:1037–1054
123

hourly rainfall of 40 mm. During the rainy season, daily
rainfall of about 30–50 mm occurring successively over a
period of several days has triggered many landslides.
Highland development in Penang Island gives rise to
severe soil erosion which is then followed by siltation
downstream, degrading surface water quality by
sedimentation. Rainfall intensities often exceed 100 mm
per hour and are therefore extremely erosive. Chan (1998a,
1998b) has shown that the rate of soil loss in Penang is
alarming. His study in the Penang hill area, for example,
indicates that an average slope of 30°gives rise to a 50-fold
increase in surface runoff and soil loss of between 700 and
10,000 tonnes/ha per year (very severe erosion). Especially
during the rainfall reason, between 35 and 80% of the
annual soil loss occurs in a single month.
Multiple field investigations were carried out (5–7 May
2005; 15–19 July 2006; 3–7 June 2007; and 12–16
November 2008) and standard mapping information on
landslide locations were collected using GPS survey
(Fig. 4).
The data set
Identification and mapping of a suitable set of instability
factors having a relationship with the slope failures require
a-priori knowledge of the main causes of landslides. These
instability factors include surface and bedrock lithology
and structure, seismicity, slope steepness and morphology,
stream evolution, groundwater conditions, climate, vege-
tation cover, land use, and human activity. The availability
of thematic data varies widely, depending on the type,
scale, and method of data acquisition. But collecting an
ideal landslide related database on a suitable scale is often
intriguing. To apply the probabilistic model, a spatial
database that considers landslide-related factors was
designed and constructed. These data are available in
Malaysia as either paper or digital maps. The spatial
database constructed is listed in Table 2.
Table 1 Major soil series of Penang Island
No Soil series (types) Texture Grain size Origin
1 Serong Clay Medium to coarse Acid sulfate soils
2 Selangor kangkong association Clay Medium Acid sulfate soils
3 Rengam Coarse sandy clay Fine to medium Soil developed on igneous and
high-grade metamorphic rocks
4 Kuala-kedah permatang association Clay Medium and coarse Developed on relatively flat areas
and mostly used for agricultural
purposes
5 Steep land Weathered materials, clay Fine to medium Weathering of granitic rocks
6 Rengam bukit temiang association Sandy clay Medium to coarse Soil developed on igneous and
high-grade metamorphic rocks
7 Local alluvium-colluvium association Loam Fine grained Very fine grained particles
developed on the river delta
8 Urban land Sand and gravel Fine to medium Weathering phenomena, mainly
composed of marine clay, sand,
and gravel, and abundant in the
shore line area
Courtesy: Guide to the soil series in peninsular Malaysia: 1993, Department of Agriculture, Malaysia
Fig. 3 Rainfall distribution for Peninsular Malaysia (courtesy:
Malaysian Meteorological Department)
Environ Earth Sci (2010) 60:1037–1054 1041
123

Fig. 4 Field photographs
illustrating the characteristics
and types of landslides:
alandslide scar on –45°slopes,
near Penang Hill. Arrows depict
the movement resulting risk to
the house (July 2006);
blandslides on a high slope near
Kampung Permatan Timbun
(November 2007); clandslide
scar near Kampung Bukit (June
2007); dlandslide scar near
Kampung Binjai (June 2007);
esmall debris slide along road,
near Pekan Balik Pulau
(November 2008); frock slide
on a steep slope near Balik
Pulau (November 2008)
Table 2 Spatial thematic layers used in the analysis
Classification GIS data type Scale or Resolution
Spatial database Data layers Spatial database Data layers Spatial database Data layers
Landslide Landslide ARC/INFO polygon coverage ARC/INFO GRID 1:25,000 10 m 910 m
Topographic map Slope ARC/INFO 1:25,000
Aspect Line and Point
Curvature Coverage
Drainage map Distance from drainage ARC/INFO
Line Coverage
Soil Map Types ARC/INFO 1:25,000
Polygon coverage
Geology Map Litho types ARC/INFO 1:63,300
Distance from lineaments Polygon, Line coverage
Land cover Land cover ARC/INFO 30 m 930m
GRID
NDVI NDVI ARC/INFO 10 m 910 m
GRID
Precipitation Precipitation GRID 10 m 910 m
1042 Environ Earth Sci (2010) 60:1037–1054
123

Remote sensing methods were used to obtain historical
records of the landslides. Archived 1:10,000–1:50,000
aerial photographs, SPOT 5 panchromatic satellite images,
and landslide reports over the past 21 years were used for
visual detection of landslide occurrences in the study area.
These aerial photographs were taken during the period
1981–2002 and were acquired by the Malaysian Remote
Sensing Agency. All historical landslide reports, newspa-
per records, and archived data were assembled for the
period under examination. The source material varies in
quality with regard to describing the precise location of the
landslide event. Based on site description, archived data-
base, and aerial photo interpretation, the locations of the
individual landslides were located on 1:25,000 maps and
the location plotted as closely as possible. Field observa-
tions were used to confirm fresh landslide scars. In the
aerial photographs and high-resolution satellite images
historical landslides could be observed as breaks in the
forest canopy, bare soil, or geomorphological features, for
example head and side scarps, flow tracks, and soil and
debris deposits below a scar. These landslides were then
classified and sorted on the basis of their modes of
occurrence. The landslide inventory map was very helpful
in understanding that different triggering factors control
different slope movement types. Most of the landslides are
shallow rotational and a few translational in type. The few
landslides that occurred in flat areas were not considered
and thus eliminated from the analysis. To assemble a
database to assess the surface area and number of land-
slides in the study area, a total of 463 landslides were
mapped in an area of 285 km
2
. The landslide inventory
map is shown in Fig. 5. Of the 463 landslides identified
from historical records in the 21 year period, most failures
(54%) were shallow rotational debris slides, 30% were
debris flows, and remaining 16% were rockfalls. The
landslide inventory map was used to evaluate the spatial
distribution of landslides in the area.
Nine factors were considered for the analysis, and the
factors were extracted from the constructed spatial data-
base. The list of all the data layers is illustrated in Fig. 6.
The factors were transformed into a vector-type spatial
database using the GIS, and landslide-related factors were
extracted using the database. The data used for the analysis
are shown in Table 2. A digital elevation model (DEM)
was created first from the topographic database. Contour
and survey base points that had elevation values from the
1:25,000-scale topographic maps were extracted, and a
DEM was constructed with a resolution of 10 m. Using this
DEM, the slope angle, slope aspect, and slope curvature
were calculated. In this study, substantial attention was
given to slope conditions. Slope configuration and steep-
ness play an important role in conjunction with lithology.
Maps have been produced showing slope steepness, slope
varies from 0 to 45°in the plain area to near vertical cliffs,
[45°, in the steep areas. The slope map was reclassified
into four classes following the standard classification
scheme of the Ministry of Science, Technology and
Environment Malaysia for hill land: class 1, \15°, class 2,
16–25°, class 3, 26–35°, and class, 4 [35°(Fig. 6a). In the
case of the aspect layer, eight directions are shown for the
different directions of slope (Fig. 6b). In the case of the
curvature, negative curvatures represent concave, zero
curvature represents flat, and positive curvatures represents
convex. The curvature map was prepared using the avenue
routine in ArcView 3.2 (Fig. 6c). In addition, the distance
from drainage was calculated using the topographic data-
base. The drainage buffer was recalculated and classified
into ten equal area classes (Fig. 6d). Using the geology
database, the lithology was extracted, and the distance from
lineament was calculated. The lithology map was obtained
from a 1:63,300-scale geological map (Fig. 6e), and the
distance from lineament map was recalculated into ten
equal area classes (Fig. 6f). The soil layer was digitised
from a Department of Irrigation and Drainage soil hard
copy map. Eight classes of soil types were extracted as
shown in Fig. 6g. A land-cover map was prepared using
SPOT 5 images (10 m spatial resolution) using the
Fig. 5 Inventory map of the study area: acentral part of Penang
island; bNorth western part of Penang Island
Environ Earth Sci (2010) 60:1037–1054 1043
123

supervised classification method and verified by field sur-
vey. The SPOT 5 scene of January 2005 was classified to
map the different land-cover classes. Nine classes of
landcover types were identified, for example urban, water,
forest, agricultural area, tin mines, rubber, and palm oil
plantation were extracted for land cover mapping (Fig. 6h).
Fig. 6 Input thematic layers:
aslope; baspect; ccurvature;
ddistance from drainage;
egeology; fdistance from
lineament; gsoil; hland cover;
ivegetation index (NDVI); and
jamount of precipitation
1044 Environ Earth Sci (2010) 60:1037–1054
123

Finally, the normalized difference vegetation index
(NDVI) map was obtained from Landsat satellite image
(Fig. 6i). The NDVI value was calculated using the for-
mula NDVI ¼IR RðÞ=IR þRðÞ;where RI value is
the infrared portion of the electromagnetic spectrum, and R
value is the red portion of the electromagnetic spectrum.
The NDVI value denotes areas of vegetation in an image.
In many climatic environments, for example the tropical
zone, the spatial and topographic distribution of precipi-
tation, especially rainfall intensity, is very difficult to map
or to model accurately. High intensity rainfall, which is
often the most efficient in terms of landslide triggering,
results from localized storm or accumulated monsoon
rainfall covering small areas of only a few square km. The
recording raingauge network is generally not dense enough
to record all these storms and localized precipitation and
therefore it is again necessary to take a probabilistic
approach based on the records of a few stations. The
precipitation map was prepared using the last 26 years
(1981–2007) of historical rainfall data. In the study area,
there are only two raingauge stations (Fig. 6j). So data used
in this study are collected from these stations and a sta-
tistical distribution of the accumulated average precipita-
tion was prepared in a GIS.
Landslide hazard analysis
In this study, frequency ratio, logistic regression, and
neural network models were used for landslide hazard
analysis using nine thematic layers.
Application of frequency ratio model
Frequency ratio approaches are based on the observed
relationships between distribution of landslides and each
landslide-related factor, to reveal the correlation between
landslide locations and the factors in the study area. Using
the frequency ratio model, the spatial relationships between
landslide-occurrence location and factors contributing to
landslide occurrence were derived. The frequency is cal-
culated from analysis of the relationship between land-
slides and the attribute factors. Therefore, the frequency
ratios of each factor’s type or range were calculated from
their relationship with landslide events as shown in
Table 3. In the relationship analysis, the ratio is that of the
area where landslides occurred to the total area, so that a
value of 1 is an average value. If the value is greater than 1,
it means a higher correlation, and a value lower than 1
means lower correlation.
To calculate the landslide hazard index (LHI), each
factor’s frequency ratio values were summed to the training
area in the neural network model as different training sites.
The landslide hazard value represents the relative hazard to
landslide occurrence. So the greater the frequency ratio
value, the higher the hazard to landslide occurrence and the
lower the value, the lower the hazard to landslide
occurrence.
LHI ¼Fr1þFr2þþFrnð3Þ
LHI: landslide hazard index; Fr: rating of each factor’s type
or range)
Application of logistic regression model
Logistic regression allows one to form a multivariate
regression relationship between a dependent variable and
several independent variables. Logistic regression, which is
a multivariate analysis model, is useful for predicting the
presence or absence of a characteristic or outcome based on
values of a set of predictor variables. The advantage of
logistic regression is that, through addition of an appro-
priate link function to the usual linear regression model, the
variables may be either continuous or discrete, or any
combination of both types, and they do not necessarily
have normal distributions. In the case of multi-regression
analysis, the factors must be numerical, and in the case of
discriminant analysis, the variables must have a normal
distribution. In the current situation, the dependent variable
is a binary variable representing presence or absence of
landslide. Where the dependent variable is binary, the
logistic link function is applicable (Atkinson and Massari
1998). For this study, the dependent variable must be input
as either 0 or 1, so the model applies well to landslide
possibility analysis. Logistic regression coefficients can be
used to estimate ratios for each of the independent vari-
ables in the model.
Quantitatively, the relationship between the occurrence
and its dependency on several variables can be expressed
as:
p¼1
1þezð4Þ
where pis the probability of an event occurring. In this
situation, the value pis the estimated probability of
landslide occurrence. The probability varies from 0 to 1 on
an S-shaped curve and zis the linear combination. It
follows that logistic regression involves fitting an equation
of the following form to the data:
z¼b0þb1x1þb2x2þþbnxnð5Þ
where b
0
is the intercept of the model, the
bii¼0;1;2;...;nðÞare the slope coefficients of the
logistic regression model, and the xiði¼0;1;2;...;nÞ
are the independent variables. The linear model formed is
then a logistic regression of the presence or absence of
Environ Earth Sci (2010) 60:1037–1054 1045
123

Table 3 Frequency ratio and logistic regression coefficient values for causative factors
Factor Class Pixels in
domain
Pixel, %
a
Landslide
occurrence points
Landslide occurrence
points, %
b
Frequency
ratio
b/a
Coefficients of
logistic regression
Slope 0–15°1709800 57.87 53 11.45 0.20 0.0554
16–25°765189 25.90 152 32.83 1.27
26–35°360229 12.19 157 33.91 2.78
[35°119564 4.05 101 21.81 5.39
Aspect Flat 1199400 40.59 13 2.80 0.07 -1.7056
North 206629 6.99 41 8.85 1.27 -0.0727
Northeast 207860 7.03 51 11.01 1.57 0.1133
East 228674 7.74 60 12.95 1.67 0.1479
Southeast 236988 8.02 82 17.71 2.21 0.4637
South 205108 6.94 58 12.53 1.80 0.2584
Southwest 206970 7.01 52 11.23 1.60 0.1679
West 228117 7.72 54 11.66 1.51 0.0322
Northwest 235036 7.95 52 11.23 1.41 0.0000
Curvature Concave 770757 26.09 50 10.80 0.41 -0.0001
Flat 1419529 48.04 45 9.72 0.20
Convex 764496 25.87 368 79.48 3.07
Distance from
drainage
0–14 m 325460 11.01 21 4.54 0.41 0.0009
20–36 m 347537 11.76 43 9.29 0.79
40–56 m 298382 10.10 61 13.17 1.30
58–76 m 285453 9.66 58 12.53 1.30
78–100 m 310971 10.52 48 10.37 0.99
101–130 m 296818 10.05 52 11.23 1.12
131–169 m 273396 9.25 48 10.37 1.12
170–222 m 275609 9.33 49 10.58 1.13
223–331 m 272270 9.21 55 11.88 1.29
332–2,064 m 268886 9.10 28 6.05 0.66
Geology Granite 2195706 76.65 461 99.57 2.29 1.0542
Marine clay, sand and gravel 668834 23.35 2 0.43 0.02 -1.8516
Soil Rengam-bukit temiang
association
289450 10.03 96 20.73 2.07 10.9673
Selangor-kangkong
association
34197 1.18 0 0.00 0.00 0.6377
Local alluvium-colluvium
association
373655 12.94 13 2.81 0.22 10.0145
Serong series 80436 2.79 0 0.00 0.00 0.5953
Steep land 1506818 52.20 341 73.65 1.41 10.1995
Kuala kedah-permatang
association
187057 6.48 0 0.00 0.00 2.7604
Urban land 413813 14.33 13 2.81 0.20 9.9792
Rengam 1329 0.05 0 0.00 0.00 0.0000
Distance from
lineament
0–89 m 297410 10.07 45 9.72 0.97 0.0001
90–180 m 307232 10.40 48 10.37 1.00
181–275 m 293932 9.95 62 13.39 1.35
276–377 m 294078 9.95 63 13.61 1.37
378–494 m 294927 9.98 54 11.66 1.17
495–640 m 294365 9.96 54 11.66 1.17
641–841 m 294059 9.95 74 15.98 1.61
842–1,150 m 292980 9.92 50 10.80 1.09
1,151–1,777 m 293193 9.92 13 2.81 0.28
1,778–5,317 m 292606 9.90 0 0.00 0.00
1046 Environ Earth Sci (2010) 60:1037–1054
123

landslides (present conditions) on the independent vari-
ables (pre-failure conditions).
Using the logistic regression model, the spatial relationship
between landslide-occurrence and factors affecting landslides
was assessed. The spatial databases of each factor were con-
verted to ASCII format files to be used in the statistical
package SPSS. The correlations between landslide and each
factor were calculated. In this analysis, the ‘‘continuous data’’
such as slope, curvature, distance from drainage, distance
from lineament, NDVI, and precipitation were treated as
‘‘scale’’ in SPSS whereas aspect, land cover, geology, and soil
layer were taken as ‘‘nominal’’ data. The logistic regression
coefficientfor each of the thematic layers was computed and is
shown in Table 3, where each of the continuous thematic
layers have only single coefficient values and, on the other
hand, nominal or discrete thematic layers show coefficients
for each discrete class.Finally, the probability that predicts the
possibility of landslide occurrence was calculated using the
spatial database, data from Table 3, and Eqs. 4and 5.How-
ever, in the second case, all factors were used and logistic
regression mathematical equations were formulated as shown
in Eqs. 5and 6for each case.
Zn¼ð0:554 SLOPEÞþASPECTcþð0:0001
CURVATUREÞþð0:0009 DRAINAGEÞ
þLITHOLOGYcþð0:0001 LINEAMENTÞ
þð0:0168 NDVIÞþSOILcþLANDCOVERc
þð0:0052 PRECIPITATIONÞ34:4228
ð6Þ
(where SLOPE is slope value; CURVATURE is curvature
value; DRAINAGE is distance from drainage value, LIN-
EAMENT is distance from lineament value, NDVI is NDVI
value, and ASPECT
c
, LITHOLOGY
c
, SOIL
c
, LAND-
COVER
c
and PRECIPITATION
c
are logistic regression
coefficient values listed in Table 3and z
n
is a parameter).
Application of the neural network model
An artificial neural network is a ‘‘computational mecha-
nism able to acquire, represent, and compute a mapping
from one multivariate space of information to another,
given a set of data representing that mapping’’ (Paola and
Schowengerdt 1995). An ANN is composed of a set of
Table 3 continued
Factor Class Pixels in
domain
Pixel, %
a
Landslide
occurrence points
Landslide occurrence
points, %
b
Frequency
ratio
b/a
Coefficients of
logistic regression
Land cover Water Body 73106 2.49 1 0.23 0.09 0.0000.
Settlement 678841 23.15 45 10.27 0.44 0.9230
Forest 1468084 50.07 295 67.35 1.35 0.4509
Urban 205767 7.02 17 3.88 0.55 0.6901
Bare Land 60200 2.05 24 5.48 2.67 0.7638
Agriculture 446124 15.22 56 12.79 0.84 0.9315
NDVI -73–18 291092 10.05 33 7.13 0.71 -0.0168
-17–1 300254 10.37 31 6.70 0.65
2–21 297248 10.26 47 10.15 0.99
22–32 315879 10.91 48 10.37 0.95
33–37 358384 12.37 44 9.50 0.77
38–40 322673 11.14 61 13.17 1.18
41–43 373180 12.89 57 12.31 0.96
44–45 226395 7.82 54 11.66 1.49
46–48 242836 8.38 45 9.72 1.16
49–61 168249 5.81 43 9.29 1.60
Precipitation 2,613–2,651 mm 310554 10.51 39 8.42 0.80 0.0052
2,652–1,676 mm 305133 10.33 13 2.81 0.27
2,677–2,695 mm 298684 10.11 31 6.70 0.66
2,696–2,707 mm 298405 10.10 24 5.18 0.51
2,708–2,718 mm 292410 9.90 49 10.58 1.07
2,719–2,730 mm 292990 9.92 44 9.50 0.96
2,731–2,742 mm 293306 9.93 41 8.86 0.89
2,743–2,753 mm 293819 9.94 73 15.77 1.59
2,754–2,763 mm 293702 9.94 68 14.69 1.48
2,764–2,772 mm 275779 9.33 81 17.49 1.87
Environ Earth Sci (2010) 60:1037–1054 1047
123

nodes and a number of interconnected processing elements.
ANNs use learning algorithms to model knowledge and
save this knowledge in weighted connections, mimicking
the function of a human brain (Turban and Aronson 2001).
The most popular ANN model used in prediction and
regression tasks is multi-layer perceptron (MLP) with the
feed-forward back-error propagation (BP) type of learning
algorithm, or simply as MLP-BP. In this study, the
back-propagation learning algorithm (MLP) having a
multi-layered neural network is used. A three-layered
interconnected neural network (Fig. 7) consisting of input
layer, hidden layer, and output layer was used. Landslide-
related causative layers were fed into the neural network as
‘‘input nodes’’ that are regarded as independent variables
and the weightage computation of the input data was pro-
cessed in the ‘‘hidden layer’’. Input nodes receive the
variables independently while hidden nodes run the learn-
ing process before passing on to output node (a dependent
variable) for prediction. Each network carries weights
ðw1;w2;w3Þ. These weighted connections act as coeffi-
cients to the input nodes. The hidden nodes (sum of the
weighted inputs) compute network output (feed-forward)
through a non-linear activation function, determining the
difference (error) to the expected output (actual output).
This error is distributed among the connections weights
(back-propagation) in order to progressively reduce the
error. In the network the hidden and output layer neurons
process their inputs by multiplying each input by a corre-
sponding weight, summing the product, and then process-
ing the sum using a nonlinear transfer function to produce a
result. There are two stages involved in using neural net-
works for multi-source classification: the training stage, in
which the internal weights are adjusted; and the classifying
stage. Typically, the back-propagation algorithm trains the
network until some targeted minimal error is achieved
between the desired and actual output values of the net-
work. When the training is complete, the network is used as
a feed-forward structure to produce a classification for the
entire data (Paola and Schowengerdt 1995).
Before running the artificial neural network program, the
training site should be selected. So, the landslide-prone
(occurrence) area and the landslide-not-prone area were
selected as training sites. Cells from each of the two classes
were randomly selected as training cells, with 463 cells
denoting areas where a landslide had not occurred or had
occurred. First, areas where a landslide had not occurred
were classified as ‘‘areas not prone to landslide’’ and areas
where a landslide was known to exist were assigned to an
‘‘areas prone to landslide’’ training set. The back-propaga-
tion algorithm was then used to calculate the weights
between the input layer and the hidden layer, and between
the hidden layer and the output layer, by modifying the
Landslide non
prone area
Landslide
prone area
GIS layers Hidden Input Output
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
Curvature
Lithology
Distance from
lineaments
Land cover
Topographic
Map
Geology Map
Land Cover
Types
Drainage Map Distance from
drainage
NDVI Map
Soil Map
Aspect
Preci
p
itation
Precipitation Map
NDVI
Slope
As
p
ect
Fig. 7 Three layered
architecture of the neural
network
1048 Environ Earth Sci (2010) 60:1037–1054
123

number of hidden nodes and the learning rate. Three-layered
feed-forward network was implemented using the MAT-
LAB software package. Here, ‘‘feed-forward’’ denotes that
the interconnections between the layers propagate forward
to the next layer. The number of hidden layers and the
number of nodes in a hidden layer required for a particular
classification problem are not easy to deduce. In this study, a
9ðinput layersÞ20 ðhidden layers) 2ðoutput layers)
structure was selected for the network, with input data nor-
malized in the range 0.1–0.9. The nominal and interval class
group data were converted to continuous values ranging
between 0.1 and 0.9. Therefore, the continuous values were
not ordinal data, but nominal data, and the numbers denote
the classification of the input data.
The learning rate was set to 0.01, and the initial weights
were randomly selected to values between 0.1 and 0.3. The
weights calculated from ten test cases were compared to
determine whether the variation in the final weights was
dependent on the selection of the initial weights. The back-
propagation algorithm was used to minimize the error
between the predicted output values and the calculated
output values. The algorithm propagated the error back-
wards, and iteratively adjusted the weights. The number of
epochs was set to 2,000, and the root mean square error
(RMSE) value used for the stopping criterion was set to
0.01. Most of the training data sets met the 0.01 RMSE
goal. However, if the RMSE value was not achieved, the
maximum number of iterations was terminated at 2,000
epochs. When the latter case occurred, the maximum
RMSE value was 0.051. The final weights between layers
acquired during training of the neural network and the
contribution or importance of each of the nine factors used
to predict landslide hazard are shown in Table 4.
For easy interpretation, the average values were calcu-
lated, and these values were divided by the average of the
weights of the same factor that had a minimum value. The
distance from drainage value was the minimum value,
0.0767, and the slope value was the maximum value,
0.2079. Finally, the weights were applied to the entire
study area and the landslide hazard map was created
(Fig. 8). The values were classified by equal areas and
grouped into four classes for visual interpretation. The
hazard index was classified into four classes (highest 10%,
Table 4 Weights of each factor estimated by the neural network
considered in this study
Factor Weight Normalized
weight
Slope (°) 0.2079 1.00
Aspect (°) 0.1123 0.27
Curvature 0.0935 0.13
Distance from drainage (m) 0.0971 0.16
Geology 0.0788 0.02
Distance from lineament (m) 0.0767 0.00
Soil 0.1021 0.19
Land cover 0.0865 0.07
NDVI 0.1450 0.52
Precipitation (mm) 0.1320 0.42
Fig. 8 Landslide hazard maps draped over the hill shaded maps for
the study area: alandslide hazard map using frequency ratio model;
blandslide hazard map using logistic regression model; clandslide
hazard map using the artificial neural network model
Environ Earth Sci (2010) 60:1037–1054 1049
123

second 10%, third 20% and remaining 60%) based on area
for visual and easy interpretation.
Verification of the hazard maps
The landslide hazard analysis was performed using the
frequency ratio, logistic regression, and artificial neural
network models, and the analysis results were verified
using the landslide locations for the study area; the result of
this is shown in Fig. 9. Two basic assumptions are needed
to verify the landslide hazard calculation methods. One is
that landslides are related to spatial information, such as
slope, aspect, curvature, distance from drainage, distance
from lineaments, geology, land cover, NDVI, and soil; the
other is that future landslides will be precipitated by a
specific impact factor such as rainfall (Chung and Fabbri
1999). In this study, the two assumptions are satisfied
because the landslides are related to the spatial information
and the landslides were precipitated by heavy rainfall in the
study area.
The verification method was performed by comparison
of existing landslide data and landslide hazard analysis
results. The comparison results shown in Table 5and as a
line graph in Fig. 9illustrate how well the estimators
perform with respect to the landslides used in constructing
those estimators. Verification of the success rate is based
on the landslide susceptibility analysis result in the Penang
Island area, using the landslide occurrence locations, for
the three kinds of analysis methods—frequency ratio,
logistic regression, and artificial neural network models.
The rate curves were created and the ‘‘areas under the
curves’’ were calculated for all three cases of hazard maps
using the existing landslide location data. The ‘‘areas under
the curves’’ constitutes one of the most commonly used
accuracy statistics for the prediction models in natural
hazard assessments (Begueria 2006). The rate explains how
well the model and factor predict the landslide (Chung and
Fabbri 1999). So, the area under the curve can be used to
assess the prediction accuracy qualitatively. To obtain the
relative ranks for each prediction pattern, the calculated
index values of all cells in the study area were sorted in
descending order. The ordered cell values were then divi-
ded into 100 classes and set on the y-axis, with accumu-
lated 1% intervals on the x-axis. The rate verification
results appear as a line in Fig. 9and in Table 5. For
example, in the case of the frequency ratio model used, 90–
100% (10%) class of the study area where the landslide
hazard index had a higher rank could explain 53% of all the
landslides in the success rate and were classified as ‘‘very
highly hazardous’’ zone (Table 5). The next 80–100%
(20%) class of the study area where the landslide hazard
index had a higher rank could explain 74% of the land-
slides in the success rate and were classified as ‘‘highly
hazardous’’ zone. Similarly, the 60–100% (40%) class of
the study area where the landslide hazard index had a
relatively lower rank could explain 93% of the landslides in
the success rate and were classified as ‘‘moderately haz-
ardous’’ zone. Finally, the remaining 0–100% (60%) class
of the study area where the landslide hazard index had a
low rank could explain 100% of the landslides were clas-
sified as ‘‘not hazardous’’ zone. The same procedure was
adopted for classification and verification of the hazard
maps obtained through logistic regression (Table 5, col-
umn 2) and neural network (Table 5, column 3) models. To
compare the result quantitatively, the areas under the curve
were re-calculated as the total area is 1 which means per-
fect prediction accuracy. So, the area under a curve can be
0
10
20
30
40
50
60
70
80
90
100
100 ~ 100
100 ~ 90
100 ~ 80
100 ~ 70
100 ~ 60
100 ~ 50
100 ~ 40
100 ~ 30
100 ~ 20
100 ~ 10
100 ~ 0
Landslide Hazard Index Rank (%)
Cumul ative p ercentage of landslide occurence
Frequency Ratio (A UC=0.8641)
Logist ic Regressi on (AUC=0. 8959)
Neural Network (A UC=0.8355)
Reference line (AUC = 0.5)
Fig. 9 Prediction accuracy
assessment and success rate
curve for three types of hazard
maps
1050 Environ Earth Sci (2010) 60:1037–1054
123

Table 5 Verification and success rate for study area
Range Success rate curve
(frequency ratio)
Success rate
curve (logistic
regression)
Success rate curve
(neural network)
100–100 0 0 0
100–99 6 10 4
100–98 15 22 10
100–97 25 31 19
100–96 32 37 25
100–95 38 45 30
100–94 39 49 34
100–93 43 53 38
100–92 48 56 41
100–91 50 58 44
100–90 53 61 46
100–89 56 63 49
100–88 59 65 52
100–87 61 69 56
100–86 63 72 59
100–85 65 75 60
100–84 67 77 62
100–83 69 77 63
100–82 71 79 64
100–81 72 81 65
100–80 74 82 67
100–79 76 84 69
100–78 77 86 69
100–77 79 87 70
100–76 79 88 73
100–75 80 89 73
100–74 82 89 75
100–73 83 90 77
100–72 85 90 78
100–71 85 91 78
100–70 86 92 79
100–69 86 92 81
100–68 87 93 81
100–67 88 93 82
100–66 89 93 83
100–65 89 94 85
100–64 90 94 85
100–63 91 95 86
100–62 92 95 87
100–61 93 95 87
100–60 93 96 88
100–59 93 97 89
100–58 93 97 90
100–57 93 97 92
100–56 94 97 92
100–55 94 97 92
100–54 94 97 93
Table 5 continued
Range Success rate curve
(frequency ratio)
Success rate
curve (logistic
regression)
Success rate curve
(neural network)
100–53 94 98 94
100–52 94 98 94
100–51 94 98 94
100–50 94 99 95
100–49 94 99 95
100–48 94 99 95
100–47 100 99 95
100–46 100 99 96
100–45 100 100 96
100–44 100 100 96
100–43 100 100 97
100–42 100 100 97
100–41 100 100 97
100–40 100 100 98
100–39 100 100 98
100–38 100 100 99
100–37 100 100 99
100–36 100 100 99
100–35 100 100 99
100–34 100 100 100
100–33 100 100 100
100–32 100 100 100
100–31 100 100 100
100–30 100 100 100
100–29 100 100 100
100–28 100 100 100
100–27 100 100 100
100–26 100 100 100
100–25 100 100 100
100–24 100 100 100
100–23 100 100 100
100–22 100 100 100
100–21 100 100 100
100–20 100 100 100
100–19 100 100 100
100–18 100 100 100
100–17 100 100 100
100–16 100 100 100
100–15 100 100 100
100–14 100 100 100
100–13 100 100 100
100–12 100 100 100
100–11 100 100 100
100–10 100 100 100
100–9 100 100 100
100–8 100 100 100
100–7 100 100 100
Environ Earth Sci (2010) 60:1037–1054 1051
123

used to assess the prediction accuracy qualitatively, as
shown in Fig. 9.
From Fig. 9, verification results show that in the fre-
quency ratio case, the area ratio was 0.8641 and the pre-
diction accuracy was 86.41%. In the logistic regression
case, the prediction accuracy was 89.59%. In the neural
network case, the area ratio was 0.8355 and the prediction
accuracy was 83.55%. So from the success rate graphs
(Fig. 9), it is quite evident that, logistic regression has the
best prediction accuracy of 89.59%, whereas the neural
network has the worst accuracy of 83.55%, with a differ-
ence of about 6%.
Although, for the first seven classes (30–100%), the
logistic regression model is a little better than those from
the frequency ratio and neural network models, except for
the remainder of the classes (70–100%), the logistic
regression model produced somewhat similar results to
those from the frequency ratio and neural network models.
The artificial neural network model was worse than the
frequency ratio and logistic regression models in all classes.
Summary and conclusion
Landslides are a significant constraint to development in
Malaysia, notably through reactivation of ancient land-
slides. The fast growth of urbanisation and the limited
space on Penang Island have created a need to construct
buildings on high-slope areas and debris fans. A series of
Government funded research projects has provided much
background information and identified suitable methods for
use of landslide hazard information in land-use planning.
However, a number of significant problems remain over the
use of this information. This study establishes a multi-
model procedure for evaluation of the landslides-hazard on
a medium scale. The use of frequency ratio, logistic
regression, and neural network models enabled simulation
of various landslide hazard maps using GIS tools and
remote sensing data.
An artificial neural network was used to estimate areas
hazardous to landslides using a spatial database for Penang
Island. Three hazard maps were prepared using frequency
ratio, logistic regression, and artificial neural network
models. The motive was to quantify the best hazard map
for the study area by assessment of various modelling
approaches. The results from use of the logistic regression
model were most accurate and were better than those from
the other models. Therefore, the hazard map produced
using the logistic regression model is found to be more
realistic. Results obtained using the artificial neural net-
work model were the worst. For application of the artificial
neural network, the relative weights for the various input
factors were calculated. The results (Table 4) show that the
slope is the most important factor; it has a weightage more
than twice those of the other factors. The average nor-
malized value for the nine factors (Table 4) shows slope
has the highest weightage of 1. NDVI is also important,
with a weightage of 0.52, and then precipitation, with a
weightage of 0.42. Distance from lineament has the lowest
weightage value (0). This shows that in the study area fault
lines and structural features do not contribute much to
landslide triggering. In addition to this, the factors found to
be important for their contribution to landsliding in the
neural network method are in agreement with field
observations.
The back-propagation training algorithm presents few
difficulties when trying to follow the internal processes of
the procedure. The method also involves a long execution
time and has a heavy computing load compared with the
other two models used in this analysis. Therefore, the
thematic data layers were converted into ASCII format to
increase the speed of the computing process. Computation
of the weightage of the factors and artificial neural network
modelling was performed in MATLAB and later the out-
puts were exported to GIS for map production and visual
interpretation. Landslide hazard maps were analyzed
qualitatively using equal area classification schemes. These
results can be used as basic data to assist general slope
management and land-use planning. This methodology
shows that the uses of GIS techniques can improve multi
hazard prediction models. The methods used in the study
can also be used for generalized planning and assessment
purposes, although they may be less useful on the site-
specific scale, where local geological and geographic het-
erogeneities may prevail. Because of the simplicity of the
models used, the current approach can be applied to hazard
assessment at a much larger areas, if they are homogenous.
Acknowledgments B. Pradhan would like to thank the Alexander
von Humboldt Foundation (AvH), Germany for awarding a visiting
scientist position and adequate funds to carry out research at Dresden
University of Technology, Germany. Thanks are due to anonymous
Table 5 continued
Range Success rate curve
(frequency ratio)
Success rate
curve (logistic
regression)
Success rate curve
(neural network)
100–6 100 100 100
100–5 100 100 100
100–4 100 100 100
100–3 100 100 100
100–2 100 100 100
100–1 100 100 100
100–0 100 100 100
1052 Environ Earth Sci (2010) 60:1037–1054
123

reviewers for their critical and valuable comments that helped to bring
the manuscript into the present form.
References
Ahmad F, Yahaya AS, Farooqi MA (2006) Characterization and
geotechnical properties of Penang residual soils with emphasis
on landslides. Am J Environ Sci 2(4):121–128
Akgul A, Bulut F (2007) GIS-based landslide susceptibility for
Arsin-Yomra (Trabzon, North Turkey) region. Environ Geol
51(8):1377–1387
Akgun A, Dag S, Bulut F (2008) Landslide susceptibility mapping for
a landslide-prone area (Findikli, NE of Turkey) by likelihood-
frequency ratio and weighted linear combination models.
Environ Geol 54(6):1127–1143(17)
Atkinson PM, Massari R (1998) Generalized linear modeling of
susceptibility to landsliding in the central Apennines. Italy
Comput Geosci 24(4):373–385
Begueria S (2006) Validation and evaluation of predictive models in
hazard assessment and risk management. Nat Hazards 37:315–329
Carro M, De Amicis M, Luzi L, Marzorati S (2003) The application
of predictive modeling techniques to landslides induced by
earthquakes: the case study of the 26 September 1997 Umbria-
Marche earthquake (Italy). Eng Geol 1(2):139–159
Chan NW (1998a) Environmental hazards associated with hill land
development in Penang Island, Malaysia: some recommenda-
tions on effective management. Disaster Prev Manage Int J
7(4):305–318
Chan NW (1998) Responding to landslide hazards in rapidly
developing Malaysia: a case of economics versus environmental
protection. Disaster Prev Manage Int J 7(1)
Cheng TA, Lateh H, Peng KS (2008) Intelligence explanation system
on landslide dissemination: a case study in Malaysia. In:
Proceddings of the first world landslide forum report: Imple-
menting the 2006 Tokyo action plan on the international
program on landslides (IPL). 330–333
Choi J, Oh HJ, Won JS, Lee S (2009) Validation of an artificial neural
network model for landslide susceptibility mapping. Environ
Earth Sci. doi:10.1007/s12665-009-0188-0
Chung CF, Fabbri AG (1999) Probabilistic prediction models for
landslide hazard mapping. Photogrammetric Eng Remote Sens
65(12):1389–1399
Clerici A, Perego S, Tellini C, Vescovi PA (2002) Procedure for
landslide susceptibility zonation by the conditional analysis
method. Geomorphology 48:349–364
Dahal RK, Hasegawa S, Nonomura S, Yamanaka M, Masuda T,
Nishino K (2008) GIS-based weights-of-evidence modelling of
rainfall-induced landslides in small catchments for landslide
susceptibility mapping. Environ Geol 54(2):314–324
Dai FC, Lee CF (2002) Landslide characteristics and slope instability
modeling using GIS, Lantau Island, Hong Kong. Geomorphol-
ogy 42:213–228
Ercanoglu M, Gokceoglu C (2002) Assessment of landslide suscep-
tibility for a landslide-prone area (north of Yenice, NW Turkye)
by fuzzy approach. Environ Geol 41:720–730
Gokceoglu C, Sonmez H, Ercanoglu M (2000) Discontinuity
controlled probabilistic slope failure risk maps of the Altindag
(settlement) region in Turkey. Eng Geol 55:277–296
Guzzetti F, Carrarra 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
Lamelas MT, Marinoni O, Hoppe A, Riva J (2008) Doline probability
map using logistic regression and GIS technology in the central
Ebro Basin (Spain). Environ Geol 54(5):963–977
Lee S (2005) Application of logistic regression model and its
validation for landslide susceptibility mapping using GIS and
remote sensing data. Int J Remote Sens 26:1477–1491
Lee S (2007) Application and verification of fuzzy algebraic operators
to landslide susceptibility mapping. Environ Geol 52:615–623
Lee S, Dan NT (2005) Probabilistic landslide susceptibility mapping
in the Lai Chau province of Vietnam: focus on the relationship
between tectonic fractures and landslides. Environ Geol 48:778–
787
Lee S, Pradhan B (2006) Probabilistic landslide risk mapping at
Penang Island. Malaysia J Earth Syst Sci 115(6):1–12
Lee S, Pradhan B (2007) Landslide hazard mapping at Selangor,
Malaysia using frequency ratio and logistic regression models.
Landslides 4:33–41
Lee S, Talib JA (2005) Probabilistic landslide susceptibility and
factor effect analysis. Environ Geol 47:982–990
Lloyd DM, Othman A, Wilkinson PL, Anderson MG (2001)
Predicting landslides: Assessment of an automated rainfall based
landslide warning system. In: K.K.S.Ho and K.S.Li, Geotechni-
cal Engineering—Meeting Society’s Needs, Balkema, 1, 135–
139
Ong WS (1993) The geology and engineering geology of Penang
Island. Geological Survey of Malaysia, Malaysia
Ooi LH (1999) Rockfall protection: Technical talks, Kuala Lumpur,
28 Oct. 1999, Geological Society of Malaysia, pp 20–32
Paola JD, Schowengerdt RA (1995) A review and analysis of
backpropagation neural networks for classificatioin of remotely
sensed multi-spectral imagery. Int J Remote Sens 16:3033–3058
Pradhan B, Lee S (2008a) Utilization of optical remote sensing data
and GIS tools for regional landslide hazard analysis by using an
artificial neural network model at Selangor. Malaysia Earth Sci
Front 14(6):143–152
Pradhan B, Lee S (2008b) Landslide risk analysis using artificial
neural network model focusing on different training sites. Int J
Phys Sci 3(11):1–15
Pradhan B, Singh RP, Buchroithner MF (2006) Estimation of stress
and its use in evaluation of landslide prone regions using remote
sensing data. Adv Space Res 37:698–709
Pradhan B, Lee S, Mansor S, Buchroithner MF, Jallaluddin N (2008)
Utilization of optical remote sensing data and geographic
information system tools for regional landslide hazard analysis
by using binomial logistic regression model. Appl Remote Sens
2:1–11
Pradhan B, Lee S, Buchroithner MF (2009) Use of geospatial data for
the development of fuzzy algebraic operators to landslide hazard
mapping: a case study in Malaysia. Appl Geomatics 1:3–15
Refice A, Capolongo D (2002) Probabilistic modeling of uncertainties
in earthquake-induced landslide hazard assessment. Comput
Geosci 28:735–749
Romeo R (2000) Seismically induced landslide displacements:
a predictive model. Eng Geol 58:337–351
Sassa K (2008) Land-use and Landslides – Human-Induced Disasters.
Proceddings of the first world landslide forum report: imple-
menting the 2006 Tokyo action plan on the international
program on landslides (IPL), pp 1–20
Sin HT, Chan NW (2004) The urban heat island phenomenon in
Penang Island: Some observations during the wet and dry season.
In: Jamaluddin Md. Jahi, Kadir Arifin, Salmijah Surif and
Shaharudin Idrus (eds.). Proceedings 2nd. Bangi World Confer-
ence on Environmental Management. Facing Changing Condi-
tions. 13–14 September 2004, Bangi, Malaysia, pp 504–516
Environ Earth Sci (2010) 60:1037–1054 1053
123

Su
¨zen ML, Doyuran VA (2004) Comparison of the GIS based
landslide susceptibility assessment methods: multivariate versus
bivariate. Environ Geol 45:665–679
Tan BK (1990) Studies on the characteristics of soils and rocks and
slope stability in urban areas of Penang island. Research Report,
July 1990, Univ. Kebangsaan Malaysia, Bangi, (in Malay). 1990.
80–86
Toh CT (1999) Influence of geology and geological structures on cut
slope stability: Technical Talks, Kuala Lumpur, 28 Oct 1999,
Geological Society of Malaysia, pp 12–25
Tunusluoglu MC, Gokceoglu C, Nefeslioglu HA, Sonmez H (2008)
Extraction of potential debris source areas by logistic regression
technique: a case study from Barla, Besparmak and Kapi
mountains (NW Taurids, Turkey). Environ Geol 54(1):9–22
Turban E, Aronson JE (2001) Decision Support Systems and
Intelligent Systems. Prentice Hall, New Jersey
Varnes DJ (1984) Landslide hazard zonation: a review of principles
and practice. Nat Hazards 3:63
Wang HB, Sassa K (2005) Comparative evaluation of landslide
susceptibility in Minamata area. Jpn Environ Geol 47:956–966
Youssef AM, Pradhan B, Gaber AFD, Buchroithner MF (2009)
Geomorphological Hazard Analysis along the Egyptian Red Sea
Coast between Safaga and Quseir. Natural Hazards and Earth
System Science, pp 751–766
Zhou G, Esaki T, Mitani Y, Xie M, Mori J (2003) Spatial probabilistic
modeling of slope failure using an integrated GIS Monte Carlo
simulation approach. Eng Geol 68:373–386
1054 Environ Earth Sci (2010) 60:1037–1054
123